Check in LLVM r95781.
diff --git a/lib/Analysis/AliasAnalysis.cpp b/lib/Analysis/AliasAnalysis.cpp
new file mode 100644
index 0000000..371dcaf
--- /dev/null
+++ b/lib/Analysis/AliasAnalysis.cpp
@@ -0,0 +1,249 @@
+//===- AliasAnalysis.cpp - Generic Alias Analysis Interface Implementation -==//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements the generic AliasAnalysis interface which is used as the
+// common interface used by all clients and implementations of alias analysis.
+//
+// This file also implements the default version of the AliasAnalysis interface
+// that is to be used when no other implementation is specified. This does some
+// simple tests that detect obvious cases: two different global pointers cannot
+// alias, a global cannot alias a malloc, two different mallocs cannot alias,
+// etc.
+//
+// This alias analysis implementation really isn't very good for anything, but
+// it is very fast, and makes a nice clean default implementation. Because it
+// handles lots of little corner cases, other, more complex, alias analysis
+// implementations may choose to rely on this pass to resolve these simple and
+// easy cases.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/Pass.h"
+#include "llvm/BasicBlock.h"
+#include "llvm/Function.h"
+#include "llvm/IntrinsicInst.h"
+#include "llvm/Instructions.h"
+#include "llvm/Type.h"
+#include "llvm/Target/TargetData.h"
+using namespace llvm;
+
+// Register the AliasAnalysis interface, providing a nice name to refer to.
+static RegisterAnalysisGroup<AliasAnalysis> Z("Alias Analysis");
+char AliasAnalysis::ID = 0;
+
+//===----------------------------------------------------------------------===//
+// Default chaining methods
+//===----------------------------------------------------------------------===//
+
+AliasAnalysis::AliasResult
+AliasAnalysis::alias(const Value *V1, unsigned V1Size,
+ const Value *V2, unsigned V2Size) {
+ assert(AA && "AA didn't call InitializeAliasAnalysis in its run method!");
+ return AA->alias(V1, V1Size, V2, V2Size);
+}
+
+bool AliasAnalysis::pointsToConstantMemory(const Value *P) {
+ assert(AA && "AA didn't call InitializeAliasAnalysis in its run method!");
+ return AA->pointsToConstantMemory(P);
+}
+
+void AliasAnalysis::deleteValue(Value *V) {
+ assert(AA && "AA didn't call InitializeAliasAnalysis in its run method!");
+ AA->deleteValue(V);
+}
+
+void AliasAnalysis::copyValue(Value *From, Value *To) {
+ assert(AA && "AA didn't call InitializeAliasAnalysis in its run method!");
+ AA->copyValue(From, To);
+}
+
+AliasAnalysis::ModRefResult
+AliasAnalysis::getModRefInfo(CallSite CS1, CallSite CS2) {
+ // FIXME: we can do better.
+ assert(AA && "AA didn't call InitializeAliasAnalysis in its run method!");
+ return AA->getModRefInfo(CS1, CS2);
+}
+
+
+//===----------------------------------------------------------------------===//
+// AliasAnalysis non-virtual helper method implementation
+//===----------------------------------------------------------------------===//
+
+AliasAnalysis::ModRefResult
+AliasAnalysis::getModRefInfo(LoadInst *L, Value *P, unsigned Size) {
+ return alias(L->getOperand(0), getTypeStoreSize(L->getType()),
+ P, Size) ? Ref : NoModRef;
+}
+
+AliasAnalysis::ModRefResult
+AliasAnalysis::getModRefInfo(StoreInst *S, Value *P, unsigned Size) {
+ // If the stored address cannot alias the pointer in question, then the
+ // pointer cannot be modified by the store.
+ if (!alias(S->getOperand(1),
+ getTypeStoreSize(S->getOperand(0)->getType()), P, Size))
+ return NoModRef;
+
+ // If the pointer is a pointer to constant memory, then it could not have been
+ // modified by this store.
+ return pointsToConstantMemory(P) ? NoModRef : Mod;
+}
+
+AliasAnalysis::ModRefBehavior
+AliasAnalysis::getModRefBehavior(CallSite CS,
+ std::vector<PointerAccessInfo> *Info) {
+ if (CS.doesNotAccessMemory())
+ // Can't do better than this.
+ return DoesNotAccessMemory;
+ ModRefBehavior MRB = getModRefBehavior(CS.getCalledFunction(), Info);
+ if (MRB != DoesNotAccessMemory && CS.onlyReadsMemory())
+ return OnlyReadsMemory;
+ return MRB;
+}
+
+AliasAnalysis::ModRefBehavior
+AliasAnalysis::getModRefBehavior(Function *F,
+ std::vector<PointerAccessInfo> *Info) {
+ if (F) {
+ if (F->doesNotAccessMemory())
+ // Can't do better than this.
+ return DoesNotAccessMemory;
+ if (F->onlyReadsMemory())
+ return OnlyReadsMemory;
+ if (unsigned id = F->getIntrinsicID())
+ return getModRefBehavior(id);
+ }
+ return UnknownModRefBehavior;
+}
+
+AliasAnalysis::ModRefBehavior AliasAnalysis::getModRefBehavior(unsigned iid) {
+#define GET_INTRINSIC_MODREF_BEHAVIOR
+#include "llvm/Intrinsics.gen"
+#undef GET_INTRINSIC_MODREF_BEHAVIOR
+}
+
+AliasAnalysis::ModRefResult
+AliasAnalysis::getModRefInfo(CallSite CS, Value *P, unsigned Size) {
+ ModRefBehavior MRB = getModRefBehavior(CS);
+ if (MRB == DoesNotAccessMemory)
+ return NoModRef;
+
+ ModRefResult Mask = ModRef;
+ if (MRB == OnlyReadsMemory)
+ Mask = Ref;
+ else if (MRB == AliasAnalysis::AccessesArguments) {
+ bool doesAlias = false;
+ for (CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
+ AI != AE; ++AI)
+ if (!isNoAlias(*AI, ~0U, P, Size)) {
+ doesAlias = true;
+ break;
+ }
+
+ if (!doesAlias)
+ return NoModRef;
+ }
+
+ if (!AA) return Mask;
+
+ // If P points to a constant memory location, the call definitely could not
+ // modify the memory location.
+ if ((Mask & Mod) && AA->pointsToConstantMemory(P))
+ Mask = ModRefResult(Mask & ~Mod);
+
+ return ModRefResult(Mask & AA->getModRefInfo(CS, P, Size));
+}
+
+// AliasAnalysis destructor: DO NOT move this to the header file for
+// AliasAnalysis or else clients of the AliasAnalysis class may not depend on
+// the AliasAnalysis.o file in the current .a file, causing alias analysis
+// support to not be included in the tool correctly!
+//
+AliasAnalysis::~AliasAnalysis() {}
+
+/// InitializeAliasAnalysis - Subclasses must call this method to initialize the
+/// AliasAnalysis interface before any other methods are called.
+///
+void AliasAnalysis::InitializeAliasAnalysis(Pass *P) {
+ TD = P->getAnalysisIfAvailable<TargetData>();
+ AA = &P->getAnalysis<AliasAnalysis>();
+}
+
+// getAnalysisUsage - All alias analysis implementations should invoke this
+// directly (using AliasAnalysis::getAnalysisUsage(AU)).
+void AliasAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.addRequired<AliasAnalysis>(); // All AA's chain
+}
+
+/// getTypeStoreSize - Return the TargetData store size for the given type,
+/// if known, or a conservative value otherwise.
+///
+unsigned AliasAnalysis::getTypeStoreSize(const Type *Ty) {
+ return TD ? TD->getTypeStoreSize(Ty) : ~0u;
+}
+
+/// canBasicBlockModify - Return true if it is possible for execution of the
+/// specified basic block to modify the value pointed to by Ptr.
+///
+bool AliasAnalysis::canBasicBlockModify(const BasicBlock &BB,
+ const Value *Ptr, unsigned Size) {
+ return canInstructionRangeModify(BB.front(), BB.back(), Ptr, Size);
+}
+
+/// canInstructionRangeModify - Return true if it is possible for the execution
+/// of the specified instructions to modify the value pointed to by Ptr. The
+/// instructions to consider are all of the instructions in the range of [I1,I2]
+/// INCLUSIVE. I1 and I2 must be in the same basic block.
+///
+bool AliasAnalysis::canInstructionRangeModify(const Instruction &I1,
+ const Instruction &I2,
+ const Value *Ptr, unsigned Size) {
+ assert(I1.getParent() == I2.getParent() &&
+ "Instructions not in same basic block!");
+ BasicBlock::iterator I = const_cast<Instruction*>(&I1);
+ BasicBlock::iterator E = const_cast<Instruction*>(&I2);
+ ++E; // Convert from inclusive to exclusive range.
+
+ for (; I != E; ++I) // Check every instruction in range
+ if (getModRefInfo(I, const_cast<Value*>(Ptr), Size) & Mod)
+ return true;
+ return false;
+}
+
+/// isNoAliasCall - Return true if this pointer is returned by a noalias
+/// function.
+bool llvm::isNoAliasCall(const Value *V) {
+ if (isa<CallInst>(V) || isa<InvokeInst>(V))
+ return CallSite(const_cast<Instruction*>(cast<Instruction>(V)))
+ .paramHasAttr(0, Attribute::NoAlias);
+ return false;
+}
+
+/// isIdentifiedObject - Return true if this pointer refers to a distinct and
+/// identifiable object. This returns true for:
+/// Global Variables and Functions (but not Global Aliases)
+/// Allocas and Mallocs
+/// ByVal and NoAlias Arguments
+/// NoAlias returns
+///
+bool llvm::isIdentifiedObject(const Value *V) {
+ if (isa<AllocaInst>(V) || isNoAliasCall(V))
+ return true;
+ if (isa<GlobalValue>(V) && !isa<GlobalAlias>(V))
+ return true;
+ if (const Argument *A = dyn_cast<Argument>(V))
+ return A->hasNoAliasAttr() || A->hasByValAttr();
+ return false;
+}
+
+// Because of the way .a files work, we must force the BasicAA implementation to
+// be pulled in if the AliasAnalysis classes are pulled in. Otherwise we run
+// the risk of AliasAnalysis being used, but the default implementation not
+// being linked into the tool that uses it.
+DEFINING_FILE_FOR(AliasAnalysis)
diff --git a/lib/Analysis/AliasAnalysisCounter.cpp b/lib/Analysis/AliasAnalysisCounter.cpp
new file mode 100644
index 0000000..761cd46
--- /dev/null
+++ b/lib/Analysis/AliasAnalysisCounter.cpp
@@ -0,0 +1,168 @@
+//===- AliasAnalysisCounter.cpp - Alias Analysis Query Counter ------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements a pass which can be used to count how many alias queries
+// are being made and how the alias analysis implementation being used responds.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/Passes.h"
+#include "llvm/Pass.h"
+#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/Assembly/Writer.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/raw_ostream.h"
+using namespace llvm;
+
+static cl::opt<bool>
+PrintAll("count-aa-print-all-queries", cl::ReallyHidden, cl::init(true));
+static cl::opt<bool>
+PrintAllFailures("count-aa-print-all-failed-queries", cl::ReallyHidden);
+
+namespace {
+ class AliasAnalysisCounter : public ModulePass, public AliasAnalysis {
+ unsigned No, May, Must;
+ unsigned NoMR, JustRef, JustMod, MR;
+ Module *M;
+ public:
+ static char ID; // Class identification, replacement for typeinfo
+ AliasAnalysisCounter() : ModulePass(&ID) {
+ No = May = Must = 0;
+ NoMR = JustRef = JustMod = MR = 0;
+ }
+
+ void printLine(const char *Desc, unsigned Val, unsigned Sum) {
+ errs() << " " << Val << " " << Desc << " responses ("
+ << Val*100/Sum << "%)\n";
+ }
+ ~AliasAnalysisCounter() {
+ unsigned AASum = No+May+Must;
+ unsigned MRSum = NoMR+JustRef+JustMod+MR;
+ if (AASum + MRSum) { // Print a report if any counted queries occurred...
+ errs() << "\n===== Alias Analysis Counter Report =====\n"
+ << " Analysis counted:\n"
+ << " " << AASum << " Total Alias Queries Performed\n";
+ if (AASum) {
+ printLine("no alias", No, AASum);
+ printLine("may alias", May, AASum);
+ printLine("must alias", Must, AASum);
+ errs() << " Alias Analysis Counter Summary: " << No*100/AASum << "%/"
+ << May*100/AASum << "%/" << Must*100/AASum<<"%\n\n";
+ }
+
+ errs() << " " << MRSum << " Total Mod/Ref Queries Performed\n";
+ if (MRSum) {
+ printLine("no mod/ref", NoMR, MRSum);
+ printLine("ref", JustRef, MRSum);
+ printLine("mod", JustMod, MRSum);
+ printLine("mod/ref", MR, MRSum);
+ errs() << " Mod/Ref Analysis Counter Summary: " <<NoMR*100/MRSum
+ << "%/" << JustRef*100/MRSum << "%/" << JustMod*100/MRSum
+ << "%/" << MR*100/MRSum <<"%\n\n";
+ }
+ }
+ }
+
+ bool runOnModule(Module &M) {
+ this->M = &M;
+ InitializeAliasAnalysis(this);
+ return false;
+ }
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AliasAnalysis::getAnalysisUsage(AU);
+ AU.addRequired<AliasAnalysis>();
+ AU.setPreservesAll();
+ }
+
+ /// getAdjustedAnalysisPointer - This method is used when a pass implements
+ /// an analysis interface through multiple inheritance. If needed, it
+ /// should override this to adjust the this pointer as needed for the
+ /// specified pass info.
+ virtual void *getAdjustedAnalysisPointer(const PassInfo *PI) {
+ if (PI->isPassID(&AliasAnalysis::ID))
+ return (AliasAnalysis*)this;
+ return this;
+ }
+
+ // FIXME: We could count these too...
+ bool pointsToConstantMemory(const Value *P) {
+ return getAnalysis<AliasAnalysis>().pointsToConstantMemory(P);
+ }
+
+ // Forwarding functions: just delegate to a real AA implementation, counting
+ // the number of responses...
+ AliasResult alias(const Value *V1, unsigned V1Size,
+ const Value *V2, unsigned V2Size);
+
+ ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size);
+ ModRefResult getModRefInfo(CallSite CS1, CallSite CS2) {
+ return AliasAnalysis::getModRefInfo(CS1,CS2);
+ }
+ };
+}
+
+char AliasAnalysisCounter::ID = 0;
+static RegisterPass<AliasAnalysisCounter>
+X("count-aa", "Count Alias Analysis Query Responses", false, true);
+static RegisterAnalysisGroup<AliasAnalysis> Y(X);
+
+ModulePass *llvm::createAliasAnalysisCounterPass() {
+ return new AliasAnalysisCounter();
+}
+
+AliasAnalysis::AliasResult
+AliasAnalysisCounter::alias(const Value *V1, unsigned V1Size,
+ const Value *V2, unsigned V2Size) {
+ AliasResult R = getAnalysis<AliasAnalysis>().alias(V1, V1Size, V2, V2Size);
+
+ const char *AliasString;
+ switch (R) {
+ default: llvm_unreachable("Unknown alias type!");
+ case NoAlias: No++; AliasString = "No alias"; break;
+ case MayAlias: May++; AliasString = "May alias"; break;
+ case MustAlias: Must++; AliasString = "Must alias"; break;
+ }
+
+ if (PrintAll || (PrintAllFailures && R == MayAlias)) {
+ errs() << AliasString << ":\t";
+ errs() << "[" << V1Size << "B] ";
+ WriteAsOperand(errs(), V1, true, M);
+ errs() << ", ";
+ errs() << "[" << V2Size << "B] ";
+ WriteAsOperand(errs(), V2, true, M);
+ errs() << "\n";
+ }
+
+ return R;
+}
+
+AliasAnalysis::ModRefResult
+AliasAnalysisCounter::getModRefInfo(CallSite CS, Value *P, unsigned Size) {
+ ModRefResult R = getAnalysis<AliasAnalysis>().getModRefInfo(CS, P, Size);
+
+ const char *MRString;
+ switch (R) {
+ default: llvm_unreachable("Unknown mod/ref type!");
+ case NoModRef: NoMR++; MRString = "NoModRef"; break;
+ case Ref: JustRef++; MRString = "JustRef"; break;
+ case Mod: JustMod++; MRString = "JustMod"; break;
+ case ModRef: MR++; MRString = "ModRef"; break;
+ }
+
+ if (PrintAll || (PrintAllFailures && R == ModRef)) {
+ errs() << MRString << ": Ptr: ";
+ errs() << "[" << Size << "B] ";
+ WriteAsOperand(errs(), P, true, M);
+ errs() << "\t<->" << *CS.getInstruction();
+ }
+ return R;
+}
diff --git a/lib/Analysis/AliasAnalysisEvaluator.cpp b/lib/Analysis/AliasAnalysisEvaluator.cpp
new file mode 100644
index 0000000..6b0a956
--- /dev/null
+++ b/lib/Analysis/AliasAnalysisEvaluator.cpp
@@ -0,0 +1,246 @@
+//===- AliasAnalysisEvaluator.cpp - Alias Analysis Accuracy Evaluator -----===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements a simple N^2 alias analysis accuracy evaluator.
+// Basically, for each function in the program, it simply queries to see how the
+// alias analysis implementation answers alias queries between each pair of
+// pointers in the function.
+//
+// This is inspired and adapted from code by: Naveen Neelakantam, Francesco
+// Spadini, and Wojciech Stryjewski.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Constants.h"
+#include "llvm/DerivedTypes.h"
+#include "llvm/Function.h"
+#include "llvm/Instructions.h"
+#include "llvm/Pass.h"
+#include "llvm/Analysis/Passes.h"
+#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/Assembly/Writer.h"
+#include "llvm/Target/TargetData.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/InstIterator.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/ADT/SetVector.h"
+using namespace llvm;
+
+static cl::opt<bool> PrintAll("print-all-alias-modref-info", cl::ReallyHidden);
+
+static cl::opt<bool> PrintNoAlias("print-no-aliases", cl::ReallyHidden);
+static cl::opt<bool> PrintMayAlias("print-may-aliases", cl::ReallyHidden);
+static cl::opt<bool> PrintMustAlias("print-must-aliases", cl::ReallyHidden);
+
+static cl::opt<bool> PrintNoModRef("print-no-modref", cl::ReallyHidden);
+static cl::opt<bool> PrintMod("print-mod", cl::ReallyHidden);
+static cl::opt<bool> PrintRef("print-ref", cl::ReallyHidden);
+static cl::opt<bool> PrintModRef("print-modref", cl::ReallyHidden);
+
+namespace {
+ class AAEval : public FunctionPass {
+ unsigned NoAlias, MayAlias, MustAlias;
+ unsigned NoModRef, Mod, Ref, ModRef;
+
+ public:
+ static char ID; // Pass identification, replacement for typeid
+ AAEval() : FunctionPass(&ID) {}
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.addRequired<AliasAnalysis>();
+ AU.setPreservesAll();
+ }
+
+ bool doInitialization(Module &M) {
+ NoAlias = MayAlias = MustAlias = 0;
+ NoModRef = Mod = Ref = ModRef = 0;
+
+ if (PrintAll) {
+ PrintNoAlias = PrintMayAlias = PrintMustAlias = true;
+ PrintNoModRef = PrintMod = PrintRef = PrintModRef = true;
+ }
+ return false;
+ }
+
+ bool runOnFunction(Function &F);
+ bool doFinalization(Module &M);
+ };
+}
+
+char AAEval::ID = 0;
+static RegisterPass<AAEval>
+X("aa-eval", "Exhaustive Alias Analysis Precision Evaluator", false, true);
+
+FunctionPass *llvm::createAAEvalPass() { return new AAEval(); }
+
+static void PrintResults(const char *Msg, bool P, const Value *V1,
+ const Value *V2, const Module *M) {
+ if (P) {
+ std::string o1, o2;
+ {
+ raw_string_ostream os1(o1), os2(o2);
+ WriteAsOperand(os1, V1, true, M);
+ WriteAsOperand(os2, V2, true, M);
+ }
+
+ if (o2 < o1)
+ std::swap(o1, o2);
+ errs() << " " << Msg << ":\t"
+ << o1 << ", "
+ << o2 << "\n";
+ }
+}
+
+static inline void
+PrintModRefResults(const char *Msg, bool P, Instruction *I, Value *Ptr,
+ Module *M) {
+ if (P) {
+ errs() << " " << Msg << ": Ptr: ";
+ WriteAsOperand(errs(), Ptr, true, M);
+ errs() << "\t<->" << *I << '\n';
+ }
+}
+
+bool AAEval::runOnFunction(Function &F) {
+ AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
+
+ SetVector<Value *> Pointers;
+ SetVector<CallSite> CallSites;
+
+ for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I)
+ if (isa<PointerType>(I->getType())) // Add all pointer arguments
+ Pointers.insert(I);
+
+ for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I) {
+ if (isa<PointerType>(I->getType())) // Add all pointer instructions
+ Pointers.insert(&*I);
+ Instruction &Inst = *I;
+ User::op_iterator OI = Inst.op_begin();
+ CallSite CS = CallSite::get(&Inst);
+ if (CS.getInstruction() &&
+ isa<Function>(CS.getCalledValue()))
+ ++OI; // Skip actual functions for direct function calls.
+ for (; OI != Inst.op_end(); ++OI)
+ if (isa<PointerType>((*OI)->getType()) && !isa<ConstantPointerNull>(*OI))
+ Pointers.insert(*OI);
+
+ if (CS.getInstruction()) CallSites.insert(CS);
+ }
+
+ if (PrintNoAlias || PrintMayAlias || PrintMustAlias ||
+ PrintNoModRef || PrintMod || PrintRef || PrintModRef)
+ errs() << "Function: " << F.getName() << ": " << Pointers.size()
+ << " pointers, " << CallSites.size() << " call sites\n";
+
+ // iterate over the worklist, and run the full (n^2)/2 disambiguations
+ for (SetVector<Value *>::iterator I1 = Pointers.begin(), E = Pointers.end();
+ I1 != E; ++I1) {
+ unsigned I1Size = ~0u;
+ const Type *I1ElTy = cast<PointerType>((*I1)->getType())->getElementType();
+ if (I1ElTy->isSized()) I1Size = AA.getTypeStoreSize(I1ElTy);
+
+ for (SetVector<Value *>::iterator I2 = Pointers.begin(); I2 != I1; ++I2) {
+ unsigned I2Size = ~0u;
+ const Type *I2ElTy =cast<PointerType>((*I2)->getType())->getElementType();
+ if (I2ElTy->isSized()) I2Size = AA.getTypeStoreSize(I2ElTy);
+
+ switch (AA.alias(*I1, I1Size, *I2, I2Size)) {
+ case AliasAnalysis::NoAlias:
+ PrintResults("NoAlias", PrintNoAlias, *I1, *I2, F.getParent());
+ ++NoAlias; break;
+ case AliasAnalysis::MayAlias:
+ PrintResults("MayAlias", PrintMayAlias, *I1, *I2, F.getParent());
+ ++MayAlias; break;
+ case AliasAnalysis::MustAlias:
+ PrintResults("MustAlias", PrintMustAlias, *I1, *I2, F.getParent());
+ ++MustAlias; break;
+ default:
+ errs() << "Unknown alias query result!\n";
+ }
+ }
+ }
+
+ // Mod/ref alias analysis: compare all pairs of calls and values
+ for (SetVector<CallSite>::iterator C = CallSites.begin(),
+ Ce = CallSites.end(); C != Ce; ++C) {
+ Instruction *I = C->getInstruction();
+
+ for (SetVector<Value *>::iterator V = Pointers.begin(), Ve = Pointers.end();
+ V != Ve; ++V) {
+ unsigned Size = ~0u;
+ const Type *ElTy = cast<PointerType>((*V)->getType())->getElementType();
+ if (ElTy->isSized()) Size = AA.getTypeStoreSize(ElTy);
+
+ switch (AA.getModRefInfo(*C, *V, Size)) {
+ case AliasAnalysis::NoModRef:
+ PrintModRefResults("NoModRef", PrintNoModRef, I, *V, F.getParent());
+ ++NoModRef; break;
+ case AliasAnalysis::Mod:
+ PrintModRefResults(" Mod", PrintMod, I, *V, F.getParent());
+ ++Mod; break;
+ case AliasAnalysis::Ref:
+ PrintModRefResults(" Ref", PrintRef, I, *V, F.getParent());
+ ++Ref; break;
+ case AliasAnalysis::ModRef:
+ PrintModRefResults(" ModRef", PrintModRef, I, *V, F.getParent());
+ ++ModRef; break;
+ default:
+ errs() << "Unknown alias query result!\n";
+ }
+ }
+ }
+
+ return false;
+}
+
+static void PrintPercent(unsigned Num, unsigned Sum) {
+ errs() << "(" << Num*100ULL/Sum << "."
+ << ((Num*1000ULL/Sum) % 10) << "%)\n";
+}
+
+bool AAEval::doFinalization(Module &M) {
+ unsigned AliasSum = NoAlias + MayAlias + MustAlias;
+ errs() << "===== Alias Analysis Evaluator Report =====\n";
+ if (AliasSum == 0) {
+ errs() << " Alias Analysis Evaluator Summary: No pointers!\n";
+ } else {
+ errs() << " " << AliasSum << " Total Alias Queries Performed\n";
+ errs() << " " << NoAlias << " no alias responses ";
+ PrintPercent(NoAlias, AliasSum);
+ errs() << " " << MayAlias << " may alias responses ";
+ PrintPercent(MayAlias, AliasSum);
+ errs() << " " << MustAlias << " must alias responses ";
+ PrintPercent(MustAlias, AliasSum);
+ errs() << " Alias Analysis Evaluator Pointer Alias Summary: "
+ << NoAlias*100/AliasSum << "%/" << MayAlias*100/AliasSum << "%/"
+ << MustAlias*100/AliasSum << "%\n";
+ }
+
+ // Display the summary for mod/ref analysis
+ unsigned ModRefSum = NoModRef + Mod + Ref + ModRef;
+ if (ModRefSum == 0) {
+ errs() << " Alias Analysis Mod/Ref Evaluator Summary: no mod/ref!\n";
+ } else {
+ errs() << " " << ModRefSum << " Total ModRef Queries Performed\n";
+ errs() << " " << NoModRef << " no mod/ref responses ";
+ PrintPercent(NoModRef, ModRefSum);
+ errs() << " " << Mod << " mod responses ";
+ PrintPercent(Mod, ModRefSum);
+ errs() << " " << Ref << " ref responses ";
+ PrintPercent(Ref, ModRefSum);
+ errs() << " " << ModRef << " mod & ref responses ";
+ PrintPercent(ModRef, ModRefSum);
+ errs() << " Alias Analysis Evaluator Mod/Ref Summary: "
+ << NoModRef*100/ModRefSum << "%/" << Mod*100/ModRefSum << "%/"
+ << Ref*100/ModRefSum << "%/" << ModRef*100/ModRefSum << "%\n";
+ }
+
+ return false;
+}
diff --git a/lib/Analysis/AliasDebugger.cpp b/lib/Analysis/AliasDebugger.cpp
new file mode 100644
index 0000000..88c2875
--- /dev/null
+++ b/lib/Analysis/AliasDebugger.cpp
@@ -0,0 +1,126 @@
+//===- AliasDebugger.cpp - Simple Alias Analysis Use Checker --------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This simple pass checks alias analysis users to ensure that if they
+// create a new value, they do not query AA without informing it of the value.
+// It acts as a shim over any other AA pass you want.
+//
+// Yes keeping track of every value in the program is expensive, but this is
+// a debugging pass.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/Passes.h"
+#include "llvm/Module.h"
+#include "llvm/Pass.h"
+#include "llvm/Instructions.h"
+#include "llvm/Constants.h"
+#include "llvm/DerivedTypes.h"
+#include "llvm/Analysis/AliasAnalysis.h"
+#include <set>
+using namespace llvm;
+
+namespace {
+
+ class AliasDebugger : public ModulePass, public AliasAnalysis {
+
+ //What we do is simple. Keep track of every value the AA could
+ //know about, and verify that queries are one of those.
+ //A query to a value that didn't exist when the AA was created
+ //means someone forgot to update the AA when creating new values
+
+ std::set<const Value*> Vals;
+
+ public:
+ static char ID; // Class identification, replacement for typeinfo
+ AliasDebugger() : ModulePass(&ID) {}
+
+ bool runOnModule(Module &M) {
+ InitializeAliasAnalysis(this); // set up super class
+
+ for(Module::global_iterator I = M.global_begin(),
+ E = M.global_end(); I != E; ++I)
+ Vals.insert(&*I);
+
+ for(Module::iterator I = M.begin(),
+ E = M.end(); I != E; ++I){
+ Vals.insert(&*I);
+ if(!I->isDeclaration()) {
+ for (Function::arg_iterator AI = I->arg_begin(), AE = I->arg_end();
+ AI != AE; ++AI)
+ Vals.insert(&*AI);
+ for (Function::const_iterator FI = I->begin(), FE = I->end();
+ FI != FE; ++FI)
+ for (BasicBlock::const_iterator BI = FI->begin(), BE = FI->end();
+ BI != BE; ++BI)
+ Vals.insert(&*BI);
+ }
+
+ }
+ return false;
+ }
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AliasAnalysis::getAnalysisUsage(AU);
+ AU.setPreservesAll(); // Does not transform code
+ }
+
+ /// getAdjustedAnalysisPointer - This method is used when a pass implements
+ /// an analysis interface through multiple inheritance. If needed, it
+ /// should override this to adjust the this pointer as needed for the
+ /// specified pass info.
+ virtual void *getAdjustedAnalysisPointer(const PassInfo *PI) {
+ if (PI->isPassID(&AliasAnalysis::ID))
+ return (AliasAnalysis*)this;
+ return this;
+ }
+
+ //------------------------------------------------
+ // Implement the AliasAnalysis API
+ //
+ AliasResult alias(const Value *V1, unsigned V1Size,
+ const Value *V2, unsigned V2Size) {
+ assert(Vals.find(V1) != Vals.end() && "Never seen value in AA before");
+ assert(Vals.find(V2) != Vals.end() && "Never seen value in AA before");
+ return AliasAnalysis::alias(V1, V1Size, V2, V2Size);
+ }
+
+ ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size) {
+ assert(Vals.find(P) != Vals.end() && "Never seen value in AA before");
+ return AliasAnalysis::getModRefInfo(CS, P, Size);
+ }
+
+ ModRefResult getModRefInfo(CallSite CS1, CallSite CS2) {
+ return AliasAnalysis::getModRefInfo(CS1,CS2);
+ }
+
+ bool pointsToConstantMemory(const Value *P) {
+ assert(Vals.find(P) != Vals.end() && "Never seen value in AA before");
+ return AliasAnalysis::pointsToConstantMemory(P);
+ }
+
+ virtual void deleteValue(Value *V) {
+ assert(Vals.find(V) != Vals.end() && "Never seen value in AA before");
+ AliasAnalysis::deleteValue(V);
+ }
+ virtual void copyValue(Value *From, Value *To) {
+ Vals.insert(To);
+ AliasAnalysis::copyValue(From, To);
+ }
+
+ };
+}
+
+char AliasDebugger::ID = 0;
+static RegisterPass<AliasDebugger>
+X("debug-aa", "AA use debugger", false, true);
+static RegisterAnalysisGroup<AliasAnalysis> Y(X);
+
+Pass *llvm::createAliasDebugger() { return new AliasDebugger(); }
+
diff --git a/lib/Analysis/AliasSetTracker.cpp b/lib/Analysis/AliasSetTracker.cpp
new file mode 100644
index 0000000..02aff50
--- /dev/null
+++ b/lib/Analysis/AliasSetTracker.cpp
@@ -0,0 +1,604 @@
+//===- AliasSetTracker.cpp - Alias Sets Tracker implementation-------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements the AliasSetTracker and AliasSet classes.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/AliasSetTracker.h"
+#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/Instructions.h"
+#include "llvm/IntrinsicInst.h"
+#include "llvm/Pass.h"
+#include "llvm/Type.h"
+#include "llvm/Target/TargetData.h"
+#include "llvm/Assembly/Writer.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/InstIterator.h"
+#include "llvm/Support/Format.h"
+#include "llvm/Support/raw_ostream.h"
+using namespace llvm;
+
+/// mergeSetIn - Merge the specified alias set into this alias set.
+///
+void AliasSet::mergeSetIn(AliasSet &AS, AliasSetTracker &AST) {
+ assert(!AS.Forward && "Alias set is already forwarding!");
+ assert(!Forward && "This set is a forwarding set!!");
+
+ // Update the alias and access types of this set...
+ AccessTy |= AS.AccessTy;
+ AliasTy |= AS.AliasTy;
+
+ if (AliasTy == MustAlias) {
+ // Check that these two merged sets really are must aliases. Since both
+ // used to be must-alias sets, we can just check any pointer from each set
+ // for aliasing.
+ AliasAnalysis &AA = AST.getAliasAnalysis();
+ PointerRec *L = getSomePointer();
+ PointerRec *R = AS.getSomePointer();
+
+ // If the pointers are not a must-alias pair, this set becomes a may alias.
+ if (AA.alias(L->getValue(), L->getSize(), R->getValue(), R->getSize())
+ != AliasAnalysis::MustAlias)
+ AliasTy = MayAlias;
+ }
+
+ if (CallSites.empty()) { // Merge call sites...
+ if (!AS.CallSites.empty())
+ std::swap(CallSites, AS.CallSites);
+ } else if (!AS.CallSites.empty()) {
+ CallSites.insert(CallSites.end(), AS.CallSites.begin(), AS.CallSites.end());
+ AS.CallSites.clear();
+ }
+
+ AS.Forward = this; // Forward across AS now...
+ addRef(); // AS is now pointing to us...
+
+ // Merge the list of constituent pointers...
+ if (AS.PtrList) {
+ *PtrListEnd = AS.PtrList;
+ AS.PtrList->setPrevInList(PtrListEnd);
+ PtrListEnd = AS.PtrListEnd;
+
+ AS.PtrList = 0;
+ AS.PtrListEnd = &AS.PtrList;
+ assert(*AS.PtrListEnd == 0 && "End of list is not null?");
+ }
+}
+
+void AliasSetTracker::removeAliasSet(AliasSet *AS) {
+ if (AliasSet *Fwd = AS->Forward) {
+ Fwd->dropRef(*this);
+ AS->Forward = 0;
+ }
+ AliasSets.erase(AS);
+}
+
+void AliasSet::removeFromTracker(AliasSetTracker &AST) {
+ assert(RefCount == 0 && "Cannot remove non-dead alias set from tracker!");
+ AST.removeAliasSet(this);
+}
+
+void AliasSet::addPointer(AliasSetTracker &AST, PointerRec &Entry,
+ unsigned Size, bool KnownMustAlias) {
+ assert(!Entry.hasAliasSet() && "Entry already in set!");
+
+ // Check to see if we have to downgrade to _may_ alias.
+ if (isMustAlias() && !KnownMustAlias)
+ if (PointerRec *P = getSomePointer()) {
+ AliasAnalysis &AA = AST.getAliasAnalysis();
+ AliasAnalysis::AliasResult Result =
+ AA.alias(P->getValue(), P->getSize(), Entry.getValue(), Size);
+ if (Result == AliasAnalysis::MayAlias)
+ AliasTy = MayAlias;
+ else // First entry of must alias must have maximum size!
+ P->updateSize(Size);
+ assert(Result != AliasAnalysis::NoAlias && "Cannot be part of must set!");
+ }
+
+ Entry.setAliasSet(this);
+ Entry.updateSize(Size);
+
+ // Add it to the end of the list...
+ assert(*PtrListEnd == 0 && "End of list is not null?");
+ *PtrListEnd = &Entry;
+ PtrListEnd = Entry.setPrevInList(PtrListEnd);
+ assert(*PtrListEnd == 0 && "End of list is not null?");
+ addRef(); // Entry points to alias set...
+}
+
+void AliasSet::addCallSite(CallSite CS, AliasAnalysis &AA) {
+ CallSites.push_back(CS);
+
+ AliasAnalysis::ModRefBehavior Behavior = AA.getModRefBehavior(CS);
+ if (Behavior == AliasAnalysis::DoesNotAccessMemory)
+ return;
+ else if (Behavior == AliasAnalysis::OnlyReadsMemory) {
+ AliasTy = MayAlias;
+ AccessTy |= Refs;
+ return;
+ }
+
+ // FIXME: This should use mod/ref information to make this not suck so bad
+ AliasTy = MayAlias;
+ AccessTy = ModRef;
+}
+
+/// aliasesPointer - Return true if the specified pointer "may" (or must)
+/// alias one of the members in the set.
+///
+bool AliasSet::aliasesPointer(const Value *Ptr, unsigned Size,
+ AliasAnalysis &AA) const {
+ if (AliasTy == MustAlias) {
+ assert(CallSites.empty() && "Illegal must alias set!");
+
+ // If this is a set of MustAliases, only check to see if the pointer aliases
+ // SOME value in the set...
+ PointerRec *SomePtr = getSomePointer();
+ assert(SomePtr && "Empty must-alias set??");
+ return AA.alias(SomePtr->getValue(), SomePtr->getSize(), Ptr, Size);
+ }
+
+ // If this is a may-alias set, we have to check all of the pointers in the set
+ // to be sure it doesn't alias the set...
+ for (iterator I = begin(), E = end(); I != E; ++I)
+ if (AA.alias(Ptr, Size, I.getPointer(), I.getSize()))
+ return true;
+
+ // Check the call sites list and invoke list...
+ if (!CallSites.empty()) {
+ for (unsigned i = 0, e = CallSites.size(); i != e; ++i)
+ if (AA.getModRefInfo(CallSites[i], const_cast<Value*>(Ptr), Size)
+ != AliasAnalysis::NoModRef)
+ return true;
+ }
+
+ return false;
+}
+
+bool AliasSet::aliasesCallSite(CallSite CS, AliasAnalysis &AA) const {
+ if (AA.doesNotAccessMemory(CS))
+ return false;
+
+ for (unsigned i = 0, e = CallSites.size(); i != e; ++i)
+ if (AA.getModRefInfo(CallSites[i], CS) != AliasAnalysis::NoModRef ||
+ AA.getModRefInfo(CS, CallSites[i]) != AliasAnalysis::NoModRef)
+ return true;
+
+ for (iterator I = begin(), E = end(); I != E; ++I)
+ if (AA.getModRefInfo(CS, I.getPointer(), I.getSize()) !=
+ AliasAnalysis::NoModRef)
+ return true;
+
+ return false;
+}
+
+void AliasSetTracker::clear() {
+ // Delete all the PointerRec entries.
+ for (PointerMapType::iterator I = PointerMap.begin(), E = PointerMap.end();
+ I != E; ++I)
+ I->second->eraseFromList();
+
+ PointerMap.clear();
+
+ // The alias sets should all be clear now.
+ AliasSets.clear();
+}
+
+
+/// findAliasSetForPointer - Given a pointer, find the one alias set to put the
+/// instruction referring to the pointer into. If there are multiple alias sets
+/// that may alias the pointer, merge them together and return the unified set.
+///
+AliasSet *AliasSetTracker::findAliasSetForPointer(const Value *Ptr,
+ unsigned Size) {
+ AliasSet *FoundSet = 0;
+ for (iterator I = begin(), E = end(); I != E; ++I)
+ if (!I->Forward && I->aliasesPointer(Ptr, Size, AA)) {
+ if (FoundSet == 0) { // If this is the first alias set ptr can go into.
+ FoundSet = I; // Remember it.
+ } else { // Otherwise, we must merge the sets.
+ FoundSet->mergeSetIn(*I, *this); // Merge in contents.
+ }
+ }
+
+ return FoundSet;
+}
+
+/// containsPointer - Return true if the specified location is represented by
+/// this alias set, false otherwise. This does not modify the AST object or
+/// alias sets.
+bool AliasSetTracker::containsPointer(Value *Ptr, unsigned Size) const {
+ for (const_iterator I = begin(), E = end(); I != E; ++I)
+ if (!I->Forward && I->aliasesPointer(Ptr, Size, AA))
+ return true;
+ return false;
+}
+
+
+
+AliasSet *AliasSetTracker::findAliasSetForCallSite(CallSite CS) {
+ AliasSet *FoundSet = 0;
+ for (iterator I = begin(), E = end(); I != E; ++I)
+ if (!I->Forward && I->aliasesCallSite(CS, AA)) {
+ if (FoundSet == 0) { // If this is the first alias set ptr can go into.
+ FoundSet = I; // Remember it.
+ } else if (!I->Forward) { // Otherwise, we must merge the sets.
+ FoundSet->mergeSetIn(*I, *this); // Merge in contents.
+ }
+ }
+
+ return FoundSet;
+}
+
+
+
+
+/// getAliasSetForPointer - Return the alias set that the specified pointer
+/// lives in.
+AliasSet &AliasSetTracker::getAliasSetForPointer(Value *Pointer, unsigned Size,
+ bool *New) {
+ AliasSet::PointerRec &Entry = getEntryFor(Pointer);
+
+ // Check to see if the pointer is already known...
+ if (Entry.hasAliasSet()) {
+ Entry.updateSize(Size);
+ // Return the set!
+ return *Entry.getAliasSet(*this)->getForwardedTarget(*this);
+ } else if (AliasSet *AS = findAliasSetForPointer(Pointer, Size)) {
+ // Add it to the alias set it aliases...
+ AS->addPointer(*this, Entry, Size);
+ return *AS;
+ } else {
+ if (New) *New = true;
+ // Otherwise create a new alias set to hold the loaded pointer...
+ AliasSets.push_back(new AliasSet());
+ AliasSets.back().addPointer(*this, Entry, Size);
+ return AliasSets.back();
+ }
+}
+
+bool AliasSetTracker::add(Value *Ptr, unsigned Size) {
+ bool NewPtr;
+ addPointer(Ptr, Size, AliasSet::NoModRef, NewPtr);
+ return NewPtr;
+}
+
+
+bool AliasSetTracker::add(LoadInst *LI) {
+ bool NewPtr;
+ AliasSet &AS = addPointer(LI->getOperand(0),
+ AA.getTypeStoreSize(LI->getType()),
+ AliasSet::Refs, NewPtr);
+ if (LI->isVolatile()) AS.setVolatile();
+ return NewPtr;
+}
+
+bool AliasSetTracker::add(StoreInst *SI) {
+ bool NewPtr;
+ Value *Val = SI->getOperand(0);
+ AliasSet &AS = addPointer(SI->getOperand(1),
+ AA.getTypeStoreSize(Val->getType()),
+ AliasSet::Mods, NewPtr);
+ if (SI->isVolatile()) AS.setVolatile();
+ return NewPtr;
+}
+
+bool AliasSetTracker::add(VAArgInst *VAAI) {
+ bool NewPtr;
+ addPointer(VAAI->getOperand(0), ~0, AliasSet::ModRef, NewPtr);
+ return NewPtr;
+}
+
+
+bool AliasSetTracker::add(CallSite CS) {
+ if (isa<DbgInfoIntrinsic>(CS.getInstruction()))
+ return true; // Ignore DbgInfo Intrinsics.
+ if (AA.doesNotAccessMemory(CS))
+ return true; // doesn't alias anything
+
+ AliasSet *AS = findAliasSetForCallSite(CS);
+ if (!AS) {
+ AliasSets.push_back(new AliasSet());
+ AS = &AliasSets.back();
+ AS->addCallSite(CS, AA);
+ return true;
+ } else {
+ AS->addCallSite(CS, AA);
+ return false;
+ }
+}
+
+bool AliasSetTracker::add(Instruction *I) {
+ // Dispatch to one of the other add methods...
+ if (LoadInst *LI = dyn_cast<LoadInst>(I))
+ return add(LI);
+ else if (StoreInst *SI = dyn_cast<StoreInst>(I))
+ return add(SI);
+ else if (CallInst *CI = dyn_cast<CallInst>(I))
+ return add(CI);
+ else if (InvokeInst *II = dyn_cast<InvokeInst>(I))
+ return add(II);
+ else if (VAArgInst *VAAI = dyn_cast<VAArgInst>(I))
+ return add(VAAI);
+ return true;
+}
+
+void AliasSetTracker::add(BasicBlock &BB) {
+ for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
+ add(I);
+}
+
+void AliasSetTracker::add(const AliasSetTracker &AST) {
+ assert(&AA == &AST.AA &&
+ "Merging AliasSetTracker objects with different Alias Analyses!");
+
+ // Loop over all of the alias sets in AST, adding the pointers contained
+ // therein into the current alias sets. This can cause alias sets to be
+ // merged together in the current AST.
+ for (const_iterator I = AST.begin(), E = AST.end(); I != E; ++I)
+ if (!I->Forward) { // Ignore forwarding alias sets
+ AliasSet &AS = const_cast<AliasSet&>(*I);
+
+ // If there are any call sites in the alias set, add them to this AST.
+ for (unsigned i = 0, e = AS.CallSites.size(); i != e; ++i)
+ add(AS.CallSites[i]);
+
+ // Loop over all of the pointers in this alias set...
+ AliasSet::iterator I = AS.begin(), E = AS.end();
+ bool X;
+ for (; I != E; ++I) {
+ AliasSet &NewAS = addPointer(I.getPointer(), I.getSize(),
+ (AliasSet::AccessType)AS.AccessTy, X);
+ if (AS.isVolatile()) NewAS.setVolatile();
+ }
+ }
+}
+
+/// remove - Remove the specified (potentially non-empty) alias set from the
+/// tracker.
+void AliasSetTracker::remove(AliasSet &AS) {
+ // Drop all call sites.
+ AS.CallSites.clear();
+
+ // Clear the alias set.
+ unsigned NumRefs = 0;
+ while (!AS.empty()) {
+ AliasSet::PointerRec *P = AS.PtrList;
+
+ Value *ValToRemove = P->getValue();
+
+ // Unlink and delete entry from the list of values.
+ P->eraseFromList();
+
+ // Remember how many references need to be dropped.
+ ++NumRefs;
+
+ // Finally, remove the entry.
+ PointerMap.erase(ValToRemove);
+ }
+
+ // Stop using the alias set, removing it.
+ AS.RefCount -= NumRefs;
+ if (AS.RefCount == 0)
+ AS.removeFromTracker(*this);
+}
+
+bool AliasSetTracker::remove(Value *Ptr, unsigned Size) {
+ AliasSet *AS = findAliasSetForPointer(Ptr, Size);
+ if (!AS) return false;
+ remove(*AS);
+ return true;
+}
+
+bool AliasSetTracker::remove(LoadInst *LI) {
+ unsigned Size = AA.getTypeStoreSize(LI->getType());
+ AliasSet *AS = findAliasSetForPointer(LI->getOperand(0), Size);
+ if (!AS) return false;
+ remove(*AS);
+ return true;
+}
+
+bool AliasSetTracker::remove(StoreInst *SI) {
+ unsigned Size = AA.getTypeStoreSize(SI->getOperand(0)->getType());
+ AliasSet *AS = findAliasSetForPointer(SI->getOperand(1), Size);
+ if (!AS) return false;
+ remove(*AS);
+ return true;
+}
+
+bool AliasSetTracker::remove(VAArgInst *VAAI) {
+ AliasSet *AS = findAliasSetForPointer(VAAI->getOperand(0), ~0);
+ if (!AS) return false;
+ remove(*AS);
+ return true;
+}
+
+bool AliasSetTracker::remove(CallSite CS) {
+ if (AA.doesNotAccessMemory(CS))
+ return false; // doesn't alias anything
+
+ AliasSet *AS = findAliasSetForCallSite(CS);
+ if (!AS) return false;
+ remove(*AS);
+ return true;
+}
+
+bool AliasSetTracker::remove(Instruction *I) {
+ // Dispatch to one of the other remove methods...
+ if (LoadInst *LI = dyn_cast<LoadInst>(I))
+ return remove(LI);
+ else if (StoreInst *SI = dyn_cast<StoreInst>(I))
+ return remove(SI);
+ else if (CallInst *CI = dyn_cast<CallInst>(I))
+ return remove(CI);
+ else if (VAArgInst *VAAI = dyn_cast<VAArgInst>(I))
+ return remove(VAAI);
+ return true;
+}
+
+
+// deleteValue method - This method is used to remove a pointer value from the
+// AliasSetTracker entirely. It should be used when an instruction is deleted
+// from the program to update the AST. If you don't use this, you would have
+// dangling pointers to deleted instructions.
+//
+void AliasSetTracker::deleteValue(Value *PtrVal) {
+ // Notify the alias analysis implementation that this value is gone.
+ AA.deleteValue(PtrVal);
+
+ // If this is a call instruction, remove the callsite from the appropriate
+ // AliasSet.
+ CallSite CS = CallSite::get(PtrVal);
+ if (CS.getInstruction())
+ if (!AA.doesNotAccessMemory(CS))
+ if (AliasSet *AS = findAliasSetForCallSite(CS))
+ AS->removeCallSite(CS);
+
+ // First, look up the PointerRec for this pointer.
+ PointerMapType::iterator I = PointerMap.find(PtrVal);
+ if (I == PointerMap.end()) return; // Noop
+
+ // If we found one, remove the pointer from the alias set it is in.
+ AliasSet::PointerRec *PtrValEnt = I->second;
+ AliasSet *AS = PtrValEnt->getAliasSet(*this);
+
+ // Unlink and delete from the list of values.
+ PtrValEnt->eraseFromList();
+
+ // Stop using the alias set.
+ AS->dropRef(*this);
+
+ PointerMap.erase(I);
+}
+
+// copyValue - This method should be used whenever a preexisting value in the
+// program is copied or cloned, introducing a new value. Note that it is ok for
+// clients that use this method to introduce the same value multiple times: if
+// the tracker already knows about a value, it will ignore the request.
+//
+void AliasSetTracker::copyValue(Value *From, Value *To) {
+ // Notify the alias analysis implementation that this value is copied.
+ AA.copyValue(From, To);
+
+ // First, look up the PointerRec for this pointer.
+ PointerMapType::iterator I = PointerMap.find(From);
+ if (I == PointerMap.end())
+ return; // Noop
+ assert(I->second->hasAliasSet() && "Dead entry?");
+
+ AliasSet::PointerRec &Entry = getEntryFor(To);
+ if (Entry.hasAliasSet()) return; // Already in the tracker!
+
+ // Add it to the alias set it aliases...
+ I = PointerMap.find(From);
+ AliasSet *AS = I->second->getAliasSet(*this);
+ AS->addPointer(*this, Entry, I->second->getSize(), true);
+}
+
+
+
+//===----------------------------------------------------------------------===//
+// AliasSet/AliasSetTracker Printing Support
+//===----------------------------------------------------------------------===//
+
+void AliasSet::print(raw_ostream &OS) const {
+ OS << " AliasSet[" << format("0x%p", (void*)this) << "," << RefCount << "] ";
+ OS << (AliasTy == MustAlias ? "must" : "may") << " alias, ";
+ switch (AccessTy) {
+ case NoModRef: OS << "No access "; break;
+ case Refs : OS << "Ref "; break;
+ case Mods : OS << "Mod "; break;
+ case ModRef : OS << "Mod/Ref "; break;
+ default: llvm_unreachable("Bad value for AccessTy!");
+ }
+ if (isVolatile()) OS << "[volatile] ";
+ if (Forward)
+ OS << " forwarding to " << (void*)Forward;
+
+
+ if (!empty()) {
+ OS << "Pointers: ";
+ for (iterator I = begin(), E = end(); I != E; ++I) {
+ if (I != begin()) OS << ", ";
+ WriteAsOperand(OS << "(", I.getPointer());
+ OS << ", " << I.getSize() << ")";
+ }
+ }
+ if (!CallSites.empty()) {
+ OS << "\n " << CallSites.size() << " Call Sites: ";
+ for (unsigned i = 0, e = CallSites.size(); i != e; ++i) {
+ if (i) OS << ", ";
+ WriteAsOperand(OS, CallSites[i].getCalledValue());
+ }
+ }
+ OS << "\n";
+}
+
+void AliasSetTracker::print(raw_ostream &OS) const {
+ OS << "Alias Set Tracker: " << AliasSets.size() << " alias sets for "
+ << PointerMap.size() << " pointer values.\n";
+ for (const_iterator I = begin(), E = end(); I != E; ++I)
+ I->print(OS);
+ OS << "\n";
+}
+
+void AliasSet::dump() const { print(dbgs()); }
+void AliasSetTracker::dump() const { print(dbgs()); }
+
+//===----------------------------------------------------------------------===//
+// ASTCallbackVH Class Implementation
+//===----------------------------------------------------------------------===//
+
+void AliasSetTracker::ASTCallbackVH::deleted() {
+ assert(AST && "ASTCallbackVH called with a null AliasSetTracker!");
+ AST->deleteValue(getValPtr());
+ // this now dangles!
+}
+
+AliasSetTracker::ASTCallbackVH::ASTCallbackVH(Value *V, AliasSetTracker *ast)
+ : CallbackVH(V), AST(ast) {}
+
+AliasSetTracker::ASTCallbackVH &
+AliasSetTracker::ASTCallbackVH::operator=(Value *V) {
+ return *this = ASTCallbackVH(V, AST);
+}
+
+//===----------------------------------------------------------------------===//
+// AliasSetPrinter Pass
+//===----------------------------------------------------------------------===//
+
+namespace {
+ class AliasSetPrinter : public FunctionPass {
+ AliasSetTracker *Tracker;
+ public:
+ static char ID; // Pass identification, replacement for typeid
+ AliasSetPrinter() : FunctionPass(&ID) {}
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.setPreservesAll();
+ AU.addRequired<AliasAnalysis>();
+ }
+
+ virtual bool runOnFunction(Function &F) {
+ Tracker = new AliasSetTracker(getAnalysis<AliasAnalysis>());
+
+ for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
+ Tracker->add(&*I);
+ Tracker->print(errs());
+ delete Tracker;
+ return false;
+ }
+ };
+}
+
+char AliasSetPrinter::ID = 0;
+static RegisterPass<AliasSetPrinter>
+X("print-alias-sets", "Alias Set Printer", false, true);
diff --git a/lib/Analysis/Analysis.cpp b/lib/Analysis/Analysis.cpp
new file mode 100644
index 0000000..398dec7
--- /dev/null
+++ b/lib/Analysis/Analysis.cpp
@@ -0,0 +1,43 @@
+//===-- Analysis.cpp ------------------------------------------------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm-c/Analysis.h"
+#include "llvm/Analysis/Verifier.h"
+#include <cstring>
+
+using namespace llvm;
+
+LLVMBool LLVMVerifyModule(LLVMModuleRef M, LLVMVerifierFailureAction Action,
+ char **OutMessages) {
+ std::string Messages;
+
+ LLVMBool Result = verifyModule(*unwrap(M),
+ static_cast<VerifierFailureAction>(Action),
+ OutMessages? &Messages : 0);
+
+ if (OutMessages)
+ *OutMessages = strdup(Messages.c_str());
+
+ return Result;
+}
+
+LLVMBool LLVMVerifyFunction(LLVMValueRef Fn, LLVMVerifierFailureAction Action) {
+ return verifyFunction(*unwrap<Function>(Fn),
+ static_cast<VerifierFailureAction>(Action));
+}
+
+void LLVMViewFunctionCFG(LLVMValueRef Fn) {
+ Function *F = unwrap<Function>(Fn);
+ F->viewCFG();
+}
+
+void LLVMViewFunctionCFGOnly(LLVMValueRef Fn) {
+ Function *F = unwrap<Function>(Fn);
+ F->viewCFGOnly();
+}
diff --git a/lib/Analysis/BasicAliasAnalysis.cpp b/lib/Analysis/BasicAliasAnalysis.cpp
new file mode 100644
index 0000000..36b831c
--- /dev/null
+++ b/lib/Analysis/BasicAliasAnalysis.cpp
@@ -0,0 +1,753 @@
+//===- BasicAliasAnalysis.cpp - Local Alias Analysis Impl -----------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file defines the default implementation of the Alias Analysis interface
+// that simply implements a few identities (two different globals cannot alias,
+// etc), but otherwise does no analysis.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/Analysis/Passes.h"
+#include "llvm/Constants.h"
+#include "llvm/DerivedTypes.h"
+#include "llvm/Function.h"
+#include "llvm/GlobalVariable.h"
+#include "llvm/Instructions.h"
+#include "llvm/IntrinsicInst.h"
+#include "llvm/Operator.h"
+#include "llvm/Pass.h"
+#include "llvm/Analysis/CaptureTracking.h"
+#include "llvm/Analysis/MemoryBuiltins.h"
+#include "llvm/Analysis/ValueTracking.h"
+#include "llvm/Target/TargetData.h"
+#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/Support/ErrorHandling.h"
+#include <algorithm>
+using namespace llvm;
+
+//===----------------------------------------------------------------------===//
+// Useful predicates
+//===----------------------------------------------------------------------===//
+
+/// isKnownNonNull - Return true if we know that the specified value is never
+/// null.
+static bool isKnownNonNull(const Value *V) {
+ // Alloca never returns null, malloc might.
+ if (isa<AllocaInst>(V)) return true;
+
+ // A byval argument is never null.
+ if (const Argument *A = dyn_cast<Argument>(V))
+ return A->hasByValAttr();
+
+ // Global values are not null unless extern weak.
+ if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
+ return !GV->hasExternalWeakLinkage();
+ return false;
+}
+
+/// isNonEscapingLocalObject - Return true if the pointer is to a function-local
+/// object that never escapes from the function.
+static bool isNonEscapingLocalObject(const Value *V) {
+ // If this is a local allocation, check to see if it escapes.
+ if (isa<AllocaInst>(V) || isNoAliasCall(V))
+ // Set StoreCaptures to True so that we can assume in our callers that the
+ // pointer is not the result of a load instruction. Currently
+ // PointerMayBeCaptured doesn't have any special analysis for the
+ // StoreCaptures=false case; if it did, our callers could be refined to be
+ // more precise.
+ return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
+
+ // If this is an argument that corresponds to a byval or noalias argument,
+ // then it has not escaped before entering the function. Check if it escapes
+ // inside the function.
+ if (const Argument *A = dyn_cast<Argument>(V))
+ if (A->hasByValAttr() || A->hasNoAliasAttr()) {
+ // Don't bother analyzing arguments already known not to escape.
+ if (A->hasNoCaptureAttr())
+ return true;
+ return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
+ }
+ return false;
+}
+
+
+/// isObjectSmallerThan - Return true if we can prove that the object specified
+/// by V is smaller than Size.
+static bool isObjectSmallerThan(const Value *V, unsigned Size,
+ const TargetData &TD) {
+ const Type *AccessTy;
+ if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
+ AccessTy = GV->getType()->getElementType();
+ } else if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
+ if (!AI->isArrayAllocation())
+ AccessTy = AI->getType()->getElementType();
+ else
+ return false;
+ } else if (const CallInst* CI = extractMallocCall(V)) {
+ if (!isArrayMalloc(V, &TD))
+ // The size is the argument to the malloc call.
+ if (const ConstantInt* C = dyn_cast<ConstantInt>(CI->getOperand(1)))
+ return (C->getZExtValue() < Size);
+ return false;
+ } else if (const Argument *A = dyn_cast<Argument>(V)) {
+ if (A->hasByValAttr())
+ AccessTy = cast<PointerType>(A->getType())->getElementType();
+ else
+ return false;
+ } else {
+ return false;
+ }
+
+ if (AccessTy->isSized())
+ return TD.getTypeAllocSize(AccessTy) < Size;
+ return false;
+}
+
+//===----------------------------------------------------------------------===//
+// NoAA Pass
+//===----------------------------------------------------------------------===//
+
+namespace {
+ /// NoAA - This class implements the -no-aa pass, which always returns "I
+ /// don't know" for alias queries. NoAA is unlike other alias analysis
+ /// implementations, in that it does not chain to a previous analysis. As
+ /// such it doesn't follow many of the rules that other alias analyses must.
+ ///
+ struct NoAA : public ImmutablePass, public AliasAnalysis {
+ static char ID; // Class identification, replacement for typeinfo
+ NoAA() : ImmutablePass(&ID) {}
+ explicit NoAA(void *PID) : ImmutablePass(PID) { }
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ }
+
+ virtual void initializePass() {
+ TD = getAnalysisIfAvailable<TargetData>();
+ }
+
+ virtual AliasResult alias(const Value *V1, unsigned V1Size,
+ const Value *V2, unsigned V2Size) {
+ return MayAlias;
+ }
+
+ virtual void getArgumentAccesses(Function *F, CallSite CS,
+ std::vector<PointerAccessInfo> &Info) {
+ llvm_unreachable("This method may not be called on this function!");
+ }
+
+ virtual bool pointsToConstantMemory(const Value *P) { return false; }
+ virtual ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size) {
+ return ModRef;
+ }
+ virtual ModRefResult getModRefInfo(CallSite CS1, CallSite CS2) {
+ return ModRef;
+ }
+
+ virtual void deleteValue(Value *V) {}
+ virtual void copyValue(Value *From, Value *To) {}
+
+ /// getAdjustedAnalysisPointer - This method is used when a pass implements
+ /// an analysis interface through multiple inheritance. If needed, it should
+ /// override this to adjust the this pointer as needed for the specified pass
+ /// info.
+ virtual void *getAdjustedAnalysisPointer(const PassInfo *PI) {
+ if (PI->isPassID(&AliasAnalysis::ID))
+ return (AliasAnalysis*)this;
+ return this;
+ }
+ };
+} // End of anonymous namespace
+
+// Register this pass...
+char NoAA::ID = 0;
+static RegisterPass<NoAA>
+U("no-aa", "No Alias Analysis (always returns 'may' alias)", true, true);
+
+// Declare that we implement the AliasAnalysis interface
+static RegisterAnalysisGroup<AliasAnalysis> V(U);
+
+ImmutablePass *llvm::createNoAAPass() { return new NoAA(); }
+
+//===----------------------------------------------------------------------===//
+// BasicAA Pass
+//===----------------------------------------------------------------------===//
+
+namespace {
+ /// BasicAliasAnalysis - This is the default alias analysis implementation.
+ /// Because it doesn't chain to a previous alias analysis (like -no-aa), it
+ /// derives from the NoAA class.
+ struct BasicAliasAnalysis : public NoAA {
+ static char ID; // Class identification, replacement for typeinfo
+ BasicAliasAnalysis() : NoAA(&ID) {}
+ AliasResult alias(const Value *V1, unsigned V1Size,
+ const Value *V2, unsigned V2Size) {
+ assert(VisitedPHIs.empty() && "VisitedPHIs must be cleared after use!");
+ AliasResult Alias = aliasCheck(V1, V1Size, V2, V2Size);
+ VisitedPHIs.clear();
+ return Alias;
+ }
+
+ ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size);
+ ModRefResult getModRefInfo(CallSite CS1, CallSite CS2);
+
+ /// pointsToConstantMemory - Chase pointers until we find a (constant
+ /// global) or not.
+ bool pointsToConstantMemory(const Value *P);
+
+ /// getAdjustedAnalysisPointer - This method is used when a pass implements
+ /// an analysis interface through multiple inheritance. If needed, it should
+ /// override this to adjust the this pointer as needed for the specified pass
+ /// info.
+ virtual void *getAdjustedAnalysisPointer(const PassInfo *PI) {
+ if (PI->isPassID(&AliasAnalysis::ID))
+ return (AliasAnalysis*)this;
+ return this;
+ }
+
+ private:
+ // VisitedPHIs - Track PHI nodes visited by a aliasCheck() call.
+ SmallPtrSet<const Value*, 16> VisitedPHIs;
+
+ // aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP
+ // instruction against another.
+ AliasResult aliasGEP(const GEPOperator *V1, unsigned V1Size,
+ const Value *V2, unsigned V2Size,
+ const Value *UnderlyingV1, const Value *UnderlyingV2);
+
+ // aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI
+ // instruction against another.
+ AliasResult aliasPHI(const PHINode *PN, unsigned PNSize,
+ const Value *V2, unsigned V2Size);
+
+ /// aliasSelect - Disambiguate a Select instruction against another value.
+ AliasResult aliasSelect(const SelectInst *SI, unsigned SISize,
+ const Value *V2, unsigned V2Size);
+
+ AliasResult aliasCheck(const Value *V1, unsigned V1Size,
+ const Value *V2, unsigned V2Size);
+ };
+} // End of anonymous namespace
+
+// Register this pass...
+char BasicAliasAnalysis::ID = 0;
+static RegisterPass<BasicAliasAnalysis>
+X("basicaa", "Basic Alias Analysis (default AA impl)", false, true);
+
+// Declare that we implement the AliasAnalysis interface
+static RegisterAnalysisGroup<AliasAnalysis, true> Y(X);
+
+ImmutablePass *llvm::createBasicAliasAnalysisPass() {
+ return new BasicAliasAnalysis();
+}
+
+
+/// pointsToConstantMemory - Chase pointers until we find a (constant
+/// global) or not.
+bool BasicAliasAnalysis::pointsToConstantMemory(const Value *P) {
+ if (const GlobalVariable *GV =
+ dyn_cast<GlobalVariable>(P->getUnderlyingObject()))
+ // Note: this doesn't require GV to be "ODR" because it isn't legal for a
+ // global to be marked constant in some modules and non-constant in others.
+ // GV may even be a declaration, not a definition.
+ return GV->isConstant();
+ return false;
+}
+
+
+/// getModRefInfo - Check to see if the specified callsite can clobber the
+/// specified memory object. Since we only look at local properties of this
+/// function, we really can't say much about this query. We do, however, use
+/// simple "address taken" analysis on local objects.
+AliasAnalysis::ModRefResult
+BasicAliasAnalysis::getModRefInfo(CallSite CS, Value *P, unsigned Size) {
+ const Value *Object = P->getUnderlyingObject();
+
+ // If this is a tail call and P points to a stack location, we know that
+ // the tail call cannot access or modify the local stack.
+ // We cannot exclude byval arguments here; these belong to the caller of
+ // the current function not to the current function, and a tail callee
+ // may reference them.
+ if (isa<AllocaInst>(Object))
+ if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction()))
+ if (CI->isTailCall())
+ return NoModRef;
+
+ // If the pointer is to a locally allocated object that does not escape,
+ // then the call can not mod/ref the pointer unless the call takes the pointer
+ // as an argument, and itself doesn't capture it.
+ if (!isa<Constant>(Object) && CS.getInstruction() != Object &&
+ isNonEscapingLocalObject(Object)) {
+ bool PassedAsArg = false;
+ unsigned ArgNo = 0;
+ for (CallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end();
+ CI != CE; ++CI, ++ArgNo) {
+ // Only look at the no-capture pointer arguments.
+ if (!isa<PointerType>((*CI)->getType()) ||
+ !CS.paramHasAttr(ArgNo+1, Attribute::NoCapture))
+ continue;
+
+ // If this is a no-capture pointer argument, see if we can tell that it
+ // is impossible to alias the pointer we're checking. If not, we have to
+ // assume that the call could touch the pointer, even though it doesn't
+ // escape.
+ if (!isNoAlias(cast<Value>(CI), ~0U, P, ~0U)) {
+ PassedAsArg = true;
+ break;
+ }
+ }
+
+ if (!PassedAsArg)
+ return NoModRef;
+ }
+
+ // Finally, handle specific knowledge of intrinsics.
+ IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction());
+ if (II == 0)
+ return AliasAnalysis::getModRefInfo(CS, P, Size);
+
+ switch (II->getIntrinsicID()) {
+ default: break;
+ case Intrinsic::memcpy:
+ case Intrinsic::memmove: {
+ unsigned Len = ~0U;
+ if (ConstantInt *LenCI = dyn_cast<ConstantInt>(II->getOperand(3)))
+ Len = LenCI->getZExtValue();
+ Value *Dest = II->getOperand(1);
+ Value *Src = II->getOperand(2);
+ if (isNoAlias(Dest, Len, P, Size)) {
+ if (isNoAlias(Src, Len, P, Size))
+ return NoModRef;
+ return Ref;
+ }
+ break;
+ }
+ case Intrinsic::memset:
+ // Since memset is 'accesses arguments' only, the AliasAnalysis base class
+ // will handle it for the variable length case.
+ if (ConstantInt *LenCI = dyn_cast<ConstantInt>(II->getOperand(3))) {
+ unsigned Len = LenCI->getZExtValue();
+ Value *Dest = II->getOperand(1);
+ if (isNoAlias(Dest, Len, P, Size))
+ return NoModRef;
+ }
+ break;
+ case Intrinsic::atomic_cmp_swap:
+ case Intrinsic::atomic_swap:
+ case Intrinsic::atomic_load_add:
+ case Intrinsic::atomic_load_sub:
+ case Intrinsic::atomic_load_and:
+ case Intrinsic::atomic_load_nand:
+ case Intrinsic::atomic_load_or:
+ case Intrinsic::atomic_load_xor:
+ case Intrinsic::atomic_load_max:
+ case Intrinsic::atomic_load_min:
+ case Intrinsic::atomic_load_umax:
+ case Intrinsic::atomic_load_umin:
+ if (TD) {
+ Value *Op1 = II->getOperand(1);
+ unsigned Op1Size = TD->getTypeStoreSize(Op1->getType());
+ if (isNoAlias(Op1, Op1Size, P, Size))
+ return NoModRef;
+ }
+ break;
+ case Intrinsic::lifetime_start:
+ case Intrinsic::lifetime_end:
+ case Intrinsic::invariant_start: {
+ unsigned PtrSize = cast<ConstantInt>(II->getOperand(1))->getZExtValue();
+ if (isNoAlias(II->getOperand(2), PtrSize, P, Size))
+ return NoModRef;
+ break;
+ }
+ case Intrinsic::invariant_end: {
+ unsigned PtrSize = cast<ConstantInt>(II->getOperand(2))->getZExtValue();
+ if (isNoAlias(II->getOperand(3), PtrSize, P, Size))
+ return NoModRef;
+ break;
+ }
+ }
+
+ // The AliasAnalysis base class has some smarts, lets use them.
+ return AliasAnalysis::getModRefInfo(CS, P, Size);
+}
+
+
+AliasAnalysis::ModRefResult
+BasicAliasAnalysis::getModRefInfo(CallSite CS1, CallSite CS2) {
+ // If CS1 or CS2 are readnone, they don't interact.
+ ModRefBehavior CS1B = AliasAnalysis::getModRefBehavior(CS1);
+ if (CS1B == DoesNotAccessMemory) return NoModRef;
+
+ ModRefBehavior CS2B = AliasAnalysis::getModRefBehavior(CS2);
+ if (CS2B == DoesNotAccessMemory) return NoModRef;
+
+ // If they both only read from memory, just return ref.
+ if (CS1B == OnlyReadsMemory && CS2B == OnlyReadsMemory)
+ return Ref;
+
+ // Otherwise, fall back to NoAA (mod+ref).
+ return NoAA::getModRefInfo(CS1, CS2);
+}
+
+/// GetIndiceDifference - Dest and Src are the variable indices from two
+/// decomposed GetElementPtr instructions GEP1 and GEP2 which have common base
+/// pointers. Subtract the GEP2 indices from GEP1 to find the symbolic
+/// difference between the two pointers.
+static void GetIndiceDifference(
+ SmallVectorImpl<std::pair<const Value*, int64_t> > &Dest,
+ const SmallVectorImpl<std::pair<const Value*, int64_t> > &Src) {
+ if (Src.empty()) return;
+
+ for (unsigned i = 0, e = Src.size(); i != e; ++i) {
+ const Value *V = Src[i].first;
+ int64_t Scale = Src[i].second;
+
+ // Find V in Dest. This is N^2, but pointer indices almost never have more
+ // than a few variable indexes.
+ for (unsigned j = 0, e = Dest.size(); j != e; ++j) {
+ if (Dest[j].first != V) continue;
+
+ // If we found it, subtract off Scale V's from the entry in Dest. If it
+ // goes to zero, remove the entry.
+ if (Dest[j].second != Scale)
+ Dest[j].second -= Scale;
+ else
+ Dest.erase(Dest.begin()+j);
+ Scale = 0;
+ break;
+ }
+
+ // If we didn't consume this entry, add it to the end of the Dest list.
+ if (Scale)
+ Dest.push_back(std::make_pair(V, -Scale));
+ }
+}
+
+/// aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP instruction
+/// against another pointer. We know that V1 is a GEP, but we don't know
+/// anything about V2. UnderlyingV1 is GEP1->getUnderlyingObject(),
+/// UnderlyingV2 is the same for V2.
+///
+AliasAnalysis::AliasResult
+BasicAliasAnalysis::aliasGEP(const GEPOperator *GEP1, unsigned V1Size,
+ const Value *V2, unsigned V2Size,
+ const Value *UnderlyingV1,
+ const Value *UnderlyingV2) {
+ int64_t GEP1BaseOffset;
+ SmallVector<std::pair<const Value*, int64_t>, 4> GEP1VariableIndices;
+
+ // If we have two gep instructions with must-alias'ing base pointers, figure
+ // out if the indexes to the GEP tell us anything about the derived pointer.
+ if (const GEPOperator *GEP2 = dyn_cast<GEPOperator>(V2)) {
+ // Do the base pointers alias?
+ AliasResult BaseAlias = aliasCheck(UnderlyingV1, ~0U, UnderlyingV2, ~0U);
+
+ // If we get a No or May, then return it immediately, no amount of analysis
+ // will improve this situation.
+ if (BaseAlias != MustAlias) return BaseAlias;
+
+ // Otherwise, we have a MustAlias. Since the base pointers alias each other
+ // exactly, see if the computed offset from the common pointer tells us
+ // about the relation of the resulting pointer.
+ const Value *GEP1BasePtr =
+ DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices, TD);
+
+ int64_t GEP2BaseOffset;
+ SmallVector<std::pair<const Value*, int64_t>, 4> GEP2VariableIndices;
+ const Value *GEP2BasePtr =
+ DecomposeGEPExpression(GEP2, GEP2BaseOffset, GEP2VariableIndices, TD);
+
+ // If DecomposeGEPExpression isn't able to look all the way through the
+ // addressing operation, we must not have TD and this is too complex for us
+ // to handle without it.
+ if (GEP1BasePtr != UnderlyingV1 || GEP2BasePtr != UnderlyingV2) {
+ assert(TD == 0 &&
+ "DecomposeGEPExpression and getUnderlyingObject disagree!");
+ return MayAlias;
+ }
+
+ // Subtract the GEP2 pointer from the GEP1 pointer to find out their
+ // symbolic difference.
+ GEP1BaseOffset -= GEP2BaseOffset;
+ GetIndiceDifference(GEP1VariableIndices, GEP2VariableIndices);
+
+ } else {
+ // Check to see if these two pointers are related by the getelementptr
+ // instruction. If one pointer is a GEP with a non-zero index of the other
+ // pointer, we know they cannot alias.
+
+ // If both accesses are unknown size, we can't do anything useful here.
+ if (V1Size == ~0U && V2Size == ~0U)
+ return MayAlias;
+
+ AliasResult R = aliasCheck(UnderlyingV1, ~0U, V2, V2Size);
+ if (R != MustAlias)
+ // If V2 may alias GEP base pointer, conservatively returns MayAlias.
+ // If V2 is known not to alias GEP base pointer, then the two values
+ // cannot alias per GEP semantics: "A pointer value formed from a
+ // getelementptr instruction is associated with the addresses associated
+ // with the first operand of the getelementptr".
+ return R;
+
+ const Value *GEP1BasePtr =
+ DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices, TD);
+
+ // If DecomposeGEPExpression isn't able to look all the way through the
+ // addressing operation, we must not have TD and this is too complex for us
+ // to handle without it.
+ if (GEP1BasePtr != UnderlyingV1) {
+ assert(TD == 0 &&
+ "DecomposeGEPExpression and getUnderlyingObject disagree!");
+ return MayAlias;
+ }
+ }
+
+ // In the two GEP Case, if there is no difference in the offsets of the
+ // computed pointers, the resultant pointers are a must alias. This
+ // hapens when we have two lexically identical GEP's (for example).
+ //
+ // In the other case, if we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2
+ // must aliases the GEP, the end result is a must alias also.
+ if (GEP1BaseOffset == 0 && GEP1VariableIndices.empty())
+ return MustAlias;
+
+ // If we have a known constant offset, see if this offset is larger than the
+ // access size being queried. If so, and if no variable indices can remove
+ // pieces of this constant, then we know we have a no-alias. For example,
+ // &A[100] != &A.
+
+ // In order to handle cases like &A[100][i] where i is an out of range
+ // subscript, we have to ignore all constant offset pieces that are a multiple
+ // of a scaled index. Do this by removing constant offsets that are a
+ // multiple of any of our variable indices. This allows us to transform
+ // things like &A[i][1] because i has a stride of (e.g.) 8 bytes but the 1
+ // provides an offset of 4 bytes (assuming a <= 4 byte access).
+ for (unsigned i = 0, e = GEP1VariableIndices.size();
+ i != e && GEP1BaseOffset;++i)
+ if (int64_t RemovedOffset = GEP1BaseOffset/GEP1VariableIndices[i].second)
+ GEP1BaseOffset -= RemovedOffset*GEP1VariableIndices[i].second;
+
+ // If our known offset is bigger than the access size, we know we don't have
+ // an alias.
+ if (GEP1BaseOffset) {
+ if (GEP1BaseOffset >= (int64_t)V2Size ||
+ GEP1BaseOffset <= -(int64_t)V1Size)
+ return NoAlias;
+ }
+
+ return MayAlias;
+}
+
+/// aliasSelect - Provide a bunch of ad-hoc rules to disambiguate a Select
+/// instruction against another.
+AliasAnalysis::AliasResult
+BasicAliasAnalysis::aliasSelect(const SelectInst *SI, unsigned SISize,
+ const Value *V2, unsigned V2Size) {
+ // If the values are Selects with the same condition, we can do a more precise
+ // check: just check for aliases between the values on corresponding arms.
+ if (const SelectInst *SI2 = dyn_cast<SelectInst>(V2))
+ if (SI->getCondition() == SI2->getCondition()) {
+ AliasResult Alias =
+ aliasCheck(SI->getTrueValue(), SISize,
+ SI2->getTrueValue(), V2Size);
+ if (Alias == MayAlias)
+ return MayAlias;
+ AliasResult ThisAlias =
+ aliasCheck(SI->getFalseValue(), SISize,
+ SI2->getFalseValue(), V2Size);
+ if (ThisAlias != Alias)
+ return MayAlias;
+ return Alias;
+ }
+
+ // If both arms of the Select node NoAlias or MustAlias V2, then returns
+ // NoAlias / MustAlias. Otherwise, returns MayAlias.
+ AliasResult Alias =
+ aliasCheck(SI->getTrueValue(), SISize, V2, V2Size);
+ if (Alias == MayAlias)
+ return MayAlias;
+ AliasResult ThisAlias =
+ aliasCheck(SI->getFalseValue(), SISize, V2, V2Size);
+ if (ThisAlias != Alias)
+ return MayAlias;
+ return Alias;
+}
+
+// aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI instruction
+// against another.
+AliasAnalysis::AliasResult
+BasicAliasAnalysis::aliasPHI(const PHINode *PN, unsigned PNSize,
+ const Value *V2, unsigned V2Size) {
+ // The PHI node has already been visited, avoid recursion any further.
+ if (!VisitedPHIs.insert(PN))
+ return MayAlias;
+
+ // If the values are PHIs in the same block, we can do a more precise
+ // as well as efficient check: just check for aliases between the values
+ // on corresponding edges.
+ if (const PHINode *PN2 = dyn_cast<PHINode>(V2))
+ if (PN2->getParent() == PN->getParent()) {
+ AliasResult Alias =
+ aliasCheck(PN->getIncomingValue(0), PNSize,
+ PN2->getIncomingValueForBlock(PN->getIncomingBlock(0)),
+ V2Size);
+ if (Alias == MayAlias)
+ return MayAlias;
+ for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
+ AliasResult ThisAlias =
+ aliasCheck(PN->getIncomingValue(i), PNSize,
+ PN2->getIncomingValueForBlock(PN->getIncomingBlock(i)),
+ V2Size);
+ if (ThisAlias != Alias)
+ return MayAlias;
+ }
+ return Alias;
+ }
+
+ SmallPtrSet<Value*, 4> UniqueSrc;
+ SmallVector<Value*, 4> V1Srcs;
+ for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
+ Value *PV1 = PN->getIncomingValue(i);
+ if (isa<PHINode>(PV1))
+ // If any of the source itself is a PHI, return MayAlias conservatively
+ // to avoid compile time explosion. The worst possible case is if both
+ // sides are PHI nodes. In which case, this is O(m x n) time where 'm'
+ // and 'n' are the number of PHI sources.
+ return MayAlias;
+ if (UniqueSrc.insert(PV1))
+ V1Srcs.push_back(PV1);
+ }
+
+ AliasResult Alias = aliasCheck(V2, V2Size, V1Srcs[0], PNSize);
+ // Early exit if the check of the first PHI source against V2 is MayAlias.
+ // Other results are not possible.
+ if (Alias == MayAlias)
+ return MayAlias;
+
+ // If all sources of the PHI node NoAlias or MustAlias V2, then returns
+ // NoAlias / MustAlias. Otherwise, returns MayAlias.
+ for (unsigned i = 1, e = V1Srcs.size(); i != e; ++i) {
+ Value *V = V1Srcs[i];
+
+ // If V2 is a PHI, the recursive case will have been caught in the
+ // above aliasCheck call, so these subsequent calls to aliasCheck
+ // don't need to assume that V2 is being visited recursively.
+ VisitedPHIs.erase(V2);
+
+ AliasResult ThisAlias = aliasCheck(V2, V2Size, V, PNSize);
+ if (ThisAlias != Alias || ThisAlias == MayAlias)
+ return MayAlias;
+ }
+
+ return Alias;
+}
+
+// aliasCheck - Provide a bunch of ad-hoc rules to disambiguate in common cases,
+// such as array references.
+//
+AliasAnalysis::AliasResult
+BasicAliasAnalysis::aliasCheck(const Value *V1, unsigned V1Size,
+ const Value *V2, unsigned V2Size) {
+ // Strip off any casts if they exist.
+ V1 = V1->stripPointerCasts();
+ V2 = V2->stripPointerCasts();
+
+ // Are we checking for alias of the same value?
+ if (V1 == V2) return MustAlias;
+
+ if (!isa<PointerType>(V1->getType()) || !isa<PointerType>(V2->getType()))
+ return NoAlias; // Scalars cannot alias each other
+
+ // Figure out what objects these things are pointing to if we can.
+ const Value *O1 = V1->getUnderlyingObject();
+ const Value *O2 = V2->getUnderlyingObject();
+
+ // Null values in the default address space don't point to any object, so they
+ // don't alias any other pointer.
+ if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O1))
+ if (CPN->getType()->getAddressSpace() == 0)
+ return NoAlias;
+ if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O2))
+ if (CPN->getType()->getAddressSpace() == 0)
+ return NoAlias;
+
+ if (O1 != O2) {
+ // If V1/V2 point to two different objects we know that we have no alias.
+ if (isIdentifiedObject(O1) && isIdentifiedObject(O2))
+ return NoAlias;
+
+ // Constant pointers can't alias with non-const isIdentifiedObject objects.
+ if ((isa<Constant>(O1) && isIdentifiedObject(O2) && !isa<Constant>(O2)) ||
+ (isa<Constant>(O2) && isIdentifiedObject(O1) && !isa<Constant>(O1)))
+ return NoAlias;
+
+ // Arguments can't alias with local allocations or noalias calls.
+ if ((isa<Argument>(O1) && (isa<AllocaInst>(O2) || isNoAliasCall(O2))) ||
+ (isa<Argument>(O2) && (isa<AllocaInst>(O1) || isNoAliasCall(O1))))
+ return NoAlias;
+
+ // Most objects can't alias null.
+ if ((isa<ConstantPointerNull>(V2) && isKnownNonNull(O1)) ||
+ (isa<ConstantPointerNull>(V1) && isKnownNonNull(O2)))
+ return NoAlias;
+ }
+
+ // If the size of one access is larger than the entire object on the other
+ // side, then we know such behavior is undefined and can assume no alias.
+ if (TD)
+ if ((V1Size != ~0U && isObjectSmallerThan(O2, V1Size, *TD)) ||
+ (V2Size != ~0U && isObjectSmallerThan(O1, V2Size, *TD)))
+ return NoAlias;
+
+ // If one pointer is the result of a call/invoke or load and the other is a
+ // non-escaping local object, then we know the object couldn't escape to a
+ // point where the call could return it. The load case works because
+ // isNonEscapingLocalObject considers all stores to be escapes (it
+ // passes true for the StoreCaptures argument to PointerMayBeCaptured).
+ if (O1 != O2) {
+ if ((isa<CallInst>(O1) || isa<InvokeInst>(O1) || isa<LoadInst>(O1) ||
+ isa<Argument>(O1)) &&
+ isNonEscapingLocalObject(O2))
+ return NoAlias;
+ if ((isa<CallInst>(O2) || isa<InvokeInst>(O2) || isa<LoadInst>(O2) ||
+ isa<Argument>(O2)) &&
+ isNonEscapingLocalObject(O1))
+ return NoAlias;
+ }
+
+ // FIXME: This isn't aggressively handling alias(GEP, PHI) for example: if the
+ // GEP can't simplify, we don't even look at the PHI cases.
+ if (!isa<GEPOperator>(V1) && isa<GEPOperator>(V2)) {
+ std::swap(V1, V2);
+ std::swap(V1Size, V2Size);
+ std::swap(O1, O2);
+ }
+ if (const GEPOperator *GV1 = dyn_cast<GEPOperator>(V1))
+ return aliasGEP(GV1, V1Size, V2, V2Size, O1, O2);
+
+ if (isa<PHINode>(V2) && !isa<PHINode>(V1)) {
+ std::swap(V1, V2);
+ std::swap(V1Size, V2Size);
+ }
+ if (const PHINode *PN = dyn_cast<PHINode>(V1))
+ return aliasPHI(PN, V1Size, V2, V2Size);
+
+ if (isa<SelectInst>(V2) && !isa<SelectInst>(V1)) {
+ std::swap(V1, V2);
+ std::swap(V1Size, V2Size);
+ }
+ if (const SelectInst *S1 = dyn_cast<SelectInst>(V1))
+ return aliasSelect(S1, V1Size, V2, V2Size);
+
+ return MayAlias;
+}
+
+// Make sure that anything that uses AliasAnalysis pulls in this file.
+DEFINING_FILE_FOR(BasicAliasAnalysis)
diff --git a/lib/Analysis/CFGPrinter.cpp b/lib/Analysis/CFGPrinter.cpp
new file mode 100644
index 0000000..e06704b
--- /dev/null
+++ b/lib/Analysis/CFGPrinter.cpp
@@ -0,0 +1,160 @@
+//===- CFGPrinter.cpp - DOT printer for the control flow graph ------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file defines a '-dot-cfg' analysis pass, which emits the
+// cfg.<fnname>.dot file for each function in the program, with a graph of the
+// CFG for that function.
+//
+// The other main feature of this file is that it implements the
+// Function::viewCFG method, which is useful for debugging passes which operate
+// on the CFG.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/CFGPrinter.h"
+
+#include "llvm/Pass.h"
+using namespace llvm;
+
+namespace {
+ struct CFGViewer : public FunctionPass {
+ static char ID; // Pass identifcation, replacement for typeid
+ CFGViewer() : FunctionPass(&ID) {}
+
+ virtual bool runOnFunction(Function &F) {
+ F.viewCFG();
+ return false;
+ }
+
+ void print(raw_ostream &OS, const Module* = 0) const {}
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.setPreservesAll();
+ }
+ };
+}
+
+char CFGViewer::ID = 0;
+static RegisterPass<CFGViewer>
+V0("view-cfg", "View CFG of function", false, true);
+
+namespace {
+ struct CFGOnlyViewer : public FunctionPass {
+ static char ID; // Pass identifcation, replacement for typeid
+ CFGOnlyViewer() : FunctionPass(&ID) {}
+
+ virtual bool runOnFunction(Function &F) {
+ F.viewCFGOnly();
+ return false;
+ }
+
+ void print(raw_ostream &OS, const Module* = 0) const {}
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.setPreservesAll();
+ }
+ };
+}
+
+char CFGOnlyViewer::ID = 0;
+static RegisterPass<CFGOnlyViewer>
+V1("view-cfg-only",
+ "View CFG of function (with no function bodies)", false, true);
+
+namespace {
+ struct CFGPrinter : public FunctionPass {
+ static char ID; // Pass identification, replacement for typeid
+ CFGPrinter() : FunctionPass(&ID) {}
+ explicit CFGPrinter(void *pid) : FunctionPass(pid) {}
+
+ virtual bool runOnFunction(Function &F) {
+ std::string Filename = "cfg." + F.getNameStr() + ".dot";
+ errs() << "Writing '" << Filename << "'...";
+
+ std::string ErrorInfo;
+ raw_fd_ostream File(Filename.c_str(), ErrorInfo);
+
+ if (ErrorInfo.empty())
+ WriteGraph(File, (const Function*)&F);
+ else
+ errs() << " error opening file for writing!";
+ errs() << "\n";
+ return false;
+ }
+
+ void print(raw_ostream &OS, const Module* = 0) const {}
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.setPreservesAll();
+ }
+ };
+}
+
+char CFGPrinter::ID = 0;
+static RegisterPass<CFGPrinter>
+P1("dot-cfg", "Print CFG of function to 'dot' file", false, true);
+
+namespace {
+ struct CFGOnlyPrinter : public FunctionPass {
+ static char ID; // Pass identification, replacement for typeid
+ CFGOnlyPrinter() : FunctionPass(&ID) {}
+ explicit CFGOnlyPrinter(void *pid) : FunctionPass(pid) {}
+ virtual bool runOnFunction(Function &F) {
+ std::string Filename = "cfg." + F.getNameStr() + ".dot";
+ errs() << "Writing '" << Filename << "'...";
+
+ std::string ErrorInfo;
+ raw_fd_ostream File(Filename.c_str(), ErrorInfo);
+
+ if (ErrorInfo.empty())
+ WriteGraph(File, (const Function*)&F, true);
+ else
+ errs() << " error opening file for writing!";
+ errs() << "\n";
+ return false;
+ }
+ void print(raw_ostream &OS, const Module* = 0) const {}
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.setPreservesAll();
+ }
+ };
+}
+
+char CFGOnlyPrinter::ID = 0;
+static RegisterPass<CFGOnlyPrinter>
+P2("dot-cfg-only",
+ "Print CFG of function to 'dot' file (with no function bodies)", false, true);
+
+/// viewCFG - This function is meant for use from the debugger. You can just
+/// say 'call F->viewCFG()' and a ghostview window should pop up from the
+/// program, displaying the CFG of the current function. This depends on there
+/// being a 'dot' and 'gv' program in your path.
+///
+void Function::viewCFG() const {
+ ViewGraph(this, "cfg" + getNameStr());
+}
+
+/// viewCFGOnly - This function is meant for use from the debugger. It works
+/// just like viewCFG, but it does not include the contents of basic blocks
+/// into the nodes, just the label. If you are only interested in the CFG t
+/// his can make the graph smaller.
+///
+void Function::viewCFGOnly() const {
+ ViewGraph(this, "cfg" + getNameStr(), true);
+}
+
+FunctionPass *llvm::createCFGPrinterPass () {
+ return new CFGPrinter();
+}
+
+FunctionPass *llvm::createCFGOnlyPrinterPass () {
+ return new CFGOnlyPrinter();
+}
+
diff --git a/lib/Analysis/CMakeLists.txt b/lib/Analysis/CMakeLists.txt
new file mode 100644
index 0000000..17c9b86
--- /dev/null
+++ b/lib/Analysis/CMakeLists.txt
@@ -0,0 +1,46 @@
+add_llvm_library(LLVMAnalysis
+ AliasAnalysis.cpp
+ AliasAnalysisCounter.cpp
+ AliasAnalysisEvaluator.cpp
+ AliasDebugger.cpp
+ AliasSetTracker.cpp
+ Analysis.cpp
+ BasicAliasAnalysis.cpp
+ CFGPrinter.cpp
+ CaptureTracking.cpp
+ ConstantFolding.cpp
+ DbgInfoPrinter.cpp
+ DebugInfo.cpp
+ DomPrinter.cpp
+ IVUsers.cpp
+ InlineCost.cpp
+ InstCount.cpp
+ InstructionSimplify.cpp
+ Interval.cpp
+ IntervalPartition.cpp
+ LazyValueInfo.cpp
+ LibCallAliasAnalysis.cpp
+ LibCallSemantics.cpp
+ LiveValues.cpp
+ LoopDependenceAnalysis.cpp
+ LoopInfo.cpp
+ LoopPass.cpp
+ MemoryBuiltins.cpp
+ MemoryDependenceAnalysis.cpp
+ PHITransAddr.cpp
+ PointerTracking.cpp
+ PostDominators.cpp
+ ProfileEstimatorPass.cpp
+ ProfileInfo.cpp
+ ProfileInfoLoader.cpp
+ ProfileInfoLoaderPass.cpp
+ ProfileVerifierPass.cpp
+ ScalarEvolution.cpp
+ ScalarEvolutionAliasAnalysis.cpp
+ ScalarEvolutionExpander.cpp
+ SparsePropagation.cpp
+ Trace.cpp
+ ValueTracking.cpp
+ )
+
+target_link_libraries (LLVMAnalysis LLVMSupport)
diff --git a/lib/Analysis/CaptureTracking.cpp b/lib/Analysis/CaptureTracking.cpp
new file mode 100644
index 0000000..10a8b11
--- /dev/null
+++ b/lib/Analysis/CaptureTracking.cpp
@@ -0,0 +1,144 @@
+//===--- CaptureTracking.cpp - Determine whether a pointer is captured ----===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file contains routines that help determine which pointers are captured.
+// A pointer value is captured if the function makes a copy of any part of the
+// pointer that outlives the call. Not being captured means, more or less, that
+// the pointer is only dereferenced and not stored in a global. Returning part
+// of the pointer as the function return value may or may not count as capturing
+// the pointer, depending on the context.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/CaptureTracking.h"
+#include "llvm/Instructions.h"
+#include "llvm/Value.h"
+#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/ADT/SmallSet.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/Support/CallSite.h"
+using namespace llvm;
+
+/// As its comment mentions, PointerMayBeCaptured can be expensive.
+/// However, it's not easy for BasicAA to cache the result, because
+/// it's an ImmutablePass. To work around this, bound queries at a
+/// fixed number of uses.
+///
+/// TODO: Write a new FunctionPass AliasAnalysis so that it can keep
+/// a cache. Then we can move the code from BasicAliasAnalysis into
+/// that path, and remove this threshold.
+static int const Threshold = 20;
+
+/// PointerMayBeCaptured - Return true if this pointer value may be captured
+/// by the enclosing function (which is required to exist). This routine can
+/// be expensive, so consider caching the results. The boolean ReturnCaptures
+/// specifies whether returning the value (or part of it) from the function
+/// counts as capturing it or not. The boolean StoreCaptures specified whether
+/// storing the value (or part of it) into memory anywhere automatically
+/// counts as capturing it or not.
+bool llvm::PointerMayBeCaptured(const Value *V,
+ bool ReturnCaptures, bool StoreCaptures) {
+ assert(isa<PointerType>(V->getType()) && "Capture is for pointers only!");
+ SmallVector<Use*, Threshold> Worklist;
+ SmallSet<Use*, Threshold> Visited;
+ int Count = 0;
+
+ for (Value::use_const_iterator UI = V->use_begin(), UE = V->use_end();
+ UI != UE; ++UI) {
+ // If there are lots of uses, conservatively say that the value
+ // is captured to avoid taking too much compile time.
+ if (Count++ >= Threshold)
+ return true;
+
+ Use *U = &UI.getUse();
+ Visited.insert(U);
+ Worklist.push_back(U);
+ }
+
+ while (!Worklist.empty()) {
+ Use *U = Worklist.pop_back_val();
+ Instruction *I = cast<Instruction>(U->getUser());
+ V = U->get();
+
+ switch (I->getOpcode()) {
+ case Instruction::Call:
+ case Instruction::Invoke: {
+ CallSite CS = CallSite::get(I);
+ // Not captured if the callee is readonly, doesn't return a copy through
+ // its return value and doesn't unwind (a readonly function can leak bits
+ // by throwing an exception or not depending on the input value).
+ if (CS.onlyReadsMemory() && CS.doesNotThrow() && I->getType()->isVoidTy())
+ break;
+
+ // Not captured if only passed via 'nocapture' arguments. Note that
+ // calling a function pointer does not in itself cause the pointer to
+ // be captured. This is a subtle point considering that (for example)
+ // the callee might return its own address. It is analogous to saying
+ // that loading a value from a pointer does not cause the pointer to be
+ // captured, even though the loaded value might be the pointer itself
+ // (think of self-referential objects).
+ CallSite::arg_iterator B = CS.arg_begin(), E = CS.arg_end();
+ for (CallSite::arg_iterator A = B; A != E; ++A)
+ if (A->get() == V && !CS.paramHasAttr(A - B + 1, Attribute::NoCapture))
+ // The parameter is not marked 'nocapture' - captured.
+ return true;
+ // Only passed via 'nocapture' arguments, or is the called function - not
+ // captured.
+ break;
+ }
+ case Instruction::Load:
+ // Loading from a pointer does not cause it to be captured.
+ break;
+ case Instruction::Ret:
+ if (ReturnCaptures)
+ return true;
+ break;
+ case Instruction::Store:
+ if (V == I->getOperand(0))
+ // Stored the pointer - conservatively assume it may be captured.
+ // TODO: If StoreCaptures is not true, we could do Fancy analysis
+ // to determine whether this store is not actually an escape point.
+ // In that case, BasicAliasAnalysis should be updated as well to
+ // take advantage of this.
+ return true;
+ // Storing to the pointee does not cause the pointer to be captured.
+ break;
+ case Instruction::BitCast:
+ case Instruction::GetElementPtr:
+ case Instruction::PHI:
+ case Instruction::Select:
+ // The original value is not captured via this if the new value isn't.
+ for (Instruction::use_iterator UI = I->use_begin(), UE = I->use_end();
+ UI != UE; ++UI) {
+ Use *U = &UI.getUse();
+ if (Visited.insert(U))
+ Worklist.push_back(U);
+ }
+ break;
+ case Instruction::ICmp:
+ // Don't count comparisons of a no-alias return value against null as
+ // captures. This allows us to ignore comparisons of malloc results
+ // with null, for example.
+ if (isNoAliasCall(V->stripPointerCasts()))
+ if (ConstantPointerNull *CPN =
+ dyn_cast<ConstantPointerNull>(I->getOperand(1)))
+ if (CPN->getType()->getAddressSpace() == 0)
+ break;
+ // Otherwise, be conservative. There are crazy ways to capture pointers
+ // using comparisons.
+ return true;
+ default:
+ // Something else - be conservative and say it is captured.
+ return true;
+ }
+ }
+
+ // All uses examined - not captured.
+ return false;
+}
diff --git a/lib/Analysis/ConstantFolding.cpp b/lib/Analysis/ConstantFolding.cpp
new file mode 100644
index 0000000..ba87040
--- /dev/null
+++ b/lib/Analysis/ConstantFolding.cpp
@@ -0,0 +1,1269 @@
+//===-- ConstantFolding.cpp - Fold instructions into constants ------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file defines routines for folding instructions into constants.
+//
+// Also, to supplement the basic VMCore ConstantExpr simplifications,
+// this file defines some additional folding routines that can make use of
+// TargetData information. These functions cannot go in VMCore due to library
+// dependency issues.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/ConstantFolding.h"
+#include "llvm/Constants.h"
+#include "llvm/DerivedTypes.h"
+#include "llvm/Function.h"
+#include "llvm/GlobalVariable.h"
+#include "llvm/Instructions.h"
+#include "llvm/Intrinsics.h"
+#include "llvm/Analysis/ValueTracking.h"
+#include "llvm/Target/TargetData.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/ADT/StringMap.h"
+#include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/GetElementPtrTypeIterator.h"
+#include "llvm/Support/MathExtras.h"
+#include <cerrno>
+#include <cmath>
+using namespace llvm;
+
+//===----------------------------------------------------------------------===//
+// Constant Folding internal helper functions
+//===----------------------------------------------------------------------===//
+
+/// FoldBitCast - Constant fold bitcast, symbolically evaluating it with
+/// TargetData. This always returns a non-null constant, but it may be a
+/// ConstantExpr if unfoldable.
+static Constant *FoldBitCast(Constant *C, const Type *DestTy,
+ const TargetData &TD) {
+
+ // This only handles casts to vectors currently.
+ const VectorType *DestVTy = dyn_cast<VectorType>(DestTy);
+ if (DestVTy == 0)
+ return ConstantExpr::getBitCast(C, DestTy);
+
+ // If this is a scalar -> vector cast, convert the input into a <1 x scalar>
+ // vector so the code below can handle it uniformly.
+ if (isa<ConstantFP>(C) || isa<ConstantInt>(C)) {
+ Constant *Ops = C; // don't take the address of C!
+ return FoldBitCast(ConstantVector::get(&Ops, 1), DestTy, TD);
+ }
+
+ // If this is a bitcast from constant vector -> vector, fold it.
+ ConstantVector *CV = dyn_cast<ConstantVector>(C);
+ if (CV == 0)
+ return ConstantExpr::getBitCast(C, DestTy);
+
+ // If the element types match, VMCore can fold it.
+ unsigned NumDstElt = DestVTy->getNumElements();
+ unsigned NumSrcElt = CV->getNumOperands();
+ if (NumDstElt == NumSrcElt)
+ return ConstantExpr::getBitCast(C, DestTy);
+
+ const Type *SrcEltTy = CV->getType()->getElementType();
+ const Type *DstEltTy = DestVTy->getElementType();
+
+ // Otherwise, we're changing the number of elements in a vector, which
+ // requires endianness information to do the right thing. For example,
+ // bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
+ // folds to (little endian):
+ // <4 x i32> <i32 0, i32 0, i32 1, i32 0>
+ // and to (big endian):
+ // <4 x i32> <i32 0, i32 0, i32 0, i32 1>
+
+ // First thing is first. We only want to think about integer here, so if
+ // we have something in FP form, recast it as integer.
+ if (DstEltTy->isFloatingPoint()) {
+ // Fold to an vector of integers with same size as our FP type.
+ unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
+ const Type *DestIVTy =
+ VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumDstElt);
+ // Recursively handle this integer conversion, if possible.
+ C = FoldBitCast(C, DestIVTy, TD);
+ if (!C) return ConstantExpr::getBitCast(C, DestTy);
+
+ // Finally, VMCore can handle this now that #elts line up.
+ return ConstantExpr::getBitCast(C, DestTy);
+ }
+
+ // Okay, we know the destination is integer, if the input is FP, convert
+ // it to integer first.
+ if (SrcEltTy->isFloatingPoint()) {
+ unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
+ const Type *SrcIVTy =
+ VectorType::get(IntegerType::get(C->getContext(), FPWidth), NumSrcElt);
+ // Ask VMCore to do the conversion now that #elts line up.
+ C = ConstantExpr::getBitCast(C, SrcIVTy);
+ CV = dyn_cast<ConstantVector>(C);
+ if (!CV) // If VMCore wasn't able to fold it, bail out.
+ return C;
+ }
+
+ // Now we know that the input and output vectors are both integer vectors
+ // of the same size, and that their #elements is not the same. Do the
+ // conversion here, which depends on whether the input or output has
+ // more elements.
+ bool isLittleEndian = TD.isLittleEndian();
+
+ SmallVector<Constant*, 32> Result;
+ if (NumDstElt < NumSrcElt) {
+ // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
+ Constant *Zero = Constant::getNullValue(DstEltTy);
+ unsigned Ratio = NumSrcElt/NumDstElt;
+ unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
+ unsigned SrcElt = 0;
+ for (unsigned i = 0; i != NumDstElt; ++i) {
+ // Build each element of the result.
+ Constant *Elt = Zero;
+ unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
+ for (unsigned j = 0; j != Ratio; ++j) {
+ Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(SrcElt++));
+ if (!Src) // Reject constantexpr elements.
+ return ConstantExpr::getBitCast(C, DestTy);
+
+ // Zero extend the element to the right size.
+ Src = ConstantExpr::getZExt(Src, Elt->getType());
+
+ // Shift it to the right place, depending on endianness.
+ Src = ConstantExpr::getShl(Src,
+ ConstantInt::get(Src->getType(), ShiftAmt));
+ ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
+
+ // Mix it in.
+ Elt = ConstantExpr::getOr(Elt, Src);
+ }
+ Result.push_back(Elt);
+ }
+ } else {
+ // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
+ unsigned Ratio = NumDstElt/NumSrcElt;
+ unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits();
+
+ // Loop over each source value, expanding into multiple results.
+ for (unsigned i = 0; i != NumSrcElt; ++i) {
+ Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(i));
+ if (!Src) // Reject constantexpr elements.
+ return ConstantExpr::getBitCast(C, DestTy);
+
+ unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
+ for (unsigned j = 0; j != Ratio; ++j) {
+ // Shift the piece of the value into the right place, depending on
+ // endianness.
+ Constant *Elt = ConstantExpr::getLShr(Src,
+ ConstantInt::get(Src->getType(), ShiftAmt));
+ ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
+
+ // Truncate and remember this piece.
+ Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
+ }
+ }
+ }
+
+ return ConstantVector::get(Result.data(), Result.size());
+}
+
+
+/// IsConstantOffsetFromGlobal - If this constant is actually a constant offset
+/// from a global, return the global and the constant. Because of
+/// constantexprs, this function is recursive.
+static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
+ int64_t &Offset, const TargetData &TD) {
+ // Trivial case, constant is the global.
+ if ((GV = dyn_cast<GlobalValue>(C))) {
+ Offset = 0;
+ return true;
+ }
+
+ // Otherwise, if this isn't a constant expr, bail out.
+ ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
+ if (!CE) return false;
+
+ // Look through ptr->int and ptr->ptr casts.
+ if (CE->getOpcode() == Instruction::PtrToInt ||
+ CE->getOpcode() == Instruction::BitCast)
+ return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD);
+
+ // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
+ if (CE->getOpcode() == Instruction::GetElementPtr) {
+ // Cannot compute this if the element type of the pointer is missing size
+ // info.
+ if (!cast<PointerType>(CE->getOperand(0)->getType())
+ ->getElementType()->isSized())
+ return false;
+
+ // If the base isn't a global+constant, we aren't either.
+ if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD))
+ return false;
+
+ // Otherwise, add any offset that our operands provide.
+ gep_type_iterator GTI = gep_type_begin(CE);
+ for (User::const_op_iterator i = CE->op_begin() + 1, e = CE->op_end();
+ i != e; ++i, ++GTI) {
+ ConstantInt *CI = dyn_cast<ConstantInt>(*i);
+ if (!CI) return false; // Index isn't a simple constant?
+ if (CI->getZExtValue() == 0) continue; // Not adding anything.
+
+ if (const StructType *ST = dyn_cast<StructType>(*GTI)) {
+ // N = N + Offset
+ Offset += TD.getStructLayout(ST)->getElementOffset(CI->getZExtValue());
+ } else {
+ const SequentialType *SQT = cast<SequentialType>(*GTI);
+ Offset += TD.getTypeAllocSize(SQT->getElementType())*CI->getSExtValue();
+ }
+ }
+ return true;
+ }
+
+ return false;
+}
+
+/// ReadDataFromGlobal - Recursive helper to read bits out of global. C is the
+/// constant being copied out of. ByteOffset is an offset into C. CurPtr is the
+/// pointer to copy results into and BytesLeft is the number of bytes left in
+/// the CurPtr buffer. TD is the target data.
+static bool ReadDataFromGlobal(Constant *C, uint64_t ByteOffset,
+ unsigned char *CurPtr, unsigned BytesLeft,
+ const TargetData &TD) {
+ assert(ByteOffset <= TD.getTypeAllocSize(C->getType()) &&
+ "Out of range access");
+
+ // If this element is zero or undefined, we can just return since *CurPtr is
+ // zero initialized.
+ if (isa<ConstantAggregateZero>(C) || isa<UndefValue>(C))
+ return true;
+
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
+ if (CI->getBitWidth() > 64 ||
+ (CI->getBitWidth() & 7) != 0)
+ return false;
+
+ uint64_t Val = CI->getZExtValue();
+ unsigned IntBytes = unsigned(CI->getBitWidth()/8);
+
+ for (unsigned i = 0; i != BytesLeft && ByteOffset != IntBytes; ++i) {
+ CurPtr[i] = (unsigned char)(Val >> (ByteOffset * 8));
+ ++ByteOffset;
+ }
+ return true;
+ }
+
+ if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
+ if (CFP->getType()->isDoubleTy()) {
+ C = FoldBitCast(C, Type::getInt64Ty(C->getContext()), TD);
+ return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
+ }
+ if (CFP->getType()->isFloatTy()){
+ C = FoldBitCast(C, Type::getInt32Ty(C->getContext()), TD);
+ return ReadDataFromGlobal(C, ByteOffset, CurPtr, BytesLeft, TD);
+ }
+ return false;
+ }
+
+ if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
+ const StructLayout *SL = TD.getStructLayout(CS->getType());
+ unsigned Index = SL->getElementContainingOffset(ByteOffset);
+ uint64_t CurEltOffset = SL->getElementOffset(Index);
+ ByteOffset -= CurEltOffset;
+
+ while (1) {
+ // If the element access is to the element itself and not to tail padding,
+ // read the bytes from the element.
+ uint64_t EltSize = TD.getTypeAllocSize(CS->getOperand(Index)->getType());
+
+ if (ByteOffset < EltSize &&
+ !ReadDataFromGlobal(CS->getOperand(Index), ByteOffset, CurPtr,
+ BytesLeft, TD))
+ return false;
+
+ ++Index;
+
+ // Check to see if we read from the last struct element, if so we're done.
+ if (Index == CS->getType()->getNumElements())
+ return true;
+
+ // If we read all of the bytes we needed from this element we're done.
+ uint64_t NextEltOffset = SL->getElementOffset(Index);
+
+ if (BytesLeft <= NextEltOffset-CurEltOffset-ByteOffset)
+ return true;
+
+ // Move to the next element of the struct.
+ CurPtr += NextEltOffset-CurEltOffset-ByteOffset;
+ BytesLeft -= NextEltOffset-CurEltOffset-ByteOffset;
+ ByteOffset = 0;
+ CurEltOffset = NextEltOffset;
+ }
+ // not reached.
+ }
+
+ if (ConstantArray *CA = dyn_cast<ConstantArray>(C)) {
+ uint64_t EltSize = TD.getTypeAllocSize(CA->getType()->getElementType());
+ uint64_t Index = ByteOffset / EltSize;
+ uint64_t Offset = ByteOffset - Index * EltSize;
+ for (; Index != CA->getType()->getNumElements(); ++Index) {
+ if (!ReadDataFromGlobal(CA->getOperand(Index), Offset, CurPtr,
+ BytesLeft, TD))
+ return false;
+ if (EltSize >= BytesLeft)
+ return true;
+
+ Offset = 0;
+ BytesLeft -= EltSize;
+ CurPtr += EltSize;
+ }
+ return true;
+ }
+
+ if (ConstantVector *CV = dyn_cast<ConstantVector>(C)) {
+ uint64_t EltSize = TD.getTypeAllocSize(CV->getType()->getElementType());
+ uint64_t Index = ByteOffset / EltSize;
+ uint64_t Offset = ByteOffset - Index * EltSize;
+ for (; Index != CV->getType()->getNumElements(); ++Index) {
+ if (!ReadDataFromGlobal(CV->getOperand(Index), Offset, CurPtr,
+ BytesLeft, TD))
+ return false;
+ if (EltSize >= BytesLeft)
+ return true;
+
+ Offset = 0;
+ BytesLeft -= EltSize;
+ CurPtr += EltSize;
+ }
+ return true;
+ }
+
+ // Otherwise, unknown initializer type.
+ return false;
+}
+
+static Constant *FoldReinterpretLoadFromConstPtr(Constant *C,
+ const TargetData &TD) {
+ const Type *LoadTy = cast<PointerType>(C->getType())->getElementType();
+ const IntegerType *IntType = dyn_cast<IntegerType>(LoadTy);
+
+ // If this isn't an integer load we can't fold it directly.
+ if (!IntType) {
+ // If this is a float/double load, we can try folding it as an int32/64 load
+ // and then bitcast the result. This can be useful for union cases. Note
+ // that address spaces don't matter here since we're not going to result in
+ // an actual new load.
+ const Type *MapTy;
+ if (LoadTy->isFloatTy())
+ MapTy = Type::getInt32PtrTy(C->getContext());
+ else if (LoadTy->isDoubleTy())
+ MapTy = Type::getInt64PtrTy(C->getContext());
+ else if (isa<VectorType>(LoadTy)) {
+ MapTy = IntegerType::get(C->getContext(),
+ TD.getTypeAllocSizeInBits(LoadTy));
+ MapTy = PointerType::getUnqual(MapTy);
+ } else
+ return 0;
+
+ C = FoldBitCast(C, MapTy, TD);
+ if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, TD))
+ return FoldBitCast(Res, LoadTy, TD);
+ return 0;
+ }
+
+ unsigned BytesLoaded = (IntType->getBitWidth() + 7) / 8;
+ if (BytesLoaded > 32 || BytesLoaded == 0) return 0;
+
+ GlobalValue *GVal;
+ int64_t Offset;
+ if (!IsConstantOffsetFromGlobal(C, GVal, Offset, TD))
+ return 0;
+
+ GlobalVariable *GV = dyn_cast<GlobalVariable>(GVal);
+ if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
+ !GV->getInitializer()->getType()->isSized())
+ return 0;
+
+ // If we're loading off the beginning of the global, some bytes may be valid,
+ // but we don't try to handle this.
+ if (Offset < 0) return 0;
+
+ // If we're not accessing anything in this constant, the result is undefined.
+ if (uint64_t(Offset) >= TD.getTypeAllocSize(GV->getInitializer()->getType()))
+ return UndefValue::get(IntType);
+
+ unsigned char RawBytes[32] = {0};
+ if (!ReadDataFromGlobal(GV->getInitializer(), Offset, RawBytes,
+ BytesLoaded, TD))
+ return 0;
+
+ APInt ResultVal = APInt(IntType->getBitWidth(), RawBytes[BytesLoaded-1]);
+ for (unsigned i = 1; i != BytesLoaded; ++i) {
+ ResultVal <<= 8;
+ ResultVal |= APInt(IntType->getBitWidth(), RawBytes[BytesLoaded-1-i]);
+ }
+
+ return ConstantInt::get(IntType->getContext(), ResultVal);
+}
+
+/// ConstantFoldLoadFromConstPtr - Return the value that a load from C would
+/// produce if it is constant and determinable. If this is not determinable,
+/// return null.
+Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C,
+ const TargetData *TD) {
+ // First, try the easy cases:
+ if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
+ if (GV->isConstant() && GV->hasDefinitiveInitializer())
+ return GV->getInitializer();
+
+ // If the loaded value isn't a constant expr, we can't handle it.
+ ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
+ if (!CE) return 0;
+
+ if (CE->getOpcode() == Instruction::GetElementPtr) {
+ if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
+ if (GV->isConstant() && GV->hasDefinitiveInitializer())
+ if (Constant *V =
+ ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
+ return V;
+ }
+
+ // Instead of loading constant c string, use corresponding integer value
+ // directly if string length is small enough.
+ std::string Str;
+ if (TD && GetConstantStringInfo(CE, Str) && !Str.empty()) {
+ unsigned StrLen = Str.length();
+ const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
+ unsigned NumBits = Ty->getPrimitiveSizeInBits();
+ // Replace LI with immediate integer store.
+ if ((NumBits >> 3) == StrLen + 1) {
+ APInt StrVal(NumBits, 0);
+ APInt SingleChar(NumBits, 0);
+ if (TD->isLittleEndian()) {
+ for (signed i = StrLen-1; i >= 0; i--) {
+ SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
+ StrVal = (StrVal << 8) | SingleChar;
+ }
+ } else {
+ for (unsigned i = 0; i < StrLen; i++) {
+ SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
+ StrVal = (StrVal << 8) | SingleChar;
+ }
+ // Append NULL at the end.
+ SingleChar = 0;
+ StrVal = (StrVal << 8) | SingleChar;
+ }
+ return ConstantInt::get(CE->getContext(), StrVal);
+ }
+ }
+
+ // If this load comes from anywhere in a constant global, and if the global
+ // is all undef or zero, we know what it loads.
+ if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getUnderlyingObject())){
+ if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
+ const Type *ResTy = cast<PointerType>(C->getType())->getElementType();
+ if (GV->getInitializer()->isNullValue())
+ return Constant::getNullValue(ResTy);
+ if (isa<UndefValue>(GV->getInitializer()))
+ return UndefValue::get(ResTy);
+ }
+ }
+
+ // Try hard to fold loads from bitcasted strange and non-type-safe things. We
+ // currently don't do any of this for big endian systems. It can be
+ // generalized in the future if someone is interested.
+ if (TD && TD->isLittleEndian())
+ return FoldReinterpretLoadFromConstPtr(CE, *TD);
+ return 0;
+}
+
+static Constant *ConstantFoldLoadInst(const LoadInst *LI, const TargetData *TD){
+ if (LI->isVolatile()) return 0;
+
+ if (Constant *C = dyn_cast<Constant>(LI->getOperand(0)))
+ return ConstantFoldLoadFromConstPtr(C, TD);
+
+ return 0;
+}
+
+/// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression.
+/// Attempt to symbolically evaluate the result of a binary operator merging
+/// these together. If target data info is available, it is provided as TD,
+/// otherwise TD is null.
+static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0,
+ Constant *Op1, const TargetData *TD){
+ // SROA
+
+ // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
+ // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
+ // bits.
+
+
+ // If the constant expr is something like &A[123] - &A[4].f, fold this into a
+ // constant. This happens frequently when iterating over a global array.
+ if (Opc == Instruction::Sub && TD) {
+ GlobalValue *GV1, *GV2;
+ int64_t Offs1, Offs2;
+
+ if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *TD))
+ if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *TD) &&
+ GV1 == GV2) {
+ // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
+ return ConstantInt::get(Op0->getType(), Offs1-Offs2);
+ }
+ }
+
+ return 0;
+}
+
+/// CastGEPIndices - If array indices are not pointer-sized integers,
+/// explicitly cast them so that they aren't implicitly casted by the
+/// getelementptr.
+static Constant *CastGEPIndices(Constant *const *Ops, unsigned NumOps,
+ const Type *ResultTy,
+ const TargetData *TD) {
+ if (!TD) return 0;
+ const Type *IntPtrTy = TD->getIntPtrType(ResultTy->getContext());
+
+ bool Any = false;
+ SmallVector<Constant*, 32> NewIdxs;
+ for (unsigned i = 1; i != NumOps; ++i) {
+ if ((i == 1 ||
+ !isa<StructType>(GetElementPtrInst::getIndexedType(Ops[0]->getType(),
+ reinterpret_cast<Value *const *>(Ops+1),
+ i-1))) &&
+ Ops[i]->getType() != IntPtrTy) {
+ Any = true;
+ NewIdxs.push_back(ConstantExpr::getCast(CastInst::getCastOpcode(Ops[i],
+ true,
+ IntPtrTy,
+ true),
+ Ops[i], IntPtrTy));
+ } else
+ NewIdxs.push_back(Ops[i]);
+ }
+ if (!Any) return 0;
+
+ Constant *C =
+ ConstantExpr::getGetElementPtr(Ops[0], &NewIdxs[0], NewIdxs.size());
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
+ if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
+ C = Folded;
+ return C;
+}
+
+/// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP
+/// constant expression, do so.
+static Constant *SymbolicallyEvaluateGEP(Constant *const *Ops, unsigned NumOps,
+ const Type *ResultTy,
+ const TargetData *TD) {
+ Constant *Ptr = Ops[0];
+ if (!TD || !cast<PointerType>(Ptr->getType())->getElementType()->isSized())
+ return 0;
+
+ unsigned BitWidth =
+ TD->getTypeSizeInBits(TD->getIntPtrType(Ptr->getContext()));
+ APInt BasePtr(BitWidth, 0);
+ bool BaseIsInt = true;
+ if (!Ptr->isNullValue()) {
+ // If this is a inttoptr from a constant int, we can fold this as the base,
+ // otherwise we can't.
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
+ if (CE->getOpcode() == Instruction::IntToPtr)
+ if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0))) {
+ BasePtr = Base->getValue();
+ BasePtr.zextOrTrunc(BitWidth);
+ }
+
+ if (BasePtr == 0)
+ BaseIsInt = false;
+ }
+
+ // If this is a constant expr gep that is effectively computing an
+ // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
+ for (unsigned i = 1; i != NumOps; ++i)
+ if (!isa<ConstantInt>(Ops[i]))
+ return 0;
+
+ APInt Offset = APInt(BitWidth,
+ TD->getIndexedOffset(Ptr->getType(),
+ (Value**)Ops+1, NumOps-1));
+ // If the base value for this address is a literal integer value, fold the
+ // getelementptr to the resulting integer value casted to the pointer type.
+ if (BaseIsInt) {
+ Constant *C = ConstantInt::get(Ptr->getContext(), Offset+BasePtr);
+ return ConstantExpr::getIntToPtr(C, ResultTy);
+ }
+
+ // Otherwise form a regular getelementptr. Recompute the indices so that
+ // we eliminate over-indexing of the notional static type array bounds.
+ // This makes it easy to determine if the getelementptr is "inbounds".
+ // Also, this helps GlobalOpt do SROA on GlobalVariables.
+ Ptr = cast<Constant>(Ptr->stripPointerCasts());
+ const Type *Ty = Ptr->getType();
+ SmallVector<Constant*, 32> NewIdxs;
+ do {
+ if (const SequentialType *ATy = dyn_cast<SequentialType>(Ty)) {
+ if (isa<PointerType>(ATy)) {
+ // The only pointer indexing we'll do is on the first index of the GEP.
+ if (!NewIdxs.empty())
+ break;
+
+ // Only handle pointers to sized types, not pointers to functions.
+ if (!ATy->getElementType()->isSized())
+ return 0;
+ }
+
+ // Determine which element of the array the offset points into.
+ APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType()));
+ if (ElemSize == 0)
+ return 0;
+ APInt NewIdx = Offset.udiv(ElemSize);
+ Offset -= NewIdx * ElemSize;
+ NewIdxs.push_back(ConstantInt::get(TD->getIntPtrType(Ty->getContext()),
+ NewIdx));
+ Ty = ATy->getElementType();
+ } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
+ // Determine which field of the struct the offset points into. The
+ // getZExtValue is at least as safe as the StructLayout API because we
+ // know the offset is within the struct at this point.
+ const StructLayout &SL = *TD->getStructLayout(STy);
+ unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
+ NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
+ ElIdx));
+ Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx));
+ Ty = STy->getTypeAtIndex(ElIdx);
+ } else {
+ // We've reached some non-indexable type.
+ break;
+ }
+ } while (Ty != cast<PointerType>(ResultTy)->getElementType());
+
+ // If we haven't used up the entire offset by descending the static
+ // type, then the offset is pointing into the middle of an indivisible
+ // member, so we can't simplify it.
+ if (Offset != 0)
+ return 0;
+
+ // Create a GEP.
+ Constant *C =
+ ConstantExpr::getGetElementPtr(Ptr, &NewIdxs[0], NewIdxs.size());
+ assert(cast<PointerType>(C->getType())->getElementType() == Ty &&
+ "Computed GetElementPtr has unexpected type!");
+
+ // If we ended up indexing a member with a type that doesn't match
+ // the type of what the original indices indexed, add a cast.
+ if (Ty != cast<PointerType>(ResultTy)->getElementType())
+ C = FoldBitCast(C, ResultTy, *TD);
+
+ return C;
+}
+
+
+
+//===----------------------------------------------------------------------===//
+// Constant Folding public APIs
+//===----------------------------------------------------------------------===//
+
+
+/// ConstantFoldInstruction - Attempt to constant fold the specified
+/// instruction. If successful, the constant result is returned, if not, null
+/// is returned. Note that this function can only fail when attempting to fold
+/// instructions like loads and stores, which have no constant expression form.
+///
+Constant *llvm::ConstantFoldInstruction(Instruction *I, const TargetData *TD) {
+ if (PHINode *PN = dyn_cast<PHINode>(I)) {
+ if (PN->getNumIncomingValues() == 0)
+ return UndefValue::get(PN->getType());
+
+ Constant *Result = dyn_cast<Constant>(PN->getIncomingValue(0));
+ if (Result == 0) return 0;
+
+ // Handle PHI nodes specially here...
+ for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)
+ if (PN->getIncomingValue(i) != Result && PN->getIncomingValue(i) != PN)
+ return 0; // Not all the same incoming constants...
+
+ // If we reach here, all incoming values are the same constant.
+ return Result;
+ }
+
+ // Scan the operand list, checking to see if they are all constants, if so,
+ // hand off to ConstantFoldInstOperands.
+ SmallVector<Constant*, 8> Ops;
+ for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
+ if (Constant *Op = dyn_cast<Constant>(*i))
+ Ops.push_back(Op);
+ else
+ return 0; // All operands not constant!
+
+ if (const CmpInst *CI = dyn_cast<CmpInst>(I))
+ return ConstantFoldCompareInstOperands(CI->getPredicate(), Ops[0], Ops[1],
+ TD);
+
+ if (const LoadInst *LI = dyn_cast<LoadInst>(I))
+ return ConstantFoldLoadInst(LI, TD);
+
+ return ConstantFoldInstOperands(I->getOpcode(), I->getType(),
+ Ops.data(), Ops.size(), TD);
+}
+
+/// ConstantFoldConstantExpression - Attempt to fold the constant expression
+/// using the specified TargetData. If successful, the constant result is
+/// result is returned, if not, null is returned.
+Constant *llvm::ConstantFoldConstantExpression(const ConstantExpr *CE,
+ const TargetData *TD) {
+ SmallVector<Constant*, 8> Ops;
+ for (User::const_op_iterator i = CE->op_begin(), e = CE->op_end(); i != e; ++i) {
+ Constant *NewC = cast<Constant>(*i);
+ // Recursively fold the ConstantExpr's operands.
+ if (ConstantExpr *NewCE = dyn_cast<ConstantExpr>(NewC))
+ NewC = ConstantFoldConstantExpression(NewCE, TD);
+ Ops.push_back(NewC);
+ }
+
+ if (CE->isCompare())
+ return ConstantFoldCompareInstOperands(CE->getPredicate(), Ops[0], Ops[1],
+ TD);
+ return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(),
+ Ops.data(), Ops.size(), TD);
+}
+
+/// ConstantFoldInstOperands - Attempt to constant fold an instruction with the
+/// specified opcode and operands. If successful, the constant result is
+/// returned, if not, null is returned. Note that this function can fail when
+/// attempting to fold instructions like loads and stores, which have no
+/// constant expression form.
+///
+/// TODO: This function neither utilizes nor preserves nsw/nuw/inbounds/etc
+/// information, due to only being passed an opcode and operands. Constant
+/// folding using this function strips this information.
+///
+Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, const Type *DestTy,
+ Constant* const* Ops, unsigned NumOps,
+ const TargetData *TD) {
+ // Handle easy binops first.
+ if (Instruction::isBinaryOp(Opcode)) {
+ if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1]))
+ if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD))
+ return C;
+
+ return ConstantExpr::get(Opcode, Ops[0], Ops[1]);
+ }
+
+ switch (Opcode) {
+ default: return 0;
+ case Instruction::ICmp:
+ case Instruction::FCmp: assert(0 && "Invalid for compares");
+ case Instruction::Call:
+ if (Function *F = dyn_cast<Function>(Ops[0]))
+ if (canConstantFoldCallTo(F))
+ return ConstantFoldCall(F, Ops+1, NumOps-1);
+ return 0;
+ case Instruction::PtrToInt:
+ // If the input is a inttoptr, eliminate the pair. This requires knowing
+ // the width of a pointer, so it can't be done in ConstantExpr::getCast.
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
+ if (TD && CE->getOpcode() == Instruction::IntToPtr) {
+ Constant *Input = CE->getOperand(0);
+ unsigned InWidth = Input->getType()->getScalarSizeInBits();
+ if (TD->getPointerSizeInBits() < InWidth) {
+ Constant *Mask =
+ ConstantInt::get(CE->getContext(), APInt::getLowBitsSet(InWidth,
+ TD->getPointerSizeInBits()));
+ Input = ConstantExpr::getAnd(Input, Mask);
+ }
+ // Do a zext or trunc to get to the dest size.
+ return ConstantExpr::getIntegerCast(Input, DestTy, false);
+ }
+ }
+ return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
+ case Instruction::IntToPtr:
+ // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
+ // the int size is >= the ptr size. This requires knowing the width of a
+ // pointer, so it can't be done in ConstantExpr::getCast.
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
+ if (TD &&
+ TD->getPointerSizeInBits() <=
+ CE->getType()->getScalarSizeInBits()) {
+ if (CE->getOpcode() == Instruction::PtrToInt)
+ return FoldBitCast(CE->getOperand(0), DestTy, *TD);
+
+ // If there's a constant offset added to the integer value before
+ // it is casted back to a pointer, see if the expression can be
+ // converted into a GEP.
+ if (CE->getOpcode() == Instruction::Add)
+ if (ConstantInt *L = dyn_cast<ConstantInt>(CE->getOperand(0)))
+ if (ConstantExpr *R = dyn_cast<ConstantExpr>(CE->getOperand(1)))
+ if (R->getOpcode() == Instruction::PtrToInt)
+ if (GlobalVariable *GV =
+ dyn_cast<GlobalVariable>(R->getOperand(0))) {
+ const PointerType *GVTy = cast<PointerType>(GV->getType());
+ if (const ArrayType *AT =
+ dyn_cast<ArrayType>(GVTy->getElementType())) {
+ const Type *ElTy = AT->getElementType();
+ uint64_t AllocSize = TD->getTypeAllocSize(ElTy);
+ APInt PSA(L->getValue().getBitWidth(), AllocSize);
+ if (ElTy == cast<PointerType>(DestTy)->getElementType() &&
+ L->getValue().urem(PSA) == 0) {
+ APInt ElemIdx = L->getValue().udiv(PSA);
+ if (ElemIdx.ult(APInt(ElemIdx.getBitWidth(),
+ AT->getNumElements()))) {
+ Constant *Index[] = {
+ Constant::getNullValue(CE->getType()),
+ ConstantInt::get(ElTy->getContext(), ElemIdx)
+ };
+ return
+ ConstantExpr::getGetElementPtr(GV, &Index[0], 2);
+ }
+ }
+ }
+ }
+ }
+ }
+ return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
+ case Instruction::Trunc:
+ case Instruction::ZExt:
+ case Instruction::SExt:
+ case Instruction::FPTrunc:
+ case Instruction::FPExt:
+ case Instruction::UIToFP:
+ case Instruction::SIToFP:
+ case Instruction::FPToUI:
+ case Instruction::FPToSI:
+ return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
+ case Instruction::BitCast:
+ if (TD)
+ return FoldBitCast(Ops[0], DestTy, *TD);
+ return ConstantExpr::getBitCast(Ops[0], DestTy);
+ case Instruction::Select:
+ return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
+ case Instruction::ExtractElement:
+ return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
+ case Instruction::InsertElement:
+ return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
+ case Instruction::ShuffleVector:
+ return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
+ case Instruction::GetElementPtr:
+ if (Constant *C = CastGEPIndices(Ops, NumOps, DestTy, TD))
+ return C;
+ if (Constant *C = SymbolicallyEvaluateGEP(Ops, NumOps, DestTy, TD))
+ return C;
+
+ return ConstantExpr::getGetElementPtr(Ops[0], Ops+1, NumOps-1);
+ }
+}
+
+/// ConstantFoldCompareInstOperands - Attempt to constant fold a compare
+/// instruction (icmp/fcmp) with the specified operands. If it fails, it
+/// returns a constant expression of the specified operands.
+///
+Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
+ Constant *Ops0, Constant *Ops1,
+ const TargetData *TD) {
+ // fold: icmp (inttoptr x), null -> icmp x, 0
+ // fold: icmp (ptrtoint x), 0 -> icmp x, null
+ // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
+ // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
+ //
+ // ConstantExpr::getCompare cannot do this, because it doesn't have TD
+ // around to know if bit truncation is happening.
+ if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops0)) {
+ if (TD && Ops1->isNullValue()) {
+ const Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
+ if (CE0->getOpcode() == Instruction::IntToPtr) {
+ // Convert the integer value to the right size to ensure we get the
+ // proper extension or truncation.
+ Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
+ IntPtrTy, false);
+ Constant *Null = Constant::getNullValue(C->getType());
+ return ConstantFoldCompareInstOperands(Predicate, C, Null, TD);
+ }
+
+ // Only do this transformation if the int is intptrty in size, otherwise
+ // there is a truncation or extension that we aren't modeling.
+ if (CE0->getOpcode() == Instruction::PtrToInt &&
+ CE0->getType() == IntPtrTy) {
+ Constant *C = CE0->getOperand(0);
+ Constant *Null = Constant::getNullValue(C->getType());
+ return ConstantFoldCompareInstOperands(Predicate, C, Null, TD);
+ }
+ }
+
+ if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops1)) {
+ if (TD && CE0->getOpcode() == CE1->getOpcode()) {
+ const Type *IntPtrTy = TD->getIntPtrType(CE0->getContext());
+
+ if (CE0->getOpcode() == Instruction::IntToPtr) {
+ // Convert the integer value to the right size to ensure we get the
+ // proper extension or truncation.
+ Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
+ IntPtrTy, false);
+ Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
+ IntPtrTy, false);
+ return ConstantFoldCompareInstOperands(Predicate, C0, C1, TD);
+ }
+
+ // Only do this transformation if the int is intptrty in size, otherwise
+ // there is a truncation or extension that we aren't modeling.
+ if ((CE0->getOpcode() == Instruction::PtrToInt &&
+ CE0->getType() == IntPtrTy &&
+ CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType()))
+ return ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0),
+ CE1->getOperand(0), TD);
+ }
+ }
+
+ // icmp eq (or x, y), 0 -> (icmp eq x, 0) & (icmp eq y, 0)
+ // icmp ne (or x, y), 0 -> (icmp ne x, 0) | (icmp ne y, 0)
+ if ((Predicate == ICmpInst::ICMP_EQ || Predicate == ICmpInst::ICMP_NE) &&
+ CE0->getOpcode() == Instruction::Or && Ops1->isNullValue()) {
+ Constant *LHS =
+ ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(0), Ops1,TD);
+ Constant *RHS =
+ ConstantFoldCompareInstOperands(Predicate, CE0->getOperand(1), Ops1,TD);
+ unsigned OpC =
+ Predicate == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
+ Constant *Ops[] = { LHS, RHS };
+ return ConstantFoldInstOperands(OpC, LHS->getType(), Ops, 2, TD);
+ }
+ }
+
+ return ConstantExpr::getCompare(Predicate, Ops0, Ops1);
+}
+
+
+/// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a
+/// getelementptr constantexpr, return the constant value being addressed by the
+/// constant expression, or null if something is funny and we can't decide.
+Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
+ ConstantExpr *CE) {
+ if (CE->getOperand(1) != Constant::getNullValue(CE->getOperand(1)->getType()))
+ return 0; // Do not allow stepping over the value!
+
+ // Loop over all of the operands, tracking down which value we are
+ // addressing...
+ gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
+ for (++I; I != E; ++I)
+ if (const StructType *STy = dyn_cast<StructType>(*I)) {
+ ConstantInt *CU = cast<ConstantInt>(I.getOperand());
+ assert(CU->getZExtValue() < STy->getNumElements() &&
+ "Struct index out of range!");
+ unsigned El = (unsigned)CU->getZExtValue();
+ if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
+ C = CS->getOperand(El);
+ } else if (isa<ConstantAggregateZero>(C)) {
+ C = Constant::getNullValue(STy->getElementType(El));
+ } else if (isa<UndefValue>(C)) {
+ C = UndefValue::get(STy->getElementType(El));
+ } else {
+ return 0;
+ }
+ } else if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand())) {
+ if (const ArrayType *ATy = dyn_cast<ArrayType>(*I)) {
+ if (CI->getZExtValue() >= ATy->getNumElements())
+ return 0;
+ if (ConstantArray *CA = dyn_cast<ConstantArray>(C))
+ C = CA->getOperand(CI->getZExtValue());
+ else if (isa<ConstantAggregateZero>(C))
+ C = Constant::getNullValue(ATy->getElementType());
+ else if (isa<UndefValue>(C))
+ C = UndefValue::get(ATy->getElementType());
+ else
+ return 0;
+ } else if (const VectorType *VTy = dyn_cast<VectorType>(*I)) {
+ if (CI->getZExtValue() >= VTy->getNumElements())
+ return 0;
+ if (ConstantVector *CP = dyn_cast<ConstantVector>(C))
+ C = CP->getOperand(CI->getZExtValue());
+ else if (isa<ConstantAggregateZero>(C))
+ C = Constant::getNullValue(VTy->getElementType());
+ else if (isa<UndefValue>(C))
+ C = UndefValue::get(VTy->getElementType());
+ else
+ return 0;
+ } else {
+ return 0;
+ }
+ } else {
+ return 0;
+ }
+ return C;
+}
+
+
+//===----------------------------------------------------------------------===//
+// Constant Folding for Calls
+//
+
+/// canConstantFoldCallTo - Return true if its even possible to fold a call to
+/// the specified function.
+bool
+llvm::canConstantFoldCallTo(const Function *F) {
+ switch (F->getIntrinsicID()) {
+ case Intrinsic::sqrt:
+ case Intrinsic::powi:
+ case Intrinsic::bswap:
+ case Intrinsic::ctpop:
+ case Intrinsic::ctlz:
+ case Intrinsic::cttz:
+ case Intrinsic::uadd_with_overflow:
+ case Intrinsic::usub_with_overflow:
+ case Intrinsic::sadd_with_overflow:
+ case Intrinsic::ssub_with_overflow:
+ return true;
+ default:
+ return false;
+ case 0: break;
+ }
+
+ if (!F->hasName()) return false;
+ StringRef Name = F->getName();
+
+ // In these cases, the check of the length is required. We don't want to
+ // return true for a name like "cos\0blah" which strcmp would return equal to
+ // "cos", but has length 8.
+ switch (Name[0]) {
+ default: return false;
+ case 'a':
+ return Name == "acos" || Name == "asin" ||
+ Name == "atan" || Name == "atan2";
+ case 'c':
+ return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh";
+ case 'e':
+ return Name == "exp";
+ case 'f':
+ return Name == "fabs" || Name == "fmod" || Name == "floor";
+ case 'l':
+ return Name == "log" || Name == "log10";
+ case 'p':
+ return Name == "pow";
+ case 's':
+ return Name == "sin" || Name == "sinh" || Name == "sqrt" ||
+ Name == "sinf" || Name == "sqrtf";
+ case 't':
+ return Name == "tan" || Name == "tanh";
+ }
+}
+
+static Constant *ConstantFoldFP(double (*NativeFP)(double), double V,
+ const Type *Ty) {
+ errno = 0;
+ V = NativeFP(V);
+ if (errno != 0) {
+ errno = 0;
+ return 0;
+ }
+
+ if (Ty->isFloatTy())
+ return ConstantFP::get(Ty->getContext(), APFloat((float)V));
+ if (Ty->isDoubleTy())
+ return ConstantFP::get(Ty->getContext(), APFloat(V));
+ llvm_unreachable("Can only constant fold float/double");
+ return 0; // dummy return to suppress warning
+}
+
+static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),
+ double V, double W, const Type *Ty) {
+ errno = 0;
+ V = NativeFP(V, W);
+ if (errno != 0) {
+ errno = 0;
+ return 0;
+ }
+
+ if (Ty->isFloatTy())
+ return ConstantFP::get(Ty->getContext(), APFloat((float)V));
+ if (Ty->isDoubleTy())
+ return ConstantFP::get(Ty->getContext(), APFloat(V));
+ llvm_unreachable("Can only constant fold float/double");
+ return 0; // dummy return to suppress warning
+}
+
+/// ConstantFoldCall - Attempt to constant fold a call to the specified function
+/// with the specified arguments, returning null if unsuccessful.
+Constant *
+llvm::ConstantFoldCall(Function *F,
+ Constant *const *Operands, unsigned NumOperands) {
+ if (!F->hasName()) return 0;
+ StringRef Name = F->getName();
+
+ const Type *Ty = F->getReturnType();
+ if (NumOperands == 1) {
+ if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) {
+ if (!Ty->isFloatTy() && !Ty->isDoubleTy())
+ return 0;
+ /// Currently APFloat versions of these functions do not exist, so we use
+ /// the host native double versions. Float versions are not called
+ /// directly but for all these it is true (float)(f((double)arg)) ==
+ /// f(arg). Long double not supported yet.
+ double V = Ty->isFloatTy() ? (double)Op->getValueAPF().convertToFloat() :
+ Op->getValueAPF().convertToDouble();
+ switch (Name[0]) {
+ case 'a':
+ if (Name == "acos")
+ return ConstantFoldFP(acos, V, Ty);
+ else if (Name == "asin")
+ return ConstantFoldFP(asin, V, Ty);
+ else if (Name == "atan")
+ return ConstantFoldFP(atan, V, Ty);
+ break;
+ case 'c':
+ if (Name == "ceil")
+ return ConstantFoldFP(ceil, V, Ty);
+ else if (Name == "cos")
+ return ConstantFoldFP(cos, V, Ty);
+ else if (Name == "cosh")
+ return ConstantFoldFP(cosh, V, Ty);
+ else if (Name == "cosf")
+ return ConstantFoldFP(cos, V, Ty);
+ break;
+ case 'e':
+ if (Name == "exp")
+ return ConstantFoldFP(exp, V, Ty);
+ break;
+ case 'f':
+ if (Name == "fabs")
+ return ConstantFoldFP(fabs, V, Ty);
+ else if (Name == "floor")
+ return ConstantFoldFP(floor, V, Ty);
+ break;
+ case 'l':
+ if (Name == "log" && V > 0)
+ return ConstantFoldFP(log, V, Ty);
+ else if (Name == "log10" && V > 0)
+ return ConstantFoldFP(log10, V, Ty);
+ else if (Name == "llvm.sqrt.f32" ||
+ Name == "llvm.sqrt.f64") {
+ if (V >= -0.0)
+ return ConstantFoldFP(sqrt, V, Ty);
+ else // Undefined
+ return Constant::getNullValue(Ty);
+ }
+ break;
+ case 's':
+ if (Name == "sin")
+ return ConstantFoldFP(sin, V, Ty);
+ else if (Name == "sinh")
+ return ConstantFoldFP(sinh, V, Ty);
+ else if (Name == "sqrt" && V >= 0)
+ return ConstantFoldFP(sqrt, V, Ty);
+ else if (Name == "sqrtf" && V >= 0)
+ return ConstantFoldFP(sqrt, V, Ty);
+ else if (Name == "sinf")
+ return ConstantFoldFP(sin, V, Ty);
+ break;
+ case 't':
+ if (Name == "tan")
+ return ConstantFoldFP(tan, V, Ty);
+ else if (Name == "tanh")
+ return ConstantFoldFP(tanh, V, Ty);
+ break;
+ default:
+ break;
+ }
+ return 0;
+ }
+
+
+ if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) {
+ if (Name.startswith("llvm.bswap"))
+ return ConstantInt::get(F->getContext(), Op->getValue().byteSwap());
+ else if (Name.startswith("llvm.ctpop"))
+ return ConstantInt::get(Ty, Op->getValue().countPopulation());
+ else if (Name.startswith("llvm.cttz"))
+ return ConstantInt::get(Ty, Op->getValue().countTrailingZeros());
+ else if (Name.startswith("llvm.ctlz"))
+ return ConstantInt::get(Ty, Op->getValue().countLeadingZeros());
+ return 0;
+ }
+
+ return 0;
+ }
+
+ if (NumOperands == 2) {
+ if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
+ if (!Ty->isFloatTy() && !Ty->isDoubleTy())
+ return 0;
+ double Op1V = Ty->isFloatTy() ?
+ (double)Op1->getValueAPF().convertToFloat() :
+ Op1->getValueAPF().convertToDouble();
+ if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
+ if (Op2->getType() != Op1->getType())
+ return 0;
+
+ double Op2V = Ty->isFloatTy() ?
+ (double)Op2->getValueAPF().convertToFloat():
+ Op2->getValueAPF().convertToDouble();
+
+ if (Name == "pow")
+ return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
+ if (Name == "fmod")
+ return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty);
+ if (Name == "atan2")
+ return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
+ } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
+ if (Name == "llvm.powi.f32")
+ return ConstantFP::get(F->getContext(),
+ APFloat((float)std::pow((float)Op1V,
+ (int)Op2C->getZExtValue())));
+ if (Name == "llvm.powi.f64")
+ return ConstantFP::get(F->getContext(),
+ APFloat((double)std::pow((double)Op1V,
+ (int)Op2C->getZExtValue())));
+ }
+ return 0;
+ }
+
+
+ if (ConstantInt *Op1 = dyn_cast<ConstantInt>(Operands[0])) {
+ if (ConstantInt *Op2 = dyn_cast<ConstantInt>(Operands[1])) {
+ switch (F->getIntrinsicID()) {
+ default: break;
+ case Intrinsic::uadd_with_overflow: {
+ Constant *Res = ConstantExpr::getAdd(Op1, Op2); // result.
+ Constant *Ops[] = {
+ Res, ConstantExpr::getICmp(CmpInst::ICMP_ULT, Res, Op1) // overflow.
+ };
+ return ConstantStruct::get(F->getContext(), Ops, 2, false);
+ }
+ case Intrinsic::usub_with_overflow: {
+ Constant *Res = ConstantExpr::getSub(Op1, Op2); // result.
+ Constant *Ops[] = {
+ Res, ConstantExpr::getICmp(CmpInst::ICMP_UGT, Res, Op1) // overflow.
+ };
+ return ConstantStruct::get(F->getContext(), Ops, 2, false);
+ }
+ case Intrinsic::sadd_with_overflow: {
+ Constant *Res = ConstantExpr::getAdd(Op1, Op2); // result.
+ Constant *Overflow = ConstantExpr::getSelect(
+ ConstantExpr::getICmp(CmpInst::ICMP_SGT,
+ ConstantInt::get(Op1->getType(), 0), Op1),
+ ConstantExpr::getICmp(CmpInst::ICMP_SGT, Res, Op2),
+ ConstantExpr::getICmp(CmpInst::ICMP_SLT, Res, Op2)); // overflow.
+
+ Constant *Ops[] = { Res, Overflow };
+ return ConstantStruct::get(F->getContext(), Ops, 2, false);
+ }
+ case Intrinsic::ssub_with_overflow: {
+ Constant *Res = ConstantExpr::getSub(Op1, Op2); // result.
+ Constant *Overflow = ConstantExpr::getSelect(
+ ConstantExpr::getICmp(CmpInst::ICMP_SGT,
+ ConstantInt::get(Op2->getType(), 0), Op2),
+ ConstantExpr::getICmp(CmpInst::ICMP_SLT, Res, Op1),
+ ConstantExpr::getICmp(CmpInst::ICMP_SGT, Res, Op1)); // overflow.
+
+ Constant *Ops[] = { Res, Overflow };
+ return ConstantStruct::get(F->getContext(), Ops, 2, false);
+ }
+ }
+ }
+
+ return 0;
+ }
+ return 0;
+ }
+ return 0;
+}
+
diff --git a/lib/Analysis/DbgInfoPrinter.cpp b/lib/Analysis/DbgInfoPrinter.cpp
new file mode 100644
index 0000000..3532b05
--- /dev/null
+++ b/lib/Analysis/DbgInfoPrinter.cpp
@@ -0,0 +1,105 @@
+//===- DbgInfoPrinter.cpp - Print debug info in a human readable form ------==//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements a pass that prints instructions, and associated debug
+// info:
+//
+// - source/line/col information
+// - original variable name
+// - original type name
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Pass.h"
+#include "llvm/Function.h"
+#include "llvm/IntrinsicInst.h"
+#include "llvm/Metadata.h"
+#include "llvm/Assembly/Writer.h"
+#include "llvm/Analysis/DebugInfo.h"
+#include "llvm/Analysis/Passes.h"
+#include "llvm/Support/CFG.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/raw_ostream.h"
+
+using namespace llvm;
+
+static cl::opt<bool>
+PrintDirectory("print-fullpath",
+ cl::desc("Print fullpath when printing debug info"),
+ cl::Hidden);
+
+namespace {
+ class PrintDbgInfo : public FunctionPass {
+ raw_ostream &Out;
+ void printVariableDeclaration(const Value *V);
+ public:
+ static char ID; // Pass identification
+ PrintDbgInfo() : FunctionPass(&ID), Out(outs()) {}
+
+ virtual bool runOnFunction(Function &F);
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.setPreservesAll();
+ }
+ };
+ char PrintDbgInfo::ID = 0;
+ static RegisterPass<PrintDbgInfo> X("print-dbginfo",
+ "Print debug info in human readable form");
+}
+
+FunctionPass *llvm::createDbgInfoPrinterPass() { return new PrintDbgInfo(); }
+
+void PrintDbgInfo::printVariableDeclaration(const Value *V) {
+ std::string DisplayName, File, Directory, Type;
+ unsigned LineNo;
+
+ if (!getLocationInfo(V, DisplayName, Type, LineNo, File, Directory))
+ return;
+
+ Out << "; ";
+ WriteAsOperand(Out, V, false, 0);
+ Out << " is variable " << DisplayName
+ << " of type " << Type << " declared at ";
+
+ if (PrintDirectory)
+ Out << Directory << "/";
+
+ Out << File << ":" << LineNo << "\n";
+}
+
+bool PrintDbgInfo::runOnFunction(Function &F) {
+ if (F.isDeclaration())
+ return false;
+
+ Out << "function " << F.getName() << "\n\n";
+
+ for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
+ BasicBlock *BB = I;
+
+ if (I != F.begin() && (pred_begin(BB) == pred_end(BB)))
+ // Skip dead blocks.
+ continue;
+
+ Out << BB->getName();
+ Out << ":";
+
+ Out << "\n";
+
+ for (BasicBlock::const_iterator i = BB->begin(), e = BB->end();
+ i != e; ++i) {
+
+ printVariableDeclaration(i);
+
+ if (const User *U = dyn_cast<User>(i)) {
+ for(unsigned i=0;i<U->getNumOperands();i++)
+ printVariableDeclaration(U->getOperand(i));
+ }
+ }
+ }
+ return false;
+}
diff --git a/lib/Analysis/DebugInfo.cpp b/lib/Analysis/DebugInfo.cpp
new file mode 100644
index 0000000..258f1db
--- /dev/null
+++ b/lib/Analysis/DebugInfo.cpp
@@ -0,0 +1,1405 @@
+//===--- DebugInfo.cpp - Debug Information Helper Classes -----------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements the helper classes used to build and interpret debug
+// information in LLVM IR form.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/DebugInfo.h"
+#include "llvm/Target/TargetMachine.h" // FIXME: LAYERING VIOLATION!
+#include "llvm/Constants.h"
+#include "llvm/DerivedTypes.h"
+#include "llvm/Intrinsics.h"
+#include "llvm/IntrinsicInst.h"
+#include "llvm/Instructions.h"
+#include "llvm/Module.h"
+#include "llvm/Analysis/ValueTracking.h"
+#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/Dwarf.h"
+#include "llvm/Support/DebugLoc.h"
+#include "llvm/Support/raw_ostream.h"
+using namespace llvm;
+using namespace llvm::dwarf;
+
+//===----------------------------------------------------------------------===//
+// DIDescriptor
+//===----------------------------------------------------------------------===//
+
+/// ValidDebugInfo - Return true if V represents valid debug info value.
+/// FIXME : Add DIDescriptor.isValid()
+bool DIDescriptor::ValidDebugInfo(MDNode *N, unsigned OptLevel) {
+ if (!N)
+ return false;
+
+ DIDescriptor DI(N);
+
+ // Check current version. Allow Version6 for now.
+ unsigned Version = DI.getVersion();
+ if (Version != LLVMDebugVersion && Version != LLVMDebugVersion6)
+ return false;
+
+ switch (DI.getTag()) {
+ case DW_TAG_variable:
+ assert(DIVariable(N).Verify() && "Invalid DebugInfo value");
+ break;
+ case DW_TAG_compile_unit:
+ assert(DICompileUnit(N).Verify() && "Invalid DebugInfo value");
+ break;
+ case DW_TAG_subprogram:
+ assert(DISubprogram(N).Verify() && "Invalid DebugInfo value");
+ break;
+ case DW_TAG_lexical_block:
+ // FIXME: This interfers with the quality of generated code during
+ // optimization.
+ if (OptLevel != CodeGenOpt::None)
+ return false;
+ // FALLTHROUGH
+ default:
+ break;
+ }
+
+ return true;
+}
+
+DIDescriptor::DIDescriptor(MDNode *N, unsigned RequiredTag) {
+ DbgNode = N;
+
+ // If this is non-null, check to see if the Tag matches. If not, set to null.
+ if (N && getTag() != RequiredTag) {
+ DbgNode = 0;
+ }
+}
+
+StringRef
+DIDescriptor::getStringField(unsigned Elt) const {
+ if (DbgNode == 0)
+ return StringRef();
+
+ if (Elt < DbgNode->getNumOperands())
+ if (MDString *MDS = dyn_cast_or_null<MDString>(DbgNode->getOperand(Elt)))
+ return MDS->getString();
+
+ return StringRef();
+}
+
+uint64_t DIDescriptor::getUInt64Field(unsigned Elt) const {
+ if (DbgNode == 0)
+ return 0;
+
+ if (Elt < DbgNode->getNumOperands())
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(DbgNode->getOperand(Elt)))
+ return CI->getZExtValue();
+
+ return 0;
+}
+
+DIDescriptor DIDescriptor::getDescriptorField(unsigned Elt) const {
+ if (DbgNode == 0)
+ return DIDescriptor();
+
+ if (Elt < DbgNode->getNumOperands() && DbgNode->getOperand(Elt))
+ return DIDescriptor(dyn_cast<MDNode>(DbgNode->getOperand(Elt)));
+
+ return DIDescriptor();
+}
+
+GlobalVariable *DIDescriptor::getGlobalVariableField(unsigned Elt) const {
+ if (DbgNode == 0)
+ return 0;
+
+ if (Elt < DbgNode->getNumOperands())
+ return dyn_cast_or_null<GlobalVariable>(DbgNode->getOperand(Elt));
+ return 0;
+}
+
+unsigned DIVariable::getNumAddrElements() const {
+ return DbgNode->getNumOperands()-6;
+}
+
+
+//===----------------------------------------------------------------------===//
+// Predicates
+//===----------------------------------------------------------------------===//
+
+/// isBasicType - Return true if the specified tag is legal for
+/// DIBasicType.
+bool DIDescriptor::isBasicType() const {
+ assert(!isNull() && "Invalid descriptor!");
+ return getTag() == dwarf::DW_TAG_base_type;
+}
+
+/// isDerivedType - Return true if the specified tag is legal for DIDerivedType.
+bool DIDescriptor::isDerivedType() const {
+ assert(!isNull() && "Invalid descriptor!");
+ switch (getTag()) {
+ case dwarf::DW_TAG_typedef:
+ case dwarf::DW_TAG_pointer_type:
+ case dwarf::DW_TAG_reference_type:
+ case dwarf::DW_TAG_const_type:
+ case dwarf::DW_TAG_volatile_type:
+ case dwarf::DW_TAG_restrict_type:
+ case dwarf::DW_TAG_member:
+ case dwarf::DW_TAG_inheritance:
+ return true;
+ default:
+ // CompositeTypes are currently modelled as DerivedTypes.
+ return isCompositeType();
+ }
+}
+
+/// isCompositeType - Return true if the specified tag is legal for
+/// DICompositeType.
+bool DIDescriptor::isCompositeType() const {
+ assert(!isNull() && "Invalid descriptor!");
+ switch (getTag()) {
+ case dwarf::DW_TAG_array_type:
+ case dwarf::DW_TAG_structure_type:
+ case dwarf::DW_TAG_union_type:
+ case dwarf::DW_TAG_enumeration_type:
+ case dwarf::DW_TAG_vector_type:
+ case dwarf::DW_TAG_subroutine_type:
+ case dwarf::DW_TAG_class_type:
+ return true;
+ default:
+ return false;
+ }
+}
+
+/// isVariable - Return true if the specified tag is legal for DIVariable.
+bool DIDescriptor::isVariable() const {
+ assert(!isNull() && "Invalid descriptor!");
+ switch (getTag()) {
+ case dwarf::DW_TAG_auto_variable:
+ case dwarf::DW_TAG_arg_variable:
+ case dwarf::DW_TAG_return_variable:
+ return true;
+ default:
+ return false;
+ }
+}
+
+/// isType - Return true if the specified tag is legal for DIType.
+bool DIDescriptor::isType() const {
+ return isBasicType() || isCompositeType() || isDerivedType();
+}
+
+/// isSubprogram - Return true if the specified tag is legal for
+/// DISubprogram.
+bool DIDescriptor::isSubprogram() const {
+ assert(!isNull() && "Invalid descriptor!");
+ return getTag() == dwarf::DW_TAG_subprogram;
+}
+
+/// isGlobalVariable - Return true if the specified tag is legal for
+/// DIGlobalVariable.
+bool DIDescriptor::isGlobalVariable() const {
+ assert(!isNull() && "Invalid descriptor!");
+ return getTag() == dwarf::DW_TAG_variable;
+}
+
+/// isGlobal - Return true if the specified tag is legal for DIGlobal.
+bool DIDescriptor::isGlobal() const {
+ return isGlobalVariable();
+}
+
+/// isScope - Return true if the specified tag is one of the scope
+/// related tag.
+bool DIDescriptor::isScope() const {
+ assert(!isNull() && "Invalid descriptor!");
+ switch (getTag()) {
+ case dwarf::DW_TAG_compile_unit:
+ case dwarf::DW_TAG_lexical_block:
+ case dwarf::DW_TAG_subprogram:
+ case dwarf::DW_TAG_namespace:
+ return true;
+ default:
+ break;
+ }
+ return false;
+}
+
+/// isCompileUnit - Return true if the specified tag is DW_TAG_compile_unit.
+bool DIDescriptor::isCompileUnit() const {
+ assert(!isNull() && "Invalid descriptor!");
+ return getTag() == dwarf::DW_TAG_compile_unit;
+}
+
+/// isNameSpace - Return true if the specified tag is DW_TAG_namespace.
+bool DIDescriptor::isNameSpace() const {
+ assert(!isNull() && "Invalid descriptor!");
+ return getTag() == dwarf::DW_TAG_namespace;
+}
+
+/// isLexicalBlock - Return true if the specified tag is DW_TAG_lexical_block.
+bool DIDescriptor::isLexicalBlock() const {
+ assert(!isNull() && "Invalid descriptor!");
+ return getTag() == dwarf::DW_TAG_lexical_block;
+}
+
+/// isSubrange - Return true if the specified tag is DW_TAG_subrange_type.
+bool DIDescriptor::isSubrange() const {
+ assert(!isNull() && "Invalid descriptor!");
+ return getTag() == dwarf::DW_TAG_subrange_type;
+}
+
+/// isEnumerator - Return true if the specified tag is DW_TAG_enumerator.
+bool DIDescriptor::isEnumerator() const {
+ assert(!isNull() && "Invalid descriptor!");
+ return getTag() == dwarf::DW_TAG_enumerator;
+}
+
+//===----------------------------------------------------------------------===//
+// Simple Descriptor Constructors and other Methods
+//===----------------------------------------------------------------------===//
+
+DIType::DIType(MDNode *N) : DIDescriptor(N) {
+ if (!N) return;
+ if (!isBasicType() && !isDerivedType() && !isCompositeType()) {
+ DbgNode = 0;
+ }
+}
+
+unsigned DIArray::getNumElements() const {
+ assert(DbgNode && "Invalid DIArray");
+ return DbgNode->getNumOperands();
+}
+
+/// replaceAllUsesWith - Replace all uses of debug info referenced by
+/// this descriptor. After this completes, the current debug info value
+/// is erased.
+void DIDerivedType::replaceAllUsesWith(DIDescriptor &D) {
+ if (isNull())
+ return;
+
+ assert(!D.isNull() && "Can not replace with null");
+
+ // Since we use a TrackingVH for the node, its easy for clients to manufacture
+ // legitimate situations where they want to replaceAllUsesWith() on something
+ // which, due to uniquing, has merged with the source. We shield clients from
+ // this detail by allowing a value to be replaced with replaceAllUsesWith()
+ // itself.
+ if (getNode() != D.getNode()) {
+ MDNode *Node = DbgNode;
+ Node->replaceAllUsesWith(D.getNode());
+ Node->destroy();
+ }
+}
+
+/// Verify - Verify that a compile unit is well formed.
+bool DICompileUnit::Verify() const {
+ if (isNull())
+ return false;
+ StringRef N = getFilename();
+ if (N.empty())
+ return false;
+ // It is possible that directory and produce string is empty.
+ return true;
+}
+
+/// Verify - Verify that a type descriptor is well formed.
+bool DIType::Verify() const {
+ if (isNull())
+ return false;
+ if (getContext().isNull())
+ return false;
+
+ DICompileUnit CU = getCompileUnit();
+ if (!CU.isNull() && !CU.Verify())
+ return false;
+ return true;
+}
+
+/// Verify - Verify that a composite type descriptor is well formed.
+bool DICompositeType::Verify() const {
+ if (isNull())
+ return false;
+ if (getContext().isNull())
+ return false;
+
+ DICompileUnit CU = getCompileUnit();
+ if (!CU.isNull() && !CU.Verify())
+ return false;
+ return true;
+}
+
+/// Verify - Verify that a subprogram descriptor is well formed.
+bool DISubprogram::Verify() const {
+ if (isNull())
+ return false;
+
+ if (getContext().isNull())
+ return false;
+
+ DICompileUnit CU = getCompileUnit();
+ if (!CU.Verify())
+ return false;
+
+ DICompositeType Ty = getType();
+ if (!Ty.isNull() && !Ty.Verify())
+ return false;
+ return true;
+}
+
+/// Verify - Verify that a global variable descriptor is well formed.
+bool DIGlobalVariable::Verify() const {
+ if (isNull())
+ return false;
+
+ if (getDisplayName().empty())
+ return false;
+
+ if (getContext().isNull())
+ return false;
+
+ DICompileUnit CU = getCompileUnit();
+ if (!CU.isNull() && !CU.Verify())
+ return false;
+
+ DIType Ty = getType();
+ if (!Ty.Verify())
+ return false;
+
+ if (!getGlobal())
+ return false;
+
+ return true;
+}
+
+/// Verify - Verify that a variable descriptor is well formed.
+bool DIVariable::Verify() const {
+ if (isNull())
+ return false;
+
+ if (getContext().isNull())
+ return false;
+
+ DIType Ty = getType();
+ if (!Ty.Verify())
+ return false;
+
+ return true;
+}
+
+/// getOriginalTypeSize - If this type is derived from a base type then
+/// return base type size.
+uint64_t DIDerivedType::getOriginalTypeSize() const {
+ unsigned Tag = getTag();
+ if (Tag == dwarf::DW_TAG_member || Tag == dwarf::DW_TAG_typedef ||
+ Tag == dwarf::DW_TAG_const_type || Tag == dwarf::DW_TAG_volatile_type ||
+ Tag == dwarf::DW_TAG_restrict_type) {
+ DIType BaseType = getTypeDerivedFrom();
+ // If this type is not derived from any type then take conservative
+ // approach.
+ if (BaseType.isNull())
+ return getSizeInBits();
+ if (BaseType.isDerivedType())
+ return DIDerivedType(BaseType.getNode()).getOriginalTypeSize();
+ else
+ return BaseType.getSizeInBits();
+ }
+
+ return getSizeInBits();
+}
+
+/// describes - Return true if this subprogram provides debugging
+/// information for the function F.
+bool DISubprogram::describes(const Function *F) {
+ assert(F && "Invalid function");
+ StringRef Name = getLinkageName();
+ if (Name.empty())
+ Name = getName();
+ if (F->getName() == Name)
+ return true;
+ return false;
+}
+
+StringRef DIScope::getFilename() const {
+ if (isLexicalBlock())
+ return DILexicalBlock(DbgNode).getFilename();
+ if (isSubprogram())
+ return DISubprogram(DbgNode).getFilename();
+ if (isCompileUnit())
+ return DICompileUnit(DbgNode).getFilename();
+ if (isNameSpace())
+ return DINameSpace(DbgNode).getFilename();
+ assert(0 && "Invalid DIScope!");
+ return StringRef();
+}
+
+StringRef DIScope::getDirectory() const {
+ if (isLexicalBlock())
+ return DILexicalBlock(DbgNode).getDirectory();
+ if (isSubprogram())
+ return DISubprogram(DbgNode).getDirectory();
+ if (isCompileUnit())
+ return DICompileUnit(DbgNode).getDirectory();
+ if (isNameSpace())
+ return DINameSpace(DbgNode).getDirectory();
+ assert(0 && "Invalid DIScope!");
+ return StringRef();
+}
+
+//===----------------------------------------------------------------------===//
+// DIDescriptor: dump routines for all descriptors.
+//===----------------------------------------------------------------------===//
+
+
+/// dump - Print descriptor.
+void DIDescriptor::dump() const {
+ dbgs() << "[" << dwarf::TagString(getTag()) << "] ";
+ dbgs().write_hex((intptr_t) &*DbgNode) << ']';
+}
+
+/// dump - Print compile unit.
+void DICompileUnit::dump() const {
+ if (getLanguage())
+ dbgs() << " [" << dwarf::LanguageString(getLanguage()) << "] ";
+
+ dbgs() << " [" << getDirectory() << "/" << getFilename() << " ]";
+}
+
+/// dump - Print type.
+void DIType::dump() const {
+ if (isNull()) return;
+
+ StringRef Res = getName();
+ if (!Res.empty())
+ dbgs() << " [" << Res << "] ";
+
+ unsigned Tag = getTag();
+ dbgs() << " [" << dwarf::TagString(Tag) << "] ";
+
+ // TODO : Print context
+ getCompileUnit().dump();
+ dbgs() << " ["
+ << getLineNumber() << ", "
+ << getSizeInBits() << ", "
+ << getAlignInBits() << ", "
+ << getOffsetInBits()
+ << "] ";
+
+ if (isPrivate())
+ dbgs() << " [private] ";
+ else if (isProtected())
+ dbgs() << " [protected] ";
+
+ if (isForwardDecl())
+ dbgs() << " [fwd] ";
+
+ if (isBasicType())
+ DIBasicType(DbgNode).dump();
+ else if (isDerivedType())
+ DIDerivedType(DbgNode).dump();
+ else if (isCompositeType())
+ DICompositeType(DbgNode).dump();
+ else {
+ dbgs() << "Invalid DIType\n";
+ return;
+ }
+
+ dbgs() << "\n";
+}
+
+/// dump - Print basic type.
+void DIBasicType::dump() const {
+ dbgs() << " [" << dwarf::AttributeEncodingString(getEncoding()) << "] ";
+}
+
+/// dump - Print derived type.
+void DIDerivedType::dump() const {
+ dbgs() << "\n\t Derived From: "; getTypeDerivedFrom().dump();
+}
+
+/// dump - Print composite type.
+void DICompositeType::dump() const {
+ DIArray A = getTypeArray();
+ if (A.isNull())
+ return;
+ dbgs() << " [" << A.getNumElements() << " elements]";
+}
+
+/// dump - Print global.
+void DIGlobal::dump() const {
+ StringRef Res = getName();
+ if (!Res.empty())
+ dbgs() << " [" << Res << "] ";
+
+ unsigned Tag = getTag();
+ dbgs() << " [" << dwarf::TagString(Tag) << "] ";
+
+ // TODO : Print context
+ getCompileUnit().dump();
+ dbgs() << " [" << getLineNumber() << "] ";
+
+ if (isLocalToUnit())
+ dbgs() << " [local] ";
+
+ if (isDefinition())
+ dbgs() << " [def] ";
+
+ if (isGlobalVariable())
+ DIGlobalVariable(DbgNode).dump();
+
+ dbgs() << "\n";
+}
+
+/// dump - Print subprogram.
+void DISubprogram::dump() const {
+ StringRef Res = getName();
+ if (!Res.empty())
+ dbgs() << " [" << Res << "] ";
+
+ unsigned Tag = getTag();
+ dbgs() << " [" << dwarf::TagString(Tag) << "] ";
+
+ // TODO : Print context
+ getCompileUnit().dump();
+ dbgs() << " [" << getLineNumber() << "] ";
+
+ if (isLocalToUnit())
+ dbgs() << " [local] ";
+
+ if (isDefinition())
+ dbgs() << " [def] ";
+
+ dbgs() << "\n";
+}
+
+/// dump - Print global variable.
+void DIGlobalVariable::dump() const {
+ dbgs() << " [";
+ getGlobal()->dump();
+ dbgs() << "] ";
+}
+
+/// dump - Print variable.
+void DIVariable::dump() const {
+ StringRef Res = getName();
+ if (!Res.empty())
+ dbgs() << " [" << Res << "] ";
+
+ getCompileUnit().dump();
+ dbgs() << " [" << getLineNumber() << "] ";
+ getType().dump();
+ dbgs() << "\n";
+
+ // FIXME: Dump complex addresses
+}
+
+//===----------------------------------------------------------------------===//
+// DIFactory: Basic Helpers
+//===----------------------------------------------------------------------===//
+
+DIFactory::DIFactory(Module &m)
+ : M(m), VMContext(M.getContext()), DeclareFn(0), ValueFn(0) {}
+
+Constant *DIFactory::GetTagConstant(unsigned TAG) {
+ assert((TAG & LLVMDebugVersionMask) == 0 &&
+ "Tag too large for debug encoding!");
+ return ConstantInt::get(Type::getInt32Ty(VMContext), TAG | LLVMDebugVersion);
+}
+
+//===----------------------------------------------------------------------===//
+// DIFactory: Primary Constructors
+//===----------------------------------------------------------------------===//
+
+/// GetOrCreateArray - Create an descriptor for an array of descriptors.
+/// This implicitly uniques the arrays created.
+DIArray DIFactory::GetOrCreateArray(DIDescriptor *Tys, unsigned NumTys) {
+ SmallVector<Value*, 16> Elts;
+
+ if (NumTys == 0)
+ Elts.push_back(llvm::Constant::getNullValue(Type::getInt32Ty(VMContext)));
+ else
+ for (unsigned i = 0; i != NumTys; ++i)
+ Elts.push_back(Tys[i].getNode());
+
+ return DIArray(MDNode::get(VMContext,Elts.data(), Elts.size()));
+}
+
+/// GetOrCreateSubrange - Create a descriptor for a value range. This
+/// implicitly uniques the values returned.
+DISubrange DIFactory::GetOrCreateSubrange(int64_t Lo, int64_t Hi) {
+ Value *Elts[] = {
+ GetTagConstant(dwarf::DW_TAG_subrange_type),
+ ConstantInt::get(Type::getInt64Ty(VMContext), Lo),
+ ConstantInt::get(Type::getInt64Ty(VMContext), Hi)
+ };
+
+ return DISubrange(MDNode::get(VMContext, &Elts[0], 3));
+}
+
+
+
+/// CreateCompileUnit - Create a new descriptor for the specified compile
+/// unit. Note that this does not unique compile units within the module.
+DICompileUnit DIFactory::CreateCompileUnit(unsigned LangID,
+ StringRef Filename,
+ StringRef Directory,
+ StringRef Producer,
+ bool isMain,
+ bool isOptimized,
+ StringRef Flags,
+ unsigned RunTimeVer) {
+ Value *Elts[] = {
+ GetTagConstant(dwarf::DW_TAG_compile_unit),
+ llvm::Constant::getNullValue(Type::getInt32Ty(VMContext)),
+ ConstantInt::get(Type::getInt32Ty(VMContext), LangID),
+ MDString::get(VMContext, Filename),
+ MDString::get(VMContext, Directory),
+ MDString::get(VMContext, Producer),
+ ConstantInt::get(Type::getInt1Ty(VMContext), isMain),
+ ConstantInt::get(Type::getInt1Ty(VMContext), isOptimized),
+ MDString::get(VMContext, Flags),
+ ConstantInt::get(Type::getInt32Ty(VMContext), RunTimeVer)
+ };
+
+ return DICompileUnit(MDNode::get(VMContext, &Elts[0], 10));
+}
+
+/// CreateEnumerator - Create a single enumerator value.
+DIEnumerator DIFactory::CreateEnumerator(StringRef Name, uint64_t Val){
+ Value *Elts[] = {
+ GetTagConstant(dwarf::DW_TAG_enumerator),
+ MDString::get(VMContext, Name),
+ ConstantInt::get(Type::getInt64Ty(VMContext), Val)
+ };
+ return DIEnumerator(MDNode::get(VMContext, &Elts[0], 3));
+}
+
+
+/// CreateBasicType - Create a basic type like int, float, etc.
+DIBasicType DIFactory::CreateBasicType(DIDescriptor Context,
+ StringRef Name,
+ DICompileUnit CompileUnit,
+ unsigned LineNumber,
+ uint64_t SizeInBits,
+ uint64_t AlignInBits,
+ uint64_t OffsetInBits, unsigned Flags,
+ unsigned Encoding) {
+ Value *Elts[] = {
+ GetTagConstant(dwarf::DW_TAG_base_type),
+ Context.getNode(),
+ MDString::get(VMContext, Name),
+ CompileUnit.getNode(),
+ ConstantInt::get(Type::getInt32Ty(VMContext), LineNumber),
+ ConstantInt::get(Type::getInt64Ty(VMContext), SizeInBits),
+ ConstantInt::get(Type::getInt64Ty(VMContext), AlignInBits),
+ ConstantInt::get(Type::getInt64Ty(VMContext), OffsetInBits),
+ ConstantInt::get(Type::getInt32Ty(VMContext), Flags),
+ ConstantInt::get(Type::getInt32Ty(VMContext), Encoding)
+ };
+ return DIBasicType(MDNode::get(VMContext, &Elts[0], 10));
+}
+
+
+/// CreateBasicType - Create a basic type like int, float, etc.
+DIBasicType DIFactory::CreateBasicTypeEx(DIDescriptor Context,
+ StringRef Name,
+ DICompileUnit CompileUnit,
+ unsigned LineNumber,
+ Constant *SizeInBits,
+ Constant *AlignInBits,
+ Constant *OffsetInBits, unsigned Flags,
+ unsigned Encoding) {
+ Value *Elts[] = {
+ GetTagConstant(dwarf::DW_TAG_base_type),
+ Context.getNode(),
+ MDString::get(VMContext, Name),
+ CompileUnit.getNode(),
+ ConstantInt::get(Type::getInt32Ty(VMContext), LineNumber),
+ SizeInBits,
+ AlignInBits,
+ OffsetInBits,
+ ConstantInt::get(Type::getInt32Ty(VMContext), Flags),
+ ConstantInt::get(Type::getInt32Ty(VMContext), Encoding)
+ };
+ return DIBasicType(MDNode::get(VMContext, &Elts[0], 10));
+}
+
+/// CreateArtificialType - Create a new DIType with "artificial" flag set.
+DIType DIFactory::CreateArtificialType(DIType Ty) {
+ if (Ty.isArtificial())
+ return Ty;
+
+ SmallVector<Value *, 9> Elts;
+ MDNode *N = Ty.getNode();
+ assert (N && "Unexpected input DIType!");
+ for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
+ if (Value *V = N->getOperand(i))
+ Elts.push_back(V);
+ else
+ Elts.push_back(Constant::getNullValue(Type::getInt32Ty(VMContext)));
+ }
+
+ unsigned CurFlags = Ty.getFlags();
+ CurFlags = CurFlags | DIType::FlagArtificial;
+
+ // Flags are stored at this slot.
+ Elts[8] = ConstantInt::get(Type::getInt32Ty(VMContext), CurFlags);
+
+ return DIType(MDNode::get(VMContext, Elts.data(), Elts.size()));
+}
+
+/// CreateDerivedType - Create a derived type like const qualified type,
+/// pointer, typedef, etc.
+DIDerivedType DIFactory::CreateDerivedType(unsigned Tag,
+ DIDescriptor Context,
+ StringRef Name,
+ DICompileUnit CompileUnit,
+ unsigned LineNumber,
+ uint64_t SizeInBits,
+ uint64_t AlignInBits,
+ uint64_t OffsetInBits,
+ unsigned Flags,
+ DIType DerivedFrom) {
+ Value *Elts[] = {
+ GetTagConstant(Tag),
+ Context.getNode(),
+ MDString::get(VMContext, Name),
+ CompileUnit.getNode(),
+ ConstantInt::get(Type::getInt32Ty(VMContext), LineNumber),
+ ConstantInt::get(Type::getInt64Ty(VMContext), SizeInBits),
+ ConstantInt::get(Type::getInt64Ty(VMContext), AlignInBits),
+ ConstantInt::get(Type::getInt64Ty(VMContext), OffsetInBits),
+ ConstantInt::get(Type::getInt32Ty(VMContext), Flags),
+ DerivedFrom.getNode(),
+ };
+ return DIDerivedType(MDNode::get(VMContext, &Elts[0], 10));
+}
+
+
+/// CreateDerivedType - Create a derived type like const qualified type,
+/// pointer, typedef, etc.
+DIDerivedType DIFactory::CreateDerivedTypeEx(unsigned Tag,
+ DIDescriptor Context,
+ StringRef Name,
+ DICompileUnit CompileUnit,
+ unsigned LineNumber,
+ Constant *SizeInBits,
+ Constant *AlignInBits,
+ Constant *OffsetInBits,
+ unsigned Flags,
+ DIType DerivedFrom) {
+ Value *Elts[] = {
+ GetTagConstant(Tag),
+ Context.getNode(),
+ MDString::get(VMContext, Name),
+ CompileUnit.getNode(),
+ ConstantInt::get(Type::getInt32Ty(VMContext), LineNumber),
+ SizeInBits,
+ AlignInBits,
+ OffsetInBits,
+ ConstantInt::get(Type::getInt32Ty(VMContext), Flags),
+ DerivedFrom.getNode(),
+ };
+ return DIDerivedType(MDNode::get(VMContext, &Elts[0], 10));
+}
+
+
+/// CreateCompositeType - Create a composite type like array, struct, etc.
+DICompositeType DIFactory::CreateCompositeType(unsigned Tag,
+ DIDescriptor Context,
+ StringRef Name,
+ DICompileUnit CompileUnit,
+ unsigned LineNumber,
+ uint64_t SizeInBits,
+ uint64_t AlignInBits,
+ uint64_t OffsetInBits,
+ unsigned Flags,
+ DIType DerivedFrom,
+ DIArray Elements,
+ unsigned RuntimeLang,
+ MDNode *ContainingType) {
+
+ Value *Elts[] = {
+ GetTagConstant(Tag),
+ Context.getNode(),
+ MDString::get(VMContext, Name),
+ CompileUnit.getNode(),
+ ConstantInt::get(Type::getInt32Ty(VMContext), LineNumber),
+ ConstantInt::get(Type::getInt64Ty(VMContext), SizeInBits),
+ ConstantInt::get(Type::getInt64Ty(VMContext), AlignInBits),
+ ConstantInt::get(Type::getInt64Ty(VMContext), OffsetInBits),
+ ConstantInt::get(Type::getInt32Ty(VMContext), Flags),
+ DerivedFrom.getNode(),
+ Elements.getNode(),
+ ConstantInt::get(Type::getInt32Ty(VMContext), RuntimeLang),
+ ContainingType
+ };
+ return DICompositeType(MDNode::get(VMContext, &Elts[0], 13));
+}
+
+
+/// CreateCompositeType - Create a composite type like array, struct, etc.
+DICompositeType DIFactory::CreateCompositeTypeEx(unsigned Tag,
+ DIDescriptor Context,
+ StringRef Name,
+ DICompileUnit CompileUnit,
+ unsigned LineNumber,
+ Constant *SizeInBits,
+ Constant *AlignInBits,
+ Constant *OffsetInBits,
+ unsigned Flags,
+ DIType DerivedFrom,
+ DIArray Elements,
+ unsigned RuntimeLang) {
+
+ Value *Elts[] = {
+ GetTagConstant(Tag),
+ Context.getNode(),
+ MDString::get(VMContext, Name),
+ CompileUnit.getNode(),
+ ConstantInt::get(Type::getInt32Ty(VMContext), LineNumber),
+ SizeInBits,
+ AlignInBits,
+ OffsetInBits,
+ ConstantInt::get(Type::getInt32Ty(VMContext), Flags),
+ DerivedFrom.getNode(),
+ Elements.getNode(),
+ ConstantInt::get(Type::getInt32Ty(VMContext), RuntimeLang)
+ };
+ return DICompositeType(MDNode::get(VMContext, &Elts[0], 12));
+}
+
+
+/// CreateSubprogram - Create a new descriptor for the specified subprogram.
+/// See comments in DISubprogram for descriptions of these fields. This
+/// method does not unique the generated descriptors.
+DISubprogram DIFactory::CreateSubprogram(DIDescriptor Context,
+ StringRef Name,
+ StringRef DisplayName,
+ StringRef LinkageName,
+ DICompileUnit CompileUnit,
+ unsigned LineNo, DIType Ty,
+ bool isLocalToUnit,
+ bool isDefinition,
+ unsigned VK, unsigned VIndex,
+ DIType ContainingType,
+ bool isArtificial) {
+
+ Value *Elts[] = {
+ GetTagConstant(dwarf::DW_TAG_subprogram),
+ llvm::Constant::getNullValue(Type::getInt32Ty(VMContext)),
+ Context.getNode(),
+ MDString::get(VMContext, Name),
+ MDString::get(VMContext, DisplayName),
+ MDString::get(VMContext, LinkageName),
+ CompileUnit.getNode(),
+ ConstantInt::get(Type::getInt32Ty(VMContext), LineNo),
+ Ty.getNode(),
+ ConstantInt::get(Type::getInt1Ty(VMContext), isLocalToUnit),
+ ConstantInt::get(Type::getInt1Ty(VMContext), isDefinition),
+ ConstantInt::get(Type::getInt32Ty(VMContext), (unsigned)VK),
+ ConstantInt::get(Type::getInt32Ty(VMContext), VIndex),
+ ContainingType.getNode(),
+ ConstantInt::get(Type::getInt1Ty(VMContext), isArtificial)
+ };
+ return DISubprogram(MDNode::get(VMContext, &Elts[0], 15));
+}
+
+/// CreateSubprogramDefinition - Create new subprogram descriptor for the
+/// given declaration.
+DISubprogram DIFactory::CreateSubprogramDefinition(DISubprogram &SPDeclaration) {
+ if (SPDeclaration.isDefinition())
+ return DISubprogram(SPDeclaration.getNode());
+
+ MDNode *DeclNode = SPDeclaration.getNode();
+ Value *Elts[] = {
+ GetTagConstant(dwarf::DW_TAG_subprogram),
+ llvm::Constant::getNullValue(Type::getInt32Ty(VMContext)),
+ DeclNode->getOperand(2), // Context
+ DeclNode->getOperand(3), // Name
+ DeclNode->getOperand(4), // DisplayName
+ DeclNode->getOperand(5), // LinkageName
+ DeclNode->getOperand(6), // CompileUnit
+ DeclNode->getOperand(7), // LineNo
+ DeclNode->getOperand(8), // Type
+ DeclNode->getOperand(9), // isLocalToUnit
+ ConstantInt::get(Type::getInt1Ty(VMContext), true),
+ DeclNode->getOperand(11), // Virtuality
+ DeclNode->getOperand(12), // VIndex
+ DeclNode->getOperand(13), // Containting Type
+ DeclNode->getOperand(14) // isArtificial
+ };
+ return DISubprogram(MDNode::get(VMContext, &Elts[0], 15));
+}
+
+/// CreateGlobalVariable - Create a new descriptor for the specified global.
+DIGlobalVariable
+DIFactory::CreateGlobalVariable(DIDescriptor Context, StringRef Name,
+ StringRef DisplayName,
+ StringRef LinkageName,
+ DICompileUnit CompileUnit,
+ unsigned LineNo, DIType Ty,bool isLocalToUnit,
+ bool isDefinition, llvm::GlobalVariable *Val) {
+ Value *Elts[] = {
+ GetTagConstant(dwarf::DW_TAG_variable),
+ llvm::Constant::getNullValue(Type::getInt32Ty(VMContext)),
+ Context.getNode(),
+ MDString::get(VMContext, Name),
+ MDString::get(VMContext, DisplayName),
+ MDString::get(VMContext, LinkageName),
+ CompileUnit.getNode(),
+ ConstantInt::get(Type::getInt32Ty(VMContext), LineNo),
+ Ty.getNode(),
+ ConstantInt::get(Type::getInt1Ty(VMContext), isLocalToUnit),
+ ConstantInt::get(Type::getInt1Ty(VMContext), isDefinition),
+ Val
+ };
+
+ Value *const *Vs = &Elts[0];
+ MDNode *Node = MDNode::get(VMContext,Vs, 12);
+
+ // Create a named metadata so that we do not lose this mdnode.
+ NamedMDNode *NMD = M.getOrInsertNamedMetadata("llvm.dbg.gv");
+ NMD->addOperand(Node);
+
+ return DIGlobalVariable(Node);
+}
+
+
+/// CreateVariable - Create a new descriptor for the specified variable.
+DIVariable DIFactory::CreateVariable(unsigned Tag, DIDescriptor Context,
+ StringRef Name,
+ DICompileUnit CompileUnit, unsigned LineNo,
+ DIType Ty) {
+ Value *Elts[] = {
+ GetTagConstant(Tag),
+ Context.getNode(),
+ MDString::get(VMContext, Name),
+ CompileUnit.getNode(),
+ ConstantInt::get(Type::getInt32Ty(VMContext), LineNo),
+ Ty.getNode(),
+ };
+ return DIVariable(MDNode::get(VMContext, &Elts[0], 6));
+}
+
+
+/// CreateComplexVariable - Create a new descriptor for the specified variable
+/// which has a complex address expression for its address.
+DIVariable DIFactory::CreateComplexVariable(unsigned Tag, DIDescriptor Context,
+ const std::string &Name,
+ DICompileUnit CompileUnit,
+ unsigned LineNo,
+ DIType Ty,
+ SmallVector<Value *, 9> &addr) {
+ SmallVector<Value *, 9> Elts;
+ Elts.push_back(GetTagConstant(Tag));
+ Elts.push_back(Context.getNode());
+ Elts.push_back(MDString::get(VMContext, Name));
+ Elts.push_back(CompileUnit.getNode());
+ Elts.push_back(ConstantInt::get(Type::getInt32Ty(VMContext), LineNo));
+ Elts.push_back(Ty.getNode());
+ Elts.insert(Elts.end(), addr.begin(), addr.end());
+
+ return DIVariable(MDNode::get(VMContext, &Elts[0], 6+addr.size()));
+}
+
+
+/// CreateBlock - This creates a descriptor for a lexical block with the
+/// specified parent VMContext.
+DILexicalBlock DIFactory::CreateLexicalBlock(DIDescriptor Context) {
+ Value *Elts[] = {
+ GetTagConstant(dwarf::DW_TAG_lexical_block),
+ Context.getNode()
+ };
+ return DILexicalBlock(MDNode::get(VMContext, &Elts[0], 2));
+}
+
+/// CreateNameSpace - This creates new descriptor for a namespace
+/// with the specified parent context.
+DINameSpace DIFactory::CreateNameSpace(DIDescriptor Context, StringRef Name,
+ DICompileUnit CompileUnit,
+ unsigned LineNo) {
+ Value *Elts[] = {
+ GetTagConstant(dwarf::DW_TAG_namespace),
+ Context.getNode(),
+ MDString::get(VMContext, Name),
+ CompileUnit.getNode(),
+ ConstantInt::get(Type::getInt32Ty(VMContext), LineNo)
+ };
+ return DINameSpace(MDNode::get(VMContext, &Elts[0], 5));
+}
+
+/// CreateLocation - Creates a debug info location.
+DILocation DIFactory::CreateLocation(unsigned LineNo, unsigned ColumnNo,
+ DIScope S, DILocation OrigLoc) {
+ Value *Elts[] = {
+ ConstantInt::get(Type::getInt32Ty(VMContext), LineNo),
+ ConstantInt::get(Type::getInt32Ty(VMContext), ColumnNo),
+ S.getNode(),
+ OrigLoc.getNode(),
+ };
+ return DILocation(MDNode::get(VMContext, &Elts[0], 4));
+}
+
+/// CreateLocation - Creates a debug info location.
+DILocation DIFactory::CreateLocation(unsigned LineNo, unsigned ColumnNo,
+ DIScope S, MDNode *OrigLoc) {
+ Value *Elts[] = {
+ ConstantInt::get(Type::getInt32Ty(VMContext), LineNo),
+ ConstantInt::get(Type::getInt32Ty(VMContext), ColumnNo),
+ S.getNode(),
+ OrigLoc
+ };
+ return DILocation(MDNode::get(VMContext, &Elts[0], 4));
+}
+
+//===----------------------------------------------------------------------===//
+// DIFactory: Routines for inserting code into a function
+//===----------------------------------------------------------------------===//
+
+/// InsertDeclare - Insert a new llvm.dbg.declare intrinsic call.
+Instruction *DIFactory::InsertDeclare(Value *Storage, DIVariable D,
+ Instruction *InsertBefore) {
+ assert(Storage && "no storage passed to dbg.declare");
+ assert(D.getNode() && "empty DIVariable passed to dbg.declare");
+ if (!DeclareFn)
+ DeclareFn = Intrinsic::getDeclaration(&M, Intrinsic::dbg_declare);
+
+ Value *Args[] = { MDNode::get(Storage->getContext(), &Storage, 1),
+ D.getNode() };
+ return CallInst::Create(DeclareFn, Args, Args+2, "", InsertBefore);
+}
+
+/// InsertDeclare - Insert a new llvm.dbg.declare intrinsic call.
+Instruction *DIFactory::InsertDeclare(Value *Storage, DIVariable D,
+ BasicBlock *InsertAtEnd) {
+ assert(Storage && "no storage passed to dbg.declare");
+ assert(D.getNode() && "empty DIVariable passed to dbg.declare");
+ if (!DeclareFn)
+ DeclareFn = Intrinsic::getDeclaration(&M, Intrinsic::dbg_declare);
+
+ Value *Args[] = { MDNode::get(Storage->getContext(), &Storage, 1),
+ D.getNode() };
+
+ // If this block already has a terminator then insert this intrinsic
+ // before the terminator.
+ if (TerminatorInst *T = InsertAtEnd->getTerminator())
+ return CallInst::Create(DeclareFn, Args, Args+2, "", T);
+ else
+ return CallInst::Create(DeclareFn, Args, Args+2, "", InsertAtEnd);}
+
+/// InsertDbgValueIntrinsic - Insert a new llvm.dbg.value intrinsic call.
+Instruction *DIFactory::InsertDbgValueIntrinsic(Value *V, uint64_t Offset,
+ DIVariable D,
+ Instruction *InsertBefore) {
+ assert(V && "no value passed to dbg.value");
+ assert(D.getNode() && "empty DIVariable passed to dbg.value");
+ if (!ValueFn)
+ ValueFn = Intrinsic::getDeclaration(&M, Intrinsic::dbg_value);
+
+ Value *Args[] = { MDNode::get(V->getContext(), &V, 1),
+ ConstantInt::get(Type::getInt64Ty(V->getContext()), Offset),
+ D.getNode() };
+ return CallInst::Create(ValueFn, Args, Args+3, "", InsertBefore);
+}
+
+/// InsertDbgValueIntrinsic - Insert a new llvm.dbg.value intrinsic call.
+Instruction *DIFactory::InsertDbgValueIntrinsic(Value *V, uint64_t Offset,
+ DIVariable D,
+ BasicBlock *InsertAtEnd) {
+ assert(V && "no value passed to dbg.value");
+ assert(D.getNode() && "empty DIVariable passed to dbg.value");
+ if (!ValueFn)
+ ValueFn = Intrinsic::getDeclaration(&M, Intrinsic::dbg_value);
+
+ Value *Args[] = { MDNode::get(V->getContext(), &V, 1),
+ ConstantInt::get(Type::getInt64Ty(V->getContext()), Offset),
+ D.getNode() };
+ return CallInst::Create(ValueFn, Args, Args+3, "", InsertAtEnd);
+}
+
+//===----------------------------------------------------------------------===//
+// DebugInfoFinder implementations.
+//===----------------------------------------------------------------------===//
+
+/// processModule - Process entire module and collect debug info.
+void DebugInfoFinder::processModule(Module &M) {
+ unsigned MDDbgKind = M.getMDKindID("dbg");
+
+ for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
+ for (Function::iterator FI = (*I).begin(), FE = (*I).end(); FI != FE; ++FI)
+ for (BasicBlock::iterator BI = (*FI).begin(), BE = (*FI).end(); BI != BE;
+ ++BI) {
+ if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(BI))
+ processDeclare(DDI);
+ else if (MDNode *L = BI->getMetadata(MDDbgKind))
+ processLocation(DILocation(L));
+ }
+
+ NamedMDNode *NMD = M.getNamedMetadata("llvm.dbg.gv");
+ if (!NMD)
+ return;
+
+ for (unsigned i = 0, e = NMD->getNumOperands(); i != e; ++i) {
+ DIGlobalVariable DIG(cast<MDNode>(NMD->getOperand(i)));
+ if (addGlobalVariable(DIG)) {
+ addCompileUnit(DIG.getCompileUnit());
+ processType(DIG.getType());
+ }
+ }
+}
+
+/// processLocation - Process DILocation.
+void DebugInfoFinder::processLocation(DILocation Loc) {
+ if (Loc.isNull()) return;
+ DIScope S(Loc.getScope().getNode());
+ if (S.isNull()) return;
+ if (S.isCompileUnit())
+ addCompileUnit(DICompileUnit(S.getNode()));
+ else if (S.isSubprogram())
+ processSubprogram(DISubprogram(S.getNode()));
+ else if (S.isLexicalBlock())
+ processLexicalBlock(DILexicalBlock(S.getNode()));
+ processLocation(Loc.getOrigLocation());
+}
+
+/// processType - Process DIType.
+void DebugInfoFinder::processType(DIType DT) {
+ if (!addType(DT))
+ return;
+
+ addCompileUnit(DT.getCompileUnit());
+ if (DT.isCompositeType()) {
+ DICompositeType DCT(DT.getNode());
+ processType(DCT.getTypeDerivedFrom());
+ DIArray DA = DCT.getTypeArray();
+ if (!DA.isNull())
+ for (unsigned i = 0, e = DA.getNumElements(); i != e; ++i) {
+ DIDescriptor D = DA.getElement(i);
+ DIType TyE = DIType(D.getNode());
+ if (!TyE.isNull())
+ processType(TyE);
+ else
+ processSubprogram(DISubprogram(D.getNode()));
+ }
+ } else if (DT.isDerivedType()) {
+ DIDerivedType DDT(DT.getNode());
+ if (!DDT.isNull())
+ processType(DDT.getTypeDerivedFrom());
+ }
+}
+
+/// processLexicalBlock
+void DebugInfoFinder::processLexicalBlock(DILexicalBlock LB) {
+ if (LB.isNull())
+ return;
+ DIScope Context = LB.getContext();
+ if (Context.isLexicalBlock())
+ return processLexicalBlock(DILexicalBlock(Context.getNode()));
+ else
+ return processSubprogram(DISubprogram(Context.getNode()));
+}
+
+/// processSubprogram - Process DISubprogram.
+void DebugInfoFinder::processSubprogram(DISubprogram SP) {
+ if (SP.isNull())
+ return;
+ if (!addSubprogram(SP))
+ return;
+ addCompileUnit(SP.getCompileUnit());
+ processType(SP.getType());
+}
+
+/// processDeclare - Process DbgDeclareInst.
+void DebugInfoFinder::processDeclare(DbgDeclareInst *DDI) {
+ DIVariable DV(cast<MDNode>(DDI->getVariable()));
+ if (DV.isNull())
+ return;
+
+ if (!NodesSeen.insert(DV.getNode()))
+ return;
+
+ addCompileUnit(DV.getCompileUnit());
+ processType(DV.getType());
+}
+
+/// addType - Add type into Tys.
+bool DebugInfoFinder::addType(DIType DT) {
+ if (DT.isNull())
+ return false;
+
+ if (!NodesSeen.insert(DT.getNode()))
+ return false;
+
+ TYs.push_back(DT.getNode());
+ return true;
+}
+
+/// addCompileUnit - Add compile unit into CUs.
+bool DebugInfoFinder::addCompileUnit(DICompileUnit CU) {
+ if (CU.isNull())
+ return false;
+
+ if (!NodesSeen.insert(CU.getNode()))
+ return false;
+
+ CUs.push_back(CU.getNode());
+ return true;
+}
+
+/// addGlobalVariable - Add global variable into GVs.
+bool DebugInfoFinder::addGlobalVariable(DIGlobalVariable DIG) {
+ if (DIG.isNull())
+ return false;
+
+ if (!NodesSeen.insert(DIG.getNode()))
+ return false;
+
+ GVs.push_back(DIG.getNode());
+ return true;
+}
+
+// addSubprogram - Add subprgoram into SPs.
+bool DebugInfoFinder::addSubprogram(DISubprogram SP) {
+ if (SP.isNull())
+ return false;
+
+ if (!NodesSeen.insert(SP.getNode()))
+ return false;
+
+ SPs.push_back(SP.getNode());
+ return true;
+}
+
+/// Find the debug info descriptor corresponding to this global variable.
+static Value *findDbgGlobalDeclare(GlobalVariable *V) {
+ const Module *M = V->getParent();
+ NamedMDNode *NMD = M->getNamedMetadata("llvm.dbg.gv");
+ if (!NMD)
+ return 0;
+
+ for (unsigned i = 0, e = NMD->getNumOperands(); i != e; ++i) {
+ DIGlobalVariable DIG(cast_or_null<MDNode>(NMD->getOperand(i)));
+ if (DIG.isNull())
+ continue;
+ if (DIG.getGlobal() == V)
+ return DIG.getNode();
+ }
+ return 0;
+}
+
+/// Finds the llvm.dbg.declare intrinsic corresponding to this value if any.
+/// It looks through pointer casts too.
+static const DbgDeclareInst *findDbgDeclare(const Value *V) {
+ V = V->stripPointerCasts();
+
+ if (!isa<Instruction>(V) && !isa<Argument>(V))
+ return 0;
+
+ const Function *F = NULL;
+ if (const Instruction *I = dyn_cast<Instruction>(V))
+ F = I->getParent()->getParent();
+ else if (const Argument *A = dyn_cast<Argument>(V))
+ F = A->getParent();
+
+ for (Function::const_iterator FI = F->begin(), FE = F->end(); FI != FE; ++FI)
+ for (BasicBlock::const_iterator BI = (*FI).begin(), BE = (*FI).end();
+ BI != BE; ++BI)
+ if (const DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(BI))
+ if (DDI->getAddress() == V)
+ return DDI;
+
+ return 0;
+}
+
+bool llvm::getLocationInfo(const Value *V, std::string &DisplayName,
+ std::string &Type, unsigned &LineNo,
+ std::string &File, std::string &Dir) {
+ DICompileUnit Unit;
+ DIType TypeD;
+
+ if (GlobalVariable *GV = dyn_cast<GlobalVariable>(const_cast<Value*>(V))) {
+ Value *DIGV = findDbgGlobalDeclare(GV);
+ if (!DIGV) return false;
+ DIGlobalVariable Var(cast<MDNode>(DIGV));
+
+ StringRef D = Var.getDisplayName();
+ if (!D.empty())
+ DisplayName = D;
+ LineNo = Var.getLineNumber();
+ Unit = Var.getCompileUnit();
+ TypeD = Var.getType();
+ } else {
+ const DbgDeclareInst *DDI = findDbgDeclare(V);
+ if (!DDI) return false;
+ DIVariable Var(cast<MDNode>(DDI->getVariable()));
+
+ StringRef D = Var.getName();
+ if (!D.empty())
+ DisplayName = D;
+ LineNo = Var.getLineNumber();
+ Unit = Var.getCompileUnit();
+ TypeD = Var.getType();
+ }
+
+ StringRef T = TypeD.getName();
+ if (!T.empty())
+ Type = T;
+ StringRef F = Unit.getFilename();
+ if (!F.empty())
+ File = F;
+ StringRef D = Unit.getDirectory();
+ if (!D.empty())
+ Dir = D;
+ return true;
+}
+
+/// ExtractDebugLocation - Extract debug location information
+/// from DILocation.
+DebugLoc llvm::ExtractDebugLocation(DILocation &Loc,
+ DebugLocTracker &DebugLocInfo) {
+ DenseMap<MDNode *, unsigned>::iterator II
+ = DebugLocInfo.DebugIdMap.find(Loc.getNode());
+ if (II != DebugLocInfo.DebugIdMap.end())
+ return DebugLoc::get(II->second);
+
+ // Add a new location entry.
+ unsigned Id = DebugLocInfo.DebugLocations.size();
+ DebugLocInfo.DebugLocations.push_back(Loc.getNode());
+ DebugLocInfo.DebugIdMap[Loc.getNode()] = Id;
+
+ return DebugLoc::get(Id);
+}
+
+/// getDISubprogram - Find subprogram that is enclosing this scope.
+DISubprogram llvm::getDISubprogram(MDNode *Scope) {
+ DIDescriptor D(Scope);
+ if (D.isNull())
+ return DISubprogram();
+
+ if (D.isCompileUnit())
+ return DISubprogram();
+
+ if (D.isSubprogram())
+ return DISubprogram(Scope);
+
+ if (D.isLexicalBlock())
+ return getDISubprogram(DILexicalBlock(Scope).getContext().getNode());
+
+ return DISubprogram();
+}
+
+/// getDICompositeType - Find underlying composite type.
+DICompositeType llvm::getDICompositeType(DIType T) {
+ if (T.isNull())
+ return DICompositeType();
+
+ if (T.isCompositeType())
+ return DICompositeType(T.getNode());
+
+ if (T.isDerivedType())
+ return getDICompositeType(DIDerivedType(T.getNode()).getTypeDerivedFrom());
+
+ return DICompositeType();
+}
diff --git a/lib/Analysis/DomPrinter.cpp b/lib/Analysis/DomPrinter.cpp
new file mode 100644
index 0000000..3af687a
--- /dev/null
+++ b/lib/Analysis/DomPrinter.cpp
@@ -0,0 +1,242 @@
+//===- DomPrinter.cpp - DOT printer for the dominance trees ------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file defines '-dot-dom' and '-dot-postdom' analysis passes, which emit
+// a dom.<fnname>.dot or postdom.<fnname>.dot file for each function in the
+// program, with a graph of the dominance/postdominance tree of that
+// function.
+//
+// There are also passes available to directly call dotty ('-view-dom' or
+// '-view-postdom'). By appending '-only' like '-dot-dom-only' only the
+// names of the bbs are printed, but the content is hidden.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/DomPrinter.h"
+
+#include "llvm/Analysis/Dominators.h"
+#include "llvm/Analysis/DOTGraphTraitsPass.h"
+#include "llvm/Analysis/PostDominators.h"
+
+using namespace llvm;
+
+namespace llvm {
+template<>
+struct DOTGraphTraits<DomTreeNode*> : public DefaultDOTGraphTraits {
+
+ DOTGraphTraits (bool isSimple=false)
+ : DefaultDOTGraphTraits(isSimple) {}
+
+ std::string getNodeLabel(DomTreeNode *Node, DomTreeNode *Graph) {
+
+ BasicBlock *BB = Node->getBlock();
+
+ if (!BB)
+ return "Post dominance root node";
+
+
+ if (isSimple())
+ return DOTGraphTraits<const Function*>
+ ::getSimpleNodeLabel(BB, BB->getParent());
+ else
+ return DOTGraphTraits<const Function*>
+ ::getCompleteNodeLabel(BB, BB->getParent());
+ }
+};
+
+template<>
+struct DOTGraphTraits<DominatorTree*> : public DOTGraphTraits<DomTreeNode*> {
+
+ DOTGraphTraits (bool isSimple=false)
+ : DOTGraphTraits<DomTreeNode*>(isSimple) {}
+
+ static std::string getGraphName(DominatorTree *DT) {
+ return "Dominator tree";
+ }
+
+ std::string getNodeLabel(DomTreeNode *Node, DominatorTree *G) {
+ return DOTGraphTraits<DomTreeNode*>::getNodeLabel(Node, G->getRootNode());
+ }
+};
+
+template<>
+struct DOTGraphTraits<PostDominatorTree*>
+ : public DOTGraphTraits<DomTreeNode*> {
+
+ DOTGraphTraits (bool isSimple=false)
+ : DOTGraphTraits<DomTreeNode*>(isSimple) {}
+
+ static std::string getGraphName(PostDominatorTree *DT) {
+ return "Post dominator tree";
+ }
+
+ std::string getNodeLabel(DomTreeNode *Node, PostDominatorTree *G ) {
+ return DOTGraphTraits<DomTreeNode*>::getNodeLabel(Node, G->getRootNode());
+ }
+};
+}
+
+namespace {
+template <class Analysis, bool OnlyBBS>
+struct GenericGraphViewer : public FunctionPass {
+ std::string Name;
+
+ GenericGraphViewer(std::string GraphName, const void *ID) : FunctionPass(ID) {
+ Name = GraphName;
+ }
+
+ virtual bool runOnFunction(Function &F) {
+ Analysis *Graph;
+ std::string Title, GraphName;
+ Graph = &getAnalysis<Analysis>();
+ GraphName = DOTGraphTraits<Analysis*>::getGraphName(Graph);
+ Title = GraphName + " for '" + F.getNameStr() + "' function";
+ ViewGraph(Graph, Name, OnlyBBS, Title);
+
+ return false;
+ }
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.setPreservesAll();
+ AU.addRequired<Analysis>();
+ }
+};
+
+struct DomViewer
+ : public DOTGraphTraitsViewer<DominatorTree, false> {
+ static char ID;
+ DomViewer() : DOTGraphTraitsViewer<DominatorTree, false>("dom", &ID){}
+};
+
+struct DomOnlyViewer
+ : public DOTGraphTraitsViewer<DominatorTree, true> {
+ static char ID;
+ DomOnlyViewer() : DOTGraphTraitsViewer<DominatorTree, true>("domonly", &ID){}
+};
+
+struct PostDomViewer
+ : public DOTGraphTraitsViewer<PostDominatorTree, false> {
+ static char ID;
+ PostDomViewer() :
+ DOTGraphTraitsViewer<PostDominatorTree, false>("postdom", &ID){}
+};
+
+struct PostDomOnlyViewer
+ : public DOTGraphTraitsViewer<PostDominatorTree, true> {
+ static char ID;
+ PostDomOnlyViewer() :
+ DOTGraphTraitsViewer<PostDominatorTree, true>("postdomonly", &ID){}
+};
+} // end anonymous namespace
+
+char DomViewer::ID = 0;
+RegisterPass<DomViewer> A("view-dom",
+ "View dominance tree of function");
+
+char DomOnlyViewer::ID = 0;
+RegisterPass<DomOnlyViewer> B("view-dom-only",
+ "View dominance tree of function "
+ "(with no function bodies)");
+
+char PostDomViewer::ID = 0;
+RegisterPass<PostDomViewer> C("view-postdom",
+ "View postdominance tree of function");
+
+char PostDomOnlyViewer::ID = 0;
+RegisterPass<PostDomOnlyViewer> D("view-postdom-only",
+ "View postdominance tree of function "
+ "(with no function bodies)");
+
+namespace {
+struct DomPrinter
+ : public DOTGraphTraitsPrinter<DominatorTree, false> {
+ static char ID;
+ DomPrinter() : DOTGraphTraitsPrinter<DominatorTree, false>("dom", &ID) {}
+};
+
+struct DomOnlyPrinter
+ : public DOTGraphTraitsPrinter<DominatorTree, true> {
+ static char ID;
+ DomOnlyPrinter() : DOTGraphTraitsPrinter<DominatorTree, true>("domonly", &ID) {}
+};
+
+struct PostDomPrinter
+ : public DOTGraphTraitsPrinter<PostDominatorTree, false> {
+ static char ID;
+ PostDomPrinter() :
+ DOTGraphTraitsPrinter<PostDominatorTree, false>("postdom", &ID) {}
+};
+
+struct PostDomOnlyPrinter
+ : public DOTGraphTraitsPrinter<PostDominatorTree, true> {
+ static char ID;
+ PostDomOnlyPrinter() :
+ DOTGraphTraitsPrinter<PostDominatorTree, true>("postdomonly", &ID) {}
+};
+} // end anonymous namespace
+
+
+
+char DomPrinter::ID = 0;
+RegisterPass<DomPrinter> E("dot-dom",
+ "Print dominance tree of function "
+ "to 'dot' file");
+
+char DomOnlyPrinter::ID = 0;
+RegisterPass<DomOnlyPrinter> F("dot-dom-only",
+ "Print dominance tree of function "
+ "to 'dot' file "
+ "(with no function bodies)");
+
+char PostDomPrinter::ID = 0;
+RegisterPass<PostDomPrinter> G("dot-postdom",
+ "Print postdominance tree of function "
+ "to 'dot' file");
+
+char PostDomOnlyPrinter::ID = 0;
+RegisterPass<PostDomOnlyPrinter> H("dot-postdom-only",
+ "Print postdominance tree of function "
+ "to 'dot' file "
+ "(with no function bodies)");
+
+// Create methods available outside of this file, to use them
+// "include/llvm/LinkAllPasses.h". Otherwise the pass would be deleted by
+// the link time optimization.
+
+FunctionPass *llvm::createDomPrinterPass() {
+ return new DomPrinter();
+}
+
+FunctionPass *llvm::createDomOnlyPrinterPass() {
+ return new DomOnlyPrinter();
+}
+
+FunctionPass *llvm::createDomViewerPass() {
+ return new DomViewer();
+}
+
+FunctionPass *llvm::createDomOnlyViewerPass() {
+ return new DomOnlyViewer();
+}
+
+FunctionPass *llvm::createPostDomPrinterPass() {
+ return new PostDomPrinter();
+}
+
+FunctionPass *llvm::createPostDomOnlyPrinterPass() {
+ return new PostDomOnlyPrinter();
+}
+
+FunctionPass *llvm::createPostDomViewerPass() {
+ return new PostDomViewer();
+}
+
+FunctionPass *llvm::createPostDomOnlyViewerPass() {
+ return new PostDomOnlyViewer();
+}
diff --git a/lib/Analysis/IPA/Andersens.cpp b/lib/Analysis/IPA/Andersens.cpp
new file mode 100644
index 0000000..4180206
--- /dev/null
+++ b/lib/Analysis/IPA/Andersens.cpp
@@ -0,0 +1,2868 @@
+//===- Andersens.cpp - Andersen's Interprocedural Alias Analysis ----------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file defines an implementation of Andersen's interprocedural alias
+// analysis
+//
+// In pointer analysis terms, this is a subset-based, flow-insensitive,
+// field-sensitive, and context-insensitive algorithm pointer algorithm.
+//
+// This algorithm is implemented as three stages:
+// 1. Object identification.
+// 2. Inclusion constraint identification.
+// 3. Offline constraint graph optimization
+// 4. Inclusion constraint solving.
+//
+// The object identification stage identifies all of the memory objects in the
+// program, which includes globals, heap allocated objects, and stack allocated
+// objects.
+//
+// The inclusion constraint identification stage finds all inclusion constraints
+// in the program by scanning the program, looking for pointer assignments and
+// other statements that effect the points-to graph. For a statement like "A =
+// B", this statement is processed to indicate that A can point to anything that
+// B can point to. Constraints can handle copies, loads, and stores, and
+// address taking.
+//
+// The offline constraint graph optimization portion includes offline variable
+// substitution algorithms intended to compute pointer and location
+// equivalences. Pointer equivalences are those pointers that will have the
+// same points-to sets, and location equivalences are those variables that
+// always appear together in points-to sets. It also includes an offline
+// cycle detection algorithm that allows cycles to be collapsed sooner
+// during solving.
+//
+// The inclusion constraint solving phase iteratively propagates the inclusion
+// constraints until a fixed point is reached. This is an O(N^3) algorithm.
+//
+// Function constraints are handled as if they were structs with X fields.
+// Thus, an access to argument X of function Y is an access to node index
+// getNode(Y) + X. This representation allows handling of indirect calls
+// without any issues. To wit, an indirect call Y(a,b) is equivalent to
+// *(Y + 1) = a, *(Y + 2) = b.
+// The return node for a function is always located at getNode(F) +
+// CallReturnPos. The arguments start at getNode(F) + CallArgPos.
+//
+// Future Improvements:
+// Use of BDD's.
+//===----------------------------------------------------------------------===//
+
+#define DEBUG_TYPE "anders-aa"
+#include "llvm/Constants.h"
+#include "llvm/DerivedTypes.h"
+#include "llvm/Instructions.h"
+#include "llvm/Module.h"
+#include "llvm/Pass.h"
+#include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/InstIterator.h"
+#include "llvm/Support/InstVisitor.h"
+#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/Analysis/MemoryBuiltins.h"
+#include "llvm/Analysis/Passes.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/System/Atomic.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/ADT/SparseBitVector.h"
+#include "llvm/ADT/DenseSet.h"
+#include <algorithm>
+#include <set>
+#include <list>
+#include <map>
+#include <stack>
+#include <vector>
+#include <queue>
+
+// Determining the actual set of nodes the universal set can consist of is very
+// expensive because it means propagating around very large sets. We rely on
+// other analysis being able to determine which nodes can never be pointed to in
+// order to disambiguate further than "points-to anything".
+#define FULL_UNIVERSAL 0
+
+using namespace llvm;
+#ifndef NDEBUG
+STATISTIC(NumIters , "Number of iterations to reach convergence");
+#endif
+STATISTIC(NumConstraints, "Number of constraints");
+STATISTIC(NumNodes , "Number of nodes");
+STATISTIC(NumUnified , "Number of variables unified");
+STATISTIC(NumErased , "Number of redundant constraints erased");
+
+static const unsigned SelfRep = (unsigned)-1;
+static const unsigned Unvisited = (unsigned)-1;
+// Position of the function return node relative to the function node.
+static const unsigned CallReturnPos = 1;
+// Position of the function call node relative to the function node.
+static const unsigned CallFirstArgPos = 2;
+
+namespace {
+ struct BitmapKeyInfo {
+ static inline SparseBitVector<> *getEmptyKey() {
+ return reinterpret_cast<SparseBitVector<> *>(-1);
+ }
+ static inline SparseBitVector<> *getTombstoneKey() {
+ return reinterpret_cast<SparseBitVector<> *>(-2);
+ }
+ static unsigned getHashValue(const SparseBitVector<> *bitmap) {
+ return bitmap->getHashValue();
+ }
+ static bool isEqual(const SparseBitVector<> *LHS,
+ const SparseBitVector<> *RHS) {
+ if (LHS == RHS)
+ return true;
+ else if (LHS == getEmptyKey() || RHS == getEmptyKey()
+ || LHS == getTombstoneKey() || RHS == getTombstoneKey())
+ return false;
+
+ return *LHS == *RHS;
+ }
+ };
+
+ class Andersens : public ModulePass, public AliasAnalysis,
+ private InstVisitor<Andersens> {
+ struct Node;
+
+ /// Constraint - Objects of this structure are used to represent the various
+ /// constraints identified by the algorithm. The constraints are 'copy',
+ /// for statements like "A = B", 'load' for statements like "A = *B",
+ /// 'store' for statements like "*A = B", and AddressOf for statements like
+ /// A = alloca; The Offset is applied as *(A + K) = B for stores,
+ /// A = *(B + K) for loads, and A = B + K for copies. It is
+ /// illegal on addressof constraints (because it is statically
+ /// resolvable to A = &C where C = B + K)
+
+ struct Constraint {
+ enum ConstraintType { Copy, Load, Store, AddressOf } Type;
+ unsigned Dest;
+ unsigned Src;
+ unsigned Offset;
+
+ Constraint(ConstraintType Ty, unsigned D, unsigned S, unsigned O = 0)
+ : Type(Ty), Dest(D), Src(S), Offset(O) {
+ assert((Offset == 0 || Ty != AddressOf) &&
+ "Offset is illegal on addressof constraints");
+ }
+
+ bool operator==(const Constraint &RHS) const {
+ return RHS.Type == Type
+ && RHS.Dest == Dest
+ && RHS.Src == Src
+ && RHS.Offset == Offset;
+ }
+
+ bool operator!=(const Constraint &RHS) const {
+ return !(*this == RHS);
+ }
+
+ bool operator<(const Constraint &RHS) const {
+ if (RHS.Type != Type)
+ return RHS.Type < Type;
+ else if (RHS.Dest != Dest)
+ return RHS.Dest < Dest;
+ else if (RHS.Src != Src)
+ return RHS.Src < Src;
+ return RHS.Offset < Offset;
+ }
+ };
+
+ // Information DenseSet requires implemented in order to be able to do
+ // it's thing
+ struct PairKeyInfo {
+ static inline std::pair<unsigned, unsigned> getEmptyKey() {
+ return std::make_pair(~0U, ~0U);
+ }
+ static inline std::pair<unsigned, unsigned> getTombstoneKey() {
+ return std::make_pair(~0U - 1, ~0U - 1);
+ }
+ static unsigned getHashValue(const std::pair<unsigned, unsigned> &P) {
+ return P.first ^ P.second;
+ }
+ static unsigned isEqual(const std::pair<unsigned, unsigned> &LHS,
+ const std::pair<unsigned, unsigned> &RHS) {
+ return LHS == RHS;
+ }
+ };
+
+ struct ConstraintKeyInfo {
+ static inline Constraint getEmptyKey() {
+ return Constraint(Constraint::Copy, ~0U, ~0U, ~0U);
+ }
+ static inline Constraint getTombstoneKey() {
+ return Constraint(Constraint::Copy, ~0U - 1, ~0U - 1, ~0U - 1);
+ }
+ static unsigned getHashValue(const Constraint &C) {
+ return C.Src ^ C.Dest ^ C.Type ^ C.Offset;
+ }
+ static bool isEqual(const Constraint &LHS,
+ const Constraint &RHS) {
+ return LHS.Type == RHS.Type && LHS.Dest == RHS.Dest
+ && LHS.Src == RHS.Src && LHS.Offset == RHS.Offset;
+ }
+ };
+
+ // Node class - This class is used to represent a node in the constraint
+ // graph. Due to various optimizations, it is not always the case that
+ // there is a mapping from a Node to a Value. In particular, we add
+ // artificial Node's that represent the set of pointed-to variables shared
+ // for each location equivalent Node.
+ struct Node {
+ private:
+ static volatile sys::cas_flag Counter;
+
+ public:
+ Value *Val;
+ SparseBitVector<> *Edges;
+ SparseBitVector<> *PointsTo;
+ SparseBitVector<> *OldPointsTo;
+ std::list<Constraint> Constraints;
+
+ // Pointer and location equivalence labels
+ unsigned PointerEquivLabel;
+ unsigned LocationEquivLabel;
+ // Predecessor edges, both real and implicit
+ SparseBitVector<> *PredEdges;
+ SparseBitVector<> *ImplicitPredEdges;
+ // Set of nodes that point to us, only use for location equivalence.
+ SparseBitVector<> *PointedToBy;
+ // Number of incoming edges, used during variable substitution to early
+ // free the points-to sets
+ unsigned NumInEdges;
+ // True if our points-to set is in the Set2PEClass map
+ bool StoredInHash;
+ // True if our node has no indirect constraints (complex or otherwise)
+ bool Direct;
+ // True if the node is address taken, *or* it is part of a group of nodes
+ // that must be kept together. This is set to true for functions and
+ // their arg nodes, which must be kept at the same position relative to
+ // their base function node.
+ bool AddressTaken;
+
+ // Nodes in cycles (or in equivalence classes) are united together using a
+ // standard union-find representation with path compression. NodeRep
+ // gives the index into GraphNodes for the representative Node.
+ unsigned NodeRep;
+
+ // Modification timestamp. Assigned from Counter.
+ // Used for work list prioritization.
+ unsigned Timestamp;
+
+ explicit Node(bool direct = true) :
+ Val(0), Edges(0), PointsTo(0), OldPointsTo(0),
+ PointerEquivLabel(0), LocationEquivLabel(0), PredEdges(0),
+ ImplicitPredEdges(0), PointedToBy(0), NumInEdges(0),
+ StoredInHash(false), Direct(direct), AddressTaken(false),
+ NodeRep(SelfRep), Timestamp(0) { }
+
+ Node *setValue(Value *V) {
+ assert(Val == 0 && "Value already set for this node!");
+ Val = V;
+ return this;
+ }
+
+ /// getValue - Return the LLVM value corresponding to this node.
+ ///
+ Value *getValue() const { return Val; }
+
+ /// addPointerTo - Add a pointer to the list of pointees of this node,
+ /// returning true if this caused a new pointer to be added, or false if
+ /// we already knew about the points-to relation.
+ bool addPointerTo(unsigned Node) {
+ return PointsTo->test_and_set(Node);
+ }
+
+ /// intersects - Return true if the points-to set of this node intersects
+ /// with the points-to set of the specified node.
+ bool intersects(Node *N) const;
+
+ /// intersectsIgnoring - Return true if the points-to set of this node
+ /// intersects with the points-to set of the specified node on any nodes
+ /// except for the specified node to ignore.
+ bool intersectsIgnoring(Node *N, unsigned) const;
+
+ // Timestamp a node (used for work list prioritization)
+ void Stamp() {
+ Timestamp = sys::AtomicIncrement(&Counter);
+ --Timestamp;
+ }
+
+ bool isRep() const {
+ return( (int) NodeRep < 0 );
+ }
+ };
+
+ struct WorkListElement {
+ Node* node;
+ unsigned Timestamp;
+ WorkListElement(Node* n, unsigned t) : node(n), Timestamp(t) {}
+
+ // Note that we reverse the sense of the comparison because we
+ // actually want to give low timestamps the priority over high,
+ // whereas priority is typically interpreted as a greater value is
+ // given high priority.
+ bool operator<(const WorkListElement& that) const {
+ return( this->Timestamp > that.Timestamp );
+ }
+ };
+
+ // Priority-queue based work list specialized for Nodes.
+ class WorkList {
+ std::priority_queue<WorkListElement> Q;
+
+ public:
+ void insert(Node* n) {
+ Q.push( WorkListElement(n, n->Timestamp) );
+ }
+
+ // We automatically discard non-representative nodes and nodes
+ // that were in the work list twice (we keep a copy of the
+ // timestamp in the work list so we can detect this situation by
+ // comparing against the node's current timestamp).
+ Node* pop() {
+ while( !Q.empty() ) {
+ WorkListElement x = Q.top(); Q.pop();
+ Node* INode = x.node;
+
+ if( INode->isRep() &&
+ INode->Timestamp == x.Timestamp ) {
+ return(x.node);
+ }
+ }
+ return(0);
+ }
+
+ bool empty() {
+ return Q.empty();
+ }
+ };
+
+ /// GraphNodes - This vector is populated as part of the object
+ /// identification stage of the analysis, which populates this vector with a
+ /// node for each memory object and fills in the ValueNodes map.
+ std::vector<Node> GraphNodes;
+
+ /// ValueNodes - This map indicates the Node that a particular Value* is
+ /// represented by. This contains entries for all pointers.
+ DenseMap<Value*, unsigned> ValueNodes;
+
+ /// ObjectNodes - This map contains entries for each memory object in the
+ /// program: globals, alloca's and mallocs.
+ DenseMap<Value*, unsigned> ObjectNodes;
+
+ /// ReturnNodes - This map contains an entry for each function in the
+ /// program that returns a value.
+ DenseMap<Function*, unsigned> ReturnNodes;
+
+ /// VarargNodes - This map contains the entry used to represent all pointers
+ /// passed through the varargs portion of a function call for a particular
+ /// function. An entry is not present in this map for functions that do not
+ /// take variable arguments.
+ DenseMap<Function*, unsigned> VarargNodes;
+
+
+ /// Constraints - This vector contains a list of all of the constraints
+ /// identified by the program.
+ std::vector<Constraint> Constraints;
+
+ // Map from graph node to maximum K value that is allowed (for functions,
+ // this is equivalent to the number of arguments + CallFirstArgPos)
+ std::map<unsigned, unsigned> MaxK;
+
+ /// This enum defines the GraphNodes indices that correspond to important
+ /// fixed sets.
+ enum {
+ UniversalSet = 0,
+ NullPtr = 1,
+ NullObject = 2,
+ NumberSpecialNodes
+ };
+ // Stack for Tarjan's
+ std::stack<unsigned> SCCStack;
+ // Map from Graph Node to DFS number
+ std::vector<unsigned> Node2DFS;
+ // Map from Graph Node to Deleted from graph.
+ std::vector<bool> Node2Deleted;
+ // Same as Node Maps, but implemented as std::map because it is faster to
+ // clear
+ std::map<unsigned, unsigned> Tarjan2DFS;
+ std::map<unsigned, bool> Tarjan2Deleted;
+ // Current DFS number
+ unsigned DFSNumber;
+
+ // Work lists.
+ WorkList w1, w2;
+ WorkList *CurrWL, *NextWL; // "current" and "next" work lists
+
+ // Offline variable substitution related things
+
+ // Temporary rep storage, used because we can't collapse SCC's in the
+ // predecessor graph by uniting the variables permanently, we can only do so
+ // for the successor graph.
+ std::vector<unsigned> VSSCCRep;
+ // Mapping from node to whether we have visited it during SCC finding yet.
+ std::vector<bool> Node2Visited;
+ // During variable substitution, we create unknowns to represent the unknown
+ // value that is a dereference of a variable. These nodes are known as
+ // "ref" nodes (since they represent the value of dereferences).
+ unsigned FirstRefNode;
+ // During HVN, we create represent address taken nodes as if they were
+ // unknown (since HVN, unlike HU, does not evaluate unions).
+ unsigned FirstAdrNode;
+ // Current pointer equivalence class number
+ unsigned PEClass;
+ // Mapping from points-to sets to equivalence classes
+ typedef DenseMap<SparseBitVector<> *, unsigned, BitmapKeyInfo> BitVectorMap;
+ BitVectorMap Set2PEClass;
+ // Mapping from pointer equivalences to the representative node. -1 if we
+ // have no representative node for this pointer equivalence class yet.
+ std::vector<int> PEClass2Node;
+ // Mapping from pointer equivalences to representative node. This includes
+ // pointer equivalent but not location equivalent variables. -1 if we have
+ // no representative node for this pointer equivalence class yet.
+ std::vector<int> PENLEClass2Node;
+ // Union/Find for HCD
+ std::vector<unsigned> HCDSCCRep;
+ // HCD's offline-detected cycles; "Statically DeTected"
+ // -1 if not part of such a cycle, otherwise a representative node.
+ std::vector<int> SDT;
+ // Whether to use SDT (UniteNodes can use it during solving, but not before)
+ bool SDTActive;
+
+ public:
+ static char ID;
+ Andersens() : ModulePass(&ID) {}
+
+ bool runOnModule(Module &M) {
+ InitializeAliasAnalysis(this);
+ IdentifyObjects(M);
+ CollectConstraints(M);
+#undef DEBUG_TYPE
+#define DEBUG_TYPE "anders-aa-constraints"
+ DEBUG(PrintConstraints());
+#undef DEBUG_TYPE
+#define DEBUG_TYPE "anders-aa"
+ SolveConstraints();
+ DEBUG(PrintPointsToGraph());
+
+ // Free the constraints list, as we don't need it to respond to alias
+ // requests.
+ std::vector<Constraint>().swap(Constraints);
+ //These are needed for Print() (-analyze in opt)
+ //ObjectNodes.clear();
+ //ReturnNodes.clear();
+ //VarargNodes.clear();
+ return false;
+ }
+
+ void releaseMemory() {
+ // FIXME: Until we have transitively required passes working correctly,
+ // this cannot be enabled! Otherwise, using -count-aa with the pass
+ // causes memory to be freed too early. :(
+#if 0
+ // The memory objects and ValueNodes data structures at the only ones that
+ // are still live after construction.
+ std::vector<Node>().swap(GraphNodes);
+ ValueNodes.clear();
+#endif
+ }
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AliasAnalysis::getAnalysisUsage(AU);
+ AU.setPreservesAll(); // Does not transform code
+ }
+
+ /// getAdjustedAnalysisPointer - This method is used when a pass implements
+ /// an analysis interface through multiple inheritance. If needed, it
+ /// should override this to adjust the this pointer as needed for the
+ /// specified pass info.
+ virtual void *getAdjustedAnalysisPointer(const PassInfo *PI) {
+ if (PI->isPassID(&AliasAnalysis::ID))
+ return (AliasAnalysis*)this;
+ return this;
+ }
+
+ //------------------------------------------------
+ // Implement the AliasAnalysis API
+ //
+ AliasResult alias(const Value *V1, unsigned V1Size,
+ const Value *V2, unsigned V2Size);
+ virtual ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size);
+ virtual ModRefResult getModRefInfo(CallSite CS1, CallSite CS2);
+ bool pointsToConstantMemory(const Value *P);
+
+ virtual void deleteValue(Value *V) {
+ ValueNodes.erase(V);
+ getAnalysis<AliasAnalysis>().deleteValue(V);
+ }
+
+ virtual void copyValue(Value *From, Value *To) {
+ ValueNodes[To] = ValueNodes[From];
+ getAnalysis<AliasAnalysis>().copyValue(From, To);
+ }
+
+ private:
+ /// getNode - Return the node corresponding to the specified pointer scalar.
+ ///
+ unsigned getNode(Value *V) {
+ if (Constant *C = dyn_cast<Constant>(V))
+ if (!isa<GlobalValue>(C))
+ return getNodeForConstantPointer(C);
+
+ DenseMap<Value*, unsigned>::iterator I = ValueNodes.find(V);
+ if (I == ValueNodes.end()) {
+#ifndef NDEBUG
+ V->dump();
+#endif
+ llvm_unreachable("Value does not have a node in the points-to graph!");
+ }
+ return I->second;
+ }
+
+ /// getObject - Return the node corresponding to the memory object for the
+ /// specified global or allocation instruction.
+ unsigned getObject(Value *V) const {
+ DenseMap<Value*, unsigned>::const_iterator I = ObjectNodes.find(V);
+ assert(I != ObjectNodes.end() &&
+ "Value does not have an object in the points-to graph!");
+ return I->second;
+ }
+
+ /// getReturnNode - Return the node representing the return value for the
+ /// specified function.
+ unsigned getReturnNode(Function *F) const {
+ DenseMap<Function*, unsigned>::const_iterator I = ReturnNodes.find(F);
+ assert(I != ReturnNodes.end() && "Function does not return a value!");
+ return I->second;
+ }
+
+ /// getVarargNode - Return the node representing the variable arguments
+ /// formal for the specified function.
+ unsigned getVarargNode(Function *F) const {
+ DenseMap<Function*, unsigned>::const_iterator I = VarargNodes.find(F);
+ assert(I != VarargNodes.end() && "Function does not take var args!");
+ return I->second;
+ }
+
+ /// getNodeValue - Get the node for the specified LLVM value and set the
+ /// value for it to be the specified value.
+ unsigned getNodeValue(Value &V) {
+ unsigned Index = getNode(&V);
+ GraphNodes[Index].setValue(&V);
+ return Index;
+ }
+
+ unsigned UniteNodes(unsigned First, unsigned Second,
+ bool UnionByRank = true);
+ unsigned FindNode(unsigned Node);
+ unsigned FindNode(unsigned Node) const;
+
+ void IdentifyObjects(Module &M);
+ void CollectConstraints(Module &M);
+ bool AnalyzeUsesOfFunction(Value *);
+ void CreateConstraintGraph();
+ void OptimizeConstraints();
+ unsigned FindEquivalentNode(unsigned, unsigned);
+ void ClumpAddressTaken();
+ void RewriteConstraints();
+ void HU();
+ void HVN();
+ void HCD();
+ void Search(unsigned Node);
+ void UnitePointerEquivalences();
+ void SolveConstraints();
+ bool QueryNode(unsigned Node);
+ void Condense(unsigned Node);
+ void HUValNum(unsigned Node);
+ void HVNValNum(unsigned Node);
+ unsigned getNodeForConstantPointer(Constant *C);
+ unsigned getNodeForConstantPointerTarget(Constant *C);
+ void AddGlobalInitializerConstraints(unsigned, Constant *C);
+
+ void AddConstraintsForNonInternalLinkage(Function *F);
+ void AddConstraintsForCall(CallSite CS, Function *F);
+ bool AddConstraintsForExternalCall(CallSite CS, Function *F);
+
+
+ void PrintNode(const Node *N) const;
+ void PrintConstraints() const ;
+ void PrintConstraint(const Constraint &) const;
+ void PrintLabels() const;
+ void PrintPointsToGraph() const;
+
+ //===------------------------------------------------------------------===//
+ // Instruction visitation methods for adding constraints
+ //
+ friend class InstVisitor<Andersens>;
+ void visitReturnInst(ReturnInst &RI);
+ void visitInvokeInst(InvokeInst &II) { visitCallSite(CallSite(&II)); }
+ void visitCallInst(CallInst &CI) {
+ if (isMalloc(&CI)) visitAlloc(CI);
+ else visitCallSite(CallSite(&CI));
+ }
+ void visitCallSite(CallSite CS);
+ void visitAllocaInst(AllocaInst &I);
+ void visitAlloc(Instruction &I);
+ void visitLoadInst(LoadInst &LI);
+ void visitStoreInst(StoreInst &SI);
+ void visitGetElementPtrInst(GetElementPtrInst &GEP);
+ void visitPHINode(PHINode &PN);
+ void visitCastInst(CastInst &CI);
+ void visitICmpInst(ICmpInst &ICI) {} // NOOP!
+ void visitFCmpInst(FCmpInst &ICI) {} // NOOP!
+ void visitSelectInst(SelectInst &SI);
+ void visitVAArg(VAArgInst &I);
+ void visitInstruction(Instruction &I);
+
+ //===------------------------------------------------------------------===//
+ // Implement Analyize interface
+ //
+ void print(raw_ostream &O, const Module*) const {
+ PrintPointsToGraph();
+ }
+ };
+}
+
+char Andersens::ID = 0;
+static RegisterPass<Andersens>
+X("anders-aa", "Andersen's Interprocedural Alias Analysis (experimental)",
+ false, true);
+static RegisterAnalysisGroup<AliasAnalysis> Y(X);
+
+// Initialize Timestamp Counter (static).
+volatile llvm::sys::cas_flag Andersens::Node::Counter = 0;
+
+ModulePass *llvm::createAndersensPass() { return new Andersens(); }
+
+//===----------------------------------------------------------------------===//
+// AliasAnalysis Interface Implementation
+//===----------------------------------------------------------------------===//
+
+AliasAnalysis::AliasResult Andersens::alias(const Value *V1, unsigned V1Size,
+ const Value *V2, unsigned V2Size) {
+ Node *N1 = &GraphNodes[FindNode(getNode(const_cast<Value*>(V1)))];
+ Node *N2 = &GraphNodes[FindNode(getNode(const_cast<Value*>(V2)))];
+
+ // Check to see if the two pointers are known to not alias. They don't alias
+ // if their points-to sets do not intersect.
+ if (!N1->intersectsIgnoring(N2, NullObject))
+ return NoAlias;
+
+ return AliasAnalysis::alias(V1, V1Size, V2, V2Size);
+}
+
+AliasAnalysis::ModRefResult
+Andersens::getModRefInfo(CallSite CS, Value *P, unsigned Size) {
+ // The only thing useful that we can contribute for mod/ref information is
+ // when calling external function calls: if we know that memory never escapes
+ // from the program, it cannot be modified by an external call.
+ //
+ // NOTE: This is not really safe, at least not when the entire program is not
+ // available. The deal is that the external function could call back into the
+ // program and modify stuff. We ignore this technical niggle for now. This
+ // is, after all, a "research quality" implementation of Andersen's analysis.
+ if (Function *F = CS.getCalledFunction())
+ if (F->isDeclaration()) {
+ Node *N1 = &GraphNodes[FindNode(getNode(P))];
+
+ if (N1->PointsTo->empty())
+ return NoModRef;
+#if FULL_UNIVERSAL
+ if (!UniversalSet->PointsTo->test(FindNode(getNode(P))))
+ return NoModRef; // Universal set does not contain P
+#else
+ if (!N1->PointsTo->test(UniversalSet))
+ return NoModRef; // P doesn't point to the universal set.
+#endif
+ }
+
+ return AliasAnalysis::getModRefInfo(CS, P, Size);
+}
+
+AliasAnalysis::ModRefResult
+Andersens::getModRefInfo(CallSite CS1, CallSite CS2) {
+ return AliasAnalysis::getModRefInfo(CS1,CS2);
+}
+
+/// pointsToConstantMemory - If we can determine that this pointer only points
+/// to constant memory, return true. In practice, this means that if the
+/// pointer can only point to constant globals, functions, or the null pointer,
+/// return true.
+///
+bool Andersens::pointsToConstantMemory(const Value *P) {
+ Node *N = &GraphNodes[FindNode(getNode(const_cast<Value*>(P)))];
+ unsigned i;
+
+ for (SparseBitVector<>::iterator bi = N->PointsTo->begin();
+ bi != N->PointsTo->end();
+ ++bi) {
+ i = *bi;
+ Node *Pointee = &GraphNodes[i];
+ if (Value *V = Pointee->getValue()) {
+ if (!isa<GlobalValue>(V) || (isa<GlobalVariable>(V) &&
+ !cast<GlobalVariable>(V)->isConstant()))
+ return AliasAnalysis::pointsToConstantMemory(P);
+ } else {
+ if (i != NullObject)
+ return AliasAnalysis::pointsToConstantMemory(P);
+ }
+ }
+
+ return true;
+}
+
+//===----------------------------------------------------------------------===//
+// Object Identification Phase
+//===----------------------------------------------------------------------===//
+
+/// IdentifyObjects - This stage scans the program, adding an entry to the
+/// GraphNodes list for each memory object in the program (global stack or
+/// heap), and populates the ValueNodes and ObjectNodes maps for these objects.
+///
+void Andersens::IdentifyObjects(Module &M) {
+ unsigned NumObjects = 0;
+
+ // Object #0 is always the universal set: the object that we don't know
+ // anything about.
+ assert(NumObjects == UniversalSet && "Something changed!");
+ ++NumObjects;
+
+ // Object #1 always represents the null pointer.
+ assert(NumObjects == NullPtr && "Something changed!");
+ ++NumObjects;
+
+ // Object #2 always represents the null object (the object pointed to by null)
+ assert(NumObjects == NullObject && "Something changed!");
+ ++NumObjects;
+
+ // Add all the globals first.
+ for (Module::global_iterator I = M.global_begin(), E = M.global_end();
+ I != E; ++I) {
+ ObjectNodes[I] = NumObjects++;
+ ValueNodes[I] = NumObjects++;
+ }
+
+ // Add nodes for all of the functions and the instructions inside of them.
+ for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
+ // The function itself is a memory object.
+ unsigned First = NumObjects;
+ ValueNodes[F] = NumObjects++;
+ if (isa<PointerType>(F->getFunctionType()->getReturnType()))
+ ReturnNodes[F] = NumObjects++;
+ if (F->getFunctionType()->isVarArg())
+ VarargNodes[F] = NumObjects++;
+
+
+ // Add nodes for all of the incoming pointer arguments.
+ for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
+ I != E; ++I)
+ {
+ if (isa<PointerType>(I->getType()))
+ ValueNodes[I] = NumObjects++;
+ }
+ MaxK[First] = NumObjects - First;
+
+ // Scan the function body, creating a memory object for each heap/stack
+ // allocation in the body of the function and a node to represent all
+ // pointer values defined by instructions and used as operands.
+ for (inst_iterator II = inst_begin(F), E = inst_end(F); II != E; ++II) {
+ // If this is an heap or stack allocation, create a node for the memory
+ // object.
+ if (isa<PointerType>(II->getType())) {
+ ValueNodes[&*II] = NumObjects++;
+ if (AllocaInst *AI = dyn_cast<AllocaInst>(&*II))
+ ObjectNodes[AI] = NumObjects++;
+ else if (isMalloc(&*II))
+ ObjectNodes[&*II] = NumObjects++;
+ }
+
+ // Calls to inline asm need to be added as well because the callee isn't
+ // referenced anywhere else.
+ if (CallInst *CI = dyn_cast<CallInst>(&*II)) {
+ Value *Callee = CI->getCalledValue();
+ if (isa<InlineAsm>(Callee))
+ ValueNodes[Callee] = NumObjects++;
+ }
+ }
+ }
+
+ // Now that we know how many objects to create, make them all now!
+ GraphNodes.resize(NumObjects);
+ NumNodes += NumObjects;
+}
+
+//===----------------------------------------------------------------------===//
+// Constraint Identification Phase
+//===----------------------------------------------------------------------===//
+
+/// getNodeForConstantPointer - Return the node corresponding to the constant
+/// pointer itself.
+unsigned Andersens::getNodeForConstantPointer(Constant *C) {
+ assert(isa<PointerType>(C->getType()) && "Not a constant pointer!");
+
+ if (isa<ConstantPointerNull>(C) || isa<UndefValue>(C))
+ return NullPtr;
+ else if (GlobalValue *GV = dyn_cast<GlobalValue>(C))
+ return getNode(GV);
+ else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
+ switch (CE->getOpcode()) {
+ case Instruction::GetElementPtr:
+ return getNodeForConstantPointer(CE->getOperand(0));
+ case Instruction::IntToPtr:
+ return UniversalSet;
+ case Instruction::BitCast:
+ return getNodeForConstantPointer(CE->getOperand(0));
+ default:
+ errs() << "Constant Expr not yet handled: " << *CE << "\n";
+ llvm_unreachable(0);
+ }
+ } else {
+ llvm_unreachable("Unknown constant pointer!");
+ }
+ return 0;
+}
+
+/// getNodeForConstantPointerTarget - Return the node POINTED TO by the
+/// specified constant pointer.
+unsigned Andersens::getNodeForConstantPointerTarget(Constant *C) {
+ assert(isa<PointerType>(C->getType()) && "Not a constant pointer!");
+
+ if (isa<ConstantPointerNull>(C))
+ return NullObject;
+ else if (GlobalValue *GV = dyn_cast<GlobalValue>(C))
+ return getObject(GV);
+ else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
+ switch (CE->getOpcode()) {
+ case Instruction::GetElementPtr:
+ return getNodeForConstantPointerTarget(CE->getOperand(0));
+ case Instruction::IntToPtr:
+ return UniversalSet;
+ case Instruction::BitCast:
+ return getNodeForConstantPointerTarget(CE->getOperand(0));
+ default:
+ errs() << "Constant Expr not yet handled: " << *CE << "\n";
+ llvm_unreachable(0);
+ }
+ } else {
+ llvm_unreachable("Unknown constant pointer!");
+ }
+ return 0;
+}
+
+/// AddGlobalInitializerConstraints - Add inclusion constraints for the memory
+/// object N, which contains values indicated by C.
+void Andersens::AddGlobalInitializerConstraints(unsigned NodeIndex,
+ Constant *C) {
+ if (C->getType()->isSingleValueType()) {
+ if (isa<PointerType>(C->getType()))
+ Constraints.push_back(Constraint(Constraint::Copy, NodeIndex,
+ getNodeForConstantPointer(C)));
+ } else if (C->isNullValue()) {
+ Constraints.push_back(Constraint(Constraint::Copy, NodeIndex,
+ NullObject));
+ return;
+ } else if (!isa<UndefValue>(C)) {
+ // If this is an array or struct, include constraints for each element.
+ assert(isa<ConstantArray>(C) || isa<ConstantStruct>(C));
+ for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i)
+ AddGlobalInitializerConstraints(NodeIndex,
+ cast<Constant>(C->getOperand(i)));
+ }
+}
+
+/// AddConstraintsForNonInternalLinkage - If this function does not have
+/// internal linkage, realize that we can't trust anything passed into or
+/// returned by this function.
+void Andersens::AddConstraintsForNonInternalLinkage(Function *F) {
+ for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
+ if (isa<PointerType>(I->getType()))
+ // If this is an argument of an externally accessible function, the
+ // incoming pointer might point to anything.
+ Constraints.push_back(Constraint(Constraint::Copy, getNode(I),
+ UniversalSet));
+}
+
+/// AddConstraintsForCall - If this is a call to a "known" function, add the
+/// constraints and return true. If this is a call to an unknown function,
+/// return false.
+bool Andersens::AddConstraintsForExternalCall(CallSite CS, Function *F) {
+ assert(F->isDeclaration() && "Not an external function!");
+
+ // These functions don't induce any points-to constraints.
+ if (F->getName() == "atoi" || F->getName() == "atof" ||
+ F->getName() == "atol" || F->getName() == "atoll" ||
+ F->getName() == "remove" || F->getName() == "unlink" ||
+ F->getName() == "rename" || F->getName() == "memcmp" ||
+ F->getName() == "llvm.memset" ||
+ F->getName() == "strcmp" || F->getName() == "strncmp" ||
+ F->getName() == "execl" || F->getName() == "execlp" ||
+ F->getName() == "execle" || F->getName() == "execv" ||
+ F->getName() == "execvp" || F->getName() == "chmod" ||
+ F->getName() == "puts" || F->getName() == "write" ||
+ F->getName() == "open" || F->getName() == "create" ||
+ F->getName() == "truncate" || F->getName() == "chdir" ||
+ F->getName() == "mkdir" || F->getName() == "rmdir" ||
+ F->getName() == "read" || F->getName() == "pipe" ||
+ F->getName() == "wait" || F->getName() == "time" ||
+ F->getName() == "stat" || F->getName() == "fstat" ||
+ F->getName() == "lstat" || F->getName() == "strtod" ||
+ F->getName() == "strtof" || F->getName() == "strtold" ||
+ F->getName() == "fopen" || F->getName() == "fdopen" ||
+ F->getName() == "freopen" ||
+ F->getName() == "fflush" || F->getName() == "feof" ||
+ F->getName() == "fileno" || F->getName() == "clearerr" ||
+ F->getName() == "rewind" || F->getName() == "ftell" ||
+ F->getName() == "ferror" || F->getName() == "fgetc" ||
+ F->getName() == "fgetc" || F->getName() == "_IO_getc" ||
+ F->getName() == "fwrite" || F->getName() == "fread" ||
+ F->getName() == "fgets" || F->getName() == "ungetc" ||
+ F->getName() == "fputc" ||
+ F->getName() == "fputs" || F->getName() == "putc" ||
+ F->getName() == "ftell" || F->getName() == "rewind" ||
+ F->getName() == "_IO_putc" || F->getName() == "fseek" ||
+ F->getName() == "fgetpos" || F->getName() == "fsetpos" ||
+ F->getName() == "printf" || F->getName() == "fprintf" ||
+ F->getName() == "sprintf" || F->getName() == "vprintf" ||
+ F->getName() == "vfprintf" || F->getName() == "vsprintf" ||
+ F->getName() == "scanf" || F->getName() == "fscanf" ||
+ F->getName() == "sscanf" || F->getName() == "__assert_fail" ||
+ F->getName() == "modf")
+ return true;
+
+
+ // These functions do induce points-to edges.
+ if (F->getName() == "llvm.memcpy" ||
+ F->getName() == "llvm.memmove" ||
+ F->getName() == "memmove") {
+
+ const FunctionType *FTy = F->getFunctionType();
+ if (FTy->getNumParams() > 1 &&
+ isa<PointerType>(FTy->getParamType(0)) &&
+ isa<PointerType>(FTy->getParamType(1))) {
+
+ // *Dest = *Src, which requires an artificial graph node to represent the
+ // constraint. It is broken up into *Dest = temp, temp = *Src
+ unsigned FirstArg = getNode(CS.getArgument(0));
+ unsigned SecondArg = getNode(CS.getArgument(1));
+ unsigned TempArg = GraphNodes.size();
+ GraphNodes.push_back(Node());
+ Constraints.push_back(Constraint(Constraint::Store,
+ FirstArg, TempArg));
+ Constraints.push_back(Constraint(Constraint::Load,
+ TempArg, SecondArg));
+ // In addition, Dest = Src
+ Constraints.push_back(Constraint(Constraint::Copy,
+ FirstArg, SecondArg));
+ return true;
+ }
+ }
+
+ // Result = Arg0
+ if (F->getName() == "realloc" || F->getName() == "strchr" ||
+ F->getName() == "strrchr" || F->getName() == "strstr" ||
+ F->getName() == "strtok") {
+ const FunctionType *FTy = F->getFunctionType();
+ if (FTy->getNumParams() > 0 &&
+ isa<PointerType>(FTy->getParamType(0))) {
+ Constraints.push_back(Constraint(Constraint::Copy,
+ getNode(CS.getInstruction()),
+ getNode(CS.getArgument(0))));
+ return true;
+ }
+ }
+
+ return false;
+}
+
+
+
+/// AnalyzeUsesOfFunction - Look at all of the users of the specified function.
+/// If this is used by anything complex (i.e., the address escapes), return
+/// true.
+bool Andersens::AnalyzeUsesOfFunction(Value *V) {
+
+ if (!isa<PointerType>(V->getType())) return true;
+
+ for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
+ if (isa<LoadInst>(*UI)) {
+ return false;
+ } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
+ if (V == SI->getOperand(1)) {
+ return false;
+ } else if (SI->getOperand(1)) {
+ return true; // Storing the pointer
+ }
+ } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) {
+ if (AnalyzeUsesOfFunction(GEP)) return true;
+ } else if (isFreeCall(*UI)) {
+ return false;
+ } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
+ // Make sure that this is just the function being called, not that it is
+ // passing into the function.
+ for (unsigned i = 1, e = CI->getNumOperands(); i != e; ++i)
+ if (CI->getOperand(i) == V) return true;
+ } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
+ // Make sure that this is just the function being called, not that it is
+ // passing into the function.
+ for (unsigned i = 3, e = II->getNumOperands(); i != e; ++i)
+ if (II->getOperand(i) == V) return true;
+ } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
+ if (CE->getOpcode() == Instruction::GetElementPtr ||
+ CE->getOpcode() == Instruction::BitCast) {
+ if (AnalyzeUsesOfFunction(CE))
+ return true;
+ } else {
+ return true;
+ }
+ } else if (ICmpInst *ICI = dyn_cast<ICmpInst>(*UI)) {
+ if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
+ return true; // Allow comparison against null.
+ } else {
+ return true;
+ }
+ return false;
+}
+
+/// CollectConstraints - This stage scans the program, adding a constraint to
+/// the Constraints list for each instruction in the program that induces a
+/// constraint, and setting up the initial points-to graph.
+///
+void Andersens::CollectConstraints(Module &M) {
+ // First, the universal set points to itself.
+ Constraints.push_back(Constraint(Constraint::AddressOf, UniversalSet,
+ UniversalSet));
+ Constraints.push_back(Constraint(Constraint::Store, UniversalSet,
+ UniversalSet));
+
+ // Next, the null pointer points to the null object.
+ Constraints.push_back(Constraint(Constraint::AddressOf, NullPtr, NullObject));
+
+ // Next, add any constraints on global variables and their initializers.
+ for (Module::global_iterator I = M.global_begin(), E = M.global_end();
+ I != E; ++I) {
+ // Associate the address of the global object as pointing to the memory for
+ // the global: &G = <G memory>
+ unsigned ObjectIndex = getObject(I);
+ Node *Object = &GraphNodes[ObjectIndex];
+ Object->setValue(I);
+ Constraints.push_back(Constraint(Constraint::AddressOf, getNodeValue(*I),
+ ObjectIndex));
+
+ if (I->hasDefinitiveInitializer()) {
+ AddGlobalInitializerConstraints(ObjectIndex, I->getInitializer());
+ } else {
+ // If it doesn't have an initializer (i.e. it's defined in another
+ // translation unit), it points to the universal set.
+ Constraints.push_back(Constraint(Constraint::Copy, ObjectIndex,
+ UniversalSet));
+ }
+ }
+
+ for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
+ // Set up the return value node.
+ if (isa<PointerType>(F->getFunctionType()->getReturnType()))
+ GraphNodes[getReturnNode(F)].setValue(F);
+ if (F->getFunctionType()->isVarArg())
+ GraphNodes[getVarargNode(F)].setValue(F);
+
+ // Set up incoming argument nodes.
+ for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
+ I != E; ++I)
+ if (isa<PointerType>(I->getType()))
+ getNodeValue(*I);
+
+ // At some point we should just add constraints for the escaping functions
+ // at solve time, but this slows down solving. For now, we simply mark
+ // address taken functions as escaping and treat them as external.
+ if (!F->hasLocalLinkage() || AnalyzeUsesOfFunction(F))
+ AddConstraintsForNonInternalLinkage(F);
+
+ if (!F->isDeclaration()) {
+ // Scan the function body, creating a memory object for each heap/stack
+ // allocation in the body of the function and a node to represent all
+ // pointer values defined by instructions and used as operands.
+ visit(F);
+ } else {
+ // External functions that return pointers return the universal set.
+ if (isa<PointerType>(F->getFunctionType()->getReturnType()))
+ Constraints.push_back(Constraint(Constraint::Copy,
+ getReturnNode(F),
+ UniversalSet));
+
+ // Any pointers that are passed into the function have the universal set
+ // stored into them.
+ for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
+ I != E; ++I)
+ if (isa<PointerType>(I->getType())) {
+ // Pointers passed into external functions could have anything stored
+ // through them.
+ Constraints.push_back(Constraint(Constraint::Store, getNode(I),
+ UniversalSet));
+ // Memory objects passed into external function calls can have the
+ // universal set point to them.
+#if FULL_UNIVERSAL
+ Constraints.push_back(Constraint(Constraint::Copy,
+ UniversalSet,
+ getNode(I)));
+#else
+ Constraints.push_back(Constraint(Constraint::Copy,
+ getNode(I),
+ UniversalSet));
+#endif
+ }
+
+ // If this is an external varargs function, it can also store pointers
+ // into any pointers passed through the varargs section.
+ if (F->getFunctionType()->isVarArg())
+ Constraints.push_back(Constraint(Constraint::Store, getVarargNode(F),
+ UniversalSet));
+ }
+ }
+ NumConstraints += Constraints.size();
+}
+
+
+void Andersens::visitInstruction(Instruction &I) {
+#ifdef NDEBUG
+ return; // This function is just a big assert.
+#endif
+ if (isa<BinaryOperator>(I))
+ return;
+ // Most instructions don't have any effect on pointer values.
+ switch (I.getOpcode()) {
+ case Instruction::Br:
+ case Instruction::Switch:
+ case Instruction::Unwind:
+ case Instruction::Unreachable:
+ case Instruction::ICmp:
+ case Instruction::FCmp:
+ return;
+ default:
+ // Is this something we aren't handling yet?
+ errs() << "Unknown instruction: " << I;
+ llvm_unreachable(0);
+ }
+}
+
+void Andersens::visitAllocaInst(AllocaInst &I) {
+ visitAlloc(I);
+}
+
+void Andersens::visitAlloc(Instruction &I) {
+ unsigned ObjectIndex = getObject(&I);
+ GraphNodes[ObjectIndex].setValue(&I);
+ Constraints.push_back(Constraint(Constraint::AddressOf, getNodeValue(I),
+ ObjectIndex));
+}
+
+void Andersens::visitReturnInst(ReturnInst &RI) {
+ if (RI.getNumOperands() && isa<PointerType>(RI.getOperand(0)->getType()))
+ // return V --> <Copy/retval{F}/v>
+ Constraints.push_back(Constraint(Constraint::Copy,
+ getReturnNode(RI.getParent()->getParent()),
+ getNode(RI.getOperand(0))));
+}
+
+void Andersens::visitLoadInst(LoadInst &LI) {
+ if (isa<PointerType>(LI.getType()))
+ // P1 = load P2 --> <Load/P1/P2>
+ Constraints.push_back(Constraint(Constraint::Load, getNodeValue(LI),
+ getNode(LI.getOperand(0))));
+}
+
+void Andersens::visitStoreInst(StoreInst &SI) {
+ if (isa<PointerType>(SI.getOperand(0)->getType()))
+ // store P1, P2 --> <Store/P2/P1>
+ Constraints.push_back(Constraint(Constraint::Store,
+ getNode(SI.getOperand(1)),
+ getNode(SI.getOperand(0))));
+}
+
+void Andersens::visitGetElementPtrInst(GetElementPtrInst &GEP) {
+ // P1 = getelementptr P2, ... --> <Copy/P1/P2>
+ Constraints.push_back(Constraint(Constraint::Copy, getNodeValue(GEP),
+ getNode(GEP.getOperand(0))));
+}
+
+void Andersens::visitPHINode(PHINode &PN) {
+ if (isa<PointerType>(PN.getType())) {
+ unsigned PNN = getNodeValue(PN);
+ for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
+ // P1 = phi P2, P3 --> <Copy/P1/P2>, <Copy/P1/P3>, ...
+ Constraints.push_back(Constraint(Constraint::Copy, PNN,
+ getNode(PN.getIncomingValue(i))));
+ }
+}
+
+void Andersens::visitCastInst(CastInst &CI) {
+ Value *Op = CI.getOperand(0);
+ if (isa<PointerType>(CI.getType())) {
+ if (isa<PointerType>(Op->getType())) {
+ // P1 = cast P2 --> <Copy/P1/P2>
+ Constraints.push_back(Constraint(Constraint::Copy, getNodeValue(CI),
+ getNode(CI.getOperand(0))));
+ } else {
+ // P1 = cast int --> <Copy/P1/Univ>
+#if 0
+ Constraints.push_back(Constraint(Constraint::Copy, getNodeValue(CI),
+ UniversalSet));
+#else
+ getNodeValue(CI);
+#endif
+ }
+ } else if (isa<PointerType>(Op->getType())) {
+ // int = cast P1 --> <Copy/Univ/P1>
+#if 0
+ Constraints.push_back(Constraint(Constraint::Copy,
+ UniversalSet,
+ getNode(CI.getOperand(0))));
+#else
+ getNode(CI.getOperand(0));
+#endif
+ }
+}
+
+void Andersens::visitSelectInst(SelectInst &SI) {
+ if (isa<PointerType>(SI.getType())) {
+ unsigned SIN = getNodeValue(SI);
+ // P1 = select C, P2, P3 ---> <Copy/P1/P2>, <Copy/P1/P3>
+ Constraints.push_back(Constraint(Constraint::Copy, SIN,
+ getNode(SI.getOperand(1))));
+ Constraints.push_back(Constraint(Constraint::Copy, SIN,
+ getNode(SI.getOperand(2))));
+ }
+}
+
+void Andersens::visitVAArg(VAArgInst &I) {
+ llvm_unreachable("vaarg not handled yet!");
+}
+
+/// AddConstraintsForCall - Add constraints for a call with actual arguments
+/// specified by CS to the function specified by F. Note that the types of
+/// arguments might not match up in the case where this is an indirect call and
+/// the function pointer has been casted. If this is the case, do something
+/// reasonable.
+void Andersens::AddConstraintsForCall(CallSite CS, Function *F) {
+ Value *CallValue = CS.getCalledValue();
+ bool IsDeref = F == NULL;
+
+ // If this is a call to an external function, try to handle it directly to get
+ // some taste of context sensitivity.
+ if (F && F->isDeclaration() && AddConstraintsForExternalCall(CS, F))
+ return;
+
+ if (isa<PointerType>(CS.getType())) {
+ unsigned CSN = getNode(CS.getInstruction());
+ if (!F || isa<PointerType>(F->getFunctionType()->getReturnType())) {
+ if (IsDeref)
+ Constraints.push_back(Constraint(Constraint::Load, CSN,
+ getNode(CallValue), CallReturnPos));
+ else
+ Constraints.push_back(Constraint(Constraint::Copy, CSN,
+ getNode(CallValue) + CallReturnPos));
+ } else {
+ // If the function returns a non-pointer value, handle this just like we
+ // treat a nonpointer cast to pointer.
+ Constraints.push_back(Constraint(Constraint::Copy, CSN,
+ UniversalSet));
+ }
+ } else if (F && isa<PointerType>(F->getFunctionType()->getReturnType())) {
+#if FULL_UNIVERSAL
+ Constraints.push_back(Constraint(Constraint::Copy,
+ UniversalSet,
+ getNode(CallValue) + CallReturnPos));
+#else
+ Constraints.push_back(Constraint(Constraint::Copy,
+ getNode(CallValue) + CallReturnPos,
+ UniversalSet));
+#endif
+
+
+ }
+
+ CallSite::arg_iterator ArgI = CS.arg_begin(), ArgE = CS.arg_end();
+ bool external = !F || F->isDeclaration();
+ if (F) {
+ // Direct Call
+ Function::arg_iterator AI = F->arg_begin(), AE = F->arg_end();
+ for (; AI != AE && ArgI != ArgE; ++AI, ++ArgI)
+ {
+#if !FULL_UNIVERSAL
+ if (external && isa<PointerType>((*ArgI)->getType()))
+ {
+ // Add constraint that ArgI can now point to anything due to
+ // escaping, as can everything it points to. The second portion of
+ // this should be taken care of by universal = *universal
+ Constraints.push_back(Constraint(Constraint::Copy,
+ getNode(*ArgI),
+ UniversalSet));
+ }
+#endif
+ if (isa<PointerType>(AI->getType())) {
+ if (isa<PointerType>((*ArgI)->getType())) {
+ // Copy the actual argument into the formal argument.
+ Constraints.push_back(Constraint(Constraint::Copy, getNode(AI),
+ getNode(*ArgI)));
+ } else {
+ Constraints.push_back(Constraint(Constraint::Copy, getNode(AI),
+ UniversalSet));
+ }
+ } else if (isa<PointerType>((*ArgI)->getType())) {
+#if FULL_UNIVERSAL
+ Constraints.push_back(Constraint(Constraint::Copy,
+ UniversalSet,
+ getNode(*ArgI)));
+#else
+ Constraints.push_back(Constraint(Constraint::Copy,
+ getNode(*ArgI),
+ UniversalSet));
+#endif
+ }
+ }
+ } else {
+ //Indirect Call
+ unsigned ArgPos = CallFirstArgPos;
+ for (; ArgI != ArgE; ++ArgI) {
+ if (isa<PointerType>((*ArgI)->getType())) {
+ // Copy the actual argument into the formal argument.
+ Constraints.push_back(Constraint(Constraint::Store,
+ getNode(CallValue),
+ getNode(*ArgI), ArgPos++));
+ } else {
+ Constraints.push_back(Constraint(Constraint::Store,
+ getNode (CallValue),
+ UniversalSet, ArgPos++));
+ }
+ }
+ }
+ // Copy all pointers passed through the varargs section to the varargs node.
+ if (F && F->getFunctionType()->isVarArg())
+ for (; ArgI != ArgE; ++ArgI)
+ if (isa<PointerType>((*ArgI)->getType()))
+ Constraints.push_back(Constraint(Constraint::Copy, getVarargNode(F),
+ getNode(*ArgI)));
+ // If more arguments are passed in than we track, just drop them on the floor.
+}
+
+void Andersens::visitCallSite(CallSite CS) {
+ if (isa<PointerType>(CS.getType()))
+ getNodeValue(*CS.getInstruction());
+
+ if (Function *F = CS.getCalledFunction()) {
+ AddConstraintsForCall(CS, F);
+ } else {
+ AddConstraintsForCall(CS, NULL);
+ }
+}
+
+//===----------------------------------------------------------------------===//
+// Constraint Solving Phase
+//===----------------------------------------------------------------------===//
+
+/// intersects - Return true if the points-to set of this node intersects
+/// with the points-to set of the specified node.
+bool Andersens::Node::intersects(Node *N) const {
+ return PointsTo->intersects(N->PointsTo);
+}
+
+/// intersectsIgnoring - Return true if the points-to set of this node
+/// intersects with the points-to set of the specified node on any nodes
+/// except for the specified node to ignore.
+bool Andersens::Node::intersectsIgnoring(Node *N, unsigned Ignoring) const {
+ // TODO: If we are only going to call this with the same value for Ignoring,
+ // we should move the special values out of the points-to bitmap.
+ bool WeHadIt = PointsTo->test(Ignoring);
+ bool NHadIt = N->PointsTo->test(Ignoring);
+ bool Result = false;
+ if (WeHadIt)
+ PointsTo->reset(Ignoring);
+ if (NHadIt)
+ N->PointsTo->reset(Ignoring);
+ Result = PointsTo->intersects(N->PointsTo);
+ if (WeHadIt)
+ PointsTo->set(Ignoring);
+ if (NHadIt)
+ N->PointsTo->set(Ignoring);
+ return Result;
+}
+
+
+/// Clump together address taken variables so that the points-to sets use up
+/// less space and can be operated on faster.
+
+void Andersens::ClumpAddressTaken() {
+#undef DEBUG_TYPE
+#define DEBUG_TYPE "anders-aa-renumber"
+ std::vector<unsigned> Translate;
+ std::vector<Node> NewGraphNodes;
+
+ Translate.resize(GraphNodes.size());
+ unsigned NewPos = 0;
+
+ for (unsigned i = 0; i < Constraints.size(); ++i) {
+ Constraint &C = Constraints[i];
+ if (C.Type == Constraint::AddressOf) {
+ GraphNodes[C.Src].AddressTaken = true;
+ }
+ }
+ for (unsigned i = 0; i < NumberSpecialNodes; ++i) {
+ unsigned Pos = NewPos++;
+ Translate[i] = Pos;
+ NewGraphNodes.push_back(GraphNodes[i]);
+ DEBUG(dbgs() << "Renumbering node " << i << " to node " << Pos << "\n");
+ }
+
+ // I believe this ends up being faster than making two vectors and splicing
+ // them.
+ for (unsigned i = NumberSpecialNodes; i < GraphNodes.size(); ++i) {
+ if (GraphNodes[i].AddressTaken) {
+ unsigned Pos = NewPos++;
+ Translate[i] = Pos;
+ NewGraphNodes.push_back(GraphNodes[i]);
+ DEBUG(dbgs() << "Renumbering node " << i << " to node " << Pos << "\n");
+ }
+ }
+
+ for (unsigned i = NumberSpecialNodes; i < GraphNodes.size(); ++i) {
+ if (!GraphNodes[i].AddressTaken) {
+ unsigned Pos = NewPos++;
+ Translate[i] = Pos;
+ NewGraphNodes.push_back(GraphNodes[i]);
+ DEBUG(dbgs() << "Renumbering node " << i << " to node " << Pos << "\n");
+ }
+ }
+
+ for (DenseMap<Value*, unsigned>::iterator Iter = ValueNodes.begin();
+ Iter != ValueNodes.end();
+ ++Iter)
+ Iter->second = Translate[Iter->second];
+
+ for (DenseMap<Value*, unsigned>::iterator Iter = ObjectNodes.begin();
+ Iter != ObjectNodes.end();
+ ++Iter)
+ Iter->second = Translate[Iter->second];
+
+ for (DenseMap<Function*, unsigned>::iterator Iter = ReturnNodes.begin();
+ Iter != ReturnNodes.end();
+ ++Iter)
+ Iter->second = Translate[Iter->second];
+
+ for (DenseMap<Function*, unsigned>::iterator Iter = VarargNodes.begin();
+ Iter != VarargNodes.end();
+ ++Iter)
+ Iter->second = Translate[Iter->second];
+
+ for (unsigned i = 0; i < Constraints.size(); ++i) {
+ Constraint &C = Constraints[i];
+ C.Src = Translate[C.Src];
+ C.Dest = Translate[C.Dest];
+ }
+
+ GraphNodes.swap(NewGraphNodes);
+#undef DEBUG_TYPE
+#define DEBUG_TYPE "anders-aa"
+}
+
+/// The technique used here is described in "Exploiting Pointer and Location
+/// Equivalence to Optimize Pointer Analysis. In the 14th International Static
+/// Analysis Symposium (SAS), August 2007." It is known as the "HVN" algorithm,
+/// and is equivalent to value numbering the collapsed constraint graph without
+/// evaluating unions. This is used as a pre-pass to HU in order to resolve
+/// first order pointer dereferences and speed up/reduce memory usage of HU.
+/// Running both is equivalent to HRU without the iteration
+/// HVN in more detail:
+/// Imagine the set of constraints was simply straight line code with no loops
+/// (we eliminate cycles, so there are no loops), such as:
+/// E = &D
+/// E = &C
+/// E = F
+/// F = G
+/// G = F
+/// Applying value numbering to this code tells us:
+/// G == F == E
+///
+/// For HVN, this is as far as it goes. We assign new value numbers to every
+/// "address node", and every "reference node".
+/// To get the optimal result for this, we use a DFS + SCC (since all nodes in a
+/// cycle must have the same value number since the = operation is really
+/// inclusion, not overwrite), and value number nodes we receive points-to sets
+/// before we value our own node.
+/// The advantage of HU over HVN is that HU considers the inclusion property, so
+/// that if you have
+/// E = &D
+/// E = &C
+/// E = F
+/// F = G
+/// F = &D
+/// G = F
+/// HU will determine that G == F == E. HVN will not, because it cannot prove
+/// that the points to information ends up being the same because they all
+/// receive &D from E anyway.
+
+void Andersens::HVN() {
+ DEBUG(dbgs() << "Beginning HVN\n");
+ // Build a predecessor graph. This is like our constraint graph with the
+ // edges going in the opposite direction, and there are edges for all the
+ // constraints, instead of just copy constraints. We also build implicit
+ // edges for constraints are implied but not explicit. I.E for the constraint
+ // a = &b, we add implicit edges *a = b. This helps us capture more cycles
+ for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
+ Constraint &C = Constraints[i];
+ if (C.Type == Constraint::AddressOf) {
+ GraphNodes[C.Src].AddressTaken = true;
+ GraphNodes[C.Src].Direct = false;
+
+ // Dest = &src edge
+ unsigned AdrNode = C.Src + FirstAdrNode;
+ if (!GraphNodes[C.Dest].PredEdges)
+ GraphNodes[C.Dest].PredEdges = new SparseBitVector<>;
+ GraphNodes[C.Dest].PredEdges->set(AdrNode);
+
+ // *Dest = src edge
+ unsigned RefNode = C.Dest + FirstRefNode;
+ if (!GraphNodes[RefNode].ImplicitPredEdges)
+ GraphNodes[RefNode].ImplicitPredEdges = new SparseBitVector<>;
+ GraphNodes[RefNode].ImplicitPredEdges->set(C.Src);
+ } else if (C.Type == Constraint::Load) {
+ if (C.Offset == 0) {
+ // dest = *src edge
+ if (!GraphNodes[C.Dest].PredEdges)
+ GraphNodes[C.Dest].PredEdges = new SparseBitVector<>;
+ GraphNodes[C.Dest].PredEdges->set(C.Src + FirstRefNode);
+ } else {
+ GraphNodes[C.Dest].Direct = false;
+ }
+ } else if (C.Type == Constraint::Store) {
+ if (C.Offset == 0) {
+ // *dest = src edge
+ unsigned RefNode = C.Dest + FirstRefNode;
+ if (!GraphNodes[RefNode].PredEdges)
+ GraphNodes[RefNode].PredEdges = new SparseBitVector<>;
+ GraphNodes[RefNode].PredEdges->set(C.Src);
+ }
+ } else {
+ // Dest = Src edge and *Dest = *Src edge
+ if (!GraphNodes[C.Dest].PredEdges)
+ GraphNodes[C.Dest].PredEdges = new SparseBitVector<>;
+ GraphNodes[C.Dest].PredEdges->set(C.Src);
+ unsigned RefNode = C.Dest + FirstRefNode;
+ if (!GraphNodes[RefNode].ImplicitPredEdges)
+ GraphNodes[RefNode].ImplicitPredEdges = new SparseBitVector<>;
+ GraphNodes[RefNode].ImplicitPredEdges->set(C.Src + FirstRefNode);
+ }
+ }
+ PEClass = 1;
+ // Do SCC finding first to condense our predecessor graph
+ DFSNumber = 0;
+ Node2DFS.insert(Node2DFS.begin(), GraphNodes.size(), 0);
+ Node2Deleted.insert(Node2Deleted.begin(), GraphNodes.size(), false);
+ Node2Visited.insert(Node2Visited.begin(), GraphNodes.size(), false);
+
+ for (unsigned i = 0; i < FirstRefNode; ++i) {
+ unsigned Node = VSSCCRep[i];
+ if (!Node2Visited[Node])
+ HVNValNum(Node);
+ }
+ for (BitVectorMap::iterator Iter = Set2PEClass.begin();
+ Iter != Set2PEClass.end();
+ ++Iter)
+ delete Iter->first;
+ Set2PEClass.clear();
+ Node2DFS.clear();
+ Node2Deleted.clear();
+ Node2Visited.clear();
+ DEBUG(dbgs() << "Finished HVN\n");
+
+}
+
+/// This is the workhorse of HVN value numbering. We combine SCC finding at the
+/// same time because it's easy.
+void Andersens::HVNValNum(unsigned NodeIndex) {
+ unsigned MyDFS = DFSNumber++;
+ Node *N = &GraphNodes[NodeIndex];
+ Node2Visited[NodeIndex] = true;
+ Node2DFS[NodeIndex] = MyDFS;
+
+ // First process all our explicit edges
+ if (N->PredEdges)
+ for (SparseBitVector<>::iterator Iter = N->PredEdges->begin();
+ Iter != N->PredEdges->end();
+ ++Iter) {
+ unsigned j = VSSCCRep[*Iter];
+ if (!Node2Deleted[j]) {
+ if (!Node2Visited[j])
+ HVNValNum(j);
+ if (Node2DFS[NodeIndex] > Node2DFS[j])
+ Node2DFS[NodeIndex] = Node2DFS[j];
+ }
+ }
+
+ // Now process all the implicit edges
+ if (N->ImplicitPredEdges)
+ for (SparseBitVector<>::iterator Iter = N->ImplicitPredEdges->begin();
+ Iter != N->ImplicitPredEdges->end();
+ ++Iter) {
+ unsigned j = VSSCCRep[*Iter];
+ if (!Node2Deleted[j]) {
+ if (!Node2Visited[j])
+ HVNValNum(j);
+ if (Node2DFS[NodeIndex] > Node2DFS[j])
+ Node2DFS[NodeIndex] = Node2DFS[j];
+ }
+ }
+
+ // See if we found any cycles
+ if (MyDFS == Node2DFS[NodeIndex]) {
+ while (!SCCStack.empty() && Node2DFS[SCCStack.top()] >= MyDFS) {
+ unsigned CycleNodeIndex = SCCStack.top();
+ Node *CycleNode = &GraphNodes[CycleNodeIndex];
+ VSSCCRep[CycleNodeIndex] = NodeIndex;
+ // Unify the nodes
+ N->Direct &= CycleNode->Direct;
+
+ if (CycleNode->PredEdges) {
+ if (!N->PredEdges)
+ N->PredEdges = new SparseBitVector<>;
+ *(N->PredEdges) |= CycleNode->PredEdges;
+ delete CycleNode->PredEdges;
+ CycleNode->PredEdges = NULL;
+ }
+ if (CycleNode->ImplicitPredEdges) {
+ if (!N->ImplicitPredEdges)
+ N->ImplicitPredEdges = new SparseBitVector<>;
+ *(N->ImplicitPredEdges) |= CycleNode->ImplicitPredEdges;
+ delete CycleNode->ImplicitPredEdges;
+ CycleNode->ImplicitPredEdges = NULL;
+ }
+
+ SCCStack.pop();
+ }
+
+ Node2Deleted[NodeIndex] = true;
+
+ if (!N->Direct) {
+ GraphNodes[NodeIndex].PointerEquivLabel = PEClass++;
+ return;
+ }
+
+ // Collect labels of successor nodes
+ bool AllSame = true;
+ unsigned First = ~0;
+ SparseBitVector<> *Labels = new SparseBitVector<>;
+ bool Used = false;
+
+ if (N->PredEdges)
+ for (SparseBitVector<>::iterator Iter = N->PredEdges->begin();
+ Iter != N->PredEdges->end();
+ ++Iter) {
+ unsigned j = VSSCCRep[*Iter];
+ unsigned Label = GraphNodes[j].PointerEquivLabel;
+ // Ignore labels that are equal to us or non-pointers
+ if (j == NodeIndex || Label == 0)
+ continue;
+ if (First == (unsigned)~0)
+ First = Label;
+ else if (First != Label)
+ AllSame = false;
+ Labels->set(Label);
+ }
+
+ // We either have a non-pointer, a copy of an existing node, or a new node.
+ // Assign the appropriate pointer equivalence label.
+ if (Labels->empty()) {
+ GraphNodes[NodeIndex].PointerEquivLabel = 0;
+ } else if (AllSame) {
+ GraphNodes[NodeIndex].PointerEquivLabel = First;
+ } else {
+ GraphNodes[NodeIndex].PointerEquivLabel = Set2PEClass[Labels];
+ if (GraphNodes[NodeIndex].PointerEquivLabel == 0) {
+ unsigned EquivClass = PEClass++;
+ Set2PEClass[Labels] = EquivClass;
+ GraphNodes[NodeIndex].PointerEquivLabel = EquivClass;
+ Used = true;
+ }
+ }
+ if (!Used)
+ delete Labels;
+ } else {
+ SCCStack.push(NodeIndex);
+ }
+}
+
+/// The technique used here is described in "Exploiting Pointer and Location
+/// Equivalence to Optimize Pointer Analysis. In the 14th International Static
+/// Analysis Symposium (SAS), August 2007." It is known as the "HU" algorithm,
+/// and is equivalent to value numbering the collapsed constraint graph
+/// including evaluating unions.
+void Andersens::HU() {
+ DEBUG(dbgs() << "Beginning HU\n");
+ // Build a predecessor graph. This is like our constraint graph with the
+ // edges going in the opposite direction, and there are edges for all the
+ // constraints, instead of just copy constraints. We also build implicit
+ // edges for constraints are implied but not explicit. I.E for the constraint
+ // a = &b, we add implicit edges *a = b. This helps us capture more cycles
+ for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
+ Constraint &C = Constraints[i];
+ if (C.Type == Constraint::AddressOf) {
+ GraphNodes[C.Src].AddressTaken = true;
+ GraphNodes[C.Src].Direct = false;
+
+ GraphNodes[C.Dest].PointsTo->set(C.Src);
+ // *Dest = src edge
+ unsigned RefNode = C.Dest + FirstRefNode;
+ if (!GraphNodes[RefNode].ImplicitPredEdges)
+ GraphNodes[RefNode].ImplicitPredEdges = new SparseBitVector<>;
+ GraphNodes[RefNode].ImplicitPredEdges->set(C.Src);
+ GraphNodes[C.Src].PointedToBy->set(C.Dest);
+ } else if (C.Type == Constraint::Load) {
+ if (C.Offset == 0) {
+ // dest = *src edge
+ if (!GraphNodes[C.Dest].PredEdges)
+ GraphNodes[C.Dest].PredEdges = new SparseBitVector<>;
+ GraphNodes[C.Dest].PredEdges->set(C.Src + FirstRefNode);
+ } else {
+ GraphNodes[C.Dest].Direct = false;
+ }
+ } else if (C.Type == Constraint::Store) {
+ if (C.Offset == 0) {
+ // *dest = src edge
+ unsigned RefNode = C.Dest + FirstRefNode;
+ if (!GraphNodes[RefNode].PredEdges)
+ GraphNodes[RefNode].PredEdges = new SparseBitVector<>;
+ GraphNodes[RefNode].PredEdges->set(C.Src);
+ }
+ } else {
+ // Dest = Src edge and *Dest = *Src edg
+ if (!GraphNodes[C.Dest].PredEdges)
+ GraphNodes[C.Dest].PredEdges = new SparseBitVector<>;
+ GraphNodes[C.Dest].PredEdges->set(C.Src);
+ unsigned RefNode = C.Dest + FirstRefNode;
+ if (!GraphNodes[RefNode].ImplicitPredEdges)
+ GraphNodes[RefNode].ImplicitPredEdges = new SparseBitVector<>;
+ GraphNodes[RefNode].ImplicitPredEdges->set(C.Src + FirstRefNode);
+ }
+ }
+ PEClass = 1;
+ // Do SCC finding first to condense our predecessor graph
+ DFSNumber = 0;
+ Node2DFS.insert(Node2DFS.begin(), GraphNodes.size(), 0);
+ Node2Deleted.insert(Node2Deleted.begin(), GraphNodes.size(), false);
+ Node2Visited.insert(Node2Visited.begin(), GraphNodes.size(), false);
+
+ for (unsigned i = 0; i < FirstRefNode; ++i) {
+ if (FindNode(i) == i) {
+ unsigned Node = VSSCCRep[i];
+ if (!Node2Visited[Node])
+ Condense(Node);
+ }
+ }
+
+ // Reset tables for actual labeling
+ Node2DFS.clear();
+ Node2Visited.clear();
+ Node2Deleted.clear();
+ // Pre-grow our densemap so that we don't get really bad behavior
+ Set2PEClass.resize(GraphNodes.size());
+
+ // Visit the condensed graph and generate pointer equivalence labels.
+ Node2Visited.insert(Node2Visited.begin(), GraphNodes.size(), false);
+ for (unsigned i = 0; i < FirstRefNode; ++i) {
+ if (FindNode(i) == i) {
+ unsigned Node = VSSCCRep[i];
+ if (!Node2Visited[Node])
+ HUValNum(Node);
+ }
+ }
+ // PEClass nodes will be deleted by the deleting of N->PointsTo in our caller.
+ Set2PEClass.clear();
+ DEBUG(dbgs() << "Finished HU\n");
+}
+
+
+/// Implementation of standard Tarjan SCC algorithm as modified by Nuutilla.
+void Andersens::Condense(unsigned NodeIndex) {
+ unsigned MyDFS = DFSNumber++;
+ Node *N = &GraphNodes[NodeIndex];
+ Node2Visited[NodeIndex] = true;
+ Node2DFS[NodeIndex] = MyDFS;
+
+ // First process all our explicit edges
+ if (N->PredEdges)
+ for (SparseBitVector<>::iterator Iter = N->PredEdges->begin();
+ Iter != N->PredEdges->end();
+ ++Iter) {
+ unsigned j = VSSCCRep[*Iter];
+ if (!Node2Deleted[j]) {
+ if (!Node2Visited[j])
+ Condense(j);
+ if (Node2DFS[NodeIndex] > Node2DFS[j])
+ Node2DFS[NodeIndex] = Node2DFS[j];
+ }
+ }
+
+ // Now process all the implicit edges
+ if (N->ImplicitPredEdges)
+ for (SparseBitVector<>::iterator Iter = N->ImplicitPredEdges->begin();
+ Iter != N->ImplicitPredEdges->end();
+ ++Iter) {
+ unsigned j = VSSCCRep[*Iter];
+ if (!Node2Deleted[j]) {
+ if (!Node2Visited[j])
+ Condense(j);
+ if (Node2DFS[NodeIndex] > Node2DFS[j])
+ Node2DFS[NodeIndex] = Node2DFS[j];
+ }
+ }
+
+ // See if we found any cycles
+ if (MyDFS == Node2DFS[NodeIndex]) {
+ while (!SCCStack.empty() && Node2DFS[SCCStack.top()] >= MyDFS) {
+ unsigned CycleNodeIndex = SCCStack.top();
+ Node *CycleNode = &GraphNodes[CycleNodeIndex];
+ VSSCCRep[CycleNodeIndex] = NodeIndex;
+ // Unify the nodes
+ N->Direct &= CycleNode->Direct;
+
+ *(N->PointsTo) |= CycleNode->PointsTo;
+ delete CycleNode->PointsTo;
+ CycleNode->PointsTo = NULL;
+ if (CycleNode->PredEdges) {
+ if (!N->PredEdges)
+ N->PredEdges = new SparseBitVector<>;
+ *(N->PredEdges) |= CycleNode->PredEdges;
+ delete CycleNode->PredEdges;
+ CycleNode->PredEdges = NULL;
+ }
+ if (CycleNode->ImplicitPredEdges) {
+ if (!N->ImplicitPredEdges)
+ N->ImplicitPredEdges = new SparseBitVector<>;
+ *(N->ImplicitPredEdges) |= CycleNode->ImplicitPredEdges;
+ delete CycleNode->ImplicitPredEdges;
+ CycleNode->ImplicitPredEdges = NULL;
+ }
+ SCCStack.pop();
+ }
+
+ Node2Deleted[NodeIndex] = true;
+
+ // Set up number of incoming edges for other nodes
+ if (N->PredEdges)
+ for (SparseBitVector<>::iterator Iter = N->PredEdges->begin();
+ Iter != N->PredEdges->end();
+ ++Iter)
+ ++GraphNodes[VSSCCRep[*Iter]].NumInEdges;
+ } else {
+ SCCStack.push(NodeIndex);
+ }
+}
+
+void Andersens::HUValNum(unsigned NodeIndex) {
+ Node *N = &GraphNodes[NodeIndex];
+ Node2Visited[NodeIndex] = true;
+
+ // Eliminate dereferences of non-pointers for those non-pointers we have
+ // already identified. These are ref nodes whose non-ref node:
+ // 1. Has already been visited determined to point to nothing (and thus, a
+ // dereference of it must point to nothing)
+ // 2. Any direct node with no predecessor edges in our graph and with no
+ // points-to set (since it can't point to anything either, being that it
+ // receives no points-to sets and has none).
+ if (NodeIndex >= FirstRefNode) {
+ unsigned j = VSSCCRep[FindNode(NodeIndex - FirstRefNode)];
+ if ((Node2Visited[j] && !GraphNodes[j].PointerEquivLabel)
+ || (GraphNodes[j].Direct && !GraphNodes[j].PredEdges
+ && GraphNodes[j].PointsTo->empty())){
+ return;
+ }
+ }
+ // Process all our explicit edges
+ if (N->PredEdges)
+ for (SparseBitVector<>::iterator Iter = N->PredEdges->begin();
+ Iter != N->PredEdges->end();
+ ++Iter) {
+ unsigned j = VSSCCRep[*Iter];
+ if (!Node2Visited[j])
+ HUValNum(j);
+
+ // If this edge turned out to be the same as us, or got no pointer
+ // equivalence label (and thus points to nothing) , just decrement our
+ // incoming edges and continue.
+ if (j == NodeIndex || GraphNodes[j].PointerEquivLabel == 0) {
+ --GraphNodes[j].NumInEdges;
+ continue;
+ }
+
+ *(N->PointsTo) |= GraphNodes[j].PointsTo;
+
+ // If we didn't end up storing this in the hash, and we're done with all
+ // the edges, we don't need the points-to set anymore.
+ --GraphNodes[j].NumInEdges;
+ if (!GraphNodes[j].NumInEdges && !GraphNodes[j].StoredInHash) {
+ delete GraphNodes[j].PointsTo;
+ GraphNodes[j].PointsTo = NULL;
+ }
+ }
+ // If this isn't a direct node, generate a fresh variable.
+ if (!N->Direct) {
+ N->PointsTo->set(FirstRefNode + NodeIndex);
+ }
+
+ // See If we have something equivalent to us, if not, generate a new
+ // equivalence class.
+ if (N->PointsTo->empty()) {
+ delete N->PointsTo;
+ N->PointsTo = NULL;
+ } else {
+ if (N->Direct) {
+ N->PointerEquivLabel = Set2PEClass[N->PointsTo];
+ if (N->PointerEquivLabel == 0) {
+ unsigned EquivClass = PEClass++;
+ N->StoredInHash = true;
+ Set2PEClass[N->PointsTo] = EquivClass;
+ N->PointerEquivLabel = EquivClass;
+ }
+ } else {
+ N->PointerEquivLabel = PEClass++;
+ }
+ }
+}
+
+/// Rewrite our list of constraints so that pointer equivalent nodes are
+/// replaced by their the pointer equivalence class representative.
+void Andersens::RewriteConstraints() {
+ std::vector<Constraint> NewConstraints;
+ DenseSet<Constraint, ConstraintKeyInfo> Seen;
+
+ PEClass2Node.clear();
+ PENLEClass2Node.clear();
+
+ // We may have from 1 to Graphnodes + 1 equivalence classes.
+ PEClass2Node.insert(PEClass2Node.begin(), GraphNodes.size() + 1, -1);
+ PENLEClass2Node.insert(PENLEClass2Node.begin(), GraphNodes.size() + 1, -1);
+
+ // Rewrite constraints, ignoring non-pointer constraints, uniting equivalent
+ // nodes, and rewriting constraints to use the representative nodes.
+ for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
+ Constraint &C = Constraints[i];
+ unsigned RHSNode = FindNode(C.Src);
+ unsigned LHSNode = FindNode(C.Dest);
+ unsigned RHSLabel = GraphNodes[VSSCCRep[RHSNode]].PointerEquivLabel;
+ unsigned LHSLabel = GraphNodes[VSSCCRep[LHSNode]].PointerEquivLabel;
+
+ // First we try to eliminate constraints for things we can prove don't point
+ // to anything.
+ if (LHSLabel == 0) {
+ DEBUG(PrintNode(&GraphNodes[LHSNode]));
+ DEBUG(dbgs() << " is a non-pointer, ignoring constraint.\n");
+ continue;
+ }
+ if (RHSLabel == 0) {
+ DEBUG(PrintNode(&GraphNodes[RHSNode]));
+ DEBUG(dbgs() << " is a non-pointer, ignoring constraint.\n");
+ continue;
+ }
+ // This constraint may be useless, and it may become useless as we translate
+ // it.
+ if (C.Src == C.Dest && C.Type == Constraint::Copy)
+ continue;
+
+ C.Src = FindEquivalentNode(RHSNode, RHSLabel);
+ C.Dest = FindEquivalentNode(FindNode(LHSNode), LHSLabel);
+ if ((C.Src == C.Dest && C.Type == Constraint::Copy)
+ || Seen.count(C))
+ continue;
+
+ Seen.insert(C);
+ NewConstraints.push_back(C);
+ }
+ Constraints.swap(NewConstraints);
+ PEClass2Node.clear();
+}
+
+/// See if we have a node that is pointer equivalent to the one being asked
+/// about, and if so, unite them and return the equivalent node. Otherwise,
+/// return the original node.
+unsigned Andersens::FindEquivalentNode(unsigned NodeIndex,
+ unsigned NodeLabel) {
+ if (!GraphNodes[NodeIndex].AddressTaken) {
+ if (PEClass2Node[NodeLabel] != -1) {
+ // We found an existing node with the same pointer label, so unify them.
+ // We specifically request that Union-By-Rank not be used so that
+ // PEClass2Node[NodeLabel] U= NodeIndex and not the other way around.
+ return UniteNodes(PEClass2Node[NodeLabel], NodeIndex, false);
+ } else {
+ PEClass2Node[NodeLabel] = NodeIndex;
+ PENLEClass2Node[NodeLabel] = NodeIndex;
+ }
+ } else if (PENLEClass2Node[NodeLabel] == -1) {
+ PENLEClass2Node[NodeLabel] = NodeIndex;
+ }
+
+ return NodeIndex;
+}
+
+void Andersens::PrintLabels() const {
+ for (unsigned i = 0; i < GraphNodes.size(); ++i) {
+ if (i < FirstRefNode) {
+ PrintNode(&GraphNodes[i]);
+ } else if (i < FirstAdrNode) {
+ DEBUG(dbgs() << "REF(");
+ PrintNode(&GraphNodes[i-FirstRefNode]);
+ DEBUG(dbgs() <<")");
+ } else {
+ DEBUG(dbgs() << "ADR(");
+ PrintNode(&GraphNodes[i-FirstAdrNode]);
+ DEBUG(dbgs() <<")");
+ }
+
+ DEBUG(dbgs() << " has pointer label " << GraphNodes[i].PointerEquivLabel
+ << " and SCC rep " << VSSCCRep[i]
+ << " and is " << (GraphNodes[i].Direct ? "Direct" : "Not direct")
+ << "\n");
+ }
+}
+
+/// The technique used here is described in "The Ant and the
+/// Grasshopper: Fast and Accurate Pointer Analysis for Millions of
+/// Lines of Code. In Programming Language Design and Implementation
+/// (PLDI), June 2007." It is known as the "HCD" (Hybrid Cycle
+/// Detection) algorithm. It is called a hybrid because it performs an
+/// offline analysis and uses its results during the solving (online)
+/// phase. This is just the offline portion; the results of this
+/// operation are stored in SDT and are later used in SolveContraints()
+/// and UniteNodes().
+void Andersens::HCD() {
+ DEBUG(dbgs() << "Starting HCD.\n");
+ HCDSCCRep.resize(GraphNodes.size());
+
+ for (unsigned i = 0; i < GraphNodes.size(); ++i) {
+ GraphNodes[i].Edges = new SparseBitVector<>;
+ HCDSCCRep[i] = i;
+ }
+
+ for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
+ Constraint &C = Constraints[i];
+ assert (C.Src < GraphNodes.size() && C.Dest < GraphNodes.size());
+ if (C.Type == Constraint::AddressOf) {
+ continue;
+ } else if (C.Type == Constraint::Load) {
+ if( C.Offset == 0 )
+ GraphNodes[C.Dest].Edges->set(C.Src + FirstRefNode);
+ } else if (C.Type == Constraint::Store) {
+ if( C.Offset == 0 )
+ GraphNodes[C.Dest + FirstRefNode].Edges->set(C.Src);
+ } else {
+ GraphNodes[C.Dest].Edges->set(C.Src);
+ }
+ }
+
+ Node2DFS.insert(Node2DFS.begin(), GraphNodes.size(), 0);
+ Node2Deleted.insert(Node2Deleted.begin(), GraphNodes.size(), false);
+ Node2Visited.insert(Node2Visited.begin(), GraphNodes.size(), false);
+ SDT.insert(SDT.begin(), GraphNodes.size() / 2, -1);
+
+ DFSNumber = 0;
+ for (unsigned i = 0; i < GraphNodes.size(); ++i) {
+ unsigned Node = HCDSCCRep[i];
+ if (!Node2Deleted[Node])
+ Search(Node);
+ }
+
+ for (unsigned i = 0; i < GraphNodes.size(); ++i)
+ if (GraphNodes[i].Edges != NULL) {
+ delete GraphNodes[i].Edges;
+ GraphNodes[i].Edges = NULL;
+ }
+
+ while( !SCCStack.empty() )
+ SCCStack.pop();
+
+ Node2DFS.clear();
+ Node2Visited.clear();
+ Node2Deleted.clear();
+ HCDSCCRep.clear();
+ DEBUG(dbgs() << "HCD complete.\n");
+}
+
+// Component of HCD:
+// Use Nuutila's variant of Tarjan's algorithm to detect
+// Strongly-Connected Components (SCCs). For non-trivial SCCs
+// containing ref nodes, insert the appropriate information in SDT.
+void Andersens::Search(unsigned Node) {
+ unsigned MyDFS = DFSNumber++;
+
+ Node2Visited[Node] = true;
+ Node2DFS[Node] = MyDFS;
+
+ for (SparseBitVector<>::iterator Iter = GraphNodes[Node].Edges->begin(),
+ End = GraphNodes[Node].Edges->end();
+ Iter != End;
+ ++Iter) {
+ unsigned J = HCDSCCRep[*Iter];
+ assert(GraphNodes[J].isRep() && "Debug check; must be representative");
+ if (!Node2Deleted[J]) {
+ if (!Node2Visited[J])
+ Search(J);
+ if (Node2DFS[Node] > Node2DFS[J])
+ Node2DFS[Node] = Node2DFS[J];
+ }
+ }
+
+ if( MyDFS != Node2DFS[Node] ) {
+ SCCStack.push(Node);
+ return;
+ }
+
+ // This node is the root of a SCC, so process it.
+ //
+ // If the SCC is "non-trivial" (not a singleton) and contains a reference
+ // node, we place this SCC into SDT. We unite the nodes in any case.
+ if (!SCCStack.empty() && Node2DFS[SCCStack.top()] >= MyDFS) {
+ SparseBitVector<> SCC;
+
+ SCC.set(Node);
+
+ bool Ref = (Node >= FirstRefNode);
+
+ Node2Deleted[Node] = true;
+
+ do {
+ unsigned P = SCCStack.top(); SCCStack.pop();
+ Ref |= (P >= FirstRefNode);
+ SCC.set(P);
+ HCDSCCRep[P] = Node;
+ } while (!SCCStack.empty() && Node2DFS[SCCStack.top()] >= MyDFS);
+
+ if (Ref) {
+ unsigned Rep = SCC.find_first();
+ assert(Rep < FirstRefNode && "The SCC didn't have a non-Ref node!");
+
+ SparseBitVector<>::iterator i = SCC.begin();
+
+ // Skip over the non-ref nodes
+ while( *i < FirstRefNode )
+ ++i;
+
+ while( i != SCC.end() )
+ SDT[ (*i++) - FirstRefNode ] = Rep;
+ }
+ }
+}
+
+
+/// Optimize the constraints by performing offline variable substitution and
+/// other optimizations.
+void Andersens::OptimizeConstraints() {
+ DEBUG(dbgs() << "Beginning constraint optimization\n");
+
+ SDTActive = false;
+
+ // Function related nodes need to stay in the same relative position and can't
+ // be location equivalent.
+ for (std::map<unsigned, unsigned>::iterator Iter = MaxK.begin();
+ Iter != MaxK.end();
+ ++Iter) {
+ for (unsigned i = Iter->first;
+ i != Iter->first + Iter->second;
+ ++i) {
+ GraphNodes[i].AddressTaken = true;
+ GraphNodes[i].Direct = false;
+ }
+ }
+
+ ClumpAddressTaken();
+ FirstRefNode = GraphNodes.size();
+ FirstAdrNode = FirstRefNode + GraphNodes.size();
+ GraphNodes.insert(GraphNodes.end(), 2 * GraphNodes.size(),
+ Node(false));
+ VSSCCRep.resize(GraphNodes.size());
+ for (unsigned i = 0; i < GraphNodes.size(); ++i) {
+ VSSCCRep[i] = i;
+ }
+ HVN();
+ for (unsigned i = 0; i < GraphNodes.size(); ++i) {
+ Node *N = &GraphNodes[i];
+ delete N->PredEdges;
+ N->PredEdges = NULL;
+ delete N->ImplicitPredEdges;
+ N->ImplicitPredEdges = NULL;
+ }
+#undef DEBUG_TYPE
+#define DEBUG_TYPE "anders-aa-labels"
+ DEBUG(PrintLabels());
+#undef DEBUG_TYPE
+#define DEBUG_TYPE "anders-aa"
+ RewriteConstraints();
+ // Delete the adr nodes.
+ GraphNodes.resize(FirstRefNode * 2);
+
+ // Now perform HU
+ for (unsigned i = 0; i < GraphNodes.size(); ++i) {
+ Node *N = &GraphNodes[i];
+ if (FindNode(i) == i) {
+ N->PointsTo = new SparseBitVector<>;
+ N->PointedToBy = new SparseBitVector<>;
+ // Reset our labels
+ }
+ VSSCCRep[i] = i;
+ N->PointerEquivLabel = 0;
+ }
+ HU();
+#undef DEBUG_TYPE
+#define DEBUG_TYPE "anders-aa-labels"
+ DEBUG(PrintLabels());
+#undef DEBUG_TYPE
+#define DEBUG_TYPE "anders-aa"
+ RewriteConstraints();
+ for (unsigned i = 0; i < GraphNodes.size(); ++i) {
+ if (FindNode(i) == i) {
+ Node *N = &GraphNodes[i];
+ delete N->PointsTo;
+ N->PointsTo = NULL;
+ delete N->PredEdges;
+ N->PredEdges = NULL;
+ delete N->ImplicitPredEdges;
+ N->ImplicitPredEdges = NULL;
+ delete N->PointedToBy;
+ N->PointedToBy = NULL;
+ }
+ }
+
+ // perform Hybrid Cycle Detection (HCD)
+ HCD();
+ SDTActive = true;
+
+ // No longer any need for the upper half of GraphNodes (for ref nodes).
+ GraphNodes.erase(GraphNodes.begin() + FirstRefNode, GraphNodes.end());
+
+ // HCD complete.
+
+ DEBUG(dbgs() << "Finished constraint optimization\n");
+ FirstRefNode = 0;
+ FirstAdrNode = 0;
+}
+
+/// Unite pointer but not location equivalent variables, now that the constraint
+/// graph is built.
+void Andersens::UnitePointerEquivalences() {
+ DEBUG(dbgs() << "Uniting remaining pointer equivalences\n");
+ for (unsigned i = 0; i < GraphNodes.size(); ++i) {
+ if (GraphNodes[i].AddressTaken && GraphNodes[i].isRep()) {
+ unsigned Label = GraphNodes[i].PointerEquivLabel;
+
+ if (Label && PENLEClass2Node[Label] != -1)
+ UniteNodes(i, PENLEClass2Node[Label]);
+ }
+ }
+ DEBUG(dbgs() << "Finished remaining pointer equivalences\n");
+ PENLEClass2Node.clear();
+}
+
+/// Create the constraint graph used for solving points-to analysis.
+///
+void Andersens::CreateConstraintGraph() {
+ for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
+ Constraint &C = Constraints[i];
+ assert (C.Src < GraphNodes.size() && C.Dest < GraphNodes.size());
+ if (C.Type == Constraint::AddressOf)
+ GraphNodes[C.Dest].PointsTo->set(C.Src);
+ else if (C.Type == Constraint::Load)
+ GraphNodes[C.Src].Constraints.push_back(C);
+ else if (C.Type == Constraint::Store)
+ GraphNodes[C.Dest].Constraints.push_back(C);
+ else if (C.Offset != 0)
+ GraphNodes[C.Src].Constraints.push_back(C);
+ else
+ GraphNodes[C.Src].Edges->set(C.Dest);
+ }
+}
+
+// Perform DFS and cycle detection.
+bool Andersens::QueryNode(unsigned Node) {
+ assert(GraphNodes[Node].isRep() && "Querying a non-rep node");
+ unsigned OurDFS = ++DFSNumber;
+ SparseBitVector<> ToErase;
+ SparseBitVector<> NewEdges;
+ Tarjan2DFS[Node] = OurDFS;
+
+ // Changed denotes a change from a recursive call that we will bubble up.
+ // Merged is set if we actually merge a node ourselves.
+ bool Changed = false, Merged = false;
+
+ for (SparseBitVector<>::iterator bi = GraphNodes[Node].Edges->begin();
+ bi != GraphNodes[Node].Edges->end();
+ ++bi) {
+ unsigned RepNode = FindNode(*bi);
+ // If this edge points to a non-representative node but we are
+ // already planning to add an edge to its representative, we have no
+ // need for this edge anymore.
+ if (RepNode != *bi && NewEdges.test(RepNode)){
+ ToErase.set(*bi);
+ continue;
+ }
+
+ // Continue about our DFS.
+ if (!Tarjan2Deleted[RepNode]){
+ if (Tarjan2DFS[RepNode] == 0) {
+ Changed |= QueryNode(RepNode);
+ // May have been changed by QueryNode
+ RepNode = FindNode(RepNode);
+ }
+ if (Tarjan2DFS[RepNode] < Tarjan2DFS[Node])
+ Tarjan2DFS[Node] = Tarjan2DFS[RepNode];
+ }
+
+ // We may have just discovered that this node is part of a cycle, in
+ // which case we can also erase it.
+ if (RepNode != *bi) {
+ ToErase.set(*bi);
+ NewEdges.set(RepNode);
+ }
+ }
+
+ GraphNodes[Node].Edges->intersectWithComplement(ToErase);
+ GraphNodes[Node].Edges |= NewEdges;
+
+ // If this node is a root of a non-trivial SCC, place it on our
+ // worklist to be processed.
+ if (OurDFS == Tarjan2DFS[Node]) {
+ while (!SCCStack.empty() && Tarjan2DFS[SCCStack.top()] >= OurDFS) {
+ Node = UniteNodes(Node, SCCStack.top());
+
+ SCCStack.pop();
+ Merged = true;
+ }
+ Tarjan2Deleted[Node] = true;
+
+ if (Merged)
+ NextWL->insert(&GraphNodes[Node]);
+ } else {
+ SCCStack.push(Node);
+ }
+
+ return(Changed | Merged);
+}
+
+/// SolveConstraints - This stage iteratively processes the constraints list
+/// propagating constraints (adding edges to the Nodes in the points-to graph)
+/// until a fixed point is reached.
+///
+/// We use a variant of the technique called "Lazy Cycle Detection", which is
+/// described in "The Ant and the Grasshopper: Fast and Accurate Pointer
+/// Analysis for Millions of Lines of Code. In Programming Language Design and
+/// Implementation (PLDI), June 2007."
+/// The paper describes performing cycle detection one node at a time, which can
+/// be expensive if there are no cycles, but there are long chains of nodes that
+/// it heuristically believes are cycles (because it will DFS from each node
+/// without state from previous nodes).
+/// Instead, we use the heuristic to build a worklist of nodes to check, then
+/// cycle detect them all at the same time to do this more cheaply. This
+/// catches cycles slightly later than the original technique did, but does it
+/// make significantly cheaper.
+
+void Andersens::SolveConstraints() {
+ CurrWL = &w1;
+ NextWL = &w2;
+
+ OptimizeConstraints();
+#undef DEBUG_TYPE
+#define DEBUG_TYPE "anders-aa-constraints"
+ DEBUG(PrintConstraints());
+#undef DEBUG_TYPE
+#define DEBUG_TYPE "anders-aa"
+
+ for (unsigned i = 0; i < GraphNodes.size(); ++i) {
+ Node *N = &GraphNodes[i];
+ N->PointsTo = new SparseBitVector<>;
+ N->OldPointsTo = new SparseBitVector<>;
+ N->Edges = new SparseBitVector<>;
+ }
+ CreateConstraintGraph();
+ UnitePointerEquivalences();
+ assert(SCCStack.empty() && "SCC Stack should be empty by now!");
+ Node2DFS.clear();
+ Node2Deleted.clear();
+ Node2DFS.insert(Node2DFS.begin(), GraphNodes.size(), 0);
+ Node2Deleted.insert(Node2Deleted.begin(), GraphNodes.size(), false);
+ DFSNumber = 0;
+ DenseSet<Constraint, ConstraintKeyInfo> Seen;
+ DenseSet<std::pair<unsigned,unsigned>, PairKeyInfo> EdgesChecked;
+
+ // Order graph and add initial nodes to work list.
+ for (unsigned i = 0; i < GraphNodes.size(); ++i) {
+ Node *INode = &GraphNodes[i];
+
+ // Add to work list if it's a representative and can contribute to the
+ // calculation right now.
+ if (INode->isRep() && !INode->PointsTo->empty()
+ && (!INode->Edges->empty() || !INode->Constraints.empty())) {
+ INode->Stamp();
+ CurrWL->insert(INode);
+ }
+ }
+ std::queue<unsigned int> TarjanWL;
+#if !FULL_UNIVERSAL
+ // "Rep and special variables" - in order for HCD to maintain conservative
+ // results when !FULL_UNIVERSAL, we need to treat the special variables in
+ // the same way that the !FULL_UNIVERSAL tweak does throughout the rest of
+ // the analysis - it's ok to add edges from the special nodes, but never
+ // *to* the special nodes.
+ std::vector<unsigned int> RSV;
+#endif
+ while( !CurrWL->empty() ) {
+ DEBUG(dbgs() << "Starting iteration #" << ++NumIters << "\n");
+
+ Node* CurrNode;
+ unsigned CurrNodeIndex;
+
+ // Actual cycle checking code. We cycle check all of the lazy cycle
+ // candidates from the last iteration in one go.
+ if (!TarjanWL.empty()) {
+ DFSNumber = 0;
+
+ Tarjan2DFS.clear();
+ Tarjan2Deleted.clear();
+ while (!TarjanWL.empty()) {
+ unsigned int ToTarjan = TarjanWL.front();
+ TarjanWL.pop();
+ if (!Tarjan2Deleted[ToTarjan]
+ && GraphNodes[ToTarjan].isRep()
+ && Tarjan2DFS[ToTarjan] == 0)
+ QueryNode(ToTarjan);
+ }
+ }
+
+ // Add to work list if it's a representative and can contribute to the
+ // calculation right now.
+ while( (CurrNode = CurrWL->pop()) != NULL ) {
+ CurrNodeIndex = CurrNode - &GraphNodes[0];
+ CurrNode->Stamp();
+
+
+ // Figure out the changed points to bits
+ SparseBitVector<> CurrPointsTo;
+ CurrPointsTo.intersectWithComplement(CurrNode->PointsTo,
+ CurrNode->OldPointsTo);
+ if (CurrPointsTo.empty())
+ continue;
+
+ *(CurrNode->OldPointsTo) |= CurrPointsTo;
+
+ // Check the offline-computed equivalencies from HCD.
+ bool SCC = false;
+ unsigned Rep;
+
+ if (SDT[CurrNodeIndex] >= 0) {
+ SCC = true;
+ Rep = FindNode(SDT[CurrNodeIndex]);
+
+#if !FULL_UNIVERSAL
+ RSV.clear();
+#endif
+ for (SparseBitVector<>::iterator bi = CurrPointsTo.begin();
+ bi != CurrPointsTo.end(); ++bi) {
+ unsigned Node = FindNode(*bi);
+#if !FULL_UNIVERSAL
+ if (Node < NumberSpecialNodes) {
+ RSV.push_back(Node);
+ continue;
+ }
+#endif
+ Rep = UniteNodes(Rep,Node);
+ }
+#if !FULL_UNIVERSAL
+ RSV.push_back(Rep);
+#endif
+
+ NextWL->insert(&GraphNodes[Rep]);
+
+ if ( ! CurrNode->isRep() )
+ continue;
+ }
+
+ Seen.clear();
+
+ /* Now process the constraints for this node. */
+ for (std::list<Constraint>::iterator li = CurrNode->Constraints.begin();
+ li != CurrNode->Constraints.end(); ) {
+ li->Src = FindNode(li->Src);
+ li->Dest = FindNode(li->Dest);
+
+ // Delete redundant constraints
+ if( Seen.count(*li) ) {
+ std::list<Constraint>::iterator lk = li; li++;
+
+ CurrNode->Constraints.erase(lk);
+ ++NumErased;
+ continue;
+ }
+ Seen.insert(*li);
+
+ // Src and Dest will be the vars we are going to process.
+ // This may look a bit ugly, but what it does is allow us to process
+ // both store and load constraints with the same code.
+ // Load constraints say that every member of our RHS solution has K
+ // added to it, and that variable gets an edge to LHS. We also union
+ // RHS+K's solution into the LHS solution.
+ // Store constraints say that every member of our LHS solution has K
+ // added to it, and that variable gets an edge from RHS. We also union
+ // RHS's solution into the LHS+K solution.
+ unsigned *Src;
+ unsigned *Dest;
+ unsigned K = li->Offset;
+ unsigned CurrMember;
+ if (li->Type == Constraint::Load) {
+ Src = &CurrMember;
+ Dest = &li->Dest;
+ } else if (li->Type == Constraint::Store) {
+ Src = &li->Src;
+ Dest = &CurrMember;
+ } else {
+ // TODO Handle offseted copy constraint
+ li++;
+ continue;
+ }
+
+ // See if we can use Hybrid Cycle Detection (that is, check
+ // if it was a statically detected offline equivalence that
+ // involves pointers; if so, remove the redundant constraints).
+ if( SCC && K == 0 ) {
+#if FULL_UNIVERSAL
+ CurrMember = Rep;
+
+ if (GraphNodes[*Src].Edges->test_and_set(*Dest))
+ if (GraphNodes[*Dest].PointsTo |= *(GraphNodes[*Src].PointsTo))
+ NextWL->insert(&GraphNodes[*Dest]);
+#else
+ for (unsigned i=0; i < RSV.size(); ++i) {
+ CurrMember = RSV[i];
+
+ if (*Dest < NumberSpecialNodes)
+ continue;
+ if (GraphNodes[*Src].Edges->test_and_set(*Dest))
+ if (GraphNodes[*Dest].PointsTo |= *(GraphNodes[*Src].PointsTo))
+ NextWL->insert(&GraphNodes[*Dest]);
+ }
+#endif
+ // since all future elements of the points-to set will be
+ // equivalent to the current ones, the complex constraints
+ // become redundant.
+ //
+ std::list<Constraint>::iterator lk = li; li++;
+#if !FULL_UNIVERSAL
+ // In this case, we can still erase the constraints when the
+ // elements of the points-to sets are referenced by *Dest,
+ // but not when they are referenced by *Src (i.e. for a Load
+ // constraint). This is because if another special variable is
+ // put into the points-to set later, we still need to add the
+ // new edge from that special variable.
+ if( lk->Type != Constraint::Load)
+#endif
+ GraphNodes[CurrNodeIndex].Constraints.erase(lk);
+ } else {
+ const SparseBitVector<> &Solution = CurrPointsTo;
+
+ for (SparseBitVector<>::iterator bi = Solution.begin();
+ bi != Solution.end();
+ ++bi) {
+ CurrMember = *bi;
+
+ // Need to increment the member by K since that is where we are
+ // supposed to copy to/from. Note that in positive weight cycles,
+ // which occur in address taking of fields, K can go past
+ // MaxK[CurrMember] elements, even though that is all it could point
+ // to.
+ if (K > 0 && K > MaxK[CurrMember])
+ continue;
+ else
+ CurrMember = FindNode(CurrMember + K);
+
+ // Add an edge to the graph, so we can just do regular
+ // bitmap ior next time. It may also let us notice a cycle.
+#if !FULL_UNIVERSAL
+ if (*Dest < NumberSpecialNodes)
+ continue;
+#endif
+ if (GraphNodes[*Src].Edges->test_and_set(*Dest))
+ if (GraphNodes[*Dest].PointsTo |= *(GraphNodes[*Src].PointsTo))
+ NextWL->insert(&GraphNodes[*Dest]);
+
+ }
+ li++;
+ }
+ }
+ SparseBitVector<> NewEdges;
+ SparseBitVector<> ToErase;
+
+ // Now all we have left to do is propagate points-to info along the
+ // edges, erasing the redundant edges.
+ for (SparseBitVector<>::iterator bi = CurrNode->Edges->begin();
+ bi != CurrNode->Edges->end();
+ ++bi) {
+
+ unsigned DestVar = *bi;
+ unsigned Rep = FindNode(DestVar);
+
+ // If we ended up with this node as our destination, or we've already
+ // got an edge for the representative, delete the current edge.
+ if (Rep == CurrNodeIndex ||
+ (Rep != DestVar && NewEdges.test(Rep))) {
+ ToErase.set(DestVar);
+ continue;
+ }
+
+ std::pair<unsigned,unsigned> edge(CurrNodeIndex,Rep);
+
+ // This is where we do lazy cycle detection.
+ // If this is a cycle candidate (equal points-to sets and this
+ // particular edge has not been cycle-checked previously), add to the
+ // list to check for cycles on the next iteration.
+ if (!EdgesChecked.count(edge) &&
+ *(GraphNodes[Rep].PointsTo) == *(CurrNode->PointsTo)) {
+ EdgesChecked.insert(edge);
+ TarjanWL.push(Rep);
+ }
+ // Union the points-to sets into the dest
+#if !FULL_UNIVERSAL
+ if (Rep >= NumberSpecialNodes)
+#endif
+ if (GraphNodes[Rep].PointsTo |= CurrPointsTo) {
+ NextWL->insert(&GraphNodes[Rep]);
+ }
+ // If this edge's destination was collapsed, rewrite the edge.
+ if (Rep != DestVar) {
+ ToErase.set(DestVar);
+ NewEdges.set(Rep);
+ }
+ }
+ CurrNode->Edges->intersectWithComplement(ToErase);
+ CurrNode->Edges |= NewEdges;
+ }
+
+ // Switch to other work list.
+ WorkList* t = CurrWL; CurrWL = NextWL; NextWL = t;
+ }
+
+
+ Node2DFS.clear();
+ Node2Deleted.clear();
+ for (unsigned i = 0; i < GraphNodes.size(); ++i) {
+ Node *N = &GraphNodes[i];
+ delete N->OldPointsTo;
+ delete N->Edges;
+ }
+ SDTActive = false;
+ SDT.clear();
+}
+
+//===----------------------------------------------------------------------===//
+// Union-Find
+//===----------------------------------------------------------------------===//
+
+// Unite nodes First and Second, returning the one which is now the
+// representative node. First and Second are indexes into GraphNodes
+unsigned Andersens::UniteNodes(unsigned First, unsigned Second,
+ bool UnionByRank) {
+ assert (First < GraphNodes.size() && Second < GraphNodes.size() &&
+ "Attempting to merge nodes that don't exist");
+
+ Node *FirstNode = &GraphNodes[First];
+ Node *SecondNode = &GraphNodes[Second];
+
+ assert (SecondNode->isRep() && FirstNode->isRep() &&
+ "Trying to unite two non-representative nodes!");
+ if (First == Second)
+ return First;
+
+ if (UnionByRank) {
+ int RankFirst = (int) FirstNode ->NodeRep;
+ int RankSecond = (int) SecondNode->NodeRep;
+
+ // Rank starts at -1 and gets decremented as it increases.
+ // Translation: higher rank, lower NodeRep value, which is always negative.
+ if (RankFirst > RankSecond) {
+ unsigned t = First; First = Second; Second = t;
+ Node* tp = FirstNode; FirstNode = SecondNode; SecondNode = tp;
+ } else if (RankFirst == RankSecond) {
+ FirstNode->NodeRep = (unsigned) (RankFirst - 1);
+ }
+ }
+
+ SecondNode->NodeRep = First;
+#if !FULL_UNIVERSAL
+ if (First >= NumberSpecialNodes)
+#endif
+ if (FirstNode->PointsTo && SecondNode->PointsTo)
+ FirstNode->PointsTo |= *(SecondNode->PointsTo);
+ if (FirstNode->Edges && SecondNode->Edges)
+ FirstNode->Edges |= *(SecondNode->Edges);
+ if (!SecondNode->Constraints.empty())
+ FirstNode->Constraints.splice(FirstNode->Constraints.begin(),
+ SecondNode->Constraints);
+ if (FirstNode->OldPointsTo) {
+ delete FirstNode->OldPointsTo;
+ FirstNode->OldPointsTo = new SparseBitVector<>;
+ }
+
+ // Destroy interesting parts of the merged-from node.
+ delete SecondNode->OldPointsTo;
+ delete SecondNode->Edges;
+ delete SecondNode->PointsTo;
+ SecondNode->Edges = NULL;
+ SecondNode->PointsTo = NULL;
+ SecondNode->OldPointsTo = NULL;
+
+ NumUnified++;
+ DEBUG(dbgs() << "Unified Node ");
+ DEBUG(PrintNode(FirstNode));
+ DEBUG(dbgs() << " and Node ");
+ DEBUG(PrintNode(SecondNode));
+ DEBUG(dbgs() << "\n");
+
+ if (SDTActive)
+ if (SDT[Second] >= 0) {
+ if (SDT[First] < 0)
+ SDT[First] = SDT[Second];
+ else {
+ UniteNodes( FindNode(SDT[First]), FindNode(SDT[Second]) );
+ First = FindNode(First);
+ }
+ }
+
+ return First;
+}
+
+// Find the index into GraphNodes of the node representing Node, performing
+// path compression along the way
+unsigned Andersens::FindNode(unsigned NodeIndex) {
+ assert (NodeIndex < GraphNodes.size()
+ && "Attempting to find a node that can't exist");
+ Node *N = &GraphNodes[NodeIndex];
+ if (N->isRep())
+ return NodeIndex;
+ else
+ return (N->NodeRep = FindNode(N->NodeRep));
+}
+
+// Find the index into GraphNodes of the node representing Node,
+// don't perform path compression along the way (for Print)
+unsigned Andersens::FindNode(unsigned NodeIndex) const {
+ assert (NodeIndex < GraphNodes.size()
+ && "Attempting to find a node that can't exist");
+ const Node *N = &GraphNodes[NodeIndex];
+ if (N->isRep())
+ return NodeIndex;
+ else
+ return FindNode(N->NodeRep);
+}
+
+//===----------------------------------------------------------------------===//
+// Debugging Output
+//===----------------------------------------------------------------------===//
+
+void Andersens::PrintNode(const Node *N) const {
+ if (N == &GraphNodes[UniversalSet]) {
+ dbgs() << "<universal>";
+ return;
+ } else if (N == &GraphNodes[NullPtr]) {
+ dbgs() << "<nullptr>";
+ return;
+ } else if (N == &GraphNodes[NullObject]) {
+ dbgs() << "<null>";
+ return;
+ }
+ if (!N->getValue()) {
+ dbgs() << "artificial" << (intptr_t) N;
+ return;
+ }
+
+ assert(N->getValue() != 0 && "Never set node label!");
+ Value *V = N->getValue();
+ if (Function *F = dyn_cast<Function>(V)) {
+ if (isa<PointerType>(F->getFunctionType()->getReturnType()) &&
+ N == &GraphNodes[getReturnNode(F)]) {
+ dbgs() << F->getName() << ":retval";
+ return;
+ } else if (F->getFunctionType()->isVarArg() &&
+ N == &GraphNodes[getVarargNode(F)]) {
+ dbgs() << F->getName() << ":vararg";
+ return;
+ }
+ }
+
+ if (Instruction *I = dyn_cast<Instruction>(V))
+ dbgs() << I->getParent()->getParent()->getName() << ":";
+ else if (Argument *Arg = dyn_cast<Argument>(V))
+ dbgs() << Arg->getParent()->getName() << ":";
+
+ if (V->hasName())
+ dbgs() << V->getName();
+ else
+ dbgs() << "(unnamed)";
+
+ if (isa<GlobalValue>(V) || isa<AllocaInst>(V) || isMalloc(V))
+ if (N == &GraphNodes[getObject(V)])
+ dbgs() << "<mem>";
+}
+void Andersens::PrintConstraint(const Constraint &C) const {
+ if (C.Type == Constraint::Store) {
+ dbgs() << "*";
+ if (C.Offset != 0)
+ dbgs() << "(";
+ }
+ PrintNode(&GraphNodes[C.Dest]);
+ if (C.Type == Constraint::Store && C.Offset != 0)
+ dbgs() << " + " << C.Offset << ")";
+ dbgs() << " = ";
+ if (C.Type == Constraint::Load) {
+ dbgs() << "*";
+ if (C.Offset != 0)
+ dbgs() << "(";
+ }
+ else if (C.Type == Constraint::AddressOf)
+ dbgs() << "&";
+ PrintNode(&GraphNodes[C.Src]);
+ if (C.Offset != 0 && C.Type != Constraint::Store)
+ dbgs() << " + " << C.Offset;
+ if (C.Type == Constraint::Load && C.Offset != 0)
+ dbgs() << ")";
+ dbgs() << "\n";
+}
+
+void Andersens::PrintConstraints() const {
+ dbgs() << "Constraints:\n";
+
+ for (unsigned i = 0, e = Constraints.size(); i != e; ++i)
+ PrintConstraint(Constraints[i]);
+}
+
+void Andersens::PrintPointsToGraph() const {
+ dbgs() << "Points-to graph:\n";
+ for (unsigned i = 0, e = GraphNodes.size(); i != e; ++i) {
+ const Node *N = &GraphNodes[i];
+ if (FindNode(i) != i) {
+ PrintNode(N);
+ dbgs() << "\t--> same as ";
+ PrintNode(&GraphNodes[FindNode(i)]);
+ dbgs() << "\n";
+ } else {
+ dbgs() << "[" << (N->PointsTo->count()) << "] ";
+ PrintNode(N);
+ dbgs() << "\t--> ";
+
+ bool first = true;
+ for (SparseBitVector<>::iterator bi = N->PointsTo->begin();
+ bi != N->PointsTo->end();
+ ++bi) {
+ if (!first)
+ dbgs() << ", ";
+ PrintNode(&GraphNodes[*bi]);
+ first = false;
+ }
+ dbgs() << "\n";
+ }
+ }
+}
diff --git a/lib/Analysis/IPA/CMakeLists.txt b/lib/Analysis/IPA/CMakeLists.txt
new file mode 100644
index 0000000..1ebb0be
--- /dev/null
+++ b/lib/Analysis/IPA/CMakeLists.txt
@@ -0,0 +1,7 @@
+add_llvm_library(LLVMipa
+ Andersens.cpp
+ CallGraph.cpp
+ CallGraphSCCPass.cpp
+ FindUsedTypes.cpp
+ GlobalsModRef.cpp
+ )
diff --git a/lib/Analysis/IPA/CallGraph.cpp b/lib/Analysis/IPA/CallGraph.cpp
new file mode 100644
index 0000000..8c43aa1
--- /dev/null
+++ b/lib/Analysis/IPA/CallGraph.cpp
@@ -0,0 +1,310 @@
+//===- CallGraph.cpp - Build a Module's call graph ------------------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements the CallGraph class and provides the BasicCallGraph
+// default implementation.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/CallGraph.h"
+#include "llvm/Module.h"
+#include "llvm/Instructions.h"
+#include "llvm/IntrinsicInst.h"
+#include "llvm/Support/CallSite.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+using namespace llvm;
+
+namespace {
+
+//===----------------------------------------------------------------------===//
+// BasicCallGraph class definition
+//
+class BasicCallGraph : public ModulePass, public CallGraph {
+ // Root is root of the call graph, or the external node if a 'main' function
+ // couldn't be found.
+ //
+ CallGraphNode *Root;
+
+ // ExternalCallingNode - This node has edges to all external functions and
+ // those internal functions that have their address taken.
+ CallGraphNode *ExternalCallingNode;
+
+ // CallsExternalNode - This node has edges to it from all functions making
+ // indirect calls or calling an external function.
+ CallGraphNode *CallsExternalNode;
+
+public:
+ static char ID; // Class identification, replacement for typeinfo
+ BasicCallGraph() : ModulePass(&ID), Root(0),
+ ExternalCallingNode(0), CallsExternalNode(0) {}
+
+ // runOnModule - Compute the call graph for the specified module.
+ virtual bool runOnModule(Module &M) {
+ CallGraph::initialize(M);
+
+ ExternalCallingNode = getOrInsertFunction(0);
+ CallsExternalNode = new CallGraphNode(0);
+ Root = 0;
+
+ // Add every function to the call graph.
+ for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
+ addToCallGraph(I);
+
+ // If we didn't find a main function, use the external call graph node
+ if (Root == 0) Root = ExternalCallingNode;
+
+ return false;
+ }
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.setPreservesAll();
+ }
+
+ virtual void print(raw_ostream &OS, const Module *) const {
+ OS << "CallGraph Root is: ";
+ if (Function *F = getRoot()->getFunction())
+ OS << F->getName() << "\n";
+ else {
+ OS << "<<null function: 0x" << getRoot() << ">>\n";
+ }
+
+ CallGraph::print(OS, 0);
+ }
+
+ virtual void releaseMemory() {
+ destroy();
+ }
+
+ /// getAdjustedAnalysisPointer - This method is used when a pass implements
+ /// an analysis interface through multiple inheritance. If needed, it should
+ /// override this to adjust the this pointer as needed for the specified pass
+ /// info.
+ virtual void *getAdjustedAnalysisPointer(const PassInfo *PI) {
+ if (PI->isPassID(&CallGraph::ID))
+ return (CallGraph*)this;
+ return this;
+ }
+
+ CallGraphNode* getExternalCallingNode() const { return ExternalCallingNode; }
+ CallGraphNode* getCallsExternalNode() const { return CallsExternalNode; }
+
+ // getRoot - Return the root of the call graph, which is either main, or if
+ // main cannot be found, the external node.
+ //
+ CallGraphNode *getRoot() { return Root; }
+ const CallGraphNode *getRoot() const { return Root; }
+
+private:
+ //===---------------------------------------------------------------------
+ // Implementation of CallGraph construction
+ //
+
+ // addToCallGraph - Add a function to the call graph, and link the node to all
+ // of the functions that it calls.
+ //
+ void addToCallGraph(Function *F) {
+ CallGraphNode *Node = getOrInsertFunction(F);
+
+ // If this function has external linkage, anything could call it.
+ if (!F->hasLocalLinkage()) {
+ ExternalCallingNode->addCalledFunction(CallSite(), Node);
+
+ // Found the entry point?
+ if (F->getName() == "main") {
+ if (Root) // Found multiple external mains? Don't pick one.
+ Root = ExternalCallingNode;
+ else
+ Root = Node; // Found a main, keep track of it!
+ }
+ }
+
+ // Loop over all of the users of the function, looking for non-call uses.
+ for (Value::use_iterator I = F->use_begin(), E = F->use_end(); I != E; ++I)
+ if ((!isa<CallInst>(I) && !isa<InvokeInst>(I))
+ || !CallSite(cast<Instruction>(I)).isCallee(I)) {
+ // Not a call, or being used as a parameter rather than as the callee.
+ ExternalCallingNode->addCalledFunction(CallSite(), Node);
+ break;
+ }
+
+ // If this function is not defined in this translation unit, it could call
+ // anything.
+ if (F->isDeclaration() && !F->isIntrinsic())
+ Node->addCalledFunction(CallSite(), CallsExternalNode);
+
+ // Look for calls by this function.
+ for (Function::iterator BB = F->begin(), BBE = F->end(); BB != BBE; ++BB)
+ for (BasicBlock::iterator II = BB->begin(), IE = BB->end();
+ II != IE; ++II) {
+ CallSite CS = CallSite::get(II);
+ if (CS.getInstruction() && !isa<DbgInfoIntrinsic>(II)) {
+ const Function *Callee = CS.getCalledFunction();
+ if (Callee)
+ Node->addCalledFunction(CS, getOrInsertFunction(Callee));
+ else
+ Node->addCalledFunction(CS, CallsExternalNode);
+ }
+ }
+ }
+
+ //
+ // destroy - Release memory for the call graph
+ virtual void destroy() {
+ /// CallsExternalNode is not in the function map, delete it explicitly.
+ delete CallsExternalNode;
+ CallsExternalNode = 0;
+ CallGraph::destroy();
+ }
+};
+
+} //End anonymous namespace
+
+static RegisterAnalysisGroup<CallGraph> X("Call Graph");
+static RegisterPass<BasicCallGraph>
+Y("basiccg", "Basic CallGraph Construction", false, true);
+static RegisterAnalysisGroup<CallGraph, true> Z(Y);
+
+char CallGraph::ID = 0;
+char BasicCallGraph::ID = 0;
+
+void CallGraph::initialize(Module &M) {
+ Mod = &M;
+}
+
+void CallGraph::destroy() {
+ if (FunctionMap.empty()) return;
+
+ for (FunctionMapTy::iterator I = FunctionMap.begin(), E = FunctionMap.end();
+ I != E; ++I)
+ delete I->second;
+ FunctionMap.clear();
+}
+
+void CallGraph::print(raw_ostream &OS, Module*) const {
+ for (CallGraph::const_iterator I = begin(), E = end(); I != E; ++I)
+ I->second->print(OS);
+}
+void CallGraph::dump() const {
+ print(dbgs(), 0);
+}
+
+//===----------------------------------------------------------------------===//
+// Implementations of public modification methods
+//
+
+// removeFunctionFromModule - Unlink the function from this module, returning
+// it. Because this removes the function from the module, the call graph node
+// is destroyed. This is only valid if the function does not call any other
+// functions (ie, there are no edges in it's CGN). The easiest way to do this
+// is to dropAllReferences before calling this.
+//
+Function *CallGraph::removeFunctionFromModule(CallGraphNode *CGN) {
+ assert(CGN->empty() && "Cannot remove function from call "
+ "graph if it references other functions!");
+ Function *F = CGN->getFunction(); // Get the function for the call graph node
+ delete CGN; // Delete the call graph node for this func
+ FunctionMap.erase(F); // Remove the call graph node from the map
+
+ Mod->getFunctionList().remove(F);
+ return F;
+}
+
+// getOrInsertFunction - This method is identical to calling operator[], but
+// it will insert a new CallGraphNode for the specified function if one does
+// not already exist.
+CallGraphNode *CallGraph::getOrInsertFunction(const Function *F) {
+ CallGraphNode *&CGN = FunctionMap[F];
+ if (CGN) return CGN;
+
+ assert((!F || F->getParent() == Mod) && "Function not in current module!");
+ return CGN = new CallGraphNode(const_cast<Function*>(F));
+}
+
+void CallGraphNode::print(raw_ostream &OS) const {
+ if (Function *F = getFunction())
+ OS << "Call graph node for function: '" << F->getName() << "'";
+ else
+ OS << "Call graph node <<null function>>";
+
+ OS << "<<0x" << this << ">> #uses=" << getNumReferences() << '\n';
+
+ for (const_iterator I = begin(), E = end(); I != E; ++I)
+ if (Function *FI = I->second->getFunction())
+ OS << " Calls function '" << FI->getName() <<"'\n";
+ else
+ OS << " Calls external node\n";
+ OS << "\n";
+}
+
+void CallGraphNode::dump() const { print(dbgs()); }
+
+/// removeCallEdgeFor - This method removes the edge in the node for the
+/// specified call site. Note that this method takes linear time, so it
+/// should be used sparingly.
+void CallGraphNode::removeCallEdgeFor(CallSite CS) {
+ for (CalledFunctionsVector::iterator I = CalledFunctions.begin(); ; ++I) {
+ assert(I != CalledFunctions.end() && "Cannot find callsite to remove!");
+ if (I->first == CS.getInstruction()) {
+ I->second->DropRef();
+ *I = CalledFunctions.back();
+ CalledFunctions.pop_back();
+ return;
+ }
+ }
+}
+
+
+// removeAnyCallEdgeTo - This method removes any call edges from this node to
+// the specified callee function. This takes more time to execute than
+// removeCallEdgeTo, so it should not be used unless necessary.
+void CallGraphNode::removeAnyCallEdgeTo(CallGraphNode *Callee) {
+ for (unsigned i = 0, e = CalledFunctions.size(); i != e; ++i)
+ if (CalledFunctions[i].second == Callee) {
+ Callee->DropRef();
+ CalledFunctions[i] = CalledFunctions.back();
+ CalledFunctions.pop_back();
+ --i; --e;
+ }
+}
+
+/// removeOneAbstractEdgeTo - Remove one edge associated with a null callsite
+/// from this node to the specified callee function.
+void CallGraphNode::removeOneAbstractEdgeTo(CallGraphNode *Callee) {
+ for (CalledFunctionsVector::iterator I = CalledFunctions.begin(); ; ++I) {
+ assert(I != CalledFunctions.end() && "Cannot find callee to remove!");
+ CallRecord &CR = *I;
+ if (CR.second == Callee && CR.first == 0) {
+ Callee->DropRef();
+ *I = CalledFunctions.back();
+ CalledFunctions.pop_back();
+ return;
+ }
+ }
+}
+
+/// replaceCallEdge - This method replaces the edge in the node for the
+/// specified call site with a new one. Note that this method takes linear
+/// time, so it should be used sparingly.
+void CallGraphNode::replaceCallEdge(CallSite CS,
+ CallSite NewCS, CallGraphNode *NewNode){
+ for (CalledFunctionsVector::iterator I = CalledFunctions.begin(); ; ++I) {
+ assert(I != CalledFunctions.end() && "Cannot find callsite to remove!");
+ if (I->first == CS.getInstruction()) {
+ I->second->DropRef();
+ I->first = NewCS.getInstruction();
+ I->second = NewNode;
+ NewNode->AddRef();
+ return;
+ }
+ }
+}
+
+// Enuse that users of CallGraph.h also link with this file
+DEFINING_FILE_FOR(CallGraph)
diff --git a/lib/Analysis/IPA/CallGraphSCCPass.cpp b/lib/Analysis/IPA/CallGraphSCCPass.cpp
new file mode 100644
index 0000000..0e333d1
--- /dev/null
+++ b/lib/Analysis/IPA/CallGraphSCCPass.cpp
@@ -0,0 +1,441 @@
+//===- CallGraphSCCPass.cpp - Pass that operates BU on call graph ---------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements the CallGraphSCCPass class, which is used for passes
+// which are implemented as bottom-up traversals on the call graph. Because
+// there may be cycles in the call graph, passes of this type operate on the
+// call-graph in SCC order: that is, they process function bottom-up, except for
+// recursive functions, which they process all at once.
+//
+//===----------------------------------------------------------------------===//
+
+#define DEBUG_TYPE "cgscc-passmgr"
+#include "llvm/CallGraphSCCPass.h"
+#include "llvm/Analysis/CallGraph.h"
+#include "llvm/ADT/SCCIterator.h"
+#include "llvm/PassManagers.h"
+#include "llvm/Function.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/IntrinsicInst.h"
+#include "llvm/Support/raw_ostream.h"
+using namespace llvm;
+
+//===----------------------------------------------------------------------===//
+// CGPassManager
+//
+/// CGPassManager manages FPPassManagers and CallGraphSCCPasses.
+
+namespace {
+
+class CGPassManager : public ModulePass, public PMDataManager {
+public:
+ static char ID;
+ explicit CGPassManager(int Depth)
+ : ModulePass(&ID), PMDataManager(Depth) { }
+
+ /// run - Execute all of the passes scheduled for execution. Keep track of
+ /// whether any of the passes modifies the module, and if so, return true.
+ bool runOnModule(Module &M);
+
+ bool doInitialization(CallGraph &CG);
+ bool doFinalization(CallGraph &CG);
+
+ /// Pass Manager itself does not invalidate any analysis info.
+ void getAnalysisUsage(AnalysisUsage &Info) const {
+ // CGPassManager walks SCC and it needs CallGraph.
+ Info.addRequired<CallGraph>();
+ Info.setPreservesAll();
+ }
+
+ virtual const char *getPassName() const {
+ return "CallGraph Pass Manager";
+ }
+
+ virtual PMDataManager *getAsPMDataManager() { return this; }
+ virtual Pass *getAsPass() { return this; }
+
+ // Print passes managed by this manager
+ void dumpPassStructure(unsigned Offset) {
+ errs().indent(Offset*2) << "Call Graph SCC Pass Manager\n";
+ for (unsigned Index = 0; Index < getNumContainedPasses(); ++Index) {
+ Pass *P = getContainedPass(Index);
+ P->dumpPassStructure(Offset + 1);
+ dumpLastUses(P, Offset+1);
+ }
+ }
+
+ Pass *getContainedPass(unsigned N) {
+ assert(N < PassVector.size() && "Pass number out of range!");
+ return static_cast<Pass *>(PassVector[N]);
+ }
+
+ virtual PassManagerType getPassManagerType() const {
+ return PMT_CallGraphPassManager;
+ }
+
+private:
+ bool RunPassOnSCC(Pass *P, std::vector<CallGraphNode*> &CurSCC,
+ CallGraph &CG, bool &CallGraphUpToDate);
+ void RefreshCallGraph(std::vector<CallGraphNode*> &CurSCC, CallGraph &CG,
+ bool IsCheckingMode);
+};
+
+} // end anonymous namespace.
+
+char CGPassManager::ID = 0;
+
+bool CGPassManager::RunPassOnSCC(Pass *P, std::vector<CallGraphNode*> &CurSCC,
+ CallGraph &CG, bool &CallGraphUpToDate) {
+ bool Changed = false;
+ PMDataManager *PM = P->getAsPMDataManager();
+
+ if (PM == 0) {
+ CallGraphSCCPass *CGSP = (CallGraphSCCPass*)P;
+ if (!CallGraphUpToDate) {
+ RefreshCallGraph(CurSCC, CG, false);
+ CallGraphUpToDate = true;
+ }
+
+ Timer *T = StartPassTimer(CGSP);
+ Changed = CGSP->runOnSCC(CurSCC);
+ StopPassTimer(CGSP, T);
+
+ // After the CGSCCPass is done, when assertions are enabled, use
+ // RefreshCallGraph to verify that the callgraph was correctly updated.
+#ifndef NDEBUG
+ if (Changed)
+ RefreshCallGraph(CurSCC, CG, true);
+#endif
+
+ return Changed;
+ }
+
+
+ assert(PM->getPassManagerType() == PMT_FunctionPassManager &&
+ "Invalid CGPassManager member");
+ FPPassManager *FPP = (FPPassManager*)P;
+
+ // Run pass P on all functions in the current SCC.
+ for (unsigned i = 0, e = CurSCC.size(); i != e; ++i) {
+ if (Function *F = CurSCC[i]->getFunction()) {
+ dumpPassInfo(P, EXECUTION_MSG, ON_FUNCTION_MSG, F->getName());
+ Timer *T = StartPassTimer(FPP);
+ Changed |= FPP->runOnFunction(*F);
+ StopPassTimer(FPP, T);
+ }
+ }
+
+ // The function pass(es) modified the IR, they may have clobbered the
+ // callgraph.
+ if (Changed && CallGraphUpToDate) {
+ DEBUG(dbgs() << "CGSCCPASSMGR: Pass Dirtied SCC: "
+ << P->getPassName() << '\n');
+ CallGraphUpToDate = false;
+ }
+ return Changed;
+}
+
+
+/// RefreshCallGraph - Scan the functions in the specified CFG and resync the
+/// callgraph with the call sites found in it. This is used after
+/// FunctionPasses have potentially munged the callgraph, and can be used after
+/// CallGraphSCC passes to verify that they correctly updated the callgraph.
+///
+void CGPassManager::RefreshCallGraph(std::vector<CallGraphNode*> &CurSCC,
+ CallGraph &CG, bool CheckingMode) {
+ DenseMap<Value*, CallGraphNode*> CallSites;
+
+ DEBUG(dbgs() << "CGSCCPASSMGR: Refreshing SCC with " << CurSCC.size()
+ << " nodes:\n";
+ for (unsigned i = 0, e = CurSCC.size(); i != e; ++i)
+ CurSCC[i]->dump();
+ );
+
+ bool MadeChange = false;
+
+ // Scan all functions in the SCC.
+ for (unsigned sccidx = 0, e = CurSCC.size(); sccidx != e; ++sccidx) {
+ CallGraphNode *CGN = CurSCC[sccidx];
+ Function *F = CGN->getFunction();
+ if (F == 0 || F->isDeclaration()) continue;
+
+ // Walk the function body looking for call sites. Sync up the call sites in
+ // CGN with those actually in the function.
+
+ // Get the set of call sites currently in the function.
+ for (CallGraphNode::iterator I = CGN->begin(), E = CGN->end(); I != E; ) {
+ // If this call site is null, then the function pass deleted the call
+ // entirely and the WeakVH nulled it out.
+ if (I->first == 0 ||
+ // If we've already seen this call site, then the FunctionPass RAUW'd
+ // one call with another, which resulted in two "uses" in the edge
+ // list of the same call.
+ CallSites.count(I->first) ||
+
+ // If the call edge is not from a call or invoke, then the function
+ // pass RAUW'd a call with another value. This can happen when
+ // constant folding happens of well known functions etc.
+ CallSite::get(I->first).getInstruction() == 0) {
+ assert(!CheckingMode &&
+ "CallGraphSCCPass did not update the CallGraph correctly!");
+
+ // Just remove the edge from the set of callees, keep track of whether
+ // I points to the last element of the vector.
+ bool WasLast = I + 1 == E;
+ CGN->removeCallEdge(I);
+
+ // If I pointed to the last element of the vector, we have to bail out:
+ // iterator checking rejects comparisons of the resultant pointer with
+ // end.
+ if (WasLast)
+ break;
+ E = CGN->end();
+ continue;
+ }
+
+ assert(!CallSites.count(I->first) &&
+ "Call site occurs in node multiple times");
+ CallSites.insert(std::make_pair(I->first, I->second));
+ ++I;
+ }
+
+ // Loop over all of the instructions in the function, getting the callsites.
+ for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
+ for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
+ CallSite CS = CallSite::get(I);
+ if (!CS.getInstruction() || isa<DbgInfoIntrinsic>(I)) continue;
+
+ // If this call site already existed in the callgraph, just verify it
+ // matches up to expectations and remove it from CallSites.
+ DenseMap<Value*, CallGraphNode*>::iterator ExistingIt =
+ CallSites.find(CS.getInstruction());
+ if (ExistingIt != CallSites.end()) {
+ CallGraphNode *ExistingNode = ExistingIt->second;
+
+ // Remove from CallSites since we have now seen it.
+ CallSites.erase(ExistingIt);
+
+ // Verify that the callee is right.
+ if (ExistingNode->getFunction() == CS.getCalledFunction())
+ continue;
+
+ // If we are in checking mode, we are not allowed to actually mutate
+ // the callgraph. If this is a case where we can infer that the
+ // callgraph is less precise than it could be (e.g. an indirect call
+ // site could be turned direct), don't reject it in checking mode, and
+ // don't tweak it to be more precise.
+ if (CheckingMode && CS.getCalledFunction() &&
+ ExistingNode->getFunction() == 0)
+ continue;
+
+ assert(!CheckingMode &&
+ "CallGraphSCCPass did not update the CallGraph correctly!");
+
+ // If not, we either went from a direct call to indirect, indirect to
+ // direct, or direct to different direct.
+ CallGraphNode *CalleeNode;
+ if (Function *Callee = CS.getCalledFunction())
+ CalleeNode = CG.getOrInsertFunction(Callee);
+ else
+ CalleeNode = CG.getCallsExternalNode();
+
+ // Update the edge target in CGN.
+ for (CallGraphNode::iterator I = CGN->begin(); ; ++I) {
+ assert(I != CGN->end() && "Didn't find call entry");
+ if (I->first == CS.getInstruction()) {
+ I->second = CalleeNode;
+ break;
+ }
+ }
+ MadeChange = true;
+ continue;
+ }
+
+ assert(!CheckingMode &&
+ "CallGraphSCCPass did not update the CallGraph correctly!");
+
+ // If the call site didn't exist in the CGN yet, add it. We assume that
+ // newly introduced call sites won't be indirect. This could be fixed
+ // in the future.
+ CallGraphNode *CalleeNode;
+ if (Function *Callee = CS.getCalledFunction())
+ CalleeNode = CG.getOrInsertFunction(Callee);
+ else
+ CalleeNode = CG.getCallsExternalNode();
+
+ CGN->addCalledFunction(CS, CalleeNode);
+ MadeChange = true;
+ }
+
+ // After scanning this function, if we still have entries in callsites, then
+ // they are dangling pointers. WeakVH should save us for this, so abort if
+ // this happens.
+ assert(CallSites.empty() && "Dangling pointers found in call sites map");
+
+ // Periodically do an explicit clear to remove tombstones when processing
+ // large scc's.
+ if ((sccidx & 15) == 0)
+ CallSites.clear();
+ }
+
+ DEBUG(if (MadeChange) {
+ dbgs() << "CGSCCPASSMGR: Refreshed SCC is now:\n";
+ for (unsigned i = 0, e = CurSCC.size(); i != e; ++i)
+ CurSCC[i]->dump();
+ } else {
+ dbgs() << "CGSCCPASSMGR: SCC Refresh didn't change call graph.\n";
+ }
+ );
+}
+
+/// run - Execute all of the passes scheduled for execution. Keep track of
+/// whether any of the passes modifies the module, and if so, return true.
+bool CGPassManager::runOnModule(Module &M) {
+ CallGraph &CG = getAnalysis<CallGraph>();
+ bool Changed = doInitialization(CG);
+
+ std::vector<CallGraphNode*> CurSCC;
+
+ // Walk the callgraph in bottom-up SCC order.
+ for (scc_iterator<CallGraph*> CGI = scc_begin(&CG), E = scc_end(&CG);
+ CGI != E;) {
+ // Copy the current SCC and increment past it so that the pass can hack
+ // on the SCC if it wants to without invalidating our iterator.
+ CurSCC = *CGI;
+ ++CGI;
+
+
+ // CallGraphUpToDate - Keep track of whether the callgraph is known to be
+ // up-to-date or not. The CGSSC pass manager runs two types of passes:
+ // CallGraphSCC Passes and other random function passes. Because other
+ // random function passes are not CallGraph aware, they may clobber the
+ // call graph by introducing new calls or deleting other ones. This flag
+ // is set to false when we run a function pass so that we know to clean up
+ // the callgraph when we need to run a CGSCCPass again.
+ bool CallGraphUpToDate = true;
+
+ // Run all passes on current SCC.
+ for (unsigned PassNo = 0, e = getNumContainedPasses();
+ PassNo != e; ++PassNo) {
+ Pass *P = getContainedPass(PassNo);
+
+ // If we're in -debug-pass=Executions mode, construct the SCC node list,
+ // otherwise avoid constructing this string as it is expensive.
+ if (isPassDebuggingExecutionsOrMore()) {
+ std::string Functions;
+#ifndef NDEBUG
+ raw_string_ostream OS(Functions);
+ for (unsigned i = 0, e = CurSCC.size(); i != e; ++i) {
+ if (i) OS << ", ";
+ CurSCC[i]->print(OS);
+ }
+ OS.flush();
+#endif
+ dumpPassInfo(P, EXECUTION_MSG, ON_CG_MSG, Functions);
+ }
+ dumpRequiredSet(P);
+
+ initializeAnalysisImpl(P);
+
+ // Actually run this pass on the current SCC.
+ Changed |= RunPassOnSCC(P, CurSCC, CG, CallGraphUpToDate);
+
+ if (Changed)
+ dumpPassInfo(P, MODIFICATION_MSG, ON_CG_MSG, "");
+ dumpPreservedSet(P);
+
+ verifyPreservedAnalysis(P);
+ removeNotPreservedAnalysis(P);
+ recordAvailableAnalysis(P);
+ removeDeadPasses(P, "", ON_CG_MSG);
+ }
+
+ // If the callgraph was left out of date (because the last pass run was a
+ // functionpass), refresh it before we move on to the next SCC.
+ if (!CallGraphUpToDate)
+ RefreshCallGraph(CurSCC, CG, false);
+ }
+ Changed |= doFinalization(CG);
+ return Changed;
+}
+
+/// Initialize CG
+bool CGPassManager::doInitialization(CallGraph &CG) {
+ bool Changed = false;
+ for (unsigned i = 0, e = getNumContainedPasses(); i != e; ++i) {
+ if (PMDataManager *PM = getContainedPass(i)->getAsPMDataManager()) {
+ assert(PM->getPassManagerType() == PMT_FunctionPassManager &&
+ "Invalid CGPassManager member");
+ Changed |= ((FPPassManager*)PM)->doInitialization(CG.getModule());
+ } else {
+ Changed |= ((CallGraphSCCPass*)getContainedPass(i))->doInitialization(CG);
+ }
+ }
+ return Changed;
+}
+
+/// Finalize CG
+bool CGPassManager::doFinalization(CallGraph &CG) {
+ bool Changed = false;
+ for (unsigned i = 0, e = getNumContainedPasses(); i != e; ++i) {
+ if (PMDataManager *PM = getContainedPass(i)->getAsPMDataManager()) {
+ assert(PM->getPassManagerType() == PMT_FunctionPassManager &&
+ "Invalid CGPassManager member");
+ Changed |= ((FPPassManager*)PM)->doFinalization(CG.getModule());
+ } else {
+ Changed |= ((CallGraphSCCPass*)getContainedPass(i))->doFinalization(CG);
+ }
+ }
+ return Changed;
+}
+
+/// Assign pass manager to manage this pass.
+void CallGraphSCCPass::assignPassManager(PMStack &PMS,
+ PassManagerType PreferredType) {
+ // Find CGPassManager
+ while (!PMS.empty() &&
+ PMS.top()->getPassManagerType() > PMT_CallGraphPassManager)
+ PMS.pop();
+
+ assert(!PMS.empty() && "Unable to handle Call Graph Pass");
+ CGPassManager *CGP;
+
+ if (PMS.top()->getPassManagerType() == PMT_CallGraphPassManager)
+ CGP = (CGPassManager*)PMS.top();
+ else {
+ // Create new Call Graph SCC Pass Manager if it does not exist.
+ assert(!PMS.empty() && "Unable to create Call Graph Pass Manager");
+ PMDataManager *PMD = PMS.top();
+
+ // [1] Create new Call Graph Pass Manager
+ CGP = new CGPassManager(PMD->getDepth() + 1);
+
+ // [2] Set up new manager's top level manager
+ PMTopLevelManager *TPM = PMD->getTopLevelManager();
+ TPM->addIndirectPassManager(CGP);
+
+ // [3] Assign manager to manage this new manager. This may create
+ // and push new managers into PMS
+ Pass *P = CGP;
+ TPM->schedulePass(P);
+
+ // [4] Push new manager into PMS
+ PMS.push(CGP);
+ }
+
+ CGP->add(this);
+}
+
+/// getAnalysisUsage - For this class, we declare that we require and preserve
+/// the call graph. If the derived class implements this method, it should
+/// always explicitly call the implementation here.
+void CallGraphSCCPass::getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.addRequired<CallGraph>();
+ AU.addPreserved<CallGraph>();
+}
diff --git a/lib/Analysis/IPA/FindUsedTypes.cpp b/lib/Analysis/IPA/FindUsedTypes.cpp
new file mode 100644
index 0000000..c4fb0b9
--- /dev/null
+++ b/lib/Analysis/IPA/FindUsedTypes.cpp
@@ -0,0 +1,103 @@
+//===- FindUsedTypes.cpp - Find all Types used by a module ----------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This pass is used to seek out all of the types in use by the program. Note
+// that this analysis explicitly does not include types only used by the symbol
+// table.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/FindUsedTypes.h"
+#include "llvm/Constants.h"
+#include "llvm/DerivedTypes.h"
+#include "llvm/Module.h"
+#include "llvm/Assembly/Writer.h"
+#include "llvm/Support/InstIterator.h"
+#include "llvm/Support/raw_ostream.h"
+using namespace llvm;
+
+char FindUsedTypes::ID = 0;
+static RegisterPass<FindUsedTypes>
+X("print-used-types", "Find Used Types", false, true);
+
+// IncorporateType - Incorporate one type and all of its subtypes into the
+// collection of used types.
+//
+void FindUsedTypes::IncorporateType(const Type *Ty) {
+ // If ty doesn't already exist in the used types map, add it now, otherwise
+ // return.
+ if (!UsedTypes.insert(Ty).second) return; // Already contain Ty.
+
+ // Make sure to add any types this type references now.
+ //
+ for (Type::subtype_iterator I = Ty->subtype_begin(), E = Ty->subtype_end();
+ I != E; ++I)
+ IncorporateType(*I);
+}
+
+void FindUsedTypes::IncorporateValue(const Value *V) {
+ IncorporateType(V->getType());
+
+ // If this is a constant, it could be using other types...
+ if (const Constant *C = dyn_cast<Constant>(V)) {
+ if (!isa<GlobalValue>(C))
+ for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end();
+ OI != OE; ++OI)
+ IncorporateValue(*OI);
+ }
+}
+
+
+// run - This incorporates all types used by the specified module
+//
+bool FindUsedTypes::runOnModule(Module &m) {
+ UsedTypes.clear(); // reset if run multiple times...
+
+ // Loop over global variables, incorporating their types
+ for (Module::const_global_iterator I = m.global_begin(), E = m.global_end();
+ I != E; ++I) {
+ IncorporateType(I->getType());
+ if (I->hasInitializer())
+ IncorporateValue(I->getInitializer());
+ }
+
+ for (Module::iterator MI = m.begin(), ME = m.end(); MI != ME; ++MI) {
+ IncorporateType(MI->getType());
+ const Function &F = *MI;
+
+ // Loop over all of the instructions in the function, adding their return
+ // type as well as the types of their operands.
+ //
+ for (const_inst_iterator II = inst_begin(F), IE = inst_end(F);
+ II != IE; ++II) {
+ const Instruction &I = *II;
+
+ IncorporateType(I.getType()); // Incorporate the type of the instruction
+ for (User::const_op_iterator OI = I.op_begin(), OE = I.op_end();
+ OI != OE; ++OI)
+ IncorporateValue(*OI); // Insert inst operand types as well
+ }
+ }
+
+ return false;
+}
+
+// Print the types found in the module. If the optional Module parameter is
+// passed in, then the types are printed symbolically if possible, using the
+// symbol table from the module.
+//
+void FindUsedTypes::print(raw_ostream &OS, const Module *M) const {
+ OS << "Types in use by this module:\n";
+ for (std::set<const Type *>::const_iterator I = UsedTypes.begin(),
+ E = UsedTypes.end(); I != E; ++I) {
+ OS << " ";
+ WriteTypeSymbolic(OS, *I, M);
+ OS << '\n';
+ }
+}
diff --git a/lib/Analysis/IPA/GlobalsModRef.cpp b/lib/Analysis/IPA/GlobalsModRef.cpp
new file mode 100644
index 0000000..ec94bc8
--- /dev/null
+++ b/lib/Analysis/IPA/GlobalsModRef.cpp
@@ -0,0 +1,579 @@
+//===- GlobalsModRef.cpp - Simple Mod/Ref Analysis for Globals ------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This simple pass provides alias and mod/ref information for global values
+// that do not have their address taken, and keeps track of whether functions
+// read or write memory (are "pure"). For this simple (but very common) case,
+// we can provide pretty accurate and useful information.
+//
+//===----------------------------------------------------------------------===//
+
+#define DEBUG_TYPE "globalsmodref-aa"
+#include "llvm/Analysis/Passes.h"
+#include "llvm/Module.h"
+#include "llvm/Pass.h"
+#include "llvm/Instructions.h"
+#include "llvm/Constants.h"
+#include "llvm/DerivedTypes.h"
+#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/Analysis/CallGraph.h"
+#include "llvm/Analysis/MemoryBuiltins.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/InstIterator.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/ADT/SCCIterator.h"
+#include <set>
+using namespace llvm;
+
+STATISTIC(NumNonAddrTakenGlobalVars,
+ "Number of global vars without address taken");
+STATISTIC(NumNonAddrTakenFunctions,"Number of functions without address taken");
+STATISTIC(NumNoMemFunctions, "Number of functions that do not access memory");
+STATISTIC(NumReadMemFunctions, "Number of functions that only read memory");
+STATISTIC(NumIndirectGlobalVars, "Number of indirect global objects");
+
+namespace {
+ /// FunctionRecord - One instance of this structure is stored for every
+ /// function in the program. Later, the entries for these functions are
+ /// removed if the function is found to call an external function (in which
+ /// case we know nothing about it.
+ struct FunctionRecord {
+ /// GlobalInfo - Maintain mod/ref info for all of the globals without
+ /// addresses taken that are read or written (transitively) by this
+ /// function.
+ std::map<GlobalValue*, unsigned> GlobalInfo;
+
+ /// MayReadAnyGlobal - May read global variables, but it is not known which.
+ bool MayReadAnyGlobal;
+
+ unsigned getInfoForGlobal(GlobalValue *GV) const {
+ unsigned Effect = MayReadAnyGlobal ? AliasAnalysis::Ref : 0;
+ std::map<GlobalValue*, unsigned>::const_iterator I = GlobalInfo.find(GV);
+ if (I != GlobalInfo.end())
+ Effect |= I->second;
+ return Effect;
+ }
+
+ /// FunctionEffect - Capture whether or not this function reads or writes to
+ /// ANY memory. If not, we can do a lot of aggressive analysis on it.
+ unsigned FunctionEffect;
+
+ FunctionRecord() : MayReadAnyGlobal (false), FunctionEffect(0) {}
+ };
+
+ /// GlobalsModRef - The actual analysis pass.
+ class GlobalsModRef : public ModulePass, public AliasAnalysis {
+ /// NonAddressTakenGlobals - The globals that do not have their addresses
+ /// taken.
+ std::set<GlobalValue*> NonAddressTakenGlobals;
+
+ /// IndirectGlobals - The memory pointed to by this global is known to be
+ /// 'owned' by the global.
+ std::set<GlobalValue*> IndirectGlobals;
+
+ /// AllocsForIndirectGlobals - If an instruction allocates memory for an
+ /// indirect global, this map indicates which one.
+ std::map<Value*, GlobalValue*> AllocsForIndirectGlobals;
+
+ /// FunctionInfo - For each function, keep track of what globals are
+ /// modified or read.
+ std::map<Function*, FunctionRecord> FunctionInfo;
+
+ public:
+ static char ID;
+ GlobalsModRef() : ModulePass(&ID) {}
+
+ bool runOnModule(Module &M) {
+ InitializeAliasAnalysis(this); // set up super class
+ AnalyzeGlobals(M); // find non-addr taken globals
+ AnalyzeCallGraph(getAnalysis<CallGraph>(), M); // Propagate on CG
+ return false;
+ }
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AliasAnalysis::getAnalysisUsage(AU);
+ AU.addRequired<CallGraph>();
+ AU.setPreservesAll(); // Does not transform code
+ }
+
+ //------------------------------------------------
+ // Implement the AliasAnalysis API
+ //
+ AliasResult alias(const Value *V1, unsigned V1Size,
+ const Value *V2, unsigned V2Size);
+ ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size);
+ ModRefResult getModRefInfo(CallSite CS1, CallSite CS2) {
+ return AliasAnalysis::getModRefInfo(CS1,CS2);
+ }
+
+ /// getModRefBehavior - Return the behavior of the specified function if
+ /// called from the specified call site. The call site may be null in which
+ /// case the most generic behavior of this function should be returned.
+ ModRefBehavior getModRefBehavior(Function *F,
+ std::vector<PointerAccessInfo> *Info) {
+ if (FunctionRecord *FR = getFunctionInfo(F)) {
+ if (FR->FunctionEffect == 0)
+ return DoesNotAccessMemory;
+ else if ((FR->FunctionEffect & Mod) == 0)
+ return OnlyReadsMemory;
+ }
+ return AliasAnalysis::getModRefBehavior(F, Info);
+ }
+
+ /// getModRefBehavior - Return the behavior of the specified function if
+ /// called from the specified call site. The call site may be null in which
+ /// case the most generic behavior of this function should be returned.
+ ModRefBehavior getModRefBehavior(CallSite CS,
+ std::vector<PointerAccessInfo> *Info) {
+ Function* F = CS.getCalledFunction();
+ if (!F) return AliasAnalysis::getModRefBehavior(CS, Info);
+ if (FunctionRecord *FR = getFunctionInfo(F)) {
+ if (FR->FunctionEffect == 0)
+ return DoesNotAccessMemory;
+ else if ((FR->FunctionEffect & Mod) == 0)
+ return OnlyReadsMemory;
+ }
+ return AliasAnalysis::getModRefBehavior(CS, Info);
+ }
+
+ virtual void deleteValue(Value *V);
+ virtual void copyValue(Value *From, Value *To);
+
+ /// getAdjustedAnalysisPointer - This method is used when a pass implements
+ /// an analysis interface through multiple inheritance. If needed, it
+ /// should override this to adjust the this pointer as needed for the
+ /// specified pass info.
+ virtual void *getAdjustedAnalysisPointer(const PassInfo *PI) {
+ if (PI->isPassID(&AliasAnalysis::ID))
+ return (AliasAnalysis*)this;
+ return this;
+ }
+
+ private:
+ /// getFunctionInfo - Return the function info for the function, or null if
+ /// we don't have anything useful to say about it.
+ FunctionRecord *getFunctionInfo(Function *F) {
+ std::map<Function*, FunctionRecord>::iterator I = FunctionInfo.find(F);
+ if (I != FunctionInfo.end())
+ return &I->second;
+ return 0;
+ }
+
+ void AnalyzeGlobals(Module &M);
+ void AnalyzeCallGraph(CallGraph &CG, Module &M);
+ bool AnalyzeUsesOfPointer(Value *V, std::vector<Function*> &Readers,
+ std::vector<Function*> &Writers,
+ GlobalValue *OkayStoreDest = 0);
+ bool AnalyzeIndirectGlobalMemory(GlobalValue *GV);
+ };
+}
+
+char GlobalsModRef::ID = 0;
+static RegisterPass<GlobalsModRef>
+X("globalsmodref-aa", "Simple mod/ref analysis for globals", false, true);
+static RegisterAnalysisGroup<AliasAnalysis> Y(X);
+
+Pass *llvm::createGlobalsModRefPass() { return new GlobalsModRef(); }
+
+/// AnalyzeGlobals - Scan through the users of all of the internal
+/// GlobalValue's in the program. If none of them have their "address taken"
+/// (really, their address passed to something nontrivial), record this fact,
+/// and record the functions that they are used directly in.
+void GlobalsModRef::AnalyzeGlobals(Module &M) {
+ std::vector<Function*> Readers, Writers;
+ for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
+ if (I->hasLocalLinkage()) {
+ if (!AnalyzeUsesOfPointer(I, Readers, Writers)) {
+ // Remember that we are tracking this global.
+ NonAddressTakenGlobals.insert(I);
+ ++NumNonAddrTakenFunctions;
+ }
+ Readers.clear(); Writers.clear();
+ }
+
+ for (Module::global_iterator I = M.global_begin(), E = M.global_end();
+ I != E; ++I)
+ if (I->hasLocalLinkage()) {
+ if (!AnalyzeUsesOfPointer(I, Readers, Writers)) {
+ // Remember that we are tracking this global, and the mod/ref fns
+ NonAddressTakenGlobals.insert(I);
+
+ for (unsigned i = 0, e = Readers.size(); i != e; ++i)
+ FunctionInfo[Readers[i]].GlobalInfo[I] |= Ref;
+
+ if (!I->isConstant()) // No need to keep track of writers to constants
+ for (unsigned i = 0, e = Writers.size(); i != e; ++i)
+ FunctionInfo[Writers[i]].GlobalInfo[I] |= Mod;
+ ++NumNonAddrTakenGlobalVars;
+
+ // If this global holds a pointer type, see if it is an indirect global.
+ if (isa<PointerType>(I->getType()->getElementType()) &&
+ AnalyzeIndirectGlobalMemory(I))
+ ++NumIndirectGlobalVars;
+ }
+ Readers.clear(); Writers.clear();
+ }
+}
+
+/// AnalyzeUsesOfPointer - Look at all of the users of the specified pointer.
+/// If this is used by anything complex (i.e., the address escapes), return
+/// true. Also, while we are at it, keep track of those functions that read and
+/// write to the value.
+///
+/// If OkayStoreDest is non-null, stores into this global are allowed.
+bool GlobalsModRef::AnalyzeUsesOfPointer(Value *V,
+ std::vector<Function*> &Readers,
+ std::vector<Function*> &Writers,
+ GlobalValue *OkayStoreDest) {
+ if (!isa<PointerType>(V->getType())) return true;
+
+ for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
+ if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
+ Readers.push_back(LI->getParent()->getParent());
+ } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
+ if (V == SI->getOperand(1)) {
+ Writers.push_back(SI->getParent()->getParent());
+ } else if (SI->getOperand(1) != OkayStoreDest) {
+ return true; // Storing the pointer
+ }
+ } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) {
+ if (AnalyzeUsesOfPointer(GEP, Readers, Writers)) return true;
+ } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) {
+ if (AnalyzeUsesOfPointer(BCI, Readers, Writers, OkayStoreDest))
+ return true;
+ } else if (isFreeCall(*UI)) {
+ Writers.push_back(cast<Instruction>(*UI)->getParent()->getParent());
+ } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
+ // Make sure that this is just the function being called, not that it is
+ // passing into the function.
+ for (unsigned i = 1, e = CI->getNumOperands(); i != e; ++i)
+ if (CI->getOperand(i) == V) return true;
+ } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
+ // Make sure that this is just the function being called, not that it is
+ // passing into the function.
+ for (unsigned i = 3, e = II->getNumOperands(); i != e; ++i)
+ if (II->getOperand(i) == V) return true;
+ } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
+ if (CE->getOpcode() == Instruction::GetElementPtr ||
+ CE->getOpcode() == Instruction::BitCast) {
+ if (AnalyzeUsesOfPointer(CE, Readers, Writers))
+ return true;
+ } else {
+ return true;
+ }
+ } else if (ICmpInst *ICI = dyn_cast<ICmpInst>(*UI)) {
+ if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
+ return true; // Allow comparison against null.
+ } else {
+ return true;
+ }
+ return false;
+}
+
+/// AnalyzeIndirectGlobalMemory - We found an non-address-taken global variable
+/// which holds a pointer type. See if the global always points to non-aliased
+/// heap memory: that is, all initializers of the globals are allocations, and
+/// those allocations have no use other than initialization of the global.
+/// Further, all loads out of GV must directly use the memory, not store the
+/// pointer somewhere. If this is true, we consider the memory pointed to by
+/// GV to be owned by GV and can disambiguate other pointers from it.
+bool GlobalsModRef::AnalyzeIndirectGlobalMemory(GlobalValue *GV) {
+ // Keep track of values related to the allocation of the memory, f.e. the
+ // value produced by the malloc call and any casts.
+ std::vector<Value*> AllocRelatedValues;
+
+ // Walk the user list of the global. If we find anything other than a direct
+ // load or store, bail out.
+ for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I){
+ if (LoadInst *LI = dyn_cast<LoadInst>(*I)) {
+ // The pointer loaded from the global can only be used in simple ways:
+ // we allow addressing of it and loading storing to it. We do *not* allow
+ // storing the loaded pointer somewhere else or passing to a function.
+ std::vector<Function*> ReadersWriters;
+ if (AnalyzeUsesOfPointer(LI, ReadersWriters, ReadersWriters))
+ return false; // Loaded pointer escapes.
+ // TODO: Could try some IP mod/ref of the loaded pointer.
+ } else if (StoreInst *SI = dyn_cast<StoreInst>(*I)) {
+ // Storing the global itself.
+ if (SI->getOperand(0) == GV) return false;
+
+ // If storing the null pointer, ignore it.
+ if (isa<ConstantPointerNull>(SI->getOperand(0)))
+ continue;
+
+ // Check the value being stored.
+ Value *Ptr = SI->getOperand(0)->getUnderlyingObject();
+
+ if (isMalloc(Ptr)) {
+ // Okay, easy case.
+ } else if (CallInst *CI = dyn_cast<CallInst>(Ptr)) {
+ Function *F = CI->getCalledFunction();
+ if (!F || !F->isDeclaration()) return false; // Too hard to analyze.
+ if (F->getName() != "calloc") return false; // Not calloc.
+ } else {
+ return false; // Too hard to analyze.
+ }
+
+ // Analyze all uses of the allocation. If any of them are used in a
+ // non-simple way (e.g. stored to another global) bail out.
+ std::vector<Function*> ReadersWriters;
+ if (AnalyzeUsesOfPointer(Ptr, ReadersWriters, ReadersWriters, GV))
+ return false; // Loaded pointer escapes.
+
+ // Remember that this allocation is related to the indirect global.
+ AllocRelatedValues.push_back(Ptr);
+ } else {
+ // Something complex, bail out.
+ return false;
+ }
+ }
+
+ // Okay, this is an indirect global. Remember all of the allocations for
+ // this global in AllocsForIndirectGlobals.
+ while (!AllocRelatedValues.empty()) {
+ AllocsForIndirectGlobals[AllocRelatedValues.back()] = GV;
+ AllocRelatedValues.pop_back();
+ }
+ IndirectGlobals.insert(GV);
+ return true;
+}
+
+/// AnalyzeCallGraph - At this point, we know the functions where globals are
+/// immediately stored to and read from. Propagate this information up the call
+/// graph to all callers and compute the mod/ref info for all memory for each
+/// function.
+void GlobalsModRef::AnalyzeCallGraph(CallGraph &CG, Module &M) {
+ // We do a bottom-up SCC traversal of the call graph. In other words, we
+ // visit all callees before callers (leaf-first).
+ for (scc_iterator<CallGraph*> I = scc_begin(&CG), E = scc_end(&CG); I != E;
+ ++I) {
+ std::vector<CallGraphNode *> &SCC = *I;
+ assert(!SCC.empty() && "SCC with no functions?");
+
+ if (!SCC[0]->getFunction()) {
+ // Calls externally - can't say anything useful. Remove any existing
+ // function records (may have been created when scanning globals).
+ for (unsigned i = 0, e = SCC.size(); i != e; ++i)
+ FunctionInfo.erase(SCC[i]->getFunction());
+ continue;
+ }
+
+ FunctionRecord &FR = FunctionInfo[SCC[0]->getFunction()];
+
+ bool KnowNothing = false;
+ unsigned FunctionEffect = 0;
+
+ // Collect the mod/ref properties due to called functions. We only compute
+ // one mod-ref set.
+ for (unsigned i = 0, e = SCC.size(); i != e && !KnowNothing; ++i) {
+ Function *F = SCC[i]->getFunction();
+ if (!F) {
+ KnowNothing = true;
+ break;
+ }
+
+ if (F->isDeclaration()) {
+ // Try to get mod/ref behaviour from function attributes.
+ if (F->doesNotAccessMemory()) {
+ // Can't do better than that!
+ } else if (F->onlyReadsMemory()) {
+ FunctionEffect |= Ref;
+ if (!F->isIntrinsic())
+ // This function might call back into the module and read a global -
+ // consider every global as possibly being read by this function.
+ FR.MayReadAnyGlobal = true;
+ } else {
+ FunctionEffect |= ModRef;
+ // Can't say anything useful unless it's an intrinsic - they don't
+ // read or write global variables of the kind considered here.
+ KnowNothing = !F->isIntrinsic();
+ }
+ continue;
+ }
+
+ for (CallGraphNode::iterator CI = SCC[i]->begin(), E = SCC[i]->end();
+ CI != E && !KnowNothing; ++CI)
+ if (Function *Callee = CI->second->getFunction()) {
+ if (FunctionRecord *CalleeFR = getFunctionInfo(Callee)) {
+ // Propagate function effect up.
+ FunctionEffect |= CalleeFR->FunctionEffect;
+
+ // Incorporate callee's effects on globals into our info.
+ for (std::map<GlobalValue*, unsigned>::iterator GI =
+ CalleeFR->GlobalInfo.begin(), E = CalleeFR->GlobalInfo.end();
+ GI != E; ++GI)
+ FR.GlobalInfo[GI->first] |= GI->second;
+ FR.MayReadAnyGlobal |= CalleeFR->MayReadAnyGlobal;
+ } else {
+ // Can't say anything about it. However, if it is inside our SCC,
+ // then nothing needs to be done.
+ CallGraphNode *CalleeNode = CG[Callee];
+ if (std::find(SCC.begin(), SCC.end(), CalleeNode) == SCC.end())
+ KnowNothing = true;
+ }
+ } else {
+ KnowNothing = true;
+ }
+ }
+
+ // If we can't say anything useful about this SCC, remove all SCC functions
+ // from the FunctionInfo map.
+ if (KnowNothing) {
+ for (unsigned i = 0, e = SCC.size(); i != e; ++i)
+ FunctionInfo.erase(SCC[i]->getFunction());
+ continue;
+ }
+
+ // Scan the function bodies for explicit loads or stores.
+ for (unsigned i = 0, e = SCC.size(); i != e && FunctionEffect != ModRef;++i)
+ for (inst_iterator II = inst_begin(SCC[i]->getFunction()),
+ E = inst_end(SCC[i]->getFunction());
+ II != E && FunctionEffect != ModRef; ++II)
+ if (isa<LoadInst>(*II)) {
+ FunctionEffect |= Ref;
+ if (cast<LoadInst>(*II).isVolatile())
+ // Volatile loads may have side-effects, so mark them as writing
+ // memory (for example, a flag inside the processor).
+ FunctionEffect |= Mod;
+ } else if (isa<StoreInst>(*II)) {
+ FunctionEffect |= Mod;
+ if (cast<StoreInst>(*II).isVolatile())
+ // Treat volatile stores as reading memory somewhere.
+ FunctionEffect |= Ref;
+ } else if (isMalloc(&cast<Instruction>(*II)) ||
+ isFreeCall(&cast<Instruction>(*II))) {
+ FunctionEffect |= ModRef;
+ }
+
+ if ((FunctionEffect & Mod) == 0)
+ ++NumReadMemFunctions;
+ if (FunctionEffect == 0)
+ ++NumNoMemFunctions;
+ FR.FunctionEffect = FunctionEffect;
+
+ // Finally, now that we know the full effect on this SCC, clone the
+ // information to each function in the SCC.
+ for (unsigned i = 1, e = SCC.size(); i != e; ++i)
+ FunctionInfo[SCC[i]->getFunction()] = FR;
+ }
+}
+
+
+
+/// alias - If one of the pointers is to a global that we are tracking, and the
+/// other is some random pointer, we know there cannot be an alias, because the
+/// address of the global isn't taken.
+AliasAnalysis::AliasResult
+GlobalsModRef::alias(const Value *V1, unsigned V1Size,
+ const Value *V2, unsigned V2Size) {
+ // Get the base object these pointers point to.
+ Value *UV1 = const_cast<Value*>(V1->getUnderlyingObject());
+ Value *UV2 = const_cast<Value*>(V2->getUnderlyingObject());
+
+ // If either of the underlying values is a global, they may be non-addr-taken
+ // globals, which we can answer queries about.
+ GlobalValue *GV1 = dyn_cast<GlobalValue>(UV1);
+ GlobalValue *GV2 = dyn_cast<GlobalValue>(UV2);
+ if (GV1 || GV2) {
+ // If the global's address is taken, pretend we don't know it's a pointer to
+ // the global.
+ if (GV1 && !NonAddressTakenGlobals.count(GV1)) GV1 = 0;
+ if (GV2 && !NonAddressTakenGlobals.count(GV2)) GV2 = 0;
+
+ // If the two pointers are derived from two different non-addr-taken
+ // globals, or if one is and the other isn't, we know these can't alias.
+ if ((GV1 || GV2) && GV1 != GV2)
+ return NoAlias;
+
+ // Otherwise if they are both derived from the same addr-taken global, we
+ // can't know the two accesses don't overlap.
+ }
+
+ // These pointers may be based on the memory owned by an indirect global. If
+ // so, we may be able to handle this. First check to see if the base pointer
+ // is a direct load from an indirect global.
+ GV1 = GV2 = 0;
+ if (LoadInst *LI = dyn_cast<LoadInst>(UV1))
+ if (GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getOperand(0)))
+ if (IndirectGlobals.count(GV))
+ GV1 = GV;
+ if (LoadInst *LI = dyn_cast<LoadInst>(UV2))
+ if (GlobalVariable *GV = dyn_cast<GlobalVariable>(LI->getOperand(0)))
+ if (IndirectGlobals.count(GV))
+ GV2 = GV;
+
+ // These pointers may also be from an allocation for the indirect global. If
+ // so, also handle them.
+ if (AllocsForIndirectGlobals.count(UV1))
+ GV1 = AllocsForIndirectGlobals[UV1];
+ if (AllocsForIndirectGlobals.count(UV2))
+ GV2 = AllocsForIndirectGlobals[UV2];
+
+ // Now that we know whether the two pointers are related to indirect globals,
+ // use this to disambiguate the pointers. If either pointer is based on an
+ // indirect global and if they are not both based on the same indirect global,
+ // they cannot alias.
+ if ((GV1 || GV2) && GV1 != GV2)
+ return NoAlias;
+
+ return AliasAnalysis::alias(V1, V1Size, V2, V2Size);
+}
+
+AliasAnalysis::ModRefResult
+GlobalsModRef::getModRefInfo(CallSite CS, Value *P, unsigned Size) {
+ unsigned Known = ModRef;
+
+ // If we are asking for mod/ref info of a direct call with a pointer to a
+ // global we are tracking, return information if we have it.
+ if (GlobalValue *GV = dyn_cast<GlobalValue>(P->getUnderlyingObject()))
+ if (GV->hasLocalLinkage())
+ if (Function *F = CS.getCalledFunction())
+ if (NonAddressTakenGlobals.count(GV))
+ if (FunctionRecord *FR = getFunctionInfo(F))
+ Known = FR->getInfoForGlobal(GV);
+
+ if (Known == NoModRef)
+ return NoModRef; // No need to query other mod/ref analyses
+ return ModRefResult(Known & AliasAnalysis::getModRefInfo(CS, P, Size));
+}
+
+
+//===----------------------------------------------------------------------===//
+// Methods to update the analysis as a result of the client transformation.
+//
+void GlobalsModRef::deleteValue(Value *V) {
+ if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
+ if (NonAddressTakenGlobals.erase(GV)) {
+ // This global might be an indirect global. If so, remove it and remove
+ // any AllocRelatedValues for it.
+ if (IndirectGlobals.erase(GV)) {
+ // Remove any entries in AllocsForIndirectGlobals for this global.
+ for (std::map<Value*, GlobalValue*>::iterator
+ I = AllocsForIndirectGlobals.begin(),
+ E = AllocsForIndirectGlobals.end(); I != E; ) {
+ if (I->second == GV) {
+ AllocsForIndirectGlobals.erase(I++);
+ } else {
+ ++I;
+ }
+ }
+ }
+ }
+ }
+
+ // Otherwise, if this is an allocation related to an indirect global, remove
+ // it.
+ AllocsForIndirectGlobals.erase(V);
+
+ AliasAnalysis::deleteValue(V);
+}
+
+void GlobalsModRef::copyValue(Value *From, Value *To) {
+ AliasAnalysis::copyValue(From, To);
+}
diff --git a/lib/Analysis/IPA/Makefile b/lib/Analysis/IPA/Makefile
new file mode 100644
index 0000000..b850c9f
--- /dev/null
+++ b/lib/Analysis/IPA/Makefile
@@ -0,0 +1,15 @@
+##===- lib/Analysis/IPA/Makefile ---------------------------*- Makefile -*-===##
+#
+# The LLVM Compiler Infrastructure
+#
+# This file is distributed under the University of Illinois Open Source
+# License. See LICENSE.TXT for details.
+#
+##===----------------------------------------------------------------------===##
+
+LEVEL = ../../..
+LIBRARYNAME = LLVMipa
+BUILD_ARCHIVE = 1
+
+include $(LEVEL)/Makefile.common
+
diff --git a/lib/Analysis/IVUsers.cpp b/lib/Analysis/IVUsers.cpp
new file mode 100644
index 0000000..9c472ae
--- /dev/null
+++ b/lib/Analysis/IVUsers.cpp
@@ -0,0 +1,411 @@
+//===- IVUsers.cpp - Induction Variable Users -------------------*- C++ -*-===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements bookkeeping for "interesting" users of expressions
+// computed from induction variables.
+//
+//===----------------------------------------------------------------------===//
+
+#define DEBUG_TYPE "iv-users"
+#include "llvm/Analysis/IVUsers.h"
+#include "llvm/Constants.h"
+#include "llvm/Instructions.h"
+#include "llvm/Type.h"
+#include "llvm/DerivedTypes.h"
+#include "llvm/Analysis/Dominators.h"
+#include "llvm/Analysis/LoopPass.h"
+#include "llvm/Analysis/ScalarEvolutionExpressions.h"
+#include "llvm/ADT/STLExtras.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+#include <algorithm>
+using namespace llvm;
+
+char IVUsers::ID = 0;
+static RegisterPass<IVUsers>
+X("iv-users", "Induction Variable Users", false, true);
+
+Pass *llvm::createIVUsersPass() {
+ return new IVUsers();
+}
+
+/// containsAddRecFromDifferentLoop - Determine whether expression S involves a
+/// subexpression that is an AddRec from a loop other than L. An outer loop
+/// of L is OK, but not an inner loop nor a disjoint loop.
+static bool containsAddRecFromDifferentLoop(const SCEV *S, Loop *L) {
+ // This is very common, put it first.
+ if (isa<SCEVConstant>(S))
+ return false;
+ if (const SCEVCommutativeExpr *AE = dyn_cast<SCEVCommutativeExpr>(S)) {
+ for (unsigned int i=0; i< AE->getNumOperands(); i++)
+ if (containsAddRecFromDifferentLoop(AE->getOperand(i), L))
+ return true;
+ return false;
+ }
+ if (const SCEVAddRecExpr *AE = dyn_cast<SCEVAddRecExpr>(S)) {
+ if (const Loop *newLoop = AE->getLoop()) {
+ if (newLoop == L)
+ return false;
+ // if newLoop is an outer loop of L, this is OK.
+ if (newLoop->contains(L))
+ return false;
+ }
+ return true;
+ }
+ if (const SCEVUDivExpr *DE = dyn_cast<SCEVUDivExpr>(S))
+ return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
+ containsAddRecFromDifferentLoop(DE->getRHS(), L);
+#if 0
+ // SCEVSDivExpr has been backed out temporarily, but will be back; we'll
+ // need this when it is.
+ if (const SCEVSDivExpr *DE = dyn_cast<SCEVSDivExpr>(S))
+ return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
+ containsAddRecFromDifferentLoop(DE->getRHS(), L);
+#endif
+ if (const SCEVCastExpr *CE = dyn_cast<SCEVCastExpr>(S))
+ return containsAddRecFromDifferentLoop(CE->getOperand(), L);
+ return false;
+}
+
+/// getSCEVStartAndStride - Compute the start and stride of this expression,
+/// returning false if the expression is not a start/stride pair, or true if it
+/// is. The stride must be a loop invariant expression, but the start may be
+/// a mix of loop invariant and loop variant expressions. The start cannot,
+/// however, contain an AddRec from a different loop, unless that loop is an
+/// outer loop of the current loop.
+static bool getSCEVStartAndStride(const SCEV *&SH, Loop *L, Loop *UseLoop,
+ const SCEV *&Start, const SCEV *&Stride,
+ ScalarEvolution *SE, DominatorTree *DT) {
+ const SCEV *TheAddRec = Start; // Initialize to zero.
+
+ // If the outer level is an AddExpr, the operands are all start values except
+ // for a nested AddRecExpr.
+ if (const SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
+ for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
+ if (const SCEVAddRecExpr *AddRec =
+ dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
+ if (AddRec->getLoop() == L)
+ TheAddRec = SE->getAddExpr(AddRec, TheAddRec);
+ else
+ return false; // Nested IV of some sort?
+ } else {
+ Start = SE->getAddExpr(Start, AE->getOperand(i));
+ }
+ } else if (isa<SCEVAddRecExpr>(SH)) {
+ TheAddRec = SH;
+ } else {
+ return false; // not analyzable.
+ }
+
+ const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
+ if (!AddRec || AddRec->getLoop() != L) return false;
+
+ // Use getSCEVAtScope to attempt to simplify other loops out of
+ // the picture.
+ const SCEV *AddRecStart = AddRec->getStart();
+ AddRecStart = SE->getSCEVAtScope(AddRecStart, UseLoop);
+ const SCEV *AddRecStride = AddRec->getStepRecurrence(*SE);
+
+ // FIXME: If Start contains an SCEVAddRecExpr from a different loop, other
+ // than an outer loop of the current loop, reject it. LSR has no concept of
+ // operating on more than one loop at a time so don't confuse it with such
+ // expressions.
+ if (containsAddRecFromDifferentLoop(AddRecStart, L))
+ return false;
+
+ Start = SE->getAddExpr(Start, AddRecStart);
+
+ // If stride is an instruction, make sure it properly dominates the header.
+ // Otherwise we could end up with a use before def situation.
+ if (!isa<SCEVConstant>(AddRecStride)) {
+ BasicBlock *Header = L->getHeader();
+ if (!AddRecStride->properlyDominates(Header, DT))
+ return false;
+
+ DEBUG(dbgs() << "[";
+ WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
+ dbgs() << "] Variable stride: " << *AddRec << "\n");
+ }
+
+ Stride = AddRecStride;
+ return true;
+}
+
+/// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
+/// and now we need to decide whether the user should use the preinc or post-inc
+/// value. If this user should use the post-inc version of the IV, return true.
+///
+/// Choosing wrong here can break dominance properties (if we choose to use the
+/// post-inc value when we cannot) or it can end up adding extra live-ranges to
+/// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
+/// should use the post-inc value).
+static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
+ Loop *L, LoopInfo *LI, DominatorTree *DT,
+ Pass *P) {
+ // If the user is in the loop, use the preinc value.
+ if (L->contains(User)) return false;
+
+ BasicBlock *LatchBlock = L->getLoopLatch();
+ if (!LatchBlock)
+ return false;
+
+ // Ok, the user is outside of the loop. If it is dominated by the latch
+ // block, use the post-inc value.
+ if (DT->dominates(LatchBlock, User->getParent()))
+ return true;
+
+ // There is one case we have to be careful of: PHI nodes. These little guys
+ // can live in blocks that are not dominated by the latch block, but (since
+ // their uses occur in the predecessor block, not the block the PHI lives in)
+ // should still use the post-inc value. Check for this case now.
+ PHINode *PN = dyn_cast<PHINode>(User);
+ if (!PN) return false; // not a phi, not dominated by latch block.
+
+ // Look at all of the uses of IV by the PHI node. If any use corresponds to
+ // a block that is not dominated by the latch block, give up and use the
+ // preincremented value.
+ unsigned NumUses = 0;
+ for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
+ if (PN->getIncomingValue(i) == IV) {
+ ++NumUses;
+ if (!DT->dominates(LatchBlock, PN->getIncomingBlock(i)))
+ return false;
+ }
+
+ // Okay, all uses of IV by PN are in predecessor blocks that really are
+ // dominated by the latch block. Use the post-incremented value.
+ return true;
+}
+
+/// AddUsersIfInteresting - Inspect the specified instruction. If it is a
+/// reducible SCEV, recursively add its users to the IVUsesByStride set and
+/// return true. Otherwise, return false.
+bool IVUsers::AddUsersIfInteresting(Instruction *I) {
+ if (!SE->isSCEVable(I->getType()))
+ return false; // Void and FP expressions cannot be reduced.
+
+ // LSR is not APInt clean, do not touch integers bigger than 64-bits.
+ if (SE->getTypeSizeInBits(I->getType()) > 64)
+ return false;
+
+ if (!Processed.insert(I))
+ return true; // Instruction already handled.
+
+ // Get the symbolic expression for this instruction.
+ const SCEV *ISE = SE->getSCEV(I);
+ if (isa<SCEVCouldNotCompute>(ISE)) return false;
+
+ // Get the start and stride for this expression.
+ Loop *UseLoop = LI->getLoopFor(I->getParent());
+ const SCEV *Start = SE->getIntegerSCEV(0, ISE->getType());
+ const SCEV *Stride = Start;
+
+ if (!getSCEVStartAndStride(ISE, L, UseLoop, Start, Stride, SE, DT))
+ return false; // Non-reducible symbolic expression, bail out.
+
+ // Keep things simple. Don't touch loop-variant strides.
+ if (!Stride->isLoopInvariant(L) && L->contains(I))
+ return false;
+
+ SmallPtrSet<Instruction *, 4> UniqueUsers;
+ for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
+ UI != E; ++UI) {
+ Instruction *User = cast<Instruction>(*UI);
+ if (!UniqueUsers.insert(User))
+ continue;
+
+ // Do not infinitely recurse on PHI nodes.
+ if (isa<PHINode>(User) && Processed.count(User))
+ continue;
+
+ // Descend recursively, but not into PHI nodes outside the current loop.
+ // It's important to see the entire expression outside the loop to get
+ // choices that depend on addressing mode use right, although we won't
+ // consider references ouside the loop in all cases.
+ // If User is already in Processed, we don't want to recurse into it again,
+ // but do want to record a second reference in the same instruction.
+ bool AddUserToIVUsers = false;
+ if (LI->getLoopFor(User->getParent()) != L) {
+ if (isa<PHINode>(User) || Processed.count(User) ||
+ !AddUsersIfInteresting(User)) {
+ DEBUG(dbgs() << "FOUND USER in other loop: " << *User << '\n'
+ << " OF SCEV: " << *ISE << '\n');
+ AddUserToIVUsers = true;
+ }
+ } else if (Processed.count(User) ||
+ !AddUsersIfInteresting(User)) {
+ DEBUG(dbgs() << "FOUND USER: " << *User << '\n'
+ << " OF SCEV: " << *ISE << '\n');
+ AddUserToIVUsers = true;
+ }
+
+ if (AddUserToIVUsers) {
+ IVUsersOfOneStride *StrideUses = IVUsesByStride[Stride];
+ if (!StrideUses) { // First occurrence of this stride?
+ StrideOrder.push_back(Stride);
+ StrideUses = new IVUsersOfOneStride(Stride);
+ IVUses.push_back(StrideUses);
+ IVUsesByStride[Stride] = StrideUses;
+ }
+
+ // Okay, we found a user that we cannot reduce. Analyze the instruction
+ // and decide what to do with it. If we are a use inside of the loop, use
+ // the value before incrementation, otherwise use it after incrementation.
+ if (IVUseShouldUsePostIncValue(User, I, L, LI, DT, this)) {
+ // The value used will be incremented by the stride more than we are
+ // expecting, so subtract this off.
+ const SCEV *NewStart = SE->getMinusSCEV(Start, Stride);
+ StrideUses->addUser(NewStart, User, I);
+ StrideUses->Users.back().setIsUseOfPostIncrementedValue(true);
+ DEBUG(dbgs() << " USING POSTINC SCEV, START=" << *NewStart<< "\n");
+ } else {
+ StrideUses->addUser(Start, User, I);
+ }
+ }
+ }
+ return true;
+}
+
+void IVUsers::AddUser(const SCEV *Stride, const SCEV *Offset,
+ Instruction *User, Value *Operand) {
+ IVUsersOfOneStride *StrideUses = IVUsesByStride[Stride];
+ if (!StrideUses) { // First occurrence of this stride?
+ StrideOrder.push_back(Stride);
+ StrideUses = new IVUsersOfOneStride(Stride);
+ IVUses.push_back(StrideUses);
+ IVUsesByStride[Stride] = StrideUses;
+ }
+ IVUsesByStride[Stride]->addUser(Offset, User, Operand);
+}
+
+IVUsers::IVUsers()
+ : LoopPass(&ID) {
+}
+
+void IVUsers::getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.addRequired<LoopInfo>();
+ AU.addRequired<DominatorTree>();
+ AU.addRequired<ScalarEvolution>();
+ AU.setPreservesAll();
+}
+
+bool IVUsers::runOnLoop(Loop *l, LPPassManager &LPM) {
+
+ L = l;
+ LI = &getAnalysis<LoopInfo>();
+ DT = &getAnalysis<DominatorTree>();
+ SE = &getAnalysis<ScalarEvolution>();
+
+ // Find all uses of induction variables in this loop, and categorize
+ // them by stride. Start by finding all of the PHI nodes in the header for
+ // this loop. If they are induction variables, inspect their uses.
+ for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
+ AddUsersIfInteresting(I);
+
+ return false;
+}
+
+/// getReplacementExpr - Return a SCEV expression which computes the
+/// value of the OperandValToReplace of the given IVStrideUse.
+const SCEV *IVUsers::getReplacementExpr(const IVStrideUse &U) const {
+ // Start with zero.
+ const SCEV *RetVal = SE->getIntegerSCEV(0, U.getParent()->Stride->getType());
+ // Create the basic add recurrence.
+ RetVal = SE->getAddRecExpr(RetVal, U.getParent()->Stride, L);
+ // Add the offset in a separate step, because it may be loop-variant.
+ RetVal = SE->getAddExpr(RetVal, U.getOffset());
+ // For uses of post-incremented values, add an extra stride to compute
+ // the actual replacement value.
+ if (U.isUseOfPostIncrementedValue())
+ RetVal = SE->getAddExpr(RetVal, U.getParent()->Stride);
+ return RetVal;
+}
+
+/// getCanonicalExpr - Return a SCEV expression which computes the
+/// value of the SCEV of the given IVStrideUse, ignoring the
+/// isUseOfPostIncrementedValue flag.
+const SCEV *IVUsers::getCanonicalExpr(const IVStrideUse &U) const {
+ // Start with zero.
+ const SCEV *RetVal = SE->getIntegerSCEV(0, U.getParent()->Stride->getType());
+ // Create the basic add recurrence.
+ RetVal = SE->getAddRecExpr(RetVal, U.getParent()->Stride, L);
+ // Add the offset in a separate step, because it may be loop-variant.
+ RetVal = SE->getAddExpr(RetVal, U.getOffset());
+ return RetVal;
+}
+
+void IVUsers::print(raw_ostream &OS, const Module *M) const {
+ OS << "IV Users for loop ";
+ WriteAsOperand(OS, L->getHeader(), false);
+ if (SE->hasLoopInvariantBackedgeTakenCount(L)) {
+ OS << " with backedge-taken count "
+ << *SE->getBackedgeTakenCount(L);
+ }
+ OS << ":\n";
+
+ for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
+ std::map<const SCEV *, IVUsersOfOneStride*>::const_iterator SI =
+ IVUsesByStride.find(StrideOrder[Stride]);
+ assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
+ OS << " Stride " << *SI->first->getType() << " " << *SI->first << ":\n";
+
+ for (ilist<IVStrideUse>::const_iterator UI = SI->second->Users.begin(),
+ E = SI->second->Users.end(); UI != E; ++UI) {
+ OS << " ";
+ WriteAsOperand(OS, UI->getOperandValToReplace(), false);
+ OS << " = ";
+ OS << *getReplacementExpr(*UI);
+ if (UI->isUseOfPostIncrementedValue())
+ OS << " (post-inc)";
+ OS << " in ";
+ UI->getUser()->print(OS);
+ OS << '\n';
+ }
+ }
+}
+
+void IVUsers::dump() const {
+ print(dbgs());
+}
+
+void IVUsers::releaseMemory() {
+ IVUsesByStride.clear();
+ StrideOrder.clear();
+ Processed.clear();
+ IVUses.clear();
+}
+
+void IVStrideUse::deleted() {
+ // Remove this user from the list.
+ Parent->Users.erase(this);
+ // this now dangles!
+}
+
+void IVUsersOfOneStride::print(raw_ostream &OS) const {
+ OS << "IV Users of one stride:\n";
+
+ if (Stride)
+ OS << " Stride: " << *Stride << '\n';
+
+ OS << " Users:\n";
+
+ unsigned Count = 1;
+
+ for (ilist<IVStrideUse>::const_iterator
+ I = Users.begin(), E = Users.end(); I != E; ++I) {
+ const IVStrideUse &SU = *I;
+ OS << " " << Count++ << '\n';
+ OS << " Offset: " << *SU.getOffset() << '\n';
+ OS << " Instr: " << *SU << '\n';
+ }
+}
+
+void IVUsersOfOneStride::dump() const {
+ print(dbgs());
+}
diff --git a/lib/Analysis/InlineCost.cpp b/lib/Analysis/InlineCost.cpp
new file mode 100644
index 0000000..972d034
--- /dev/null
+++ b/lib/Analysis/InlineCost.cpp
@@ -0,0 +1,390 @@
+//===- InlineCost.cpp - Cost analysis for inliner -------------------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements inline cost analysis.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/InlineCost.h"
+#include "llvm/Support/CallSite.h"
+#include "llvm/CallingConv.h"
+#include "llvm/IntrinsicInst.h"
+#include "llvm/ADT/SmallPtrSet.h"
+using namespace llvm;
+
+// CountCodeReductionForConstant - Figure out an approximation for how many
+// instructions will be constant folded if the specified value is constant.
+//
+unsigned InlineCostAnalyzer::FunctionInfo::
+ CountCodeReductionForConstant(Value *V) {
+ unsigned Reduction = 0;
+ for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
+ if (isa<BranchInst>(*UI) || isa<SwitchInst>(*UI)) {
+ // We will be able to eliminate all but one of the successors.
+ const TerminatorInst &TI = cast<TerminatorInst>(**UI);
+ const unsigned NumSucc = TI.getNumSuccessors();
+ unsigned Instrs = 0;
+ for (unsigned I = 0; I != NumSucc; ++I)
+ Instrs += TI.getSuccessor(I)->size();
+ // We don't know which blocks will be eliminated, so use the average size.
+ Reduction += InlineConstants::InstrCost*Instrs*(NumSucc-1)/NumSucc;
+ } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
+ // Turning an indirect call into a direct call is a BIG win
+ if (CI->getCalledValue() == V)
+ Reduction += InlineConstants::IndirectCallBonus;
+ } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
+ // Turning an indirect call into a direct call is a BIG win
+ if (II->getCalledValue() == V)
+ Reduction += InlineConstants::IndirectCallBonus;
+ } else {
+ // Figure out if this instruction will be removed due to simple constant
+ // propagation.
+ Instruction &Inst = cast<Instruction>(**UI);
+
+ // We can't constant propagate instructions which have effects or
+ // read memory.
+ //
+ // FIXME: It would be nice to capture the fact that a load from a
+ // pointer-to-constant-global is actually a *really* good thing to zap.
+ // Unfortunately, we don't know the pointer that may get propagated here,
+ // so we can't make this decision.
+ if (Inst.mayReadFromMemory() || Inst.mayHaveSideEffects() ||
+ isa<AllocaInst>(Inst))
+ continue;
+
+ bool AllOperandsConstant = true;
+ for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
+ if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
+ AllOperandsConstant = false;
+ break;
+ }
+
+ if (AllOperandsConstant) {
+ // We will get to remove this instruction...
+ Reduction += InlineConstants::InstrCost;
+
+ // And any other instructions that use it which become constants
+ // themselves.
+ Reduction += CountCodeReductionForConstant(&Inst);
+ }
+ }
+
+ return Reduction;
+}
+
+// CountCodeReductionForAlloca - Figure out an approximation of how much smaller
+// the function will be if it is inlined into a context where an argument
+// becomes an alloca.
+//
+unsigned InlineCostAnalyzer::FunctionInfo::
+ CountCodeReductionForAlloca(Value *V) {
+ if (!isa<PointerType>(V->getType())) return 0; // Not a pointer
+ unsigned Reduction = 0;
+ for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
+ Instruction *I = cast<Instruction>(*UI);
+ if (isa<LoadInst>(I) || isa<StoreInst>(I))
+ Reduction += InlineConstants::InstrCost;
+ else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
+ // If the GEP has variable indices, we won't be able to do much with it.
+ if (GEP->hasAllConstantIndices())
+ Reduction += CountCodeReductionForAlloca(GEP);
+ } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(I)) {
+ // Track pointer through bitcasts.
+ Reduction += CountCodeReductionForAlloca(BCI);
+ } else {
+ // If there is some other strange instruction, we're not going to be able
+ // to do much if we inline this.
+ return 0;
+ }
+ }
+
+ return Reduction;
+}
+
+// callIsSmall - If a call is likely to lower to a single target instruction, or
+// is otherwise deemed small return true.
+// TODO: Perhaps calls like memcpy, strcpy, etc?
+static bool callIsSmall(const Function *F) {
+ if (!F) return false;
+
+ if (F->hasLocalLinkage()) return false;
+
+ if (!F->hasName()) return false;
+
+ StringRef Name = F->getName();
+
+ // These will all likely lower to a single selection DAG node.
+ if (Name == "copysign" || Name == "copysignf" ||
+ Name == "fabs" || Name == "fabsf" || Name == "fabsl" ||
+ Name == "sin" || Name == "sinf" || Name == "sinl" ||
+ Name == "cos" || Name == "cosf" || Name == "cosl" ||
+ Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl" )
+ return true;
+
+ // These are all likely to be optimized into something smaller.
+ if (Name == "pow" || Name == "powf" || Name == "powl" ||
+ Name == "exp2" || Name == "exp2l" || Name == "exp2f" ||
+ Name == "floor" || Name == "floorf" || Name == "ceil" ||
+ Name == "round" || Name == "ffs" || Name == "ffsl" ||
+ Name == "abs" || Name == "labs" || Name == "llabs")
+ return true;
+
+ return false;
+}
+
+/// analyzeBasicBlock - Fill in the current structure with information gleaned
+/// from the specified block.
+void CodeMetrics::analyzeBasicBlock(const BasicBlock *BB) {
+ ++NumBlocks;
+
+ for (BasicBlock::const_iterator II = BB->begin(), E = BB->end();
+ II != E; ++II) {
+ if (isa<PHINode>(II)) continue; // PHI nodes don't count.
+
+ // Special handling for calls.
+ if (isa<CallInst>(II) || isa<InvokeInst>(II)) {
+ if (isa<DbgInfoIntrinsic>(II))
+ continue; // Debug intrinsics don't count as size.
+
+ CallSite CS = CallSite::get(const_cast<Instruction*>(&*II));
+
+ // If this function contains a call to setjmp or _setjmp, never inline
+ // it. This is a hack because we depend on the user marking their local
+ // variables as volatile if they are live across a setjmp call, and they
+ // probably won't do this in callers.
+ if (Function *F = CS.getCalledFunction())
+ if (F->isDeclaration() &&
+ (F->getName() == "setjmp" || F->getName() == "_setjmp"))
+ NeverInline = true;
+
+ if (!isa<IntrinsicInst>(II) && !callIsSmall(CS.getCalledFunction())) {
+ // Each argument to a call takes on average one instruction to set up.
+ NumInsts += CS.arg_size();
+ ++NumCalls;
+ }
+ }
+
+ if (const AllocaInst *AI = dyn_cast<AllocaInst>(II)) {
+ if (!AI->isStaticAlloca())
+ this->usesDynamicAlloca = true;
+ }
+
+ if (isa<ExtractElementInst>(II) || isa<VectorType>(II->getType()))
+ ++NumVectorInsts;
+
+ if (const CastInst *CI = dyn_cast<CastInst>(II)) {
+ // Noop casts, including ptr <-> int, don't count.
+ if (CI->isLosslessCast() || isa<IntToPtrInst>(CI) ||
+ isa<PtrToIntInst>(CI))
+ continue;
+ // Result of a cmp instruction is often extended (to be used by other
+ // cmp instructions, logical or return instructions). These are usually
+ // nop on most sane targets.
+ if (isa<CmpInst>(CI->getOperand(0)))
+ continue;
+ } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(II)){
+ // If a GEP has all constant indices, it will probably be folded with
+ // a load/store.
+ if (GEPI->hasAllConstantIndices())
+ continue;
+ }
+
+ ++NumInsts;
+ }
+
+ if (isa<ReturnInst>(BB->getTerminator()))
+ ++NumRets;
+
+ // We never want to inline functions that contain an indirectbr. This is
+ // incorrect because all the blockaddress's (in static global initializers
+ // for example) would be referring to the original function, and this indirect
+ // jump would jump from the inlined copy of the function into the original
+ // function which is extremely undefined behavior.
+ if (isa<IndirectBrInst>(BB->getTerminator()))
+ NeverInline = true;
+}
+
+/// analyzeFunction - Fill in the current structure with information gleaned
+/// from the specified function.
+void CodeMetrics::analyzeFunction(Function *F) {
+ // Look at the size of the callee.
+ for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
+ analyzeBasicBlock(&*BB);
+}
+
+/// analyzeFunction - Fill in the current structure with information gleaned
+/// from the specified function.
+void InlineCostAnalyzer::FunctionInfo::analyzeFunction(Function *F) {
+ Metrics.analyzeFunction(F);
+
+ // A function with exactly one return has it removed during the inlining
+ // process (see InlineFunction), so don't count it.
+ // FIXME: This knowledge should really be encoded outside of FunctionInfo.
+ if (Metrics.NumRets==1)
+ --Metrics.NumInsts;
+
+ // Don't bother calculating argument weights if we are never going to inline
+ // the function anyway.
+ if (Metrics.NeverInline)
+ return;
+
+ // Check out all of the arguments to the function, figuring out how much
+ // code can be eliminated if one of the arguments is a constant.
+ ArgumentWeights.reserve(F->arg_size());
+ for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
+ ArgumentWeights.push_back(ArgInfo(CountCodeReductionForConstant(I),
+ CountCodeReductionForAlloca(I)));
+}
+
+// getInlineCost - The heuristic used to determine if we should inline the
+// function call or not.
+//
+InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS,
+ SmallPtrSet<const Function *, 16> &NeverInline) {
+ Instruction *TheCall = CS.getInstruction();
+ Function *Callee = CS.getCalledFunction();
+ Function *Caller = TheCall->getParent()->getParent();
+
+ // Don't inline functions which can be redefined at link-time to mean
+ // something else. Don't inline functions marked noinline.
+ if (Callee->mayBeOverridden() ||
+ Callee->hasFnAttr(Attribute::NoInline) || NeverInline.count(Callee))
+ return llvm::InlineCost::getNever();
+
+ // InlineCost - This value measures how good of an inline candidate this call
+ // site is to inline. A lower inline cost make is more likely for the call to
+ // be inlined. This value may go negative.
+ //
+ int InlineCost = 0;
+
+ // If there is only one call of the function, and it has internal linkage,
+ // make it almost guaranteed to be inlined.
+ //
+ if (Callee->hasLocalLinkage() && Callee->hasOneUse())
+ InlineCost += InlineConstants::LastCallToStaticBonus;
+
+ // If this function uses the coldcc calling convention, prefer not to inline
+ // it.
+ if (Callee->getCallingConv() == CallingConv::Cold)
+ InlineCost += InlineConstants::ColdccPenalty;
+
+ // If the instruction after the call, or if the normal destination of the
+ // invoke is an unreachable instruction, the function is noreturn. As such,
+ // there is little point in inlining this.
+ if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
+ if (isa<UnreachableInst>(II->getNormalDest()->begin()))
+ InlineCost += InlineConstants::NoreturnPenalty;
+ } else if (isa<UnreachableInst>(++BasicBlock::iterator(TheCall)))
+ InlineCost += InlineConstants::NoreturnPenalty;
+
+ // Get information about the callee...
+ FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
+
+ // If we haven't calculated this information yet, do so now.
+ if (CalleeFI.Metrics.NumBlocks == 0)
+ CalleeFI.analyzeFunction(Callee);
+
+ // If we should never inline this, return a huge cost.
+ if (CalleeFI.Metrics.NeverInline)
+ return InlineCost::getNever();
+
+ // FIXME: It would be nice to kill off CalleeFI.NeverInline. Then we
+ // could move this up and avoid computing the FunctionInfo for
+ // things we are going to just return always inline for. This
+ // requires handling setjmp somewhere else, however.
+ if (!Callee->isDeclaration() && Callee->hasFnAttr(Attribute::AlwaysInline))
+ return InlineCost::getAlways();
+
+ if (CalleeFI.Metrics.usesDynamicAlloca) {
+ // Get infomation about the caller...
+ FunctionInfo &CallerFI = CachedFunctionInfo[Caller];
+
+ // If we haven't calculated this information yet, do so now.
+ if (CallerFI.Metrics.NumBlocks == 0)
+ CallerFI.analyzeFunction(Caller);
+
+ // Don't inline a callee with dynamic alloca into a caller without them.
+ // Functions containing dynamic alloca's are inefficient in various ways;
+ // don't create more inefficiency.
+ if (!CallerFI.Metrics.usesDynamicAlloca)
+ return InlineCost::getNever();
+ }
+
+ // Add to the inline quality for properties that make the call valuable to
+ // inline. This includes factors that indicate that the result of inlining
+ // the function will be optimizable. Currently this just looks at arguments
+ // passed into the function.
+ //
+ unsigned ArgNo = 0;
+ for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
+ I != E; ++I, ++ArgNo) {
+ // Each argument passed in has a cost at both the caller and the callee
+ // sides. Measurements show that each argument costs about the same as an
+ // instruction.
+ InlineCost -= InlineConstants::InstrCost;
+
+ // If an alloca is passed in, inlining this function is likely to allow
+ // significant future optimization possibilities (like scalar promotion, and
+ // scalarization), so encourage the inlining of the function.
+ //
+ if (isa<AllocaInst>(I)) {
+ if (ArgNo < CalleeFI.ArgumentWeights.size())
+ InlineCost -= CalleeFI.ArgumentWeights[ArgNo].AllocaWeight;
+
+ // If this is a constant being passed into the function, use the argument
+ // weights calculated for the callee to determine how much will be folded
+ // away with this information.
+ } else if (isa<Constant>(I)) {
+ if (ArgNo < CalleeFI.ArgumentWeights.size())
+ InlineCost -= CalleeFI.ArgumentWeights[ArgNo].ConstantWeight;
+ }
+ }
+
+ // Now that we have considered all of the factors that make the call site more
+ // likely to be inlined, look at factors that make us not want to inline it.
+
+ // Calls usually take a long time, so they make the inlining gain smaller.
+ InlineCost += CalleeFI.Metrics.NumCalls * InlineConstants::CallPenalty;
+
+ // Don't inline into something too big, which would make it bigger.
+ // "size" here is the number of basic blocks, not instructions.
+ //
+ InlineCost += Caller->size()/15;
+
+ // Look at the size of the callee. Each instruction counts as 5.
+ InlineCost += CalleeFI.Metrics.NumInsts*InlineConstants::InstrCost;
+
+ return llvm::InlineCost::get(InlineCost);
+}
+
+// getInlineFudgeFactor - Return a > 1.0 factor if the inliner should use a
+// higher threshold to determine if the function call should be inlined.
+float InlineCostAnalyzer::getInlineFudgeFactor(CallSite CS) {
+ Function *Callee = CS.getCalledFunction();
+
+ // Get information about the callee...
+ FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
+
+ // If we haven't calculated this information yet, do so now.
+ if (CalleeFI.Metrics.NumBlocks == 0)
+ CalleeFI.analyzeFunction(Callee);
+
+ float Factor = 1.0f;
+ // Single BB functions are often written to be inlined.
+ if (CalleeFI.Metrics.NumBlocks == 1)
+ Factor += 0.5f;
+
+ // Be more aggressive if the function contains a good chunk (if it mades up
+ // at least 10% of the instructions) of vector instructions.
+ if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/2)
+ Factor += 2.0f;
+ else if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/10)
+ Factor += 1.5f;
+ return Factor;
+}
diff --git a/lib/Analysis/InstCount.cpp b/lib/Analysis/InstCount.cpp
new file mode 100644
index 0000000..bb2cf53
--- /dev/null
+++ b/lib/Analysis/InstCount.cpp
@@ -0,0 +1,85 @@
+//===-- InstCount.cpp - Collects the count of all instructions ------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This pass collects the count of all instructions and reports them
+//
+//===----------------------------------------------------------------------===//
+
+#define DEBUG_TYPE "instcount"
+#include "llvm/Analysis/Passes.h"
+#include "llvm/Pass.h"
+#include "llvm/Function.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/InstVisitor.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/ADT/Statistic.h"
+using namespace llvm;
+
+STATISTIC(TotalInsts , "Number of instructions (of all types)");
+STATISTIC(TotalBlocks, "Number of basic blocks");
+STATISTIC(TotalFuncs , "Number of non-external functions");
+STATISTIC(TotalMemInst, "Number of memory instructions");
+
+#define HANDLE_INST(N, OPCODE, CLASS) \
+ STATISTIC(Num ## OPCODE ## Inst, "Number of " #OPCODE " insts");
+
+#include "llvm/Instruction.def"
+
+
+namespace {
+ class InstCount : public FunctionPass, public InstVisitor<InstCount> {
+ friend class InstVisitor<InstCount>;
+
+ void visitFunction (Function &F) { ++TotalFuncs; }
+ void visitBasicBlock(BasicBlock &BB) { ++TotalBlocks; }
+
+#define HANDLE_INST(N, OPCODE, CLASS) \
+ void visit##OPCODE(CLASS &) { ++Num##OPCODE##Inst; ++TotalInsts; }
+
+#include "llvm/Instruction.def"
+
+ void visitInstruction(Instruction &I) {
+ errs() << "Instruction Count does not know about " << I;
+ llvm_unreachable(0);
+ }
+ public:
+ static char ID; // Pass identification, replacement for typeid
+ InstCount() : FunctionPass(&ID) {}
+
+ virtual bool runOnFunction(Function &F);
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.setPreservesAll();
+ }
+ virtual void print(raw_ostream &O, const Module *M) const {}
+
+ };
+}
+
+char InstCount::ID = 0;
+static RegisterPass<InstCount>
+X("instcount", "Counts the various types of Instructions", false, true);
+
+FunctionPass *llvm::createInstCountPass() { return new InstCount(); }
+
+// InstCount::run - This is the main Analysis entry point for a
+// function.
+//
+bool InstCount::runOnFunction(Function &F) {
+ unsigned StartMemInsts =
+ NumGetElementPtrInst + NumLoadInst + NumStoreInst + NumCallInst +
+ NumInvokeInst + NumAllocaInst;
+ visit(F);
+ unsigned EndMemInsts =
+ NumGetElementPtrInst + NumLoadInst + NumStoreInst + NumCallInst +
+ NumInvokeInst + NumAllocaInst;
+ TotalMemInst += EndMemInsts-StartMemInsts;
+ return false;
+}
diff --git a/lib/Analysis/InstructionSimplify.cpp b/lib/Analysis/InstructionSimplify.cpp
new file mode 100644
index 0000000..b53ac13
--- /dev/null
+++ b/lib/Analysis/InstructionSimplify.cpp
@@ -0,0 +1,409 @@
+//===- InstructionSimplify.cpp - Fold instruction operands ----------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements routines for folding instructions into simpler forms
+// that do not require creating new instructions. For example, this does
+// constant folding, and can handle identities like (X&0)->0.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/InstructionSimplify.h"
+#include "llvm/Analysis/ConstantFolding.h"
+#include "llvm/Support/ValueHandle.h"
+#include "llvm/Instructions.h"
+#include "llvm/Support/PatternMatch.h"
+using namespace llvm;
+using namespace llvm::PatternMatch;
+
+/// SimplifyAddInst - Given operands for an Add, see if we can
+/// fold the result. If not, this returns null.
+Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
+ const TargetData *TD) {
+ if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
+ if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
+ Constant *Ops[] = { CLHS, CRHS };
+ return ConstantFoldInstOperands(Instruction::Add, CLHS->getType(),
+ Ops, 2, TD);
+ }
+
+ // Canonicalize the constant to the RHS.
+ std::swap(Op0, Op1);
+ }
+
+ if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
+ // X + undef -> undef
+ if (isa<UndefValue>(Op1C))
+ return Op1C;
+
+ // X + 0 --> X
+ if (Op1C->isNullValue())
+ return Op0;
+ }
+
+ // FIXME: Could pull several more out of instcombine.
+ return 0;
+}
+
+/// SimplifyAndInst - Given operands for an And, see if we can
+/// fold the result. If not, this returns null.
+Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const TargetData *TD) {
+ if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
+ if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
+ Constant *Ops[] = { CLHS, CRHS };
+ return ConstantFoldInstOperands(Instruction::And, CLHS->getType(),
+ Ops, 2, TD);
+ }
+
+ // Canonicalize the constant to the RHS.
+ std::swap(Op0, Op1);
+ }
+
+ // X & undef -> 0
+ if (isa<UndefValue>(Op1))
+ return Constant::getNullValue(Op0->getType());
+
+ // X & X = X
+ if (Op0 == Op1)
+ return Op0;
+
+ // X & <0,0> = <0,0>
+ if (isa<ConstantAggregateZero>(Op1))
+ return Op1;
+
+ // X & <-1,-1> = X
+ if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1))
+ if (CP->isAllOnesValue())
+ return Op0;
+
+ if (ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1)) {
+ // X & 0 = 0
+ if (Op1CI->isZero())
+ return Op1CI;
+ // X & -1 = X
+ if (Op1CI->isAllOnesValue())
+ return Op0;
+ }
+
+ // A & ~A = ~A & A = 0
+ Value *A, *B;
+ if ((match(Op0, m_Not(m_Value(A))) && A == Op1) ||
+ (match(Op1, m_Not(m_Value(A))) && A == Op0))
+ return Constant::getNullValue(Op0->getType());
+
+ // (A | ?) & A = A
+ if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
+ (A == Op1 || B == Op1))
+ return Op1;
+
+ // A & (A | ?) = A
+ if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
+ (A == Op0 || B == Op0))
+ return Op0;
+
+ return 0;
+}
+
+/// SimplifyOrInst - Given operands for an Or, see if we can
+/// fold the result. If not, this returns null.
+Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const TargetData *TD) {
+ if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
+ if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
+ Constant *Ops[] = { CLHS, CRHS };
+ return ConstantFoldInstOperands(Instruction::Or, CLHS->getType(),
+ Ops, 2, TD);
+ }
+
+ // Canonicalize the constant to the RHS.
+ std::swap(Op0, Op1);
+ }
+
+ // X | undef -> -1
+ if (isa<UndefValue>(Op1))
+ return Constant::getAllOnesValue(Op0->getType());
+
+ // X | X = X
+ if (Op0 == Op1)
+ return Op0;
+
+ // X | <0,0> = X
+ if (isa<ConstantAggregateZero>(Op1))
+ return Op0;
+
+ // X | <-1,-1> = <-1,-1>
+ if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1))
+ if (CP->isAllOnesValue())
+ return Op1;
+
+ if (ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1)) {
+ // X | 0 = X
+ if (Op1CI->isZero())
+ return Op0;
+ // X | -1 = -1
+ if (Op1CI->isAllOnesValue())
+ return Op1CI;
+ }
+
+ // A | ~A = ~A | A = -1
+ Value *A, *B;
+ if ((match(Op0, m_Not(m_Value(A))) && A == Op1) ||
+ (match(Op1, m_Not(m_Value(A))) && A == Op0))
+ return Constant::getAllOnesValue(Op0->getType());
+
+ // (A & ?) | A = A
+ if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
+ (A == Op1 || B == Op1))
+ return Op1;
+
+ // A | (A & ?) = A
+ if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
+ (A == Op0 || B == Op0))
+ return Op0;
+
+ return 0;
+}
+
+
+static const Type *GetCompareTy(Value *Op) {
+ return CmpInst::makeCmpResultType(Op->getType());
+}
+
+
+/// SimplifyICmpInst - Given operands for an ICmpInst, see if we can
+/// fold the result. If not, this returns null.
+Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
+ const TargetData *TD) {
+ CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
+ assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
+
+ if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
+ if (Constant *CRHS = dyn_cast<Constant>(RHS))
+ return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, TD);
+
+ // If we have a constant, make sure it is on the RHS.
+ std::swap(LHS, RHS);
+ Pred = CmpInst::getSwappedPredicate(Pred);
+ }
+
+ // ITy - This is the return type of the compare we're considering.
+ const Type *ITy = GetCompareTy(LHS);
+
+ // icmp X, X -> true/false
+ if (LHS == RHS)
+ return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
+
+ if (isa<UndefValue>(RHS)) // X icmp undef -> undef
+ return UndefValue::get(ITy);
+
+ // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
+ // addresses never equal each other! We already know that Op0 != Op1.
+ if ((isa<GlobalValue>(LHS) || isa<AllocaInst>(LHS) ||
+ isa<ConstantPointerNull>(LHS)) &&
+ (isa<GlobalValue>(RHS) || isa<AllocaInst>(RHS) ||
+ isa<ConstantPointerNull>(RHS)))
+ return ConstantInt::get(ITy, CmpInst::isFalseWhenEqual(Pred));
+
+ // See if we are doing a comparison with a constant.
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
+ // If we have an icmp le or icmp ge instruction, turn it into the
+ // appropriate icmp lt or icmp gt instruction. This allows us to rely on
+ // them being folded in the code below.
+ switch (Pred) {
+ default: break;
+ case ICmpInst::ICMP_ULE:
+ if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
+ return ConstantInt::getTrue(CI->getContext());
+ break;
+ case ICmpInst::ICMP_SLE:
+ if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
+ return ConstantInt::getTrue(CI->getContext());
+ break;
+ case ICmpInst::ICMP_UGE:
+ if (CI->isMinValue(false)) // A >=u MIN -> TRUE
+ return ConstantInt::getTrue(CI->getContext());
+ break;
+ case ICmpInst::ICMP_SGE:
+ if (CI->isMinValue(true)) // A >=s MIN -> TRUE
+ return ConstantInt::getTrue(CI->getContext());
+ break;
+ }
+ }
+
+
+ return 0;
+}
+
+/// SimplifyFCmpInst - Given operands for an FCmpInst, see if we can
+/// fold the result. If not, this returns null.
+Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
+ const TargetData *TD) {
+ CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
+ assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
+
+ if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
+ if (Constant *CRHS = dyn_cast<Constant>(RHS))
+ return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, TD);
+
+ // If we have a constant, make sure it is on the RHS.
+ std::swap(LHS, RHS);
+ Pred = CmpInst::getSwappedPredicate(Pred);
+ }
+
+ // Fold trivial predicates.
+ if (Pred == FCmpInst::FCMP_FALSE)
+ return ConstantInt::get(GetCompareTy(LHS), 0);
+ if (Pred == FCmpInst::FCMP_TRUE)
+ return ConstantInt::get(GetCompareTy(LHS), 1);
+
+ if (isa<UndefValue>(RHS)) // fcmp pred X, undef -> undef
+ return UndefValue::get(GetCompareTy(LHS));
+
+ // fcmp x,x -> true/false. Not all compares are foldable.
+ if (LHS == RHS) {
+ if (CmpInst::isTrueWhenEqual(Pred))
+ return ConstantInt::get(GetCompareTy(LHS), 1);
+ if (CmpInst::isFalseWhenEqual(Pred))
+ return ConstantInt::get(GetCompareTy(LHS), 0);
+ }
+
+ // Handle fcmp with constant RHS
+ if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
+ // If the constant is a nan, see if we can fold the comparison based on it.
+ if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
+ if (CFP->getValueAPF().isNaN()) {
+ if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo"
+ return ConstantInt::getFalse(CFP->getContext());
+ assert(FCmpInst::isUnordered(Pred) &&
+ "Comparison must be either ordered or unordered!");
+ // True if unordered.
+ return ConstantInt::getTrue(CFP->getContext());
+ }
+ }
+ }
+
+ return 0;
+}
+
+/// SimplifyGEPInst - Given operands for an GetElementPtrInst, see if we can
+/// fold the result. If not, this returns null.
+Value *llvm::SimplifyGEPInst(Value *const *Ops, unsigned NumOps,
+ const TargetData *TD) {
+ // getelementptr P -> P.
+ if (NumOps == 1)
+ return Ops[0];
+
+ // TODO.
+ //if (isa<UndefValue>(Ops[0]))
+ // return UndefValue::get(GEP.getType());
+
+ // getelementptr P, 0 -> P.
+ if (NumOps == 2)
+ if (ConstantInt *C = dyn_cast<ConstantInt>(Ops[1]))
+ if (C->isZero())
+ return Ops[0];
+
+ // Check to see if this is constant foldable.
+ for (unsigned i = 0; i != NumOps; ++i)
+ if (!isa<Constant>(Ops[i]))
+ return 0;
+
+ return ConstantExpr::getGetElementPtr(cast<Constant>(Ops[0]),
+ (Constant *const*)Ops+1, NumOps-1);
+}
+
+
+//=== Helper functions for higher up the class hierarchy.
+
+/// SimplifyBinOp - Given operands for a BinaryOperator, see if we can
+/// fold the result. If not, this returns null.
+Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
+ const TargetData *TD) {
+ switch (Opcode) {
+ case Instruction::And: return SimplifyAndInst(LHS, RHS, TD);
+ case Instruction::Or: return SimplifyOrInst(LHS, RHS, TD);
+ default:
+ if (Constant *CLHS = dyn_cast<Constant>(LHS))
+ if (Constant *CRHS = dyn_cast<Constant>(RHS)) {
+ Constant *COps[] = {CLHS, CRHS};
+ return ConstantFoldInstOperands(Opcode, LHS->getType(), COps, 2, TD);
+ }
+ return 0;
+ }
+}
+
+/// SimplifyCmpInst - Given operands for a CmpInst, see if we can
+/// fold the result.
+Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
+ const TargetData *TD) {
+ if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
+ return SimplifyICmpInst(Predicate, LHS, RHS, TD);
+ return SimplifyFCmpInst(Predicate, LHS, RHS, TD);
+}
+
+
+/// SimplifyInstruction - See if we can compute a simplified version of this
+/// instruction. If not, this returns null.
+Value *llvm::SimplifyInstruction(Instruction *I, const TargetData *TD) {
+ switch (I->getOpcode()) {
+ default:
+ return ConstantFoldInstruction(I, TD);
+ case Instruction::Add:
+ return SimplifyAddInst(I->getOperand(0), I->getOperand(1),
+ cast<BinaryOperator>(I)->hasNoSignedWrap(),
+ cast<BinaryOperator>(I)->hasNoUnsignedWrap(), TD);
+ case Instruction::And:
+ return SimplifyAndInst(I->getOperand(0), I->getOperand(1), TD);
+ case Instruction::Or:
+ return SimplifyOrInst(I->getOperand(0), I->getOperand(1), TD);
+ case Instruction::ICmp:
+ return SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(),
+ I->getOperand(0), I->getOperand(1), TD);
+ case Instruction::FCmp:
+ return SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(),
+ I->getOperand(0), I->getOperand(1), TD);
+ case Instruction::GetElementPtr: {
+ SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
+ return SimplifyGEPInst(&Ops[0], Ops.size(), TD);
+ }
+ }
+}
+
+/// ReplaceAndSimplifyAllUses - Perform From->replaceAllUsesWith(To) and then
+/// delete the From instruction. In addition to a basic RAUW, this does a
+/// recursive simplification of the newly formed instructions. This catches
+/// things where one simplification exposes other opportunities. This only
+/// simplifies and deletes scalar operations, it does not change the CFG.
+///
+void llvm::ReplaceAndSimplifyAllUses(Instruction *From, Value *To,
+ const TargetData *TD) {
+ assert(From != To && "ReplaceAndSimplifyAllUses(X,X) is not valid!");
+
+ // FromHandle - This keeps a weakvh on the from value so that we can know if
+ // it gets deleted out from under us in a recursive simplification.
+ WeakVH FromHandle(From);
+
+ while (!From->use_empty()) {
+ // Update the instruction to use the new value.
+ Use &U = From->use_begin().getUse();
+ Instruction *User = cast<Instruction>(U.getUser());
+ U = To;
+
+ // See if we can simplify it.
+ if (Value *V = SimplifyInstruction(User, TD)) {
+ // Recursively simplify this.
+ ReplaceAndSimplifyAllUses(User, V, TD);
+
+ // If the recursive simplification ended up revisiting and deleting 'From'
+ // then we're done.
+ if (FromHandle == 0)
+ return;
+ }
+ }
+ From->eraseFromParent();
+}
+
diff --git a/lib/Analysis/Interval.cpp b/lib/Analysis/Interval.cpp
new file mode 100644
index 0000000..ca9cdca
--- /dev/null
+++ b/lib/Analysis/Interval.cpp
@@ -0,0 +1,58 @@
+//===- Interval.cpp - Interval class code ---------------------------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file contains the definition of the Interval class, which represents a
+// partition of a control flow graph of some kind.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/Interval.h"
+#include "llvm/BasicBlock.h"
+#include "llvm/Support/CFG.h"
+#include "llvm/Support/raw_ostream.h"
+#include <algorithm>
+
+using namespace llvm;
+
+//===----------------------------------------------------------------------===//
+// Interval Implementation
+//===----------------------------------------------------------------------===//
+
+// isLoop - Find out if there is a back edge in this interval...
+//
+bool Interval::isLoop() const {
+ // There is a loop in this interval iff one of the predecessors of the header
+ // node lives in the interval.
+ for (::pred_iterator I = ::pred_begin(HeaderNode), E = ::pred_end(HeaderNode);
+ I != E; ++I)
+ if (contains(*I))
+ return true;
+ return false;
+}
+
+
+void Interval::print(raw_ostream &OS) const {
+ OS << "-------------------------------------------------------------\n"
+ << "Interval Contents:\n";
+
+ // Print out all of the basic blocks in the interval...
+ for (std::vector<BasicBlock*>::const_iterator I = Nodes.begin(),
+ E = Nodes.end(); I != E; ++I)
+ OS << **I << "\n";
+
+ OS << "Interval Predecessors:\n";
+ for (std::vector<BasicBlock*>::const_iterator I = Predecessors.begin(),
+ E = Predecessors.end(); I != E; ++I)
+ OS << **I << "\n";
+
+ OS << "Interval Successors:\n";
+ for (std::vector<BasicBlock*>::const_iterator I = Successors.begin(),
+ E = Successors.end(); I != E; ++I)
+ OS << **I << "\n";
+}
diff --git a/lib/Analysis/IntervalPartition.cpp b/lib/Analysis/IntervalPartition.cpp
new file mode 100644
index 0000000..1f17b77
--- /dev/null
+++ b/lib/Analysis/IntervalPartition.cpp
@@ -0,0 +1,114 @@
+//===- IntervalPartition.cpp - Interval Partition module code -------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file contains the definition of the IntervalPartition class, which
+// calculates and represent the interval partition of a function.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/IntervalIterator.h"
+using namespace llvm;
+
+char IntervalPartition::ID = 0;
+static RegisterPass<IntervalPartition>
+X("intervals", "Interval Partition Construction", true, true);
+
+//===----------------------------------------------------------------------===//
+// IntervalPartition Implementation
+//===----------------------------------------------------------------------===//
+
+// releaseMemory - Reset state back to before function was analyzed
+void IntervalPartition::releaseMemory() {
+ for (unsigned i = 0, e = Intervals.size(); i != e; ++i)
+ delete Intervals[i];
+ IntervalMap.clear();
+ Intervals.clear();
+ RootInterval = 0;
+}
+
+void IntervalPartition::print(raw_ostream &O, const Module*) const {
+ for(unsigned i = 0, e = Intervals.size(); i != e; ++i)
+ Intervals[i]->print(O);
+}
+
+// addIntervalToPartition - Add an interval to the internal list of intervals,
+// and then add mappings from all of the basic blocks in the interval to the
+// interval itself (in the IntervalMap).
+//
+void IntervalPartition::addIntervalToPartition(Interval *I) {
+ Intervals.push_back(I);
+
+ // Add mappings for all of the basic blocks in I to the IntervalPartition
+ for (Interval::node_iterator It = I->Nodes.begin(), End = I->Nodes.end();
+ It != End; ++It)
+ IntervalMap.insert(std::make_pair(*It, I));
+}
+
+// updatePredecessors - Interval generation only sets the successor fields of
+// the interval data structures. After interval generation is complete,
+// run through all of the intervals and propagate successor info as
+// predecessor info.
+//
+void IntervalPartition::updatePredecessors(Interval *Int) {
+ BasicBlock *Header = Int->getHeaderNode();
+ for (Interval::succ_iterator I = Int->Successors.begin(),
+ E = Int->Successors.end(); I != E; ++I)
+ getBlockInterval(*I)->Predecessors.push_back(Header);
+}
+
+// IntervalPartition ctor - Build the first level interval partition for the
+// specified function...
+//
+bool IntervalPartition::runOnFunction(Function &F) {
+ // Pass false to intervals_begin because we take ownership of it's memory
+ function_interval_iterator I = intervals_begin(&F, false);
+ assert(I != intervals_end(&F) && "No intervals in function!?!?!");
+
+ addIntervalToPartition(RootInterval = *I);
+
+ ++I; // After the first one...
+
+ // Add the rest of the intervals to the partition.
+ for (function_interval_iterator E = intervals_end(&F); I != E; ++I)
+ addIntervalToPartition(*I);
+
+ // Now that we know all of the successor information, propagate this to the
+ // predecessors for each block.
+ for (unsigned i = 0, e = Intervals.size(); i != e; ++i)
+ updatePredecessors(Intervals[i]);
+ return false;
+}
+
+
+// IntervalPartition ctor - Build a reduced interval partition from an
+// existing interval graph. This takes an additional boolean parameter to
+// distinguish it from a copy constructor. Always pass in false for now.
+//
+IntervalPartition::IntervalPartition(IntervalPartition &IP, bool)
+ : FunctionPass(&ID) {
+ assert(IP.getRootInterval() && "Cannot operate on empty IntervalPartitions!");
+
+ // Pass false to intervals_begin because we take ownership of it's memory
+ interval_part_interval_iterator I = intervals_begin(IP, false);
+ assert(I != intervals_end(IP) && "No intervals in interval partition!?!?!");
+
+ addIntervalToPartition(RootInterval = *I);
+
+ ++I; // After the first one...
+
+ // Add the rest of the intervals to the partition.
+ for (interval_part_interval_iterator E = intervals_end(IP); I != E; ++I)
+ addIntervalToPartition(*I);
+
+ // Now that we know all of the successor information, propagate this to the
+ // predecessors for each block.
+ for (unsigned i = 0, e = Intervals.size(); i != e; ++i)
+ updatePredecessors(Intervals[i]);
+}
+
diff --git a/lib/Analysis/LazyValueInfo.cpp b/lib/Analysis/LazyValueInfo.cpp
new file mode 100644
index 0000000..ff9026b
--- /dev/null
+++ b/lib/Analysis/LazyValueInfo.cpp
@@ -0,0 +1,582 @@
+//===- LazyValueInfo.cpp - Value constraint analysis ----------------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file defines the interface for lazy computation of value constraint
+// information.
+//
+//===----------------------------------------------------------------------===//
+
+#define DEBUG_TYPE "lazy-value-info"
+#include "llvm/Analysis/LazyValueInfo.h"
+#include "llvm/Constants.h"
+#include "llvm/Instructions.h"
+#include "llvm/Analysis/ConstantFolding.h"
+#include "llvm/Target/TargetData.h"
+#include "llvm/Support/CFG.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/PointerIntPair.h"
+#include "llvm/ADT/STLExtras.h"
+using namespace llvm;
+
+char LazyValueInfo::ID = 0;
+static RegisterPass<LazyValueInfo>
+X("lazy-value-info", "Lazy Value Information Analysis", false, true);
+
+namespace llvm {
+ FunctionPass *createLazyValueInfoPass() { return new LazyValueInfo(); }
+}
+
+
+//===----------------------------------------------------------------------===//
+// LVILatticeVal
+//===----------------------------------------------------------------------===//
+
+/// LVILatticeVal - This is the information tracked by LazyValueInfo for each
+/// value.
+///
+/// FIXME: This is basically just for bringup, this can be made a lot more rich
+/// in the future.
+///
+namespace {
+class LVILatticeVal {
+ enum LatticeValueTy {
+ /// undefined - This LLVM Value has no known value yet.
+ undefined,
+ /// constant - This LLVM Value has a specific constant value.
+ constant,
+
+ /// notconstant - This LLVM value is known to not have the specified value.
+ notconstant,
+
+ /// overdefined - This instruction is not known to be constant, and we know
+ /// it has a value.
+ overdefined
+ };
+
+ /// Val: This stores the current lattice value along with the Constant* for
+ /// the constant if this is a 'constant' or 'notconstant' value.
+ PointerIntPair<Constant *, 2, LatticeValueTy> Val;
+
+public:
+ LVILatticeVal() : Val(0, undefined) {}
+
+ static LVILatticeVal get(Constant *C) {
+ LVILatticeVal Res;
+ Res.markConstant(C);
+ return Res;
+ }
+ static LVILatticeVal getNot(Constant *C) {
+ LVILatticeVal Res;
+ Res.markNotConstant(C);
+ return Res;
+ }
+
+ bool isUndefined() const { return Val.getInt() == undefined; }
+ bool isConstant() const { return Val.getInt() == constant; }
+ bool isNotConstant() const { return Val.getInt() == notconstant; }
+ bool isOverdefined() const { return Val.getInt() == overdefined; }
+
+ Constant *getConstant() const {
+ assert(isConstant() && "Cannot get the constant of a non-constant!");
+ return Val.getPointer();
+ }
+
+ Constant *getNotConstant() const {
+ assert(isNotConstant() && "Cannot get the constant of a non-notconstant!");
+ return Val.getPointer();
+ }
+
+ /// markOverdefined - Return true if this is a change in status.
+ bool markOverdefined() {
+ if (isOverdefined())
+ return false;
+ Val.setInt(overdefined);
+ return true;
+ }
+
+ /// markConstant - Return true if this is a change in status.
+ bool markConstant(Constant *V) {
+ if (isConstant()) {
+ assert(getConstant() == V && "Marking constant with different value");
+ return false;
+ }
+
+ assert(isUndefined());
+ Val.setInt(constant);
+ assert(V && "Marking constant with NULL");
+ Val.setPointer(V);
+ return true;
+ }
+
+ /// markNotConstant - Return true if this is a change in status.
+ bool markNotConstant(Constant *V) {
+ if (isNotConstant()) {
+ assert(getNotConstant() == V && "Marking !constant with different value");
+ return false;
+ }
+
+ if (isConstant())
+ assert(getConstant() != V && "Marking not constant with different value");
+ else
+ assert(isUndefined());
+
+ Val.setInt(notconstant);
+ assert(V && "Marking constant with NULL");
+ Val.setPointer(V);
+ return true;
+ }
+
+ /// mergeIn - Merge the specified lattice value into this one, updating this
+ /// one and returning true if anything changed.
+ bool mergeIn(const LVILatticeVal &RHS) {
+ if (RHS.isUndefined() || isOverdefined()) return false;
+ if (RHS.isOverdefined()) return markOverdefined();
+
+ if (RHS.isNotConstant()) {
+ if (isNotConstant()) {
+ if (getNotConstant() != RHS.getNotConstant() ||
+ isa<ConstantExpr>(getNotConstant()) ||
+ isa<ConstantExpr>(RHS.getNotConstant()))
+ return markOverdefined();
+ return false;
+ }
+ if (isConstant()) {
+ if (getConstant() == RHS.getNotConstant() ||
+ isa<ConstantExpr>(RHS.getNotConstant()) ||
+ isa<ConstantExpr>(getConstant()))
+ return markOverdefined();
+ return markNotConstant(RHS.getNotConstant());
+ }
+
+ assert(isUndefined() && "Unexpected lattice");
+ return markNotConstant(RHS.getNotConstant());
+ }
+
+ // RHS must be a constant, we must be undef, constant, or notconstant.
+ if (isUndefined())
+ return markConstant(RHS.getConstant());
+
+ if (isConstant()) {
+ if (getConstant() != RHS.getConstant())
+ return markOverdefined();
+ return false;
+ }
+
+ // If we are known "!=4" and RHS is "==5", stay at "!=4".
+ if (getNotConstant() == RHS.getConstant() ||
+ isa<ConstantExpr>(getNotConstant()) ||
+ isa<ConstantExpr>(RHS.getConstant()))
+ return markOverdefined();
+ return false;
+ }
+
+};
+
+} // end anonymous namespace.
+
+namespace llvm {
+raw_ostream &operator<<(raw_ostream &OS, const LVILatticeVal &Val) {
+ if (Val.isUndefined())
+ return OS << "undefined";
+ if (Val.isOverdefined())
+ return OS << "overdefined";
+
+ if (Val.isNotConstant())
+ return OS << "notconstant<" << *Val.getNotConstant() << '>';
+ return OS << "constant<" << *Val.getConstant() << '>';
+}
+}
+
+//===----------------------------------------------------------------------===//
+// LazyValueInfoCache Decl
+//===----------------------------------------------------------------------===//
+
+namespace {
+ /// LazyValueInfoCache - This is the cache kept by LazyValueInfo which
+ /// maintains information about queries across the clients' queries.
+ class LazyValueInfoCache {
+ public:
+ /// BlockCacheEntryTy - This is a computed lattice value at the end of the
+ /// specified basic block for a Value* that depends on context.
+ typedef std::pair<BasicBlock*, LVILatticeVal> BlockCacheEntryTy;
+
+ /// ValueCacheEntryTy - This is all of the cached block information for
+ /// exactly one Value*. The entries are sorted by the BasicBlock* of the
+ /// entries, allowing us to do a lookup with a binary search.
+ typedef std::vector<BlockCacheEntryTy> ValueCacheEntryTy;
+
+ private:
+ /// ValueCache - This is all of the cached information for all values,
+ /// mapped from Value* to key information.
+ DenseMap<Value*, ValueCacheEntryTy> ValueCache;
+ public:
+
+ /// getValueInBlock - This is the query interface to determine the lattice
+ /// value for the specified Value* at the end of the specified block.
+ LVILatticeVal getValueInBlock(Value *V, BasicBlock *BB);
+
+ /// getValueOnEdge - This is the query interface to determine the lattice
+ /// value for the specified Value* that is true on the specified edge.
+ LVILatticeVal getValueOnEdge(Value *V, BasicBlock *FromBB,BasicBlock *ToBB);
+ };
+} // end anonymous namespace
+
+namespace {
+ struct BlockCacheEntryComparator {
+ static int Compare(const void *LHSv, const void *RHSv) {
+ const LazyValueInfoCache::BlockCacheEntryTy *LHS =
+ static_cast<const LazyValueInfoCache::BlockCacheEntryTy *>(LHSv);
+ const LazyValueInfoCache::BlockCacheEntryTy *RHS =
+ static_cast<const LazyValueInfoCache::BlockCacheEntryTy *>(RHSv);
+ if (LHS->first < RHS->first)
+ return -1;
+ if (LHS->first > RHS->first)
+ return 1;
+ return 0;
+ }
+
+ bool operator()(const LazyValueInfoCache::BlockCacheEntryTy &LHS,
+ const LazyValueInfoCache::BlockCacheEntryTy &RHS) const {
+ return LHS.first < RHS.first;
+ }
+ };
+}
+
+//===----------------------------------------------------------------------===//
+// LVIQuery Impl
+//===----------------------------------------------------------------------===//
+
+namespace {
+ /// LVIQuery - This is a transient object that exists while a query is
+ /// being performed.
+ ///
+ /// TODO: Reuse LVIQuery instead of recreating it for every query, this avoids
+ /// reallocation of the densemap on every query.
+ class LVIQuery {
+ typedef LazyValueInfoCache::BlockCacheEntryTy BlockCacheEntryTy;
+ typedef LazyValueInfoCache::ValueCacheEntryTy ValueCacheEntryTy;
+
+ /// This is the current value being queried for.
+ Value *Val;
+
+ /// This is all of the cached information about this value.
+ ValueCacheEntryTy &Cache;
+
+ /// NewBlocks - This is a mapping of the new BasicBlocks which have been
+ /// added to cache but that are not in sorted order.
+ DenseMap<BasicBlock*, LVILatticeVal> NewBlockInfo;
+ public:
+
+ LVIQuery(Value *V, ValueCacheEntryTy &VC) : Val(V), Cache(VC) {
+ }
+
+ ~LVIQuery() {
+ // When the query is done, insert the newly discovered facts into the
+ // cache in sorted order.
+ if (NewBlockInfo.empty()) return;
+
+ // Grow the cache to exactly fit the new data.
+ Cache.reserve(Cache.size() + NewBlockInfo.size());
+
+ // If we only have one new entry, insert it instead of doing a full-on
+ // sort.
+ if (NewBlockInfo.size() == 1) {
+ BlockCacheEntryTy Entry = *NewBlockInfo.begin();
+ ValueCacheEntryTy::iterator I =
+ std::lower_bound(Cache.begin(), Cache.end(), Entry,
+ BlockCacheEntryComparator());
+ assert((I == Cache.end() || I->first != Entry.first) &&
+ "Entry already in map!");
+
+ Cache.insert(I, Entry);
+ return;
+ }
+
+ // TODO: If we only have two new elements, INSERT them both.
+
+ Cache.insert(Cache.end(), NewBlockInfo.begin(), NewBlockInfo.end());
+ array_pod_sort(Cache.begin(), Cache.end(),
+ BlockCacheEntryComparator::Compare);
+
+ }
+
+ LVILatticeVal getBlockValue(BasicBlock *BB);
+ LVILatticeVal getEdgeValue(BasicBlock *FromBB, BasicBlock *ToBB);
+
+ private:
+ LVILatticeVal &getCachedEntryForBlock(BasicBlock *BB);
+ };
+} // end anonymous namespace
+
+/// getCachedEntryForBlock - See if we already have a value for this block. If
+/// so, return it, otherwise create a new entry in the NewBlockInfo map to use.
+LVILatticeVal &LVIQuery::getCachedEntryForBlock(BasicBlock *BB) {
+
+ // Do a binary search to see if we already have an entry for this block in
+ // the cache set. If so, find it.
+ if (!Cache.empty()) {
+ ValueCacheEntryTy::iterator Entry =
+ std::lower_bound(Cache.begin(), Cache.end(),
+ BlockCacheEntryTy(BB, LVILatticeVal()),
+ BlockCacheEntryComparator());
+ if (Entry != Cache.end() && Entry->first == BB)
+ return Entry->second;
+ }
+
+ // Otherwise, check to see if it's in NewBlockInfo or create a new entry if
+ // not.
+ return NewBlockInfo[BB];
+}
+
+LVILatticeVal LVIQuery::getBlockValue(BasicBlock *BB) {
+ // See if we already have a value for this block.
+ LVILatticeVal &BBLV = getCachedEntryForBlock(BB);
+
+ // If we've already computed this block's value, return it.
+ if (!BBLV.isUndefined()) {
+ DEBUG(dbgs() << " reuse BB '" << BB->getName() << "' val=" << BBLV <<'\n');
+ return BBLV;
+ }
+
+ // Otherwise, this is the first time we're seeing this block. Reset the
+ // lattice value to overdefined, so that cycles will terminate and be
+ // conservatively correct.
+ BBLV.markOverdefined();
+
+ // If V is live into BB, see if our predecessors know anything about it.
+ Instruction *BBI = dyn_cast<Instruction>(Val);
+ if (BBI == 0 || BBI->getParent() != BB) {
+ LVILatticeVal Result; // Start Undefined.
+ unsigned NumPreds = 0;
+
+ // Loop over all of our predecessors, merging what we know from them into
+ // result.
+ for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
+ Result.mergeIn(getEdgeValue(*PI, BB));
+
+ // If we hit overdefined, exit early. The BlockVals entry is already set
+ // to overdefined.
+ if (Result.isOverdefined()) {
+ DEBUG(dbgs() << " compute BB '" << BB->getName()
+ << "' - overdefined because of pred.\n");
+ return Result;
+ }
+ ++NumPreds;
+ }
+
+ // If this is the entry block, we must be asking about an argument. The
+ // value is overdefined.
+ if (NumPreds == 0 && BB == &BB->getParent()->front()) {
+ assert(isa<Argument>(Val) && "Unknown live-in to the entry block");
+ Result.markOverdefined();
+ return Result;
+ }
+
+ // Return the merged value, which is more precise than 'overdefined'.
+ assert(!Result.isOverdefined());
+ return getCachedEntryForBlock(BB) = Result;
+ }
+
+ // If this value is defined by an instruction in this block, we have to
+ // process it here somehow or return overdefined.
+ if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
+ (void)PN;
+ // TODO: PHI Translation in preds.
+ } else {
+
+ }
+
+ DEBUG(dbgs() << " compute BB '" << BB->getName()
+ << "' - overdefined because inst def found.\n");
+
+ LVILatticeVal Result;
+ Result.markOverdefined();
+ return getCachedEntryForBlock(BB) = Result;
+}
+
+
+/// getEdgeValue - This method attempts to infer more complex
+LVILatticeVal LVIQuery::getEdgeValue(BasicBlock *BBFrom, BasicBlock *BBTo) {
+ // TODO: Handle more complex conditionals. If (v == 0 || v2 < 1) is false, we
+ // know that v != 0.
+ if (BranchInst *BI = dyn_cast<BranchInst>(BBFrom->getTerminator())) {
+ // If this is a conditional branch and only one successor goes to BBTo, then
+ // we maybe able to infer something from the condition.
+ if (BI->isConditional() &&
+ BI->getSuccessor(0) != BI->getSuccessor(1)) {
+ bool isTrueDest = BI->getSuccessor(0) == BBTo;
+ assert(BI->getSuccessor(!isTrueDest) == BBTo &&
+ "BBTo isn't a successor of BBFrom");
+
+ // If V is the condition of the branch itself, then we know exactly what
+ // it is.
+ if (BI->getCondition() == Val)
+ return LVILatticeVal::get(ConstantInt::get(
+ Type::getInt1Ty(Val->getContext()), isTrueDest));
+
+ // If the condition of the branch is an equality comparison, we may be
+ // able to infer the value.
+ if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition()))
+ if (ICI->isEquality() && ICI->getOperand(0) == Val &&
+ isa<Constant>(ICI->getOperand(1))) {
+ // We know that V has the RHS constant if this is a true SETEQ or
+ // false SETNE.
+ if (isTrueDest == (ICI->getPredicate() == ICmpInst::ICMP_EQ))
+ return LVILatticeVal::get(cast<Constant>(ICI->getOperand(1)));
+ return LVILatticeVal::getNot(cast<Constant>(ICI->getOperand(1)));
+ }
+ }
+ }
+
+ // If the edge was formed by a switch on the value, then we may know exactly
+ // what it is.
+ if (SwitchInst *SI = dyn_cast<SwitchInst>(BBFrom->getTerminator())) {
+ // If BBTo is the default destination of the switch, we don't know anything.
+ // Given a more powerful range analysis we could know stuff.
+ if (SI->getCondition() == Val && SI->getDefaultDest() != BBTo) {
+ // We only know something if there is exactly one value that goes from
+ // BBFrom to BBTo.
+ unsigned NumEdges = 0;
+ ConstantInt *EdgeVal = 0;
+ for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i) {
+ if (SI->getSuccessor(i) != BBTo) continue;
+ if (NumEdges++) break;
+ EdgeVal = SI->getCaseValue(i);
+ }
+ assert(EdgeVal && "Missing successor?");
+ if (NumEdges == 1)
+ return LVILatticeVal::get(EdgeVal);
+ }
+ }
+
+ // Otherwise see if the value is known in the block.
+ return getBlockValue(BBFrom);
+}
+
+
+//===----------------------------------------------------------------------===//
+// LazyValueInfoCache Impl
+//===----------------------------------------------------------------------===//
+
+LVILatticeVal LazyValueInfoCache::getValueInBlock(Value *V, BasicBlock *BB) {
+ // If already a constant, there is nothing to compute.
+ if (Constant *VC = dyn_cast<Constant>(V))
+ return LVILatticeVal::get(VC);
+
+ DEBUG(dbgs() << "LVI Getting block end value " << *V << " at '"
+ << BB->getName() << "'\n");
+
+ LVILatticeVal Result = LVIQuery(V, ValueCache[V]).getBlockValue(BB);
+
+ DEBUG(dbgs() << " Result = " << Result << "\n");
+ return Result;
+}
+
+LVILatticeVal LazyValueInfoCache::
+getValueOnEdge(Value *V, BasicBlock *FromBB, BasicBlock *ToBB) {
+ // If already a constant, there is nothing to compute.
+ if (Constant *VC = dyn_cast<Constant>(V))
+ return LVILatticeVal::get(VC);
+
+ DEBUG(dbgs() << "LVI Getting edge value " << *V << " from '"
+ << FromBB->getName() << "' to '" << ToBB->getName() << "'\n");
+ LVILatticeVal Result =
+ LVIQuery(V, ValueCache[V]).getEdgeValue(FromBB, ToBB);
+
+ DEBUG(dbgs() << " Result = " << Result << "\n");
+
+ return Result;
+}
+
+//===----------------------------------------------------------------------===//
+// LazyValueInfo Impl
+//===----------------------------------------------------------------------===//
+
+bool LazyValueInfo::runOnFunction(Function &F) {
+ TD = getAnalysisIfAvailable<TargetData>();
+ // Fully lazy.
+ return false;
+}
+
+/// getCache - This lazily constructs the LazyValueInfoCache.
+static LazyValueInfoCache &getCache(void *&PImpl) {
+ if (!PImpl)
+ PImpl = new LazyValueInfoCache();
+ return *static_cast<LazyValueInfoCache*>(PImpl);
+}
+
+void LazyValueInfo::releaseMemory() {
+ // If the cache was allocated, free it.
+ if (PImpl) {
+ delete &getCache(PImpl);
+ PImpl = 0;
+ }
+}
+
+Constant *LazyValueInfo::getConstant(Value *V, BasicBlock *BB) {
+ LVILatticeVal Result = getCache(PImpl).getValueInBlock(V, BB);
+
+ if (Result.isConstant())
+ return Result.getConstant();
+ return 0;
+}
+
+/// getConstantOnEdge - Determine whether the specified value is known to be a
+/// constant on the specified edge. Return null if not.
+Constant *LazyValueInfo::getConstantOnEdge(Value *V, BasicBlock *FromBB,
+ BasicBlock *ToBB) {
+ LVILatticeVal Result = getCache(PImpl).getValueOnEdge(V, FromBB, ToBB);
+
+ if (Result.isConstant())
+ return Result.getConstant();
+ return 0;
+}
+
+/// getPredicateOnEdge - Determine whether the specified value comparison
+/// with a constant is known to be true or false on the specified CFG edge.
+/// Pred is a CmpInst predicate.
+LazyValueInfo::Tristate
+LazyValueInfo::getPredicateOnEdge(unsigned Pred, Value *V, Constant *C,
+ BasicBlock *FromBB, BasicBlock *ToBB) {
+ LVILatticeVal Result = getCache(PImpl).getValueOnEdge(V, FromBB, ToBB);
+
+ // If we know the value is a constant, evaluate the conditional.
+ Constant *Res = 0;
+ if (Result.isConstant()) {
+ Res = ConstantFoldCompareInstOperands(Pred, Result.getConstant(), C, TD);
+ if (ConstantInt *ResCI = dyn_cast_or_null<ConstantInt>(Res))
+ return ResCI->isZero() ? False : True;
+ return Unknown;
+ }
+
+ if (Result.isNotConstant()) {
+ // If this is an equality comparison, we can try to fold it knowing that
+ // "V != C1".
+ if (Pred == ICmpInst::ICMP_EQ) {
+ // !C1 == C -> false iff C1 == C.
+ Res = ConstantFoldCompareInstOperands(ICmpInst::ICMP_NE,
+ Result.getNotConstant(), C, TD);
+ if (Res->isNullValue())
+ return False;
+ } else if (Pred == ICmpInst::ICMP_NE) {
+ // !C1 != C -> true iff C1 == C.
+ Res = ConstantFoldCompareInstOperands(ICmpInst::ICMP_NE,
+ Result.getNotConstant(), C, TD);
+ if (Res->isNullValue())
+ return True;
+ }
+ return Unknown;
+ }
+
+ return Unknown;
+}
+
+
diff --git a/lib/Analysis/LibCallAliasAnalysis.cpp b/lib/Analysis/LibCallAliasAnalysis.cpp
new file mode 100644
index 0000000..7419659
--- /dev/null
+++ b/lib/Analysis/LibCallAliasAnalysis.cpp
@@ -0,0 +1,139 @@
+//===- LibCallAliasAnalysis.cpp - Implement AliasAnalysis for libcalls ----===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements the LibCallAliasAnalysis class.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/LibCallAliasAnalysis.h"
+#include "llvm/Analysis/Passes.h"
+#include "llvm/Analysis/LibCallSemantics.h"
+#include "llvm/Function.h"
+#include "llvm/Pass.h"
+using namespace llvm;
+
+// Register this pass...
+char LibCallAliasAnalysis::ID = 0;
+static RegisterPass<LibCallAliasAnalysis>
+X("libcall-aa", "LibCall Alias Analysis", false, true);
+
+// Declare that we implement the AliasAnalysis interface
+static RegisterAnalysisGroup<AliasAnalysis> Y(X);
+
+FunctionPass *llvm::createLibCallAliasAnalysisPass(LibCallInfo *LCI) {
+ return new LibCallAliasAnalysis(LCI);
+}
+
+LibCallAliasAnalysis::~LibCallAliasAnalysis() {
+ delete LCI;
+}
+
+void LibCallAliasAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
+ AliasAnalysis::getAnalysisUsage(AU);
+ AU.setPreservesAll(); // Does not transform code
+}
+
+
+
+/// AnalyzeLibCallDetails - Given a call to a function with the specified
+/// LibCallFunctionInfo, see if we can improve the mod/ref footprint of the call
+/// vs the specified pointer/size.
+AliasAnalysis::ModRefResult
+LibCallAliasAnalysis::AnalyzeLibCallDetails(const LibCallFunctionInfo *FI,
+ CallSite CS, Value *P,
+ unsigned Size) {
+ // If we have a function, check to see what kind of mod/ref effects it
+ // has. Start by including any info globally known about the function.
+ AliasAnalysis::ModRefResult MRInfo = FI->UniversalBehavior;
+ if (MRInfo == NoModRef) return MRInfo;
+
+ // If that didn't tell us that the function is 'readnone', check to see
+ // if we have detailed info and if 'P' is any of the locations we know
+ // about.
+ const LibCallFunctionInfo::LocationMRInfo *Details = FI->LocationDetails;
+ if (Details == 0)
+ return MRInfo;
+
+ // If the details array is of the 'DoesNot' kind, we only know something if
+ // the pointer is a match for one of the locations in 'Details'. If we find a
+ // match, we can prove some interactions cannot happen.
+ //
+ if (FI->DetailsType == LibCallFunctionInfo::DoesNot) {
+ // Find out if the pointer refers to a known location.
+ for (unsigned i = 0; Details[i].LocationID != ~0U; ++i) {
+ const LibCallLocationInfo &Loc =
+ LCI->getLocationInfo(Details[i].LocationID);
+ LibCallLocationInfo::LocResult Res = Loc.isLocation(CS, P, Size);
+ if (Res != LibCallLocationInfo::Yes) continue;
+
+ // If we find a match against a location that we 'do not' interact with,
+ // learn this info into MRInfo.
+ return ModRefResult(MRInfo & ~Details[i].MRInfo);
+ }
+ return MRInfo;
+ }
+
+ // If the details are of the 'DoesOnly' sort, we know something if the pointer
+ // is a match for one of the locations in 'Details'. Also, if we can prove
+ // that the pointers is *not* one of the locations in 'Details', we know that
+ // the call is NoModRef.
+ assert(FI->DetailsType == LibCallFunctionInfo::DoesOnly);
+
+ // Find out if the pointer refers to a known location.
+ bool NoneMatch = true;
+ for (unsigned i = 0; Details[i].LocationID != ~0U; ++i) {
+ const LibCallLocationInfo &Loc =
+ LCI->getLocationInfo(Details[i].LocationID);
+ LibCallLocationInfo::LocResult Res = Loc.isLocation(CS, P, Size);
+ if (Res == LibCallLocationInfo::No) continue;
+
+ // If we don't know if this pointer points to the location, then we have to
+ // assume it might alias in some case.
+ if (Res == LibCallLocationInfo::Unknown) {
+ NoneMatch = false;
+ continue;
+ }
+
+ // If we know that this pointer definitely is pointing into the location,
+ // merge in this information.
+ return ModRefResult(MRInfo & Details[i].MRInfo);
+ }
+
+ // If we found that the pointer is guaranteed to not match any of the
+ // locations in our 'DoesOnly' rule, then we know that the pointer must point
+ // to some other location. Since the libcall doesn't mod/ref any other
+ // locations, return NoModRef.
+ if (NoneMatch)
+ return NoModRef;
+
+ // Otherwise, return any other info gained so far.
+ return MRInfo;
+}
+
+// getModRefInfo - Check to see if the specified callsite can clobber the
+// specified memory object.
+//
+AliasAnalysis::ModRefResult
+LibCallAliasAnalysis::getModRefInfo(CallSite CS, Value *P, unsigned Size) {
+ ModRefResult MRInfo = ModRef;
+
+ // If this is a direct call to a function that LCI knows about, get the
+ // information about the runtime function.
+ if (LCI) {
+ if (Function *F = CS.getCalledFunction()) {
+ if (const LibCallFunctionInfo *FI = LCI->getFunctionInfo(F)) {
+ MRInfo = ModRefResult(MRInfo & AnalyzeLibCallDetails(FI, CS, P, Size));
+ if (MRInfo == NoModRef) return NoModRef;
+ }
+ }
+ }
+
+ // The AliasAnalysis base class has some smarts, lets use them.
+ return (ModRefResult)(MRInfo | AliasAnalysis::getModRefInfo(CS, P, Size));
+}
diff --git a/lib/Analysis/LibCallSemantics.cpp b/lib/Analysis/LibCallSemantics.cpp
new file mode 100644
index 0000000..e0060c3
--- /dev/null
+++ b/lib/Analysis/LibCallSemantics.cpp
@@ -0,0 +1,62 @@
+//===- LibCallSemantics.cpp - Describe library semantics ------------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements interfaces that can be used to describe language
+// specific runtime library interfaces (e.g. libc, libm, etc) to LLVM
+// optimizers.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/LibCallSemantics.h"
+#include "llvm/ADT/StringMap.h"
+#include "llvm/Function.h"
+using namespace llvm;
+
+/// getMap - This impl pointer in ~LibCallInfo is actually a StringMap. This
+/// helper does the cast.
+static StringMap<const LibCallFunctionInfo*> *getMap(void *Ptr) {
+ return static_cast<StringMap<const LibCallFunctionInfo*> *>(Ptr);
+}
+
+LibCallInfo::~LibCallInfo() {
+ delete getMap(Impl);
+}
+
+const LibCallLocationInfo &LibCallInfo::getLocationInfo(unsigned LocID) const {
+ // Get location info on the first call.
+ if (NumLocations == 0)
+ NumLocations = getLocationInfo(Locations);
+
+ assert(LocID < NumLocations && "Invalid location ID!");
+ return Locations[LocID];
+}
+
+
+/// getFunctionInfo - Return the LibCallFunctionInfo object corresponding to
+/// the specified function if we have it. If not, return null.
+const LibCallFunctionInfo *LibCallInfo::getFunctionInfo(Function *F) const {
+ StringMap<const LibCallFunctionInfo*> *Map = getMap(Impl);
+
+ /// If this is the first time we are querying for this info, lazily construct
+ /// the StringMap to index it.
+ if (Map == 0) {
+ Impl = Map = new StringMap<const LibCallFunctionInfo*>();
+
+ const LibCallFunctionInfo *Array = getFunctionInfoArray();
+ if (Array == 0) return 0;
+
+ // We now have the array of entries. Populate the StringMap.
+ for (unsigned i = 0; Array[i].Name; ++i)
+ (*Map)[Array[i].Name] = Array+i;
+ }
+
+ // Look up this function in the string map.
+ return Map->lookup(F->getName());
+}
+
diff --git a/lib/Analysis/LiveValues.cpp b/lib/Analysis/LiveValues.cpp
new file mode 100644
index 0000000..1b91d93
--- /dev/null
+++ b/lib/Analysis/LiveValues.cpp
@@ -0,0 +1,193 @@
+//===- LiveValues.cpp - Liveness information for LLVM IR Values. ----------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file defines the implementation for the LLVM IR Value liveness
+// analysis pass.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/LiveValues.h"
+#include "llvm/Analysis/Dominators.h"
+#include "llvm/Analysis/LoopInfo.h"
+using namespace llvm;
+
+namespace llvm {
+ FunctionPass *createLiveValuesPass() { return new LiveValues(); }
+}
+
+char LiveValues::ID = 0;
+static RegisterPass<LiveValues>
+X("live-values", "Value Liveness Analysis", false, true);
+
+LiveValues::LiveValues() : FunctionPass(&ID) {}
+
+void LiveValues::getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.addRequired<DominatorTree>();
+ AU.addRequired<LoopInfo>();
+ AU.setPreservesAll();
+}
+
+bool LiveValues::runOnFunction(Function &F) {
+ DT = &getAnalysis<DominatorTree>();
+ LI = &getAnalysis<LoopInfo>();
+
+ // This pass' values are computed lazily, so there's nothing to do here.
+
+ return false;
+}
+
+void LiveValues::releaseMemory() {
+ Memos.clear();
+}
+
+/// isUsedInBlock - Test if the given value is used in the given block.
+///
+bool LiveValues::isUsedInBlock(const Value *V, const BasicBlock *BB) {
+ Memo &M = getMemo(V);
+ return M.Used.count(BB);
+}
+
+/// isLiveThroughBlock - Test if the given value is known to be
+/// live-through the given block, meaning that the block is properly
+/// dominated by the value's definition, and there exists a block
+/// reachable from it that contains a use. This uses a conservative
+/// approximation that errs on the side of returning false.
+///
+bool LiveValues::isLiveThroughBlock(const Value *V,
+ const BasicBlock *BB) {
+ Memo &M = getMemo(V);
+ return M.LiveThrough.count(BB);
+}
+
+/// isKilledInBlock - Test if the given value is known to be killed in
+/// the given block, meaning that the block contains a use of the value,
+/// and no blocks reachable from the block contain a use. This uses a
+/// conservative approximation that errs on the side of returning false.
+///
+bool LiveValues::isKilledInBlock(const Value *V, const BasicBlock *BB) {
+ Memo &M = getMemo(V);
+ return M.Killed.count(BB);
+}
+
+/// getMemo - Retrieve an existing Memo for the given value if one
+/// is available, otherwise compute a new one.
+///
+LiveValues::Memo &LiveValues::getMemo(const Value *V) {
+ DenseMap<const Value *, Memo>::iterator I = Memos.find(V);
+ if (I != Memos.end())
+ return I->second;
+ return compute(V);
+}
+
+/// getImmediateDominator - A handy utility for the specific DominatorTree
+/// query that we need here.
+///
+static const BasicBlock *getImmediateDominator(const BasicBlock *BB,
+ const DominatorTree *DT) {
+ DomTreeNode *Node = DT->getNode(const_cast<BasicBlock *>(BB))->getIDom();
+ return Node ? Node->getBlock() : 0;
+}
+
+/// compute - Compute a new Memo for the given value.
+///
+LiveValues::Memo &LiveValues::compute(const Value *V) {
+ Memo &M = Memos[V];
+
+ // Determine the block containing the definition.
+ const BasicBlock *DefBB;
+ // Instructions define values with meaningful live ranges.
+ if (const Instruction *I = dyn_cast<Instruction>(V))
+ DefBB = I->getParent();
+ // Arguments can be analyzed as values defined in the entry block.
+ else if (const Argument *A = dyn_cast<Argument>(V))
+ DefBB = &A->getParent()->getEntryBlock();
+ // Constants and other things aren't meaningful here, so just
+ // return having computed an empty Memo so that we don't come
+ // here again. The assumption here is that client code won't
+ // be asking about such values very often.
+ else
+ return M;
+
+ // Determine if the value is defined inside a loop. This is used
+ // to track whether the value is ever used outside the loop, so
+ // it'll be set to null if the value is either not defined in a
+ // loop or used outside the loop in which it is defined.
+ const Loop *L = LI->getLoopFor(DefBB);
+
+ // Track whether the value is used anywhere outside of the block
+ // in which it is defined.
+ bool LiveOutOfDefBB = false;
+
+ // Examine each use of the value.
+ for (Value::use_const_iterator I = V->use_begin(), E = V->use_end();
+ I != E; ++I) {
+ const User *U = *I;
+ const BasicBlock *UseBB = cast<Instruction>(U)->getParent();
+
+ // Note the block in which this use occurs.
+ M.Used.insert(UseBB);
+
+ // If the use block doesn't have successors, the value can be
+ // considered killed.
+ if (succ_begin(UseBB) == succ_end(UseBB))
+ M.Killed.insert(UseBB);
+
+ // Observe whether the value is used outside of the loop in which
+ // it is defined. Switch to an enclosing loop if necessary.
+ for (; L; L = L->getParentLoop())
+ if (L->contains(UseBB))
+ break;
+
+ // Search for live-through blocks.
+ const BasicBlock *BB;
+ if (const PHINode *PHI = dyn_cast<PHINode>(U)) {
+ // For PHI nodes, start the search at the incoming block paired with the
+ // incoming value, which must be dominated by the definition.
+ unsigned Num = PHI->getIncomingValueNumForOperand(I.getOperandNo());
+ BB = PHI->getIncomingBlock(Num);
+
+ // A PHI-node use means the value is live-out of it's defining block
+ // even if that block also contains the only use.
+ LiveOutOfDefBB = true;
+ } else {
+ // Otherwise just start the search at the use.
+ BB = UseBB;
+
+ // Note if the use is outside the defining block.
+ LiveOutOfDefBB |= UseBB != DefBB;
+ }
+
+ // Climb the immediate dominator tree from the use to the definition
+ // and mark all intermediate blocks as live-through.
+ for (; BB != DefBB; BB = getImmediateDominator(BB, DT)) {
+ if (BB != UseBB && !M.LiveThrough.insert(BB))
+ break;
+ }
+ }
+
+ // If the value is defined inside a loop and is not live outside
+ // the loop, then each exit block of the loop in which the value
+ // is used is a kill block.
+ if (L) {
+ SmallVector<BasicBlock *, 4> ExitingBlocks;
+ L->getExitingBlocks(ExitingBlocks);
+ for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
+ const BasicBlock *ExitingBlock = ExitingBlocks[i];
+ if (M.Used.count(ExitingBlock))
+ M.Killed.insert(ExitingBlock);
+ }
+ }
+
+ // If the value was never used outside the block in which it was
+ // defined, it's killed in that block.
+ if (!LiveOutOfDefBB)
+ M.Killed.insert(DefBB);
+
+ return M;
+}
diff --git a/lib/Analysis/LoopDependenceAnalysis.cpp b/lib/Analysis/LoopDependenceAnalysis.cpp
new file mode 100644
index 0000000..bb4f46d
--- /dev/null
+++ b/lib/Analysis/LoopDependenceAnalysis.cpp
@@ -0,0 +1,352 @@
+//===- LoopDependenceAnalysis.cpp - LDA Implementation ----------*- C++ -*-===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This is the (beginning) of an implementation of a loop dependence analysis
+// framework, which is used to detect dependences in memory accesses in loops.
+//
+// Please note that this is work in progress and the interface is subject to
+// change.
+//
+// TODO: adapt as implementation progresses.
+//
+// TODO: document lingo (pair, subscript, index)
+//
+//===----------------------------------------------------------------------===//
+
+#define DEBUG_TYPE "lda"
+#include "llvm/ADT/DenseSet.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/Analysis/LoopDependenceAnalysis.h"
+#include "llvm/Analysis/LoopPass.h"
+#include "llvm/Analysis/ScalarEvolution.h"
+#include "llvm/Analysis/ScalarEvolutionExpressions.h"
+#include "llvm/Instructions.h"
+#include "llvm/Operator.h"
+#include "llvm/Support/Allocator.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/Target/TargetData.h"
+using namespace llvm;
+
+STATISTIC(NumAnswered, "Number of dependence queries answered");
+STATISTIC(NumAnalysed, "Number of distinct dependence pairs analysed");
+STATISTIC(NumDependent, "Number of pairs with dependent accesses");
+STATISTIC(NumIndependent, "Number of pairs with independent accesses");
+STATISTIC(NumUnknown, "Number of pairs with unknown accesses");
+
+LoopPass *llvm::createLoopDependenceAnalysisPass() {
+ return new LoopDependenceAnalysis();
+}
+
+static RegisterPass<LoopDependenceAnalysis>
+R("lda", "Loop Dependence Analysis", false, true);
+char LoopDependenceAnalysis::ID = 0;
+
+//===----------------------------------------------------------------------===//
+// Utility Functions
+//===----------------------------------------------------------------------===//
+
+static inline bool IsMemRefInstr(const Value *V) {
+ const Instruction *I = dyn_cast<const Instruction>(V);
+ return I && (I->mayReadFromMemory() || I->mayWriteToMemory());
+}
+
+static void GetMemRefInstrs(const Loop *L,
+ SmallVectorImpl<Instruction*> &Memrefs) {
+ for (Loop::block_iterator b = L->block_begin(), be = L->block_end();
+ b != be; ++b)
+ for (BasicBlock::iterator i = (*b)->begin(), ie = (*b)->end();
+ i != ie; ++i)
+ if (IsMemRefInstr(i))
+ Memrefs.push_back(i);
+}
+
+static bool IsLoadOrStoreInst(Value *I) {
+ return isa<LoadInst>(I) || isa<StoreInst>(I);
+}
+
+static Value *GetPointerOperand(Value *I) {
+ if (LoadInst *i = dyn_cast<LoadInst>(I))
+ return i->getPointerOperand();
+ if (StoreInst *i = dyn_cast<StoreInst>(I))
+ return i->getPointerOperand();
+ llvm_unreachable("Value is no load or store instruction!");
+ // Never reached.
+ return 0;
+}
+
+static AliasAnalysis::AliasResult UnderlyingObjectsAlias(AliasAnalysis *AA,
+ const Value *A,
+ const Value *B) {
+ const Value *aObj = A->getUnderlyingObject();
+ const Value *bObj = B->getUnderlyingObject();
+ return AA->alias(aObj, AA->getTypeStoreSize(aObj->getType()),
+ bObj, AA->getTypeStoreSize(bObj->getType()));
+}
+
+static inline const SCEV *GetZeroSCEV(ScalarEvolution *SE) {
+ return SE->getConstant(Type::getInt32Ty(SE->getContext()), 0L);
+}
+
+//===----------------------------------------------------------------------===//
+// Dependence Testing
+//===----------------------------------------------------------------------===//
+
+bool LoopDependenceAnalysis::isDependencePair(const Value *A,
+ const Value *B) const {
+ return IsMemRefInstr(A) &&
+ IsMemRefInstr(B) &&
+ (cast<const Instruction>(A)->mayWriteToMemory() ||
+ cast<const Instruction>(B)->mayWriteToMemory());
+}
+
+bool LoopDependenceAnalysis::findOrInsertDependencePair(Value *A,
+ Value *B,
+ DependencePair *&P) {
+ void *insertPos = 0;
+ FoldingSetNodeID id;
+ id.AddPointer(A);
+ id.AddPointer(B);
+
+ P = Pairs.FindNodeOrInsertPos(id, insertPos);
+ if (P) return true;
+
+ P = PairAllocator.Allocate<DependencePair>();
+ new (P) DependencePair(id, A, B);
+ Pairs.InsertNode(P, insertPos);
+ return false;
+}
+
+void LoopDependenceAnalysis::getLoops(const SCEV *S,
+ DenseSet<const Loop*>* Loops) const {
+ // Refactor this into an SCEVVisitor, if efficiency becomes a concern.
+ for (const Loop *L = this->L; L != 0; L = L->getParentLoop())
+ if (!S->isLoopInvariant(L))
+ Loops->insert(L);
+}
+
+bool LoopDependenceAnalysis::isLoopInvariant(const SCEV *S) const {
+ DenseSet<const Loop*> loops;
+ getLoops(S, &loops);
+ return loops.empty();
+}
+
+bool LoopDependenceAnalysis::isAffine(const SCEV *S) const {
+ const SCEVAddRecExpr *rec = dyn_cast<SCEVAddRecExpr>(S);
+ return isLoopInvariant(S) || (rec && rec->isAffine());
+}
+
+bool LoopDependenceAnalysis::isZIVPair(const SCEV *A, const SCEV *B) const {
+ return isLoopInvariant(A) && isLoopInvariant(B);
+}
+
+bool LoopDependenceAnalysis::isSIVPair(const SCEV *A, const SCEV *B) const {
+ DenseSet<const Loop*> loops;
+ getLoops(A, &loops);
+ getLoops(B, &loops);
+ return loops.size() == 1;
+}
+
+LoopDependenceAnalysis::DependenceResult
+LoopDependenceAnalysis::analyseZIV(const SCEV *A,
+ const SCEV *B,
+ Subscript *S) const {
+ assert(isZIVPair(A, B) && "Attempted to ZIV-test non-ZIV SCEVs!");
+ return A == B ? Dependent : Independent;
+}
+
+LoopDependenceAnalysis::DependenceResult
+LoopDependenceAnalysis::analyseSIV(const SCEV *A,
+ const SCEV *B,
+ Subscript *S) const {
+ return Unknown; // TODO: Implement.
+}
+
+LoopDependenceAnalysis::DependenceResult
+LoopDependenceAnalysis::analyseMIV(const SCEV *A,
+ const SCEV *B,
+ Subscript *S) const {
+ return Unknown; // TODO: Implement.
+}
+
+LoopDependenceAnalysis::DependenceResult
+LoopDependenceAnalysis::analyseSubscript(const SCEV *A,
+ const SCEV *B,
+ Subscript *S) const {
+ DEBUG(dbgs() << " Testing subscript: " << *A << ", " << *B << "\n");
+
+ if (A == B) {
+ DEBUG(dbgs() << " -> [D] same SCEV\n");
+ return Dependent;
+ }
+
+ if (!isAffine(A) || !isAffine(B)) {
+ DEBUG(dbgs() << " -> [?] not affine\n");
+ return Unknown;
+ }
+
+ if (isZIVPair(A, B))
+ return analyseZIV(A, B, S);
+
+ if (isSIVPair(A, B))
+ return analyseSIV(A, B, S);
+
+ return analyseMIV(A, B, S);
+}
+
+LoopDependenceAnalysis::DependenceResult
+LoopDependenceAnalysis::analysePair(DependencePair *P) const {
+ DEBUG(dbgs() << "Analysing:\n" << *P->A << "\n" << *P->B << "\n");
+
+ // We only analyse loads and stores but no possible memory accesses by e.g.
+ // free, call, or invoke instructions.
+ if (!IsLoadOrStoreInst(P->A) || !IsLoadOrStoreInst(P->B)) {
+ DEBUG(dbgs() << "--> [?] no load/store\n");
+ return Unknown;
+ }
+
+ Value *aPtr = GetPointerOperand(P->A);
+ Value *bPtr = GetPointerOperand(P->B);
+
+ switch (UnderlyingObjectsAlias(AA, aPtr, bPtr)) {
+ case AliasAnalysis::MayAlias:
+ // We can not analyse objects if we do not know about their aliasing.
+ DEBUG(dbgs() << "---> [?] may alias\n");
+ return Unknown;
+
+ case AliasAnalysis::NoAlias:
+ // If the objects noalias, they are distinct, accesses are independent.
+ DEBUG(dbgs() << "---> [I] no alias\n");
+ return Independent;
+
+ case AliasAnalysis::MustAlias:
+ break; // The underlying objects alias, test accesses for dependence.
+ }
+
+ const GEPOperator *aGEP = dyn_cast<GEPOperator>(aPtr);
+ const GEPOperator *bGEP = dyn_cast<GEPOperator>(bPtr);
+
+ if (!aGEP || !bGEP)
+ return Unknown;
+
+ // FIXME: Is filtering coupled subscripts necessary?
+
+ // Collect GEP operand pairs (FIXME: use GetGEPOperands from BasicAA), adding
+ // trailing zeroes to the smaller GEP, if needed.
+ typedef SmallVector<std::pair<const SCEV*, const SCEV*>, 4> GEPOpdPairsTy;
+ GEPOpdPairsTy opds;
+ for(GEPOperator::const_op_iterator aIdx = aGEP->idx_begin(),
+ aEnd = aGEP->idx_end(),
+ bIdx = bGEP->idx_begin(),
+ bEnd = bGEP->idx_end();
+ aIdx != aEnd && bIdx != bEnd;
+ aIdx += (aIdx != aEnd), bIdx += (bIdx != bEnd)) {
+ const SCEV* aSCEV = (aIdx != aEnd) ? SE->getSCEV(*aIdx) : GetZeroSCEV(SE);
+ const SCEV* bSCEV = (bIdx != bEnd) ? SE->getSCEV(*bIdx) : GetZeroSCEV(SE);
+ opds.push_back(std::make_pair(aSCEV, bSCEV));
+ }
+
+ if (!opds.empty() && opds[0].first != opds[0].second) {
+ // We cannot (yet) handle arbitrary GEP pointer offsets. By limiting
+ //
+ // TODO: this could be relaxed by adding the size of the underlying object
+ // to the first subscript. If we have e.g. (GEP x,0,i; GEP x,2,-i) and we
+ // know that x is a [100 x i8]*, we could modify the first subscript to be
+ // (i, 200-i) instead of (i, -i).
+ return Unknown;
+ }
+
+ // Now analyse the collected operand pairs (skipping the GEP ptr offsets).
+ for (GEPOpdPairsTy::const_iterator i = opds.begin() + 1, end = opds.end();
+ i != end; ++i) {
+ Subscript subscript;
+ DependenceResult result = analyseSubscript(i->first, i->second, &subscript);
+ if (result != Dependent) {
+ // We either proved independence or failed to analyse this subscript.
+ // Further subscripts will not improve the situation, so abort early.
+ return result;
+ }
+ P->Subscripts.push_back(subscript);
+ }
+ // We successfully analysed all subscripts but failed to prove independence.
+ return Dependent;
+}
+
+bool LoopDependenceAnalysis::depends(Value *A, Value *B) {
+ assert(isDependencePair(A, B) && "Values form no dependence pair!");
+ ++NumAnswered;
+
+ DependencePair *p;
+ if (!findOrInsertDependencePair(A, B, p)) {
+ // The pair is not cached, so analyse it.
+ ++NumAnalysed;
+ switch (p->Result = analysePair(p)) {
+ case Dependent: ++NumDependent; break;
+ case Independent: ++NumIndependent; break;
+ case Unknown: ++NumUnknown; break;
+ }
+ }
+ return p->Result != Independent;
+}
+
+//===----------------------------------------------------------------------===//
+// LoopDependenceAnalysis Implementation
+//===----------------------------------------------------------------------===//
+
+bool LoopDependenceAnalysis::runOnLoop(Loop *L, LPPassManager &) {
+ this->L = L;
+ AA = &getAnalysis<AliasAnalysis>();
+ SE = &getAnalysis<ScalarEvolution>();
+ return false;
+}
+
+void LoopDependenceAnalysis::releaseMemory() {
+ Pairs.clear();
+ PairAllocator.Reset();
+}
+
+void LoopDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.setPreservesAll();
+ AU.addRequiredTransitive<AliasAnalysis>();
+ AU.addRequiredTransitive<ScalarEvolution>();
+}
+
+static void PrintLoopInfo(raw_ostream &OS,
+ LoopDependenceAnalysis *LDA, const Loop *L) {
+ if (!L->empty()) return; // ignore non-innermost loops
+
+ SmallVector<Instruction*, 8> memrefs;
+ GetMemRefInstrs(L, memrefs);
+
+ OS << "Loop at depth " << L->getLoopDepth() << ", header block: ";
+ WriteAsOperand(OS, L->getHeader(), false);
+ OS << "\n";
+
+ OS << " Load/store instructions: " << memrefs.size() << "\n";
+ for (SmallVector<Instruction*, 8>::const_iterator x = memrefs.begin(),
+ end = memrefs.end(); x != end; ++x)
+ OS << "\t" << (x - memrefs.begin()) << ": " << **x << "\n";
+
+ OS << " Pairwise dependence results:\n";
+ for (SmallVector<Instruction*, 8>::const_iterator x = memrefs.begin(),
+ end = memrefs.end(); x != end; ++x)
+ for (SmallVector<Instruction*, 8>::const_iterator y = x + 1;
+ y != end; ++y)
+ if (LDA->isDependencePair(*x, *y))
+ OS << "\t" << (x - memrefs.begin()) << "," << (y - memrefs.begin())
+ << ": " << (LDA->depends(*x, *y) ? "dependent" : "independent")
+ << "\n";
+}
+
+void LoopDependenceAnalysis::print(raw_ostream &OS, const Module*) const {
+ // TODO: doc why const_cast is safe
+ PrintLoopInfo(OS, const_cast<LoopDependenceAnalysis*>(this), this->L);
+}
diff --git a/lib/Analysis/LoopInfo.cpp b/lib/Analysis/LoopInfo.cpp
new file mode 100644
index 0000000..453af5a
--- /dev/null
+++ b/lib/Analysis/LoopInfo.cpp
@@ -0,0 +1,427 @@
+//===- LoopInfo.cpp - Natural Loop Calculator -----------------------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file defines the LoopInfo class that is used to identify natural loops
+// and determine the loop depth of various nodes of the CFG. Note that the
+// loops identified may actually be several natural loops that share the same
+// header node... not just a single natural loop.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/LoopInfo.h"
+#include "llvm/Constants.h"
+#include "llvm/Instructions.h"
+#include "llvm/Analysis/Dominators.h"
+#include "llvm/Assembly/Writer.h"
+#include "llvm/Support/CFG.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/ADT/DepthFirstIterator.h"
+#include "llvm/ADT/SmallPtrSet.h"
+#include <algorithm>
+using namespace llvm;
+
+// Always verify loopinfo if expensive checking is enabled.
+#ifdef XDEBUG
+bool VerifyLoopInfo = true;
+#else
+bool VerifyLoopInfo = false;
+#endif
+static cl::opt<bool,true>
+VerifyLoopInfoX("verify-loop-info", cl::location(VerifyLoopInfo),
+ cl::desc("Verify loop info (time consuming)"));
+
+char LoopInfo::ID = 0;
+static RegisterPass<LoopInfo>
+X("loops", "Natural Loop Information", true, true);
+
+//===----------------------------------------------------------------------===//
+// Loop implementation
+//
+
+/// isLoopInvariant - Return true if the specified value is loop invariant
+///
+bool Loop::isLoopInvariant(Value *V) const {
+ if (Instruction *I = dyn_cast<Instruction>(V))
+ return isLoopInvariant(I);
+ return true; // All non-instructions are loop invariant
+}
+
+/// isLoopInvariant - Return true if the specified instruction is
+/// loop-invariant.
+///
+bool Loop::isLoopInvariant(Instruction *I) const {
+ return !contains(I);
+}
+
+/// makeLoopInvariant - If the given value is an instruciton inside of the
+/// loop and it can be hoisted, do so to make it trivially loop-invariant.
+/// Return true if the value after any hoisting is loop invariant. This
+/// function can be used as a slightly more aggressive replacement for
+/// isLoopInvariant.
+///
+/// If InsertPt is specified, it is the point to hoist instructions to.
+/// If null, the terminator of the loop preheader is used.
+///
+bool Loop::makeLoopInvariant(Value *V, bool &Changed,
+ Instruction *InsertPt) const {
+ if (Instruction *I = dyn_cast<Instruction>(V))
+ return makeLoopInvariant(I, Changed, InsertPt);
+ return true; // All non-instructions are loop-invariant.
+}
+
+/// makeLoopInvariant - If the given instruction is inside of the
+/// loop and it can be hoisted, do so to make it trivially loop-invariant.
+/// Return true if the instruction after any hoisting is loop invariant. This
+/// function can be used as a slightly more aggressive replacement for
+/// isLoopInvariant.
+///
+/// If InsertPt is specified, it is the point to hoist instructions to.
+/// If null, the terminator of the loop preheader is used.
+///
+bool Loop::makeLoopInvariant(Instruction *I, bool &Changed,
+ Instruction *InsertPt) const {
+ // Test if the value is already loop-invariant.
+ if (isLoopInvariant(I))
+ return true;
+ if (!I->isSafeToSpeculativelyExecute())
+ return false;
+ if (I->mayReadFromMemory())
+ return false;
+ // Determine the insertion point, unless one was given.
+ if (!InsertPt) {
+ BasicBlock *Preheader = getLoopPreheader();
+ // Without a preheader, hoisting is not feasible.
+ if (!Preheader)
+ return false;
+ InsertPt = Preheader->getTerminator();
+ }
+ // Don't hoist instructions with loop-variant operands.
+ for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
+ if (!makeLoopInvariant(I->getOperand(i), Changed, InsertPt))
+ return false;
+ // Hoist.
+ I->moveBefore(InsertPt);
+ Changed = true;
+ return true;
+}
+
+/// getCanonicalInductionVariable - Check to see if the loop has a canonical
+/// induction variable: an integer recurrence that starts at 0 and increments
+/// by one each time through the loop. If so, return the phi node that
+/// corresponds to it.
+///
+/// The IndVarSimplify pass transforms loops to have a canonical induction
+/// variable.
+///
+PHINode *Loop::getCanonicalInductionVariable() const {
+ BasicBlock *H = getHeader();
+
+ BasicBlock *Incoming = 0, *Backedge = 0;
+ typedef GraphTraits<Inverse<BasicBlock*> > InvBlockTraits;
+ InvBlockTraits::ChildIteratorType PI = InvBlockTraits::child_begin(H);
+ assert(PI != InvBlockTraits::child_end(H) &&
+ "Loop must have at least one backedge!");
+ Backedge = *PI++;
+ if (PI == InvBlockTraits::child_end(H)) return 0; // dead loop
+ Incoming = *PI++;
+ if (PI != InvBlockTraits::child_end(H)) return 0; // multiple backedges?
+
+ if (contains(Incoming)) {
+ if (contains(Backedge))
+ return 0;
+ std::swap(Incoming, Backedge);
+ } else if (!contains(Backedge))
+ return 0;
+
+ // Loop over all of the PHI nodes, looking for a canonical indvar.
+ for (BasicBlock::iterator I = H->begin(); isa<PHINode>(I); ++I) {
+ PHINode *PN = cast<PHINode>(I);
+ if (ConstantInt *CI =
+ dyn_cast<ConstantInt>(PN->getIncomingValueForBlock(Incoming)))
+ if (CI->isNullValue())
+ if (Instruction *Inc =
+ dyn_cast<Instruction>(PN->getIncomingValueForBlock(Backedge)))
+ if (Inc->getOpcode() == Instruction::Add &&
+ Inc->getOperand(0) == PN)
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(Inc->getOperand(1)))
+ if (CI->equalsInt(1))
+ return PN;
+ }
+ return 0;
+}
+
+/// getCanonicalInductionVariableIncrement - Return the LLVM value that holds
+/// the canonical induction variable value for the "next" iteration of the
+/// loop. This always succeeds if getCanonicalInductionVariable succeeds.
+///
+Instruction *Loop::getCanonicalInductionVariableIncrement() const {
+ if (PHINode *PN = getCanonicalInductionVariable()) {
+ bool P1InLoop = contains(PN->getIncomingBlock(1));
+ return cast<Instruction>(PN->getIncomingValue(P1InLoop));
+ }
+ return 0;
+}
+
+/// getTripCount - Return a loop-invariant LLVM value indicating the number of
+/// times the loop will be executed. Note that this means that the backedge
+/// of the loop executes N-1 times. If the trip-count cannot be determined,
+/// this returns null.
+///
+/// The IndVarSimplify pass transforms loops to have a form that this
+/// function easily understands.
+///
+Value *Loop::getTripCount() const {
+ // Canonical loops will end with a 'cmp ne I, V', where I is the incremented
+ // canonical induction variable and V is the trip count of the loop.
+ Instruction *Inc = getCanonicalInductionVariableIncrement();
+ if (Inc == 0) return 0;
+ PHINode *IV = cast<PHINode>(Inc->getOperand(0));
+
+ BasicBlock *BackedgeBlock =
+ IV->getIncomingBlock(contains(IV->getIncomingBlock(1)));
+
+ if (BranchInst *BI = dyn_cast<BranchInst>(BackedgeBlock->getTerminator()))
+ if (BI->isConditional()) {
+ if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
+ if (ICI->getOperand(0) == Inc) {
+ if (BI->getSuccessor(0) == getHeader()) {
+ if (ICI->getPredicate() == ICmpInst::ICMP_NE)
+ return ICI->getOperand(1);
+ } else if (ICI->getPredicate() == ICmpInst::ICMP_EQ) {
+ return ICI->getOperand(1);
+ }
+ }
+ }
+ }
+
+ return 0;
+}
+
+/// getSmallConstantTripCount - Returns the trip count of this loop as a
+/// normal unsigned value, if possible. Returns 0 if the trip count is unknown
+/// of not constant. Will also return 0 if the trip count is very large
+/// (>= 2^32)
+unsigned Loop::getSmallConstantTripCount() const {
+ Value* TripCount = this->getTripCount();
+ if (TripCount) {
+ if (ConstantInt *TripCountC = dyn_cast<ConstantInt>(TripCount)) {
+ // Guard against huge trip counts.
+ if (TripCountC->getValue().getActiveBits() <= 32) {
+ return (unsigned)TripCountC->getZExtValue();
+ }
+ }
+ }
+ return 0;
+}
+
+/// getSmallConstantTripMultiple - Returns the largest constant divisor of the
+/// trip count of this loop as a normal unsigned value, if possible. This
+/// means that the actual trip count is always a multiple of the returned
+/// value (don't forget the trip count could very well be zero as well!).
+///
+/// Returns 1 if the trip count is unknown or not guaranteed to be the
+/// multiple of a constant (which is also the case if the trip count is simply
+/// constant, use getSmallConstantTripCount for that case), Will also return 1
+/// if the trip count is very large (>= 2^32).
+unsigned Loop::getSmallConstantTripMultiple() const {
+ Value* TripCount = this->getTripCount();
+ // This will hold the ConstantInt result, if any
+ ConstantInt *Result = NULL;
+ if (TripCount) {
+ // See if the trip count is constant itself
+ Result = dyn_cast<ConstantInt>(TripCount);
+ // if not, see if it is a multiplication
+ if (!Result)
+ if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TripCount)) {
+ switch (BO->getOpcode()) {
+ case BinaryOperator::Mul:
+ Result = dyn_cast<ConstantInt>(BO->getOperand(1));
+ break;
+ case BinaryOperator::Shl:
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1)))
+ if (CI->getValue().getActiveBits() <= 5)
+ return 1u << CI->getZExtValue();
+ break;
+ default:
+ break;
+ }
+ }
+ }
+ // Guard against huge trip counts.
+ if (Result && Result->getValue().getActiveBits() <= 32) {
+ return (unsigned)Result->getZExtValue();
+ } else {
+ return 1;
+ }
+}
+
+/// isLCSSAForm - Return true if the Loop is in LCSSA form
+bool Loop::isLCSSAForm() const {
+ // Sort the blocks vector so that we can use binary search to do quick
+ // lookups.
+ SmallPtrSet<BasicBlock *, 16> LoopBBs(block_begin(), block_end());
+
+ for (block_iterator BI = block_begin(), E = block_end(); BI != E; ++BI) {
+ BasicBlock *BB = *BI;
+ for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;++I)
+ for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
+ ++UI) {
+ BasicBlock *UserBB = cast<Instruction>(*UI)->getParent();
+ if (PHINode *P = dyn_cast<PHINode>(*UI))
+ UserBB = P->getIncomingBlock(UI);
+
+ // Check the current block, as a fast-path. Most values are used in
+ // the same block they are defined in.
+ if (UserBB != BB && !LoopBBs.count(UserBB))
+ return false;
+ }
+ }
+
+ return true;
+}
+
+/// isLoopSimplifyForm - Return true if the Loop is in the form that
+/// the LoopSimplify form transforms loops to, which is sometimes called
+/// normal form.
+bool Loop::isLoopSimplifyForm() const {
+ // Normal-form loops have a preheader, a single backedge, and all of their
+ // exits have all their predecessors inside the loop.
+ return getLoopPreheader() && getLoopLatch() && hasDedicatedExits();
+}
+
+/// hasDedicatedExits - Return true if no exit block for the loop
+/// has a predecessor that is outside the loop.
+bool Loop::hasDedicatedExits() const {
+ // Sort the blocks vector so that we can use binary search to do quick
+ // lookups.
+ SmallPtrSet<BasicBlock *, 16> LoopBBs(block_begin(), block_end());
+ // Each predecessor of each exit block of a normal loop is contained
+ // within the loop.
+ SmallVector<BasicBlock *, 4> ExitBlocks;
+ getExitBlocks(ExitBlocks);
+ for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
+ for (pred_iterator PI = pred_begin(ExitBlocks[i]),
+ PE = pred_end(ExitBlocks[i]); PI != PE; ++PI)
+ if (!LoopBBs.count(*PI))
+ return false;
+ // All the requirements are met.
+ return true;
+}
+
+/// getUniqueExitBlocks - Return all unique successor blocks of this loop.
+/// These are the blocks _outside of the current loop_ which are branched to.
+/// This assumes that loop exits are in canonical form.
+///
+void
+Loop::getUniqueExitBlocks(SmallVectorImpl<BasicBlock *> &ExitBlocks) const {
+ assert(hasDedicatedExits() &&
+ "getUniqueExitBlocks assumes the loop has canonical form exits!");
+
+ // Sort the blocks vector so that we can use binary search to do quick
+ // lookups.
+ SmallVector<BasicBlock *, 128> LoopBBs(block_begin(), block_end());
+ std::sort(LoopBBs.begin(), LoopBBs.end());
+
+ SmallVector<BasicBlock *, 32> switchExitBlocks;
+
+ for (block_iterator BI = block_begin(), BE = block_end(); BI != BE; ++BI) {
+
+ BasicBlock *current = *BI;
+ switchExitBlocks.clear();
+
+ typedef GraphTraits<BasicBlock *> BlockTraits;
+ typedef GraphTraits<Inverse<BasicBlock *> > InvBlockTraits;
+ for (BlockTraits::ChildIteratorType I =
+ BlockTraits::child_begin(*BI), E = BlockTraits::child_end(*BI);
+ I != E; ++I) {
+ // If block is inside the loop then it is not a exit block.
+ if (std::binary_search(LoopBBs.begin(), LoopBBs.end(), *I))
+ continue;
+
+ InvBlockTraits::ChildIteratorType PI = InvBlockTraits::child_begin(*I);
+ BasicBlock *firstPred = *PI;
+
+ // If current basic block is this exit block's first predecessor
+ // then only insert exit block in to the output ExitBlocks vector.
+ // This ensures that same exit block is not inserted twice into
+ // ExitBlocks vector.
+ if (current != firstPred)
+ continue;
+
+ // If a terminator has more then two successors, for example SwitchInst,
+ // then it is possible that there are multiple edges from current block
+ // to one exit block.
+ if (std::distance(BlockTraits::child_begin(current),
+ BlockTraits::child_end(current)) <= 2) {
+ ExitBlocks.push_back(*I);
+ continue;
+ }
+
+ // In case of multiple edges from current block to exit block, collect
+ // only one edge in ExitBlocks. Use switchExitBlocks to keep track of
+ // duplicate edges.
+ if (std::find(switchExitBlocks.begin(), switchExitBlocks.end(), *I)
+ == switchExitBlocks.end()) {
+ switchExitBlocks.push_back(*I);
+ ExitBlocks.push_back(*I);
+ }
+ }
+ }
+}
+
+/// getUniqueExitBlock - If getUniqueExitBlocks would return exactly one
+/// block, return that block. Otherwise return null.
+BasicBlock *Loop::getUniqueExitBlock() const {
+ SmallVector<BasicBlock *, 8> UniqueExitBlocks;
+ getUniqueExitBlocks(UniqueExitBlocks);
+ if (UniqueExitBlocks.size() == 1)
+ return UniqueExitBlocks[0];
+ return 0;
+}
+
+void Loop::dump() const {
+ print(dbgs());
+}
+
+//===----------------------------------------------------------------------===//
+// LoopInfo implementation
+//
+bool LoopInfo::runOnFunction(Function &) {
+ releaseMemory();
+ LI.Calculate(getAnalysis<DominatorTree>().getBase()); // Update
+ return false;
+}
+
+void LoopInfo::verifyAnalysis() const {
+ // LoopInfo is a FunctionPass, but verifying every loop in the function
+ // each time verifyAnalysis is called is very expensive. The
+ // -verify-loop-info option can enable this. In order to perform some
+ // checking by default, LoopPass has been taught to call verifyLoop
+ // manually during loop pass sequences.
+
+ if (!VerifyLoopInfo) return;
+
+ for (iterator I = begin(), E = end(); I != E; ++I) {
+ assert(!(*I)->getParentLoop() && "Top-level loop has a parent!");
+ (*I)->verifyLoopNest();
+ }
+
+ // TODO: check BBMap consistency.
+}
+
+void LoopInfo::getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.setPreservesAll();
+ AU.addRequired<DominatorTree>();
+}
+
+void LoopInfo::print(raw_ostream &OS, const Module*) const {
+ LI.print(OS);
+}
+
diff --git a/lib/Analysis/LoopPass.cpp b/lib/Analysis/LoopPass.cpp
new file mode 100644
index 0000000..2d613f6
--- /dev/null
+++ b/lib/Analysis/LoopPass.cpp
@@ -0,0 +1,362 @@
+//===- LoopPass.cpp - Loop Pass and Loop Pass Manager ---------------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements LoopPass and LPPassManager. All loop optimization
+// and transformation passes are derived from LoopPass. LPPassManager is
+// responsible for managing LoopPasses.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/LoopPass.h"
+using namespace llvm;
+
+//===----------------------------------------------------------------------===//
+// LPPassManager
+//
+
+char LPPassManager::ID = 0;
+
+LPPassManager::LPPassManager(int Depth)
+ : FunctionPass(&ID), PMDataManager(Depth) {
+ skipThisLoop = false;
+ redoThisLoop = false;
+ LI = NULL;
+ CurrentLoop = NULL;
+}
+
+/// Delete loop from the loop queue and loop hierarchy (LoopInfo).
+void LPPassManager::deleteLoopFromQueue(Loop *L) {
+
+ if (Loop *ParentLoop = L->getParentLoop()) { // Not a top-level loop.
+ // Reparent all of the blocks in this loop. Since BBLoop had a parent,
+ // they are now all in it.
+ for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
+ I != E; ++I)
+ if (LI->getLoopFor(*I) == L) // Don't change blocks in subloops.
+ LI->changeLoopFor(*I, ParentLoop);
+
+ // Remove the loop from its parent loop.
+ for (Loop::iterator I = ParentLoop->begin(), E = ParentLoop->end();;
+ ++I) {
+ assert(I != E && "Couldn't find loop");
+ if (*I == L) {
+ ParentLoop->removeChildLoop(I);
+ break;
+ }
+ }
+
+ // Move all subloops into the parent loop.
+ while (!L->empty())
+ ParentLoop->addChildLoop(L->removeChildLoop(L->end()-1));
+ } else {
+ // Reparent all of the blocks in this loop. Since BBLoop had no parent,
+ // they no longer in a loop at all.
+
+ for (unsigned i = 0; i != L->getBlocks().size(); ++i) {
+ // Don't change blocks in subloops.
+ if (LI->getLoopFor(L->getBlocks()[i]) == L) {
+ LI->removeBlock(L->getBlocks()[i]);
+ --i;
+ }
+ }
+
+ // Remove the loop from the top-level LoopInfo object.
+ for (LoopInfo::iterator I = LI->begin(), E = LI->end();; ++I) {
+ assert(I != E && "Couldn't find loop");
+ if (*I == L) {
+ LI->removeLoop(I);
+ break;
+ }
+ }
+
+ // Move all of the subloops to the top-level.
+ while (!L->empty())
+ LI->addTopLevelLoop(L->removeChildLoop(L->end()-1));
+ }
+
+ delete L;
+
+ // If L is current loop then skip rest of the passes and let
+ // runOnFunction remove L from LQ. Otherwise, remove L from LQ now
+ // and continue applying other passes on CurrentLoop.
+ if (CurrentLoop == L) {
+ skipThisLoop = true;
+ return;
+ }
+
+ for (std::deque<Loop *>::iterator I = LQ.begin(),
+ E = LQ.end(); I != E; ++I) {
+ if (*I == L) {
+ LQ.erase(I);
+ break;
+ }
+ }
+}
+
+// Inset loop into loop nest (LoopInfo) and loop queue (LQ).
+void LPPassManager::insertLoop(Loop *L, Loop *ParentLoop) {
+
+ assert (CurrentLoop != L && "Cannot insert CurrentLoop");
+
+ // Insert into loop nest
+ if (ParentLoop)
+ ParentLoop->addChildLoop(L);
+ else
+ LI->addTopLevelLoop(L);
+
+ insertLoopIntoQueue(L);
+}
+
+void LPPassManager::insertLoopIntoQueue(Loop *L) {
+ // Insert L into loop queue
+ if (L == CurrentLoop)
+ redoLoop(L);
+ else if (!L->getParentLoop())
+ // This is top level loop.
+ LQ.push_front(L);
+ else {
+ // Insert L after the parent loop.
+ for (std::deque<Loop *>::iterator I = LQ.begin(),
+ E = LQ.end(); I != E; ++I) {
+ if (*I == L->getParentLoop()) {
+ // deque does not support insert after.
+ ++I;
+ LQ.insert(I, 1, L);
+ break;
+ }
+ }
+ }
+}
+
+// Reoptimize this loop. LPPassManager will re-insert this loop into the
+// queue. This allows LoopPass to change loop nest for the loop. This
+// utility may send LPPassManager into infinite loops so use caution.
+void LPPassManager::redoLoop(Loop *L) {
+ assert (CurrentLoop == L && "Can redo only CurrentLoop");
+ redoThisLoop = true;
+}
+
+/// cloneBasicBlockSimpleAnalysis - Invoke cloneBasicBlockAnalysis hook for
+/// all loop passes.
+void LPPassManager::cloneBasicBlockSimpleAnalysis(BasicBlock *From,
+ BasicBlock *To, Loop *L) {
+ for (unsigned Index = 0; Index < getNumContainedPasses(); ++Index) {
+ LoopPass *LP = (LoopPass *)getContainedPass(Index);
+ LP->cloneBasicBlockAnalysis(From, To, L);
+ }
+}
+
+/// deleteSimpleAnalysisValue - Invoke deleteAnalysisValue hook for all passes.
+void LPPassManager::deleteSimpleAnalysisValue(Value *V, Loop *L) {
+ if (BasicBlock *BB = dyn_cast<BasicBlock>(V)) {
+ for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE;
+ ++BI) {
+ Instruction &I = *BI;
+ deleteSimpleAnalysisValue(&I, L);
+ }
+ }
+ for (unsigned Index = 0; Index < getNumContainedPasses(); ++Index) {
+ LoopPass *LP = (LoopPass *)getContainedPass(Index);
+ LP->deleteAnalysisValue(V, L);
+ }
+}
+
+
+// Recurse through all subloops and all loops into LQ.
+static void addLoopIntoQueue(Loop *L, std::deque<Loop *> &LQ) {
+ LQ.push_back(L);
+ for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
+ addLoopIntoQueue(*I, LQ);
+}
+
+/// Pass Manager itself does not invalidate any analysis info.
+void LPPassManager::getAnalysisUsage(AnalysisUsage &Info) const {
+ // LPPassManager needs LoopInfo. In the long term LoopInfo class will
+ // become part of LPPassManager.
+ Info.addRequired<LoopInfo>();
+ Info.setPreservesAll();
+}
+
+/// run - Execute all of the passes scheduled for execution. Keep track of
+/// whether any of the passes modifies the function, and if so, return true.
+bool LPPassManager::runOnFunction(Function &F) {
+ LI = &getAnalysis<LoopInfo>();
+ bool Changed = false;
+
+ // Collect inherited analysis from Module level pass manager.
+ populateInheritedAnalysis(TPM->activeStack);
+
+ // Populate Loop Queue
+ for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
+ addLoopIntoQueue(*I, LQ);
+
+ if (LQ.empty()) // No loops, skip calling finalizers
+ return false;
+
+ // Initialization
+ for (std::deque<Loop *>::const_iterator I = LQ.begin(), E = LQ.end();
+ I != E; ++I) {
+ Loop *L = *I;
+ for (unsigned Index = 0; Index < getNumContainedPasses(); ++Index) {
+ LoopPass *P = (LoopPass*)getContainedPass(Index);
+ Changed |= P->doInitialization(L, *this);
+ }
+ }
+
+ // Walk Loops
+ while (!LQ.empty()) {
+
+ CurrentLoop = LQ.back();
+ skipThisLoop = false;
+ redoThisLoop = false;
+
+ // Run all passes on the current Loop.
+ for (unsigned Index = 0; Index < getNumContainedPasses(); ++Index) {
+ LoopPass *P = (LoopPass*)getContainedPass(Index);
+
+ dumpPassInfo(P, EXECUTION_MSG, ON_LOOP_MSG,
+ CurrentLoop->getHeader()->getNameStr());
+ dumpRequiredSet(P);
+
+ initializeAnalysisImpl(P);
+
+ {
+ PassManagerPrettyStackEntry X(P, *CurrentLoop->getHeader());
+ Timer *T = StartPassTimer(P);
+ Changed |= P->runOnLoop(CurrentLoop, *this);
+ StopPassTimer(P, T);
+ }
+
+ if (Changed)
+ dumpPassInfo(P, MODIFICATION_MSG, ON_LOOP_MSG,
+ skipThisLoop ? "<deleted>" :
+ CurrentLoop->getHeader()->getNameStr());
+ dumpPreservedSet(P);
+
+ if (!skipThisLoop) {
+ // Manually check that this loop is still healthy. This is done
+ // instead of relying on LoopInfo::verifyLoop since LoopInfo
+ // is a function pass and it's really expensive to verify every
+ // loop in the function every time. That level of checking can be
+ // enabled with the -verify-loop-info option.
+ Timer *T = StartPassTimer(LI);
+ CurrentLoop->verifyLoop();
+ StopPassTimer(LI, T);
+
+ // Then call the regular verifyAnalysis functions.
+ verifyPreservedAnalysis(P);
+ }
+
+ removeNotPreservedAnalysis(P);
+ recordAvailableAnalysis(P);
+ removeDeadPasses(P,
+ skipThisLoop ? "<deleted>" :
+ CurrentLoop->getHeader()->getNameStr(),
+ ON_LOOP_MSG);
+
+ if (skipThisLoop)
+ // Do not run other passes on this loop.
+ break;
+ }
+
+ // If the loop was deleted, release all the loop passes. This frees up
+ // some memory, and avoids trouble with the pass manager trying to call
+ // verifyAnalysis on them.
+ if (skipThisLoop)
+ for (unsigned Index = 0; Index < getNumContainedPasses(); ++Index) {
+ Pass *P = getContainedPass(Index);
+ freePass(P, "<deleted>", ON_LOOP_MSG);
+ }
+
+ // Pop the loop from queue after running all passes.
+ LQ.pop_back();
+
+ if (redoThisLoop)
+ LQ.push_back(CurrentLoop);
+ }
+
+ // Finalization
+ for (unsigned Index = 0; Index < getNumContainedPasses(); ++Index) {
+ LoopPass *P = (LoopPass *)getContainedPass(Index);
+ Changed |= P->doFinalization();
+ }
+
+ return Changed;
+}
+
+/// Print passes managed by this manager
+void LPPassManager::dumpPassStructure(unsigned Offset) {
+ errs().indent(Offset*2) << "Loop Pass Manager\n";
+ for (unsigned Index = 0; Index < getNumContainedPasses(); ++Index) {
+ Pass *P = getContainedPass(Index);
+ P->dumpPassStructure(Offset + 1);
+ dumpLastUses(P, Offset+1);
+ }
+}
+
+
+//===----------------------------------------------------------------------===//
+// LoopPass
+
+// Check if this pass is suitable for the current LPPassManager, if
+// available. This pass P is not suitable for a LPPassManager if P
+// is not preserving higher level analysis info used by other
+// LPPassManager passes. In such case, pop LPPassManager from the
+// stack. This will force assignPassManager() to create new
+// LPPassManger as expected.
+void LoopPass::preparePassManager(PMStack &PMS) {
+
+ // Find LPPassManager
+ while (!PMS.empty() &&
+ PMS.top()->getPassManagerType() > PMT_LoopPassManager)
+ PMS.pop();
+
+ // If this pass is destroying high level information that is used
+ // by other passes that are managed by LPM then do not insert
+ // this pass in current LPM. Use new LPPassManager.
+ if (PMS.top()->getPassManagerType() == PMT_LoopPassManager &&
+ !PMS.top()->preserveHigherLevelAnalysis(this))
+ PMS.pop();
+}
+
+/// Assign pass manager to manage this pass.
+void LoopPass::assignPassManager(PMStack &PMS,
+ PassManagerType PreferredType) {
+ // Find LPPassManager
+ while (!PMS.empty() &&
+ PMS.top()->getPassManagerType() > PMT_LoopPassManager)
+ PMS.pop();
+
+ LPPassManager *LPPM;
+ if (PMS.top()->getPassManagerType() == PMT_LoopPassManager)
+ LPPM = (LPPassManager*)PMS.top();
+ else {
+ // Create new Loop Pass Manager if it does not exist.
+ assert (!PMS.empty() && "Unable to create Loop Pass Manager");
+ PMDataManager *PMD = PMS.top();
+
+ // [1] Create new Call Graph Pass Manager
+ LPPM = new LPPassManager(PMD->getDepth() + 1);
+ LPPM->populateInheritedAnalysis(PMS);
+
+ // [2] Set up new manager's top level manager
+ PMTopLevelManager *TPM = PMD->getTopLevelManager();
+ TPM->addIndirectPassManager(LPPM);
+
+ // [3] Assign manager to manage this new manager. This may create
+ // and push new managers into PMS
+ Pass *P = LPPM->getAsPass();
+ TPM->schedulePass(P);
+
+ // [4] Push new manager into PMS
+ PMS.push(LPPM);
+ }
+
+ LPPM->add(this);
+}
diff --git a/lib/Analysis/Makefile b/lib/Analysis/Makefile
new file mode 100644
index 0000000..4af6d35
--- /dev/null
+++ b/lib/Analysis/Makefile
@@ -0,0 +1,16 @@
+##===- lib/Analysis/Makefile -------------------------------*- Makefile -*-===##
+#
+# The LLVM Compiler Infrastructure
+#
+# This file is distributed under the University of Illinois Open Source
+# License. See LICENSE.TXT for details.
+#
+##===----------------------------------------------------------------------===##
+
+LEVEL = ../..
+LIBRARYNAME = LLVMAnalysis
+DIRS = IPA
+BUILD_ARCHIVE = 1
+
+include $(LEVEL)/Makefile.common
+
diff --git a/lib/Analysis/MemoryBuiltins.cpp b/lib/Analysis/MemoryBuiltins.cpp
new file mode 100644
index 0000000..297b588
--- /dev/null
+++ b/lib/Analysis/MemoryBuiltins.cpp
@@ -0,0 +1,207 @@
+//===------ MemoryBuiltins.cpp - Identify calls to memory builtins --------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This family of functions identifies calls to builtin functions that allocate
+// or free memory.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/MemoryBuiltins.h"
+#include "llvm/Constants.h"
+#include "llvm/Instructions.h"
+#include "llvm/Module.h"
+#include "llvm/Analysis/ValueTracking.h"
+#include "llvm/Target/TargetData.h"
+using namespace llvm;
+
+//===----------------------------------------------------------------------===//
+// malloc Call Utility Functions.
+//
+
+/// isMalloc - Returns true if the value is either a malloc call or a
+/// bitcast of the result of a malloc call.
+bool llvm::isMalloc(const Value *I) {
+ return extractMallocCall(I) || extractMallocCallFromBitCast(I);
+}
+
+static bool isMallocCall(const CallInst *CI) {
+ if (!CI)
+ return false;
+
+ Function *Callee = CI->getCalledFunction();
+ if (Callee == 0 || !Callee->isDeclaration() || Callee->getName() != "malloc")
+ return false;
+
+ // Check malloc prototype.
+ // FIXME: workaround for PR5130, this will be obsolete when a nobuiltin
+ // attribute will exist.
+ const FunctionType *FTy = Callee->getFunctionType();
+ if (FTy->getNumParams() != 1)
+ return false;
+ if (IntegerType *ITy = dyn_cast<IntegerType>(FTy->param_begin()->get())) {
+ if (ITy->getBitWidth() != 32 && ITy->getBitWidth() != 64)
+ return false;
+ return true;
+ }
+
+ return false;
+}
+
+/// extractMallocCall - Returns the corresponding CallInst if the instruction
+/// is a malloc call. Since CallInst::CreateMalloc() only creates calls, we
+/// ignore InvokeInst here.
+const CallInst *llvm::extractMallocCall(const Value *I) {
+ const CallInst *CI = dyn_cast<CallInst>(I);
+ return (isMallocCall(CI)) ? CI : NULL;
+}
+
+CallInst *llvm::extractMallocCall(Value *I) {
+ CallInst *CI = dyn_cast<CallInst>(I);
+ return (isMallocCall(CI)) ? CI : NULL;
+}
+
+static bool isBitCastOfMallocCall(const BitCastInst *BCI) {
+ if (!BCI)
+ return false;
+
+ return isMallocCall(dyn_cast<CallInst>(BCI->getOperand(0)));
+}
+
+/// extractMallocCallFromBitCast - Returns the corresponding CallInst if the
+/// instruction is a bitcast of the result of a malloc call.
+CallInst *llvm::extractMallocCallFromBitCast(Value *I) {
+ BitCastInst *BCI = dyn_cast<BitCastInst>(I);
+ return (isBitCastOfMallocCall(BCI)) ? cast<CallInst>(BCI->getOperand(0))
+ : NULL;
+}
+
+const CallInst *llvm::extractMallocCallFromBitCast(const Value *I) {
+ const BitCastInst *BCI = dyn_cast<BitCastInst>(I);
+ return (isBitCastOfMallocCall(BCI)) ? cast<CallInst>(BCI->getOperand(0))
+ : NULL;
+}
+
+static Value *computeArraySize(const CallInst *CI, const TargetData *TD,
+ bool LookThroughSExt = false) {
+ if (!CI)
+ return NULL;
+
+ // The size of the malloc's result type must be known to determine array size.
+ const Type *T = getMallocAllocatedType(CI);
+ if (!T || !T->isSized() || !TD)
+ return NULL;
+
+ unsigned ElementSize = TD->getTypeAllocSize(T);
+ if (const StructType *ST = dyn_cast<StructType>(T))
+ ElementSize = TD->getStructLayout(ST)->getSizeInBytes();
+
+ // If malloc calls' arg can be determined to be a multiple of ElementSize,
+ // return the multiple. Otherwise, return NULL.
+ Value *MallocArg = CI->getOperand(1);
+ Value *Multiple = NULL;
+ if (ComputeMultiple(MallocArg, ElementSize, Multiple,
+ LookThroughSExt))
+ return Multiple;
+
+ return NULL;
+}
+
+/// isArrayMalloc - Returns the corresponding CallInst if the instruction
+/// is a call to malloc whose array size can be determined and the array size
+/// is not constant 1. Otherwise, return NULL.
+const CallInst *llvm::isArrayMalloc(const Value *I, const TargetData *TD) {
+ const CallInst *CI = extractMallocCall(I);
+ Value *ArraySize = computeArraySize(CI, TD);
+
+ if (ArraySize &&
+ ArraySize != ConstantInt::get(CI->getOperand(1)->getType(), 1))
+ return CI;
+
+ // CI is a non-array malloc or we can't figure out that it is an array malloc.
+ return NULL;
+}
+
+/// getMallocType - Returns the PointerType resulting from the malloc call.
+/// The PointerType depends on the number of bitcast uses of the malloc call:
+/// 0: PointerType is the calls' return type.
+/// 1: PointerType is the bitcast's result type.
+/// >1: Unique PointerType cannot be determined, return NULL.
+const PointerType *llvm::getMallocType(const CallInst *CI) {
+ assert(isMalloc(CI) && "getMallocType and not malloc call");
+
+ const PointerType *MallocType = NULL;
+ unsigned NumOfBitCastUses = 0;
+
+ // Determine if CallInst has a bitcast use.
+ for (Value::use_const_iterator UI = CI->use_begin(), E = CI->use_end();
+ UI != E; )
+ if (const BitCastInst *BCI = dyn_cast<BitCastInst>(*UI++)) {
+ MallocType = cast<PointerType>(BCI->getDestTy());
+ NumOfBitCastUses++;
+ }
+
+ // Malloc call has 1 bitcast use, so type is the bitcast's destination type.
+ if (NumOfBitCastUses == 1)
+ return MallocType;
+
+ // Malloc call was not bitcast, so type is the malloc function's return type.
+ if (NumOfBitCastUses == 0)
+ return cast<PointerType>(CI->getType());
+
+ // Type could not be determined.
+ return NULL;
+}
+
+/// getMallocAllocatedType - Returns the Type allocated by malloc call.
+/// The Type depends on the number of bitcast uses of the malloc call:
+/// 0: PointerType is the malloc calls' return type.
+/// 1: PointerType is the bitcast's result type.
+/// >1: Unique PointerType cannot be determined, return NULL.
+const Type *llvm::getMallocAllocatedType(const CallInst *CI) {
+ const PointerType *PT = getMallocType(CI);
+ return PT ? PT->getElementType() : NULL;
+}
+
+/// getMallocArraySize - Returns the array size of a malloc call. If the
+/// argument passed to malloc is a multiple of the size of the malloced type,
+/// then return that multiple. For non-array mallocs, the multiple is
+/// constant 1. Otherwise, return NULL for mallocs whose array size cannot be
+/// determined.
+Value *llvm::getMallocArraySize(CallInst *CI, const TargetData *TD,
+ bool LookThroughSExt) {
+ assert(isMalloc(CI) && "getMallocArraySize and not malloc call");
+ return computeArraySize(CI, TD, LookThroughSExt);
+}
+
+//===----------------------------------------------------------------------===//
+// free Call Utility Functions.
+//
+
+/// isFreeCall - Returns true if the value is a call to the builtin free()
+bool llvm::isFreeCall(const Value *I) {
+ const CallInst *CI = dyn_cast<CallInst>(I);
+ if (!CI)
+ return false;
+ Function *Callee = CI->getCalledFunction();
+ if (Callee == 0 || !Callee->isDeclaration() || Callee->getName() != "free")
+ return false;
+
+ // Check free prototype.
+ // FIXME: workaround for PR5130, this will be obsolete when a nobuiltin
+ // attribute will exist.
+ const FunctionType *FTy = Callee->getFunctionType();
+ if (!FTy->getReturnType()->isVoidTy())
+ return false;
+ if (FTy->getNumParams() != 1)
+ return false;
+ if (FTy->param_begin()->get() != Type::getInt8PtrTy(Callee->getContext()))
+ return false;
+
+ return true;
+}
diff --git a/lib/Analysis/MemoryDependenceAnalysis.cpp b/lib/Analysis/MemoryDependenceAnalysis.cpp
new file mode 100644
index 0000000..2d74709d
--- /dev/null
+++ b/lib/Analysis/MemoryDependenceAnalysis.cpp
@@ -0,0 +1,1245 @@
+//===- MemoryDependenceAnalysis.cpp - Mem Deps Implementation --*- C++ -*-===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements an analysis that determines, for a given memory
+// operation, what preceding memory operations it depends on. It builds on
+// alias analysis information, and tries to provide a lazy, caching interface to
+// a common kind of alias information query.
+//
+//===----------------------------------------------------------------------===//
+
+#define DEBUG_TYPE "memdep"
+#include "llvm/Analysis/MemoryDependenceAnalysis.h"
+#include "llvm/Instructions.h"
+#include "llvm/IntrinsicInst.h"
+#include "llvm/Function.h"
+#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/Analysis/Dominators.h"
+#include "llvm/Analysis/InstructionSimplify.h"
+#include "llvm/Analysis/MemoryBuiltins.h"
+#include "llvm/Analysis/PHITransAddr.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/ADT/STLExtras.h"
+#include "llvm/Support/PredIteratorCache.h"
+#include "llvm/Support/Debug.h"
+using namespace llvm;
+
+STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses");
+STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses");
+STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses");
+
+STATISTIC(NumCacheNonLocalPtr,
+ "Number of fully cached non-local ptr responses");
+STATISTIC(NumCacheDirtyNonLocalPtr,
+ "Number of cached, but dirty, non-local ptr responses");
+STATISTIC(NumUncacheNonLocalPtr,
+ "Number of uncached non-local ptr responses");
+STATISTIC(NumCacheCompleteNonLocalPtr,
+ "Number of block queries that were completely cached");
+
+char MemoryDependenceAnalysis::ID = 0;
+
+// Register this pass...
+static RegisterPass<MemoryDependenceAnalysis> X("memdep",
+ "Memory Dependence Analysis", false, true);
+
+MemoryDependenceAnalysis::MemoryDependenceAnalysis()
+: FunctionPass(&ID), PredCache(0) {
+}
+MemoryDependenceAnalysis::~MemoryDependenceAnalysis() {
+}
+
+/// Clean up memory in between runs
+void MemoryDependenceAnalysis::releaseMemory() {
+ LocalDeps.clear();
+ NonLocalDeps.clear();
+ NonLocalPointerDeps.clear();
+ ReverseLocalDeps.clear();
+ ReverseNonLocalDeps.clear();
+ ReverseNonLocalPtrDeps.clear();
+ PredCache->clear();
+}
+
+
+
+/// getAnalysisUsage - Does not modify anything. It uses Alias Analysis.
+///
+void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.setPreservesAll();
+ AU.addRequiredTransitive<AliasAnalysis>();
+}
+
+bool MemoryDependenceAnalysis::runOnFunction(Function &) {
+ AA = &getAnalysis<AliasAnalysis>();
+ if (PredCache == 0)
+ PredCache.reset(new PredIteratorCache());
+ return false;
+}
+
+/// RemoveFromReverseMap - This is a helper function that removes Val from
+/// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry.
+template <typename KeyTy>
+static void RemoveFromReverseMap(DenseMap<Instruction*,
+ SmallPtrSet<KeyTy, 4> > &ReverseMap,
+ Instruction *Inst, KeyTy Val) {
+ typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::iterator
+ InstIt = ReverseMap.find(Inst);
+ assert(InstIt != ReverseMap.end() && "Reverse map out of sync?");
+ bool Found = InstIt->second.erase(Val);
+ assert(Found && "Invalid reverse map!"); Found=Found;
+ if (InstIt->second.empty())
+ ReverseMap.erase(InstIt);
+}
+
+
+/// getCallSiteDependencyFrom - Private helper for finding the local
+/// dependencies of a call site.
+MemDepResult MemoryDependenceAnalysis::
+getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall,
+ BasicBlock::iterator ScanIt, BasicBlock *BB) {
+ // Walk backwards through the block, looking for dependencies
+ while (ScanIt != BB->begin()) {
+ Instruction *Inst = --ScanIt;
+
+ // If this inst is a memory op, get the pointer it accessed
+ Value *Pointer = 0;
+ uint64_t PointerSize = 0;
+ if (StoreInst *S = dyn_cast<StoreInst>(Inst)) {
+ Pointer = S->getPointerOperand();
+ PointerSize = AA->getTypeStoreSize(S->getOperand(0)->getType());
+ } else if (VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
+ Pointer = V->getOperand(0);
+ PointerSize = AA->getTypeStoreSize(V->getType());
+ } else if (isFreeCall(Inst)) {
+ Pointer = Inst->getOperand(1);
+ // calls to free() erase the entire structure
+ PointerSize = ~0ULL;
+ } else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
+ // Debug intrinsics don't cause dependences.
+ if (isa<DbgInfoIntrinsic>(Inst)) continue;
+ CallSite InstCS = CallSite::get(Inst);
+ // If these two calls do not interfere, look past it.
+ switch (AA->getModRefInfo(CS, InstCS)) {
+ case AliasAnalysis::NoModRef:
+ // If the two calls don't interact (e.g. InstCS is readnone) keep
+ // scanning.
+ continue;
+ case AliasAnalysis::Ref:
+ // If the two calls read the same memory locations and CS is a readonly
+ // function, then we have two cases: 1) the calls may not interfere with
+ // each other at all. 2) the calls may produce the same value. In case
+ // #1 we want to ignore the values, in case #2, we want to return Inst
+ // as a Def dependence. This allows us to CSE in cases like:
+ // X = strlen(P);
+ // memchr(...);
+ // Y = strlen(P); // Y = X
+ if (isReadOnlyCall) {
+ if (CS.getCalledFunction() != 0 &&
+ CS.getCalledFunction() == InstCS.getCalledFunction())
+ return MemDepResult::getDef(Inst);
+ // Ignore unrelated read/read call dependences.
+ continue;
+ }
+ // FALL THROUGH
+ default:
+ return MemDepResult::getClobber(Inst);
+ }
+ } else {
+ // Non-memory instruction.
+ continue;
+ }
+
+ if (AA->getModRefInfo(CS, Pointer, PointerSize) != AliasAnalysis::NoModRef)
+ return MemDepResult::getClobber(Inst);
+ }
+
+ // No dependence found. If this is the entry block of the function, it is a
+ // clobber, otherwise it is non-local.
+ if (BB != &BB->getParent()->getEntryBlock())
+ return MemDepResult::getNonLocal();
+ return MemDepResult::getClobber(ScanIt);
+}
+
+/// getPointerDependencyFrom - Return the instruction on which a memory
+/// location depends. If isLoad is true, this routine ignore may-aliases with
+/// read-only operations.
+MemDepResult MemoryDependenceAnalysis::
+getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad,
+ BasicBlock::iterator ScanIt, BasicBlock *BB) {
+
+ Value *InvariantTag = 0;
+
+ // Walk backwards through the basic block, looking for dependencies.
+ while (ScanIt != BB->begin()) {
+ Instruction *Inst = --ScanIt;
+
+ // If we're in an invariant region, no dependencies can be found before
+ // we pass an invariant-begin marker.
+ if (InvariantTag == Inst) {
+ InvariantTag = 0;
+ continue;
+ }
+
+ if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
+ // Debug intrinsics don't cause dependences.
+ if (isa<DbgInfoIntrinsic>(Inst)) continue;
+
+ // If we pass an invariant-end marker, then we've just entered an
+ // invariant region and can start ignoring dependencies.
+ if (II->getIntrinsicID() == Intrinsic::invariant_end) {
+ // FIXME: This only considers queries directly on the invariant-tagged
+ // pointer, not on query pointers that are indexed off of them. It'd
+ // be nice to handle that at some point.
+ AliasAnalysis::AliasResult R =
+ AA->alias(II->getOperand(3), ~0U, MemPtr, ~0U);
+ if (R == AliasAnalysis::MustAlias) {
+ InvariantTag = II->getOperand(1);
+ continue;
+ }
+
+ // If we reach a lifetime begin or end marker, then the query ends here
+ // because the value is undefined.
+ } else if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
+ // FIXME: This only considers queries directly on the invariant-tagged
+ // pointer, not on query pointers that are indexed off of them. It'd
+ // be nice to handle that at some point.
+ AliasAnalysis::AliasResult R =
+ AA->alias(II->getOperand(2), ~0U, MemPtr, ~0U);
+ if (R == AliasAnalysis::MustAlias)
+ return MemDepResult::getDef(II);
+ }
+ }
+
+ // If we're querying on a load and we're in an invariant region, we're done
+ // at this point. Nothing a load depends on can live in an invariant region.
+ if (isLoad && InvariantTag) continue;
+
+ // Values depend on loads if the pointers are must aliased. This means that
+ // a load depends on another must aliased load from the same value.
+ if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
+ Value *Pointer = LI->getPointerOperand();
+ uint64_t PointerSize = AA->getTypeStoreSize(LI->getType());
+
+ // If we found a pointer, check if it could be the same as our pointer.
+ AliasAnalysis::AliasResult R =
+ AA->alias(Pointer, PointerSize, MemPtr, MemSize);
+ if (R == AliasAnalysis::NoAlias)
+ continue;
+
+ // May-alias loads don't depend on each other without a dependence.
+ if (isLoad && R == AliasAnalysis::MayAlias)
+ continue;
+ // Stores depend on may and must aliased loads, loads depend on must-alias
+ // loads.
+ return MemDepResult::getDef(Inst);
+ }
+
+ if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
+ // There can't be stores to the value we care about inside an
+ // invariant region.
+ if (InvariantTag) continue;
+
+ // If alias analysis can tell that this store is guaranteed to not modify
+ // the query pointer, ignore it. Use getModRefInfo to handle cases where
+ // the query pointer points to constant memory etc.
+ if (AA->getModRefInfo(SI, MemPtr, MemSize) == AliasAnalysis::NoModRef)
+ continue;
+
+ // Ok, this store might clobber the query pointer. Check to see if it is
+ // a must alias: in this case, we want to return this as a def.
+ Value *Pointer = SI->getPointerOperand();
+ uint64_t PointerSize = AA->getTypeStoreSize(SI->getOperand(0)->getType());
+
+ // If we found a pointer, check if it could be the same as our pointer.
+ AliasAnalysis::AliasResult R =
+ AA->alias(Pointer, PointerSize, MemPtr, MemSize);
+
+ if (R == AliasAnalysis::NoAlias)
+ continue;
+ if (R == AliasAnalysis::MayAlias)
+ return MemDepResult::getClobber(Inst);
+ return MemDepResult::getDef(Inst);
+ }
+
+ // If this is an allocation, and if we know that the accessed pointer is to
+ // the allocation, return Def. This means that there is no dependence and
+ // the access can be optimized based on that. For example, a load could
+ // turn into undef.
+ // Note: Only determine this to be a malloc if Inst is the malloc call, not
+ // a subsequent bitcast of the malloc call result. There can be stores to
+ // the malloced memory between the malloc call and its bitcast uses, and we
+ // need to continue scanning until the malloc call.
+ if (isa<AllocaInst>(Inst) ||
+ (isa<CallInst>(Inst) && extractMallocCall(Inst))) {
+ Value *AccessPtr = MemPtr->getUnderlyingObject();
+
+ if (AccessPtr == Inst ||
+ AA->alias(Inst, 1, AccessPtr, 1) == AliasAnalysis::MustAlias)
+ return MemDepResult::getDef(Inst);
+ continue;
+ }
+
+ // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
+ switch (AA->getModRefInfo(Inst, MemPtr, MemSize)) {
+ case AliasAnalysis::NoModRef:
+ // If the call has no effect on the queried pointer, just ignore it.
+ continue;
+ case AliasAnalysis::Mod:
+ // If we're in an invariant region, we can ignore calls that ONLY
+ // modify the pointer.
+ if (InvariantTag) continue;
+ return MemDepResult::getClobber(Inst);
+ case AliasAnalysis::Ref:
+ // If the call is known to never store to the pointer, and if this is a
+ // load query, we can safely ignore it (scan past it).
+ if (isLoad)
+ continue;
+ default:
+ // Otherwise, there is a potential dependence. Return a clobber.
+ return MemDepResult::getClobber(Inst);
+ }
+ }
+
+ // No dependence found. If this is the entry block of the function, it is a
+ // clobber, otherwise it is non-local.
+ if (BB != &BB->getParent()->getEntryBlock())
+ return MemDepResult::getNonLocal();
+ return MemDepResult::getClobber(ScanIt);
+}
+
+/// getDependency - Return the instruction on which a memory operation
+/// depends.
+MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
+ Instruction *ScanPos = QueryInst;
+
+ // Check for a cached result
+ MemDepResult &LocalCache = LocalDeps[QueryInst];
+
+ // If the cached entry is non-dirty, just return it. Note that this depends
+ // on MemDepResult's default constructing to 'dirty'.
+ if (!LocalCache.isDirty())
+ return LocalCache;
+
+ // Otherwise, if we have a dirty entry, we know we can start the scan at that
+ // instruction, which may save us some work.
+ if (Instruction *Inst = LocalCache.getInst()) {
+ ScanPos = Inst;
+
+ RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
+ }
+
+ BasicBlock *QueryParent = QueryInst->getParent();
+
+ Value *MemPtr = 0;
+ uint64_t MemSize = 0;
+
+ // Do the scan.
+ if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
+ // No dependence found. If this is the entry block of the function, it is a
+ // clobber, otherwise it is non-local.
+ if (QueryParent != &QueryParent->getParent()->getEntryBlock())
+ LocalCache = MemDepResult::getNonLocal();
+ else
+ LocalCache = MemDepResult::getClobber(QueryInst);
+ } else if (StoreInst *SI = dyn_cast<StoreInst>(QueryInst)) {
+ // If this is a volatile store, don't mess around with it. Just return the
+ // previous instruction as a clobber.
+ if (SI->isVolatile())
+ LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
+ else {
+ MemPtr = SI->getPointerOperand();
+ MemSize = AA->getTypeStoreSize(SI->getOperand(0)->getType());
+ }
+ } else if (LoadInst *LI = dyn_cast<LoadInst>(QueryInst)) {
+ // If this is a volatile load, don't mess around with it. Just return the
+ // previous instruction as a clobber.
+ if (LI->isVolatile())
+ LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
+ else {
+ MemPtr = LI->getPointerOperand();
+ MemSize = AA->getTypeStoreSize(LI->getType());
+ }
+ } else if (isFreeCall(QueryInst)) {
+ MemPtr = QueryInst->getOperand(1);
+ // calls to free() erase the entire structure, not just a field.
+ MemSize = ~0UL;
+ } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
+ int IntrinsicID = 0; // Intrinsic IDs start at 1.
+ if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst))
+ IntrinsicID = II->getIntrinsicID();
+
+ switch (IntrinsicID) {
+ case Intrinsic::lifetime_start:
+ case Intrinsic::lifetime_end:
+ case Intrinsic::invariant_start:
+ MemPtr = QueryInst->getOperand(2);
+ MemSize = cast<ConstantInt>(QueryInst->getOperand(1))->getZExtValue();
+ break;
+ case Intrinsic::invariant_end:
+ MemPtr = QueryInst->getOperand(3);
+ MemSize = cast<ConstantInt>(QueryInst->getOperand(2))->getZExtValue();
+ break;
+ default:
+ CallSite QueryCS = CallSite::get(QueryInst);
+ bool isReadOnly = AA->onlyReadsMemory(QueryCS);
+ LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos,
+ QueryParent);
+ break;
+ }
+ } else {
+ // Non-memory instruction.
+ LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
+ }
+
+ // If we need to do a pointer scan, make it happen.
+ if (MemPtr) {
+ bool isLoad = !QueryInst->mayWriteToMemory();
+ if (IntrinsicInst *II = dyn_cast<MemoryUseIntrinsic>(QueryInst)) {
+ isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_end;
+ }
+ LocalCache = getPointerDependencyFrom(MemPtr, MemSize, isLoad, ScanPos,
+ QueryParent);
+ }
+
+ // Remember the result!
+ if (Instruction *I = LocalCache.getInst())
+ ReverseLocalDeps[I].insert(QueryInst);
+
+ return LocalCache;
+}
+
+#ifndef NDEBUG
+/// AssertSorted - This method is used when -debug is specified to verify that
+/// cache arrays are properly kept sorted.
+static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
+ int Count = -1) {
+ if (Count == -1) Count = Cache.size();
+ if (Count == 0) return;
+
+ for (unsigned i = 1; i != unsigned(Count); ++i)
+ assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!");
+}
+#endif
+
+/// getNonLocalCallDependency - Perform a full dependency query for the
+/// specified call, returning the set of blocks that the value is
+/// potentially live across. The returned set of results will include a
+/// "NonLocal" result for all blocks where the value is live across.
+///
+/// This method assumes the instruction returns a "NonLocal" dependency
+/// within its own block.
+///
+/// This returns a reference to an internal data structure that may be
+/// invalidated on the next non-local query or when an instruction is
+/// removed. Clients must copy this data if they want it around longer than
+/// that.
+const MemoryDependenceAnalysis::NonLocalDepInfo &
+MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
+ assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
+ "getNonLocalCallDependency should only be used on calls with non-local deps!");
+ PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
+ NonLocalDepInfo &Cache = CacheP.first;
+
+ /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
+ /// the cached case, this can happen due to instructions being deleted etc. In
+ /// the uncached case, this starts out as the set of predecessors we care
+ /// about.
+ SmallVector<BasicBlock*, 32> DirtyBlocks;
+
+ if (!Cache.empty()) {
+ // Okay, we have a cache entry. If we know it is not dirty, just return it
+ // with no computation.
+ if (!CacheP.second) {
+ NumCacheNonLocal++;
+ return Cache;
+ }
+
+ // If we already have a partially computed set of results, scan them to
+ // determine what is dirty, seeding our initial DirtyBlocks worklist.
+ for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
+ I != E; ++I)
+ if (I->getResult().isDirty())
+ DirtyBlocks.push_back(I->getBB());
+
+ // Sort the cache so that we can do fast binary search lookups below.
+ std::sort(Cache.begin(), Cache.end());
+
+ ++NumCacheDirtyNonLocal;
+ //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
+ // << Cache.size() << " cached: " << *QueryInst;
+ } else {
+ // Seed DirtyBlocks with each of the preds of QueryInst's block.
+ BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
+ for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI)
+ DirtyBlocks.push_back(*PI);
+ NumUncacheNonLocal++;
+ }
+
+ // isReadonlyCall - If this is a read-only call, we can be more aggressive.
+ bool isReadonlyCall = AA->onlyReadsMemory(QueryCS);
+
+ SmallPtrSet<BasicBlock*, 64> Visited;
+
+ unsigned NumSortedEntries = Cache.size();
+ DEBUG(AssertSorted(Cache));
+
+ // Iterate while we still have blocks to update.
+ while (!DirtyBlocks.empty()) {
+ BasicBlock *DirtyBB = DirtyBlocks.back();
+ DirtyBlocks.pop_back();
+
+ // Already processed this block?
+ if (!Visited.insert(DirtyBB))
+ continue;
+
+ // Do a binary search to see if we already have an entry for this block in
+ // the cache set. If so, find it.
+ DEBUG(AssertSorted(Cache, NumSortedEntries));
+ NonLocalDepInfo::iterator Entry =
+ std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
+ NonLocalDepEntry(DirtyBB));
+ if (Entry != Cache.begin() && prior(Entry)->getBB() == DirtyBB)
+ --Entry;
+
+ NonLocalDepEntry *ExistingResult = 0;
+ if (Entry != Cache.begin()+NumSortedEntries &&
+ Entry->getBB() == DirtyBB) {
+ // If we already have an entry, and if it isn't already dirty, the block
+ // is done.
+ if (!Entry->getResult().isDirty())
+ continue;
+
+ // Otherwise, remember this slot so we can update the value.
+ ExistingResult = &*Entry;
+ }
+
+ // If the dirty entry has a pointer, start scanning from it so we don't have
+ // to rescan the entire block.
+ BasicBlock::iterator ScanPos = DirtyBB->end();
+ if (ExistingResult) {
+ if (Instruction *Inst = ExistingResult->getResult().getInst()) {
+ ScanPos = Inst;
+ // We're removing QueryInst's use of Inst.
+ RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
+ QueryCS.getInstruction());
+ }
+ }
+
+ // Find out if this block has a local dependency for QueryInst.
+ MemDepResult Dep;
+
+ if (ScanPos != DirtyBB->begin()) {
+ Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB);
+ } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
+ // No dependence found. If this is the entry block of the function, it is
+ // a clobber, otherwise it is non-local.
+ Dep = MemDepResult::getNonLocal();
+ } else {
+ Dep = MemDepResult::getClobber(ScanPos);
+ }
+
+ // If we had a dirty entry for the block, update it. Otherwise, just add
+ // a new entry.
+ if (ExistingResult)
+ ExistingResult->setResult(Dep);
+ else
+ Cache.push_back(NonLocalDepEntry(DirtyBB, Dep));
+
+ // If the block has a dependency (i.e. it isn't completely transparent to
+ // the value), remember the association!
+ if (!Dep.isNonLocal()) {
+ // Keep the ReverseNonLocalDeps map up to date so we can efficiently
+ // update this when we remove instructions.
+ if (Instruction *Inst = Dep.getInst())
+ ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
+ } else {
+
+ // If the block *is* completely transparent to the load, we need to check
+ // the predecessors of this block. Add them to our worklist.
+ for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI)
+ DirtyBlocks.push_back(*PI);
+ }
+ }
+
+ return Cache;
+}
+
+/// getNonLocalPointerDependency - Perform a full dependency query for an
+/// access to the specified (non-volatile) memory location, returning the
+/// set of instructions that either define or clobber the value.
+///
+/// This method assumes the pointer has a "NonLocal" dependency within its
+/// own block.
+///
+void MemoryDependenceAnalysis::
+getNonLocalPointerDependency(Value *Pointer, bool isLoad, BasicBlock *FromBB,
+ SmallVectorImpl<NonLocalDepResult> &Result) {
+ assert(isa<PointerType>(Pointer->getType()) &&
+ "Can't get pointer deps of a non-pointer!");
+ Result.clear();
+
+ // We know that the pointer value is live into FromBB find the def/clobbers
+ // from presecessors.
+ const Type *EltTy = cast<PointerType>(Pointer->getType())->getElementType();
+ uint64_t PointeeSize = AA->getTypeStoreSize(EltTy);
+
+ PHITransAddr Address(Pointer, TD);
+
+ // This is the set of blocks we've inspected, and the pointer we consider in
+ // each block. Because of critical edges, we currently bail out if querying
+ // a block with multiple different pointers. This can happen during PHI
+ // translation.
+ DenseMap<BasicBlock*, Value*> Visited;
+ if (!getNonLocalPointerDepFromBB(Address, PointeeSize, isLoad, FromBB,
+ Result, Visited, true))
+ return;
+ Result.clear();
+ Result.push_back(NonLocalDepResult(FromBB,
+ MemDepResult::getClobber(FromBB->begin()),
+ Pointer));
+}
+
+/// GetNonLocalInfoForBlock - Compute the memdep value for BB with
+/// Pointer/PointeeSize using either cached information in Cache or by doing a
+/// lookup (which may use dirty cache info if available). If we do a lookup,
+/// add the result to the cache.
+MemDepResult MemoryDependenceAnalysis::
+GetNonLocalInfoForBlock(Value *Pointer, uint64_t PointeeSize,
+ bool isLoad, BasicBlock *BB,
+ NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
+
+ // Do a binary search to see if we already have an entry for this block in
+ // the cache set. If so, find it.
+ NonLocalDepInfo::iterator Entry =
+ std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
+ NonLocalDepEntry(BB));
+ if (Entry != Cache->begin() && (Entry-1)->getBB() == BB)
+ --Entry;
+
+ NonLocalDepEntry *ExistingResult = 0;
+ if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB)
+ ExistingResult = &*Entry;
+
+ // If we have a cached entry, and it is non-dirty, use it as the value for
+ // this dependency.
+ if (ExistingResult && !ExistingResult->getResult().isDirty()) {
+ ++NumCacheNonLocalPtr;
+ return ExistingResult->getResult();
+ }
+
+ // Otherwise, we have to scan for the value. If we have a dirty cache
+ // entry, start scanning from its position, otherwise we scan from the end
+ // of the block.
+ BasicBlock::iterator ScanPos = BB->end();
+ if (ExistingResult && ExistingResult->getResult().getInst()) {
+ assert(ExistingResult->getResult().getInst()->getParent() == BB &&
+ "Instruction invalidated?");
+ ++NumCacheDirtyNonLocalPtr;
+ ScanPos = ExistingResult->getResult().getInst();
+
+ // Eliminating the dirty entry from 'Cache', so update the reverse info.
+ ValueIsLoadPair CacheKey(Pointer, isLoad);
+ RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey);
+ } else {
+ ++NumUncacheNonLocalPtr;
+ }
+
+ // Scan the block for the dependency.
+ MemDepResult Dep = getPointerDependencyFrom(Pointer, PointeeSize, isLoad,
+ ScanPos, BB);
+
+ // If we had a dirty entry for the block, update it. Otherwise, just add
+ // a new entry.
+ if (ExistingResult)
+ ExistingResult->setResult(Dep);
+ else
+ Cache->push_back(NonLocalDepEntry(BB, Dep));
+
+ // If the block has a dependency (i.e. it isn't completely transparent to
+ // the value), remember the reverse association because we just added it
+ // to Cache!
+ if (Dep.isNonLocal())
+ return Dep;
+
+ // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
+ // update MemDep when we remove instructions.
+ Instruction *Inst = Dep.getInst();
+ assert(Inst && "Didn't depend on anything?");
+ ValueIsLoadPair CacheKey(Pointer, isLoad);
+ ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
+ return Dep;
+}
+
+/// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain
+/// number of elements in the array that are already properly ordered. This is
+/// optimized for the case when only a few entries are added.
+static void
+SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
+ unsigned NumSortedEntries) {
+ switch (Cache.size() - NumSortedEntries) {
+ case 0:
+ // done, no new entries.
+ break;
+ case 2: {
+ // Two new entries, insert the last one into place.
+ NonLocalDepEntry Val = Cache.back();
+ Cache.pop_back();
+ MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
+ std::upper_bound(Cache.begin(), Cache.end()-1, Val);
+ Cache.insert(Entry, Val);
+ // FALL THROUGH.
+ }
+ case 1:
+ // One new entry, Just insert the new value at the appropriate position.
+ if (Cache.size() != 1) {
+ NonLocalDepEntry Val = Cache.back();
+ Cache.pop_back();
+ MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
+ std::upper_bound(Cache.begin(), Cache.end(), Val);
+ Cache.insert(Entry, Val);
+ }
+ break;
+ default:
+ // Added many values, do a full scale sort.
+ std::sort(Cache.begin(), Cache.end());
+ break;
+ }
+}
+
+/// getNonLocalPointerDepFromBB - Perform a dependency query based on
+/// pointer/pointeesize starting at the end of StartBB. Add any clobber/def
+/// results to the results vector and keep track of which blocks are visited in
+/// 'Visited'.
+///
+/// This has special behavior for the first block queries (when SkipFirstBlock
+/// is true). In this special case, it ignores the contents of the specified
+/// block and starts returning dependence info for its predecessors.
+///
+/// This function returns false on success, or true to indicate that it could
+/// not compute dependence information for some reason. This should be treated
+/// as a clobber dependence on the first instruction in the predecessor block.
+bool MemoryDependenceAnalysis::
+getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, uint64_t PointeeSize,
+ bool isLoad, BasicBlock *StartBB,
+ SmallVectorImpl<NonLocalDepResult> &Result,
+ DenseMap<BasicBlock*, Value*> &Visited,
+ bool SkipFirstBlock) {
+
+ // Look up the cached info for Pointer.
+ ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad);
+
+ std::pair<BBSkipFirstBlockPair, NonLocalDepInfo> *CacheInfo =
+ &NonLocalPointerDeps[CacheKey];
+ NonLocalDepInfo *Cache = &CacheInfo->second;
+
+ // If we have valid cached information for exactly the block we are
+ // investigating, just return it with no recomputation.
+ if (CacheInfo->first == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
+ // We have a fully cached result for this query then we can just return the
+ // cached results and populate the visited set. However, we have to verify
+ // that we don't already have conflicting results for these blocks. Check
+ // to ensure that if a block in the results set is in the visited set that
+ // it was for the same pointer query.
+ if (!Visited.empty()) {
+ for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
+ I != E; ++I) {
+ DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB());
+ if (VI == Visited.end() || VI->second == Pointer.getAddr())
+ continue;
+
+ // We have a pointer mismatch in a block. Just return clobber, saying
+ // that something was clobbered in this result. We could also do a
+ // non-fully cached query, but there is little point in doing this.
+ return true;
+ }
+ }
+
+ Value *Addr = Pointer.getAddr();
+ for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
+ I != E; ++I) {
+ Visited.insert(std::make_pair(I->getBB(), Addr));
+ if (!I->getResult().isNonLocal())
+ Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr));
+ }
+ ++NumCacheCompleteNonLocalPtr;
+ return false;
+ }
+
+ // Otherwise, either this is a new block, a block with an invalid cache
+ // pointer or one that we're about to invalidate by putting more info into it
+ // than its valid cache info. If empty, the result will be valid cache info,
+ // otherwise it isn't.
+ if (Cache->empty())
+ CacheInfo->first = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
+ else
+ CacheInfo->first = BBSkipFirstBlockPair();
+
+ SmallVector<BasicBlock*, 32> Worklist;
+ Worklist.push_back(StartBB);
+
+ // Keep track of the entries that we know are sorted. Previously cached
+ // entries will all be sorted. The entries we add we only sort on demand (we
+ // don't insert every element into its sorted position). We know that we
+ // won't get any reuse from currently inserted values, because we don't
+ // revisit blocks after we insert info for them.
+ unsigned NumSortedEntries = Cache->size();
+ DEBUG(AssertSorted(*Cache));
+
+ while (!Worklist.empty()) {
+ BasicBlock *BB = Worklist.pop_back_val();
+
+ // Skip the first block if we have it.
+ if (!SkipFirstBlock) {
+ // Analyze the dependency of *Pointer in FromBB. See if we already have
+ // been here.
+ assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
+
+ // Get the dependency info for Pointer in BB. If we have cached
+ // information, we will use it, otherwise we compute it.
+ DEBUG(AssertSorted(*Cache, NumSortedEntries));
+ MemDepResult Dep = GetNonLocalInfoForBlock(Pointer.getAddr(), PointeeSize,
+ isLoad, BB, Cache,
+ NumSortedEntries);
+
+ // If we got a Def or Clobber, add this to the list of results.
+ if (!Dep.isNonLocal()) {
+ Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr()));
+ continue;
+ }
+ }
+
+ // If 'Pointer' is an instruction defined in this block, then we need to do
+ // phi translation to change it into a value live in the predecessor block.
+ // If not, we just add the predecessors to the worklist and scan them with
+ // the same Pointer.
+ if (!Pointer.NeedsPHITranslationFromBlock(BB)) {
+ SkipFirstBlock = false;
+ for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
+ // Verify that we haven't looked at this block yet.
+ std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
+ InsertRes = Visited.insert(std::make_pair(*PI, Pointer.getAddr()));
+ if (InsertRes.second) {
+ // First time we've looked at *PI.
+ Worklist.push_back(*PI);
+ continue;
+ }
+
+ // If we have seen this block before, but it was with a different
+ // pointer then we have a phi translation failure and we have to treat
+ // this as a clobber.
+ if (InsertRes.first->second != Pointer.getAddr())
+ goto PredTranslationFailure;
+ }
+ continue;
+ }
+
+ // We do need to do phi translation, if we know ahead of time we can't phi
+ // translate this value, don't even try.
+ if (!Pointer.IsPotentiallyPHITranslatable())
+ goto PredTranslationFailure;
+
+ // We may have added values to the cache list before this PHI translation.
+ // If so, we haven't done anything to ensure that the cache remains sorted.
+ // Sort it now (if needed) so that recursive invocations of
+ // getNonLocalPointerDepFromBB and other routines that could reuse the cache
+ // value will only see properly sorted cache arrays.
+ if (Cache && NumSortedEntries != Cache->size()) {
+ SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
+ NumSortedEntries = Cache->size();
+ }
+ Cache = 0;
+
+ for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
+ BasicBlock *Pred = *PI;
+
+ // Get the PHI translated pointer in this predecessor. This can fail if
+ // not translatable, in which case the getAddr() returns null.
+ PHITransAddr PredPointer(Pointer);
+ PredPointer.PHITranslateValue(BB, Pred);
+
+ Value *PredPtrVal = PredPointer.getAddr();
+
+ // Check to see if we have already visited this pred block with another
+ // pointer. If so, we can't do this lookup. This failure can occur
+ // with PHI translation when a critical edge exists and the PHI node in
+ // the successor translates to a pointer value different than the
+ // pointer the block was first analyzed with.
+ std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
+ InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal));
+
+ if (!InsertRes.second) {
+ // If the predecessor was visited with PredPtr, then we already did
+ // the analysis and can ignore it.
+ if (InsertRes.first->second == PredPtrVal)
+ continue;
+
+ // Otherwise, the block was previously analyzed with a different
+ // pointer. We can't represent the result of this case, so we just
+ // treat this as a phi translation failure.
+ goto PredTranslationFailure;
+ }
+
+ // If PHI translation was unable to find an available pointer in this
+ // predecessor, then we have to assume that the pointer is clobbered in
+ // that predecessor. We can still do PRE of the load, which would insert
+ // a computation of the pointer in this predecessor.
+ if (PredPtrVal == 0) {
+ // Add the entry to the Result list.
+ NonLocalDepResult Entry(Pred,
+ MemDepResult::getClobber(Pred->getTerminator()),
+ PredPtrVal);
+ Result.push_back(Entry);
+
+ // Since we had a phi translation failure, the cache for CacheKey won't
+ // include all of the entries that we need to immediately satisfy future
+ // queries. Mark this in NonLocalPointerDeps by setting the
+ // BBSkipFirstBlockPair pointer to null. This requires reuse of the
+ // cached value to do more work but not miss the phi trans failure.
+ NonLocalPointerDeps[CacheKey].first = BBSkipFirstBlockPair();
+ continue;
+ }
+
+ // FIXME: it is entirely possible that PHI translating will end up with
+ // the same value. Consider PHI translating something like:
+ // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
+ // to recurse here, pedantically speaking.
+
+ // If we have a problem phi translating, fall through to the code below
+ // to handle the failure condition.
+ if (getNonLocalPointerDepFromBB(PredPointer, PointeeSize, isLoad, Pred,
+ Result, Visited))
+ goto PredTranslationFailure;
+ }
+
+ // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
+ CacheInfo = &NonLocalPointerDeps[CacheKey];
+ Cache = &CacheInfo->second;
+ NumSortedEntries = Cache->size();
+
+ // Since we did phi translation, the "Cache" set won't contain all of the
+ // results for the query. This is ok (we can still use it to accelerate
+ // specific block queries) but we can't do the fastpath "return all
+ // results from the set" Clear out the indicator for this.
+ CacheInfo->first = BBSkipFirstBlockPair();
+ SkipFirstBlock = false;
+ continue;
+
+ PredTranslationFailure:
+
+ if (Cache == 0) {
+ // Refresh the CacheInfo/Cache pointer if it got invalidated.
+ CacheInfo = &NonLocalPointerDeps[CacheKey];
+ Cache = &CacheInfo->second;
+ NumSortedEntries = Cache->size();
+ }
+
+ // Since we failed phi translation, the "Cache" set won't contain all of the
+ // results for the query. This is ok (we can still use it to accelerate
+ // specific block queries) but we can't do the fastpath "return all
+ // results from the set". Clear out the indicator for this.
+ CacheInfo->first = BBSkipFirstBlockPair();
+
+ // If *nothing* works, mark the pointer as being clobbered by the first
+ // instruction in this block.
+ //
+ // If this is the magic first block, return this as a clobber of the whole
+ // incoming value. Since we can't phi translate to one of the predecessors,
+ // we have to bail out.
+ if (SkipFirstBlock)
+ return true;
+
+ for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) {
+ assert(I != Cache->rend() && "Didn't find current block??");
+ if (I->getBB() != BB)
+ continue;
+
+ assert(I->getResult().isNonLocal() &&
+ "Should only be here with transparent block");
+ I->setResult(MemDepResult::getClobber(BB->begin()));
+ ReverseNonLocalPtrDeps[BB->begin()].insert(CacheKey);
+ Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(),
+ Pointer.getAddr()));
+ break;
+ }
+ }
+
+ // Okay, we're done now. If we added new values to the cache, re-sort it.
+ SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
+ DEBUG(AssertSorted(*Cache));
+ return false;
+}
+
+/// RemoveCachedNonLocalPointerDependencies - If P exists in
+/// CachedNonLocalPointerInfo, remove it.
+void MemoryDependenceAnalysis::
+RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) {
+ CachedNonLocalPointerInfo::iterator It =
+ NonLocalPointerDeps.find(P);
+ if (It == NonLocalPointerDeps.end()) return;
+
+ // Remove all of the entries in the BB->val map. This involves removing
+ // instructions from the reverse map.
+ NonLocalDepInfo &PInfo = It->second.second;
+
+ for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
+ Instruction *Target = PInfo[i].getResult().getInst();
+ if (Target == 0) continue; // Ignore non-local dep results.
+ assert(Target->getParent() == PInfo[i].getBB());
+
+ // Eliminating the dirty entry from 'Cache', so update the reverse info.
+ RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
+ }
+
+ // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
+ NonLocalPointerDeps.erase(It);
+}
+
+
+/// invalidateCachedPointerInfo - This method is used to invalidate cached
+/// information about the specified pointer, because it may be too
+/// conservative in memdep. This is an optional call that can be used when
+/// the client detects an equivalence between the pointer and some other
+/// value and replaces the other value with ptr. This can make Ptr available
+/// in more places that cached info does not necessarily keep.
+void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) {
+ // If Ptr isn't really a pointer, just ignore it.
+ if (!isa<PointerType>(Ptr->getType())) return;
+ // Flush store info for the pointer.
+ RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
+ // Flush load info for the pointer.
+ RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
+}
+
+/// removeInstruction - Remove an instruction from the dependence analysis,
+/// updating the dependence of instructions that previously depended on it.
+/// This method attempts to keep the cache coherent using the reverse map.
+void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
+ // Walk through the Non-local dependencies, removing this one as the value
+ // for any cached queries.
+ NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
+ if (NLDI != NonLocalDeps.end()) {
+ NonLocalDepInfo &BlockMap = NLDI->second.first;
+ for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end();
+ DI != DE; ++DI)
+ if (Instruction *Inst = DI->getResult().getInst())
+ RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
+ NonLocalDeps.erase(NLDI);
+ }
+
+ // If we have a cached local dependence query for this instruction, remove it.
+ //
+ LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
+ if (LocalDepEntry != LocalDeps.end()) {
+ // Remove us from DepInst's reverse set now that the local dep info is gone.
+ if (Instruction *Inst = LocalDepEntry->second.getInst())
+ RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
+
+ // Remove this local dependency info.
+ LocalDeps.erase(LocalDepEntry);
+ }
+
+ // If we have any cached pointer dependencies on this instruction, remove
+ // them. If the instruction has non-pointer type, then it can't be a pointer
+ // base.
+
+ // Remove it from both the load info and the store info. The instruction
+ // can't be in either of these maps if it is non-pointer.
+ if (isa<PointerType>(RemInst->getType())) {
+ RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
+ RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
+ }
+
+ // Loop over all of the things that depend on the instruction we're removing.
+ //
+ SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
+
+ // If we find RemInst as a clobber or Def in any of the maps for other values,
+ // we need to replace its entry with a dirty version of the instruction after
+ // it. If RemInst is a terminator, we use a null dirty value.
+ //
+ // Using a dirty version of the instruction after RemInst saves having to scan
+ // the entire block to get to this point.
+ MemDepResult NewDirtyVal;
+ if (!RemInst->isTerminator())
+ NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst));
+
+ ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
+ if (ReverseDepIt != ReverseLocalDeps.end()) {
+ SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second;
+ // RemInst can't be the terminator if it has local stuff depending on it.
+ assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) &&
+ "Nothing can locally depend on a terminator");
+
+ for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(),
+ E = ReverseDeps.end(); I != E; ++I) {
+ Instruction *InstDependingOnRemInst = *I;
+ assert(InstDependingOnRemInst != RemInst &&
+ "Already removed our local dep info");
+
+ LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
+
+ // Make sure to remember that new things depend on NewDepInst.
+ assert(NewDirtyVal.getInst() && "There is no way something else can have "
+ "a local dep on this if it is a terminator!");
+ ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(),
+ InstDependingOnRemInst));
+ }
+
+ ReverseLocalDeps.erase(ReverseDepIt);
+
+ // Add new reverse deps after scanning the set, to avoid invalidating the
+ // 'ReverseDeps' reference.
+ while (!ReverseDepsToAdd.empty()) {
+ ReverseLocalDeps[ReverseDepsToAdd.back().first]
+ .insert(ReverseDepsToAdd.back().second);
+ ReverseDepsToAdd.pop_back();
+ }
+ }
+
+ ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
+ if (ReverseDepIt != ReverseNonLocalDeps.end()) {
+ SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second;
+ for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end();
+ I != E; ++I) {
+ assert(*I != RemInst && "Already removed NonLocalDep info for RemInst");
+
+ PerInstNLInfo &INLD = NonLocalDeps[*I];
+ // The information is now dirty!
+ INLD.second = true;
+
+ for (NonLocalDepInfo::iterator DI = INLD.first.begin(),
+ DE = INLD.first.end(); DI != DE; ++DI) {
+ if (DI->getResult().getInst() != RemInst) continue;
+
+ // Convert to a dirty entry for the subsequent instruction.
+ DI->setResult(NewDirtyVal);
+
+ if (Instruction *NextI = NewDirtyVal.getInst())
+ ReverseDepsToAdd.push_back(std::make_pair(NextI, *I));
+ }
+ }
+
+ ReverseNonLocalDeps.erase(ReverseDepIt);
+
+ // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
+ while (!ReverseDepsToAdd.empty()) {
+ ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
+ .insert(ReverseDepsToAdd.back().second);
+ ReverseDepsToAdd.pop_back();
+ }
+ }
+
+ // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
+ // value in the NonLocalPointerDeps info.
+ ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
+ ReverseNonLocalPtrDeps.find(RemInst);
+ if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
+ SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second;
+ SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd;
+
+ for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(),
+ E = Set.end(); I != E; ++I) {
+ ValueIsLoadPair P = *I;
+ assert(P.getPointer() != RemInst &&
+ "Already removed NonLocalPointerDeps info for RemInst");
+
+ NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].second;
+
+ // The cache is not valid for any specific block anymore.
+ NonLocalPointerDeps[P].first = BBSkipFirstBlockPair();
+
+ // Update any entries for RemInst to use the instruction after it.
+ for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end();
+ DI != DE; ++DI) {
+ if (DI->getResult().getInst() != RemInst) continue;
+
+ // Convert to a dirty entry for the subsequent instruction.
+ DI->setResult(NewDirtyVal);
+
+ if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
+ ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
+ }
+
+ // Re-sort the NonLocalDepInfo. Changing the dirty entry to its
+ // subsequent value may invalidate the sortedness.
+ std::sort(NLPDI.begin(), NLPDI.end());
+ }
+
+ ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
+
+ while (!ReversePtrDepsToAdd.empty()) {
+ ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first]
+ .insert(ReversePtrDepsToAdd.back().second);
+ ReversePtrDepsToAdd.pop_back();
+ }
+ }
+
+
+ assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
+ AA->deleteValue(RemInst);
+ DEBUG(verifyRemoved(RemInst));
+}
+/// verifyRemoved - Verify that the specified instruction does not occur
+/// in our internal data structures.
+void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
+ for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
+ E = LocalDeps.end(); I != E; ++I) {
+ assert(I->first != D && "Inst occurs in data structures");
+ assert(I->second.getInst() != D &&
+ "Inst occurs in data structures");
+ }
+
+ for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(),
+ E = NonLocalPointerDeps.end(); I != E; ++I) {
+ assert(I->first.getPointer() != D && "Inst occurs in NLPD map key");
+ const NonLocalDepInfo &Val = I->second.second;
+ for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end();
+ II != E; ++II)
+ assert(II->getResult().getInst() != D && "Inst occurs as NLPD value");
+ }
+
+ for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
+ E = NonLocalDeps.end(); I != E; ++I) {
+ assert(I->first != D && "Inst occurs in data structures");
+ const PerInstNLInfo &INLD = I->second;
+ for (NonLocalDepInfo::const_iterator II = INLD.first.begin(),
+ EE = INLD.first.end(); II != EE; ++II)
+ assert(II->getResult().getInst() != D && "Inst occurs in data structures");
+ }
+
+ for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
+ E = ReverseLocalDeps.end(); I != E; ++I) {
+ assert(I->first != D && "Inst occurs in data structures");
+ for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
+ EE = I->second.end(); II != EE; ++II)
+ assert(*II != D && "Inst occurs in data structures");
+ }
+
+ for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
+ E = ReverseNonLocalDeps.end();
+ I != E; ++I) {
+ assert(I->first != D && "Inst occurs in data structures");
+ for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
+ EE = I->second.end(); II != EE; ++II)
+ assert(*II != D && "Inst occurs in data structures");
+ }
+
+ for (ReverseNonLocalPtrDepTy::const_iterator
+ I = ReverseNonLocalPtrDeps.begin(),
+ E = ReverseNonLocalPtrDeps.end(); I != E; ++I) {
+ assert(I->first != D && "Inst occurs in rev NLPD map");
+
+ for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(),
+ E = I->second.end(); II != E; ++II)
+ assert(*II != ValueIsLoadPair(D, false) &&
+ *II != ValueIsLoadPair(D, true) &&
+ "Inst occurs in ReverseNonLocalPtrDeps map");
+ }
+
+}
diff --git a/lib/Analysis/PHITransAddr.cpp b/lib/Analysis/PHITransAddr.cpp
new file mode 100644
index 0000000..334a188
--- /dev/null
+++ b/lib/Analysis/PHITransAddr.cpp
@@ -0,0 +1,433 @@
+//===- PHITransAddr.cpp - PHI Translation for Addresses -------------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements the PHITransAddr class.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/PHITransAddr.h"
+#include "llvm/Analysis/Dominators.h"
+#include "llvm/Analysis/InstructionSimplify.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+using namespace llvm;
+
+static bool CanPHITrans(Instruction *Inst) {
+ if (isa<PHINode>(Inst) ||
+ isa<BitCastInst>(Inst) ||
+ isa<GetElementPtrInst>(Inst))
+ return true;
+
+ if (Inst->getOpcode() == Instruction::Add &&
+ isa<ConstantInt>(Inst->getOperand(1)))
+ return true;
+
+ // cerr << "MEMDEP: Could not PHI translate: " << *Pointer;
+ // if (isa<BitCastInst>(PtrInst) || isa<GetElementPtrInst>(PtrInst))
+ // cerr << "OP:\t\t\t\t" << *PtrInst->getOperand(0);
+ return false;
+}
+
+void PHITransAddr::dump() const {
+ if (Addr == 0) {
+ dbgs() << "PHITransAddr: null\n";
+ return;
+ }
+ dbgs() << "PHITransAddr: " << *Addr << "\n";
+ for (unsigned i = 0, e = InstInputs.size(); i != e; ++i)
+ dbgs() << " Input #" << i << " is " << *InstInputs[i] << "\n";
+}
+
+
+static bool VerifySubExpr(Value *Expr,
+ SmallVectorImpl<Instruction*> &InstInputs) {
+ // If this is a non-instruction value, there is nothing to do.
+ Instruction *I = dyn_cast<Instruction>(Expr);
+ if (I == 0) return true;
+
+ // If it's an instruction, it is either in Tmp or its operands recursively
+ // are.
+ SmallVectorImpl<Instruction*>::iterator Entry =
+ std::find(InstInputs.begin(), InstInputs.end(), I);
+ if (Entry != InstInputs.end()) {
+ InstInputs.erase(Entry);
+ return true;
+ }
+
+ // If it isn't in the InstInputs list it is a subexpr incorporated into the
+ // address. Sanity check that it is phi translatable.
+ if (!CanPHITrans(I)) {
+ errs() << "Non phi translatable instruction found in PHITransAddr, either "
+ "something is missing from InstInputs or CanPHITrans is wrong:\n";
+ errs() << *I << '\n';
+ return false;
+ }
+
+ // Validate the operands of the instruction.
+ for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
+ if (!VerifySubExpr(I->getOperand(i), InstInputs))
+ return false;
+
+ return true;
+}
+
+/// Verify - Check internal consistency of this data structure. If the
+/// structure is valid, it returns true. If invalid, it prints errors and
+/// returns false.
+bool PHITransAddr::Verify() const {
+ if (Addr == 0) return true;
+
+ SmallVector<Instruction*, 8> Tmp(InstInputs.begin(), InstInputs.end());
+
+ if (!VerifySubExpr(Addr, Tmp))
+ return false;
+
+ if (!Tmp.empty()) {
+ errs() << "PHITransAddr inconsistent, contains extra instructions:\n";
+ for (unsigned i = 0, e = InstInputs.size(); i != e; ++i)
+ errs() << " InstInput #" << i << " is " << *InstInputs[i] << "\n";
+ return false;
+ }
+
+ // a-ok.
+ return true;
+}
+
+
+/// IsPotentiallyPHITranslatable - If this needs PHI translation, return true
+/// if we have some hope of doing it. This should be used as a filter to
+/// avoid calling PHITranslateValue in hopeless situations.
+bool PHITransAddr::IsPotentiallyPHITranslatable() const {
+ // If the input value is not an instruction, or if it is not defined in CurBB,
+ // then we don't need to phi translate it.
+ Instruction *Inst = dyn_cast<Instruction>(Addr);
+ return Inst == 0 || CanPHITrans(Inst);
+}
+
+
+static void RemoveInstInputs(Value *V,
+ SmallVectorImpl<Instruction*> &InstInputs) {
+ Instruction *I = dyn_cast<Instruction>(V);
+ if (I == 0) return;
+
+ // If the instruction is in the InstInputs list, remove it.
+ SmallVectorImpl<Instruction*>::iterator Entry =
+ std::find(InstInputs.begin(), InstInputs.end(), I);
+ if (Entry != InstInputs.end()) {
+ InstInputs.erase(Entry);
+ return;
+ }
+
+ assert(!isa<PHINode>(I) && "Error, removing something that isn't an input");
+
+ // Otherwise, it must have instruction inputs itself. Zap them recursively.
+ for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
+ if (Instruction *Op = dyn_cast<Instruction>(I->getOperand(i)))
+ RemoveInstInputs(Op, InstInputs);
+ }
+}
+
+Value *PHITransAddr::PHITranslateSubExpr(Value *V, BasicBlock *CurBB,
+ BasicBlock *PredBB) {
+ // If this is a non-instruction value, it can't require PHI translation.
+ Instruction *Inst = dyn_cast<Instruction>(V);
+ if (Inst == 0) return V;
+
+ // Determine whether 'Inst' is an input to our PHI translatable expression.
+ bool isInput = std::count(InstInputs.begin(), InstInputs.end(), Inst);
+
+ // Handle inputs instructions if needed.
+ if (isInput) {
+ if (Inst->getParent() != CurBB) {
+ // If it is an input defined in a different block, then it remains an
+ // input.
+ return Inst;
+ }
+
+ // If 'Inst' is defined in this block and is an input that needs to be phi
+ // translated, we need to incorporate the value into the expression or fail.
+
+ // In either case, the instruction itself isn't an input any longer.
+ InstInputs.erase(std::find(InstInputs.begin(), InstInputs.end(), Inst));
+
+ // If this is a PHI, go ahead and translate it.
+ if (PHINode *PN = dyn_cast<PHINode>(Inst))
+ return AddAsInput(PN->getIncomingValueForBlock(PredBB));
+
+ // If this is a non-phi value, and it is analyzable, we can incorporate it
+ // into the expression by making all instruction operands be inputs.
+ if (!CanPHITrans(Inst))
+ return 0;
+
+ // All instruction operands are now inputs (and of course, they may also be
+ // defined in this block, so they may need to be phi translated themselves.
+ for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
+ if (Instruction *Op = dyn_cast<Instruction>(Inst->getOperand(i)))
+ InstInputs.push_back(Op);
+ }
+
+ // Ok, it must be an intermediate result (either because it started that way
+ // or because we just incorporated it into the expression). See if its
+ // operands need to be phi translated, and if so, reconstruct it.
+
+ if (BitCastInst *BC = dyn_cast<BitCastInst>(Inst)) {
+ Value *PHIIn = PHITranslateSubExpr(BC->getOperand(0), CurBB, PredBB);
+ if (PHIIn == 0) return 0;
+ if (PHIIn == BC->getOperand(0))
+ return BC;
+
+ // Find an available version of this cast.
+
+ // Constants are trivial to find.
+ if (Constant *C = dyn_cast<Constant>(PHIIn))
+ return AddAsInput(ConstantExpr::getBitCast(C, BC->getType()));
+
+ // Otherwise we have to see if a bitcasted version of the incoming pointer
+ // is available. If so, we can use it, otherwise we have to fail.
+ for (Value::use_iterator UI = PHIIn->use_begin(), E = PHIIn->use_end();
+ UI != E; ++UI) {
+ if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI))
+ if (BCI->getType() == BC->getType())
+ return BCI;
+ }
+ return 0;
+ }
+
+ // Handle getelementptr with at least one PHI translatable operand.
+ if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Inst)) {
+ SmallVector<Value*, 8> GEPOps;
+ bool AnyChanged = false;
+ for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i) {
+ Value *GEPOp = PHITranslateSubExpr(GEP->getOperand(i), CurBB, PredBB);
+ if (GEPOp == 0) return 0;
+
+ AnyChanged |= GEPOp != GEP->getOperand(i);
+ GEPOps.push_back(GEPOp);
+ }
+
+ if (!AnyChanged)
+ return GEP;
+
+ // Simplify the GEP to handle 'gep x, 0' -> x etc.
+ if (Value *V = SimplifyGEPInst(&GEPOps[0], GEPOps.size(), TD)) {
+ for (unsigned i = 0, e = GEPOps.size(); i != e; ++i)
+ RemoveInstInputs(GEPOps[i], InstInputs);
+
+ return AddAsInput(V);
+ }
+
+ // Scan to see if we have this GEP available.
+ Value *APHIOp = GEPOps[0];
+ for (Value::use_iterator UI = APHIOp->use_begin(), E = APHIOp->use_end();
+ UI != E; ++UI) {
+ if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI))
+ if (GEPI->getType() == GEP->getType() &&
+ GEPI->getNumOperands() == GEPOps.size() &&
+ GEPI->getParent()->getParent() == CurBB->getParent()) {
+ bool Mismatch = false;
+ for (unsigned i = 0, e = GEPOps.size(); i != e; ++i)
+ if (GEPI->getOperand(i) != GEPOps[i]) {
+ Mismatch = true;
+ break;
+ }
+ if (!Mismatch)
+ return GEPI;
+ }
+ }
+ return 0;
+ }
+
+ // Handle add with a constant RHS.
+ if (Inst->getOpcode() == Instruction::Add &&
+ isa<ConstantInt>(Inst->getOperand(1))) {
+ // PHI translate the LHS.
+ Constant *RHS = cast<ConstantInt>(Inst->getOperand(1));
+ bool isNSW = cast<BinaryOperator>(Inst)->hasNoSignedWrap();
+ bool isNUW = cast<BinaryOperator>(Inst)->hasNoUnsignedWrap();
+
+ Value *LHS = PHITranslateSubExpr(Inst->getOperand(0), CurBB, PredBB);
+ if (LHS == 0) return 0;
+
+ // If the PHI translated LHS is an add of a constant, fold the immediates.
+ if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(LHS))
+ if (BOp->getOpcode() == Instruction::Add)
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(BOp->getOperand(1))) {
+ LHS = BOp->getOperand(0);
+ RHS = ConstantExpr::getAdd(RHS, CI);
+ isNSW = isNUW = false;
+
+ // If the old 'LHS' was an input, add the new 'LHS' as an input.
+ if (std::count(InstInputs.begin(), InstInputs.end(), BOp)) {
+ RemoveInstInputs(BOp, InstInputs);
+ AddAsInput(LHS);
+ }
+ }
+
+ // See if the add simplifies away.
+ if (Value *Res = SimplifyAddInst(LHS, RHS, isNSW, isNUW, TD)) {
+ // If we simplified the operands, the LHS is no longer an input, but Res
+ // is.
+ RemoveInstInputs(LHS, InstInputs);
+ return AddAsInput(Res);
+ }
+
+ // If we didn't modify the add, just return it.
+ if (LHS == Inst->getOperand(0) && RHS == Inst->getOperand(1))
+ return Inst;
+
+ // Otherwise, see if we have this add available somewhere.
+ for (Value::use_iterator UI = LHS->use_begin(), E = LHS->use_end();
+ UI != E; ++UI) {
+ if (BinaryOperator *BO = dyn_cast<BinaryOperator>(*UI))
+ if (BO->getOpcode() == Instruction::Add &&
+ BO->getOperand(0) == LHS && BO->getOperand(1) == RHS &&
+ BO->getParent()->getParent() == CurBB->getParent())
+ return BO;
+ }
+
+ return 0;
+ }
+
+ // Otherwise, we failed.
+ return 0;
+}
+
+
+/// PHITranslateValue - PHI translate the current address up the CFG from
+/// CurBB to Pred, updating our state the reflect any needed changes. This
+/// returns true on failure and sets Addr to null.
+bool PHITransAddr::PHITranslateValue(BasicBlock *CurBB, BasicBlock *PredBB) {
+ assert(Verify() && "Invalid PHITransAddr!");
+ Addr = PHITranslateSubExpr(Addr, CurBB, PredBB);
+ assert(Verify() && "Invalid PHITransAddr!");
+ return Addr == 0;
+}
+
+/// GetAvailablePHITranslatedSubExpr - Return the value computed by
+/// PHITranslateSubExpr if it dominates PredBB, otherwise return null.
+Value *PHITransAddr::
+GetAvailablePHITranslatedSubExpr(Value *V, BasicBlock *CurBB,BasicBlock *PredBB,
+ const DominatorTree &DT) const {
+ PHITransAddr Tmp(V, TD);
+ Tmp.PHITranslateValue(CurBB, PredBB);
+
+ // See if PHI translation succeeds.
+ V = Tmp.getAddr();
+
+ // Make sure the value is live in the predecessor.
+ if (Instruction *Inst = dyn_cast_or_null<Instruction>(V))
+ if (!DT.dominates(Inst->getParent(), PredBB))
+ return 0;
+ return V;
+}
+
+
+/// PHITranslateWithInsertion - PHI translate this value into the specified
+/// predecessor block, inserting a computation of the value if it is
+/// unavailable.
+///
+/// All newly created instructions are added to the NewInsts list. This
+/// returns null on failure.
+///
+Value *PHITransAddr::
+PHITranslateWithInsertion(BasicBlock *CurBB, BasicBlock *PredBB,
+ const DominatorTree &DT,
+ SmallVectorImpl<Instruction*> &NewInsts) {
+ unsigned NISize = NewInsts.size();
+
+ // Attempt to PHI translate with insertion.
+ Addr = InsertPHITranslatedSubExpr(Addr, CurBB, PredBB, DT, NewInsts);
+
+ // If successful, return the new value.
+ if (Addr) return Addr;
+
+ // If not, destroy any intermediate instructions inserted.
+ while (NewInsts.size() != NISize)
+ NewInsts.pop_back_val()->eraseFromParent();
+ return 0;
+}
+
+
+/// InsertPHITranslatedPointer - Insert a computation of the PHI translated
+/// version of 'V' for the edge PredBB->CurBB into the end of the PredBB
+/// block. All newly created instructions are added to the NewInsts list.
+/// This returns null on failure.
+///
+Value *PHITransAddr::
+InsertPHITranslatedSubExpr(Value *InVal, BasicBlock *CurBB,
+ BasicBlock *PredBB, const DominatorTree &DT,
+ SmallVectorImpl<Instruction*> &NewInsts) {
+ // See if we have a version of this value already available and dominating
+ // PredBB. If so, there is no need to insert a new instance of it.
+ if (Value *Res = GetAvailablePHITranslatedSubExpr(InVal, CurBB, PredBB, DT))
+ return Res;
+
+ // If we don't have an available version of this value, it must be an
+ // instruction.
+ Instruction *Inst = cast<Instruction>(InVal);
+
+ // Handle bitcast of PHI translatable value.
+ if (BitCastInst *BC = dyn_cast<BitCastInst>(Inst)) {
+ Value *OpVal = InsertPHITranslatedSubExpr(BC->getOperand(0),
+ CurBB, PredBB, DT, NewInsts);
+ if (OpVal == 0) return 0;
+
+ // Otherwise insert a bitcast at the end of PredBB.
+ BitCastInst *New = new BitCastInst(OpVal, InVal->getType(),
+ InVal->getName()+".phi.trans.insert",
+ PredBB->getTerminator());
+ NewInsts.push_back(New);
+ return New;
+ }
+
+ // Handle getelementptr with at least one PHI operand.
+ if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Inst)) {
+ SmallVector<Value*, 8> GEPOps;
+ BasicBlock *CurBB = GEP->getParent();
+ for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i) {
+ Value *OpVal = InsertPHITranslatedSubExpr(GEP->getOperand(i),
+ CurBB, PredBB, DT, NewInsts);
+ if (OpVal == 0) return 0;
+ GEPOps.push_back(OpVal);
+ }
+
+ GetElementPtrInst *Result =
+ GetElementPtrInst::Create(GEPOps[0], GEPOps.begin()+1, GEPOps.end(),
+ InVal->getName()+".phi.trans.insert",
+ PredBB->getTerminator());
+ Result->setIsInBounds(GEP->isInBounds());
+ NewInsts.push_back(Result);
+ return Result;
+ }
+
+#if 0
+ // FIXME: This code works, but it is unclear that we actually want to insert
+ // a big chain of computation in order to make a value available in a block.
+ // This needs to be evaluated carefully to consider its cost trade offs.
+
+ // Handle add with a constant RHS.
+ if (Inst->getOpcode() == Instruction::Add &&
+ isa<ConstantInt>(Inst->getOperand(1))) {
+ // PHI translate the LHS.
+ Value *OpVal = InsertPHITranslatedSubExpr(Inst->getOperand(0),
+ CurBB, PredBB, DT, NewInsts);
+ if (OpVal == 0) return 0;
+
+ BinaryOperator *Res = BinaryOperator::CreateAdd(OpVal, Inst->getOperand(1),
+ InVal->getName()+".phi.trans.insert",
+ PredBB->getTerminator());
+ Res->setHasNoSignedWrap(cast<BinaryOperator>(Inst)->hasNoSignedWrap());
+ Res->setHasNoUnsignedWrap(cast<BinaryOperator>(Inst)->hasNoUnsignedWrap());
+ NewInsts.push_back(Res);
+ return Res;
+ }
+#endif
+
+ return 0;
+}
diff --git a/lib/Analysis/PointerTracking.cpp b/lib/Analysis/PointerTracking.cpp
new file mode 100644
index 0000000..8da07e7
--- /dev/null
+++ b/lib/Analysis/PointerTracking.cpp
@@ -0,0 +1,267 @@
+//===- PointerTracking.cpp - Pointer Bounds Tracking ------------*- C++ -*-===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements tracking of pointer bounds.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/ConstantFolding.h"
+#include "llvm/Analysis/Dominators.h"
+#include "llvm/Analysis/LoopInfo.h"
+#include "llvm/Analysis/MemoryBuiltins.h"
+#include "llvm/Analysis/PointerTracking.h"
+#include "llvm/Analysis/ScalarEvolution.h"
+#include "llvm/Analysis/ScalarEvolutionExpressions.h"
+#include "llvm/Constants.h"
+#include "llvm/Module.h"
+#include "llvm/Value.h"
+#include "llvm/Support/CallSite.h"
+#include "llvm/Support/InstIterator.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/Target/TargetData.h"
+using namespace llvm;
+
+char PointerTracking::ID = 0;
+PointerTracking::PointerTracking() : FunctionPass(&ID) {}
+
+bool PointerTracking::runOnFunction(Function &F) {
+ predCache.clear();
+ assert(analyzing.empty());
+ FF = &F;
+ TD = getAnalysisIfAvailable<TargetData>();
+ SE = &getAnalysis<ScalarEvolution>();
+ LI = &getAnalysis<LoopInfo>();
+ DT = &getAnalysis<DominatorTree>();
+ return false;
+}
+
+void PointerTracking::getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.addRequiredTransitive<DominatorTree>();
+ AU.addRequiredTransitive<LoopInfo>();
+ AU.addRequiredTransitive<ScalarEvolution>();
+ AU.setPreservesAll();
+}
+
+bool PointerTracking::doInitialization(Module &M) {
+ const Type *PTy = Type::getInt8PtrTy(M.getContext());
+
+ // Find calloc(i64, i64) or calloc(i32, i32).
+ callocFunc = M.getFunction("calloc");
+ if (callocFunc) {
+ const FunctionType *Ty = callocFunc->getFunctionType();
+
+ std::vector<const Type*> args, args2;
+ args.push_back(Type::getInt64Ty(M.getContext()));
+ args.push_back(Type::getInt64Ty(M.getContext()));
+ args2.push_back(Type::getInt32Ty(M.getContext()));
+ args2.push_back(Type::getInt32Ty(M.getContext()));
+ const FunctionType *Calloc1Type =
+ FunctionType::get(PTy, args, false);
+ const FunctionType *Calloc2Type =
+ FunctionType::get(PTy, args2, false);
+ if (Ty != Calloc1Type && Ty != Calloc2Type)
+ callocFunc = 0; // Give up
+ }
+
+ // Find realloc(i8*, i64) or realloc(i8*, i32).
+ reallocFunc = M.getFunction("realloc");
+ if (reallocFunc) {
+ const FunctionType *Ty = reallocFunc->getFunctionType();
+ std::vector<const Type*> args, args2;
+ args.push_back(PTy);
+ args.push_back(Type::getInt64Ty(M.getContext()));
+ args2.push_back(PTy);
+ args2.push_back(Type::getInt32Ty(M.getContext()));
+
+ const FunctionType *Realloc1Type =
+ FunctionType::get(PTy, args, false);
+ const FunctionType *Realloc2Type =
+ FunctionType::get(PTy, args2, false);
+ if (Ty != Realloc1Type && Ty != Realloc2Type)
+ reallocFunc = 0; // Give up
+ }
+ return false;
+}
+
+// Calculates the number of elements allocated for pointer P,
+// the type of the element is stored in Ty.
+const SCEV *PointerTracking::computeAllocationCount(Value *P,
+ const Type *&Ty) const {
+ Value *V = P->stripPointerCasts();
+ if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
+ Value *arraySize = AI->getArraySize();
+ Ty = AI->getAllocatedType();
+ // arraySize elements of type Ty.
+ return SE->getSCEV(arraySize);
+ }
+
+ if (CallInst *CI = extractMallocCall(V)) {
+ Value *arraySize = getMallocArraySize(CI, TD);
+ const Type* AllocTy = getMallocAllocatedType(CI);
+ if (!AllocTy || !arraySize) return SE->getCouldNotCompute();
+ Ty = AllocTy;
+ // arraySize elements of type Ty.
+ return SE->getSCEV(arraySize);
+ }
+
+ if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
+ if (GV->hasDefinitiveInitializer()) {
+ Constant *C = GV->getInitializer();
+ if (const ArrayType *ATy = dyn_cast<ArrayType>(C->getType())) {
+ Ty = ATy->getElementType();
+ return SE->getConstant(Type::getInt32Ty(P->getContext()),
+ ATy->getNumElements());
+ }
+ }
+ Ty = GV->getType();
+ return SE->getConstant(Type::getInt32Ty(P->getContext()), 1);
+ //TODO: implement more tracking for globals
+ }
+
+ if (CallInst *CI = dyn_cast<CallInst>(V)) {
+ CallSite CS(CI);
+ Function *F = dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
+ const Loop *L = LI->getLoopFor(CI->getParent());
+ if (F == callocFunc) {
+ Ty = Type::getInt8Ty(P->getContext());
+ // calloc allocates arg0*arg1 bytes.
+ return SE->getSCEVAtScope(SE->getMulExpr(SE->getSCEV(CS.getArgument(0)),
+ SE->getSCEV(CS.getArgument(1))),
+ L);
+ } else if (F == reallocFunc) {
+ Ty = Type::getInt8Ty(P->getContext());
+ // realloc allocates arg1 bytes.
+ return SE->getSCEVAtScope(CS.getArgument(1), L);
+ }
+ }
+
+ return SE->getCouldNotCompute();
+}
+
+// Calculates the number of elements of type Ty allocated for P.
+const SCEV *PointerTracking::computeAllocationCountForType(Value *P,
+ const Type *Ty)
+ const {
+ const Type *elementTy;
+ const SCEV *Count = computeAllocationCount(P, elementTy);
+ if (isa<SCEVCouldNotCompute>(Count))
+ return Count;
+ if (elementTy == Ty)
+ return Count;
+
+ if (!TD) // need TargetData from this point forward
+ return SE->getCouldNotCompute();
+
+ uint64_t elementSize = TD->getTypeAllocSize(elementTy);
+ uint64_t wantSize = TD->getTypeAllocSize(Ty);
+ if (elementSize == wantSize)
+ return Count;
+ if (elementSize % wantSize) //fractional counts not possible
+ return SE->getCouldNotCompute();
+ return SE->getMulExpr(Count, SE->getConstant(Count->getType(),
+ elementSize/wantSize));
+}
+
+const SCEV *PointerTracking::getAllocationElementCount(Value *V) const {
+ // We only deal with pointers.
+ const PointerType *PTy = cast<PointerType>(V->getType());
+ return computeAllocationCountForType(V, PTy->getElementType());
+}
+
+const SCEV *PointerTracking::getAllocationSizeInBytes(Value *V) const {
+ return computeAllocationCountForType(V, Type::getInt8Ty(V->getContext()));
+}
+
+// Helper for isLoopGuardedBy that checks the swapped and inverted predicate too
+enum SolverResult PointerTracking::isLoopGuardedBy(const Loop *L,
+ Predicate Pred,
+ const SCEV *A,
+ const SCEV *B) const {
+ if (SE->isLoopGuardedByCond(L, Pred, A, B))
+ return AlwaysTrue;
+ Pred = ICmpInst::getSwappedPredicate(Pred);
+ if (SE->isLoopGuardedByCond(L, Pred, B, A))
+ return AlwaysTrue;
+
+ Pred = ICmpInst::getInversePredicate(Pred);
+ if (SE->isLoopGuardedByCond(L, Pred, B, A))
+ return AlwaysFalse;
+ Pred = ICmpInst::getSwappedPredicate(Pred);
+ if (SE->isLoopGuardedByCond(L, Pred, A, B))
+ return AlwaysTrue;
+ return Unknown;
+}
+
+enum SolverResult PointerTracking::checkLimits(const SCEV *Offset,
+ const SCEV *Limit,
+ BasicBlock *BB)
+{
+ //FIXME: merge implementation
+ return Unknown;
+}
+
+void PointerTracking::getPointerOffset(Value *Pointer, Value *&Base,
+ const SCEV *&Limit,
+ const SCEV *&Offset) const
+{
+ Pointer = Pointer->stripPointerCasts();
+ Base = Pointer->getUnderlyingObject();
+ Limit = getAllocationSizeInBytes(Base);
+ if (isa<SCEVCouldNotCompute>(Limit)) {
+ Base = 0;
+ Offset = Limit;
+ return;
+ }
+
+ Offset = SE->getMinusSCEV(SE->getSCEV(Pointer), SE->getSCEV(Base));
+ if (isa<SCEVCouldNotCompute>(Offset)) {
+ Base = 0;
+ Limit = Offset;
+ }
+}
+
+void PointerTracking::print(raw_ostream &OS, const Module* M) const {
+ // Calling some PT methods may cause caches to be updated, however
+ // this should be safe for the same reason its safe for SCEV.
+ PointerTracking &PT = *const_cast<PointerTracking*>(this);
+ for (inst_iterator I=inst_begin(*FF), E=inst_end(*FF); I != E; ++I) {
+ if (!isa<PointerType>(I->getType()))
+ continue;
+ Value *Base;
+ const SCEV *Limit, *Offset;
+ getPointerOffset(&*I, Base, Limit, Offset);
+ if (!Base)
+ continue;
+
+ if (Base == &*I) {
+ const SCEV *S = getAllocationElementCount(Base);
+ OS << *Base << " ==> " << *S << " elements, ";
+ OS << *Limit << " bytes allocated\n";
+ continue;
+ }
+ OS << &*I << " -- base: " << *Base;
+ OS << " offset: " << *Offset;
+
+ enum SolverResult res = PT.checkLimits(Offset, Limit, I->getParent());
+ switch (res) {
+ case AlwaysTrue:
+ OS << " always safe\n";
+ break;
+ case AlwaysFalse:
+ OS << " always unsafe\n";
+ break;
+ case Unknown:
+ OS << " <<unknown>>\n";
+ break;
+ }
+ }
+}
+
+static RegisterPass<PointerTracking> X("pointertracking",
+ "Track pointer bounds", false, true);
diff --git a/lib/Analysis/PostDominators.cpp b/lib/Analysis/PostDominators.cpp
new file mode 100644
index 0000000..c38e050
--- /dev/null
+++ b/lib/Analysis/PostDominators.cpp
@@ -0,0 +1,98 @@
+//===- PostDominators.cpp - Post-Dominator Calculation --------------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements the post-dominator construction algorithms.
+//
+//===----------------------------------------------------------------------===//
+
+#define DEBUG_TYPE "postdomtree"
+
+#include "llvm/Analysis/PostDominators.h"
+#include "llvm/Instructions.h"
+#include "llvm/Support/CFG.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/ADT/DepthFirstIterator.h"
+#include "llvm/ADT/SetOperations.h"
+#include "llvm/Analysis/DominatorInternals.h"
+using namespace llvm;
+
+//===----------------------------------------------------------------------===//
+// PostDominatorTree Implementation
+//===----------------------------------------------------------------------===//
+
+char PostDominatorTree::ID = 0;
+char PostDominanceFrontier::ID = 0;
+static RegisterPass<PostDominatorTree>
+F("postdomtree", "Post-Dominator Tree Construction", true, true);
+
+bool PostDominatorTree::runOnFunction(Function &F) {
+ DT->recalculate(F);
+ DEBUG(DT->print(dbgs()));
+ return false;
+}
+
+PostDominatorTree::~PostDominatorTree() {
+ delete DT;
+}
+
+void PostDominatorTree::print(raw_ostream &OS, const Module *) const {
+ DT->print(OS);
+}
+
+
+FunctionPass* llvm::createPostDomTree() {
+ return new PostDominatorTree();
+}
+
+//===----------------------------------------------------------------------===//
+// PostDominanceFrontier Implementation
+//===----------------------------------------------------------------------===//
+
+static RegisterPass<PostDominanceFrontier>
+H("postdomfrontier", "Post-Dominance Frontier Construction", true, true);
+
+const DominanceFrontier::DomSetType &
+PostDominanceFrontier::calculate(const PostDominatorTree &DT,
+ const DomTreeNode *Node) {
+ // Loop over CFG successors to calculate DFlocal[Node]
+ BasicBlock *BB = Node->getBlock();
+ DomSetType &S = Frontiers[BB]; // The new set to fill in...
+ if (getRoots().empty()) return S;
+
+ if (BB)
+ for (pred_iterator SI = pred_begin(BB), SE = pred_end(BB);
+ SI != SE; ++SI) {
+ // Does Node immediately dominate this predecessor?
+ DomTreeNode *SINode = DT[*SI];
+ if (SINode && SINode->getIDom() != Node)
+ S.insert(*SI);
+ }
+
+ // At this point, S is DFlocal. Now we union in DFup's of our children...
+ // Loop through and visit the nodes that Node immediately dominates (Node's
+ // children in the IDomTree)
+ //
+ for (DomTreeNode::const_iterator
+ NI = Node->begin(), NE = Node->end(); NI != NE; ++NI) {
+ DomTreeNode *IDominee = *NI;
+ const DomSetType &ChildDF = calculate(DT, IDominee);
+
+ DomSetType::const_iterator CDFI = ChildDF.begin(), CDFE = ChildDF.end();
+ for (; CDFI != CDFE; ++CDFI) {
+ if (!DT.properlyDominates(Node, DT[*CDFI]))
+ S.insert(*CDFI);
+ }
+ }
+
+ return S;
+}
+
+FunctionPass* llvm::createPostDomFrontier() {
+ return new PostDominanceFrontier();
+}
diff --git a/lib/Analysis/ProfileEstimatorPass.cpp b/lib/Analysis/ProfileEstimatorPass.cpp
new file mode 100644
index 0000000..bce6b31
--- /dev/null
+++ b/lib/Analysis/ProfileEstimatorPass.cpp
@@ -0,0 +1,423 @@
+//===- ProfileEstimatorPass.cpp - LLVM Pass to estimate profile info ------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements a concrete implementation of profiling information that
+// estimates the profiling information in a very crude and unimaginative way.
+//
+//===----------------------------------------------------------------------===//
+#define DEBUG_TYPE "profile-estimator"
+#include "llvm/Pass.h"
+#include "llvm/Analysis/Passes.h"
+#include "llvm/Analysis/ProfileInfo.h"
+#include "llvm/Analysis/LoopInfo.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/Support/Format.h"
+using namespace llvm;
+
+static cl::opt<double>
+LoopWeight(
+ "profile-estimator-loop-weight", cl::init(10),
+ cl::value_desc("loop-weight"),
+ cl::desc("Number of loop executions used for profile-estimator")
+);
+
+namespace {
+ class ProfileEstimatorPass : public FunctionPass, public ProfileInfo {
+ double ExecCount;
+ LoopInfo *LI;
+ std::set<BasicBlock*> BBToVisit;
+ std::map<Loop*,double> LoopExitWeights;
+ std::map<Edge,double> MinimalWeight;
+ public:
+ static char ID; // Class identification, replacement for typeinfo
+ explicit ProfileEstimatorPass(const double execcount = 0)
+ : FunctionPass(&ID), ExecCount(execcount) {
+ if (execcount == 0) ExecCount = LoopWeight;
+ }
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.setPreservesAll();
+ AU.addRequired<LoopInfo>();
+ }
+
+ virtual const char *getPassName() const {
+ return "Profiling information estimator";
+ }
+
+ /// run - Estimate the profile information from the specified file.
+ virtual bool runOnFunction(Function &F);
+
+ /// getAdjustedAnalysisPointer - This method is used when a pass implements
+ /// an analysis interface through multiple inheritance. If needed, it
+ /// should override this to adjust the this pointer as needed for the
+ /// specified pass info.
+ virtual void *getAdjustedAnalysisPointer(const PassInfo *PI) {
+ if (PI->isPassID(&ProfileInfo::ID))
+ return (ProfileInfo*)this;
+ return this;
+ }
+
+ virtual void recurseBasicBlock(BasicBlock *BB);
+
+ void inline printEdgeWeight(Edge);
+ };
+} // End of anonymous namespace
+
+char ProfileEstimatorPass::ID = 0;
+static RegisterPass<ProfileEstimatorPass>
+X("profile-estimator", "Estimate profiling information", false, true);
+
+static RegisterAnalysisGroup<ProfileInfo> Y(X);
+
+namespace llvm {
+ const PassInfo *ProfileEstimatorPassID = &X;
+
+ FunctionPass *createProfileEstimatorPass() {
+ return new ProfileEstimatorPass();
+ }
+
+ /// createProfileEstimatorPass - This function returns a Pass that estimates
+ /// profiling information using the given loop execution count.
+ Pass *createProfileEstimatorPass(const unsigned execcount) {
+ return new ProfileEstimatorPass(execcount);
+ }
+}
+
+static double ignoreMissing(double w) {
+ if (w == ProfileInfo::MissingValue) return 0;
+ return w;
+}
+
+static void inline printEdgeError(ProfileInfo::Edge e, const char *M) {
+ DEBUG(dbgs() << "-- Edge " << e << " is not calculated, " << M << "\n");
+}
+
+void inline ProfileEstimatorPass::printEdgeWeight(Edge E) {
+ DEBUG(dbgs() << "-- Weight of Edge " << E << ":"
+ << format("%20.20g", getEdgeWeight(E)) << "\n");
+}
+
+// recurseBasicBlock() - This calculates the ProfileInfo estimation for a
+// single block and then recurses into the successors.
+// The algorithm preserves the flow condition, meaning that the sum of the
+// weight of the incoming edges must be equal the block weight which must in
+// turn be equal to the sume of the weights of the outgoing edges.
+// Since the flow of an block is deterimined from the current state of the
+// flow, once an edge has a flow assigned this flow is never changed again,
+// otherwise it would be possible to violate the flow condition in another
+// block.
+void ProfileEstimatorPass::recurseBasicBlock(BasicBlock *BB) {
+
+ // Break the recursion if this BasicBlock was already visited.
+ if (BBToVisit.find(BB) == BBToVisit.end()) return;
+
+ // Read the LoopInfo for this block.
+ bool BBisHeader = LI->isLoopHeader(BB);
+ Loop* BBLoop = LI->getLoopFor(BB);
+
+ // To get the block weight, read all incoming edges.
+ double BBWeight = 0;
+ std::set<BasicBlock*> ProcessedPreds;
+ for ( pred_iterator bbi = pred_begin(BB), bbe = pred_end(BB);
+ bbi != bbe; ++bbi ) {
+ // If this block was not considered already, add weight.
+ Edge edge = getEdge(*bbi,BB);
+ double w = getEdgeWeight(edge);
+ if (ProcessedPreds.insert(*bbi).second) {
+ BBWeight += ignoreMissing(w);
+ }
+ // If this block is a loop header and the predecessor is contained in this
+ // loop, thus the edge is a backedge, continue and do not check if the
+ // value is valid.
+ if (BBisHeader && BBLoop->contains(*bbi)) {
+ printEdgeError(edge, "but is backedge, continueing");
+ continue;
+ }
+ // If the edges value is missing (and this is no loop header, and this is
+ // no backedge) return, this block is currently non estimatable.
+ if (w == MissingValue) {
+ printEdgeError(edge, "returning");
+ return;
+ }
+ }
+ if (getExecutionCount(BB) != MissingValue) {
+ BBWeight = getExecutionCount(BB);
+ }
+
+ // Fetch all necessary information for current block.
+ SmallVector<Edge, 8> ExitEdges;
+ SmallVector<Edge, 8> Edges;
+ if (BBLoop) {
+ BBLoop->getExitEdges(ExitEdges);
+ }
+
+ // If this is a loop header, consider the following:
+ // Exactly the flow that is entering this block, must exit this block too. So
+ // do the following:
+ // *) get all the exit edges, read the flow that is already leaving this
+ // loop, remember the edges that do not have any flow on them right now.
+ // (The edges that have already flow on them are most likely exiting edges of
+ // other loops, do not touch those flows because the previously caclulated
+ // loopheaders would not be exact anymore.)
+ // *) In case there is not a single exiting edge left, create one at the loop
+ // latch to prevent the flow from building up in the loop.
+ // *) Take the flow that is not leaving the loop already and distribute it on
+ // the remaining exiting edges.
+ // (This ensures that all flow that enters the loop also leaves it.)
+ // *) Increase the flow into the loop by increasing the weight of this block.
+ // There is at least one incoming backedge that will bring us this flow later
+ // on. (So that the flow condition in this node is valid again.)
+ if (BBisHeader) {
+ double incoming = BBWeight;
+ // Subtract the flow leaving the loop.
+ std::set<Edge> ProcessedExits;
+ for (SmallVector<Edge, 8>::iterator ei = ExitEdges.begin(),
+ ee = ExitEdges.end(); ei != ee; ++ei) {
+ if (ProcessedExits.insert(*ei).second) {
+ double w = getEdgeWeight(*ei);
+ if (w == MissingValue) {
+ Edges.push_back(*ei);
+ // Check if there is a necessary minimal weight, if yes, subtract it
+ // from weight.
+ if (MinimalWeight.find(*ei) != MinimalWeight.end()) {
+ incoming -= MinimalWeight[*ei];
+ DEBUG(dbgs() << "Reserving " << format("%.20g",MinimalWeight[*ei]) << " at " << (*ei) << "\n");
+ }
+ } else {
+ incoming -= w;
+ }
+ }
+ }
+ // If no exit edges, create one:
+ if (Edges.size() == 0) {
+ BasicBlock *Latch = BBLoop->getLoopLatch();
+ if (Latch) {
+ Edge edge = getEdge(Latch,0);
+ EdgeInformation[BB->getParent()][edge] = BBWeight;
+ printEdgeWeight(edge);
+ edge = getEdge(Latch, BB);
+ EdgeInformation[BB->getParent()][edge] = BBWeight * ExecCount;
+ printEdgeWeight(edge);
+ }
+ }
+
+ // Distribute remaining weight to the exting edges. To prevent fractions
+ // from building up and provoking precision problems the weight which is to
+ // be distributed is split and the rounded, the last edge gets a somewhat
+ // bigger value, but we are close enough for an estimation.
+ double fraction = floor(incoming/Edges.size());
+ for (SmallVector<Edge, 8>::iterator ei = Edges.begin(), ee = Edges.end();
+ ei != ee; ++ei) {
+ double w = 0;
+ if (ei != (ee-1)) {
+ w = fraction;
+ incoming -= fraction;
+ } else {
+ w = incoming;
+ }
+ EdgeInformation[BB->getParent()][*ei] += w;
+ // Read necessary minimal weight.
+ if (MinimalWeight.find(*ei) != MinimalWeight.end()) {
+ EdgeInformation[BB->getParent()][*ei] += MinimalWeight[*ei];
+ DEBUG(dbgs() << "Additionally " << format("%.20g",MinimalWeight[*ei]) << " at " << (*ei) << "\n");
+ }
+ printEdgeWeight(*ei);
+
+ // Add minimal weight to paths to all exit edges, this is used to ensure
+ // that enough flow is reaching this edges.
+ Path p;
+ const BasicBlock *Dest = GetPath(BB, (*ei).first, p, GetPathToDest);
+ while (Dest != BB) {
+ const BasicBlock *Parent = p.find(Dest)->second;
+ Edge e = getEdge(Parent, Dest);
+ if (MinimalWeight.find(e) == MinimalWeight.end()) {
+ MinimalWeight[e] = 0;
+ }
+ MinimalWeight[e] += w;
+ DEBUG(dbgs() << "Minimal Weight for " << e << ": " << format("%.20g",MinimalWeight[e]) << "\n");
+ Dest = Parent;
+ }
+ }
+ // Increase flow into the loop.
+ BBWeight *= (ExecCount+1);
+ }
+
+ BlockInformation[BB->getParent()][BB] = BBWeight;
+ // Up until now we considered only the loop exiting edges, now we have a
+ // definite block weight and must distribute this onto the outgoing edges.
+ // Since there may be already flow attached to some of the edges, read this
+ // flow first and remember the edges that have still now flow attached.
+ Edges.clear();
+ std::set<BasicBlock*> ProcessedSuccs;
+
+ succ_iterator bbi = succ_begin(BB), bbe = succ_end(BB);
+ // Also check for (BB,0) edges that may already contain some flow. (But only
+ // in case there are no successors.)
+ if (bbi == bbe) {
+ Edge edge = getEdge(BB,0);
+ EdgeInformation[BB->getParent()][edge] = BBWeight;
+ printEdgeWeight(edge);
+ }
+ for ( ; bbi != bbe; ++bbi ) {
+ if (ProcessedSuccs.insert(*bbi).second) {
+ Edge edge = getEdge(BB,*bbi);
+ double w = getEdgeWeight(edge);
+ if (w != MissingValue) {
+ BBWeight -= getEdgeWeight(edge);
+ } else {
+ Edges.push_back(edge);
+ // If minimal weight is necessary, reserve weight by subtracting weight
+ // from block weight, this is readded later on.
+ if (MinimalWeight.find(edge) != MinimalWeight.end()) {
+ BBWeight -= MinimalWeight[edge];
+ DEBUG(dbgs() << "Reserving " << format("%.20g",MinimalWeight[edge]) << " at " << edge << "\n");
+ }
+ }
+ }
+ }
+
+ double fraction = floor(BBWeight/Edges.size());
+ // Finally we know what flow is still not leaving the block, distribute this
+ // flow onto the empty edges.
+ for (SmallVector<Edge, 8>::iterator ei = Edges.begin(), ee = Edges.end();
+ ei != ee; ++ei) {
+ if (ei != (ee-1)) {
+ EdgeInformation[BB->getParent()][*ei] += fraction;
+ BBWeight -= fraction;
+ } else {
+ EdgeInformation[BB->getParent()][*ei] += BBWeight;
+ }
+ // Readd minial necessary weight.
+ if (MinimalWeight.find(*ei) != MinimalWeight.end()) {
+ EdgeInformation[BB->getParent()][*ei] += MinimalWeight[*ei];
+ DEBUG(dbgs() << "Additionally " << format("%.20g",MinimalWeight[*ei]) << " at " << (*ei) << "\n");
+ }
+ printEdgeWeight(*ei);
+ }
+
+ // This block is visited, mark this before the recursion.
+ BBToVisit.erase(BB);
+
+ // Recurse into successors.
+ for (succ_iterator bbi = succ_begin(BB), bbe = succ_end(BB);
+ bbi != bbe; ++bbi) {
+ recurseBasicBlock(*bbi);
+ }
+}
+
+bool ProfileEstimatorPass::runOnFunction(Function &F) {
+ if (F.isDeclaration()) return false;
+
+ // Fetch LoopInfo and clear ProfileInfo for this function.
+ LI = &getAnalysis<LoopInfo>();
+ FunctionInformation.erase(&F);
+ BlockInformation[&F].clear();
+ EdgeInformation[&F].clear();
+
+ // Mark all blocks as to visit.
+ for (Function::iterator bi = F.begin(), be = F.end(); bi != be; ++bi)
+ BBToVisit.insert(bi);
+
+ // Clear Minimal Edges.
+ MinimalWeight.clear();
+
+ DEBUG(dbgs() << "Working on function " << F.getNameStr() << "\n");
+
+ // Since the entry block is the first one and has no predecessors, the edge
+ // (0,entry) is inserted with the starting weight of 1.
+ BasicBlock *entry = &F.getEntryBlock();
+ BlockInformation[&F][entry] = pow(2.0, 32.0);
+ Edge edge = getEdge(0,entry);
+ EdgeInformation[&F][edge] = BlockInformation[&F][entry];
+ printEdgeWeight(edge);
+
+ // Since recurseBasicBlock() maybe returns with a block which was not fully
+ // estimated, use recurseBasicBlock() until everything is calculated.
+ bool cleanup = false;
+ recurseBasicBlock(entry);
+ while (BBToVisit.size() > 0 && !cleanup) {
+ // Remember number of open blocks, this is later used to check if progress
+ // was made.
+ unsigned size = BBToVisit.size();
+
+ // Try to calculate all blocks in turn.
+ for (std::set<BasicBlock*>::iterator bi = BBToVisit.begin(),
+ be = BBToVisit.end(); bi != be; ++bi) {
+ recurseBasicBlock(*bi);
+ // If at least one block was finished, break because iterator may be
+ // invalid.
+ if (BBToVisit.size() < size) break;
+ }
+
+ // If there was not a single block resolved, make some assumptions.
+ if (BBToVisit.size() == size) {
+ bool found = false;
+ for (std::set<BasicBlock*>::iterator BBI = BBToVisit.begin(), BBE = BBToVisit.end();
+ (BBI != BBE) && (!found); ++BBI) {
+ BasicBlock *BB = *BBI;
+ // Try each predecessor if it can be assumend.
+ for (pred_iterator bbi = pred_begin(BB), bbe = pred_end(BB);
+ (bbi != bbe) && (!found); ++bbi) {
+ Edge e = getEdge(*bbi,BB);
+ double w = getEdgeWeight(e);
+ // Check that edge from predecessor is still free.
+ if (w == MissingValue) {
+ // Check if there is a circle from this block to predecessor.
+ Path P;
+ const BasicBlock *Dest = GetPath(BB, *bbi, P, GetPathToDest);
+ if (Dest != *bbi) {
+ // If there is no circle, just set edge weight to 0
+ EdgeInformation[&F][e] = 0;
+ DEBUG(dbgs() << "Assuming edge weight: ");
+ printEdgeWeight(e);
+ found = true;
+ }
+ }
+ }
+ }
+ if (!found) {
+ cleanup = true;
+ DEBUG(dbgs() << "No assumption possible in Fuction "<<F.getName()<<", setting all to zero\n");
+ }
+ }
+ }
+ // In case there was no safe way to assume edges, set as a last measure,
+ // set _everything_ to zero.
+ if (cleanup) {
+ FunctionInformation[&F] = 0;
+ BlockInformation[&F].clear();
+ EdgeInformation[&F].clear();
+ for (Function::const_iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
+ const BasicBlock *BB = &(*FI);
+ BlockInformation[&F][BB] = 0;
+ pred_const_iterator predi = pred_begin(BB), prede = pred_end(BB);
+ if (predi == prede) {
+ Edge e = getEdge(0,BB);
+ setEdgeWeight(e,0);
+ }
+ for (;predi != prede; ++predi) {
+ Edge e = getEdge(*predi,BB);
+ setEdgeWeight(e,0);
+ }
+ succ_const_iterator succi = succ_begin(BB), succe = succ_end(BB);
+ if (succi == succe) {
+ Edge e = getEdge(BB,0);
+ setEdgeWeight(e,0);
+ }
+ for (;succi != succe; ++succi) {
+ Edge e = getEdge(*succi,BB);
+ setEdgeWeight(e,0);
+ }
+ }
+ }
+
+ return false;
+}
diff --git a/lib/Analysis/ProfileInfo.cpp b/lib/Analysis/ProfileInfo.cpp
new file mode 100644
index 0000000..85531be
--- /dev/null
+++ b/lib/Analysis/ProfileInfo.cpp
@@ -0,0 +1,1107 @@
+//===- ProfileInfo.cpp - Profile Info Interface ---------------------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements the abstract ProfileInfo interface, and the default
+// "no profile" implementation.
+//
+//===----------------------------------------------------------------------===//
+#define DEBUG_TYPE "profile-info"
+#include "llvm/Analysis/Passes.h"
+#include "llvm/Analysis/ProfileInfo.h"
+#include "llvm/CodeGen/MachineBasicBlock.h"
+#include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/Pass.h"
+#include "llvm/Support/CFG.h"
+#include "llvm/ADT/SmallSet.h"
+#include <set>
+#include <queue>
+#include <limits>
+using namespace llvm;
+
+// Register the ProfileInfo interface, providing a nice name to refer to.
+static RegisterAnalysisGroup<ProfileInfo> Z("Profile Information");
+
+namespace llvm {
+
+template <>
+ProfileInfoT<MachineFunction, MachineBasicBlock>::ProfileInfoT() {}
+template <>
+ProfileInfoT<MachineFunction, MachineBasicBlock>::~ProfileInfoT() {}
+
+template <>
+ProfileInfoT<Function, BasicBlock>::ProfileInfoT() {
+ MachineProfile = 0;
+}
+template <>
+ProfileInfoT<Function, BasicBlock>::~ProfileInfoT() {
+ if (MachineProfile) delete MachineProfile;
+}
+
+template<>
+char ProfileInfoT<Function,BasicBlock>::ID = 0;
+
+template<>
+char ProfileInfoT<MachineFunction, MachineBasicBlock>::ID = 0;
+
+template<>
+const double ProfileInfoT<Function,BasicBlock>::MissingValue = -1;
+
+template<> const
+double ProfileInfoT<MachineFunction, MachineBasicBlock>::MissingValue = -1;
+
+template<> double
+ProfileInfoT<Function,BasicBlock>::getExecutionCount(const BasicBlock *BB) {
+ std::map<const Function*, BlockCounts>::iterator J =
+ BlockInformation.find(BB->getParent());
+ if (J != BlockInformation.end()) {
+ BlockCounts::iterator I = J->second.find(BB);
+ if (I != J->second.end())
+ return I->second;
+ }
+
+ double Count = MissingValue;
+
+ pred_const_iterator PI = pred_begin(BB), PE = pred_end(BB);
+
+ // Are there zero predecessors of this block?
+ if (PI == PE) {
+ Edge e = getEdge(0,BB);
+ Count = getEdgeWeight(e);
+ } else {
+ // Otherwise, if there are predecessors, the execution count of this block is
+ // the sum of the edge frequencies from the incoming edges.
+ std::set<const BasicBlock*> ProcessedPreds;
+ Count = 0;
+ for (; PI != PE; ++PI)
+ if (ProcessedPreds.insert(*PI).second) {
+ double w = getEdgeWeight(getEdge(*PI, BB));
+ if (w == MissingValue) {
+ Count = MissingValue;
+ break;
+ }
+ Count += w;
+ }
+ }
+
+ // If the predecessors did not suffice to get block weight, try successors.
+ if (Count == MissingValue) {
+
+ succ_const_iterator SI = succ_begin(BB), SE = succ_end(BB);
+
+ // Are there zero successors of this block?
+ if (SI == SE) {
+ Edge e = getEdge(BB,0);
+ Count = getEdgeWeight(e);
+ } else {
+ std::set<const BasicBlock*> ProcessedSuccs;
+ Count = 0;
+ for (; SI != SE; ++SI)
+ if (ProcessedSuccs.insert(*SI).second) {
+ double w = getEdgeWeight(getEdge(BB, *SI));
+ if (w == MissingValue) {
+ Count = MissingValue;
+ break;
+ }
+ Count += w;
+ }
+ }
+ }
+
+ if (Count != MissingValue) BlockInformation[BB->getParent()][BB] = Count;
+ return Count;
+}
+
+template<>
+double ProfileInfoT<MachineFunction, MachineBasicBlock>::
+ getExecutionCount(const MachineBasicBlock *MBB) {
+ std::map<const MachineFunction*, BlockCounts>::iterator J =
+ BlockInformation.find(MBB->getParent());
+ if (J != BlockInformation.end()) {
+ BlockCounts::iterator I = J->second.find(MBB);
+ if (I != J->second.end())
+ return I->second;
+ }
+
+ return MissingValue;
+}
+
+template<>
+double ProfileInfoT<Function,BasicBlock>::getExecutionCount(const Function *F) {
+ std::map<const Function*, double>::iterator J =
+ FunctionInformation.find(F);
+ if (J != FunctionInformation.end())
+ return J->second;
+
+ // isDeclaration() is checked here and not at start of function to allow
+ // functions without a body still to have a execution count.
+ if (F->isDeclaration()) return MissingValue;
+
+ double Count = getExecutionCount(&F->getEntryBlock());
+ if (Count != MissingValue) FunctionInformation[F] = Count;
+ return Count;
+}
+
+template<>
+double ProfileInfoT<MachineFunction, MachineBasicBlock>::
+ getExecutionCount(const MachineFunction *MF) {
+ std::map<const MachineFunction*, double>::iterator J =
+ FunctionInformation.find(MF);
+ if (J != FunctionInformation.end())
+ return J->second;
+
+ double Count = getExecutionCount(&MF->front());
+ if (Count != MissingValue) FunctionInformation[MF] = Count;
+ return Count;
+}
+
+template<>
+void ProfileInfoT<Function,BasicBlock>::
+ setExecutionCount(const BasicBlock *BB, double w) {
+ DEBUG(dbgs() << "Creating Block " << BB->getName()
+ << " (weight: " << format("%.20g",w) << ")\n");
+ BlockInformation[BB->getParent()][BB] = w;
+}
+
+template<>
+void ProfileInfoT<MachineFunction, MachineBasicBlock>::
+ setExecutionCount(const MachineBasicBlock *MBB, double w) {
+ DEBUG(dbgs() << "Creating Block " << MBB->getBasicBlock()->getName()
+ << " (weight: " << format("%.20g",w) << ")\n");
+ BlockInformation[MBB->getParent()][MBB] = w;
+}
+
+template<>
+void ProfileInfoT<Function,BasicBlock>::addEdgeWeight(Edge e, double w) {
+ double oldw = getEdgeWeight(e);
+ assert (oldw != MissingValue && "Adding weight to Edge with no previous weight");
+ DEBUG(dbgs() << "Adding to Edge " << e
+ << " (new weight: " << format("%.20g",oldw + w) << ")\n");
+ EdgeInformation[getFunction(e)][e] = oldw + w;
+}
+
+template<>
+void ProfileInfoT<Function,BasicBlock>::
+ addExecutionCount(const BasicBlock *BB, double w) {
+ double oldw = getExecutionCount(BB);
+ assert (oldw != MissingValue && "Adding weight to Block with no previous weight");
+ DEBUG(dbgs() << "Adding to Block " << BB->getName()
+ << " (new weight: " << format("%.20g",oldw + w) << ")\n");
+ BlockInformation[BB->getParent()][BB] = oldw + w;
+}
+
+template<>
+void ProfileInfoT<Function,BasicBlock>::removeBlock(const BasicBlock *BB) {
+ std::map<const Function*, BlockCounts>::iterator J =
+ BlockInformation.find(BB->getParent());
+ if (J == BlockInformation.end()) return;
+
+ DEBUG(dbgs() << "Deleting " << BB->getName() << "\n");
+ J->second.erase(BB);
+}
+
+template<>
+void ProfileInfoT<Function,BasicBlock>::removeEdge(Edge e) {
+ std::map<const Function*, EdgeWeights>::iterator J =
+ EdgeInformation.find(getFunction(e));
+ if (J == EdgeInformation.end()) return;
+
+ DEBUG(dbgs() << "Deleting" << e << "\n");
+ J->second.erase(e);
+}
+
+template<>
+void ProfileInfoT<Function,BasicBlock>::
+ replaceEdge(const Edge &oldedge, const Edge &newedge) {
+ double w;
+ if ((w = getEdgeWeight(newedge)) == MissingValue) {
+ w = getEdgeWeight(oldedge);
+ DEBUG(dbgs() << "Replacing " << oldedge << " with " << newedge << "\n");
+ } else {
+ w += getEdgeWeight(oldedge);
+ DEBUG(dbgs() << "Adding " << oldedge << " to " << newedge << "\n");
+ }
+ setEdgeWeight(newedge,w);
+ removeEdge(oldedge);
+}
+
+template<>
+const BasicBlock *ProfileInfoT<Function,BasicBlock>::
+ GetPath(const BasicBlock *Src, const BasicBlock *Dest,
+ Path &P, unsigned Mode) {
+ const BasicBlock *BB = 0;
+ bool hasFoundPath = false;
+
+ std::queue<const BasicBlock *> BFS;
+ BFS.push(Src);
+
+ while(BFS.size() && !hasFoundPath) {
+ BB = BFS.front();
+ BFS.pop();
+
+ succ_const_iterator Succ = succ_begin(BB), End = succ_end(BB);
+ if (Succ == End) {
+ P[0] = BB;
+ if (Mode & GetPathToExit) {
+ hasFoundPath = true;
+ BB = 0;
+ }
+ }
+ for(;Succ != End; ++Succ) {
+ if (P.find(*Succ) != P.end()) continue;
+ Edge e = getEdge(BB,*Succ);
+ if ((Mode & GetPathWithNewEdges) && (getEdgeWeight(e) != MissingValue)) continue;
+ P[*Succ] = BB;
+ BFS.push(*Succ);
+ if ((Mode & GetPathToDest) && *Succ == Dest) {
+ hasFoundPath = true;
+ BB = *Succ;
+ break;
+ }
+ if ((Mode & GetPathToValue) && (getExecutionCount(*Succ) != MissingValue)) {
+ hasFoundPath = true;
+ BB = *Succ;
+ break;
+ }
+ }
+ }
+
+ return BB;
+}
+
+template<>
+void ProfileInfoT<Function,BasicBlock>::
+ divertFlow(const Edge &oldedge, const Edge &newedge) {
+ DEBUG(dbgs() << "Diverting " << oldedge << " via " << newedge );
+
+ // First check if the old edge was taken, if not, just delete it...
+ if (getEdgeWeight(oldedge) == 0) {
+ removeEdge(oldedge);
+ return;
+ }
+
+ Path P;
+ P[newedge.first] = 0;
+ P[newedge.second] = newedge.first;
+ const BasicBlock *BB = GetPath(newedge.second,oldedge.second,P,GetPathToExit | GetPathToDest);
+
+ double w = getEdgeWeight (oldedge);
+ DEBUG(dbgs() << ", Weight: " << format("%.20g",w) << "\n");
+ do {
+ const BasicBlock *Parent = P.find(BB)->second;
+ Edge e = getEdge(Parent,BB);
+ double oldw = getEdgeWeight(e);
+ double oldc = getExecutionCount(e.first);
+ setEdgeWeight(e, w+oldw);
+ if (Parent != oldedge.first) {
+ setExecutionCount(e.first, w+oldc);
+ }
+ BB = Parent;
+ } while (BB != newedge.first);
+ removeEdge(oldedge);
+}
+
+/// Replaces all occurences of RmBB in the ProfilingInfo with DestBB.
+/// This checks all edges of the function the blocks reside in and replaces the
+/// occurences of RmBB with DestBB.
+template<>
+void ProfileInfoT<Function,BasicBlock>::
+ replaceAllUses(const BasicBlock *RmBB, const BasicBlock *DestBB) {
+ DEBUG(dbgs() << "Replacing " << RmBB->getName()
+ << " with " << DestBB->getName() << "\n");
+ const Function *F = DestBB->getParent();
+ std::map<const Function*, EdgeWeights>::iterator J =
+ EdgeInformation.find(F);
+ if (J == EdgeInformation.end()) return;
+
+ Edge e, newedge;
+ bool erasededge = false;
+ EdgeWeights::iterator I = J->second.begin(), E = J->second.end();
+ while(I != E) {
+ e = (I++)->first;
+ bool foundedge = false; bool eraseedge = false;
+ if (e.first == RmBB) {
+ if (e.second == DestBB) {
+ eraseedge = true;
+ } else {
+ newedge = getEdge(DestBB, e.second);
+ foundedge = true;
+ }
+ }
+ if (e.second == RmBB) {
+ if (e.first == DestBB) {
+ eraseedge = true;
+ } else {
+ newedge = getEdge(e.first, DestBB);
+ foundedge = true;
+ }
+ }
+ if (foundedge) {
+ replaceEdge(e, newedge);
+ }
+ if (eraseedge) {
+ if (erasededge) {
+ Edge newedge = getEdge(DestBB, DestBB);
+ replaceEdge(e, newedge);
+ } else {
+ removeEdge(e);
+ erasededge = true;
+ }
+ }
+ }
+}
+
+/// Splits an edge in the ProfileInfo and redirects flow over NewBB.
+/// Since its possible that there is more than one edge in the CFG from FristBB
+/// to SecondBB its necessary to redirect the flow proporionally.
+template<>
+void ProfileInfoT<Function,BasicBlock>::splitEdge(const BasicBlock *FirstBB,
+ const BasicBlock *SecondBB,
+ const BasicBlock *NewBB,
+ bool MergeIdenticalEdges) {
+ const Function *F = FirstBB->getParent();
+ std::map<const Function*, EdgeWeights>::iterator J =
+ EdgeInformation.find(F);
+ if (J == EdgeInformation.end()) return;
+
+ // Generate edges and read current weight.
+ Edge e = getEdge(FirstBB, SecondBB);
+ Edge n1 = getEdge(FirstBB, NewBB);
+ Edge n2 = getEdge(NewBB, SecondBB);
+ EdgeWeights &ECs = J->second;
+ double w = ECs[e];
+
+ int succ_count = 0;
+ if (!MergeIdenticalEdges) {
+ // First count the edges from FristBB to SecondBB, if there is more than
+ // one, only slice out a proporional part for NewBB.
+ for(succ_const_iterator BBI = succ_begin(FirstBB), BBE = succ_end(FirstBB);
+ BBI != BBE; ++BBI) {
+ if (*BBI == SecondBB) succ_count++;
+ }
+ // When the NewBB is completely new, increment the count by one so that
+ // the counts are properly distributed.
+ if (getExecutionCount(NewBB) == ProfileInfo::MissingValue) succ_count++;
+ } else {
+ // When the edges are merged anyway, then redirect all flow.
+ succ_count = 1;
+ }
+
+ // We know now how many edges there are from FirstBB to SecondBB, reroute a
+ // proportional part of the edge weight over NewBB.
+ double neww = floor(w / succ_count);
+ ECs[n1] += neww;
+ ECs[n2] += neww;
+ BlockInformation[F][NewBB] += neww;
+ if (succ_count == 1) {
+ ECs.erase(e);
+ } else {
+ ECs[e] -= neww;
+ }
+}
+
+template<>
+void ProfileInfoT<Function,BasicBlock>::splitBlock(const BasicBlock *Old,
+ const BasicBlock* New) {
+ const Function *F = Old->getParent();
+ std::map<const Function*, EdgeWeights>::iterator J =
+ EdgeInformation.find(F);
+ if (J == EdgeInformation.end()) return;
+
+ DEBUG(dbgs() << "Splitting " << Old->getName() << " to " << New->getName() << "\n");
+
+ std::set<Edge> Edges;
+ for (EdgeWeights::iterator ewi = J->second.begin(), ewe = J->second.end();
+ ewi != ewe; ++ewi) {
+ Edge old = ewi->first;
+ if (old.first == Old) {
+ Edges.insert(old);
+ }
+ }
+ for (std::set<Edge>::iterator EI = Edges.begin(), EE = Edges.end();
+ EI != EE; ++EI) {
+ Edge newedge = getEdge(New, EI->second);
+ replaceEdge(*EI, newedge);
+ }
+
+ double w = getExecutionCount(Old);
+ setEdgeWeight(getEdge(Old, New), w);
+ setExecutionCount(New, w);
+}
+
+template<>
+void ProfileInfoT<Function,BasicBlock>::splitBlock(const BasicBlock *BB,
+ const BasicBlock* NewBB,
+ BasicBlock *const *Preds,
+ unsigned NumPreds) {
+ const Function *F = BB->getParent();
+ std::map<const Function*, EdgeWeights>::iterator J =
+ EdgeInformation.find(F);
+ if (J == EdgeInformation.end()) return;
+
+ DEBUG(dbgs() << "Splitting " << NumPreds << " Edges from " << BB->getName()
+ << " to " << NewBB->getName() << "\n");
+
+ // Collect weight that was redirected over NewBB.
+ double newweight = 0;
+
+ std::set<const BasicBlock *> ProcessedPreds;
+ // For all requestes Predecessors.
+ for (unsigned pred = 0; pred < NumPreds; ++pred) {
+ const BasicBlock * Pred = Preds[pred];
+ if (ProcessedPreds.insert(Pred).second) {
+ // Create edges and read old weight.
+ Edge oldedge = getEdge(Pred, BB);
+ Edge newedge = getEdge(Pred, NewBB);
+
+ // Remember how much weight was redirected.
+ newweight += getEdgeWeight(oldedge);
+
+ replaceEdge(oldedge,newedge);
+ }
+ }
+
+ Edge newedge = getEdge(NewBB,BB);
+ setEdgeWeight(newedge, newweight);
+ setExecutionCount(NewBB, newweight);
+}
+
+template<>
+void ProfileInfoT<Function,BasicBlock>::transfer(const Function *Old,
+ const Function *New) {
+ DEBUG(dbgs() << "Replacing Function " << Old->getName() << " with "
+ << New->getName() << "\n");
+ std::map<const Function*, EdgeWeights>::iterator J =
+ EdgeInformation.find(Old);
+ if(J != EdgeInformation.end()) {
+ EdgeInformation[New] = J->second;
+ }
+ EdgeInformation.erase(Old);
+ BlockInformation.erase(Old);
+ FunctionInformation.erase(Old);
+}
+
+static double readEdgeOrRemember(ProfileInfo::Edge edge, double w,
+ ProfileInfo::Edge &tocalc, unsigned &uncalc) {
+ if (w == ProfileInfo::MissingValue) {
+ tocalc = edge;
+ uncalc++;
+ return 0;
+ } else {
+ return w;
+ }
+}
+
+template<>
+bool ProfileInfoT<Function,BasicBlock>::
+ CalculateMissingEdge(const BasicBlock *BB, Edge &removed,
+ bool assumeEmptySelf) {
+ Edge edgetocalc;
+ unsigned uncalculated = 0;
+
+ // collect weights of all incoming and outgoing edges, rememer edges that
+ // have no value
+ double incount = 0;
+ SmallSet<const BasicBlock*,8> pred_visited;
+ pred_const_iterator bbi = pred_begin(BB), bbe = pred_end(BB);
+ if (bbi==bbe) {
+ Edge e = getEdge(0,BB);
+ incount += readEdgeOrRemember(e, getEdgeWeight(e) ,edgetocalc,uncalculated);
+ }
+ for (;bbi != bbe; ++bbi) {
+ if (pred_visited.insert(*bbi)) {
+ Edge e = getEdge(*bbi,BB);
+ incount += readEdgeOrRemember(e, getEdgeWeight(e) ,edgetocalc,uncalculated);
+ }
+ }
+
+ double outcount = 0;
+ SmallSet<const BasicBlock*,8> succ_visited;
+ succ_const_iterator sbbi = succ_begin(BB), sbbe = succ_end(BB);
+ if (sbbi==sbbe) {
+ Edge e = getEdge(BB,0);
+ if (getEdgeWeight(e) == MissingValue) {
+ double w = getExecutionCount(BB);
+ if (w != MissingValue) {
+ setEdgeWeight(e,w);
+ removed = e;
+ }
+ }
+ outcount += readEdgeOrRemember(e, getEdgeWeight(e), edgetocalc, uncalculated);
+ }
+ for (;sbbi != sbbe; ++sbbi) {
+ if (succ_visited.insert(*sbbi)) {
+ Edge e = getEdge(BB,*sbbi);
+ outcount += readEdgeOrRemember(e, getEdgeWeight(e), edgetocalc, uncalculated);
+ }
+ }
+
+ // if exactly one edge weight was missing, calculate it and remove it from
+ // spanning tree
+ if (uncalculated == 0 ) {
+ return true;
+ } else
+ if (uncalculated == 1) {
+ if (incount < outcount) {
+ EdgeInformation[BB->getParent()][edgetocalc] = outcount-incount;
+ } else {
+ EdgeInformation[BB->getParent()][edgetocalc] = incount-outcount;
+ }
+ DEBUG(dbgs() << "--Calc Edge Counter for " << edgetocalc << ": "
+ << format("%.20g", getEdgeWeight(edgetocalc)) << "\n");
+ removed = edgetocalc;
+ return true;
+ } else
+ if (uncalculated == 2 && assumeEmptySelf && edgetocalc.first == edgetocalc.second && incount == outcount) {
+ setEdgeWeight(edgetocalc, incount * 10);
+ removed = edgetocalc;
+ return true;
+ } else {
+ return false;
+ }
+}
+
+static void readEdge(ProfileInfo *PI, ProfileInfo::Edge e, double &calcw, std::set<ProfileInfo::Edge> &misscount) {
+ double w = PI->getEdgeWeight(e);
+ if (w != ProfileInfo::MissingValue) {
+ calcw += w;
+ } else {
+ misscount.insert(e);
+ }
+}
+
+template<>
+bool ProfileInfoT<Function,BasicBlock>::EstimateMissingEdges(const BasicBlock *BB) {
+ bool hasNoSuccessors = false;
+
+ double inWeight = 0;
+ std::set<Edge> inMissing;
+ std::set<const BasicBlock*> ProcessedPreds;
+ pred_const_iterator bbi = pred_begin(BB), bbe = pred_end(BB);
+ if (bbi == bbe) {
+ readEdge(this,getEdge(0,BB),inWeight,inMissing);
+ }
+ for( ; bbi != bbe; ++bbi ) {
+ if (ProcessedPreds.insert(*bbi).second) {
+ readEdge(this,getEdge(*bbi,BB),inWeight,inMissing);
+ }
+ }
+
+ double outWeight = 0;
+ std::set<Edge> outMissing;
+ std::set<const BasicBlock*> ProcessedSuccs;
+ succ_const_iterator sbbi = succ_begin(BB), sbbe = succ_end(BB);
+ if (sbbi == sbbe) {
+ readEdge(this,getEdge(BB,0),outWeight,outMissing);
+ hasNoSuccessors = true;
+ }
+ for ( ; sbbi != sbbe; ++sbbi ) {
+ if (ProcessedSuccs.insert(*sbbi).second) {
+ readEdge(this,getEdge(BB,*sbbi),outWeight,outMissing);
+ }
+ }
+
+ double share;
+ std::set<Edge>::iterator ei,ee;
+ if (inMissing.size() == 0 && outMissing.size() > 0) {
+ ei = outMissing.begin();
+ ee = outMissing.end();
+ share = inWeight/outMissing.size();
+ setExecutionCount(BB,inWeight);
+ } else
+ if (inMissing.size() > 0 && outMissing.size() == 0 && outWeight == 0) {
+ ei = inMissing.begin();
+ ee = inMissing.end();
+ share = 0;
+ setExecutionCount(BB,0);
+ } else
+ if (inMissing.size() == 0 && outMissing.size() == 0) {
+ setExecutionCount(BB,outWeight);
+ return true;
+ } else {
+ return false;
+ }
+ for ( ; ei != ee; ++ei ) {
+ setEdgeWeight(*ei,share);
+ }
+ return true;
+}
+
+template<>
+void ProfileInfoT<Function,BasicBlock>::repair(const Function *F) {
+// if (getExecutionCount(&(F->getEntryBlock())) == 0) {
+// for (Function::const_iterator FI = F->begin(), FE = F->end();
+// FI != FE; ++FI) {
+// const BasicBlock* BB = &(*FI);
+// {
+// pred_const_iterator NBB = pred_begin(BB), End = pred_end(BB);
+// if (NBB == End) {
+// setEdgeWeight(getEdge(0,BB),0);
+// }
+// for(;NBB != End; ++NBB) {
+// setEdgeWeight(getEdge(*NBB,BB),0);
+// }
+// }
+// {
+// succ_const_iterator NBB = succ_begin(BB), End = succ_end(BB);
+// if (NBB == End) {
+// setEdgeWeight(getEdge(0,BB),0);
+// }
+// for(;NBB != End; ++NBB) {
+// setEdgeWeight(getEdge(*NBB,BB),0);
+// }
+// }
+// }
+// return;
+// }
+ // The set of BasicBlocks that are still unvisited.
+ std::set<const BasicBlock*> Unvisited;
+
+ // The set of return edges (Edges with no successors).
+ std::set<Edge> ReturnEdges;
+ double ReturnWeight = 0;
+
+ // First iterate over the whole function and collect:
+ // 1) The blocks in this function in the Unvisited set.
+ // 2) The return edges in the ReturnEdges set.
+ // 3) The flow that is leaving the function already via return edges.
+
+ // Data structure for searching the function.
+ std::queue<const BasicBlock *> BFS;
+ const BasicBlock *BB = &(F->getEntryBlock());
+ BFS.push(BB);
+ Unvisited.insert(BB);
+
+ while (BFS.size()) {
+ BB = BFS.front(); BFS.pop();
+ succ_const_iterator NBB = succ_begin(BB), End = succ_end(BB);
+ if (NBB == End) {
+ Edge e = getEdge(BB,0);
+ double w = getEdgeWeight(e);
+ if (w == MissingValue) {
+ // If the return edge has no value, try to read value from block.
+ double bw = getExecutionCount(BB);
+ if (bw != MissingValue) {
+ setEdgeWeight(e,bw);
+ ReturnWeight += bw;
+ } else {
+ // If both return edge and block provide no value, collect edge.
+ ReturnEdges.insert(e);
+ }
+ } else {
+ // If the return edge has a proper value, collect it.
+ ReturnWeight += w;
+ }
+ }
+ for (;NBB != End; ++NBB) {
+ if (Unvisited.insert(*NBB).second) {
+ BFS.push(*NBB);
+ }
+ }
+ }
+
+ while (Unvisited.size() > 0) {
+ unsigned oldUnvisitedCount = Unvisited.size();
+ bool FoundPath = false;
+
+ // If there is only one edge left, calculate it.
+ if (ReturnEdges.size() == 1) {
+ ReturnWeight = getExecutionCount(&(F->getEntryBlock())) - ReturnWeight;
+
+ Edge e = *ReturnEdges.begin();
+ setEdgeWeight(e,ReturnWeight);
+ setExecutionCount(e.first,ReturnWeight);
+
+ Unvisited.erase(e.first);
+ ReturnEdges.erase(e);
+ continue;
+ }
+
+ // Calculate all blocks where only one edge is missing, this may also
+ // resolve furhter return edges.
+ std::set<const BasicBlock *>::iterator FI = Unvisited.begin(), FE = Unvisited.end();
+ while(FI != FE) {
+ const BasicBlock *BB = *FI; ++FI;
+ Edge e;
+ if(CalculateMissingEdge(BB,e,true)) {
+ if (BlockInformation[F].find(BB) == BlockInformation[F].end()) {
+ setExecutionCount(BB,getExecutionCount(BB));
+ }
+ Unvisited.erase(BB);
+ if (e.first != 0 && e.second == 0) {
+ ReturnEdges.erase(e);
+ ReturnWeight += getEdgeWeight(e);
+ }
+ }
+ }
+ if (oldUnvisitedCount > Unvisited.size()) continue;
+
+ // Estimate edge weights by dividing the flow proportionally.
+ FI = Unvisited.begin(), FE = Unvisited.end();
+ while(FI != FE) {
+ const BasicBlock *BB = *FI; ++FI;
+ const BasicBlock *Dest = 0;
+ bool AllEdgesHaveSameReturn = true;
+ // Check each Successor, these must all end up in the same or an empty
+ // return block otherwise its dangerous to do an estimation on them.
+ for (succ_const_iterator Succ = succ_begin(BB), End = succ_end(BB);
+ Succ != End; ++Succ) {
+ Path P;
+ GetPath(*Succ, 0, P, GetPathToExit);
+ if (Dest && Dest != P[0]) {
+ AllEdgesHaveSameReturn = false;
+ }
+ Dest = P[0];
+ }
+ if (AllEdgesHaveSameReturn) {
+ if(EstimateMissingEdges(BB)) {
+ Unvisited.erase(BB);
+ break;
+ }
+ }
+ }
+ if (oldUnvisitedCount > Unvisited.size()) continue;
+
+ // Check if there is a path to an block that has a known value and redirect
+ // flow accordingly.
+ FI = Unvisited.begin(), FE = Unvisited.end();
+ while(FI != FE && !FoundPath) {
+ // Fetch path.
+ const BasicBlock *BB = *FI; ++FI;
+ Path P;
+ const BasicBlock *Dest = GetPath(BB, 0, P, GetPathToValue);
+
+ // Calculate incoming flow.
+ double iw = 0; unsigned inmissing = 0; unsigned incount = 0; unsigned invalid = 0;
+ std::set<const BasicBlock *> Processed;
+ for (pred_const_iterator NBB = pred_begin(BB), End = pred_end(BB);
+ NBB != End; ++NBB) {
+ if (Processed.insert(*NBB).second) {
+ Edge e = getEdge(*NBB, BB);
+ double ew = getEdgeWeight(e);
+ if (ew != MissingValue) {
+ iw += ew;
+ invalid++;
+ } else {
+ // If the path contains the successor, this means its a backedge,
+ // do not count as missing.
+ if (P.find(*NBB) == P.end())
+ inmissing++;
+ }
+ incount++;
+ }
+ }
+ if (inmissing == incount) continue;
+ if (invalid == 0) continue;
+
+ // Subtract (already) outgoing flow.
+ Processed.clear();
+ for (succ_const_iterator NBB = succ_begin(BB), End = succ_end(BB);
+ NBB != End; ++NBB) {
+ if (Processed.insert(*NBB).second) {
+ Edge e = getEdge(BB, *NBB);
+ double ew = getEdgeWeight(e);
+ if (ew != MissingValue) {
+ iw -= ew;
+ }
+ }
+ }
+ if (iw < 0) continue;
+
+ // Check the recieving end of the path if it can handle the flow.
+ double ow = getExecutionCount(Dest);
+ Processed.clear();
+ for (succ_const_iterator NBB = succ_begin(BB), End = succ_end(BB);
+ NBB != End; ++NBB) {
+ if (Processed.insert(*NBB).second) {
+ Edge e = getEdge(BB, *NBB);
+ double ew = getEdgeWeight(e);
+ if (ew != MissingValue) {
+ ow -= ew;
+ }
+ }
+ }
+ if (ow < 0) continue;
+
+ // Determine how much flow shall be used.
+ double ew = getEdgeWeight(getEdge(P[Dest],Dest));
+ if (ew != MissingValue) {
+ ew = ew<ow?ew:ow;
+ ew = ew<iw?ew:iw;
+ } else {
+ if (inmissing == 0)
+ ew = iw;
+ }
+
+ // Create flow.
+ if (ew != MissingValue) {
+ do {
+ Edge e = getEdge(P[Dest],Dest);
+ if (getEdgeWeight(e) == MissingValue) {
+ setEdgeWeight(e,ew);
+ FoundPath = true;
+ }
+ Dest = P[Dest];
+ } while (Dest != BB);
+ }
+ }
+ if (FoundPath) continue;
+
+ // Calculate a block with self loop.
+ FI = Unvisited.begin(), FE = Unvisited.end();
+ while(FI != FE && !FoundPath) {
+ const BasicBlock *BB = *FI; ++FI;
+ bool SelfEdgeFound = false;
+ for (succ_const_iterator NBB = succ_begin(BB), End = succ_end(BB);
+ NBB != End; ++NBB) {
+ if (*NBB == BB) {
+ SelfEdgeFound = true;
+ break;
+ }
+ }
+ if (SelfEdgeFound) {
+ Edge e = getEdge(BB,BB);
+ if (getEdgeWeight(e) == MissingValue) {
+ double iw = 0;
+ std::set<const BasicBlock *> Processed;
+ for (pred_const_iterator NBB = pred_begin(BB), End = pred_end(BB);
+ NBB != End; ++NBB) {
+ if (Processed.insert(*NBB).second) {
+ Edge e = getEdge(*NBB, BB);
+ double ew = getEdgeWeight(e);
+ if (ew != MissingValue) {
+ iw += ew;
+ }
+ }
+ }
+ setEdgeWeight(e,iw * 10);
+ FoundPath = true;
+ }
+ }
+ }
+ if (FoundPath) continue;
+
+ // Determine backedges, set them to zero.
+ FI = Unvisited.begin(), FE = Unvisited.end();
+ while(FI != FE && !FoundPath) {
+ const BasicBlock *BB = *FI; ++FI;
+ const BasicBlock *Dest;
+ Path P;
+ bool BackEdgeFound = false;
+ for (pred_const_iterator NBB = pred_begin(BB), End = pred_end(BB);
+ NBB != End; ++NBB) {
+ Dest = GetPath(BB, *NBB, P, GetPathToDest | GetPathWithNewEdges);
+ if (Dest == *NBB) {
+ BackEdgeFound = true;
+ break;
+ }
+ }
+ if (BackEdgeFound) {
+ Edge e = getEdge(Dest,BB);
+ double w = getEdgeWeight(e);
+ if (w == MissingValue) {
+ setEdgeWeight(e,0);
+ FoundPath = true;
+ }
+ do {
+ Edge e = getEdge(P[Dest], Dest);
+ double w = getEdgeWeight(e);
+ if (w == MissingValue) {
+ setEdgeWeight(e,0);
+ FoundPath = true;
+ }
+ Dest = P[Dest];
+ } while (Dest != BB);
+ }
+ }
+ if (FoundPath) continue;
+
+ // Channel flow to return block.
+ FI = Unvisited.begin(), FE = Unvisited.end();
+ while(FI != FE && !FoundPath) {
+ const BasicBlock *BB = *FI; ++FI;
+
+ Path P;
+ const BasicBlock *Dest = GetPath(BB, 0, P, GetPathToExit | GetPathWithNewEdges);
+ Dest = P[0];
+ if (!Dest) continue;
+
+ if (getEdgeWeight(getEdge(Dest,0)) == MissingValue) {
+ // Calculate incoming flow.
+ double iw = 0;
+ std::set<const BasicBlock *> Processed;
+ for (pred_const_iterator NBB = pred_begin(BB), End = pred_end(BB);
+ NBB != End; ++NBB) {
+ if (Processed.insert(*NBB).second) {
+ Edge e = getEdge(*NBB, BB);
+ double ew = getEdgeWeight(e);
+ if (ew != MissingValue) {
+ iw += ew;
+ }
+ }
+ }
+ do {
+ Edge e = getEdge(P[Dest], Dest);
+ double w = getEdgeWeight(e);
+ if (w == MissingValue) {
+ setEdgeWeight(e,iw);
+ FoundPath = true;
+ } else {
+ assert(0 && "Edge should not have value already!");
+ }
+ Dest = P[Dest];
+ } while (Dest != BB);
+ }
+ }
+ if (FoundPath) continue;
+
+ // Speculatively set edges to zero.
+ FI = Unvisited.begin(), FE = Unvisited.end();
+ while(FI != FE && !FoundPath) {
+ const BasicBlock *BB = *FI; ++FI;
+
+ for (pred_const_iterator NBB = pred_begin(BB), End = pred_end(BB);
+ NBB != End; ++NBB) {
+ Edge e = getEdge(*NBB,BB);
+ double w = getEdgeWeight(e);
+ if (w == MissingValue) {
+ setEdgeWeight(e,0);
+ FoundPath = true;
+ break;
+ }
+ }
+ }
+ if (FoundPath) continue;
+
+ errs() << "{";
+ FI = Unvisited.begin(), FE = Unvisited.end();
+ while(FI != FE) {
+ const BasicBlock *BB = *FI; ++FI;
+ dbgs() << BB->getName();
+ if (FI != FE)
+ dbgs() << ",";
+ }
+ errs() << "}";
+
+ errs() << "ASSERT: could not repair function";
+ assert(0 && "could not repair function");
+ }
+
+ EdgeWeights J = EdgeInformation[F];
+ for (EdgeWeights::iterator EI = J.begin(), EE = J.end(); EI != EE; ++EI) {
+ Edge e = EI->first;
+
+ bool SuccFound = false;
+ if (e.first != 0) {
+ succ_const_iterator NBB = succ_begin(e.first), End = succ_end(e.first);
+ if (NBB == End) {
+ if (0 == e.second) {
+ SuccFound = true;
+ }
+ }
+ for (;NBB != End; ++NBB) {
+ if (*NBB == e.second) {
+ SuccFound = true;
+ break;
+ }
+ }
+ if (!SuccFound) {
+ removeEdge(e);
+ }
+ }
+ }
+}
+
+raw_ostream& operator<<(raw_ostream &O, const Function *F) {
+ return O << F->getName();
+}
+
+raw_ostream& operator<<(raw_ostream &O, const MachineFunction *MF) {
+ return O << MF->getFunction()->getName() << "(MF)";
+}
+
+raw_ostream& operator<<(raw_ostream &O, const BasicBlock *BB) {
+ return O << BB->getName();
+}
+
+raw_ostream& operator<<(raw_ostream &O, const MachineBasicBlock *MBB) {
+ return O << MBB->getBasicBlock()->getName() << "(MB)";
+}
+
+raw_ostream& operator<<(raw_ostream &O, std::pair<const BasicBlock *, const BasicBlock *> E) {
+ O << "(";
+
+ if (E.first)
+ O << E.first;
+ else
+ O << "0";
+
+ O << ",";
+
+ if (E.second)
+ O << E.second;
+ else
+ O << "0";
+
+ return O << ")";
+}
+
+raw_ostream& operator<<(raw_ostream &O, std::pair<const MachineBasicBlock *, const MachineBasicBlock *> E) {
+ O << "(";
+
+ if (E.first)
+ O << E.first;
+ else
+ O << "0";
+
+ O << ",";
+
+ if (E.second)
+ O << E.second;
+ else
+ O << "0";
+
+ return O << ")";
+}
+
+} // namespace llvm
+
+//===----------------------------------------------------------------------===//
+// NoProfile ProfileInfo implementation
+//
+
+namespace {
+ struct NoProfileInfo : public ImmutablePass, public ProfileInfo {
+ static char ID; // Class identification, replacement for typeinfo
+ NoProfileInfo() : ImmutablePass(&ID) {}
+
+ /// getAdjustedAnalysisPointer - This method is used when a pass implements
+ /// an analysis interface through multiple inheritance. If needed, it
+ /// should override this to adjust the this pointer as needed for the
+ /// specified pass info.
+ virtual void *getAdjustedAnalysisPointer(const PassInfo *PI) {
+ if (PI->isPassID(&ProfileInfo::ID))
+ return (ProfileInfo*)this;
+ return this;
+ }
+
+ virtual const char *getPassName() const {
+ return "NoProfileInfo";
+ }
+ };
+} // End of anonymous namespace
+
+char NoProfileInfo::ID = 0;
+// Register this pass...
+static RegisterPass<NoProfileInfo>
+X("no-profile", "No Profile Information", false, true);
+
+// Declare that we implement the ProfileInfo interface
+static RegisterAnalysisGroup<ProfileInfo, true> Y(X);
+
+ImmutablePass *llvm::createNoProfileInfoPass() { return new NoProfileInfo(); }
diff --git a/lib/Analysis/ProfileInfoLoader.cpp b/lib/Analysis/ProfileInfoLoader.cpp
new file mode 100644
index 0000000..25481b2
--- /dev/null
+++ b/lib/Analysis/ProfileInfoLoader.cpp
@@ -0,0 +1,158 @@
+//===- ProfileInfoLoad.cpp - Load profile information from disk -----------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// The ProfileInfoLoader class is used to load and represent profiling
+// information read in from the dump file.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/ProfileInfoLoader.h"
+#include "llvm/Analysis/ProfileInfoTypes.h"
+#include "llvm/Module.h"
+#include "llvm/InstrTypes.h"
+#include "llvm/Support/raw_ostream.h"
+#include <cstdio>
+#include <cstdlib>
+#include <map>
+using namespace llvm;
+
+// ByteSwap - Byteswap 'Var' if 'Really' is true.
+//
+static inline unsigned ByteSwap(unsigned Var, bool Really) {
+ if (!Really) return Var;
+ return ((Var & (255U<< 0U)) << 24U) |
+ ((Var & (255U<< 8U)) << 8U) |
+ ((Var & (255U<<16U)) >> 8U) |
+ ((Var & (255U<<24U)) >> 24U);
+}
+
+static unsigned AddCounts(unsigned A, unsigned B) {
+ // If either value is undefined, use the other.
+ if (A == ProfileInfoLoader::Uncounted) return B;
+ if (B == ProfileInfoLoader::Uncounted) return A;
+ return A + B;
+}
+
+static void ReadProfilingBlock(const char *ToolName, FILE *F,
+ bool ShouldByteSwap,
+ std::vector<unsigned> &Data) {
+ // Read the number of entries...
+ unsigned NumEntries;
+ if (fread(&NumEntries, sizeof(unsigned), 1, F) != 1) {
+ errs() << ToolName << ": data packet truncated!\n";
+ perror(0);
+ exit(1);
+ }
+ NumEntries = ByteSwap(NumEntries, ShouldByteSwap);
+
+ // Read the counts...
+ std::vector<unsigned> TempSpace(NumEntries);
+
+ // Read in the block of data...
+ if (fread(&TempSpace[0], sizeof(unsigned)*NumEntries, 1, F) != 1) {
+ errs() << ToolName << ": data packet truncated!\n";
+ perror(0);
+ exit(1);
+ }
+
+ // Make sure we have enough space... The space is initialised to -1 to
+ // facitiltate the loading of missing values for OptimalEdgeProfiling.
+ if (Data.size() < NumEntries)
+ Data.resize(NumEntries, ProfileInfoLoader::Uncounted);
+
+ // Accumulate the data we just read into the data.
+ if (!ShouldByteSwap) {
+ for (unsigned i = 0; i != NumEntries; ++i) {
+ Data[i] = AddCounts(TempSpace[i], Data[i]);
+ }
+ } else {
+ for (unsigned i = 0; i != NumEntries; ++i) {
+ Data[i] = AddCounts(ByteSwap(TempSpace[i], true), Data[i]);
+ }
+ }
+}
+
+const unsigned ProfileInfoLoader::Uncounted = ~0U;
+
+// ProfileInfoLoader ctor - Read the specified profiling data file, exiting the
+// program if the file is invalid or broken.
+//
+ProfileInfoLoader::ProfileInfoLoader(const char *ToolName,
+ const std::string &Filename,
+ Module &TheModule) :
+ Filename(Filename),
+ M(TheModule), Warned(false) {
+ FILE *F = fopen(Filename.c_str(), "rb");
+ if (F == 0) {
+ errs() << ToolName << ": Error opening '" << Filename << "': ";
+ perror(0);
+ exit(1);
+ }
+
+ // Keep reading packets until we run out of them.
+ unsigned PacketType;
+ while (fread(&PacketType, sizeof(unsigned), 1, F) == 1) {
+ // If the low eight bits of the packet are zero, we must be dealing with an
+ // endianness mismatch. Byteswap all words read from the profiling
+ // information.
+ bool ShouldByteSwap = (char)PacketType == 0;
+ PacketType = ByteSwap(PacketType, ShouldByteSwap);
+
+ switch (PacketType) {
+ case ArgumentInfo: {
+ unsigned ArgLength;
+ if (fread(&ArgLength, sizeof(unsigned), 1, F) != 1) {
+ errs() << ToolName << ": arguments packet truncated!\n";
+ perror(0);
+ exit(1);
+ }
+ ArgLength = ByteSwap(ArgLength, ShouldByteSwap);
+
+ // Read in the arguments...
+ std::vector<char> Chars(ArgLength+4);
+
+ if (ArgLength)
+ if (fread(&Chars[0], (ArgLength+3) & ~3, 1, F) != 1) {
+ errs() << ToolName << ": arguments packet truncated!\n";
+ perror(0);
+ exit(1);
+ }
+ CommandLines.push_back(std::string(&Chars[0], &Chars[ArgLength]));
+ break;
+ }
+
+ case FunctionInfo:
+ ReadProfilingBlock(ToolName, F, ShouldByteSwap, FunctionCounts);
+ break;
+
+ case BlockInfo:
+ ReadProfilingBlock(ToolName, F, ShouldByteSwap, BlockCounts);
+ break;
+
+ case EdgeInfo:
+ ReadProfilingBlock(ToolName, F, ShouldByteSwap, EdgeCounts);
+ break;
+
+ case OptEdgeInfo:
+ ReadProfilingBlock(ToolName, F, ShouldByteSwap, OptimalEdgeCounts);
+ break;
+
+ case BBTraceInfo:
+ ReadProfilingBlock(ToolName, F, ShouldByteSwap, BBTrace);
+ break;
+
+ default:
+ errs() << ToolName << ": Unknown packet type #" << PacketType << "!\n";
+ exit(1);
+ }
+ }
+
+ fclose(F);
+}
+
diff --git a/lib/Analysis/ProfileInfoLoaderPass.cpp b/lib/Analysis/ProfileInfoLoaderPass.cpp
new file mode 100644
index 0000000..ac9ed52
--- /dev/null
+++ b/lib/Analysis/ProfileInfoLoaderPass.cpp
@@ -0,0 +1,268 @@
+//===- ProfileInfoLoaderPass.cpp - LLVM Pass to load profile info ---------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements a concrete implementation of profiling information that
+// loads the information from a profile dump file.
+//
+//===----------------------------------------------------------------------===//
+#define DEBUG_TYPE "profile-loader"
+#include "llvm/BasicBlock.h"
+#include "llvm/InstrTypes.h"
+#include "llvm/Module.h"
+#include "llvm/Pass.h"
+#include "llvm/Analysis/Passes.h"
+#include "llvm/Analysis/ProfileInfo.h"
+#include "llvm/Analysis/ProfileInfoLoader.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/CFG.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/Support/Format.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/ADT/SmallSet.h"
+#include <set>
+using namespace llvm;
+
+STATISTIC(NumEdgesRead, "The # of edges read.");
+
+static cl::opt<std::string>
+ProfileInfoFilename("profile-info-file", cl::init("llvmprof.out"),
+ cl::value_desc("filename"),
+ cl::desc("Profile file loaded by -profile-loader"));
+
+namespace {
+ class LoaderPass : public ModulePass, public ProfileInfo {
+ std::string Filename;
+ std::set<Edge> SpanningTree;
+ std::set<const BasicBlock*> BBisUnvisited;
+ unsigned ReadCount;
+ public:
+ static char ID; // Class identification, replacement for typeinfo
+ explicit LoaderPass(const std::string &filename = "")
+ : ModulePass(&ID), Filename(filename) {
+ if (filename.empty()) Filename = ProfileInfoFilename;
+ }
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.setPreservesAll();
+ }
+
+ virtual const char *getPassName() const {
+ return "Profiling information loader";
+ }
+
+ // recurseBasicBlock() - Calculates the edge weights for as much basic
+ // blocks as possbile.
+ virtual void recurseBasicBlock(const BasicBlock *BB);
+ virtual void readEdgeOrRemember(Edge, Edge&, unsigned &, double &);
+ virtual void readEdge(ProfileInfo::Edge, std::vector<unsigned>&);
+
+ /// getAdjustedAnalysisPointer - This method is used when a pass implements
+ /// an analysis interface through multiple inheritance. If needed, it
+ /// should override this to adjust the this pointer as needed for the
+ /// specified pass info.
+ virtual void *getAdjustedAnalysisPointer(const PassInfo *PI) {
+ if (PI->isPassID(&ProfileInfo::ID))
+ return (ProfileInfo*)this;
+ return this;
+ }
+
+ /// run - Load the profile information from the specified file.
+ virtual bool runOnModule(Module &M);
+ };
+} // End of anonymous namespace
+
+char LoaderPass::ID = 0;
+static RegisterPass<LoaderPass>
+X("profile-loader", "Load profile information from llvmprof.out", false, true);
+
+static RegisterAnalysisGroup<ProfileInfo> Y(X);
+
+const PassInfo *llvm::ProfileLoaderPassID = &X;
+
+ModulePass *llvm::createProfileLoaderPass() { return new LoaderPass(); }
+
+/// createProfileLoaderPass - This function returns a Pass that loads the
+/// profiling information for the module from the specified filename, making it
+/// available to the optimizers.
+Pass *llvm::createProfileLoaderPass(const std::string &Filename) {
+ return new LoaderPass(Filename);
+}
+
+void LoaderPass::readEdgeOrRemember(Edge edge, Edge &tocalc,
+ unsigned &uncalc, double &count) {
+ double w;
+ if ((w = getEdgeWeight(edge)) == MissingValue) {
+ tocalc = edge;
+ uncalc++;
+ } else {
+ count+=w;
+ }
+}
+
+// recurseBasicBlock - Visits all neighbours of a block and then tries to
+// calculate the missing edge values.
+void LoaderPass::recurseBasicBlock(const BasicBlock *BB) {
+
+ // break recursion if already visited
+ if (BBisUnvisited.find(BB) == BBisUnvisited.end()) return;
+ BBisUnvisited.erase(BB);
+ if (!BB) return;
+
+ for (succ_const_iterator bbi = succ_begin(BB), bbe = succ_end(BB);
+ bbi != bbe; ++bbi) {
+ recurseBasicBlock(*bbi);
+ }
+ for (pred_const_iterator bbi = pred_begin(BB), bbe = pred_end(BB);
+ bbi != bbe; ++bbi) {
+ recurseBasicBlock(*bbi);
+ }
+
+ Edge tocalc;
+ if (CalculateMissingEdge(BB, tocalc)) {
+ SpanningTree.erase(tocalc);
+ }
+}
+
+void LoaderPass::readEdge(ProfileInfo::Edge e,
+ std::vector<unsigned> &ECs) {
+ if (ReadCount < ECs.size()) {
+ double weight = ECs[ReadCount++];
+ if (weight != ProfileInfoLoader::Uncounted) {
+ // Here the data realm changes from the unsigned of the file to the
+ // double of the ProfileInfo. This conversion is save because we know
+ // that everything thats representable in unsinged is also representable
+ // in double.
+ EdgeInformation[getFunction(e)][e] += (double)weight;
+
+ DEBUG(dbgs() << "--Read Edge Counter for " << e
+ << " (# "<< (ReadCount-1) << "): "
+ << (unsigned)getEdgeWeight(e) << "\n");
+ } else {
+ // This happens only if reading optimal profiling information, not when
+ // reading regular profiling information.
+ SpanningTree.insert(e);
+ }
+ }
+}
+
+bool LoaderPass::runOnModule(Module &M) {
+ ProfileInfoLoader PIL("profile-loader", Filename, M);
+
+ EdgeInformation.clear();
+ std::vector<unsigned> Counters = PIL.getRawEdgeCounts();
+ if (Counters.size() > 0) {
+ ReadCount = 0;
+ for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
+ if (F->isDeclaration()) continue;
+ DEBUG(dbgs()<<"Working on "<<F->getNameStr()<<"\n");
+ readEdge(getEdge(0,&F->getEntryBlock()), Counters);
+ for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
+ TerminatorInst *TI = BB->getTerminator();
+ for (unsigned s = 0, e = TI->getNumSuccessors(); s != e; ++s) {
+ readEdge(getEdge(BB,TI->getSuccessor(s)), Counters);
+ }
+ }
+ }
+ if (ReadCount != Counters.size()) {
+ errs() << "WARNING: profile information is inconsistent with "
+ << "the current program!\n";
+ }
+ NumEdgesRead = ReadCount;
+ }
+
+ Counters = PIL.getRawOptimalEdgeCounts();
+ if (Counters.size() > 0) {
+ ReadCount = 0;
+ for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
+ if (F->isDeclaration()) continue;
+ DEBUG(dbgs()<<"Working on "<<F->getNameStr()<<"\n");
+ readEdge(getEdge(0,&F->getEntryBlock()), Counters);
+ for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
+ TerminatorInst *TI = BB->getTerminator();
+ if (TI->getNumSuccessors() == 0) {
+ readEdge(getEdge(BB,0), Counters);
+ }
+ for (unsigned s = 0, e = TI->getNumSuccessors(); s != e; ++s) {
+ readEdge(getEdge(BB,TI->getSuccessor(s)), Counters);
+ }
+ }
+ while (SpanningTree.size() > 0) {
+
+ unsigned size = SpanningTree.size();
+
+ BBisUnvisited.clear();
+ for (std::set<Edge>::iterator ei = SpanningTree.begin(),
+ ee = SpanningTree.end(); ei != ee; ++ei) {
+ BBisUnvisited.insert(ei->first);
+ BBisUnvisited.insert(ei->second);
+ }
+ while (BBisUnvisited.size() > 0) {
+ recurseBasicBlock(*BBisUnvisited.begin());
+ }
+
+ if (SpanningTree.size() == size) {
+ DEBUG(dbgs()<<"{");
+ for (std::set<Edge>::iterator ei = SpanningTree.begin(),
+ ee = SpanningTree.end(); ei != ee; ++ei) {
+ DEBUG(dbgs()<< *ei <<",");
+ }
+ assert(0 && "No edge calculated!");
+ }
+
+ }
+ }
+ if (ReadCount != Counters.size()) {
+ errs() << "WARNING: profile information is inconsistent with "
+ << "the current program!\n";
+ }
+ NumEdgesRead = ReadCount;
+ }
+
+ BlockInformation.clear();
+ Counters = PIL.getRawBlockCounts();
+ if (Counters.size() > 0) {
+ ReadCount = 0;
+ for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
+ if (F->isDeclaration()) continue;
+ for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
+ if (ReadCount < Counters.size())
+ // Here the data realm changes from the unsigned of the file to the
+ // double of the ProfileInfo. This conversion is save because we know
+ // that everything thats representable in unsinged is also
+ // representable in double.
+ BlockInformation[F][BB] = (double)Counters[ReadCount++];
+ }
+ if (ReadCount != Counters.size()) {
+ errs() << "WARNING: profile information is inconsistent with "
+ << "the current program!\n";
+ }
+ }
+
+ FunctionInformation.clear();
+ Counters = PIL.getRawFunctionCounts();
+ if (Counters.size() > 0) {
+ ReadCount = 0;
+ for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
+ if (F->isDeclaration()) continue;
+ if (ReadCount < Counters.size())
+ // Here the data realm changes from the unsigned of the file to the
+ // double of the ProfileInfo. This conversion is save because we know
+ // that everything thats representable in unsinged is also
+ // representable in double.
+ FunctionInformation[F] = (double)Counters[ReadCount++];
+ }
+ if (ReadCount != Counters.size()) {
+ errs() << "WARNING: profile information is inconsistent with "
+ << "the current program!\n";
+ }
+ }
+
+ return false;
+}
diff --git a/lib/Analysis/ProfileVerifierPass.cpp b/lib/Analysis/ProfileVerifierPass.cpp
new file mode 100644
index 0000000..a2ddc8e
--- /dev/null
+++ b/lib/Analysis/ProfileVerifierPass.cpp
@@ -0,0 +1,377 @@
+//===- ProfileVerifierPass.cpp - LLVM Pass to estimate profile info -------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements a pass that checks profiling information for
+// plausibility.
+//
+//===----------------------------------------------------------------------===//
+#define DEBUG_TYPE "profile-verifier"
+#include "llvm/Instructions.h"
+#include "llvm/Module.h"
+#include "llvm/Pass.h"
+#include "llvm/Analysis/ProfileInfo.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/CallSite.h"
+#include "llvm/Support/CFG.h"
+#include "llvm/Support/InstIterator.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/Support/Format.h"
+#include "llvm/Support/Debug.h"
+#include <set>
+using namespace llvm;
+
+static cl::opt<bool,false>
+ProfileVerifierDisableAssertions("profile-verifier-noassert",
+ cl::desc("Disable assertions"));
+
+namespace llvm {
+ template<class FType, class BType>
+ class ProfileVerifierPassT : public FunctionPass {
+
+ struct DetailedBlockInfo {
+ const BType *BB;
+ double BBWeight;
+ double inWeight;
+ int inCount;
+ double outWeight;
+ int outCount;
+ };
+
+ ProfileInfoT<FType, BType> *PI;
+ std::set<const BType*> BBisVisited;
+ std::set<const FType*> FisVisited;
+ bool DisableAssertions;
+
+ // When debugging is enabled, the verifier prints a whole slew of debug
+ // information, otherwise its just the assert. These are all the helper
+ // functions.
+ bool PrintedDebugTree;
+ std::set<const BType*> BBisPrinted;
+ void debugEntry(DetailedBlockInfo*);
+ void printDebugInfo(const BType *BB);
+
+ public:
+ static char ID; // Class identification, replacement for typeinfo
+
+ explicit ProfileVerifierPassT () : FunctionPass(&ID) {
+ DisableAssertions = ProfileVerifierDisableAssertions;
+ }
+ explicit ProfileVerifierPassT (bool da) : FunctionPass(&ID),
+ DisableAssertions(da) {
+ }
+
+ void getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.setPreservesAll();
+ AU.addRequired<ProfileInfoT<FType, BType> >();
+ }
+
+ const char *getPassName() const {
+ return "Profiling information verifier";
+ }
+
+ /// run - Verify the profile information.
+ bool runOnFunction(FType &F);
+ void recurseBasicBlock(const BType*);
+
+ bool exitReachable(const FType*);
+ double ReadOrAssert(typename ProfileInfoT<FType, BType>::Edge);
+ void CheckValue(bool, const char*, DetailedBlockInfo*);
+ };
+
+ typedef ProfileVerifierPassT<Function, BasicBlock> ProfileVerifierPass;
+
+ template<class FType, class BType>
+ void ProfileVerifierPassT<FType, BType>::printDebugInfo(const BType *BB) {
+
+ if (BBisPrinted.find(BB) != BBisPrinted.end()) return;
+
+ double BBWeight = PI->getExecutionCount(BB);
+ if (BBWeight == ProfileInfoT<FType, BType>::MissingValue) { BBWeight = 0; }
+ double inWeight = 0;
+ int inCount = 0;
+ std::set<const BType*> ProcessedPreds;
+ for ( pred_const_iterator bbi = pred_begin(BB), bbe = pred_end(BB);
+ bbi != bbe; ++bbi ) {
+ if (ProcessedPreds.insert(*bbi).second) {
+ typename ProfileInfoT<FType, BType>::Edge E = PI->getEdge(*bbi,BB);
+ double EdgeWeight = PI->getEdgeWeight(E);
+ if (EdgeWeight == ProfileInfoT<FType, BType>::MissingValue) { EdgeWeight = 0; }
+ dbgs() << "calculated in-edge " << E << ": "
+ << format("%20.20g",EdgeWeight) << "\n";
+ inWeight += EdgeWeight;
+ inCount++;
+ }
+ }
+ double outWeight = 0;
+ int outCount = 0;
+ std::set<const BType*> ProcessedSuccs;
+ for ( succ_const_iterator bbi = succ_begin(BB), bbe = succ_end(BB);
+ bbi != bbe; ++bbi ) {
+ if (ProcessedSuccs.insert(*bbi).second) {
+ typename ProfileInfoT<FType, BType>::Edge E = PI->getEdge(BB,*bbi);
+ double EdgeWeight = PI->getEdgeWeight(E);
+ if (EdgeWeight == ProfileInfoT<FType, BType>::MissingValue) { EdgeWeight = 0; }
+ dbgs() << "calculated out-edge " << E << ": "
+ << format("%20.20g",EdgeWeight) << "\n";
+ outWeight += EdgeWeight;
+ outCount++;
+ }
+ }
+ dbgs() << "Block " << BB->getNameStr() << " in "
+ << BB->getParent()->getNameStr() << ":"
+ << "BBWeight=" << format("%20.20g",BBWeight) << ","
+ << "inWeight=" << format("%20.20g",inWeight) << ","
+ << "inCount=" << inCount << ","
+ << "outWeight=" << format("%20.20g",outWeight) << ","
+ << "outCount" << outCount << "\n";
+
+ // mark as visited and recurse into subnodes
+ BBisPrinted.insert(BB);
+ for ( succ_const_iterator bbi = succ_begin(BB), bbe = succ_end(BB);
+ bbi != bbe; ++bbi ) {
+ printDebugInfo(*bbi);
+ }
+ }
+
+ template<class FType, class BType>
+ void ProfileVerifierPassT<FType, BType>::debugEntry (DetailedBlockInfo *DI) {
+ dbgs() << "TROUBLE: Block " << DI->BB->getNameStr() << " in "
+ << DI->BB->getParent()->getNameStr() << ":"
+ << "BBWeight=" << format("%20.20g",DI->BBWeight) << ","
+ << "inWeight=" << format("%20.20g",DI->inWeight) << ","
+ << "inCount=" << DI->inCount << ","
+ << "outWeight=" << format("%20.20g",DI->outWeight) << ","
+ << "outCount=" << DI->outCount << "\n";
+ if (!PrintedDebugTree) {
+ PrintedDebugTree = true;
+ printDebugInfo(&(DI->BB->getParent()->getEntryBlock()));
+ }
+ }
+
+ // This compares A and B for equality.
+ static bool Equals(double A, double B) {
+ return A == B;
+ }
+
+ // This checks if the function "exit" is reachable from an given function
+ // via calls, this is necessary to check if a profile is valid despite the
+ // counts not fitting exactly.
+ template<class FType, class BType>
+ bool ProfileVerifierPassT<FType, BType>::exitReachable(const FType *F) {
+ if (!F) return false;
+
+ if (FisVisited.count(F)) return false;
+
+ FType *Exit = F->getParent()->getFunction("exit");
+ if (Exit == F) {
+ return true;
+ }
+
+ FisVisited.insert(F);
+ bool exits = false;
+ for (const_inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I) {
+ if (const CallInst *CI = dyn_cast<CallInst>(&*I)) {
+ FType *F = CI->getCalledFunction();
+ if (F) {
+ exits |= exitReachable(F);
+ } else {
+ // This is a call to a pointer, all bets are off...
+ exits = true;
+ }
+ if (exits) break;
+ }
+ }
+ return exits;
+ }
+
+ #define ASSERTMESSAGE(M) \
+ { dbgs() << "ASSERT:" << (M) << "\n"; \
+ if (!DisableAssertions) assert(0 && (M)); }
+
+ template<class FType, class BType>
+ double ProfileVerifierPassT<FType, BType>::ReadOrAssert(typename ProfileInfoT<FType, BType>::Edge E) {
+ double EdgeWeight = PI->getEdgeWeight(E);
+ if (EdgeWeight == ProfileInfoT<FType, BType>::MissingValue) {
+ dbgs() << "Edge " << E << " in Function "
+ << ProfileInfoT<FType, BType>::getFunction(E)->getNameStr() << ": ";
+ ASSERTMESSAGE("Edge has missing value");
+ return 0;
+ } else {
+ if (EdgeWeight < 0) {
+ dbgs() << "Edge " << E << " in Function "
+ << ProfileInfoT<FType, BType>::getFunction(E)->getNameStr() << ": ";
+ ASSERTMESSAGE("Edge has negative value");
+ }
+ return EdgeWeight;
+ }
+ }
+
+ template<class FType, class BType>
+ void ProfileVerifierPassT<FType, BType>::CheckValue(bool Error,
+ const char *Message,
+ DetailedBlockInfo *DI) {
+ if (Error) {
+ DEBUG(debugEntry(DI));
+ dbgs() << "Block " << DI->BB->getNameStr() << " in Function "
+ << DI->BB->getParent()->getNameStr() << ": ";
+ ASSERTMESSAGE(Message);
+ }
+ return;
+ }
+
+ // This calculates the Information for a block and then recurses into the
+ // successors.
+ template<class FType, class BType>
+ void ProfileVerifierPassT<FType, BType>::recurseBasicBlock(const BType *BB) {
+
+ // Break the recursion by remembering all visited blocks.
+ if (BBisVisited.find(BB) != BBisVisited.end()) return;
+
+ // Use a data structure to store all the information, this can then be handed
+ // to debug printers.
+ DetailedBlockInfo DI;
+ DI.BB = BB;
+ DI.outCount = DI.inCount = 0;
+ DI.inWeight = DI.outWeight = 0;
+
+ // Read predecessors.
+ std::set<const BType*> ProcessedPreds;
+ pred_const_iterator bpi = pred_begin(BB), bpe = pred_end(BB);
+ // If there are none, check for (0,BB) edge.
+ if (bpi == bpe) {
+ DI.inWeight += ReadOrAssert(PI->getEdge(0,BB));
+ DI.inCount++;
+ }
+ for (;bpi != bpe; ++bpi) {
+ if (ProcessedPreds.insert(*bpi).second) {
+ DI.inWeight += ReadOrAssert(PI->getEdge(*bpi,BB));
+ DI.inCount++;
+ }
+ }
+
+ // Read successors.
+ std::set<const BType*> ProcessedSuccs;
+ succ_const_iterator bbi = succ_begin(BB), bbe = succ_end(BB);
+ // If there is an (0,BB) edge, consider it too. (This is done not only when
+ // there are no successors, but every time; not every function contains
+ // return blocks with no successors (think loop latch as return block)).
+ double w = PI->getEdgeWeight(PI->getEdge(BB,0));
+ if (w != ProfileInfoT<FType, BType>::MissingValue) {
+ DI.outWeight += w;
+ DI.outCount++;
+ }
+ for (;bbi != bbe; ++bbi) {
+ if (ProcessedSuccs.insert(*bbi).second) {
+ DI.outWeight += ReadOrAssert(PI->getEdge(BB,*bbi));
+ DI.outCount++;
+ }
+ }
+
+ // Read block weight.
+ DI.BBWeight = PI->getExecutionCount(BB);
+ CheckValue(DI.BBWeight == ProfileInfoT<FType, BType>::MissingValue,
+ "BasicBlock has missing value", &DI);
+ CheckValue(DI.BBWeight < 0,
+ "BasicBlock has negative value", &DI);
+
+ // Check if this block is a setjmp target.
+ bool isSetJmpTarget = false;
+ if (DI.outWeight > DI.inWeight) {
+ for (typename BType::const_iterator i = BB->begin(), ie = BB->end();
+ i != ie; ++i) {
+ if (const CallInst *CI = dyn_cast<CallInst>(&*i)) {
+ FType *F = CI->getCalledFunction();
+ if (F && (F->getNameStr() == "_setjmp")) {
+ isSetJmpTarget = true; break;
+ }
+ }
+ }
+ }
+ // Check if this block is eventually reaching exit.
+ bool isExitReachable = false;
+ if (DI.inWeight > DI.outWeight) {
+ for (typename BType::const_iterator i = BB->begin(), ie = BB->end();
+ i != ie; ++i) {
+ if (const CallInst *CI = dyn_cast<CallInst>(&*i)) {
+ FType *F = CI->getCalledFunction();
+ if (F) {
+ FisVisited.clear();
+ isExitReachable |= exitReachable(F);
+ } else {
+ // This is a call to a pointer, all bets are off...
+ isExitReachable = true;
+ }
+ if (isExitReachable) break;
+ }
+ }
+ }
+
+ if (DI.inCount > 0 && DI.outCount == 0) {
+ // If this is a block with no successors.
+ if (!isSetJmpTarget) {
+ CheckValue(!Equals(DI.inWeight,DI.BBWeight),
+ "inWeight and BBWeight do not match", &DI);
+ }
+ } else if (DI.inCount == 0 && DI.outCount > 0) {
+ // If this is a block with no predecessors.
+ if (!isExitReachable)
+ CheckValue(!Equals(DI.BBWeight,DI.outWeight),
+ "BBWeight and outWeight do not match", &DI);
+ } else {
+ // If this block has successors and predecessors.
+ if (DI.inWeight > DI.outWeight && !isExitReachable)
+ CheckValue(!Equals(DI.inWeight,DI.outWeight),
+ "inWeight and outWeight do not match", &DI);
+ if (DI.inWeight < DI.outWeight && !isSetJmpTarget)
+ CheckValue(!Equals(DI.inWeight,DI.outWeight),
+ "inWeight and outWeight do not match", &DI);
+ }
+
+
+ // Mark this block as visited, rescurse into successors.
+ BBisVisited.insert(BB);
+ for ( succ_const_iterator bbi = succ_begin(BB), bbe = succ_end(BB);
+ bbi != bbe; ++bbi ) {
+ recurseBasicBlock(*bbi);
+ }
+ }
+
+ template<class FType, class BType>
+ bool ProfileVerifierPassT<FType, BType>::runOnFunction(FType &F) {
+ PI = getAnalysisIfAvailable<ProfileInfoT<FType, BType> >();
+ if (!PI)
+ ASSERTMESSAGE("No ProfileInfo available");
+
+ // Prepare global variables.
+ PrintedDebugTree = false;
+ BBisVisited.clear();
+
+ // Fetch entry block and recurse into it.
+ const BType *entry = &F.getEntryBlock();
+ recurseBasicBlock(entry);
+
+ if (PI->getExecutionCount(&F) != PI->getExecutionCount(entry))
+ ASSERTMESSAGE("Function count and entry block count do not match");
+
+ return false;
+ }
+
+ template<class FType, class BType>
+ char ProfileVerifierPassT<FType, BType>::ID = 0;
+}
+
+static RegisterPass<ProfileVerifierPass>
+X("profile-verifier", "Verify profiling information", false, true);
+
+namespace llvm {
+ FunctionPass *createProfileVerifierPass() {
+ return new ProfileVerifierPass(ProfileVerifierDisableAssertions);
+ }
+}
+
diff --git a/lib/Analysis/README.txt b/lib/Analysis/README.txt
new file mode 100644
index 0000000..c401090
--- /dev/null
+++ b/lib/Analysis/README.txt
@@ -0,0 +1,18 @@
+Analysis Opportunities:
+
+//===---------------------------------------------------------------------===//
+
+In test/Transforms/LoopStrengthReduce/quadradic-exit-value.ll, the
+ScalarEvolution expression for %r is this:
+
+ {1,+,3,+,2}<loop>
+
+Outside the loop, this could be evaluated simply as (%n * %n), however
+ScalarEvolution currently evaluates it as
+
+ (-2 + (2 * (trunc i65 (((zext i64 (-2 + %n) to i65) * (zext i64 (-1 + %n) to i65)) /u 2) to i64)) + (3 * %n))
+
+In addition to being much more complicated, it involves i65 arithmetic,
+which is very inefficient when expanded into code.
+
+//===---------------------------------------------------------------------===//
diff --git a/lib/Analysis/ScalarEvolution.cpp b/lib/Analysis/ScalarEvolution.cpp
new file mode 100644
index 0000000..82be9cd
--- /dev/null
+++ b/lib/Analysis/ScalarEvolution.cpp
@@ -0,0 +1,5412 @@
+//===- ScalarEvolution.cpp - Scalar Evolution Analysis ----------*- C++ -*-===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file contains the implementation of the scalar evolution analysis
+// engine, which is used primarily to analyze expressions involving induction
+// variables in loops.
+//
+// There are several aspects to this library. First is the representation of
+// scalar expressions, which are represented as subclasses of the SCEV class.
+// These classes are used to represent certain types of subexpressions that we
+// can handle. We only create one SCEV of a particular shape, so
+// pointer-comparisons for equality are legal.
+//
+// One important aspect of the SCEV objects is that they are never cyclic, even
+// if there is a cycle in the dataflow for an expression (ie, a PHI node). If
+// the PHI node is one of the idioms that we can represent (e.g., a polynomial
+// recurrence) then we represent it directly as a recurrence node, otherwise we
+// represent it as a SCEVUnknown node.
+//
+// In addition to being able to represent expressions of various types, we also
+// have folders that are used to build the *canonical* representation for a
+// particular expression. These folders are capable of using a variety of
+// rewrite rules to simplify the expressions.
+//
+// Once the folders are defined, we can implement the more interesting
+// higher-level code, such as the code that recognizes PHI nodes of various
+// types, computes the execution count of a loop, etc.
+//
+// TODO: We should use these routines and value representations to implement
+// dependence analysis!
+//
+//===----------------------------------------------------------------------===//
+//
+// There are several good references for the techniques used in this analysis.
+//
+// Chains of recurrences -- a method to expedite the evaluation
+// of closed-form functions
+// Olaf Bachmann, Paul S. Wang, Eugene V. Zima
+//
+// On computational properties of chains of recurrences
+// Eugene V. Zima
+//
+// Symbolic Evaluation of Chains of Recurrences for Loop Optimization
+// Robert A. van Engelen
+//
+// Efficient Symbolic Analysis for Optimizing Compilers
+// Robert A. van Engelen
+//
+// Using the chains of recurrences algebra for data dependence testing and
+// induction variable substitution
+// MS Thesis, Johnie Birch
+//
+//===----------------------------------------------------------------------===//
+
+#define DEBUG_TYPE "scalar-evolution"
+#include "llvm/Analysis/ScalarEvolutionExpressions.h"
+#include "llvm/Constants.h"
+#include "llvm/DerivedTypes.h"
+#include "llvm/GlobalVariable.h"
+#include "llvm/GlobalAlias.h"
+#include "llvm/Instructions.h"
+#include "llvm/LLVMContext.h"
+#include "llvm/Operator.h"
+#include "llvm/Analysis/ConstantFolding.h"
+#include "llvm/Analysis/Dominators.h"
+#include "llvm/Analysis/LoopInfo.h"
+#include "llvm/Analysis/ValueTracking.h"
+#include "llvm/Assembly/Writer.h"
+#include "llvm/Target/TargetData.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/ConstantRange.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/GetElementPtrTypeIterator.h"
+#include "llvm/Support/InstIterator.h"
+#include "llvm/Support/MathExtras.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/ADT/STLExtras.h"
+#include "llvm/ADT/SmallPtrSet.h"
+#include <algorithm>
+using namespace llvm;
+
+STATISTIC(NumArrayLenItCounts,
+ "Number of trip counts computed with array length");
+STATISTIC(NumTripCountsComputed,
+ "Number of loops with predictable loop counts");
+STATISTIC(NumTripCountsNotComputed,
+ "Number of loops without predictable loop counts");
+STATISTIC(NumBruteForceTripCountsComputed,
+ "Number of loops with trip counts computed by force");
+
+static cl::opt<unsigned>
+MaxBruteForceIterations("scalar-evolution-max-iterations", cl::ReallyHidden,
+ cl::desc("Maximum number of iterations SCEV will "
+ "symbolically execute a constant "
+ "derived loop"),
+ cl::init(100));
+
+static RegisterPass<ScalarEvolution>
+R("scalar-evolution", "Scalar Evolution Analysis", false, true);
+char ScalarEvolution::ID = 0;
+
+//===----------------------------------------------------------------------===//
+// SCEV class definitions
+//===----------------------------------------------------------------------===//
+
+//===----------------------------------------------------------------------===//
+// Implementation of the SCEV class.
+//
+
+SCEV::~SCEV() {}
+
+void SCEV::dump() const {
+ print(dbgs());
+ dbgs() << '\n';
+}
+
+bool SCEV::isZero() const {
+ if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
+ return SC->getValue()->isZero();
+ return false;
+}
+
+bool SCEV::isOne() const {
+ if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
+ return SC->getValue()->isOne();
+ return false;
+}
+
+bool SCEV::isAllOnesValue() const {
+ if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this))
+ return SC->getValue()->isAllOnesValue();
+ return false;
+}
+
+SCEVCouldNotCompute::SCEVCouldNotCompute() :
+ SCEV(FoldingSetNodeID(), scCouldNotCompute) {}
+
+bool SCEVCouldNotCompute::isLoopInvariant(const Loop *L) const {
+ llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
+ return false;
+}
+
+const Type *SCEVCouldNotCompute::getType() const {
+ llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
+ return 0;
+}
+
+bool SCEVCouldNotCompute::hasComputableLoopEvolution(const Loop *L) const {
+ llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
+ return false;
+}
+
+bool SCEVCouldNotCompute::hasOperand(const SCEV *) const {
+ llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
+ return false;
+}
+
+void SCEVCouldNotCompute::print(raw_ostream &OS) const {
+ OS << "***COULDNOTCOMPUTE***";
+}
+
+bool SCEVCouldNotCompute::classof(const SCEV *S) {
+ return S->getSCEVType() == scCouldNotCompute;
+}
+
+const SCEV *ScalarEvolution::getConstant(ConstantInt *V) {
+ FoldingSetNodeID ID;
+ ID.AddInteger(scConstant);
+ ID.AddPointer(V);
+ void *IP = 0;
+ if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
+ SCEV *S = SCEVAllocator.Allocate<SCEVConstant>();
+ new (S) SCEVConstant(ID, V);
+ UniqueSCEVs.InsertNode(S, IP);
+ return S;
+}
+
+const SCEV *ScalarEvolution::getConstant(const APInt& Val) {
+ return getConstant(ConstantInt::get(getContext(), Val));
+}
+
+const SCEV *
+ScalarEvolution::getConstant(const Type *Ty, uint64_t V, bool isSigned) {
+ return getConstant(
+ ConstantInt::get(cast<IntegerType>(Ty), V, isSigned));
+}
+
+const Type *SCEVConstant::getType() const { return V->getType(); }
+
+void SCEVConstant::print(raw_ostream &OS) const {
+ WriteAsOperand(OS, V, false);
+}
+
+SCEVCastExpr::SCEVCastExpr(const FoldingSetNodeID &ID,
+ unsigned SCEVTy, const SCEV *op, const Type *ty)
+ : SCEV(ID, SCEVTy), Op(op), Ty(ty) {}
+
+bool SCEVCastExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
+ return Op->dominates(BB, DT);
+}
+
+bool SCEVCastExpr::properlyDominates(BasicBlock *BB, DominatorTree *DT) const {
+ return Op->properlyDominates(BB, DT);
+}
+
+SCEVTruncateExpr::SCEVTruncateExpr(const FoldingSetNodeID &ID,
+ const SCEV *op, const Type *ty)
+ : SCEVCastExpr(ID, scTruncate, op, ty) {
+ assert((Op->getType()->isInteger() || isa<PointerType>(Op->getType())) &&
+ (Ty->isInteger() || isa<PointerType>(Ty)) &&
+ "Cannot truncate non-integer value!");
+}
+
+void SCEVTruncateExpr::print(raw_ostream &OS) const {
+ OS << "(trunc " << *Op->getType() << " " << *Op << " to " << *Ty << ")";
+}
+
+SCEVZeroExtendExpr::SCEVZeroExtendExpr(const FoldingSetNodeID &ID,
+ const SCEV *op, const Type *ty)
+ : SCEVCastExpr(ID, scZeroExtend, op, ty) {
+ assert((Op->getType()->isInteger() || isa<PointerType>(Op->getType())) &&
+ (Ty->isInteger() || isa<PointerType>(Ty)) &&
+ "Cannot zero extend non-integer value!");
+}
+
+void SCEVZeroExtendExpr::print(raw_ostream &OS) const {
+ OS << "(zext " << *Op->getType() << " " << *Op << " to " << *Ty << ")";
+}
+
+SCEVSignExtendExpr::SCEVSignExtendExpr(const FoldingSetNodeID &ID,
+ const SCEV *op, const Type *ty)
+ : SCEVCastExpr(ID, scSignExtend, op, ty) {
+ assert((Op->getType()->isInteger() || isa<PointerType>(Op->getType())) &&
+ (Ty->isInteger() || isa<PointerType>(Ty)) &&
+ "Cannot sign extend non-integer value!");
+}
+
+void SCEVSignExtendExpr::print(raw_ostream &OS) const {
+ OS << "(sext " << *Op->getType() << " " << *Op << " to " << *Ty << ")";
+}
+
+void SCEVCommutativeExpr::print(raw_ostream &OS) const {
+ assert(Operands.size() > 1 && "This plus expr shouldn't exist!");
+ const char *OpStr = getOperationStr();
+ OS << "(" << *Operands[0];
+ for (unsigned i = 1, e = Operands.size(); i != e; ++i)
+ OS << OpStr << *Operands[i];
+ OS << ")";
+}
+
+bool SCEVNAryExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
+ for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
+ if (!getOperand(i)->dominates(BB, DT))
+ return false;
+ }
+ return true;
+}
+
+bool SCEVNAryExpr::properlyDominates(BasicBlock *BB, DominatorTree *DT) const {
+ for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
+ if (!getOperand(i)->properlyDominates(BB, DT))
+ return false;
+ }
+ return true;
+}
+
+bool SCEVUDivExpr::dominates(BasicBlock *BB, DominatorTree *DT) const {
+ return LHS->dominates(BB, DT) && RHS->dominates(BB, DT);
+}
+
+bool SCEVUDivExpr::properlyDominates(BasicBlock *BB, DominatorTree *DT) const {
+ return LHS->properlyDominates(BB, DT) && RHS->properlyDominates(BB, DT);
+}
+
+void SCEVUDivExpr::print(raw_ostream &OS) const {
+ OS << "(" << *LHS << " /u " << *RHS << ")";
+}
+
+const Type *SCEVUDivExpr::getType() const {
+ // In most cases the types of LHS and RHS will be the same, but in some
+ // crazy cases one or the other may be a pointer. ScalarEvolution doesn't
+ // depend on the type for correctness, but handling types carefully can
+ // avoid extra casts in the SCEVExpander. The LHS is more likely to be
+ // a pointer type than the RHS, so use the RHS' type here.
+ return RHS->getType();
+}
+
+bool SCEVAddRecExpr::isLoopInvariant(const Loop *QueryLoop) const {
+ // Add recurrences are never invariant in the function-body (null loop).
+ if (!QueryLoop)
+ return false;
+
+ // This recurrence is variant w.r.t. QueryLoop if QueryLoop contains L.
+ if (QueryLoop->contains(L))
+ return false;
+
+ // This recurrence is variant w.r.t. QueryLoop if any of its operands
+ // are variant.
+ for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
+ if (!getOperand(i)->isLoopInvariant(QueryLoop))
+ return false;
+
+ // Otherwise it's loop-invariant.
+ return true;
+}
+
+void SCEVAddRecExpr::print(raw_ostream &OS) const {
+ OS << "{" << *Operands[0];
+ for (unsigned i = 1, e = Operands.size(); i != e; ++i)
+ OS << ",+," << *Operands[i];
+ OS << "}<";
+ WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
+ OS << ">";
+}
+
+bool SCEVUnknown::isLoopInvariant(const Loop *L) const {
+ // All non-instruction values are loop invariant. All instructions are loop
+ // invariant if they are not contained in the specified loop.
+ // Instructions are never considered invariant in the function body
+ // (null loop) because they are defined within the "loop".
+ if (Instruction *I = dyn_cast<Instruction>(V))
+ return L && !L->contains(I);
+ return true;
+}
+
+bool SCEVUnknown::dominates(BasicBlock *BB, DominatorTree *DT) const {
+ if (Instruction *I = dyn_cast<Instruction>(getValue()))
+ return DT->dominates(I->getParent(), BB);
+ return true;
+}
+
+bool SCEVUnknown::properlyDominates(BasicBlock *BB, DominatorTree *DT) const {
+ if (Instruction *I = dyn_cast<Instruction>(getValue()))
+ return DT->properlyDominates(I->getParent(), BB);
+ return true;
+}
+
+const Type *SCEVUnknown::getType() const {
+ return V->getType();
+}
+
+bool SCEVUnknown::isSizeOf(const Type *&AllocTy) const {
+ if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(V))
+ if (VCE->getOpcode() == Instruction::PtrToInt)
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
+ if (CE->getOpcode() == Instruction::GetElementPtr &&
+ CE->getOperand(0)->isNullValue() &&
+ CE->getNumOperands() == 2)
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(1)))
+ if (CI->isOne()) {
+ AllocTy = cast<PointerType>(CE->getOperand(0)->getType())
+ ->getElementType();
+ return true;
+ }
+
+ return false;
+}
+
+bool SCEVUnknown::isAlignOf(const Type *&AllocTy) const {
+ if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(V))
+ if (VCE->getOpcode() == Instruction::PtrToInt)
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
+ if (CE->getOpcode() == Instruction::GetElementPtr &&
+ CE->getOperand(0)->isNullValue()) {
+ const Type *Ty =
+ cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
+ if (const StructType *STy = dyn_cast<StructType>(Ty))
+ if (!STy->isPacked() &&
+ CE->getNumOperands() == 3 &&
+ CE->getOperand(1)->isNullValue()) {
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(2)))
+ if (CI->isOne() &&
+ STy->getNumElements() == 2 &&
+ STy->getElementType(0)->isInteger(1)) {
+ AllocTy = STy->getElementType(1);
+ return true;
+ }
+ }
+ }
+
+ return false;
+}
+
+bool SCEVUnknown::isOffsetOf(const Type *&CTy, Constant *&FieldNo) const {
+ if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(V))
+ if (VCE->getOpcode() == Instruction::PtrToInt)
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
+ if (CE->getOpcode() == Instruction::GetElementPtr &&
+ CE->getNumOperands() == 3 &&
+ CE->getOperand(0)->isNullValue() &&
+ CE->getOperand(1)->isNullValue()) {
+ const Type *Ty =
+ cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
+ // Ignore vector types here so that ScalarEvolutionExpander doesn't
+ // emit getelementptrs that index into vectors.
+ if (isa<StructType>(Ty) || isa<ArrayType>(Ty)) {
+ CTy = Ty;
+ FieldNo = CE->getOperand(2);
+ return true;
+ }
+ }
+
+ return false;
+}
+
+void SCEVUnknown::print(raw_ostream &OS) const {
+ const Type *AllocTy;
+ if (isSizeOf(AllocTy)) {
+ OS << "sizeof(" << *AllocTy << ")";
+ return;
+ }
+ if (isAlignOf(AllocTy)) {
+ OS << "alignof(" << *AllocTy << ")";
+ return;
+ }
+
+ const Type *CTy;
+ Constant *FieldNo;
+ if (isOffsetOf(CTy, FieldNo)) {
+ OS << "offsetof(" << *CTy << ", ";
+ WriteAsOperand(OS, FieldNo, false);
+ OS << ")";
+ return;
+ }
+
+ // Otherwise just print it normally.
+ WriteAsOperand(OS, V, false);
+}
+
+//===----------------------------------------------------------------------===//
+// SCEV Utilities
+//===----------------------------------------------------------------------===//
+
+static bool CompareTypes(const Type *A, const Type *B) {
+ if (A->getTypeID() != B->getTypeID())
+ return A->getTypeID() < B->getTypeID();
+ if (const IntegerType *AI = dyn_cast<IntegerType>(A)) {
+ const IntegerType *BI = cast<IntegerType>(B);
+ return AI->getBitWidth() < BI->getBitWidth();
+ }
+ if (const PointerType *AI = dyn_cast<PointerType>(A)) {
+ const PointerType *BI = cast<PointerType>(B);
+ return CompareTypes(AI->getElementType(), BI->getElementType());
+ }
+ if (const ArrayType *AI = dyn_cast<ArrayType>(A)) {
+ const ArrayType *BI = cast<ArrayType>(B);
+ if (AI->getNumElements() != BI->getNumElements())
+ return AI->getNumElements() < BI->getNumElements();
+ return CompareTypes(AI->getElementType(), BI->getElementType());
+ }
+ if (const VectorType *AI = dyn_cast<VectorType>(A)) {
+ const VectorType *BI = cast<VectorType>(B);
+ if (AI->getNumElements() != BI->getNumElements())
+ return AI->getNumElements() < BI->getNumElements();
+ return CompareTypes(AI->getElementType(), BI->getElementType());
+ }
+ if (const StructType *AI = dyn_cast<StructType>(A)) {
+ const StructType *BI = cast<StructType>(B);
+ if (AI->getNumElements() != BI->getNumElements())
+ return AI->getNumElements() < BI->getNumElements();
+ for (unsigned i = 0, e = AI->getNumElements(); i != e; ++i)
+ if (CompareTypes(AI->getElementType(i), BI->getElementType(i)) ||
+ CompareTypes(BI->getElementType(i), AI->getElementType(i)))
+ return CompareTypes(AI->getElementType(i), BI->getElementType(i));
+ }
+ return false;
+}
+
+namespace {
+ /// SCEVComplexityCompare - Return true if the complexity of the LHS is less
+ /// than the complexity of the RHS. This comparator is used to canonicalize
+ /// expressions.
+ class SCEVComplexityCompare {
+ LoopInfo *LI;
+ public:
+ explicit SCEVComplexityCompare(LoopInfo *li) : LI(li) {}
+
+ bool operator()(const SCEV *LHS, const SCEV *RHS) const {
+ // Fast-path: SCEVs are uniqued so we can do a quick equality check.
+ if (LHS == RHS)
+ return false;
+
+ // Primarily, sort the SCEVs by their getSCEVType().
+ if (LHS->getSCEVType() != RHS->getSCEVType())
+ return LHS->getSCEVType() < RHS->getSCEVType();
+
+ // Aside from the getSCEVType() ordering, the particular ordering
+ // isn't very important except that it's beneficial to be consistent,
+ // so that (a + b) and (b + a) don't end up as different expressions.
+
+ // Sort SCEVUnknown values with some loose heuristics. TODO: This is
+ // not as complete as it could be.
+ if (const SCEVUnknown *LU = dyn_cast<SCEVUnknown>(LHS)) {
+ const SCEVUnknown *RU = cast<SCEVUnknown>(RHS);
+
+ // Order pointer values after integer values. This helps SCEVExpander
+ // form GEPs.
+ if (isa<PointerType>(LU->getType()) && !isa<PointerType>(RU->getType()))
+ return false;
+ if (isa<PointerType>(RU->getType()) && !isa<PointerType>(LU->getType()))
+ return true;
+
+ // Compare getValueID values.
+ if (LU->getValue()->getValueID() != RU->getValue()->getValueID())
+ return LU->getValue()->getValueID() < RU->getValue()->getValueID();
+
+ // Sort arguments by their position.
+ if (const Argument *LA = dyn_cast<Argument>(LU->getValue())) {
+ const Argument *RA = cast<Argument>(RU->getValue());
+ return LA->getArgNo() < RA->getArgNo();
+ }
+
+ // For instructions, compare their loop depth, and their opcode.
+ // This is pretty loose.
+ if (Instruction *LV = dyn_cast<Instruction>(LU->getValue())) {
+ Instruction *RV = cast<Instruction>(RU->getValue());
+
+ // Compare loop depths.
+ if (LI->getLoopDepth(LV->getParent()) !=
+ LI->getLoopDepth(RV->getParent()))
+ return LI->getLoopDepth(LV->getParent()) <
+ LI->getLoopDepth(RV->getParent());
+
+ // Compare opcodes.
+ if (LV->getOpcode() != RV->getOpcode())
+ return LV->getOpcode() < RV->getOpcode();
+
+ // Compare the number of operands.
+ if (LV->getNumOperands() != RV->getNumOperands())
+ return LV->getNumOperands() < RV->getNumOperands();
+ }
+
+ return false;
+ }
+
+ // Compare constant values.
+ if (const SCEVConstant *LC = dyn_cast<SCEVConstant>(LHS)) {
+ const SCEVConstant *RC = cast<SCEVConstant>(RHS);
+ if (LC->getValue()->getBitWidth() != RC->getValue()->getBitWidth())
+ return LC->getValue()->getBitWidth() < RC->getValue()->getBitWidth();
+ return LC->getValue()->getValue().ult(RC->getValue()->getValue());
+ }
+
+ // Compare addrec loop depths.
+ if (const SCEVAddRecExpr *LA = dyn_cast<SCEVAddRecExpr>(LHS)) {
+ const SCEVAddRecExpr *RA = cast<SCEVAddRecExpr>(RHS);
+ if (LA->getLoop()->getLoopDepth() != RA->getLoop()->getLoopDepth())
+ return LA->getLoop()->getLoopDepth() < RA->getLoop()->getLoopDepth();
+ }
+
+ // Lexicographically compare n-ary expressions.
+ if (const SCEVNAryExpr *LC = dyn_cast<SCEVNAryExpr>(LHS)) {
+ const SCEVNAryExpr *RC = cast<SCEVNAryExpr>(RHS);
+ for (unsigned i = 0, e = LC->getNumOperands(); i != e; ++i) {
+ if (i >= RC->getNumOperands())
+ return false;
+ if (operator()(LC->getOperand(i), RC->getOperand(i)))
+ return true;
+ if (operator()(RC->getOperand(i), LC->getOperand(i)))
+ return false;
+ }
+ return LC->getNumOperands() < RC->getNumOperands();
+ }
+
+ // Lexicographically compare udiv expressions.
+ if (const SCEVUDivExpr *LC = dyn_cast<SCEVUDivExpr>(LHS)) {
+ const SCEVUDivExpr *RC = cast<SCEVUDivExpr>(RHS);
+ if (operator()(LC->getLHS(), RC->getLHS()))
+ return true;
+ if (operator()(RC->getLHS(), LC->getLHS()))
+ return false;
+ if (operator()(LC->getRHS(), RC->getRHS()))
+ return true;
+ if (operator()(RC->getRHS(), LC->getRHS()))
+ return false;
+ return false;
+ }
+
+ // Compare cast expressions by operand.
+ if (const SCEVCastExpr *LC = dyn_cast<SCEVCastExpr>(LHS)) {
+ const SCEVCastExpr *RC = cast<SCEVCastExpr>(RHS);
+ return operator()(LC->getOperand(), RC->getOperand());
+ }
+
+ llvm_unreachable("Unknown SCEV kind!");
+ return false;
+ }
+ };
+}
+
+/// GroupByComplexity - Given a list of SCEV objects, order them by their
+/// complexity, and group objects of the same complexity together by value.
+/// When this routine is finished, we know that any duplicates in the vector are
+/// consecutive and that complexity is monotonically increasing.
+///
+/// Note that we go take special precautions to ensure that we get determinstic
+/// results from this routine. In other words, we don't want the results of
+/// this to depend on where the addresses of various SCEV objects happened to
+/// land in memory.
+///
+static void GroupByComplexity(SmallVectorImpl<const SCEV *> &Ops,
+ LoopInfo *LI) {
+ if (Ops.size() < 2) return; // Noop
+ if (Ops.size() == 2) {
+ // This is the common case, which also happens to be trivially simple.
+ // Special case it.
+ if (SCEVComplexityCompare(LI)(Ops[1], Ops[0]))
+ std::swap(Ops[0], Ops[1]);
+ return;
+ }
+
+ // Do the rough sort by complexity.
+ std::stable_sort(Ops.begin(), Ops.end(), SCEVComplexityCompare(LI));
+
+ // Now that we are sorted by complexity, group elements of the same
+ // complexity. Note that this is, at worst, N^2, but the vector is likely to
+ // be extremely short in practice. Note that we take this approach because we
+ // do not want to depend on the addresses of the objects we are grouping.
+ for (unsigned i = 0, e = Ops.size(); i != e-2; ++i) {
+ const SCEV *S = Ops[i];
+ unsigned Complexity = S->getSCEVType();
+
+ // If there are any objects of the same complexity and same value as this
+ // one, group them.
+ for (unsigned j = i+1; j != e && Ops[j]->getSCEVType() == Complexity; ++j) {
+ if (Ops[j] == S) { // Found a duplicate.
+ // Move it to immediately after i'th element.
+ std::swap(Ops[i+1], Ops[j]);
+ ++i; // no need to rescan it.
+ if (i == e-2) return; // Done!
+ }
+ }
+ }
+}
+
+
+
+//===----------------------------------------------------------------------===//
+// Simple SCEV method implementations
+//===----------------------------------------------------------------------===//
+
+/// BinomialCoefficient - Compute BC(It, K). The result has width W.
+/// Assume, K > 0.
+static const SCEV *BinomialCoefficient(const SCEV *It, unsigned K,
+ ScalarEvolution &SE,
+ const Type* ResultTy) {
+ // Handle the simplest case efficiently.
+ if (K == 1)
+ return SE.getTruncateOrZeroExtend(It, ResultTy);
+
+ // We are using the following formula for BC(It, K):
+ //
+ // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / K!
+ //
+ // Suppose, W is the bitwidth of the return value. We must be prepared for
+ // overflow. Hence, we must assure that the result of our computation is
+ // equal to the accurate one modulo 2^W. Unfortunately, division isn't
+ // safe in modular arithmetic.
+ //
+ // However, this code doesn't use exactly that formula; the formula it uses
+ // is something like the following, where T is the number of factors of 2 in
+ // K! (i.e. trailing zeros in the binary representation of K!), and ^ is
+ // exponentiation:
+ //
+ // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / 2^T / (K! / 2^T)
+ //
+ // This formula is trivially equivalent to the previous formula. However,
+ // this formula can be implemented much more efficiently. The trick is that
+ // K! / 2^T is odd, and exact division by an odd number *is* safe in modular
+ // arithmetic. To do exact division in modular arithmetic, all we have
+ // to do is multiply by the inverse. Therefore, this step can be done at
+ // width W.
+ //
+ // The next issue is how to safely do the division by 2^T. The way this
+ // is done is by doing the multiplication step at a width of at least W + T
+ // bits. This way, the bottom W+T bits of the product are accurate. Then,
+ // when we perform the division by 2^T (which is equivalent to a right shift
+ // by T), the bottom W bits are accurate. Extra bits are okay; they'll get
+ // truncated out after the division by 2^T.
+ //
+ // In comparison to just directly using the first formula, this technique
+ // is much more efficient; using the first formula requires W * K bits,
+ // but this formula less than W + K bits. Also, the first formula requires
+ // a division step, whereas this formula only requires multiplies and shifts.
+ //
+ // It doesn't matter whether the subtraction step is done in the calculation
+ // width or the input iteration count's width; if the subtraction overflows,
+ // the result must be zero anyway. We prefer here to do it in the width of
+ // the induction variable because it helps a lot for certain cases; CodeGen
+ // isn't smart enough to ignore the overflow, which leads to much less
+ // efficient code if the width of the subtraction is wider than the native
+ // register width.
+ //
+ // (It's possible to not widen at all by pulling out factors of 2 before
+ // the multiplication; for example, K=2 can be calculated as
+ // It/2*(It+(It*INT_MIN/INT_MIN)+-1). However, it requires
+ // extra arithmetic, so it's not an obvious win, and it gets
+ // much more complicated for K > 3.)
+
+ // Protection from insane SCEVs; this bound is conservative,
+ // but it probably doesn't matter.
+ if (K > 1000)
+ return SE.getCouldNotCompute();
+
+ unsigned W = SE.getTypeSizeInBits(ResultTy);
+
+ // Calculate K! / 2^T and T; we divide out the factors of two before
+ // multiplying for calculating K! / 2^T to avoid overflow.
+ // Other overflow doesn't matter because we only care about the bottom
+ // W bits of the result.
+ APInt OddFactorial(W, 1);
+ unsigned T = 1;
+ for (unsigned i = 3; i <= K; ++i) {
+ APInt Mult(W, i);
+ unsigned TwoFactors = Mult.countTrailingZeros();
+ T += TwoFactors;
+ Mult = Mult.lshr(TwoFactors);
+ OddFactorial *= Mult;
+ }
+
+ // We need at least W + T bits for the multiplication step
+ unsigned CalculationBits = W + T;
+
+ // Calcuate 2^T, at width T+W.
+ APInt DivFactor = APInt(CalculationBits, 1).shl(T);
+
+ // Calculate the multiplicative inverse of K! / 2^T;
+ // this multiplication factor will perform the exact division by
+ // K! / 2^T.
+ APInt Mod = APInt::getSignedMinValue(W+1);
+ APInt MultiplyFactor = OddFactorial.zext(W+1);
+ MultiplyFactor = MultiplyFactor.multiplicativeInverse(Mod);
+ MultiplyFactor = MultiplyFactor.trunc(W);
+
+ // Calculate the product, at width T+W
+ const IntegerType *CalculationTy = IntegerType::get(SE.getContext(),
+ CalculationBits);
+ const SCEV *Dividend = SE.getTruncateOrZeroExtend(It, CalculationTy);
+ for (unsigned i = 1; i != K; ++i) {
+ const SCEV *S = SE.getMinusSCEV(It, SE.getIntegerSCEV(i, It->getType()));
+ Dividend = SE.getMulExpr(Dividend,
+ SE.getTruncateOrZeroExtend(S, CalculationTy));
+ }
+
+ // Divide by 2^T
+ const SCEV *DivResult = SE.getUDivExpr(Dividend, SE.getConstant(DivFactor));
+
+ // Truncate the result, and divide by K! / 2^T.
+
+ return SE.getMulExpr(SE.getConstant(MultiplyFactor),
+ SE.getTruncateOrZeroExtend(DivResult, ResultTy));
+}
+
+/// evaluateAtIteration - Return the value of this chain of recurrences at
+/// the specified iteration number. We can evaluate this recurrence by
+/// multiplying each element in the chain by the binomial coefficient
+/// corresponding to it. In other words, we can evaluate {A,+,B,+,C,+,D} as:
+///
+/// A*BC(It, 0) + B*BC(It, 1) + C*BC(It, 2) + D*BC(It, 3)
+///
+/// where BC(It, k) stands for binomial coefficient.
+///
+const SCEV *SCEVAddRecExpr::evaluateAtIteration(const SCEV *It,
+ ScalarEvolution &SE) const {
+ const SCEV *Result = getStart();
+ for (unsigned i = 1, e = getNumOperands(); i != e; ++i) {
+ // The computation is correct in the face of overflow provided that the
+ // multiplication is performed _after_ the evaluation of the binomial
+ // coefficient.
+ const SCEV *Coeff = BinomialCoefficient(It, i, SE, getType());
+ if (isa<SCEVCouldNotCompute>(Coeff))
+ return Coeff;
+
+ Result = SE.getAddExpr(Result, SE.getMulExpr(getOperand(i), Coeff));
+ }
+ return Result;
+}
+
+//===----------------------------------------------------------------------===//
+// SCEV Expression folder implementations
+//===----------------------------------------------------------------------===//
+
+const SCEV *ScalarEvolution::getTruncateExpr(const SCEV *Op,
+ const Type *Ty) {
+ assert(getTypeSizeInBits(Op->getType()) > getTypeSizeInBits(Ty) &&
+ "This is not a truncating conversion!");
+ assert(isSCEVable(Ty) &&
+ "This is not a conversion to a SCEVable type!");
+ Ty = getEffectiveSCEVType(Ty);
+
+ FoldingSetNodeID ID;
+ ID.AddInteger(scTruncate);
+ ID.AddPointer(Op);
+ ID.AddPointer(Ty);
+ void *IP = 0;
+ if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
+
+ // Fold if the operand is constant.
+ if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
+ return getConstant(
+ cast<ConstantInt>(ConstantExpr::getTrunc(SC->getValue(), Ty)));
+
+ // trunc(trunc(x)) --> trunc(x)
+ if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op))
+ return getTruncateExpr(ST->getOperand(), Ty);
+
+ // trunc(sext(x)) --> sext(x) if widening or trunc(x) if narrowing
+ if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
+ return getTruncateOrSignExtend(SS->getOperand(), Ty);
+
+ // trunc(zext(x)) --> zext(x) if widening or trunc(x) if narrowing
+ if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
+ return getTruncateOrZeroExtend(SZ->getOperand(), Ty);
+
+ // If the input value is a chrec scev, truncate the chrec's operands.
+ if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Op)) {
+ SmallVector<const SCEV *, 4> Operands;
+ for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
+ Operands.push_back(getTruncateExpr(AddRec->getOperand(i), Ty));
+ return getAddRecExpr(Operands, AddRec->getLoop());
+ }
+
+ // The cast wasn't folded; create an explicit cast node.
+ // Recompute the insert position, as it may have been invalidated.
+ if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
+ SCEV *S = SCEVAllocator.Allocate<SCEVTruncateExpr>();
+ new (S) SCEVTruncateExpr(ID, Op, Ty);
+ UniqueSCEVs.InsertNode(S, IP);
+ return S;
+}
+
+const SCEV *ScalarEvolution::getZeroExtendExpr(const SCEV *Op,
+ const Type *Ty) {
+ assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
+ "This is not an extending conversion!");
+ assert(isSCEVable(Ty) &&
+ "This is not a conversion to a SCEVable type!");
+ Ty = getEffectiveSCEVType(Ty);
+
+ // Fold if the operand is constant.
+ if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op)) {
+ const Type *IntTy = getEffectiveSCEVType(Ty);
+ Constant *C = ConstantExpr::getZExt(SC->getValue(), IntTy);
+ if (IntTy != Ty) C = ConstantExpr::getIntToPtr(C, Ty);
+ return getConstant(cast<ConstantInt>(C));
+ }
+
+ // zext(zext(x)) --> zext(x)
+ if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
+ return getZeroExtendExpr(SZ->getOperand(), Ty);
+
+ // Before doing any expensive analysis, check to see if we've already
+ // computed a SCEV for this Op and Ty.
+ FoldingSetNodeID ID;
+ ID.AddInteger(scZeroExtend);
+ ID.AddPointer(Op);
+ ID.AddPointer(Ty);
+ void *IP = 0;
+ if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
+
+ // If the input value is a chrec scev, and we can prove that the value
+ // did not overflow the old, smaller, value, we can zero extend all of the
+ // operands (often constants). This allows analysis of something like
+ // this: for (unsigned char X = 0; X < 100; ++X) { int Y = X; }
+ if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
+ if (AR->isAffine()) {
+ const SCEV *Start = AR->getStart();
+ const SCEV *Step = AR->getStepRecurrence(*this);
+ unsigned BitWidth = getTypeSizeInBits(AR->getType());
+ const Loop *L = AR->getLoop();
+
+ // If we have special knowledge that this addrec won't overflow,
+ // we don't need to do any further analysis.
+ if (AR->hasNoUnsignedWrap())
+ return getAddRecExpr(getZeroExtendExpr(Start, Ty),
+ getZeroExtendExpr(Step, Ty),
+ L);
+
+ // Check whether the backedge-taken count is SCEVCouldNotCompute.
+ // Note that this serves two purposes: It filters out loops that are
+ // simply not analyzable, and it covers the case where this code is
+ // being called from within backedge-taken count analysis, such that
+ // attempting to ask for the backedge-taken count would likely result
+ // in infinite recursion. In the later case, the analysis code will
+ // cope with a conservative value, and it will take care to purge
+ // that value once it has finished.
+ const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
+ if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
+ // Manually compute the final value for AR, checking for
+ // overflow.
+
+ // Check whether the backedge-taken count can be losslessly casted to
+ // the addrec's type. The count is always unsigned.
+ const SCEV *CastedMaxBECount =
+ getTruncateOrZeroExtend(MaxBECount, Start->getType());
+ const SCEV *RecastedMaxBECount =
+ getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
+ if (MaxBECount == RecastedMaxBECount) {
+ const Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
+ // Check whether Start+Step*MaxBECount has no unsigned overflow.
+ const SCEV *ZMul =
+ getMulExpr(CastedMaxBECount,
+ getTruncateOrZeroExtend(Step, Start->getType()));
+ const SCEV *Add = getAddExpr(Start, ZMul);
+ const SCEV *OperandExtendedAdd =
+ getAddExpr(getZeroExtendExpr(Start, WideTy),
+ getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
+ getZeroExtendExpr(Step, WideTy)));
+ if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd)
+ // Return the expression with the addrec on the outside.
+ return getAddRecExpr(getZeroExtendExpr(Start, Ty),
+ getZeroExtendExpr(Step, Ty),
+ L);
+
+ // Similar to above, only this time treat the step value as signed.
+ // This covers loops that count down.
+ const SCEV *SMul =
+ getMulExpr(CastedMaxBECount,
+ getTruncateOrSignExtend(Step, Start->getType()));
+ Add = getAddExpr(Start, SMul);
+ OperandExtendedAdd =
+ getAddExpr(getZeroExtendExpr(Start, WideTy),
+ getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
+ getSignExtendExpr(Step, WideTy)));
+ if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd)
+ // Return the expression with the addrec on the outside.
+ return getAddRecExpr(getZeroExtendExpr(Start, Ty),
+ getSignExtendExpr(Step, Ty),
+ L);
+ }
+
+ // If the backedge is guarded by a comparison with the pre-inc value
+ // the addrec is safe. Also, if the entry is guarded by a comparison
+ // with the start value and the backedge is guarded by a comparison
+ // with the post-inc value, the addrec is safe.
+ if (isKnownPositive(Step)) {
+ const SCEV *N = getConstant(APInt::getMinValue(BitWidth) -
+ getUnsignedRange(Step).getUnsignedMax());
+ if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT, AR, N) ||
+ (isLoopGuardedByCond(L, ICmpInst::ICMP_ULT, Start, N) &&
+ isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT,
+ AR->getPostIncExpr(*this), N)))
+ // Return the expression with the addrec on the outside.
+ return getAddRecExpr(getZeroExtendExpr(Start, Ty),
+ getZeroExtendExpr(Step, Ty),
+ L);
+ } else if (isKnownNegative(Step)) {
+ const SCEV *N = getConstant(APInt::getMaxValue(BitWidth) -
+ getSignedRange(Step).getSignedMin());
+ if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT, AR, N) &&
+ (isLoopGuardedByCond(L, ICmpInst::ICMP_UGT, Start, N) ||
+ isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT,
+ AR->getPostIncExpr(*this), N)))
+ // Return the expression with the addrec on the outside.
+ return getAddRecExpr(getZeroExtendExpr(Start, Ty),
+ getSignExtendExpr(Step, Ty),
+ L);
+ }
+ }
+ }
+
+ // The cast wasn't folded; create an explicit cast node.
+ // Recompute the insert position, as it may have been invalidated.
+ if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
+ SCEV *S = SCEVAllocator.Allocate<SCEVZeroExtendExpr>();
+ new (S) SCEVZeroExtendExpr(ID, Op, Ty);
+ UniqueSCEVs.InsertNode(S, IP);
+ return S;
+}
+
+const SCEV *ScalarEvolution::getSignExtendExpr(const SCEV *Op,
+ const Type *Ty) {
+ assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
+ "This is not an extending conversion!");
+ assert(isSCEVable(Ty) &&
+ "This is not a conversion to a SCEVable type!");
+ Ty = getEffectiveSCEVType(Ty);
+
+ // Fold if the operand is constant.
+ if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op)) {
+ const Type *IntTy = getEffectiveSCEVType(Ty);
+ Constant *C = ConstantExpr::getSExt(SC->getValue(), IntTy);
+ if (IntTy != Ty) C = ConstantExpr::getIntToPtr(C, Ty);
+ return getConstant(cast<ConstantInt>(C));
+ }
+
+ // sext(sext(x)) --> sext(x)
+ if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
+ return getSignExtendExpr(SS->getOperand(), Ty);
+
+ // Before doing any expensive analysis, check to see if we've already
+ // computed a SCEV for this Op and Ty.
+ FoldingSetNodeID ID;
+ ID.AddInteger(scSignExtend);
+ ID.AddPointer(Op);
+ ID.AddPointer(Ty);
+ void *IP = 0;
+ if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
+
+ // If the input value is a chrec scev, and we can prove that the value
+ // did not overflow the old, smaller, value, we can sign extend all of the
+ // operands (often constants). This allows analysis of something like
+ // this: for (signed char X = 0; X < 100; ++X) { int Y = X; }
+ if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op))
+ if (AR->isAffine()) {
+ const SCEV *Start = AR->getStart();
+ const SCEV *Step = AR->getStepRecurrence(*this);
+ unsigned BitWidth = getTypeSizeInBits(AR->getType());
+ const Loop *L = AR->getLoop();
+
+ // If we have special knowledge that this addrec won't overflow,
+ // we don't need to do any further analysis.
+ if (AR->hasNoSignedWrap())
+ return getAddRecExpr(getSignExtendExpr(Start, Ty),
+ getSignExtendExpr(Step, Ty),
+ L);
+
+ // Check whether the backedge-taken count is SCEVCouldNotCompute.
+ // Note that this serves two purposes: It filters out loops that are
+ // simply not analyzable, and it covers the case where this code is
+ // being called from within backedge-taken count analysis, such that
+ // attempting to ask for the backedge-taken count would likely result
+ // in infinite recursion. In the later case, the analysis code will
+ // cope with a conservative value, and it will take care to purge
+ // that value once it has finished.
+ const SCEV *MaxBECount = getMaxBackedgeTakenCount(L);
+ if (!isa<SCEVCouldNotCompute>(MaxBECount)) {
+ // Manually compute the final value for AR, checking for
+ // overflow.
+
+ // Check whether the backedge-taken count can be losslessly casted to
+ // the addrec's type. The count is always unsigned.
+ const SCEV *CastedMaxBECount =
+ getTruncateOrZeroExtend(MaxBECount, Start->getType());
+ const SCEV *RecastedMaxBECount =
+ getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
+ if (MaxBECount == RecastedMaxBECount) {
+ const Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
+ // Check whether Start+Step*MaxBECount has no signed overflow.
+ const SCEV *SMul =
+ getMulExpr(CastedMaxBECount,
+ getTruncateOrSignExtend(Step, Start->getType()));
+ const SCEV *Add = getAddExpr(Start, SMul);
+ const SCEV *OperandExtendedAdd =
+ getAddExpr(getSignExtendExpr(Start, WideTy),
+ getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
+ getSignExtendExpr(Step, WideTy)));
+ if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd)
+ // Return the expression with the addrec on the outside.
+ return getAddRecExpr(getSignExtendExpr(Start, Ty),
+ getSignExtendExpr(Step, Ty),
+ L);
+
+ // Similar to above, only this time treat the step value as unsigned.
+ // This covers loops that count up with an unsigned step.
+ const SCEV *UMul =
+ getMulExpr(CastedMaxBECount,
+ getTruncateOrZeroExtend(Step, Start->getType()));
+ Add = getAddExpr(Start, UMul);
+ OperandExtendedAdd =
+ getAddExpr(getSignExtendExpr(Start, WideTy),
+ getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
+ getZeroExtendExpr(Step, WideTy)));
+ if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd)
+ // Return the expression with the addrec on the outside.
+ return getAddRecExpr(getSignExtendExpr(Start, Ty),
+ getZeroExtendExpr(Step, Ty),
+ L);
+ }
+
+ // If the backedge is guarded by a comparison with the pre-inc value
+ // the addrec is safe. Also, if the entry is guarded by a comparison
+ // with the start value and the backedge is guarded by a comparison
+ // with the post-inc value, the addrec is safe.
+ if (isKnownPositive(Step)) {
+ const SCEV *N = getConstant(APInt::getSignedMinValue(BitWidth) -
+ getSignedRange(Step).getSignedMax());
+ if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SLT, AR, N) ||
+ (isLoopGuardedByCond(L, ICmpInst::ICMP_SLT, Start, N) &&
+ isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SLT,
+ AR->getPostIncExpr(*this), N)))
+ // Return the expression with the addrec on the outside.
+ return getAddRecExpr(getSignExtendExpr(Start, Ty),
+ getSignExtendExpr(Step, Ty),
+ L);
+ } else if (isKnownNegative(Step)) {
+ const SCEV *N = getConstant(APInt::getSignedMaxValue(BitWidth) -
+ getSignedRange(Step).getSignedMin());
+ if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SGT, AR, N) ||
+ (isLoopGuardedByCond(L, ICmpInst::ICMP_SGT, Start, N) &&
+ isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SGT,
+ AR->getPostIncExpr(*this), N)))
+ // Return the expression with the addrec on the outside.
+ return getAddRecExpr(getSignExtendExpr(Start, Ty),
+ getSignExtendExpr(Step, Ty),
+ L);
+ }
+ }
+ }
+
+ // The cast wasn't folded; create an explicit cast node.
+ // Recompute the insert position, as it may have been invalidated.
+ if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
+ SCEV *S = SCEVAllocator.Allocate<SCEVSignExtendExpr>();
+ new (S) SCEVSignExtendExpr(ID, Op, Ty);
+ UniqueSCEVs.InsertNode(S, IP);
+ return S;
+}
+
+/// getAnyExtendExpr - Return a SCEV for the given operand extended with
+/// unspecified bits out to the given type.
+///
+const SCEV *ScalarEvolution::getAnyExtendExpr(const SCEV *Op,
+ const Type *Ty) {
+ assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
+ "This is not an extending conversion!");
+ assert(isSCEVable(Ty) &&
+ "This is not a conversion to a SCEVable type!");
+ Ty = getEffectiveSCEVType(Ty);
+
+ // Sign-extend negative constants.
+ if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
+ if (SC->getValue()->getValue().isNegative())
+ return getSignExtendExpr(Op, Ty);
+
+ // Peel off a truncate cast.
+ if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Op)) {
+ const SCEV *NewOp = T->getOperand();
+ if (getTypeSizeInBits(NewOp->getType()) < getTypeSizeInBits(Ty))
+ return getAnyExtendExpr(NewOp, Ty);
+ return getTruncateOrNoop(NewOp, Ty);
+ }
+
+ // Next try a zext cast. If the cast is folded, use it.
+ const SCEV *ZExt = getZeroExtendExpr(Op, Ty);
+ if (!isa<SCEVZeroExtendExpr>(ZExt))
+ return ZExt;
+
+ // Next try a sext cast. If the cast is folded, use it.
+ const SCEV *SExt = getSignExtendExpr(Op, Ty);
+ if (!isa<SCEVSignExtendExpr>(SExt))
+ return SExt;
+
+ // Force the cast to be folded into the operands of an addrec.
+ if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op)) {
+ SmallVector<const SCEV *, 4> Ops;
+ for (SCEVAddRecExpr::op_iterator I = AR->op_begin(), E = AR->op_end();
+ I != E; ++I)
+ Ops.push_back(getAnyExtendExpr(*I, Ty));
+ return getAddRecExpr(Ops, AR->getLoop());
+ }
+
+ // If the expression is obviously signed, use the sext cast value.
+ if (isa<SCEVSMaxExpr>(Op))
+ return SExt;
+
+ // Absent any other information, use the zext cast value.
+ return ZExt;
+}
+
+/// CollectAddOperandsWithScales - Process the given Ops list, which is
+/// a list of operands to be added under the given scale, update the given
+/// map. This is a helper function for getAddRecExpr. As an example of
+/// what it does, given a sequence of operands that would form an add
+/// expression like this:
+///
+/// m + n + 13 + (A * (o + p + (B * q + m + 29))) + r + (-1 * r)
+///
+/// where A and B are constants, update the map with these values:
+///
+/// (m, 1+A*B), (n, 1), (o, A), (p, A), (q, A*B), (r, 0)
+///
+/// and add 13 + A*B*29 to AccumulatedConstant.
+/// This will allow getAddRecExpr to produce this:
+///
+/// 13+A*B*29 + n + (m * (1+A*B)) + ((o + p) * A) + (q * A*B)
+///
+/// This form often exposes folding opportunities that are hidden in
+/// the original operand list.
+///
+/// Return true iff it appears that any interesting folding opportunities
+/// may be exposed. This helps getAddRecExpr short-circuit extra work in
+/// the common case where no interesting opportunities are present, and
+/// is also used as a check to avoid infinite recursion.
+///
+static bool
+CollectAddOperandsWithScales(DenseMap<const SCEV *, APInt> &M,
+ SmallVector<const SCEV *, 8> &NewOps,
+ APInt &AccumulatedConstant,
+ const SmallVectorImpl<const SCEV *> &Ops,
+ const APInt &Scale,
+ ScalarEvolution &SE) {
+ bool Interesting = false;
+
+ // Iterate over the add operands.
+ for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
+ const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[i]);
+ if (Mul && isa<SCEVConstant>(Mul->getOperand(0))) {
+ APInt NewScale =
+ Scale * cast<SCEVConstant>(Mul->getOperand(0))->getValue()->getValue();
+ if (Mul->getNumOperands() == 2 && isa<SCEVAddExpr>(Mul->getOperand(1))) {
+ // A multiplication of a constant with another add; recurse.
+ Interesting |=
+ CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
+ cast<SCEVAddExpr>(Mul->getOperand(1))
+ ->getOperands(),
+ NewScale, SE);
+ } else {
+ // A multiplication of a constant with some other value. Update
+ // the map.
+ SmallVector<const SCEV *, 4> MulOps(Mul->op_begin()+1, Mul->op_end());
+ const SCEV *Key = SE.getMulExpr(MulOps);
+ std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
+ M.insert(std::make_pair(Key, NewScale));
+ if (Pair.second) {
+ NewOps.push_back(Pair.first->first);
+ } else {
+ Pair.first->second += NewScale;
+ // The map already had an entry for this value, which may indicate
+ // a folding opportunity.
+ Interesting = true;
+ }
+ }
+ } else if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
+ // Pull a buried constant out to the outside.
+ if (Scale != 1 || AccumulatedConstant != 0 || C->isZero())
+ Interesting = true;
+ AccumulatedConstant += Scale * C->getValue()->getValue();
+ } else {
+ // An ordinary operand. Update the map.
+ std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
+ M.insert(std::make_pair(Ops[i], Scale));
+ if (Pair.second) {
+ NewOps.push_back(Pair.first->first);
+ } else {
+ Pair.first->second += Scale;
+ // The map already had an entry for this value, which may indicate
+ // a folding opportunity.
+ Interesting = true;
+ }
+ }
+ }
+
+ return Interesting;
+}
+
+namespace {
+ struct APIntCompare {
+ bool operator()(const APInt &LHS, const APInt &RHS) const {
+ return LHS.ult(RHS);
+ }
+ };
+}
+
+/// getAddExpr - Get a canonical add expression, or something simpler if
+/// possible.
+const SCEV *ScalarEvolution::getAddExpr(SmallVectorImpl<const SCEV *> &Ops,
+ bool HasNUW, bool HasNSW) {
+ assert(!Ops.empty() && "Cannot get empty add!");
+ if (Ops.size() == 1) return Ops[0];
+#ifndef NDEBUG
+ for (unsigned i = 1, e = Ops.size(); i != e; ++i)
+ assert(getEffectiveSCEVType(Ops[i]->getType()) ==
+ getEffectiveSCEVType(Ops[0]->getType()) &&
+ "SCEVAddExpr operand types don't match!");
+#endif
+
+ // If HasNSW is true and all the operands are non-negative, infer HasNUW.
+ if (!HasNUW && HasNSW) {
+ bool All = true;
+ for (unsigned i = 0, e = Ops.size(); i != e; ++i)
+ if (!isKnownNonNegative(Ops[i])) {
+ All = false;
+ break;
+ }
+ if (All) HasNUW = true;
+ }
+
+ // Sort by complexity, this groups all similar expression types together.
+ GroupByComplexity(Ops, LI);
+
+ // If there are any constants, fold them together.
+ unsigned Idx = 0;
+ if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
+ ++Idx;
+ assert(Idx < Ops.size());
+ while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
+ // We found two constants, fold them together!
+ Ops[0] = getConstant(LHSC->getValue()->getValue() +
+ RHSC->getValue()->getValue());
+ if (Ops.size() == 2) return Ops[0];
+ Ops.erase(Ops.begin()+1); // Erase the folded element
+ LHSC = cast<SCEVConstant>(Ops[0]);
+ }
+
+ // If we are left with a constant zero being added, strip it off.
+ if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
+ Ops.erase(Ops.begin());
+ --Idx;
+ }
+ }
+
+ if (Ops.size() == 1) return Ops[0];
+
+ // Okay, check to see if the same value occurs in the operand list twice. If
+ // so, merge them together into an multiply expression. Since we sorted the
+ // list, these values are required to be adjacent.
+ const Type *Ty = Ops[0]->getType();
+ for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
+ if (Ops[i] == Ops[i+1]) { // X + Y + Y --> X + Y*2
+ // Found a match, merge the two values into a multiply, and add any
+ // remaining values to the result.
+ const SCEV *Two = getIntegerSCEV(2, Ty);
+ const SCEV *Mul = getMulExpr(Ops[i], Two);
+ if (Ops.size() == 2)
+ return Mul;
+ Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
+ Ops.push_back(Mul);
+ return getAddExpr(Ops, HasNUW, HasNSW);
+ }
+
+ // Check for truncates. If all the operands are truncated from the same
+ // type, see if factoring out the truncate would permit the result to be
+ // folded. eg., trunc(x) + m*trunc(n) --> trunc(x + trunc(m)*n)
+ // if the contents of the resulting outer trunc fold to something simple.
+ for (; Idx < Ops.size() && isa<SCEVTruncateExpr>(Ops[Idx]); ++Idx) {
+ const SCEVTruncateExpr *Trunc = cast<SCEVTruncateExpr>(Ops[Idx]);
+ const Type *DstType = Trunc->getType();
+ const Type *SrcType = Trunc->getOperand()->getType();
+ SmallVector<const SCEV *, 8> LargeOps;
+ bool Ok = true;
+ // Check all the operands to see if they can be represented in the
+ // source type of the truncate.
+ for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
+ if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Ops[i])) {
+ if (T->getOperand()->getType() != SrcType) {
+ Ok = false;
+ break;
+ }
+ LargeOps.push_back(T->getOperand());
+ } else if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
+ // This could be either sign or zero extension, but sign extension
+ // is much more likely to be foldable here.
+ LargeOps.push_back(getSignExtendExpr(C, SrcType));
+ } else if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(Ops[i])) {
+ SmallVector<const SCEV *, 8> LargeMulOps;
+ for (unsigned j = 0, f = M->getNumOperands(); j != f && Ok; ++j) {
+ if (const SCEVTruncateExpr *T =
+ dyn_cast<SCEVTruncateExpr>(M->getOperand(j))) {
+ if (T->getOperand()->getType() != SrcType) {
+ Ok = false;
+ break;
+ }
+ LargeMulOps.push_back(T->getOperand());
+ } else if (const SCEVConstant *C =
+ dyn_cast<SCEVConstant>(M->getOperand(j))) {
+ // This could be either sign or zero extension, but sign extension
+ // is much more likely to be foldable here.
+ LargeMulOps.push_back(getSignExtendExpr(C, SrcType));
+ } else {
+ Ok = false;
+ break;
+ }
+ }
+ if (Ok)
+ LargeOps.push_back(getMulExpr(LargeMulOps));
+ } else {
+ Ok = false;
+ break;
+ }
+ }
+ if (Ok) {
+ // Evaluate the expression in the larger type.
+ const SCEV *Fold = getAddExpr(LargeOps, HasNUW, HasNSW);
+ // If it folds to something simple, use it. Otherwise, don't.
+ if (isa<SCEVConstant>(Fold) || isa<SCEVUnknown>(Fold))
+ return getTruncateExpr(Fold, DstType);
+ }
+ }
+
+ // Skip past any other cast SCEVs.
+ while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddExpr)
+ ++Idx;
+
+ // If there are add operands they would be next.
+ if (Idx < Ops.size()) {
+ bool DeletedAdd = false;
+ while (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[Idx])) {
+ // If we have an add, expand the add operands onto the end of the operands
+ // list.
+ Ops.insert(Ops.end(), Add->op_begin(), Add->op_end());
+ Ops.erase(Ops.begin()+Idx);
+ DeletedAdd = true;
+ }
+
+ // If we deleted at least one add, we added operands to the end of the list,
+ // and they are not necessarily sorted. Recurse to resort and resimplify
+ // any operands we just aquired.
+ if (DeletedAdd)
+ return getAddExpr(Ops);
+ }
+
+ // Skip over the add expression until we get to a multiply.
+ while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
+ ++Idx;
+
+ // Check to see if there are any folding opportunities present with
+ // operands multiplied by constant values.
+ if (Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx])) {
+ uint64_t BitWidth = getTypeSizeInBits(Ty);
+ DenseMap<const SCEV *, APInt> M;
+ SmallVector<const SCEV *, 8> NewOps;
+ APInt AccumulatedConstant(BitWidth, 0);
+ if (CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant,
+ Ops, APInt(BitWidth, 1), *this)) {
+ // Some interesting folding opportunity is present, so its worthwhile to
+ // re-generate the operands list. Group the operands by constant scale,
+ // to avoid multiplying by the same constant scale multiple times.
+ std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare> MulOpLists;
+ for (SmallVector<const SCEV *, 8>::iterator I = NewOps.begin(),
+ E = NewOps.end(); I != E; ++I)
+ MulOpLists[M.find(*I)->second].push_back(*I);
+ // Re-generate the operands list.
+ Ops.clear();
+ if (AccumulatedConstant != 0)
+ Ops.push_back(getConstant(AccumulatedConstant));
+ for (std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare>::iterator
+ I = MulOpLists.begin(), E = MulOpLists.end(); I != E; ++I)
+ if (I->first != 0)
+ Ops.push_back(getMulExpr(getConstant(I->first),
+ getAddExpr(I->second)));
+ if (Ops.empty())
+ return getIntegerSCEV(0, Ty);
+ if (Ops.size() == 1)
+ return Ops[0];
+ return getAddExpr(Ops);
+ }
+ }
+
+ // If we are adding something to a multiply expression, make sure the
+ // something is not already an operand of the multiply. If so, merge it into
+ // the multiply.
+ for (; Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx]); ++Idx) {
+ const SCEVMulExpr *Mul = cast<SCEVMulExpr>(Ops[Idx]);
+ for (unsigned MulOp = 0, e = Mul->getNumOperands(); MulOp != e; ++MulOp) {
+ const SCEV *MulOpSCEV = Mul->getOperand(MulOp);
+ for (unsigned AddOp = 0, e = Ops.size(); AddOp != e; ++AddOp)
+ if (MulOpSCEV == Ops[AddOp] && !isa<SCEVConstant>(Ops[AddOp])) {
+ // Fold W + X + (X * Y * Z) --> W + (X * ((Y*Z)+1))
+ const SCEV *InnerMul = Mul->getOperand(MulOp == 0);
+ if (Mul->getNumOperands() != 2) {
+ // If the multiply has more than two operands, we must get the
+ // Y*Z term.
+ SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(), Mul->op_end());
+ MulOps.erase(MulOps.begin()+MulOp);
+ InnerMul = getMulExpr(MulOps);
+ }
+ const SCEV *One = getIntegerSCEV(1, Ty);
+ const SCEV *AddOne = getAddExpr(InnerMul, One);
+ const SCEV *OuterMul = getMulExpr(AddOne, Ops[AddOp]);
+ if (Ops.size() == 2) return OuterMul;
+ if (AddOp < Idx) {
+ Ops.erase(Ops.begin()+AddOp);
+ Ops.erase(Ops.begin()+Idx-1);
+ } else {
+ Ops.erase(Ops.begin()+Idx);
+ Ops.erase(Ops.begin()+AddOp-1);
+ }
+ Ops.push_back(OuterMul);
+ return getAddExpr(Ops);
+ }
+
+ // Check this multiply against other multiplies being added together.
+ for (unsigned OtherMulIdx = Idx+1;
+ OtherMulIdx < Ops.size() && isa<SCEVMulExpr>(Ops[OtherMulIdx]);
+ ++OtherMulIdx) {
+ const SCEVMulExpr *OtherMul = cast<SCEVMulExpr>(Ops[OtherMulIdx]);
+ // If MulOp occurs in OtherMul, we can fold the two multiplies
+ // together.
+ for (unsigned OMulOp = 0, e = OtherMul->getNumOperands();
+ OMulOp != e; ++OMulOp)
+ if (OtherMul->getOperand(OMulOp) == MulOpSCEV) {
+ // Fold X + (A*B*C) + (A*D*E) --> X + (A*(B*C+D*E))
+ const SCEV *InnerMul1 = Mul->getOperand(MulOp == 0);
+ if (Mul->getNumOperands() != 2) {
+ SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(),
+ Mul->op_end());
+ MulOps.erase(MulOps.begin()+MulOp);
+ InnerMul1 = getMulExpr(MulOps);
+ }
+ const SCEV *InnerMul2 = OtherMul->getOperand(OMulOp == 0);
+ if (OtherMul->getNumOperands() != 2) {
+ SmallVector<const SCEV *, 4> MulOps(OtherMul->op_begin(),
+ OtherMul->op_end());
+ MulOps.erase(MulOps.begin()+OMulOp);
+ InnerMul2 = getMulExpr(MulOps);
+ }
+ const SCEV *InnerMulSum = getAddExpr(InnerMul1,InnerMul2);
+ const SCEV *OuterMul = getMulExpr(MulOpSCEV, InnerMulSum);
+ if (Ops.size() == 2) return OuterMul;
+ Ops.erase(Ops.begin()+Idx);
+ Ops.erase(Ops.begin()+OtherMulIdx-1);
+ Ops.push_back(OuterMul);
+ return getAddExpr(Ops);
+ }
+ }
+ }
+ }
+
+ // If there are any add recurrences in the operands list, see if any other
+ // added values are loop invariant. If so, we can fold them into the
+ // recurrence.
+ while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
+ ++Idx;
+
+ // Scan over all recurrences, trying to fold loop invariants into them.
+ for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
+ // Scan all of the other operands to this add and add them to the vector if
+ // they are loop invariant w.r.t. the recurrence.
+ SmallVector<const SCEV *, 8> LIOps;
+ const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
+ for (unsigned i = 0, e = Ops.size(); i != e; ++i)
+ if (Ops[i]->isLoopInvariant(AddRec->getLoop())) {
+ LIOps.push_back(Ops[i]);
+ Ops.erase(Ops.begin()+i);
+ --i; --e;
+ }
+
+ // If we found some loop invariants, fold them into the recurrence.
+ if (!LIOps.empty()) {
+ // NLI + LI + {Start,+,Step} --> NLI + {LI+Start,+,Step}
+ LIOps.push_back(AddRec->getStart());
+
+ SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(),
+ AddRec->op_end());
+ AddRecOps[0] = getAddExpr(LIOps);
+
+ // It's tempting to propagate NUW/NSW flags here, but nuw/nsw addition
+ // is not associative so this isn't necessarily safe.
+ const SCEV *NewRec = getAddRecExpr(AddRecOps, AddRec->getLoop());
+
+ // If all of the other operands were loop invariant, we are done.
+ if (Ops.size() == 1) return NewRec;
+
+ // Otherwise, add the folded AddRec by the non-liv parts.
+ for (unsigned i = 0;; ++i)
+ if (Ops[i] == AddRec) {
+ Ops[i] = NewRec;
+ break;
+ }
+ return getAddExpr(Ops);
+ }
+
+ // Okay, if there weren't any loop invariants to be folded, check to see if
+ // there are multiple AddRec's with the same loop induction variable being
+ // added together. If so, we can fold them.
+ for (unsigned OtherIdx = Idx+1;
+ OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);++OtherIdx)
+ if (OtherIdx != Idx) {
+ const SCEVAddRecExpr *OtherAddRec = cast<SCEVAddRecExpr>(Ops[OtherIdx]);
+ if (AddRec->getLoop() == OtherAddRec->getLoop()) {
+ // Other + {A,+,B} + {C,+,D} --> Other + {A+C,+,B+D}
+ SmallVector<const SCEV *, 4> NewOps(AddRec->op_begin(),
+ AddRec->op_end());
+ for (unsigned i = 0, e = OtherAddRec->getNumOperands(); i != e; ++i) {
+ if (i >= NewOps.size()) {
+ NewOps.insert(NewOps.end(), OtherAddRec->op_begin()+i,
+ OtherAddRec->op_end());
+ break;
+ }
+ NewOps[i] = getAddExpr(NewOps[i], OtherAddRec->getOperand(i));
+ }
+ const SCEV *NewAddRec = getAddRecExpr(NewOps, AddRec->getLoop());
+
+ if (Ops.size() == 2) return NewAddRec;
+
+ Ops.erase(Ops.begin()+Idx);
+ Ops.erase(Ops.begin()+OtherIdx-1);
+ Ops.push_back(NewAddRec);
+ return getAddExpr(Ops);
+ }
+ }
+
+ // Otherwise couldn't fold anything into this recurrence. Move onto the
+ // next one.
+ }
+
+ // Okay, it looks like we really DO need an add expr. Check to see if we
+ // already have one, otherwise create a new one.
+ FoldingSetNodeID ID;
+ ID.AddInteger(scAddExpr);
+ ID.AddInteger(Ops.size());
+ for (unsigned i = 0, e = Ops.size(); i != e; ++i)
+ ID.AddPointer(Ops[i]);
+ void *IP = 0;
+ SCEVAddExpr *S =
+ static_cast<SCEVAddExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
+ if (!S) {
+ S = SCEVAllocator.Allocate<SCEVAddExpr>();
+ new (S) SCEVAddExpr(ID, Ops);
+ UniqueSCEVs.InsertNode(S, IP);
+ }
+ if (HasNUW) S->setHasNoUnsignedWrap(true);
+ if (HasNSW) S->setHasNoSignedWrap(true);
+ return S;
+}
+
+/// getMulExpr - Get a canonical multiply expression, or something simpler if
+/// possible.
+const SCEV *ScalarEvolution::getMulExpr(SmallVectorImpl<const SCEV *> &Ops,
+ bool HasNUW, bool HasNSW) {
+ assert(!Ops.empty() && "Cannot get empty mul!");
+ if (Ops.size() == 1) return Ops[0];
+#ifndef NDEBUG
+ for (unsigned i = 1, e = Ops.size(); i != e; ++i)
+ assert(getEffectiveSCEVType(Ops[i]->getType()) ==
+ getEffectiveSCEVType(Ops[0]->getType()) &&
+ "SCEVMulExpr operand types don't match!");
+#endif
+
+ // If HasNSW is true and all the operands are non-negative, infer HasNUW.
+ if (!HasNUW && HasNSW) {
+ bool All = true;
+ for (unsigned i = 0, e = Ops.size(); i != e; ++i)
+ if (!isKnownNonNegative(Ops[i])) {
+ All = false;
+ break;
+ }
+ if (All) HasNUW = true;
+ }
+
+ // Sort by complexity, this groups all similar expression types together.
+ GroupByComplexity(Ops, LI);
+
+ // If there are any constants, fold them together.
+ unsigned Idx = 0;
+ if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
+
+ // C1*(C2+V) -> C1*C2 + C1*V
+ if (Ops.size() == 2)
+ if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1]))
+ if (Add->getNumOperands() == 2 &&
+ isa<SCEVConstant>(Add->getOperand(0)))
+ return getAddExpr(getMulExpr(LHSC, Add->getOperand(0)),
+ getMulExpr(LHSC, Add->getOperand(1)));
+
+ ++Idx;
+ while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
+ // We found two constants, fold them together!
+ ConstantInt *Fold = ConstantInt::get(getContext(),
+ LHSC->getValue()->getValue() *
+ RHSC->getValue()->getValue());
+ Ops[0] = getConstant(Fold);
+ Ops.erase(Ops.begin()+1); // Erase the folded element
+ if (Ops.size() == 1) return Ops[0];
+ LHSC = cast<SCEVConstant>(Ops[0]);
+ }
+
+ // If we are left with a constant one being multiplied, strip it off.
+ if (cast<SCEVConstant>(Ops[0])->getValue()->equalsInt(1)) {
+ Ops.erase(Ops.begin());
+ --Idx;
+ } else if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) {
+ // If we have a multiply of zero, it will always be zero.
+ return Ops[0];
+ } else if (Ops[0]->isAllOnesValue()) {
+ // If we have a mul by -1 of an add, try distributing the -1 among the
+ // add operands.
+ if (Ops.size() == 2)
+ if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1])) {
+ SmallVector<const SCEV *, 4> NewOps;
+ bool AnyFolded = false;
+ for (SCEVAddRecExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
+ I != E; ++I) {
+ const SCEV *Mul = getMulExpr(Ops[0], *I);
+ if (!isa<SCEVMulExpr>(Mul)) AnyFolded = true;
+ NewOps.push_back(Mul);
+ }
+ if (AnyFolded)
+ return getAddExpr(NewOps);
+ }
+ }
+ }
+
+ // Skip over the add expression until we get to a multiply.
+ while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr)
+ ++Idx;
+
+ if (Ops.size() == 1)
+ return Ops[0];
+
+ // If there are mul operands inline them all into this expression.
+ if (Idx < Ops.size()) {
+ bool DeletedMul = false;
+ while (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[Idx])) {
+ // If we have an mul, expand the mul operands onto the end of the operands
+ // list.
+ Ops.insert(Ops.end(), Mul->op_begin(), Mul->op_end());
+ Ops.erase(Ops.begin()+Idx);
+ DeletedMul = true;
+ }
+
+ // If we deleted at least one mul, we added operands to the end of the list,
+ // and they are not necessarily sorted. Recurse to resort and resimplify
+ // any operands we just aquired.
+ if (DeletedMul)
+ return getMulExpr(Ops);
+ }
+
+ // If there are any add recurrences in the operands list, see if any other
+ // added values are loop invariant. If so, we can fold them into the
+ // recurrence.
+ while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr)
+ ++Idx;
+
+ // Scan over all recurrences, trying to fold loop invariants into them.
+ for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) {
+ // Scan all of the other operands to this mul and add them to the vector if
+ // they are loop invariant w.r.t. the recurrence.
+ SmallVector<const SCEV *, 8> LIOps;
+ const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]);
+ for (unsigned i = 0, e = Ops.size(); i != e; ++i)
+ if (Ops[i]->isLoopInvariant(AddRec->getLoop())) {
+ LIOps.push_back(Ops[i]);
+ Ops.erase(Ops.begin()+i);
+ --i; --e;
+ }
+
+ // If we found some loop invariants, fold them into the recurrence.
+ if (!LIOps.empty()) {
+ // NLI * LI * {Start,+,Step} --> NLI * {LI*Start,+,LI*Step}
+ SmallVector<const SCEV *, 4> NewOps;
+ NewOps.reserve(AddRec->getNumOperands());
+ if (LIOps.size() == 1) {
+ const SCEV *Scale = LIOps[0];
+ for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
+ NewOps.push_back(getMulExpr(Scale, AddRec->getOperand(i)));
+ } else {
+ for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
+ SmallVector<const SCEV *, 4> MulOps(LIOps.begin(), LIOps.end());
+ MulOps.push_back(AddRec->getOperand(i));
+ NewOps.push_back(getMulExpr(MulOps));
+ }
+ }
+
+ // It's tempting to propagate the NSW flag here, but nsw multiplication
+ // is not associative so this isn't necessarily safe.
+ const SCEV *NewRec = getAddRecExpr(NewOps, AddRec->getLoop(),
+ HasNUW && AddRec->hasNoUnsignedWrap(),
+ /*HasNSW=*/false);
+
+ // If all of the other operands were loop invariant, we are done.
+ if (Ops.size() == 1) return NewRec;
+
+ // Otherwise, multiply the folded AddRec by the non-liv parts.
+ for (unsigned i = 0;; ++i)
+ if (Ops[i] == AddRec) {
+ Ops[i] = NewRec;
+ break;
+ }
+ return getMulExpr(Ops);
+ }
+
+ // Okay, if there weren't any loop invariants to be folded, check to see if
+ // there are multiple AddRec's with the same loop induction variable being
+ // multiplied together. If so, we can fold them.
+ for (unsigned OtherIdx = Idx+1;
+ OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);++OtherIdx)
+ if (OtherIdx != Idx) {
+ const SCEVAddRecExpr *OtherAddRec = cast<SCEVAddRecExpr>(Ops[OtherIdx]);
+ if (AddRec->getLoop() == OtherAddRec->getLoop()) {
+ // F * G --> {A,+,B} * {C,+,D} --> {A*C,+,F*D + G*B + B*D}
+ const SCEVAddRecExpr *F = AddRec, *G = OtherAddRec;
+ const SCEV *NewStart = getMulExpr(F->getStart(),
+ G->getStart());
+ const SCEV *B = F->getStepRecurrence(*this);
+ const SCEV *D = G->getStepRecurrence(*this);
+ const SCEV *NewStep = getAddExpr(getMulExpr(F, D),
+ getMulExpr(G, B),
+ getMulExpr(B, D));
+ const SCEV *NewAddRec = getAddRecExpr(NewStart, NewStep,
+ F->getLoop());
+ if (Ops.size() == 2) return NewAddRec;
+
+ Ops.erase(Ops.begin()+Idx);
+ Ops.erase(Ops.begin()+OtherIdx-1);
+ Ops.push_back(NewAddRec);
+ return getMulExpr(Ops);
+ }
+ }
+
+ // Otherwise couldn't fold anything into this recurrence. Move onto the
+ // next one.
+ }
+
+ // Okay, it looks like we really DO need an mul expr. Check to see if we
+ // already have one, otherwise create a new one.
+ FoldingSetNodeID ID;
+ ID.AddInteger(scMulExpr);
+ ID.AddInteger(Ops.size());
+ for (unsigned i = 0, e = Ops.size(); i != e; ++i)
+ ID.AddPointer(Ops[i]);
+ void *IP = 0;
+ SCEVMulExpr *S =
+ static_cast<SCEVMulExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
+ if (!S) {
+ S = SCEVAllocator.Allocate<SCEVMulExpr>();
+ new (S) SCEVMulExpr(ID, Ops);
+ UniqueSCEVs.InsertNode(S, IP);
+ }
+ if (HasNUW) S->setHasNoUnsignedWrap(true);
+ if (HasNSW) S->setHasNoSignedWrap(true);
+ return S;
+}
+
+/// getUDivExpr - Get a canonical unsigned division expression, or something
+/// simpler if possible.
+const SCEV *ScalarEvolution::getUDivExpr(const SCEV *LHS,
+ const SCEV *RHS) {
+ assert(getEffectiveSCEVType(LHS->getType()) ==
+ getEffectiveSCEVType(RHS->getType()) &&
+ "SCEVUDivExpr operand types don't match!");
+
+ if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
+ if (RHSC->getValue()->equalsInt(1))
+ return LHS; // X udiv 1 --> x
+ if (RHSC->isZero())
+ return getIntegerSCEV(0, LHS->getType()); // value is undefined
+
+ // Determine if the division can be folded into the operands of
+ // its operands.
+ // TODO: Generalize this to non-constants by using known-bits information.
+ const Type *Ty = LHS->getType();
+ unsigned LZ = RHSC->getValue()->getValue().countLeadingZeros();
+ unsigned MaxShiftAmt = getTypeSizeInBits(Ty) - LZ;
+ // For non-power-of-two values, effectively round the value up to the
+ // nearest power of two.
+ if (!RHSC->getValue()->getValue().isPowerOf2())
+ ++MaxShiftAmt;
+ const IntegerType *ExtTy =
+ IntegerType::get(getContext(), getTypeSizeInBits(Ty) + MaxShiftAmt);
+ // {X,+,N}/C --> {X/C,+,N/C} if safe and N/C can be folded.
+ if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
+ if (const SCEVConstant *Step =
+ dyn_cast<SCEVConstant>(AR->getStepRecurrence(*this)))
+ if (!Step->getValue()->getValue()
+ .urem(RHSC->getValue()->getValue()) &&
+ getZeroExtendExpr(AR, ExtTy) ==
+ getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy),
+ getZeroExtendExpr(Step, ExtTy),
+ AR->getLoop())) {
+ SmallVector<const SCEV *, 4> Operands;
+ for (unsigned i = 0, e = AR->getNumOperands(); i != e; ++i)
+ Operands.push_back(getUDivExpr(AR->getOperand(i), RHS));
+ return getAddRecExpr(Operands, AR->getLoop());
+ }
+ // (A*B)/C --> A*(B/C) if safe and B/C can be folded.
+ if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(LHS)) {
+ SmallVector<const SCEV *, 4> Operands;
+ for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i)
+ Operands.push_back(getZeroExtendExpr(M->getOperand(i), ExtTy));
+ if (getZeroExtendExpr(M, ExtTy) == getMulExpr(Operands))
+ // Find an operand that's safely divisible.
+ for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
+ const SCEV *Op = M->getOperand(i);
+ const SCEV *Div = getUDivExpr(Op, RHSC);
+ if (!isa<SCEVUDivExpr>(Div) && getMulExpr(Div, RHSC) == Op) {
+ const SmallVectorImpl<const SCEV *> &MOperands = M->getOperands();
+ Operands = SmallVector<const SCEV *, 4>(MOperands.begin(),
+ MOperands.end());
+ Operands[i] = Div;
+ return getMulExpr(Operands);
+ }
+ }
+ }
+ // (A+B)/C --> (A/C + B/C) if safe and A/C and B/C can be folded.
+ if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(LHS)) {
+ SmallVector<const SCEV *, 4> Operands;
+ for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i)
+ Operands.push_back(getZeroExtendExpr(A->getOperand(i), ExtTy));
+ if (getZeroExtendExpr(A, ExtTy) == getAddExpr(Operands)) {
+ Operands.clear();
+ for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i) {
+ const SCEV *Op = getUDivExpr(A->getOperand(i), RHS);
+ if (isa<SCEVUDivExpr>(Op) || getMulExpr(Op, RHS) != A->getOperand(i))
+ break;
+ Operands.push_back(Op);
+ }
+ if (Operands.size() == A->getNumOperands())
+ return getAddExpr(Operands);
+ }
+ }
+
+ // Fold if both operands are constant.
+ if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
+ Constant *LHSCV = LHSC->getValue();
+ Constant *RHSCV = RHSC->getValue();
+ return getConstant(cast<ConstantInt>(ConstantExpr::getUDiv(LHSCV,
+ RHSCV)));
+ }
+ }
+
+ FoldingSetNodeID ID;
+ ID.AddInteger(scUDivExpr);
+ ID.AddPointer(LHS);
+ ID.AddPointer(RHS);
+ void *IP = 0;
+ if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
+ SCEV *S = SCEVAllocator.Allocate<SCEVUDivExpr>();
+ new (S) SCEVUDivExpr(ID, LHS, RHS);
+ UniqueSCEVs.InsertNode(S, IP);
+ return S;
+}
+
+
+/// getAddRecExpr - Get an add recurrence expression for the specified loop.
+/// Simplify the expression as much as possible.
+const SCEV *ScalarEvolution::getAddRecExpr(const SCEV *Start,
+ const SCEV *Step, const Loop *L,
+ bool HasNUW, bool HasNSW) {
+ SmallVector<const SCEV *, 4> Operands;
+ Operands.push_back(Start);
+ if (const SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step))
+ if (StepChrec->getLoop() == L) {
+ Operands.insert(Operands.end(), StepChrec->op_begin(),
+ StepChrec->op_end());
+ return getAddRecExpr(Operands, L);
+ }
+
+ Operands.push_back(Step);
+ return getAddRecExpr(Operands, L, HasNUW, HasNSW);
+}
+
+/// getAddRecExpr - Get an add recurrence expression for the specified loop.
+/// Simplify the expression as much as possible.
+const SCEV *
+ScalarEvolution::getAddRecExpr(SmallVectorImpl<const SCEV *> &Operands,
+ const Loop *L,
+ bool HasNUW, bool HasNSW) {
+ if (Operands.size() == 1) return Operands[0];
+#ifndef NDEBUG
+ for (unsigned i = 1, e = Operands.size(); i != e; ++i)
+ assert(getEffectiveSCEVType(Operands[i]->getType()) ==
+ getEffectiveSCEVType(Operands[0]->getType()) &&
+ "SCEVAddRecExpr operand types don't match!");
+#endif
+
+ if (Operands.back()->isZero()) {
+ Operands.pop_back();
+ return getAddRecExpr(Operands, L, HasNUW, HasNSW); // {X,+,0} --> X
+ }
+
+ // If HasNSW is true and all the operands are non-negative, infer HasNUW.
+ if (!HasNUW && HasNSW) {
+ bool All = true;
+ for (unsigned i = 0, e = Operands.size(); i != e; ++i)
+ if (!isKnownNonNegative(Operands[i])) {
+ All = false;
+ break;
+ }
+ if (All) HasNUW = true;
+ }
+
+ // Canonicalize nested AddRecs in by nesting them in order of loop depth.
+ if (const SCEVAddRecExpr *NestedAR = dyn_cast<SCEVAddRecExpr>(Operands[0])) {
+ const Loop *NestedLoop = NestedAR->getLoop();
+ if (L->contains(NestedLoop->getHeader()) ?
+ (L->getLoopDepth() < NestedLoop->getLoopDepth()) :
+ (!NestedLoop->contains(L->getHeader()) &&
+ DT->dominates(L->getHeader(), NestedLoop->getHeader()))) {
+ SmallVector<const SCEV *, 4> NestedOperands(NestedAR->op_begin(),
+ NestedAR->op_end());
+ Operands[0] = NestedAR->getStart();
+ // AddRecs require their operands be loop-invariant with respect to their
+ // loops. Don't perform this transformation if it would break this
+ // requirement.
+ bool AllInvariant = true;
+ for (unsigned i = 0, e = Operands.size(); i != e; ++i)
+ if (!Operands[i]->isLoopInvariant(L)) {
+ AllInvariant = false;
+ break;
+ }
+ if (AllInvariant) {
+ NestedOperands[0] = getAddRecExpr(Operands, L);
+ AllInvariant = true;
+ for (unsigned i = 0, e = NestedOperands.size(); i != e; ++i)
+ if (!NestedOperands[i]->isLoopInvariant(NestedLoop)) {
+ AllInvariant = false;
+ break;
+ }
+ if (AllInvariant)
+ // Ok, both add recurrences are valid after the transformation.
+ return getAddRecExpr(NestedOperands, NestedLoop, HasNUW, HasNSW);
+ }
+ // Reset Operands to its original state.
+ Operands[0] = NestedAR;
+ }
+ }
+
+ // Okay, it looks like we really DO need an addrec expr. Check to see if we
+ // already have one, otherwise create a new one.
+ FoldingSetNodeID ID;
+ ID.AddInteger(scAddRecExpr);
+ ID.AddInteger(Operands.size());
+ for (unsigned i = 0, e = Operands.size(); i != e; ++i)
+ ID.AddPointer(Operands[i]);
+ ID.AddPointer(L);
+ void *IP = 0;
+ SCEVAddRecExpr *S =
+ static_cast<SCEVAddRecExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP));
+ if (!S) {
+ S = SCEVAllocator.Allocate<SCEVAddRecExpr>();
+ new (S) SCEVAddRecExpr(ID, Operands, L);
+ UniqueSCEVs.InsertNode(S, IP);
+ }
+ if (HasNUW) S->setHasNoUnsignedWrap(true);
+ if (HasNSW) S->setHasNoSignedWrap(true);
+ return S;
+}
+
+const SCEV *ScalarEvolution::getSMaxExpr(const SCEV *LHS,
+ const SCEV *RHS) {
+ SmallVector<const SCEV *, 2> Ops;
+ Ops.push_back(LHS);
+ Ops.push_back(RHS);
+ return getSMaxExpr(Ops);
+}
+
+const SCEV *
+ScalarEvolution::getSMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
+ assert(!Ops.empty() && "Cannot get empty smax!");
+ if (Ops.size() == 1) return Ops[0];
+#ifndef NDEBUG
+ for (unsigned i = 1, e = Ops.size(); i != e; ++i)
+ assert(getEffectiveSCEVType(Ops[i]->getType()) ==
+ getEffectiveSCEVType(Ops[0]->getType()) &&
+ "SCEVSMaxExpr operand types don't match!");
+#endif
+
+ // Sort by complexity, this groups all similar expression types together.
+ GroupByComplexity(Ops, LI);
+
+ // If there are any constants, fold them together.
+ unsigned Idx = 0;
+ if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
+ ++Idx;
+ assert(Idx < Ops.size());
+ while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
+ // We found two constants, fold them together!
+ ConstantInt *Fold = ConstantInt::get(getContext(),
+ APIntOps::smax(LHSC->getValue()->getValue(),
+ RHSC->getValue()->getValue()));
+ Ops[0] = getConstant(Fold);
+ Ops.erase(Ops.begin()+1); // Erase the folded element
+ if (Ops.size() == 1) return Ops[0];
+ LHSC = cast<SCEVConstant>(Ops[0]);
+ }
+
+ // If we are left with a constant minimum-int, strip it off.
+ if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(true)) {
+ Ops.erase(Ops.begin());
+ --Idx;
+ } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(true)) {
+ // If we have an smax with a constant maximum-int, it will always be
+ // maximum-int.
+ return Ops[0];
+ }
+ }
+
+ if (Ops.size() == 1) return Ops[0];
+
+ // Find the first SMax
+ while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scSMaxExpr)
+ ++Idx;
+
+ // Check to see if one of the operands is an SMax. If so, expand its operands
+ // onto our operand list, and recurse to simplify.
+ if (Idx < Ops.size()) {
+ bool DeletedSMax = false;
+ while (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(Ops[Idx])) {
+ Ops.insert(Ops.end(), SMax->op_begin(), SMax->op_end());
+ Ops.erase(Ops.begin()+Idx);
+ DeletedSMax = true;
+ }
+
+ if (DeletedSMax)
+ return getSMaxExpr(Ops);
+ }
+
+ // Okay, check to see if the same value occurs in the operand list twice. If
+ // so, delete one. Since we sorted the list, these values are required to
+ // be adjacent.
+ for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
+ if (Ops[i] == Ops[i+1]) { // X smax Y smax Y --> X smax Y
+ Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
+ --i; --e;
+ }
+
+ if (Ops.size() == 1) return Ops[0];
+
+ assert(!Ops.empty() && "Reduced smax down to nothing!");
+
+ // Okay, it looks like we really DO need an smax expr. Check to see if we
+ // already have one, otherwise create a new one.
+ FoldingSetNodeID ID;
+ ID.AddInteger(scSMaxExpr);
+ ID.AddInteger(Ops.size());
+ for (unsigned i = 0, e = Ops.size(); i != e; ++i)
+ ID.AddPointer(Ops[i]);
+ void *IP = 0;
+ if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
+ SCEV *S = SCEVAllocator.Allocate<SCEVSMaxExpr>();
+ new (S) SCEVSMaxExpr(ID, Ops);
+ UniqueSCEVs.InsertNode(S, IP);
+ return S;
+}
+
+const SCEV *ScalarEvolution::getUMaxExpr(const SCEV *LHS,
+ const SCEV *RHS) {
+ SmallVector<const SCEV *, 2> Ops;
+ Ops.push_back(LHS);
+ Ops.push_back(RHS);
+ return getUMaxExpr(Ops);
+}
+
+const SCEV *
+ScalarEvolution::getUMaxExpr(SmallVectorImpl<const SCEV *> &Ops) {
+ assert(!Ops.empty() && "Cannot get empty umax!");
+ if (Ops.size() == 1) return Ops[0];
+#ifndef NDEBUG
+ for (unsigned i = 1, e = Ops.size(); i != e; ++i)
+ assert(getEffectiveSCEVType(Ops[i]->getType()) ==
+ getEffectiveSCEVType(Ops[0]->getType()) &&
+ "SCEVUMaxExpr operand types don't match!");
+#endif
+
+ // Sort by complexity, this groups all similar expression types together.
+ GroupByComplexity(Ops, LI);
+
+ // If there are any constants, fold them together.
+ unsigned Idx = 0;
+ if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
+ ++Idx;
+ assert(Idx < Ops.size());
+ while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
+ // We found two constants, fold them together!
+ ConstantInt *Fold = ConstantInt::get(getContext(),
+ APIntOps::umax(LHSC->getValue()->getValue(),
+ RHSC->getValue()->getValue()));
+ Ops[0] = getConstant(Fold);
+ Ops.erase(Ops.begin()+1); // Erase the folded element
+ if (Ops.size() == 1) return Ops[0];
+ LHSC = cast<SCEVConstant>(Ops[0]);
+ }
+
+ // If we are left with a constant minimum-int, strip it off.
+ if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(false)) {
+ Ops.erase(Ops.begin());
+ --Idx;
+ } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(false)) {
+ // If we have an umax with a constant maximum-int, it will always be
+ // maximum-int.
+ return Ops[0];
+ }
+ }
+
+ if (Ops.size() == 1) return Ops[0];
+
+ // Find the first UMax
+ while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scUMaxExpr)
+ ++Idx;
+
+ // Check to see if one of the operands is a UMax. If so, expand its operands
+ // onto our operand list, and recurse to simplify.
+ if (Idx < Ops.size()) {
+ bool DeletedUMax = false;
+ while (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(Ops[Idx])) {
+ Ops.insert(Ops.end(), UMax->op_begin(), UMax->op_end());
+ Ops.erase(Ops.begin()+Idx);
+ DeletedUMax = true;
+ }
+
+ if (DeletedUMax)
+ return getUMaxExpr(Ops);
+ }
+
+ // Okay, check to see if the same value occurs in the operand list twice. If
+ // so, delete one. Since we sorted the list, these values are required to
+ // be adjacent.
+ for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
+ if (Ops[i] == Ops[i+1]) { // X umax Y umax Y --> X umax Y
+ Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
+ --i; --e;
+ }
+
+ if (Ops.size() == 1) return Ops[0];
+
+ assert(!Ops.empty() && "Reduced umax down to nothing!");
+
+ // Okay, it looks like we really DO need a umax expr. Check to see if we
+ // already have one, otherwise create a new one.
+ FoldingSetNodeID ID;
+ ID.AddInteger(scUMaxExpr);
+ ID.AddInteger(Ops.size());
+ for (unsigned i = 0, e = Ops.size(); i != e; ++i)
+ ID.AddPointer(Ops[i]);
+ void *IP = 0;
+ if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
+ SCEV *S = SCEVAllocator.Allocate<SCEVUMaxExpr>();
+ new (S) SCEVUMaxExpr(ID, Ops);
+ UniqueSCEVs.InsertNode(S, IP);
+ return S;
+}
+
+const SCEV *ScalarEvolution::getSMinExpr(const SCEV *LHS,
+ const SCEV *RHS) {
+ // ~smax(~x, ~y) == smin(x, y).
+ return getNotSCEV(getSMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
+}
+
+const SCEV *ScalarEvolution::getUMinExpr(const SCEV *LHS,
+ const SCEV *RHS) {
+ // ~umax(~x, ~y) == umin(x, y)
+ return getNotSCEV(getUMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
+}
+
+const SCEV *ScalarEvolution::getSizeOfExpr(const Type *AllocTy) {
+ Constant *C = ConstantExpr::getSizeOf(AllocTy);
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
+ C = ConstantFoldConstantExpression(CE, TD);
+ const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
+ return getTruncateOrZeroExtend(getSCEV(C), Ty);
+}
+
+const SCEV *ScalarEvolution::getAlignOfExpr(const Type *AllocTy) {
+ Constant *C = ConstantExpr::getAlignOf(AllocTy);
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
+ C = ConstantFoldConstantExpression(CE, TD);
+ const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
+ return getTruncateOrZeroExtend(getSCEV(C), Ty);
+}
+
+const SCEV *ScalarEvolution::getOffsetOfExpr(const StructType *STy,
+ unsigned FieldNo) {
+ Constant *C = ConstantExpr::getOffsetOf(STy, FieldNo);
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
+ C = ConstantFoldConstantExpression(CE, TD);
+ const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(STy));
+ return getTruncateOrZeroExtend(getSCEV(C), Ty);
+}
+
+const SCEV *ScalarEvolution::getOffsetOfExpr(const Type *CTy,
+ Constant *FieldNo) {
+ Constant *C = ConstantExpr::getOffsetOf(CTy, FieldNo);
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
+ C = ConstantFoldConstantExpression(CE, TD);
+ const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(CTy));
+ return getTruncateOrZeroExtend(getSCEV(C), Ty);
+}
+
+const SCEV *ScalarEvolution::getUnknown(Value *V) {
+ // Don't attempt to do anything other than create a SCEVUnknown object
+ // here. createSCEV only calls getUnknown after checking for all other
+ // interesting possibilities, and any other code that calls getUnknown
+ // is doing so in order to hide a value from SCEV canonicalization.
+
+ FoldingSetNodeID ID;
+ ID.AddInteger(scUnknown);
+ ID.AddPointer(V);
+ void *IP = 0;
+ if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
+ SCEV *S = SCEVAllocator.Allocate<SCEVUnknown>();
+ new (S) SCEVUnknown(ID, V);
+ UniqueSCEVs.InsertNode(S, IP);
+ return S;
+}
+
+//===----------------------------------------------------------------------===//
+// Basic SCEV Analysis and PHI Idiom Recognition Code
+//
+
+/// isSCEVable - Test if values of the given type are analyzable within
+/// the SCEV framework. This primarily includes integer types, and it
+/// can optionally include pointer types if the ScalarEvolution class
+/// has access to target-specific information.
+bool ScalarEvolution::isSCEVable(const Type *Ty) const {
+ // Integers and pointers are always SCEVable.
+ return Ty->isInteger() || isa<PointerType>(Ty);
+}
+
+/// getTypeSizeInBits - Return the size in bits of the specified type,
+/// for which isSCEVable must return true.
+uint64_t ScalarEvolution::getTypeSizeInBits(const Type *Ty) const {
+ assert(isSCEVable(Ty) && "Type is not SCEVable!");
+
+ // If we have a TargetData, use it!
+ if (TD)
+ return TD->getTypeSizeInBits(Ty);
+
+ // Integer types have fixed sizes.
+ if (Ty->isInteger())
+ return Ty->getPrimitiveSizeInBits();
+
+ // The only other support type is pointer. Without TargetData, conservatively
+ // assume pointers are 64-bit.
+ assert(isa<PointerType>(Ty) && "isSCEVable permitted a non-SCEVable type!");
+ return 64;
+}
+
+/// getEffectiveSCEVType - Return a type with the same bitwidth as
+/// the given type and which represents how SCEV will treat the given
+/// type, for which isSCEVable must return true. For pointer types,
+/// this is the pointer-sized integer type.
+const Type *ScalarEvolution::getEffectiveSCEVType(const Type *Ty) const {
+ assert(isSCEVable(Ty) && "Type is not SCEVable!");
+
+ if (Ty->isInteger())
+ return Ty;
+
+ // The only other support type is pointer.
+ assert(isa<PointerType>(Ty) && "Unexpected non-pointer non-integer type!");
+ if (TD) return TD->getIntPtrType(getContext());
+
+ // Without TargetData, conservatively assume pointers are 64-bit.
+ return Type::getInt64Ty(getContext());
+}
+
+const SCEV *ScalarEvolution::getCouldNotCompute() {
+ return &CouldNotCompute;
+}
+
+/// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
+/// expression and create a new one.
+const SCEV *ScalarEvolution::getSCEV(Value *V) {
+ assert(isSCEVable(V->getType()) && "Value is not SCEVable!");
+
+ std::map<SCEVCallbackVH, const SCEV *>::iterator I = Scalars.find(V);
+ if (I != Scalars.end()) return I->second;
+ const SCEV *S = createSCEV(V);
+ Scalars.insert(std::make_pair(SCEVCallbackVH(V, this), S));
+ return S;
+}
+
+/// getIntegerSCEV - Given a SCEVable type, create a constant for the
+/// specified signed integer value and return a SCEV for the constant.
+const SCEV *ScalarEvolution::getIntegerSCEV(int64_t Val, const Type *Ty) {
+ const IntegerType *ITy = cast<IntegerType>(getEffectiveSCEVType(Ty));
+ return getConstant(ConstantInt::get(ITy, Val));
+}
+
+/// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V
+///
+const SCEV *ScalarEvolution::getNegativeSCEV(const SCEV *V) {
+ if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
+ return getConstant(
+ cast<ConstantInt>(ConstantExpr::getNeg(VC->getValue())));
+
+ const Type *Ty = V->getType();
+ Ty = getEffectiveSCEVType(Ty);
+ return getMulExpr(V,
+ getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty))));
+}
+
+/// getNotSCEV - Return a SCEV corresponding to ~V = -1-V
+const SCEV *ScalarEvolution::getNotSCEV(const SCEV *V) {
+ if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
+ return getConstant(
+ cast<ConstantInt>(ConstantExpr::getNot(VC->getValue())));
+
+ const Type *Ty = V->getType();
+ Ty = getEffectiveSCEVType(Ty);
+ const SCEV *AllOnes =
+ getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty)));
+ return getMinusSCEV(AllOnes, V);
+}
+
+/// getMinusSCEV - Return a SCEV corresponding to LHS - RHS.
+///
+const SCEV *ScalarEvolution::getMinusSCEV(const SCEV *LHS,
+ const SCEV *RHS) {
+ // X - Y --> X + -Y
+ return getAddExpr(LHS, getNegativeSCEV(RHS));
+}
+
+/// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion of the
+/// input value to the specified type. If the type must be extended, it is zero
+/// extended.
+const SCEV *
+ScalarEvolution::getTruncateOrZeroExtend(const SCEV *V,
+ const Type *Ty) {
+ const Type *SrcTy = V->getType();
+ assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
+ (Ty->isInteger() || isa<PointerType>(Ty)) &&
+ "Cannot truncate or zero extend with non-integer arguments!");
+ if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
+ return V; // No conversion
+ if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
+ return getTruncateExpr(V, Ty);
+ return getZeroExtendExpr(V, Ty);
+}
+
+/// getTruncateOrSignExtend - Return a SCEV corresponding to a conversion of the
+/// input value to the specified type. If the type must be extended, it is sign
+/// extended.
+const SCEV *
+ScalarEvolution::getTruncateOrSignExtend(const SCEV *V,
+ const Type *Ty) {
+ const Type *SrcTy = V->getType();
+ assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
+ (Ty->isInteger() || isa<PointerType>(Ty)) &&
+ "Cannot truncate or zero extend with non-integer arguments!");
+ if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
+ return V; // No conversion
+ if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty))
+ return getTruncateExpr(V, Ty);
+ return getSignExtendExpr(V, Ty);
+}
+
+/// getNoopOrZeroExtend - Return a SCEV corresponding to a conversion of the
+/// input value to the specified type. If the type must be extended, it is zero
+/// extended. The conversion must not be narrowing.
+const SCEV *
+ScalarEvolution::getNoopOrZeroExtend(const SCEV *V, const Type *Ty) {
+ const Type *SrcTy = V->getType();
+ assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
+ (Ty->isInteger() || isa<PointerType>(Ty)) &&
+ "Cannot noop or zero extend with non-integer arguments!");
+ assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
+ "getNoopOrZeroExtend cannot truncate!");
+ if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
+ return V; // No conversion
+ return getZeroExtendExpr(V, Ty);
+}
+
+/// getNoopOrSignExtend - Return a SCEV corresponding to a conversion of the
+/// input value to the specified type. If the type must be extended, it is sign
+/// extended. The conversion must not be narrowing.
+const SCEV *
+ScalarEvolution::getNoopOrSignExtend(const SCEV *V, const Type *Ty) {
+ const Type *SrcTy = V->getType();
+ assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
+ (Ty->isInteger() || isa<PointerType>(Ty)) &&
+ "Cannot noop or sign extend with non-integer arguments!");
+ assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
+ "getNoopOrSignExtend cannot truncate!");
+ if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
+ return V; // No conversion
+ return getSignExtendExpr(V, Ty);
+}
+
+/// getNoopOrAnyExtend - Return a SCEV corresponding to a conversion of
+/// the input value to the specified type. If the type must be extended,
+/// it is extended with unspecified bits. The conversion must not be
+/// narrowing.
+const SCEV *
+ScalarEvolution::getNoopOrAnyExtend(const SCEV *V, const Type *Ty) {
+ const Type *SrcTy = V->getType();
+ assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
+ (Ty->isInteger() || isa<PointerType>(Ty)) &&
+ "Cannot noop or any extend with non-integer arguments!");
+ assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) &&
+ "getNoopOrAnyExtend cannot truncate!");
+ if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
+ return V; // No conversion
+ return getAnyExtendExpr(V, Ty);
+}
+
+/// getTruncateOrNoop - Return a SCEV corresponding to a conversion of the
+/// input value to the specified type. The conversion must not be widening.
+const SCEV *
+ScalarEvolution::getTruncateOrNoop(const SCEV *V, const Type *Ty) {
+ const Type *SrcTy = V->getType();
+ assert((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
+ (Ty->isInteger() || isa<PointerType>(Ty)) &&
+ "Cannot truncate or noop with non-integer arguments!");
+ assert(getTypeSizeInBits(SrcTy) >= getTypeSizeInBits(Ty) &&
+ "getTruncateOrNoop cannot extend!");
+ if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty))
+ return V; // No conversion
+ return getTruncateExpr(V, Ty);
+}
+
+/// getUMaxFromMismatchedTypes - Promote the operands to the wider of
+/// the types using zero-extension, and then perform a umax operation
+/// with them.
+const SCEV *ScalarEvolution::getUMaxFromMismatchedTypes(const SCEV *LHS,
+ const SCEV *RHS) {
+ const SCEV *PromotedLHS = LHS;
+ const SCEV *PromotedRHS = RHS;
+
+ if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
+ PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
+ else
+ PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
+
+ return getUMaxExpr(PromotedLHS, PromotedRHS);
+}
+
+/// getUMinFromMismatchedTypes - Promote the operands to the wider of
+/// the types using zero-extension, and then perform a umin operation
+/// with them.
+const SCEV *ScalarEvolution::getUMinFromMismatchedTypes(const SCEV *LHS,
+ const SCEV *RHS) {
+ const SCEV *PromotedLHS = LHS;
+ const SCEV *PromotedRHS = RHS;
+
+ if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType()))
+ PromotedRHS = getZeroExtendExpr(RHS, LHS->getType());
+ else
+ PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType());
+
+ return getUMinExpr(PromotedLHS, PromotedRHS);
+}
+
+/// PushDefUseChildren - Push users of the given Instruction
+/// onto the given Worklist.
+static void
+PushDefUseChildren(Instruction *I,
+ SmallVectorImpl<Instruction *> &Worklist) {
+ // Push the def-use children onto the Worklist stack.
+ for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
+ UI != UE; ++UI)
+ Worklist.push_back(cast<Instruction>(UI));
+}
+
+/// ForgetSymbolicValue - This looks up computed SCEV values for all
+/// instructions that depend on the given instruction and removes them from
+/// the Scalars map if they reference SymName. This is used during PHI
+/// resolution.
+void
+ScalarEvolution::ForgetSymbolicName(Instruction *I, const SCEV *SymName) {
+ SmallVector<Instruction *, 16> Worklist;
+ PushDefUseChildren(I, Worklist);
+
+ SmallPtrSet<Instruction *, 8> Visited;
+ Visited.insert(I);
+ while (!Worklist.empty()) {
+ Instruction *I = Worklist.pop_back_val();
+ if (!Visited.insert(I)) continue;
+
+ std::map<SCEVCallbackVH, const SCEV *>::iterator It =
+ Scalars.find(static_cast<Value *>(I));
+ if (It != Scalars.end()) {
+ // Short-circuit the def-use traversal if the symbolic name
+ // ceases to appear in expressions.
+ if (!It->second->hasOperand(SymName))
+ continue;
+
+ // SCEVUnknown for a PHI either means that it has an unrecognized
+ // structure, or it's a PHI that's in the progress of being computed
+ // by createNodeForPHI. In the former case, additional loop trip
+ // count information isn't going to change anything. In the later
+ // case, createNodeForPHI will perform the necessary updates on its
+ // own when it gets to that point.
+ if (!isa<PHINode>(I) || !isa<SCEVUnknown>(It->second)) {
+ ValuesAtScopes.erase(It->second);
+ Scalars.erase(It);
+ }
+ }
+
+ PushDefUseChildren(I, Worklist);
+ }
+}
+
+/// createNodeForPHI - PHI nodes have two cases. Either the PHI node exists in
+/// a loop header, making it a potential recurrence, or it doesn't.
+///
+const SCEV *ScalarEvolution::createNodeForPHI(PHINode *PN) {
+ if (PN->getNumIncomingValues() == 2) // The loops have been canonicalized.
+ if (const Loop *L = LI->getLoopFor(PN->getParent()))
+ if (L->getHeader() == PN->getParent()) {
+ // If it lives in the loop header, it has two incoming values, one
+ // from outside the loop, and one from inside.
+ unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
+ unsigned BackEdge = IncomingEdge^1;
+
+ // While we are analyzing this PHI node, handle its value symbolically.
+ const SCEV *SymbolicName = getUnknown(PN);
+ assert(Scalars.find(PN) == Scalars.end() &&
+ "PHI node already processed?");
+ Scalars.insert(std::make_pair(SCEVCallbackVH(PN, this), SymbolicName));
+
+ // Using this symbolic name for the PHI, analyze the value coming around
+ // the back-edge.
+ Value *BEValueV = PN->getIncomingValue(BackEdge);
+ const SCEV *BEValue = getSCEV(BEValueV);
+
+ // NOTE: If BEValue is loop invariant, we know that the PHI node just
+ // has a special value for the first iteration of the loop.
+
+ // If the value coming around the backedge is an add with the symbolic
+ // value we just inserted, then we found a simple induction variable!
+ if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(BEValue)) {
+ // If there is a single occurrence of the symbolic value, replace it
+ // with a recurrence.
+ unsigned FoundIndex = Add->getNumOperands();
+ for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
+ if (Add->getOperand(i) == SymbolicName)
+ if (FoundIndex == e) {
+ FoundIndex = i;
+ break;
+ }
+
+ if (FoundIndex != Add->getNumOperands()) {
+ // Create an add with everything but the specified operand.
+ SmallVector<const SCEV *, 8> Ops;
+ for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
+ if (i != FoundIndex)
+ Ops.push_back(Add->getOperand(i));
+ const SCEV *Accum = getAddExpr(Ops);
+
+ // This is not a valid addrec if the step amount is varying each
+ // loop iteration, but is not itself an addrec in this loop.
+ if (Accum->isLoopInvariant(L) ||
+ (isa<SCEVAddRecExpr>(Accum) &&
+ cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) {
+ bool HasNUW = false;
+ bool HasNSW = false;
+
+ // If the increment doesn't overflow, then neither the addrec nor
+ // the post-increment will overflow.
+ if (const AddOperator *OBO = dyn_cast<AddOperator>(BEValueV)) {
+ if (OBO->hasNoUnsignedWrap())
+ HasNUW = true;
+ if (OBO->hasNoSignedWrap())
+ HasNSW = true;
+ }
+
+ const SCEV *StartVal =
+ getSCEV(PN->getIncomingValue(IncomingEdge));
+ const SCEV *PHISCEV =
+ getAddRecExpr(StartVal, Accum, L, HasNUW, HasNSW);
+
+ // Since the no-wrap flags are on the increment, they apply to the
+ // post-incremented value as well.
+ if (Accum->isLoopInvariant(L))
+ (void)getAddRecExpr(getAddExpr(StartVal, Accum),
+ Accum, L, HasNUW, HasNSW);
+
+ // Okay, for the entire analysis of this edge we assumed the PHI
+ // to be symbolic. We now need to go back and purge all of the
+ // entries for the scalars that use the symbolic expression.
+ ForgetSymbolicName(PN, SymbolicName);
+ Scalars[SCEVCallbackVH(PN, this)] = PHISCEV;
+ return PHISCEV;
+ }
+ }
+ } else if (const SCEVAddRecExpr *AddRec =
+ dyn_cast<SCEVAddRecExpr>(BEValue)) {
+ // Otherwise, this could be a loop like this:
+ // i = 0; for (j = 1; ..; ++j) { .... i = j; }
+ // In this case, j = {1,+,1} and BEValue is j.
+ // Because the other in-value of i (0) fits the evolution of BEValue
+ // i really is an addrec evolution.
+ if (AddRec->getLoop() == L && AddRec->isAffine()) {
+ const SCEV *StartVal = getSCEV(PN->getIncomingValue(IncomingEdge));
+
+ // If StartVal = j.start - j.stride, we can use StartVal as the
+ // initial step of the addrec evolution.
+ if (StartVal == getMinusSCEV(AddRec->getOperand(0),
+ AddRec->getOperand(1))) {
+ const SCEV *PHISCEV =
+ getAddRecExpr(StartVal, AddRec->getOperand(1), L);
+
+ // Okay, for the entire analysis of this edge we assumed the PHI
+ // to be symbolic. We now need to go back and purge all of the
+ // entries for the scalars that use the symbolic expression.
+ ForgetSymbolicName(PN, SymbolicName);
+ Scalars[SCEVCallbackVH(PN, this)] = PHISCEV;
+ return PHISCEV;
+ }
+ }
+ }
+
+ return SymbolicName;
+ }
+
+ // It's tempting to recognize PHIs with a unique incoming value, however
+ // this leads passes like indvars to break LCSSA form. Fortunately, such
+ // PHIs are rare, as instcombine zaps them.
+
+ // If it's not a loop phi, we can't handle it yet.
+ return getUnknown(PN);
+}
+
+/// createNodeForGEP - Expand GEP instructions into add and multiply
+/// operations. This allows them to be analyzed by regular SCEV code.
+///
+const SCEV *ScalarEvolution::createNodeForGEP(GEPOperator *GEP) {
+
+ bool InBounds = GEP->isInBounds();
+ const Type *IntPtrTy = getEffectiveSCEVType(GEP->getType());
+ Value *Base = GEP->getOperand(0);
+ // Don't attempt to analyze GEPs over unsized objects.
+ if (!cast<PointerType>(Base->getType())->getElementType()->isSized())
+ return getUnknown(GEP);
+ const SCEV *TotalOffset = getIntegerSCEV(0, IntPtrTy);
+ gep_type_iterator GTI = gep_type_begin(GEP);
+ for (GetElementPtrInst::op_iterator I = next(GEP->op_begin()),
+ E = GEP->op_end();
+ I != E; ++I) {
+ Value *Index = *I;
+ // Compute the (potentially symbolic) offset in bytes for this index.
+ if (const StructType *STy = dyn_cast<StructType>(*GTI++)) {
+ // For a struct, add the member offset.
+ unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
+ TotalOffset = getAddExpr(TotalOffset,
+ getOffsetOfExpr(STy, FieldNo),
+ /*HasNUW=*/false, /*HasNSW=*/InBounds);
+ } else {
+ // For an array, add the element offset, explicitly scaled.
+ const SCEV *LocalOffset = getSCEV(Index);
+ // Getelementptr indicies are signed.
+ LocalOffset = getTruncateOrSignExtend(LocalOffset, IntPtrTy);
+ // Lower "inbounds" GEPs to NSW arithmetic.
+ LocalOffset = getMulExpr(LocalOffset, getSizeOfExpr(*GTI),
+ /*HasNUW=*/false, /*HasNSW=*/InBounds);
+ TotalOffset = getAddExpr(TotalOffset, LocalOffset,
+ /*HasNUW=*/false, /*HasNSW=*/InBounds);
+ }
+ }
+ return getAddExpr(getSCEV(Base), TotalOffset,
+ /*HasNUW=*/false, /*HasNSW=*/InBounds);
+}
+
+/// GetMinTrailingZeros - Determine the minimum number of zero bits that S is
+/// guaranteed to end in (at every loop iteration). It is, at the same time,
+/// the minimum number of times S is divisible by 2. For example, given {4,+,8}
+/// it returns 2. If S is guaranteed to be 0, it returns the bitwidth of S.
+uint32_t
+ScalarEvolution::GetMinTrailingZeros(const SCEV *S) {
+ if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
+ return C->getValue()->getValue().countTrailingZeros();
+
+ if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(S))
+ return std::min(GetMinTrailingZeros(T->getOperand()),
+ (uint32_t)getTypeSizeInBits(T->getType()));
+
+ if (const SCEVZeroExtendExpr *E = dyn_cast<SCEVZeroExtendExpr>(S)) {
+ uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
+ return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
+ getTypeSizeInBits(E->getType()) : OpRes;
+ }
+
+ if (const SCEVSignExtendExpr *E = dyn_cast<SCEVSignExtendExpr>(S)) {
+ uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
+ return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ?
+ getTypeSizeInBits(E->getType()) : OpRes;
+ }
+
+ if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(S)) {
+ // The result is the min of all operands results.
+ uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
+ for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
+ MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
+ return MinOpRes;
+ }
+
+ if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
+ // The result is the sum of all operands results.
+ uint32_t SumOpRes = GetMinTrailingZeros(M->getOperand(0));
+ uint32_t BitWidth = getTypeSizeInBits(M->getType());
+ for (unsigned i = 1, e = M->getNumOperands();
+ SumOpRes != BitWidth && i != e; ++i)
+ SumOpRes = std::min(SumOpRes + GetMinTrailingZeros(M->getOperand(i)),
+ BitWidth);
+ return SumOpRes;
+ }
+
+ if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
+ // The result is the min of all operands results.
+ uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
+ for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
+ MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
+ return MinOpRes;
+ }
+
+ if (const SCEVSMaxExpr *M = dyn_cast<SCEVSMaxExpr>(S)) {
+ // The result is the min of all operands results.
+ uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
+ for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
+ MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
+ return MinOpRes;
+ }
+
+ if (const SCEVUMaxExpr *M = dyn_cast<SCEVUMaxExpr>(S)) {
+ // The result is the min of all operands results.
+ uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
+ for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
+ MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
+ return MinOpRes;
+ }
+
+ if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
+ // For a SCEVUnknown, ask ValueTracking.
+ unsigned BitWidth = getTypeSizeInBits(U->getType());
+ APInt Mask = APInt::getAllOnesValue(BitWidth);
+ APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
+ ComputeMaskedBits(U->getValue(), Mask, Zeros, Ones);
+ return Zeros.countTrailingOnes();
+ }
+
+ // SCEVUDivExpr
+ return 0;
+}
+
+/// getUnsignedRange - Determine the unsigned range for a particular SCEV.
+///
+ConstantRange
+ScalarEvolution::getUnsignedRange(const SCEV *S) {
+
+ if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
+ return ConstantRange(C->getValue()->getValue());
+
+ unsigned BitWidth = getTypeSizeInBits(S->getType());
+ ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
+
+ // If the value has known zeros, the maximum unsigned value will have those
+ // known zeros as well.
+ uint32_t TZ = GetMinTrailingZeros(S);
+ if (TZ != 0)
+ ConservativeResult =
+ ConstantRange(APInt::getMinValue(BitWidth),
+ APInt::getMaxValue(BitWidth).lshr(TZ).shl(TZ) + 1);
+
+ if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
+ ConstantRange X = getUnsignedRange(Add->getOperand(0));
+ for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
+ X = X.add(getUnsignedRange(Add->getOperand(i)));
+ return ConservativeResult.intersectWith(X);
+ }
+
+ if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
+ ConstantRange X = getUnsignedRange(Mul->getOperand(0));
+ for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
+ X = X.multiply(getUnsignedRange(Mul->getOperand(i)));
+ return ConservativeResult.intersectWith(X);
+ }
+
+ if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
+ ConstantRange X = getUnsignedRange(SMax->getOperand(0));
+ for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
+ X = X.smax(getUnsignedRange(SMax->getOperand(i)));
+ return ConservativeResult.intersectWith(X);
+ }
+
+ if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
+ ConstantRange X = getUnsignedRange(UMax->getOperand(0));
+ for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
+ X = X.umax(getUnsignedRange(UMax->getOperand(i)));
+ return ConservativeResult.intersectWith(X);
+ }
+
+ if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
+ ConstantRange X = getUnsignedRange(UDiv->getLHS());
+ ConstantRange Y = getUnsignedRange(UDiv->getRHS());
+ return ConservativeResult.intersectWith(X.udiv(Y));
+ }
+
+ if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
+ ConstantRange X = getUnsignedRange(ZExt->getOperand());
+ return ConservativeResult.intersectWith(X.zeroExtend(BitWidth));
+ }
+
+ if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
+ ConstantRange X = getUnsignedRange(SExt->getOperand());
+ return ConservativeResult.intersectWith(X.signExtend(BitWidth));
+ }
+
+ if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
+ ConstantRange X = getUnsignedRange(Trunc->getOperand());
+ return ConservativeResult.intersectWith(X.truncate(BitWidth));
+ }
+
+ if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
+ // If there's no unsigned wrap, the value will never be less than its
+ // initial value.
+ if (AddRec->hasNoUnsignedWrap())
+ if (const SCEVConstant *C = dyn_cast<SCEVConstant>(AddRec->getStart()))
+ ConservativeResult =
+ ConstantRange(C->getValue()->getValue(),
+ APInt(getTypeSizeInBits(C->getType()), 0));
+
+ // TODO: non-affine addrec
+ if (AddRec->isAffine()) {
+ const Type *Ty = AddRec->getType();
+ const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
+ if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
+ getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
+ MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
+
+ const SCEV *Start = AddRec->getStart();
+ const SCEV *End = AddRec->evaluateAtIteration(MaxBECount, *this);
+
+ // Check for overflow.
+ if (!AddRec->hasNoUnsignedWrap())
+ return ConservativeResult;
+
+ ConstantRange StartRange = getUnsignedRange(Start);
+ ConstantRange EndRange = getUnsignedRange(End);
+ APInt Min = APIntOps::umin(StartRange.getUnsignedMin(),
+ EndRange.getUnsignedMin());
+ APInt Max = APIntOps::umax(StartRange.getUnsignedMax(),
+ EndRange.getUnsignedMax());
+ if (Min.isMinValue() && Max.isMaxValue())
+ return ConservativeResult;
+ return ConservativeResult.intersectWith(ConstantRange(Min, Max+1));
+ }
+ }
+
+ return ConservativeResult;
+ }
+
+ if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
+ // For a SCEVUnknown, ask ValueTracking.
+ unsigned BitWidth = getTypeSizeInBits(U->getType());
+ APInt Mask = APInt::getAllOnesValue(BitWidth);
+ APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
+ ComputeMaskedBits(U->getValue(), Mask, Zeros, Ones, TD);
+ if (Ones == ~Zeros + 1)
+ return ConservativeResult;
+ return ConservativeResult.intersectWith(ConstantRange(Ones, ~Zeros + 1));
+ }
+
+ return ConservativeResult;
+}
+
+/// getSignedRange - Determine the signed range for a particular SCEV.
+///
+ConstantRange
+ScalarEvolution::getSignedRange(const SCEV *S) {
+
+ if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
+ return ConstantRange(C->getValue()->getValue());
+
+ unsigned BitWidth = getTypeSizeInBits(S->getType());
+ ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true);
+
+ // If the value has known zeros, the maximum signed value will have those
+ // known zeros as well.
+ uint32_t TZ = GetMinTrailingZeros(S);
+ if (TZ != 0)
+ ConservativeResult =
+ ConstantRange(APInt::getSignedMinValue(BitWidth),
+ APInt::getSignedMaxValue(BitWidth).ashr(TZ).shl(TZ) + 1);
+
+ if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
+ ConstantRange X = getSignedRange(Add->getOperand(0));
+ for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i)
+ X = X.add(getSignedRange(Add->getOperand(i)));
+ return ConservativeResult.intersectWith(X);
+ }
+
+ if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
+ ConstantRange X = getSignedRange(Mul->getOperand(0));
+ for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i)
+ X = X.multiply(getSignedRange(Mul->getOperand(i)));
+ return ConservativeResult.intersectWith(X);
+ }
+
+ if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
+ ConstantRange X = getSignedRange(SMax->getOperand(0));
+ for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i)
+ X = X.smax(getSignedRange(SMax->getOperand(i)));
+ return ConservativeResult.intersectWith(X);
+ }
+
+ if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
+ ConstantRange X = getSignedRange(UMax->getOperand(0));
+ for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i)
+ X = X.umax(getSignedRange(UMax->getOperand(i)));
+ return ConservativeResult.intersectWith(X);
+ }
+
+ if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
+ ConstantRange X = getSignedRange(UDiv->getLHS());
+ ConstantRange Y = getSignedRange(UDiv->getRHS());
+ return ConservativeResult.intersectWith(X.udiv(Y));
+ }
+
+ if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
+ ConstantRange X = getSignedRange(ZExt->getOperand());
+ return ConservativeResult.intersectWith(X.zeroExtend(BitWidth));
+ }
+
+ if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
+ ConstantRange X = getSignedRange(SExt->getOperand());
+ return ConservativeResult.intersectWith(X.signExtend(BitWidth));
+ }
+
+ if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
+ ConstantRange X = getSignedRange(Trunc->getOperand());
+ return ConservativeResult.intersectWith(X.truncate(BitWidth));
+ }
+
+ if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
+ // If there's no signed wrap, and all the operands have the same sign or
+ // zero, the value won't ever change sign.
+ if (AddRec->hasNoSignedWrap()) {
+ bool AllNonNeg = true;
+ bool AllNonPos = true;
+ for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
+ if (!isKnownNonNegative(AddRec->getOperand(i))) AllNonNeg = false;
+ if (!isKnownNonPositive(AddRec->getOperand(i))) AllNonPos = false;
+ }
+ if (AllNonNeg)
+ ConservativeResult = ConservativeResult.intersectWith(
+ ConstantRange(APInt(BitWidth, 0),
+ APInt::getSignedMinValue(BitWidth)));
+ else if (AllNonPos)
+ ConservativeResult = ConservativeResult.intersectWith(
+ ConstantRange(APInt::getSignedMinValue(BitWidth),
+ APInt(BitWidth, 1)));
+ }
+
+ // TODO: non-affine addrec
+ if (AddRec->isAffine()) {
+ const Type *Ty = AddRec->getType();
+ const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
+ if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
+ getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
+ MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty);
+
+ const SCEV *Start = AddRec->getStart();
+ const SCEV *End = AddRec->evaluateAtIteration(MaxBECount, *this);
+
+ // Check for overflow.
+ if (!AddRec->hasNoSignedWrap())
+ return ConservativeResult;
+
+ ConstantRange StartRange = getSignedRange(Start);
+ ConstantRange EndRange = getSignedRange(End);
+ APInt Min = APIntOps::smin(StartRange.getSignedMin(),
+ EndRange.getSignedMin());
+ APInt Max = APIntOps::smax(StartRange.getSignedMax(),
+ EndRange.getSignedMax());
+ if (Min.isMinSignedValue() && Max.isMaxSignedValue())
+ return ConservativeResult;
+ return ConservativeResult.intersectWith(ConstantRange(Min, Max+1));
+ }
+ }
+
+ return ConservativeResult;
+ }
+
+ if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
+ // For a SCEVUnknown, ask ValueTracking.
+ if (!U->getValue()->getType()->isInteger() && !TD)
+ return ConservativeResult;
+ unsigned NS = ComputeNumSignBits(U->getValue(), TD);
+ if (NS == 1)
+ return ConservativeResult;
+ return ConservativeResult.intersectWith(
+ ConstantRange(APInt::getSignedMinValue(BitWidth).ashr(NS - 1),
+ APInt::getSignedMaxValue(BitWidth).ashr(NS - 1)+1));
+ }
+
+ return ConservativeResult;
+}
+
+/// createSCEV - We know that there is no SCEV for the specified value.
+/// Analyze the expression.
+///
+const SCEV *ScalarEvolution::createSCEV(Value *V) {
+ if (!isSCEVable(V->getType()))
+ return getUnknown(V);
+
+ unsigned Opcode = Instruction::UserOp1;
+ if (Instruction *I = dyn_cast<Instruction>(V))
+ Opcode = I->getOpcode();
+ else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
+ Opcode = CE->getOpcode();
+ else if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
+ return getConstant(CI);
+ else if (isa<ConstantPointerNull>(V))
+ return getIntegerSCEV(0, V->getType());
+ else if (isa<UndefValue>(V))
+ return getIntegerSCEV(0, V->getType());
+ else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V))
+ return GA->mayBeOverridden() ? getUnknown(V) : getSCEV(GA->getAliasee());
+ else
+ return getUnknown(V);
+
+ Operator *U = cast<Operator>(V);
+ switch (Opcode) {
+ case Instruction::Add:
+ // Don't transfer the NSW and NUW bits from the Add instruction to the
+ // Add expression, because the Instruction may be guarded by control
+ // flow and the no-overflow bits may not be valid for the expression in
+ // any context.
+ return getAddExpr(getSCEV(U->getOperand(0)),
+ getSCEV(U->getOperand(1)));
+ case Instruction::Mul:
+ // Don't transfer the NSW and NUW bits from the Mul instruction to the
+ // Mul expression, as with Add.
+ return getMulExpr(getSCEV(U->getOperand(0)),
+ getSCEV(U->getOperand(1)));
+ case Instruction::UDiv:
+ return getUDivExpr(getSCEV(U->getOperand(0)),
+ getSCEV(U->getOperand(1)));
+ case Instruction::Sub:
+ return getMinusSCEV(getSCEV(U->getOperand(0)),
+ getSCEV(U->getOperand(1)));
+ case Instruction::And:
+ // For an expression like x&255 that merely masks off the high bits,
+ // use zext(trunc(x)) as the SCEV expression.
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
+ if (CI->isNullValue())
+ return getSCEV(U->getOperand(1));
+ if (CI->isAllOnesValue())
+ return getSCEV(U->getOperand(0));
+ const APInt &A = CI->getValue();
+
+ // Instcombine's ShrinkDemandedConstant may strip bits out of
+ // constants, obscuring what would otherwise be a low-bits mask.
+ // Use ComputeMaskedBits to compute what ShrinkDemandedConstant
+ // knew about to reconstruct a low-bits mask value.
+ unsigned LZ = A.countLeadingZeros();
+ unsigned BitWidth = A.getBitWidth();
+ APInt AllOnes = APInt::getAllOnesValue(BitWidth);
+ APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
+ ComputeMaskedBits(U->getOperand(0), AllOnes, KnownZero, KnownOne, TD);
+
+ APInt EffectiveMask = APInt::getLowBitsSet(BitWidth, BitWidth - LZ);
+
+ if (LZ != 0 && !((~A & ~KnownZero) & EffectiveMask))
+ return
+ getZeroExtendExpr(getTruncateExpr(getSCEV(U->getOperand(0)),
+ IntegerType::get(getContext(), BitWidth - LZ)),
+ U->getType());
+ }
+ break;
+
+ case Instruction::Or:
+ // If the RHS of the Or is a constant, we may have something like:
+ // X*4+1 which got turned into X*4|1. Handle this as an Add so loop
+ // optimizations will transparently handle this case.
+ //
+ // In order for this transformation to be safe, the LHS must be of the
+ // form X*(2^n) and the Or constant must be less than 2^n.
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
+ const SCEV *LHS = getSCEV(U->getOperand(0));
+ const APInt &CIVal = CI->getValue();
+ if (GetMinTrailingZeros(LHS) >=
+ (CIVal.getBitWidth() - CIVal.countLeadingZeros())) {
+ // Build a plain add SCEV.
+ const SCEV *S = getAddExpr(LHS, getSCEV(CI));
+ // If the LHS of the add was an addrec and it has no-wrap flags,
+ // transfer the no-wrap flags, since an or won't introduce a wrap.
+ if (const SCEVAddRecExpr *NewAR = dyn_cast<SCEVAddRecExpr>(S)) {
+ const SCEVAddRecExpr *OldAR = cast<SCEVAddRecExpr>(LHS);
+ if (OldAR->hasNoUnsignedWrap())
+ const_cast<SCEVAddRecExpr *>(NewAR)->setHasNoUnsignedWrap(true);
+ if (OldAR->hasNoSignedWrap())
+ const_cast<SCEVAddRecExpr *>(NewAR)->setHasNoSignedWrap(true);
+ }
+ return S;
+ }
+ }
+ break;
+ case Instruction::Xor:
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
+ // If the RHS of the xor is a signbit, then this is just an add.
+ // Instcombine turns add of signbit into xor as a strength reduction step.
+ if (CI->getValue().isSignBit())
+ return getAddExpr(getSCEV(U->getOperand(0)),
+ getSCEV(U->getOperand(1)));
+
+ // If the RHS of xor is -1, then this is a not operation.
+ if (CI->isAllOnesValue())
+ return getNotSCEV(getSCEV(U->getOperand(0)));
+
+ // Model xor(and(x, C), C) as and(~x, C), if C is a low-bits mask.
+ // This is a variant of the check for xor with -1, and it handles
+ // the case where instcombine has trimmed non-demanded bits out
+ // of an xor with -1.
+ if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U->getOperand(0)))
+ if (ConstantInt *LCI = dyn_cast<ConstantInt>(BO->getOperand(1)))
+ if (BO->getOpcode() == Instruction::And &&
+ LCI->getValue() == CI->getValue())
+ if (const SCEVZeroExtendExpr *Z =
+ dyn_cast<SCEVZeroExtendExpr>(getSCEV(U->getOperand(0)))) {
+ const Type *UTy = U->getType();
+ const SCEV *Z0 = Z->getOperand();
+ const Type *Z0Ty = Z0->getType();
+ unsigned Z0TySize = getTypeSizeInBits(Z0Ty);
+
+ // If C is a low-bits mask, the zero extend is zerving to
+ // mask off the high bits. Complement the operand and
+ // re-apply the zext.
+ if (APIntOps::isMask(Z0TySize, CI->getValue()))
+ return getZeroExtendExpr(getNotSCEV(Z0), UTy);
+
+ // If C is a single bit, it may be in the sign-bit position
+ // before the zero-extend. In this case, represent the xor
+ // using an add, which is equivalent, and re-apply the zext.
+ APInt Trunc = APInt(CI->getValue()).trunc(Z0TySize);
+ if (APInt(Trunc).zext(getTypeSizeInBits(UTy)) == CI->getValue() &&
+ Trunc.isSignBit())
+ return getZeroExtendExpr(getAddExpr(Z0, getConstant(Trunc)),
+ UTy);
+ }
+ }
+ break;
+
+ case Instruction::Shl:
+ // Turn shift left of a constant amount into a multiply.
+ if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
+ uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth();
+ Constant *X = ConstantInt::get(getContext(),
+ APInt(BitWidth, 1).shl(SA->getLimitedValue(BitWidth)));
+ return getMulExpr(getSCEV(U->getOperand(0)), getSCEV(X));
+ }
+ break;
+
+ case Instruction::LShr:
+ // Turn logical shift right of a constant into a unsigned divide.
+ if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
+ uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth();
+ Constant *X = ConstantInt::get(getContext(),
+ APInt(BitWidth, 1).shl(SA->getLimitedValue(BitWidth)));
+ return getUDivExpr(getSCEV(U->getOperand(0)), getSCEV(X));
+ }
+ break;
+
+ case Instruction::AShr:
+ // For a two-shift sext-inreg, use sext(trunc(x)) as the SCEV expression.
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1)))
+ if (Instruction *L = dyn_cast<Instruction>(U->getOperand(0)))
+ if (L->getOpcode() == Instruction::Shl &&
+ L->getOperand(1) == U->getOperand(1)) {
+ unsigned BitWidth = getTypeSizeInBits(U->getType());
+ uint64_t Amt = BitWidth - CI->getZExtValue();
+ if (Amt == BitWidth)
+ return getSCEV(L->getOperand(0)); // shift by zero --> noop
+ if (Amt > BitWidth)
+ return getIntegerSCEV(0, U->getType()); // value is undefined
+ return
+ getSignExtendExpr(getTruncateExpr(getSCEV(L->getOperand(0)),
+ IntegerType::get(getContext(), Amt)),
+ U->getType());
+ }
+ break;
+
+ case Instruction::Trunc:
+ return getTruncateExpr(getSCEV(U->getOperand(0)), U->getType());
+
+ case Instruction::ZExt:
+ return getZeroExtendExpr(getSCEV(U->getOperand(0)), U->getType());
+
+ case Instruction::SExt:
+ return getSignExtendExpr(getSCEV(U->getOperand(0)), U->getType());
+
+ case Instruction::BitCast:
+ // BitCasts are no-op casts so we just eliminate the cast.
+ if (isSCEVable(U->getType()) && isSCEVable(U->getOperand(0)->getType()))
+ return getSCEV(U->getOperand(0));
+ break;
+
+ // It's tempting to handle inttoptr and ptrtoint as no-ops, however this can
+ // lead to pointer expressions which cannot safely be expanded to GEPs,
+ // because ScalarEvolution doesn't respect the GEP aliasing rules when
+ // simplifying integer expressions.
+
+ case Instruction::GetElementPtr:
+ return createNodeForGEP(cast<GEPOperator>(U));
+
+ case Instruction::PHI:
+ return createNodeForPHI(cast<PHINode>(U));
+
+ case Instruction::Select:
+ // This could be a smax or umax that was lowered earlier.
+ // Try to recover it.
+ if (ICmpInst *ICI = dyn_cast<ICmpInst>(U->getOperand(0))) {
+ Value *LHS = ICI->getOperand(0);
+ Value *RHS = ICI->getOperand(1);
+ switch (ICI->getPredicate()) {
+ case ICmpInst::ICMP_SLT:
+ case ICmpInst::ICMP_SLE:
+ std::swap(LHS, RHS);
+ // fall through
+ case ICmpInst::ICMP_SGT:
+ case ICmpInst::ICMP_SGE:
+ if (LHS == U->getOperand(1) && RHS == U->getOperand(2))
+ return getSMaxExpr(getSCEV(LHS), getSCEV(RHS));
+ else if (LHS == U->getOperand(2) && RHS == U->getOperand(1))
+ return getSMinExpr(getSCEV(LHS), getSCEV(RHS));
+ break;
+ case ICmpInst::ICMP_ULT:
+ case ICmpInst::ICMP_ULE:
+ std::swap(LHS, RHS);
+ // fall through
+ case ICmpInst::ICMP_UGT:
+ case ICmpInst::ICMP_UGE:
+ if (LHS == U->getOperand(1) && RHS == U->getOperand(2))
+ return getUMaxExpr(getSCEV(LHS), getSCEV(RHS));
+ else if (LHS == U->getOperand(2) && RHS == U->getOperand(1))
+ return getUMinExpr(getSCEV(LHS), getSCEV(RHS));
+ break;
+ case ICmpInst::ICMP_NE:
+ // n != 0 ? n : 1 -> umax(n, 1)
+ if (LHS == U->getOperand(1) &&
+ isa<ConstantInt>(U->getOperand(2)) &&
+ cast<ConstantInt>(U->getOperand(2))->isOne() &&
+ isa<ConstantInt>(RHS) &&
+ cast<ConstantInt>(RHS)->isZero())
+ return getUMaxExpr(getSCEV(LHS), getSCEV(U->getOperand(2)));
+ break;
+ case ICmpInst::ICMP_EQ:
+ // n == 0 ? 1 : n -> umax(n, 1)
+ if (LHS == U->getOperand(2) &&
+ isa<ConstantInt>(U->getOperand(1)) &&
+ cast<ConstantInt>(U->getOperand(1))->isOne() &&
+ isa<ConstantInt>(RHS) &&
+ cast<ConstantInt>(RHS)->isZero())
+ return getUMaxExpr(getSCEV(LHS), getSCEV(U->getOperand(1)));
+ break;
+ default:
+ break;
+ }
+ }
+
+ default: // We cannot analyze this expression.
+ break;
+ }
+
+ return getUnknown(V);
+}
+
+
+
+//===----------------------------------------------------------------------===//
+// Iteration Count Computation Code
+//
+
+/// getBackedgeTakenCount - If the specified loop has a predictable
+/// backedge-taken count, return it, otherwise return a SCEVCouldNotCompute
+/// object. The backedge-taken count is the number of times the loop header
+/// will be branched to from within the loop. This is one less than the
+/// trip count of the loop, since it doesn't count the first iteration,
+/// when the header is branched to from outside the loop.
+///
+/// Note that it is not valid to call this method on a loop without a
+/// loop-invariant backedge-taken count (see
+/// hasLoopInvariantBackedgeTakenCount).
+///
+const SCEV *ScalarEvolution::getBackedgeTakenCount(const Loop *L) {
+ return getBackedgeTakenInfo(L).Exact;
+}
+
+/// getMaxBackedgeTakenCount - Similar to getBackedgeTakenCount, except
+/// return the least SCEV value that is known never to be less than the
+/// actual backedge taken count.
+const SCEV *ScalarEvolution::getMaxBackedgeTakenCount(const Loop *L) {
+ return getBackedgeTakenInfo(L).Max;
+}
+
+/// PushLoopPHIs - Push PHI nodes in the header of the given loop
+/// onto the given Worklist.
+static void
+PushLoopPHIs(const Loop *L, SmallVectorImpl<Instruction *> &Worklist) {
+ BasicBlock *Header = L->getHeader();
+
+ // Push all Loop-header PHIs onto the Worklist stack.
+ for (BasicBlock::iterator I = Header->begin();
+ PHINode *PN = dyn_cast<PHINode>(I); ++I)
+ Worklist.push_back(PN);
+}
+
+const ScalarEvolution::BackedgeTakenInfo &
+ScalarEvolution::getBackedgeTakenInfo(const Loop *L) {
+ // Initially insert a CouldNotCompute for this loop. If the insertion
+ // succeeds, procede to actually compute a backedge-taken count and
+ // update the value. The temporary CouldNotCompute value tells SCEV
+ // code elsewhere that it shouldn't attempt to request a new
+ // backedge-taken count, which could result in infinite recursion.
+ std::pair<std::map<const Loop *, BackedgeTakenInfo>::iterator, bool> Pair =
+ BackedgeTakenCounts.insert(std::make_pair(L, getCouldNotCompute()));
+ if (Pair.second) {
+ BackedgeTakenInfo BECount = ComputeBackedgeTakenCount(L);
+ if (BECount.Exact != getCouldNotCompute()) {
+ assert(BECount.Exact->isLoopInvariant(L) &&
+ BECount.Max->isLoopInvariant(L) &&
+ "Computed backedge-taken count isn't loop invariant for loop!");
+ ++NumTripCountsComputed;
+
+ // Update the value in the map.
+ Pair.first->second = BECount;
+ } else {
+ if (BECount.Max != getCouldNotCompute())
+ // Update the value in the map.
+ Pair.first->second = BECount;
+ if (isa<PHINode>(L->getHeader()->begin()))
+ // Only count loops that have phi nodes as not being computable.
+ ++NumTripCountsNotComputed;
+ }
+
+ // Now that we know more about the trip count for this loop, forget any
+ // existing SCEV values for PHI nodes in this loop since they are only
+ // conservative estimates made without the benefit of trip count
+ // information. This is similar to the code in forgetLoop, except that
+ // it handles SCEVUnknown PHI nodes specially.
+ if (BECount.hasAnyInfo()) {
+ SmallVector<Instruction *, 16> Worklist;
+ PushLoopPHIs(L, Worklist);
+
+ SmallPtrSet<Instruction *, 8> Visited;
+ while (!Worklist.empty()) {
+ Instruction *I = Worklist.pop_back_val();
+ if (!Visited.insert(I)) continue;
+
+ std::map<SCEVCallbackVH, const SCEV *>::iterator It =
+ Scalars.find(static_cast<Value *>(I));
+ if (It != Scalars.end()) {
+ // SCEVUnknown for a PHI either means that it has an unrecognized
+ // structure, or it's a PHI that's in the progress of being computed
+ // by createNodeForPHI. In the former case, additional loop trip
+ // count information isn't going to change anything. In the later
+ // case, createNodeForPHI will perform the necessary updates on its
+ // own when it gets to that point.
+ if (!isa<PHINode>(I) || !isa<SCEVUnknown>(It->second)) {
+ ValuesAtScopes.erase(It->second);
+ Scalars.erase(It);
+ }
+ if (PHINode *PN = dyn_cast<PHINode>(I))
+ ConstantEvolutionLoopExitValue.erase(PN);
+ }
+
+ PushDefUseChildren(I, Worklist);
+ }
+ }
+ }
+ return Pair.first->second;
+}
+
+/// forgetLoop - This method should be called by the client when it has
+/// changed a loop in a way that may effect ScalarEvolution's ability to
+/// compute a trip count, or if the loop is deleted.
+void ScalarEvolution::forgetLoop(const Loop *L) {
+ // Drop any stored trip count value.
+ BackedgeTakenCounts.erase(L);
+
+ // Drop information about expressions based on loop-header PHIs.
+ SmallVector<Instruction *, 16> Worklist;
+ PushLoopPHIs(L, Worklist);
+
+ SmallPtrSet<Instruction *, 8> Visited;
+ while (!Worklist.empty()) {
+ Instruction *I = Worklist.pop_back_val();
+ if (!Visited.insert(I)) continue;
+
+ std::map<SCEVCallbackVH, const SCEV *>::iterator It =
+ Scalars.find(static_cast<Value *>(I));
+ if (It != Scalars.end()) {
+ ValuesAtScopes.erase(It->second);
+ Scalars.erase(It);
+ if (PHINode *PN = dyn_cast<PHINode>(I))
+ ConstantEvolutionLoopExitValue.erase(PN);
+ }
+
+ PushDefUseChildren(I, Worklist);
+ }
+}
+
+/// ComputeBackedgeTakenCount - Compute the number of times the backedge
+/// of the specified loop will execute.
+ScalarEvolution::BackedgeTakenInfo
+ScalarEvolution::ComputeBackedgeTakenCount(const Loop *L) {
+ SmallVector<BasicBlock *, 8> ExitingBlocks;
+ L->getExitingBlocks(ExitingBlocks);
+
+ // Examine all exits and pick the most conservative values.
+ const SCEV *BECount = getCouldNotCompute();
+ const SCEV *MaxBECount = getCouldNotCompute();
+ bool CouldNotComputeBECount = false;
+ for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
+ BackedgeTakenInfo NewBTI =
+ ComputeBackedgeTakenCountFromExit(L, ExitingBlocks[i]);
+
+ if (NewBTI.Exact == getCouldNotCompute()) {
+ // We couldn't compute an exact value for this exit, so
+ // we won't be able to compute an exact value for the loop.
+ CouldNotComputeBECount = true;
+ BECount = getCouldNotCompute();
+ } else if (!CouldNotComputeBECount) {
+ if (BECount == getCouldNotCompute())
+ BECount = NewBTI.Exact;
+ else
+ BECount = getUMinFromMismatchedTypes(BECount, NewBTI.Exact);
+ }
+ if (MaxBECount == getCouldNotCompute())
+ MaxBECount = NewBTI.Max;
+ else if (NewBTI.Max != getCouldNotCompute())
+ MaxBECount = getUMinFromMismatchedTypes(MaxBECount, NewBTI.Max);
+ }
+
+ return BackedgeTakenInfo(BECount, MaxBECount);
+}
+
+/// ComputeBackedgeTakenCountFromExit - Compute the number of times the backedge
+/// of the specified loop will execute if it exits via the specified block.
+ScalarEvolution::BackedgeTakenInfo
+ScalarEvolution::ComputeBackedgeTakenCountFromExit(const Loop *L,
+ BasicBlock *ExitingBlock) {
+
+ // Okay, we've chosen an exiting block. See what condition causes us to
+ // exit at this block.
+ //
+ // FIXME: we should be able to handle switch instructions (with a single exit)
+ BranchInst *ExitBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
+ if (ExitBr == 0) return getCouldNotCompute();
+ assert(ExitBr->isConditional() && "If unconditional, it can't be in loop!");
+
+ // At this point, we know we have a conditional branch that determines whether
+ // the loop is exited. However, we don't know if the branch is executed each
+ // time through the loop. If not, then the execution count of the branch will
+ // not be equal to the trip count of the loop.
+ //
+ // Currently we check for this by checking to see if the Exit branch goes to
+ // the loop header. If so, we know it will always execute the same number of
+ // times as the loop. We also handle the case where the exit block *is* the
+ // loop header. This is common for un-rotated loops.
+ //
+ // If both of those tests fail, walk up the unique predecessor chain to the
+ // header, stopping if there is an edge that doesn't exit the loop. If the
+ // header is reached, the execution count of the branch will be equal to the
+ // trip count of the loop.
+ //
+ // More extensive analysis could be done to handle more cases here.
+ //
+ if (ExitBr->getSuccessor(0) != L->getHeader() &&
+ ExitBr->getSuccessor(1) != L->getHeader() &&
+ ExitBr->getParent() != L->getHeader()) {
+ // The simple checks failed, try climbing the unique predecessor chain
+ // up to the header.
+ bool Ok = false;
+ for (BasicBlock *BB = ExitBr->getParent(); BB; ) {
+ BasicBlock *Pred = BB->getUniquePredecessor();
+ if (!Pred)
+ return getCouldNotCompute();
+ TerminatorInst *PredTerm = Pred->getTerminator();
+ for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) {
+ BasicBlock *PredSucc = PredTerm->getSuccessor(i);
+ if (PredSucc == BB)
+ continue;
+ // If the predecessor has a successor that isn't BB and isn't
+ // outside the loop, assume the worst.
+ if (L->contains(PredSucc))
+ return getCouldNotCompute();
+ }
+ if (Pred == L->getHeader()) {
+ Ok = true;
+ break;
+ }
+ BB = Pred;
+ }
+ if (!Ok)
+ return getCouldNotCompute();
+ }
+
+ // Procede to the next level to examine the exit condition expression.
+ return ComputeBackedgeTakenCountFromExitCond(L, ExitBr->getCondition(),
+ ExitBr->getSuccessor(0),
+ ExitBr->getSuccessor(1));
+}
+
+/// ComputeBackedgeTakenCountFromExitCond - Compute the number of times the
+/// backedge of the specified loop will execute if its exit condition
+/// were a conditional branch of ExitCond, TBB, and FBB.
+ScalarEvolution::BackedgeTakenInfo
+ScalarEvolution::ComputeBackedgeTakenCountFromExitCond(const Loop *L,
+ Value *ExitCond,
+ BasicBlock *TBB,
+ BasicBlock *FBB) {
+ // Check if the controlling expression for this loop is an And or Or.
+ if (BinaryOperator *BO = dyn_cast<BinaryOperator>(ExitCond)) {
+ if (BO->getOpcode() == Instruction::And) {
+ // Recurse on the operands of the and.
+ BackedgeTakenInfo BTI0 =
+ ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(0), TBB, FBB);
+ BackedgeTakenInfo BTI1 =
+ ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(1), TBB, FBB);
+ const SCEV *BECount = getCouldNotCompute();
+ const SCEV *MaxBECount = getCouldNotCompute();
+ if (L->contains(TBB)) {
+ // Both conditions must be true for the loop to continue executing.
+ // Choose the less conservative count.
+ if (BTI0.Exact == getCouldNotCompute() ||
+ BTI1.Exact == getCouldNotCompute())
+ BECount = getCouldNotCompute();
+ else
+ BECount = getUMinFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
+ if (BTI0.Max == getCouldNotCompute())
+ MaxBECount = BTI1.Max;
+ else if (BTI1.Max == getCouldNotCompute())
+ MaxBECount = BTI0.Max;
+ else
+ MaxBECount = getUMinFromMismatchedTypes(BTI0.Max, BTI1.Max);
+ } else {
+ // Both conditions must be true for the loop to exit.
+ assert(L->contains(FBB) && "Loop block has no successor in loop!");
+ if (BTI0.Exact != getCouldNotCompute() &&
+ BTI1.Exact != getCouldNotCompute())
+ BECount = getUMaxFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
+ if (BTI0.Max != getCouldNotCompute() &&
+ BTI1.Max != getCouldNotCompute())
+ MaxBECount = getUMaxFromMismatchedTypes(BTI0.Max, BTI1.Max);
+ }
+
+ return BackedgeTakenInfo(BECount, MaxBECount);
+ }
+ if (BO->getOpcode() == Instruction::Or) {
+ // Recurse on the operands of the or.
+ BackedgeTakenInfo BTI0 =
+ ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(0), TBB, FBB);
+ BackedgeTakenInfo BTI1 =
+ ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(1), TBB, FBB);
+ const SCEV *BECount = getCouldNotCompute();
+ const SCEV *MaxBECount = getCouldNotCompute();
+ if (L->contains(FBB)) {
+ // Both conditions must be false for the loop to continue executing.
+ // Choose the less conservative count.
+ if (BTI0.Exact == getCouldNotCompute() ||
+ BTI1.Exact == getCouldNotCompute())
+ BECount = getCouldNotCompute();
+ else
+ BECount = getUMinFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
+ if (BTI0.Max == getCouldNotCompute())
+ MaxBECount = BTI1.Max;
+ else if (BTI1.Max == getCouldNotCompute())
+ MaxBECount = BTI0.Max;
+ else
+ MaxBECount = getUMinFromMismatchedTypes(BTI0.Max, BTI1.Max);
+ } else {
+ // Both conditions must be false for the loop to exit.
+ assert(L->contains(TBB) && "Loop block has no successor in loop!");
+ if (BTI0.Exact != getCouldNotCompute() &&
+ BTI1.Exact != getCouldNotCompute())
+ BECount = getUMaxFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
+ if (BTI0.Max != getCouldNotCompute() &&
+ BTI1.Max != getCouldNotCompute())
+ MaxBECount = getUMaxFromMismatchedTypes(BTI0.Max, BTI1.Max);
+ }
+
+ return BackedgeTakenInfo(BECount, MaxBECount);
+ }
+ }
+
+ // With an icmp, it may be feasible to compute an exact backedge-taken count.
+ // Procede to the next level to examine the icmp.
+ if (ICmpInst *ExitCondICmp = dyn_cast<ICmpInst>(ExitCond))
+ return ComputeBackedgeTakenCountFromExitCondICmp(L, ExitCondICmp, TBB, FBB);
+
+ // If it's not an integer or pointer comparison then compute it the hard way.
+ return ComputeBackedgeTakenCountExhaustively(L, ExitCond, !L->contains(TBB));
+}
+
+/// ComputeBackedgeTakenCountFromExitCondICmp - Compute the number of times the
+/// backedge of the specified loop will execute if its exit condition
+/// were a conditional branch of the ICmpInst ExitCond, TBB, and FBB.
+ScalarEvolution::BackedgeTakenInfo
+ScalarEvolution::ComputeBackedgeTakenCountFromExitCondICmp(const Loop *L,
+ ICmpInst *ExitCond,
+ BasicBlock *TBB,
+ BasicBlock *FBB) {
+
+ // If the condition was exit on true, convert the condition to exit on false
+ ICmpInst::Predicate Cond;
+ if (!L->contains(FBB))
+ Cond = ExitCond->getPredicate();
+ else
+ Cond = ExitCond->getInversePredicate();
+
+ // Handle common loops like: for (X = "string"; *X; ++X)
+ if (LoadInst *LI = dyn_cast<LoadInst>(ExitCond->getOperand(0)))
+ if (Constant *RHS = dyn_cast<Constant>(ExitCond->getOperand(1))) {
+ const SCEV *ItCnt =
+ ComputeLoadConstantCompareBackedgeTakenCount(LI, RHS, L, Cond);
+ if (!isa<SCEVCouldNotCompute>(ItCnt)) {
+ unsigned BitWidth = getTypeSizeInBits(ItCnt->getType());
+ return BackedgeTakenInfo(ItCnt,
+ isa<SCEVConstant>(ItCnt) ? ItCnt :
+ getConstant(APInt::getMaxValue(BitWidth)-1));
+ }
+ }
+
+ const SCEV *LHS = getSCEV(ExitCond->getOperand(0));
+ const SCEV *RHS = getSCEV(ExitCond->getOperand(1));
+
+ // Try to evaluate any dependencies out of the loop.
+ LHS = getSCEVAtScope(LHS, L);
+ RHS = getSCEVAtScope(RHS, L);
+
+ // At this point, we would like to compute how many iterations of the
+ // loop the predicate will return true for these inputs.
+ if (LHS->isLoopInvariant(L) && !RHS->isLoopInvariant(L)) {
+ // If there is a loop-invariant, force it into the RHS.
+ std::swap(LHS, RHS);
+ Cond = ICmpInst::getSwappedPredicate(Cond);
+ }
+
+ // If we have a comparison of a chrec against a constant, try to use value
+ // ranges to answer this query.
+ if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS))
+ if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS))
+ if (AddRec->getLoop() == L) {
+ // Form the constant range.
+ ConstantRange CompRange(
+ ICmpInst::makeConstantRange(Cond, RHSC->getValue()->getValue()));
+
+ const SCEV *Ret = AddRec->getNumIterationsInRange(CompRange, *this);
+ if (!isa<SCEVCouldNotCompute>(Ret)) return Ret;
+ }
+
+ switch (Cond) {
+ case ICmpInst::ICMP_NE: { // while (X != Y)
+ // Convert to: while (X-Y != 0)
+ const SCEV *TC = HowFarToZero(getMinusSCEV(LHS, RHS), L);
+ if (!isa<SCEVCouldNotCompute>(TC)) return TC;
+ break;
+ }
+ case ICmpInst::ICMP_EQ: { // while (X == Y)
+ // Convert to: while (X-Y == 0)
+ const SCEV *TC = HowFarToNonZero(getMinusSCEV(LHS, RHS), L);
+ if (!isa<SCEVCouldNotCompute>(TC)) return TC;
+ break;
+ }
+ case ICmpInst::ICMP_SLT: {
+ BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, true);
+ if (BTI.hasAnyInfo()) return BTI;
+ break;
+ }
+ case ICmpInst::ICMP_SGT: {
+ BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
+ getNotSCEV(RHS), L, true);
+ if (BTI.hasAnyInfo()) return BTI;
+ break;
+ }
+ case ICmpInst::ICMP_ULT: {
+ BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, false);
+ if (BTI.hasAnyInfo()) return BTI;
+ break;
+ }
+ case ICmpInst::ICMP_UGT: {
+ BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
+ getNotSCEV(RHS), L, false);
+ if (BTI.hasAnyInfo()) return BTI;
+ break;
+ }
+ default:
+#if 0
+ dbgs() << "ComputeBackedgeTakenCount ";
+ if (ExitCond->getOperand(0)->getType()->isUnsigned())
+ dbgs() << "[unsigned] ";
+ dbgs() << *LHS << " "
+ << Instruction::getOpcodeName(Instruction::ICmp)
+ << " " << *RHS << "\n";
+#endif
+ break;
+ }
+ return
+ ComputeBackedgeTakenCountExhaustively(L, ExitCond, !L->contains(TBB));
+}
+
+static ConstantInt *
+EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, ConstantInt *C,
+ ScalarEvolution &SE) {
+ const SCEV *InVal = SE.getConstant(C);
+ const SCEV *Val = AddRec->evaluateAtIteration(InVal, SE);
+ assert(isa<SCEVConstant>(Val) &&
+ "Evaluation of SCEV at constant didn't fold correctly?");
+ return cast<SCEVConstant>(Val)->getValue();
+}
+
+/// GetAddressedElementFromGlobal - Given a global variable with an initializer
+/// and a GEP expression (missing the pointer index) indexing into it, return
+/// the addressed element of the initializer or null if the index expression is
+/// invalid.
+static Constant *
+GetAddressedElementFromGlobal(GlobalVariable *GV,
+ const std::vector<ConstantInt*> &Indices) {
+ Constant *Init = GV->getInitializer();
+ for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
+ uint64_t Idx = Indices[i]->getZExtValue();
+ if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
+ assert(Idx < CS->getNumOperands() && "Bad struct index!");
+ Init = cast<Constant>(CS->getOperand(Idx));
+ } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
+ if (Idx >= CA->getNumOperands()) return 0; // Bogus program
+ Init = cast<Constant>(CA->getOperand(Idx));
+ } else if (isa<ConstantAggregateZero>(Init)) {
+ if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
+ assert(Idx < STy->getNumElements() && "Bad struct index!");
+ Init = Constant::getNullValue(STy->getElementType(Idx));
+ } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Init->getType())) {
+ if (Idx >= ATy->getNumElements()) return 0; // Bogus program
+ Init = Constant::getNullValue(ATy->getElementType());
+ } else {
+ llvm_unreachable("Unknown constant aggregate type!");
+ }
+ return 0;
+ } else {
+ return 0; // Unknown initializer type
+ }
+ }
+ return Init;
+}
+
+/// ComputeLoadConstantCompareBackedgeTakenCount - Given an exit condition of
+/// 'icmp op load X, cst', try to see if we can compute the backedge
+/// execution count.
+const SCEV *
+ScalarEvolution::ComputeLoadConstantCompareBackedgeTakenCount(
+ LoadInst *LI,
+ Constant *RHS,
+ const Loop *L,
+ ICmpInst::Predicate predicate) {
+ if (LI->isVolatile()) return getCouldNotCompute();
+
+ // Check to see if the loaded pointer is a getelementptr of a global.
+ GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0));
+ if (!GEP) return getCouldNotCompute();
+
+ // Make sure that it is really a constant global we are gepping, with an
+ // initializer, and make sure the first IDX is really 0.
+ GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0));
+ if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
+ GEP->getNumOperands() < 3 || !isa<Constant>(GEP->getOperand(1)) ||
+ !cast<Constant>(GEP->getOperand(1))->isNullValue())
+ return getCouldNotCompute();
+
+ // Okay, we allow one non-constant index into the GEP instruction.
+ Value *VarIdx = 0;
+ std::vector<ConstantInt*> Indexes;
+ unsigned VarIdxNum = 0;
+ for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i)
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
+ Indexes.push_back(CI);
+ } else if (!isa<ConstantInt>(GEP->getOperand(i))) {
+ if (VarIdx) return getCouldNotCompute(); // Multiple non-constant idx's.
+ VarIdx = GEP->getOperand(i);
+ VarIdxNum = i-2;
+ Indexes.push_back(0);
+ }
+
+ // Okay, we know we have a (load (gep GV, 0, X)) comparison with a constant.
+ // Check to see if X is a loop variant variable value now.
+ const SCEV *Idx = getSCEV(VarIdx);
+ Idx = getSCEVAtScope(Idx, L);
+
+ // We can only recognize very limited forms of loop index expressions, in
+ // particular, only affine AddRec's like {C1,+,C2}.
+ const SCEVAddRecExpr *IdxExpr = dyn_cast<SCEVAddRecExpr>(Idx);
+ if (!IdxExpr || !IdxExpr->isAffine() || IdxExpr->isLoopInvariant(L) ||
+ !isa<SCEVConstant>(IdxExpr->getOperand(0)) ||
+ !isa<SCEVConstant>(IdxExpr->getOperand(1)))
+ return getCouldNotCompute();
+
+ unsigned MaxSteps = MaxBruteForceIterations;
+ for (unsigned IterationNum = 0; IterationNum != MaxSteps; ++IterationNum) {
+ ConstantInt *ItCst = ConstantInt::get(
+ cast<IntegerType>(IdxExpr->getType()), IterationNum);
+ ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst, *this);
+
+ // Form the GEP offset.
+ Indexes[VarIdxNum] = Val;
+
+ Constant *Result = GetAddressedElementFromGlobal(GV, Indexes);
+ if (Result == 0) break; // Cannot compute!
+
+ // Evaluate the condition for this iteration.
+ Result = ConstantExpr::getICmp(predicate, Result, RHS);
+ if (!isa<ConstantInt>(Result)) break; // Couldn't decide for sure
+ if (cast<ConstantInt>(Result)->getValue().isMinValue()) {
+#if 0
+ dbgs() << "\n***\n*** Computed loop count " << *ItCst
+ << "\n*** From global " << *GV << "*** BB: " << *L->getHeader()
+ << "***\n";
+#endif
+ ++NumArrayLenItCounts;
+ return getConstant(ItCst); // Found terminating iteration!
+ }
+ }
+ return getCouldNotCompute();
+}
+
+
+/// CanConstantFold - Return true if we can constant fold an instruction of the
+/// specified type, assuming that all operands were constants.
+static bool CanConstantFold(const Instruction *I) {
+ if (isa<BinaryOperator>(I) || isa<CmpInst>(I) ||
+ isa<SelectInst>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I))
+ return true;
+
+ if (const CallInst *CI = dyn_cast<CallInst>(I))
+ if (const Function *F = CI->getCalledFunction())
+ return canConstantFoldCallTo(F);
+ return false;
+}
+
+/// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node
+/// in the loop that V is derived from. We allow arbitrary operations along the
+/// way, but the operands of an operation must either be constants or a value
+/// derived from a constant PHI. If this expression does not fit with these
+/// constraints, return null.
+static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) {
+ // If this is not an instruction, or if this is an instruction outside of the
+ // loop, it can't be derived from a loop PHI.
+ Instruction *I = dyn_cast<Instruction>(V);
+ if (I == 0 || !L->contains(I)) return 0;
+
+ if (PHINode *PN = dyn_cast<PHINode>(I)) {
+ if (L->getHeader() == I->getParent())
+ return PN;
+ else
+ // We don't currently keep track of the control flow needed to evaluate
+ // PHIs, so we cannot handle PHIs inside of loops.
+ return 0;
+ }
+
+ // If we won't be able to constant fold this expression even if the operands
+ // are constants, return early.
+ if (!CanConstantFold(I)) return 0;
+
+ // Otherwise, we can evaluate this instruction if all of its operands are
+ // constant or derived from a PHI node themselves.
+ PHINode *PHI = 0;
+ for (unsigned Op = 0, e = I->getNumOperands(); Op != e; ++Op)
+ if (!(isa<Constant>(I->getOperand(Op)) ||
+ isa<GlobalValue>(I->getOperand(Op)))) {
+ PHINode *P = getConstantEvolvingPHI(I->getOperand(Op), L);
+ if (P == 0) return 0; // Not evolving from PHI
+ if (PHI == 0)
+ PHI = P;
+ else if (PHI != P)
+ return 0; // Evolving from multiple different PHIs.
+ }
+
+ // This is a expression evolving from a constant PHI!
+ return PHI;
+}
+
+/// EvaluateExpression - Given an expression that passes the
+/// getConstantEvolvingPHI predicate, evaluate its value assuming the PHI node
+/// in the loop has the value PHIVal. If we can't fold this expression for some
+/// reason, return null.
+static Constant *EvaluateExpression(Value *V, Constant *PHIVal,
+ const TargetData *TD) {
+ if (isa<PHINode>(V)) return PHIVal;
+ if (Constant *C = dyn_cast<Constant>(V)) return C;
+ if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) return GV;
+ Instruction *I = cast<Instruction>(V);
+
+ std::vector<Constant*> Operands;
+ Operands.resize(I->getNumOperands());
+
+ for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
+ Operands[i] = EvaluateExpression(I->getOperand(i), PHIVal, TD);
+ if (Operands[i] == 0) return 0;
+ }
+
+ if (const CmpInst *CI = dyn_cast<CmpInst>(I))
+ return ConstantFoldCompareInstOperands(CI->getPredicate(), Operands[0],
+ Operands[1], TD);
+ return ConstantFoldInstOperands(I->getOpcode(), I->getType(),
+ &Operands[0], Operands.size(), TD);
+}
+
+/// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
+/// in the header of its containing loop, we know the loop executes a
+/// constant number of times, and the PHI node is just a recurrence
+/// involving constants, fold it.
+Constant *
+ScalarEvolution::getConstantEvolutionLoopExitValue(PHINode *PN,
+ const APInt &BEs,
+ const Loop *L) {
+ std::map<PHINode*, Constant*>::iterator I =
+ ConstantEvolutionLoopExitValue.find(PN);
+ if (I != ConstantEvolutionLoopExitValue.end())
+ return I->second;
+
+ if (BEs.ugt(APInt(BEs.getBitWidth(),MaxBruteForceIterations)))
+ return ConstantEvolutionLoopExitValue[PN] = 0; // Not going to evaluate it.
+
+ Constant *&RetVal = ConstantEvolutionLoopExitValue[PN];
+
+ // Since the loop is canonicalized, the PHI node must have two entries. One
+ // entry must be a constant (coming in from outside of the loop), and the
+ // second must be derived from the same PHI.
+ bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
+ Constant *StartCST =
+ dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
+ if (StartCST == 0)
+ return RetVal = 0; // Must be a constant.
+
+ Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
+ PHINode *PN2 = getConstantEvolvingPHI(BEValue, L);
+ if (PN2 != PN)
+ return RetVal = 0; // Not derived from same PHI.
+
+ // Execute the loop symbolically to determine the exit value.
+ if (BEs.getActiveBits() >= 32)
+ return RetVal = 0; // More than 2^32-1 iterations?? Not doing it!
+
+ unsigned NumIterations = BEs.getZExtValue(); // must be in range
+ unsigned IterationNum = 0;
+ for (Constant *PHIVal = StartCST; ; ++IterationNum) {
+ if (IterationNum == NumIterations)
+ return RetVal = PHIVal; // Got exit value!
+
+ // Compute the value of the PHI node for the next iteration.
+ Constant *NextPHI = EvaluateExpression(BEValue, PHIVal, TD);
+ if (NextPHI == PHIVal)
+ return RetVal = NextPHI; // Stopped evolving!
+ if (NextPHI == 0)
+ return 0; // Couldn't evaluate!
+ PHIVal = NextPHI;
+ }
+}
+
+/// ComputeBackedgeTakenCountExhaustively - If the loop is known to execute a
+/// constant number of times (the condition evolves only from constants),
+/// try to evaluate a few iterations of the loop until we get the exit
+/// condition gets a value of ExitWhen (true or false). If we cannot
+/// evaluate the trip count of the loop, return getCouldNotCompute().
+const SCEV *
+ScalarEvolution::ComputeBackedgeTakenCountExhaustively(const Loop *L,
+ Value *Cond,
+ bool ExitWhen) {
+ PHINode *PN = getConstantEvolvingPHI(Cond, L);
+ if (PN == 0) return getCouldNotCompute();
+
+ // Since the loop is canonicalized, the PHI node must have two entries. One
+ // entry must be a constant (coming in from outside of the loop), and the
+ // second must be derived from the same PHI.
+ bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
+ Constant *StartCST =
+ dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
+ if (StartCST == 0) return getCouldNotCompute(); // Must be a constant.
+
+ Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
+ PHINode *PN2 = getConstantEvolvingPHI(BEValue, L);
+ if (PN2 != PN) return getCouldNotCompute(); // Not derived from same PHI.
+
+ // Okay, we find a PHI node that defines the trip count of this loop. Execute
+ // the loop symbolically to determine when the condition gets a value of
+ // "ExitWhen".
+ unsigned IterationNum = 0;
+ unsigned MaxIterations = MaxBruteForceIterations; // Limit analysis.
+ for (Constant *PHIVal = StartCST;
+ IterationNum != MaxIterations; ++IterationNum) {
+ ConstantInt *CondVal =
+ dyn_cast_or_null<ConstantInt>(EvaluateExpression(Cond, PHIVal, TD));
+
+ // Couldn't symbolically evaluate.
+ if (!CondVal) return getCouldNotCompute();
+
+ if (CondVal->getValue() == uint64_t(ExitWhen)) {
+ ++NumBruteForceTripCountsComputed;
+ return getConstant(Type::getInt32Ty(getContext()), IterationNum);
+ }
+
+ // Compute the value of the PHI node for the next iteration.
+ Constant *NextPHI = EvaluateExpression(BEValue, PHIVal, TD);
+ if (NextPHI == 0 || NextPHI == PHIVal)
+ return getCouldNotCompute();// Couldn't evaluate or not making progress...
+ PHIVal = NextPHI;
+ }
+
+ // Too many iterations were needed to evaluate.
+ return getCouldNotCompute();
+}
+
+/// getSCEVAtScope - Return a SCEV expression for the specified value
+/// at the specified scope in the program. The L value specifies a loop
+/// nest to evaluate the expression at, where null is the top-level or a
+/// specified loop is immediately inside of the loop.
+///
+/// This method can be used to compute the exit value for a variable defined
+/// in a loop by querying what the value will hold in the parent loop.
+///
+/// In the case that a relevant loop exit value cannot be computed, the
+/// original value V is returned.
+const SCEV *ScalarEvolution::getSCEVAtScope(const SCEV *V, const Loop *L) {
+ // Check to see if we've folded this expression at this loop before.
+ std::map<const Loop *, const SCEV *> &Values = ValuesAtScopes[V];
+ std::pair<std::map<const Loop *, const SCEV *>::iterator, bool> Pair =
+ Values.insert(std::make_pair(L, static_cast<const SCEV *>(0)));
+ if (!Pair.second)
+ return Pair.first->second ? Pair.first->second : V;
+
+ // Otherwise compute it.
+ const SCEV *C = computeSCEVAtScope(V, L);
+ ValuesAtScopes[V][L] = C;
+ return C;
+}
+
+const SCEV *ScalarEvolution::computeSCEVAtScope(const SCEV *V, const Loop *L) {
+ if (isa<SCEVConstant>(V)) return V;
+
+ // If this instruction is evolved from a constant-evolving PHI, compute the
+ // exit value from the loop without using SCEVs.
+ if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) {
+ if (Instruction *I = dyn_cast<Instruction>(SU->getValue())) {
+ const Loop *LI = (*this->LI)[I->getParent()];
+ if (LI && LI->getParentLoop() == L) // Looking for loop exit value.
+ if (PHINode *PN = dyn_cast<PHINode>(I))
+ if (PN->getParent() == LI->getHeader()) {
+ // Okay, there is no closed form solution for the PHI node. Check
+ // to see if the loop that contains it has a known backedge-taken
+ // count. If so, we may be able to force computation of the exit
+ // value.
+ const SCEV *BackedgeTakenCount = getBackedgeTakenCount(LI);
+ if (const SCEVConstant *BTCC =
+ dyn_cast<SCEVConstant>(BackedgeTakenCount)) {
+ // Okay, we know how many times the containing loop executes. If
+ // this is a constant evolving PHI node, get the final value at
+ // the specified iteration number.
+ Constant *RV = getConstantEvolutionLoopExitValue(PN,
+ BTCC->getValue()->getValue(),
+ LI);
+ if (RV) return getSCEV(RV);
+ }
+ }
+
+ // Okay, this is an expression that we cannot symbolically evaluate
+ // into a SCEV. Check to see if it's possible to symbolically evaluate
+ // the arguments into constants, and if so, try to constant propagate the
+ // result. This is particularly useful for computing loop exit values.
+ if (CanConstantFold(I)) {
+ std::vector<Constant*> Operands;
+ Operands.reserve(I->getNumOperands());
+ for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
+ Value *Op = I->getOperand(i);
+ if (Constant *C = dyn_cast<Constant>(Op)) {
+ Operands.push_back(C);
+ } else {
+ // If any of the operands is non-constant and if they are
+ // non-integer and non-pointer, don't even try to analyze them
+ // with scev techniques.
+ if (!isSCEVable(Op->getType()))
+ return V;
+
+ const SCEV *OpV = getSCEVAtScope(Op, L);
+ if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(OpV)) {
+ Constant *C = SC->getValue();
+ if (C->getType() != Op->getType())
+ C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
+ Op->getType(),
+ false),
+ C, Op->getType());
+ Operands.push_back(C);
+ } else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(OpV)) {
+ if (Constant *C = dyn_cast<Constant>(SU->getValue())) {
+ if (C->getType() != Op->getType())
+ C =
+ ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
+ Op->getType(),
+ false),
+ C, Op->getType());
+ Operands.push_back(C);
+ } else
+ return V;
+ } else {
+ return V;
+ }
+ }
+ }
+
+ Constant *C;
+ if (const CmpInst *CI = dyn_cast<CmpInst>(I))
+ C = ConstantFoldCompareInstOperands(CI->getPredicate(),
+ Operands[0], Operands[1], TD);
+ else
+ C = ConstantFoldInstOperands(I->getOpcode(), I->getType(),
+ &Operands[0], Operands.size(), TD);
+ return getSCEV(C);
+ }
+ }
+
+ // This is some other type of SCEVUnknown, just return it.
+ return V;
+ }
+
+ if (const SCEVCommutativeExpr *Comm = dyn_cast<SCEVCommutativeExpr>(V)) {
+ // Avoid performing the look-up in the common case where the specified
+ // expression has no loop-variant portions.
+ for (unsigned i = 0, e = Comm->getNumOperands(); i != e; ++i) {
+ const SCEV *OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
+ if (OpAtScope != Comm->getOperand(i)) {
+ // Okay, at least one of these operands is loop variant but might be
+ // foldable. Build a new instance of the folded commutative expression.
+ SmallVector<const SCEV *, 8> NewOps(Comm->op_begin(),
+ Comm->op_begin()+i);
+ NewOps.push_back(OpAtScope);
+
+ for (++i; i != e; ++i) {
+ OpAtScope = getSCEVAtScope(Comm->getOperand(i), L);
+ NewOps.push_back(OpAtScope);
+ }
+ if (isa<SCEVAddExpr>(Comm))
+ return getAddExpr(NewOps);
+ if (isa<SCEVMulExpr>(Comm))
+ return getMulExpr(NewOps);
+ if (isa<SCEVSMaxExpr>(Comm))
+ return getSMaxExpr(NewOps);
+ if (isa<SCEVUMaxExpr>(Comm))
+ return getUMaxExpr(NewOps);
+ llvm_unreachable("Unknown commutative SCEV type!");
+ }
+ }
+ // If we got here, all operands are loop invariant.
+ return Comm;
+ }
+
+ if (const SCEVUDivExpr *Div = dyn_cast<SCEVUDivExpr>(V)) {
+ const SCEV *LHS = getSCEVAtScope(Div->getLHS(), L);
+ const SCEV *RHS = getSCEVAtScope(Div->getRHS(), L);
+ if (LHS == Div->getLHS() && RHS == Div->getRHS())
+ return Div; // must be loop invariant
+ return getUDivExpr(LHS, RHS);
+ }
+
+ // If this is a loop recurrence for a loop that does not contain L, then we
+ // are dealing with the final value computed by the loop.
+ if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) {
+ if (!L || !AddRec->getLoop()->contains(L)) {
+ // To evaluate this recurrence, we need to know how many times the AddRec
+ // loop iterates. Compute this now.
+ const SCEV *BackedgeTakenCount = getBackedgeTakenCount(AddRec->getLoop());
+ if (BackedgeTakenCount == getCouldNotCompute()) return AddRec;
+
+ // Then, evaluate the AddRec.
+ return AddRec->evaluateAtIteration(BackedgeTakenCount, *this);
+ }
+ return AddRec;
+ }
+
+ if (const SCEVZeroExtendExpr *Cast = dyn_cast<SCEVZeroExtendExpr>(V)) {
+ const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
+ if (Op == Cast->getOperand())
+ return Cast; // must be loop invariant
+ return getZeroExtendExpr(Op, Cast->getType());
+ }
+
+ if (const SCEVSignExtendExpr *Cast = dyn_cast<SCEVSignExtendExpr>(V)) {
+ const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
+ if (Op == Cast->getOperand())
+ return Cast; // must be loop invariant
+ return getSignExtendExpr(Op, Cast->getType());
+ }
+
+ if (const SCEVTruncateExpr *Cast = dyn_cast<SCEVTruncateExpr>(V)) {
+ const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L);
+ if (Op == Cast->getOperand())
+ return Cast; // must be loop invariant
+ return getTruncateExpr(Op, Cast->getType());
+ }
+
+ llvm_unreachable("Unknown SCEV type!");
+ return 0;
+}
+
+/// getSCEVAtScope - This is a convenience function which does
+/// getSCEVAtScope(getSCEV(V), L).
+const SCEV *ScalarEvolution::getSCEVAtScope(Value *V, const Loop *L) {
+ return getSCEVAtScope(getSCEV(V), L);
+}
+
+/// SolveLinEquationWithOverflow - Finds the minimum unsigned root of the
+/// following equation:
+///
+/// A * X = B (mod N)
+///
+/// where N = 2^BW and BW is the common bit width of A and B. The signedness of
+/// A and B isn't important.
+///
+/// If the equation does not have a solution, SCEVCouldNotCompute is returned.
+static const SCEV *SolveLinEquationWithOverflow(const APInt &A, const APInt &B,
+ ScalarEvolution &SE) {
+ uint32_t BW = A.getBitWidth();
+ assert(BW == B.getBitWidth() && "Bit widths must be the same.");
+ assert(A != 0 && "A must be non-zero.");
+
+ // 1. D = gcd(A, N)
+ //
+ // The gcd of A and N may have only one prime factor: 2. The number of
+ // trailing zeros in A is its multiplicity
+ uint32_t Mult2 = A.countTrailingZeros();
+ // D = 2^Mult2
+
+ // 2. Check if B is divisible by D.
+ //
+ // B is divisible by D if and only if the multiplicity of prime factor 2 for B
+ // is not less than multiplicity of this prime factor for D.
+ if (B.countTrailingZeros() < Mult2)
+ return SE.getCouldNotCompute();
+
+ // 3. Compute I: the multiplicative inverse of (A / D) in arithmetic
+ // modulo (N / D).
+ //
+ // (N / D) may need BW+1 bits in its representation. Hence, we'll use this
+ // bit width during computations.
+ APInt AD = A.lshr(Mult2).zext(BW + 1); // AD = A / D
+ APInt Mod(BW + 1, 0);
+ Mod.set(BW - Mult2); // Mod = N / D
+ APInt I = AD.multiplicativeInverse(Mod);
+
+ // 4. Compute the minimum unsigned root of the equation:
+ // I * (B / D) mod (N / D)
+ APInt Result = (I * B.lshr(Mult2).zext(BW + 1)).urem(Mod);
+
+ // The result is guaranteed to be less than 2^BW so we may truncate it to BW
+ // bits.
+ return SE.getConstant(Result.trunc(BW));
+}
+
+/// SolveQuadraticEquation - Find the roots of the quadratic equation for the
+/// given quadratic chrec {L,+,M,+,N}. This returns either the two roots (which
+/// might be the same) or two SCEVCouldNotCompute objects.
+///
+static std::pair<const SCEV *,const SCEV *>
+SolveQuadraticEquation(const SCEVAddRecExpr *AddRec, ScalarEvolution &SE) {
+ assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!");
+ const SCEVConstant *LC = dyn_cast<SCEVConstant>(AddRec->getOperand(0));
+ const SCEVConstant *MC = dyn_cast<SCEVConstant>(AddRec->getOperand(1));
+ const SCEVConstant *NC = dyn_cast<SCEVConstant>(AddRec->getOperand(2));
+
+ // We currently can only solve this if the coefficients are constants.
+ if (!LC || !MC || !NC) {
+ const SCEV *CNC = SE.getCouldNotCompute();
+ return std::make_pair(CNC, CNC);
+ }
+
+ uint32_t BitWidth = LC->getValue()->getValue().getBitWidth();
+ const APInt &L = LC->getValue()->getValue();
+ const APInt &M = MC->getValue()->getValue();
+ const APInt &N = NC->getValue()->getValue();
+ APInt Two(BitWidth, 2);
+ APInt Four(BitWidth, 4);
+
+ {
+ using namespace APIntOps;
+ const APInt& C = L;
+ // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C
+ // The B coefficient is M-N/2
+ APInt B(M);
+ B -= sdiv(N,Two);
+
+ // The A coefficient is N/2
+ APInt A(N.sdiv(Two));
+
+ // Compute the B^2-4ac term.
+ APInt SqrtTerm(B);
+ SqrtTerm *= B;
+ SqrtTerm -= Four * (A * C);
+
+ // Compute sqrt(B^2-4ac). This is guaranteed to be the nearest
+ // integer value or else APInt::sqrt() will assert.
+ APInt SqrtVal(SqrtTerm.sqrt());
+
+ // Compute the two solutions for the quadratic formula.
+ // The divisions must be performed as signed divisions.
+ APInt NegB(-B);
+ APInt TwoA( A << 1 );
+ if (TwoA.isMinValue()) {
+ const SCEV *CNC = SE.getCouldNotCompute();
+ return std::make_pair(CNC, CNC);
+ }
+
+ LLVMContext &Context = SE.getContext();
+
+ ConstantInt *Solution1 =
+ ConstantInt::get(Context, (NegB + SqrtVal).sdiv(TwoA));
+ ConstantInt *Solution2 =
+ ConstantInt::get(Context, (NegB - SqrtVal).sdiv(TwoA));
+
+ return std::make_pair(SE.getConstant(Solution1),
+ SE.getConstant(Solution2));
+ } // end APIntOps namespace
+}
+
+/// HowFarToZero - Return the number of times a backedge comparing the specified
+/// value to zero will execute. If not computable, return CouldNotCompute.
+const SCEV *ScalarEvolution::HowFarToZero(const SCEV *V, const Loop *L) {
+ // If the value is a constant
+ if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
+ // If the value is already zero, the branch will execute zero times.
+ if (C->getValue()->isZero()) return C;
+ return getCouldNotCompute(); // Otherwise it will loop infinitely.
+ }
+
+ const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V);
+ if (!AddRec || AddRec->getLoop() != L)
+ return getCouldNotCompute();
+
+ if (AddRec->isAffine()) {
+ // If this is an affine expression, the execution count of this branch is
+ // the minimum unsigned root of the following equation:
+ //
+ // Start + Step*N = 0 (mod 2^BW)
+ //
+ // equivalent to:
+ //
+ // Step*N = -Start (mod 2^BW)
+ //
+ // where BW is the common bit width of Start and Step.
+
+ // Get the initial value for the loop.
+ const SCEV *Start = getSCEVAtScope(AddRec->getStart(),
+ L->getParentLoop());
+ const SCEV *Step = getSCEVAtScope(AddRec->getOperand(1),
+ L->getParentLoop());
+
+ if (const SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step)) {
+ // For now we handle only constant steps.
+
+ // First, handle unitary steps.
+ if (StepC->getValue()->equalsInt(1)) // 1*N = -Start (mod 2^BW), so:
+ return getNegativeSCEV(Start); // N = -Start (as unsigned)
+ if (StepC->getValue()->isAllOnesValue()) // -1*N = -Start (mod 2^BW), so:
+ return Start; // N = Start (as unsigned)
+
+ // Then, try to solve the above equation provided that Start is constant.
+ if (const SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start))
+ return SolveLinEquationWithOverflow(StepC->getValue()->getValue(),
+ -StartC->getValue()->getValue(),
+ *this);
+ }
+ } else if (AddRec->isQuadratic() && AddRec->getType()->isInteger()) {
+ // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of
+ // the quadratic equation to solve it.
+ std::pair<const SCEV *,const SCEV *> Roots = SolveQuadraticEquation(AddRec,
+ *this);
+ const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
+ const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
+ if (R1) {
+#if 0
+ dbgs() << "HFTZ: " << *V << " - sol#1: " << *R1
+ << " sol#2: " << *R2 << "\n";
+#endif
+ // Pick the smallest positive root value.
+ if (ConstantInt *CB =
+ dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
+ R1->getValue(), R2->getValue()))) {
+ if (CB->getZExtValue() == false)
+ std::swap(R1, R2); // R1 is the minimum root now.
+
+ // We can only use this value if the chrec ends up with an exact zero
+ // value at this index. When solving for "X*X != 5", for example, we
+ // should not accept a root of 2.
+ const SCEV *Val = AddRec->evaluateAtIteration(R1, *this);
+ if (Val->isZero())
+ return R1; // We found a quadratic root!
+ }
+ }
+ }
+
+ return getCouldNotCompute();
+}
+
+/// HowFarToNonZero - Return the number of times a backedge checking the
+/// specified value for nonzero will execute. If not computable, return
+/// CouldNotCompute
+const SCEV *ScalarEvolution::HowFarToNonZero(const SCEV *V, const Loop *L) {
+ // Loops that look like: while (X == 0) are very strange indeed. We don't
+ // handle them yet except for the trivial case. This could be expanded in the
+ // future as needed.
+
+ // If the value is a constant, check to see if it is known to be non-zero
+ // already. If so, the backedge will execute zero times.
+ if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
+ if (!C->getValue()->isNullValue())
+ return getIntegerSCEV(0, C->getType());
+ return getCouldNotCompute(); // Otherwise it will loop infinitely.
+ }
+
+ // We could implement others, but I really doubt anyone writes loops like
+ // this, and if they did, they would already be constant folded.
+ return getCouldNotCompute();
+}
+
+/// getLoopPredecessor - If the given loop's header has exactly one unique
+/// predecessor outside the loop, return it. Otherwise return null.
+///
+BasicBlock *ScalarEvolution::getLoopPredecessor(const Loop *L) {
+ BasicBlock *Header = L->getHeader();
+ BasicBlock *Pred = 0;
+ for (pred_iterator PI = pred_begin(Header), E = pred_end(Header);
+ PI != E; ++PI)
+ if (!L->contains(*PI)) {
+ if (Pred && Pred != *PI) return 0; // Multiple predecessors.
+ Pred = *PI;
+ }
+ return Pred;
+}
+
+/// getPredecessorWithUniqueSuccessorForBB - Return a predecessor of BB
+/// (which may not be an immediate predecessor) which has exactly one
+/// successor from which BB is reachable, or null if no such block is
+/// found.
+///
+BasicBlock *
+ScalarEvolution::getPredecessorWithUniqueSuccessorForBB(BasicBlock *BB) {
+ // If the block has a unique predecessor, then there is no path from the
+ // predecessor to the block that does not go through the direct edge
+ // from the predecessor to the block.
+ if (BasicBlock *Pred = BB->getSinglePredecessor())
+ return Pred;
+
+ // A loop's header is defined to be a block that dominates the loop.
+ // If the header has a unique predecessor outside the loop, it must be
+ // a block that has exactly one successor that can reach the loop.
+ if (Loop *L = LI->getLoopFor(BB))
+ return getLoopPredecessor(L);
+
+ return 0;
+}
+
+/// HasSameValue - SCEV structural equivalence is usually sufficient for
+/// testing whether two expressions are equal, however for the purposes of
+/// looking for a condition guarding a loop, it can be useful to be a little
+/// more general, since a front-end may have replicated the controlling
+/// expression.
+///
+static bool HasSameValue(const SCEV *A, const SCEV *B) {
+ // Quick check to see if they are the same SCEV.
+ if (A == B) return true;
+
+ // Otherwise, if they're both SCEVUnknown, it's possible that they hold
+ // two different instructions with the same value. Check for this case.
+ if (const SCEVUnknown *AU = dyn_cast<SCEVUnknown>(A))
+ if (const SCEVUnknown *BU = dyn_cast<SCEVUnknown>(B))
+ if (const Instruction *AI = dyn_cast<Instruction>(AU->getValue()))
+ if (const Instruction *BI = dyn_cast<Instruction>(BU->getValue()))
+ if (AI->isIdenticalTo(BI) && !AI->mayReadFromMemory())
+ return true;
+
+ // Otherwise assume they may have a different value.
+ return false;
+}
+
+bool ScalarEvolution::isKnownNegative(const SCEV *S) {
+ return getSignedRange(S).getSignedMax().isNegative();
+}
+
+bool ScalarEvolution::isKnownPositive(const SCEV *S) {
+ return getSignedRange(S).getSignedMin().isStrictlyPositive();
+}
+
+bool ScalarEvolution::isKnownNonNegative(const SCEV *S) {
+ return !getSignedRange(S).getSignedMin().isNegative();
+}
+
+bool ScalarEvolution::isKnownNonPositive(const SCEV *S) {
+ return !getSignedRange(S).getSignedMax().isStrictlyPositive();
+}
+
+bool ScalarEvolution::isKnownNonZero(const SCEV *S) {
+ return isKnownNegative(S) || isKnownPositive(S);
+}
+
+bool ScalarEvolution::isKnownPredicate(ICmpInst::Predicate Pred,
+ const SCEV *LHS, const SCEV *RHS) {
+
+ if (HasSameValue(LHS, RHS))
+ return ICmpInst::isTrueWhenEqual(Pred);
+
+ switch (Pred) {
+ default:
+ llvm_unreachable("Unexpected ICmpInst::Predicate value!");
+ break;
+ case ICmpInst::ICMP_SGT:
+ Pred = ICmpInst::ICMP_SLT;
+ std::swap(LHS, RHS);
+ case ICmpInst::ICMP_SLT: {
+ ConstantRange LHSRange = getSignedRange(LHS);
+ ConstantRange RHSRange = getSignedRange(RHS);
+ if (LHSRange.getSignedMax().slt(RHSRange.getSignedMin()))
+ return true;
+ if (LHSRange.getSignedMin().sge(RHSRange.getSignedMax()))
+ return false;
+ break;
+ }
+ case ICmpInst::ICMP_SGE:
+ Pred = ICmpInst::ICMP_SLE;
+ std::swap(LHS, RHS);
+ case ICmpInst::ICMP_SLE: {
+ ConstantRange LHSRange = getSignedRange(LHS);
+ ConstantRange RHSRange = getSignedRange(RHS);
+ if (LHSRange.getSignedMax().sle(RHSRange.getSignedMin()))
+ return true;
+ if (LHSRange.getSignedMin().sgt(RHSRange.getSignedMax()))
+ return false;
+ break;
+ }
+ case ICmpInst::ICMP_UGT:
+ Pred = ICmpInst::ICMP_ULT;
+ std::swap(LHS, RHS);
+ case ICmpInst::ICMP_ULT: {
+ ConstantRange LHSRange = getUnsignedRange(LHS);
+ ConstantRange RHSRange = getUnsignedRange(RHS);
+ if (LHSRange.getUnsignedMax().ult(RHSRange.getUnsignedMin()))
+ return true;
+ if (LHSRange.getUnsignedMin().uge(RHSRange.getUnsignedMax()))
+ return false;
+ break;
+ }
+ case ICmpInst::ICMP_UGE:
+ Pred = ICmpInst::ICMP_ULE;
+ std::swap(LHS, RHS);
+ case ICmpInst::ICMP_ULE: {
+ ConstantRange LHSRange = getUnsignedRange(LHS);
+ ConstantRange RHSRange = getUnsignedRange(RHS);
+ if (LHSRange.getUnsignedMax().ule(RHSRange.getUnsignedMin()))
+ return true;
+ if (LHSRange.getUnsignedMin().ugt(RHSRange.getUnsignedMax()))
+ return false;
+ break;
+ }
+ case ICmpInst::ICMP_NE: {
+ if (getUnsignedRange(LHS).intersectWith(getUnsignedRange(RHS)).isEmptySet())
+ return true;
+ if (getSignedRange(LHS).intersectWith(getSignedRange(RHS)).isEmptySet())
+ return true;
+
+ const SCEV *Diff = getMinusSCEV(LHS, RHS);
+ if (isKnownNonZero(Diff))
+ return true;
+ break;
+ }
+ case ICmpInst::ICMP_EQ:
+ // The check at the top of the function catches the case where
+ // the values are known to be equal.
+ break;
+ }
+ return false;
+}
+
+/// isLoopBackedgeGuardedByCond - Test whether the backedge of the loop is
+/// protected by a conditional between LHS and RHS. This is used to
+/// to eliminate casts.
+bool
+ScalarEvolution::isLoopBackedgeGuardedByCond(const Loop *L,
+ ICmpInst::Predicate Pred,
+ const SCEV *LHS, const SCEV *RHS) {
+ // Interpret a null as meaning no loop, where there is obviously no guard
+ // (interprocedural conditions notwithstanding).
+ if (!L) return true;
+
+ BasicBlock *Latch = L->getLoopLatch();
+ if (!Latch)
+ return false;
+
+ BranchInst *LoopContinuePredicate =
+ dyn_cast<BranchInst>(Latch->getTerminator());
+ if (!LoopContinuePredicate ||
+ LoopContinuePredicate->isUnconditional())
+ return false;
+
+ return isImpliedCond(LoopContinuePredicate->getCondition(), Pred, LHS, RHS,
+ LoopContinuePredicate->getSuccessor(0) != L->getHeader());
+}
+
+/// isLoopGuardedByCond - Test whether entry to the loop is protected
+/// by a conditional between LHS and RHS. This is used to help avoid max
+/// expressions in loop trip counts, and to eliminate casts.
+bool
+ScalarEvolution::isLoopGuardedByCond(const Loop *L,
+ ICmpInst::Predicate Pred,
+ const SCEV *LHS, const SCEV *RHS) {
+ // Interpret a null as meaning no loop, where there is obviously no guard
+ // (interprocedural conditions notwithstanding).
+ if (!L) return false;
+
+ BasicBlock *Predecessor = getLoopPredecessor(L);
+ BasicBlock *PredecessorDest = L->getHeader();
+
+ // Starting at the loop predecessor, climb up the predecessor chain, as long
+ // as there are predecessors that can be found that have unique successors
+ // leading to the original header.
+ for (; Predecessor;
+ PredecessorDest = Predecessor,
+ Predecessor = getPredecessorWithUniqueSuccessorForBB(Predecessor)) {
+
+ BranchInst *LoopEntryPredicate =
+ dyn_cast<BranchInst>(Predecessor->getTerminator());
+ if (!LoopEntryPredicate ||
+ LoopEntryPredicate->isUnconditional())
+ continue;
+
+ if (isImpliedCond(LoopEntryPredicate->getCondition(), Pred, LHS, RHS,
+ LoopEntryPredicate->getSuccessor(0) != PredecessorDest))
+ return true;
+ }
+
+ return false;
+}
+
+/// isImpliedCond - Test whether the condition described by Pred, LHS,
+/// and RHS is true whenever the given Cond value evaluates to true.
+bool ScalarEvolution::isImpliedCond(Value *CondValue,
+ ICmpInst::Predicate Pred,
+ const SCEV *LHS, const SCEV *RHS,
+ bool Inverse) {
+ // Recursivly handle And and Or conditions.
+ if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CondValue)) {
+ if (BO->getOpcode() == Instruction::And) {
+ if (!Inverse)
+ return isImpliedCond(BO->getOperand(0), Pred, LHS, RHS, Inverse) ||
+ isImpliedCond(BO->getOperand(1), Pred, LHS, RHS, Inverse);
+ } else if (BO->getOpcode() == Instruction::Or) {
+ if (Inverse)
+ return isImpliedCond(BO->getOperand(0), Pred, LHS, RHS, Inverse) ||
+ isImpliedCond(BO->getOperand(1), Pred, LHS, RHS, Inverse);
+ }
+ }
+
+ ICmpInst *ICI = dyn_cast<ICmpInst>(CondValue);
+ if (!ICI) return false;
+
+ // Bail if the ICmp's operands' types are wider than the needed type
+ // before attempting to call getSCEV on them. This avoids infinite
+ // recursion, since the analysis of widening casts can require loop
+ // exit condition information for overflow checking, which would
+ // lead back here.
+ if (getTypeSizeInBits(LHS->getType()) <
+ getTypeSizeInBits(ICI->getOperand(0)->getType()))
+ return false;
+
+ // Now that we found a conditional branch that dominates the loop, check to
+ // see if it is the comparison we are looking for.
+ ICmpInst::Predicate FoundPred;
+ if (Inverse)
+ FoundPred = ICI->getInversePredicate();
+ else
+ FoundPred = ICI->getPredicate();
+
+ const SCEV *FoundLHS = getSCEV(ICI->getOperand(0));
+ const SCEV *FoundRHS = getSCEV(ICI->getOperand(1));
+
+ // Balance the types. The case where FoundLHS' type is wider than
+ // LHS' type is checked for above.
+ if (getTypeSizeInBits(LHS->getType()) >
+ getTypeSizeInBits(FoundLHS->getType())) {
+ if (CmpInst::isSigned(Pred)) {
+ FoundLHS = getSignExtendExpr(FoundLHS, LHS->getType());
+ FoundRHS = getSignExtendExpr(FoundRHS, LHS->getType());
+ } else {
+ FoundLHS = getZeroExtendExpr(FoundLHS, LHS->getType());
+ FoundRHS = getZeroExtendExpr(FoundRHS, LHS->getType());
+ }
+ }
+
+ // Canonicalize the query to match the way instcombine will have
+ // canonicalized the comparison.
+ // First, put a constant operand on the right.
+ if (isa<SCEVConstant>(LHS)) {
+ std::swap(LHS, RHS);
+ Pred = ICmpInst::getSwappedPredicate(Pred);
+ }
+ // Then, canonicalize comparisons with boundary cases.
+ if (const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS)) {
+ const APInt &RA = RC->getValue()->getValue();
+ switch (Pred) {
+ default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
+ case ICmpInst::ICMP_EQ:
+ case ICmpInst::ICMP_NE:
+ break;
+ case ICmpInst::ICMP_UGE:
+ if ((RA - 1).isMinValue()) {
+ Pred = ICmpInst::ICMP_NE;
+ RHS = getConstant(RA - 1);
+ break;
+ }
+ if (RA.isMaxValue()) {
+ Pred = ICmpInst::ICMP_EQ;
+ break;
+ }
+ if (RA.isMinValue()) return true;
+ break;
+ case ICmpInst::ICMP_ULE:
+ if ((RA + 1).isMaxValue()) {
+ Pred = ICmpInst::ICMP_NE;
+ RHS = getConstant(RA + 1);
+ break;
+ }
+ if (RA.isMinValue()) {
+ Pred = ICmpInst::ICMP_EQ;
+ break;
+ }
+ if (RA.isMaxValue()) return true;
+ break;
+ case ICmpInst::ICMP_SGE:
+ if ((RA - 1).isMinSignedValue()) {
+ Pred = ICmpInst::ICMP_NE;
+ RHS = getConstant(RA - 1);
+ break;
+ }
+ if (RA.isMaxSignedValue()) {
+ Pred = ICmpInst::ICMP_EQ;
+ break;
+ }
+ if (RA.isMinSignedValue()) return true;
+ break;
+ case ICmpInst::ICMP_SLE:
+ if ((RA + 1).isMaxSignedValue()) {
+ Pred = ICmpInst::ICMP_NE;
+ RHS = getConstant(RA + 1);
+ break;
+ }
+ if (RA.isMinSignedValue()) {
+ Pred = ICmpInst::ICMP_EQ;
+ break;
+ }
+ if (RA.isMaxSignedValue()) return true;
+ break;
+ case ICmpInst::ICMP_UGT:
+ if (RA.isMinValue()) {
+ Pred = ICmpInst::ICMP_NE;
+ break;
+ }
+ if ((RA + 1).isMaxValue()) {
+ Pred = ICmpInst::ICMP_EQ;
+ RHS = getConstant(RA + 1);
+ break;
+ }
+ if (RA.isMaxValue()) return false;
+ break;
+ case ICmpInst::ICMP_ULT:
+ if (RA.isMaxValue()) {
+ Pred = ICmpInst::ICMP_NE;
+ break;
+ }
+ if ((RA - 1).isMinValue()) {
+ Pred = ICmpInst::ICMP_EQ;
+ RHS = getConstant(RA - 1);
+ break;
+ }
+ if (RA.isMinValue()) return false;
+ break;
+ case ICmpInst::ICMP_SGT:
+ if (RA.isMinSignedValue()) {
+ Pred = ICmpInst::ICMP_NE;
+ break;
+ }
+ if ((RA + 1).isMaxSignedValue()) {
+ Pred = ICmpInst::ICMP_EQ;
+ RHS = getConstant(RA + 1);
+ break;
+ }
+ if (RA.isMaxSignedValue()) return false;
+ break;
+ case ICmpInst::ICMP_SLT:
+ if (RA.isMaxSignedValue()) {
+ Pred = ICmpInst::ICMP_NE;
+ break;
+ }
+ if ((RA - 1).isMinSignedValue()) {
+ Pred = ICmpInst::ICMP_EQ;
+ RHS = getConstant(RA - 1);
+ break;
+ }
+ if (RA.isMinSignedValue()) return false;
+ break;
+ }
+ }
+
+ // Check to see if we can make the LHS or RHS match.
+ if (LHS == FoundRHS || RHS == FoundLHS) {
+ if (isa<SCEVConstant>(RHS)) {
+ std::swap(FoundLHS, FoundRHS);
+ FoundPred = ICmpInst::getSwappedPredicate(FoundPred);
+ } else {
+ std::swap(LHS, RHS);
+ Pred = ICmpInst::getSwappedPredicate(Pred);
+ }
+ }
+
+ // Check whether the found predicate is the same as the desired predicate.
+ if (FoundPred == Pred)
+ return isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS);
+
+ // Check whether swapping the found predicate makes it the same as the
+ // desired predicate.
+ if (ICmpInst::getSwappedPredicate(FoundPred) == Pred) {
+ if (isa<SCEVConstant>(RHS))
+ return isImpliedCondOperands(Pred, LHS, RHS, FoundRHS, FoundLHS);
+ else
+ return isImpliedCondOperands(ICmpInst::getSwappedPredicate(Pred),
+ RHS, LHS, FoundLHS, FoundRHS);
+ }
+
+ // Check whether the actual condition is beyond sufficient.
+ if (FoundPred == ICmpInst::ICMP_EQ)
+ if (ICmpInst::isTrueWhenEqual(Pred))
+ if (isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS))
+ return true;
+ if (Pred == ICmpInst::ICMP_NE)
+ if (!ICmpInst::isTrueWhenEqual(FoundPred))
+ if (isImpliedCondOperands(FoundPred, LHS, RHS, FoundLHS, FoundRHS))
+ return true;
+
+ // Otherwise assume the worst.
+ return false;
+}
+
+/// isImpliedCondOperands - Test whether the condition described by Pred,
+/// LHS, and RHS is true whenever the condition desribed by Pred, FoundLHS,
+/// and FoundRHS is true.
+bool ScalarEvolution::isImpliedCondOperands(ICmpInst::Predicate Pred,
+ const SCEV *LHS, const SCEV *RHS,
+ const SCEV *FoundLHS,
+ const SCEV *FoundRHS) {
+ return isImpliedCondOperandsHelper(Pred, LHS, RHS,
+ FoundLHS, FoundRHS) ||
+ // ~x < ~y --> x > y
+ isImpliedCondOperandsHelper(Pred, LHS, RHS,
+ getNotSCEV(FoundRHS),
+ getNotSCEV(FoundLHS));
+}
+
+/// isImpliedCondOperandsHelper - Test whether the condition described by
+/// Pred, LHS, and RHS is true whenever the condition desribed by Pred,
+/// FoundLHS, and FoundRHS is true.
+bool
+ScalarEvolution::isImpliedCondOperandsHelper(ICmpInst::Predicate Pred,
+ const SCEV *LHS, const SCEV *RHS,
+ const SCEV *FoundLHS,
+ const SCEV *FoundRHS) {
+ switch (Pred) {
+ default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
+ case ICmpInst::ICMP_EQ:
+ case ICmpInst::ICMP_NE:
+ if (HasSameValue(LHS, FoundLHS) && HasSameValue(RHS, FoundRHS))
+ return true;
+ break;
+ case ICmpInst::ICMP_SLT:
+ case ICmpInst::ICMP_SLE:
+ if (isKnownPredicate(ICmpInst::ICMP_SLE, LHS, FoundLHS) &&
+ isKnownPredicate(ICmpInst::ICMP_SGE, RHS, FoundRHS))
+ return true;
+ break;
+ case ICmpInst::ICMP_SGT:
+ case ICmpInst::ICMP_SGE:
+ if (isKnownPredicate(ICmpInst::ICMP_SGE, LHS, FoundLHS) &&
+ isKnownPredicate(ICmpInst::ICMP_SLE, RHS, FoundRHS))
+ return true;
+ break;
+ case ICmpInst::ICMP_ULT:
+ case ICmpInst::ICMP_ULE:
+ if (isKnownPredicate(ICmpInst::ICMP_ULE, LHS, FoundLHS) &&
+ isKnownPredicate(ICmpInst::ICMP_UGE, RHS, FoundRHS))
+ return true;
+ break;
+ case ICmpInst::ICMP_UGT:
+ case ICmpInst::ICMP_UGE:
+ if (isKnownPredicate(ICmpInst::ICMP_UGE, LHS, FoundLHS) &&
+ isKnownPredicate(ICmpInst::ICMP_ULE, RHS, FoundRHS))
+ return true;
+ break;
+ }
+
+ return false;
+}
+
+/// getBECount - Subtract the end and start values and divide by the step,
+/// rounding up, to get the number of times the backedge is executed. Return
+/// CouldNotCompute if an intermediate computation overflows.
+const SCEV *ScalarEvolution::getBECount(const SCEV *Start,
+ const SCEV *End,
+ const SCEV *Step,
+ bool NoWrap) {
+ assert(!isKnownNegative(Step) &&
+ "This code doesn't handle negative strides yet!");
+
+ const Type *Ty = Start->getType();
+ const SCEV *NegOne = getIntegerSCEV(-1, Ty);
+ const SCEV *Diff = getMinusSCEV(End, Start);
+ const SCEV *RoundUp = getAddExpr(Step, NegOne);
+
+ // Add an adjustment to the difference between End and Start so that
+ // the division will effectively round up.
+ const SCEV *Add = getAddExpr(Diff, RoundUp);
+
+ if (!NoWrap) {
+ // Check Add for unsigned overflow.
+ // TODO: More sophisticated things could be done here.
+ const Type *WideTy = IntegerType::get(getContext(),
+ getTypeSizeInBits(Ty) + 1);
+ const SCEV *EDiff = getZeroExtendExpr(Diff, WideTy);
+ const SCEV *ERoundUp = getZeroExtendExpr(RoundUp, WideTy);
+ const SCEV *OperandExtendedAdd = getAddExpr(EDiff, ERoundUp);
+ if (getZeroExtendExpr(Add, WideTy) != OperandExtendedAdd)
+ return getCouldNotCompute();
+ }
+
+ return getUDivExpr(Add, Step);
+}
+
+/// HowManyLessThans - Return the number of times a backedge containing the
+/// specified less-than comparison will execute. If not computable, return
+/// CouldNotCompute.
+ScalarEvolution::BackedgeTakenInfo
+ScalarEvolution::HowManyLessThans(const SCEV *LHS, const SCEV *RHS,
+ const Loop *L, bool isSigned) {
+ // Only handle: "ADDREC < LoopInvariant".
+ if (!RHS->isLoopInvariant(L)) return getCouldNotCompute();
+
+ const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS);
+ if (!AddRec || AddRec->getLoop() != L)
+ return getCouldNotCompute();
+
+ // Check to see if we have a flag which makes analysis easy.
+ bool NoWrap = isSigned ? AddRec->hasNoSignedWrap() :
+ AddRec->hasNoUnsignedWrap();
+
+ if (AddRec->isAffine()) {
+ unsigned BitWidth = getTypeSizeInBits(AddRec->getType());
+ const SCEV *Step = AddRec->getStepRecurrence(*this);
+
+ if (Step->isZero())
+ return getCouldNotCompute();
+ if (Step->isOne()) {
+ // With unit stride, the iteration never steps past the limit value.
+ } else if (isKnownPositive(Step)) {
+ // Test whether a positive iteration can step past the limit
+ // value and past the maximum value for its type in a single step.
+ // Note that it's not sufficient to check NoWrap here, because even
+ // though the value after a wrap is undefined, it's not undefined
+ // behavior, so if wrap does occur, the loop could either terminate or
+ // loop infinitely, but in either case, the loop is guaranteed to
+ // iterate at least until the iteration where the wrapping occurs.
+ const SCEV *One = getIntegerSCEV(1, Step->getType());
+ if (isSigned) {
+ APInt Max = APInt::getSignedMaxValue(BitWidth);
+ if ((Max - getSignedRange(getMinusSCEV(Step, One)).getSignedMax())
+ .slt(getSignedRange(RHS).getSignedMax()))
+ return getCouldNotCompute();
+ } else {
+ APInt Max = APInt::getMaxValue(BitWidth);
+ if ((Max - getUnsignedRange(getMinusSCEV(Step, One)).getUnsignedMax())
+ .ult(getUnsignedRange(RHS).getUnsignedMax()))
+ return getCouldNotCompute();
+ }
+ } else
+ // TODO: Handle negative strides here and below.
+ return getCouldNotCompute();
+
+ // We know the LHS is of the form {n,+,s} and the RHS is some loop-invariant
+ // m. So, we count the number of iterations in which {n,+,s} < m is true.
+ // Note that we cannot simply return max(m-n,0)/s because it's not safe to
+ // treat m-n as signed nor unsigned due to overflow possibility.
+
+ // First, we get the value of the LHS in the first iteration: n
+ const SCEV *Start = AddRec->getOperand(0);
+
+ // Determine the minimum constant start value.
+ const SCEV *MinStart = getConstant(isSigned ?
+ getSignedRange(Start).getSignedMin() :
+ getUnsignedRange(Start).getUnsignedMin());
+
+ // If we know that the condition is true in order to enter the loop,
+ // then we know that it will run exactly (m-n)/s times. Otherwise, we
+ // only know that it will execute (max(m,n)-n)/s times. In both cases,
+ // the division must round up.
+ const SCEV *End = RHS;
+ if (!isLoopGuardedByCond(L,
+ isSigned ? ICmpInst::ICMP_SLT :
+ ICmpInst::ICMP_ULT,
+ getMinusSCEV(Start, Step), RHS))
+ End = isSigned ? getSMaxExpr(RHS, Start)
+ : getUMaxExpr(RHS, Start);
+
+ // Determine the maximum constant end value.
+ const SCEV *MaxEnd = getConstant(isSigned ?
+ getSignedRange(End).getSignedMax() :
+ getUnsignedRange(End).getUnsignedMax());
+
+ // If MaxEnd is within a step of the maximum integer value in its type,
+ // adjust it down to the minimum value which would produce the same effect.
+ // This allows the subsequent ceiling divison of (N+(step-1))/step to
+ // compute the correct value.
+ const SCEV *StepMinusOne = getMinusSCEV(Step,
+ getIntegerSCEV(1, Step->getType()));
+ MaxEnd = isSigned ?
+ getSMinExpr(MaxEnd,
+ getMinusSCEV(getConstant(APInt::getSignedMaxValue(BitWidth)),
+ StepMinusOne)) :
+ getUMinExpr(MaxEnd,
+ getMinusSCEV(getConstant(APInt::getMaxValue(BitWidth)),
+ StepMinusOne));
+
+ // Finally, we subtract these two values and divide, rounding up, to get
+ // the number of times the backedge is executed.
+ const SCEV *BECount = getBECount(Start, End, Step, NoWrap);
+
+ // The maximum backedge count is similar, except using the minimum start
+ // value and the maximum end value.
+ const SCEV *MaxBECount = getBECount(MinStart, MaxEnd, Step, NoWrap);
+
+ return BackedgeTakenInfo(BECount, MaxBECount);
+ }
+
+ return getCouldNotCompute();
+}
+
+/// getNumIterationsInRange - Return the number of iterations of this loop that
+/// produce values in the specified constant range. Another way of looking at
+/// this is that it returns the first iteration number where the value is not in
+/// the condition, thus computing the exit count. If the iteration count can't
+/// be computed, an instance of SCEVCouldNotCompute is returned.
+const SCEV *SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range,
+ ScalarEvolution &SE) const {
+ if (Range.isFullSet()) // Infinite loop.
+ return SE.getCouldNotCompute();
+
+ // If the start is a non-zero constant, shift the range to simplify things.
+ if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart()))
+ if (!SC->getValue()->isZero()) {
+ SmallVector<const SCEV *, 4> Operands(op_begin(), op_end());
+ Operands[0] = SE.getIntegerSCEV(0, SC->getType());
+ const SCEV *Shifted = SE.getAddRecExpr(Operands, getLoop());
+ if (const SCEVAddRecExpr *ShiftedAddRec =
+ dyn_cast<SCEVAddRecExpr>(Shifted))
+ return ShiftedAddRec->getNumIterationsInRange(
+ Range.subtract(SC->getValue()->getValue()), SE);
+ // This is strange and shouldn't happen.
+ return SE.getCouldNotCompute();
+ }
+
+ // The only time we can solve this is when we have all constant indices.
+ // Otherwise, we cannot determine the overflow conditions.
+ for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
+ if (!isa<SCEVConstant>(getOperand(i)))
+ return SE.getCouldNotCompute();
+
+
+ // Okay at this point we know that all elements of the chrec are constants and
+ // that the start element is zero.
+
+ // First check to see if the range contains zero. If not, the first
+ // iteration exits.
+ unsigned BitWidth = SE.getTypeSizeInBits(getType());
+ if (!Range.contains(APInt(BitWidth, 0)))
+ return SE.getIntegerSCEV(0, getType());
+
+ if (isAffine()) {
+ // If this is an affine expression then we have this situation:
+ // Solve {0,+,A} in Range === Ax in Range
+
+ // We know that zero is in the range. If A is positive then we know that
+ // the upper value of the range must be the first possible exit value.
+ // If A is negative then the lower of the range is the last possible loop
+ // value. Also note that we already checked for a full range.
+ APInt One(BitWidth,1);
+ APInt A = cast<SCEVConstant>(getOperand(1))->getValue()->getValue();
+ APInt End = A.sge(One) ? (Range.getUpper() - One) : Range.getLower();
+
+ // The exit value should be (End+A)/A.
+ APInt ExitVal = (End + A).udiv(A);
+ ConstantInt *ExitValue = ConstantInt::get(SE.getContext(), ExitVal);
+
+ // Evaluate at the exit value. If we really did fall out of the valid
+ // range, then we computed our trip count, otherwise wrap around or other
+ // things must have happened.
+ ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue, SE);
+ if (Range.contains(Val->getValue()))
+ return SE.getCouldNotCompute(); // Something strange happened
+
+ // Ensure that the previous value is in the range. This is a sanity check.
+ assert(Range.contains(
+ EvaluateConstantChrecAtConstant(this,
+ ConstantInt::get(SE.getContext(), ExitVal - One), SE)->getValue()) &&
+ "Linear scev computation is off in a bad way!");
+ return SE.getConstant(ExitValue);
+ } else if (isQuadratic()) {
+ // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of the
+ // quadratic equation to solve it. To do this, we must frame our problem in
+ // terms of figuring out when zero is crossed, instead of when
+ // Range.getUpper() is crossed.
+ SmallVector<const SCEV *, 4> NewOps(op_begin(), op_end());
+ NewOps[0] = SE.getNegativeSCEV(SE.getConstant(Range.getUpper()));
+ const SCEV *NewAddRec = SE.getAddRecExpr(NewOps, getLoop());
+
+ // Next, solve the constructed addrec
+ std::pair<const SCEV *,const SCEV *> Roots =
+ SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec), SE);
+ const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
+ const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
+ if (R1) {
+ // Pick the smallest positive root value.
+ if (ConstantInt *CB =
+ dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT,
+ R1->getValue(), R2->getValue()))) {
+ if (CB->getZExtValue() == false)
+ std::swap(R1, R2); // R1 is the minimum root now.
+
+ // Make sure the root is not off by one. The returned iteration should
+ // not be in the range, but the previous one should be. When solving
+ // for "X*X < 5", for example, we should not return a root of 2.
+ ConstantInt *R1Val = EvaluateConstantChrecAtConstant(this,
+ R1->getValue(),
+ SE);
+ if (Range.contains(R1Val->getValue())) {
+ // The next iteration must be out of the range...
+ ConstantInt *NextVal =
+ ConstantInt::get(SE.getContext(), R1->getValue()->getValue()+1);
+
+ R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
+ if (!Range.contains(R1Val->getValue()))
+ return SE.getConstant(NextVal);
+ return SE.getCouldNotCompute(); // Something strange happened
+ }
+
+ // If R1 was not in the range, then it is a good return value. Make
+ // sure that R1-1 WAS in the range though, just in case.
+ ConstantInt *NextVal =
+ ConstantInt::get(SE.getContext(), R1->getValue()->getValue()-1);
+ R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
+ if (Range.contains(R1Val->getValue()))
+ return R1;
+ return SE.getCouldNotCompute(); // Something strange happened
+ }
+ }
+ }
+
+ return SE.getCouldNotCompute();
+}
+
+
+
+//===----------------------------------------------------------------------===//
+// SCEVCallbackVH Class Implementation
+//===----------------------------------------------------------------------===//
+
+void ScalarEvolution::SCEVCallbackVH::deleted() {
+ assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
+ if (PHINode *PN = dyn_cast<PHINode>(getValPtr()))
+ SE->ConstantEvolutionLoopExitValue.erase(PN);
+ SE->Scalars.erase(getValPtr());
+ // this now dangles!
+}
+
+void ScalarEvolution::SCEVCallbackVH::allUsesReplacedWith(Value *) {
+ assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
+
+ // Forget all the expressions associated with users of the old value,
+ // so that future queries will recompute the expressions using the new
+ // value.
+ SmallVector<User *, 16> Worklist;
+ SmallPtrSet<User *, 8> Visited;
+ Value *Old = getValPtr();
+ bool DeleteOld = false;
+ for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end();
+ UI != UE; ++UI)
+ Worklist.push_back(*UI);
+ while (!Worklist.empty()) {
+ User *U = Worklist.pop_back_val();
+ // Deleting the Old value will cause this to dangle. Postpone
+ // that until everything else is done.
+ if (U == Old) {
+ DeleteOld = true;
+ continue;
+ }
+ if (!Visited.insert(U))
+ continue;
+ if (PHINode *PN = dyn_cast<PHINode>(U))
+ SE->ConstantEvolutionLoopExitValue.erase(PN);
+ SE->Scalars.erase(U);
+ for (Value::use_iterator UI = U->use_begin(), UE = U->use_end();
+ UI != UE; ++UI)
+ Worklist.push_back(*UI);
+ }
+ // Delete the Old value if it (indirectly) references itself.
+ if (DeleteOld) {
+ if (PHINode *PN = dyn_cast<PHINode>(Old))
+ SE->ConstantEvolutionLoopExitValue.erase(PN);
+ SE->Scalars.erase(Old);
+ // this now dangles!
+ }
+ // this may dangle!
+}
+
+ScalarEvolution::SCEVCallbackVH::SCEVCallbackVH(Value *V, ScalarEvolution *se)
+ : CallbackVH(V), SE(se) {}
+
+//===----------------------------------------------------------------------===//
+// ScalarEvolution Class Implementation
+//===----------------------------------------------------------------------===//
+
+ScalarEvolution::ScalarEvolution()
+ : FunctionPass(&ID) {
+}
+
+bool ScalarEvolution::runOnFunction(Function &F) {
+ this->F = &F;
+ LI = &getAnalysis<LoopInfo>();
+ DT = &getAnalysis<DominatorTree>();
+ TD = getAnalysisIfAvailable<TargetData>();
+ return false;
+}
+
+void ScalarEvolution::releaseMemory() {
+ Scalars.clear();
+ BackedgeTakenCounts.clear();
+ ConstantEvolutionLoopExitValue.clear();
+ ValuesAtScopes.clear();
+ UniqueSCEVs.clear();
+ SCEVAllocator.Reset();
+}
+
+void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.setPreservesAll();
+ AU.addRequiredTransitive<LoopInfo>();
+ AU.addRequiredTransitive<DominatorTree>();
+}
+
+bool ScalarEvolution::hasLoopInvariantBackedgeTakenCount(const Loop *L) {
+ return !isa<SCEVCouldNotCompute>(getBackedgeTakenCount(L));
+}
+
+static void PrintLoopInfo(raw_ostream &OS, ScalarEvolution *SE,
+ const Loop *L) {
+ // Print all inner loops first
+ for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
+ PrintLoopInfo(OS, SE, *I);
+
+ OS << "Loop ";
+ WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
+ OS << ": ";
+
+ SmallVector<BasicBlock *, 8> ExitBlocks;
+ L->getExitBlocks(ExitBlocks);
+ if (ExitBlocks.size() != 1)
+ OS << "<multiple exits> ";
+
+ if (SE->hasLoopInvariantBackedgeTakenCount(L)) {
+ OS << "backedge-taken count is " << *SE->getBackedgeTakenCount(L);
+ } else {
+ OS << "Unpredictable backedge-taken count. ";
+ }
+
+ OS << "\n"
+ "Loop ";
+ WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
+ OS << ": ";
+
+ if (!isa<SCEVCouldNotCompute>(SE->getMaxBackedgeTakenCount(L))) {
+ OS << "max backedge-taken count is " << *SE->getMaxBackedgeTakenCount(L);
+ } else {
+ OS << "Unpredictable max backedge-taken count. ";
+ }
+
+ OS << "\n";
+}
+
+void ScalarEvolution::print(raw_ostream &OS, const Module *) const {
+ // ScalarEvolution's implementaiton of the print method is to print
+ // out SCEV values of all instructions that are interesting. Doing
+ // this potentially causes it to create new SCEV objects though,
+ // which technically conflicts with the const qualifier. This isn't
+ // observable from outside the class though, so casting away the
+ // const isn't dangerous.
+ ScalarEvolution &SE = *const_cast<ScalarEvolution *>(this);
+
+ OS << "Classifying expressions for: ";
+ WriteAsOperand(OS, F, /*PrintType=*/false);
+ OS << "\n";
+ for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
+ if (isSCEVable(I->getType())) {
+ OS << *I << '\n';
+ OS << " --> ";
+ const SCEV *SV = SE.getSCEV(&*I);
+ SV->print(OS);
+
+ const Loop *L = LI->getLoopFor((*I).getParent());
+
+ const SCEV *AtUse = SE.getSCEVAtScope(SV, L);
+ if (AtUse != SV) {
+ OS << " --> ";
+ AtUse->print(OS);
+ }
+
+ if (L) {
+ OS << "\t\t" "Exits: ";
+ const SCEV *ExitValue = SE.getSCEVAtScope(SV, L->getParentLoop());
+ if (!ExitValue->isLoopInvariant(L)) {
+ OS << "<<Unknown>>";
+ } else {
+ OS << *ExitValue;
+ }
+ }
+
+ OS << "\n";
+ }
+
+ OS << "Determining loop execution counts for: ";
+ WriteAsOperand(OS, F, /*PrintType=*/false);
+ OS << "\n";
+ for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
+ PrintLoopInfo(OS, &SE, *I);
+}
+
diff --git a/lib/Analysis/ScalarEvolutionAliasAnalysis.cpp b/lib/Analysis/ScalarEvolutionAliasAnalysis.cpp
new file mode 100644
index 0000000..498c4a8
--- /dev/null
+++ b/lib/Analysis/ScalarEvolutionAliasAnalysis.cpp
@@ -0,0 +1,139 @@
+//===- ScalarEvolutionAliasAnalysis.cpp - SCEV-based Alias Analysis -------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file defines the ScalarEvolutionAliasAnalysis pass, which implements a
+// simple alias analysis implemented in terms of ScalarEvolution queries.
+//
+// ScalarEvolution has a more complete understanding of pointer arithmetic
+// than BasicAliasAnalysis' collection of ad-hoc analyses.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/Analysis/ScalarEvolutionExpressions.h"
+#include "llvm/Analysis/Passes.h"
+#include "llvm/Pass.h"
+using namespace llvm;
+
+namespace {
+ /// ScalarEvolutionAliasAnalysis - This is a simple alias analysis
+ /// implementation that uses ScalarEvolution to answer queries.
+ class ScalarEvolutionAliasAnalysis : public FunctionPass,
+ public AliasAnalysis {
+ ScalarEvolution *SE;
+
+ public:
+ static char ID; // Class identification, replacement for typeinfo
+ ScalarEvolutionAliasAnalysis() : FunctionPass(&ID), SE(0) {}
+
+ /// getAdjustedAnalysisPointer - This method is used when a pass implements
+ /// an analysis interface through multiple inheritance. If needed, it
+ /// should override this to adjust the this pointer as needed for the
+ /// specified pass info.
+ virtual void *getAdjustedAnalysisPointer(const PassInfo *PI) {
+ if (PI->isPassID(&AliasAnalysis::ID))
+ return (AliasAnalysis*)this;
+ return this;
+ }
+
+ private:
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const;
+ virtual bool runOnFunction(Function &F);
+ virtual AliasResult alias(const Value *V1, unsigned V1Size,
+ const Value *V2, unsigned V2Size);
+
+ Value *GetBaseValue(const SCEV *S);
+ };
+} // End of anonymous namespace
+
+// Register this pass...
+char ScalarEvolutionAliasAnalysis::ID = 0;
+static RegisterPass<ScalarEvolutionAliasAnalysis>
+X("scev-aa", "ScalarEvolution-based Alias Analysis", false, true);
+
+// Declare that we implement the AliasAnalysis interface
+static RegisterAnalysisGroup<AliasAnalysis> Y(X);
+
+FunctionPass *llvm::createScalarEvolutionAliasAnalysisPass() {
+ return new ScalarEvolutionAliasAnalysis();
+}
+
+void
+ScalarEvolutionAliasAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.addRequiredTransitive<ScalarEvolution>();
+ AU.setPreservesAll();
+ AliasAnalysis::getAnalysisUsage(AU);
+}
+
+bool
+ScalarEvolutionAliasAnalysis::runOnFunction(Function &F) {
+ InitializeAliasAnalysis(this);
+ SE = &getAnalysis<ScalarEvolution>();
+ return false;
+}
+
+/// GetBaseValue - Given an expression, try to find a
+/// base value. Return null is none was found.
+Value *
+ScalarEvolutionAliasAnalysis::GetBaseValue(const SCEV *S) {
+ if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
+ // In an addrec, assume that the base will be in the start, rather
+ // than the step.
+ return GetBaseValue(AR->getStart());
+ } else if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(S)) {
+ // If there's a pointer operand, it'll be sorted at the end of the list.
+ const SCEV *Last = A->getOperand(A->getNumOperands()-1);
+ if (isa<PointerType>(Last->getType()))
+ return GetBaseValue(Last);
+ } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
+ // This is a leaf node.
+ return U->getValue();
+ }
+ // No Identified object found.
+ return 0;
+}
+
+AliasAnalysis::AliasResult
+ScalarEvolutionAliasAnalysis::alias(const Value *A, unsigned ASize,
+ const Value *B, unsigned BSize) {
+ // This is ScalarEvolutionAliasAnalysis. Get the SCEVs!
+ const SCEV *AS = SE->getSCEV(const_cast<Value *>(A));
+ const SCEV *BS = SE->getSCEV(const_cast<Value *>(B));
+
+ // If they evaluate to the same expression, it's a MustAlias.
+ if (AS == BS) return MustAlias;
+
+ // If something is known about the difference between the two addresses,
+ // see if it's enough to prove a NoAlias.
+ if (SE->getEffectiveSCEVType(AS->getType()) ==
+ SE->getEffectiveSCEVType(BS->getType())) {
+ unsigned BitWidth = SE->getTypeSizeInBits(AS->getType());
+ APInt AI(BitWidth, ASize);
+ const SCEV *BA = SE->getMinusSCEV(BS, AS);
+ if (AI.ule(SE->getUnsignedRange(BA).getUnsignedMin())) {
+ APInt BI(BitWidth, BSize);
+ const SCEV *AB = SE->getMinusSCEV(AS, BS);
+ if (BI.ule(SE->getUnsignedRange(AB).getUnsignedMin()))
+ return NoAlias;
+ }
+ }
+
+ // If ScalarEvolution can find an underlying object, form a new query.
+ // The correctness of this depends on ScalarEvolution not recognizing
+ // inttoptr and ptrtoint operators.
+ Value *AO = GetBaseValue(AS);
+ Value *BO = GetBaseValue(BS);
+ if ((AO && AO != A) || (BO && BO != B))
+ if (alias(AO ? AO : A, AO ? ~0u : ASize,
+ BO ? BO : B, BO ? ~0u : BSize) == NoAlias)
+ return NoAlias;
+
+ // Forward the query to the next analysis.
+ return AliasAnalysis::alias(A, ASize, B, BSize);
+}
diff --git a/lib/Analysis/ScalarEvolutionExpander.cpp b/lib/Analysis/ScalarEvolutionExpander.cpp
new file mode 100644
index 0000000..4310e3c
--- /dev/null
+++ b/lib/Analysis/ScalarEvolutionExpander.cpp
@@ -0,0 +1,1076 @@
+//===- ScalarEvolutionExpander.cpp - Scalar Evolution Analysis --*- C++ -*-===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file contains the implementation of the scalar evolution expander,
+// which is used to generate the code corresponding to a given scalar evolution
+// expression.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/ScalarEvolutionExpander.h"
+#include "llvm/Analysis/LoopInfo.h"
+#include "llvm/LLVMContext.h"
+#include "llvm/Target/TargetData.h"
+#include "llvm/ADT/STLExtras.h"
+using namespace llvm;
+
+/// InsertNoopCastOfTo - Insert a cast of V to the specified type,
+/// which must be possible with a noop cast, doing what we can to share
+/// the casts.
+Value *SCEVExpander::InsertNoopCastOfTo(Value *V, const Type *Ty) {
+ Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false);
+ assert((Op == Instruction::BitCast ||
+ Op == Instruction::PtrToInt ||
+ Op == Instruction::IntToPtr) &&
+ "InsertNoopCastOfTo cannot perform non-noop casts!");
+ assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) &&
+ "InsertNoopCastOfTo cannot change sizes!");
+
+ // Short-circuit unnecessary bitcasts.
+ if (Op == Instruction::BitCast && V->getType() == Ty)
+ return V;
+
+ // Short-circuit unnecessary inttoptr<->ptrtoint casts.
+ if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) &&
+ SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) {
+ if (CastInst *CI = dyn_cast<CastInst>(V))
+ if ((CI->getOpcode() == Instruction::PtrToInt ||
+ CI->getOpcode() == Instruction::IntToPtr) &&
+ SE.getTypeSizeInBits(CI->getType()) ==
+ SE.getTypeSizeInBits(CI->getOperand(0)->getType()))
+ return CI->getOperand(0);
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
+ if ((CE->getOpcode() == Instruction::PtrToInt ||
+ CE->getOpcode() == Instruction::IntToPtr) &&
+ SE.getTypeSizeInBits(CE->getType()) ==
+ SE.getTypeSizeInBits(CE->getOperand(0)->getType()))
+ return CE->getOperand(0);
+ }
+
+ if (Constant *C = dyn_cast<Constant>(V))
+ return ConstantExpr::getCast(Op, C, Ty);
+
+ if (Argument *A = dyn_cast<Argument>(V)) {
+ // Check to see if there is already a cast!
+ for (Value::use_iterator UI = A->use_begin(), E = A->use_end();
+ UI != E; ++UI)
+ if ((*UI)->getType() == Ty)
+ if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI)))
+ if (CI->getOpcode() == Op) {
+ // If the cast isn't the first instruction of the function, move it.
+ if (BasicBlock::iterator(CI) !=
+ A->getParent()->getEntryBlock().begin()) {
+ // Recreate the cast at the beginning of the entry block.
+ // The old cast is left in place in case it is being used
+ // as an insert point.
+ Instruction *NewCI =
+ CastInst::Create(Op, V, Ty, "",
+ A->getParent()->getEntryBlock().begin());
+ NewCI->takeName(CI);
+ CI->replaceAllUsesWith(NewCI);
+ return NewCI;
+ }
+ return CI;
+ }
+
+ Instruction *I = CastInst::Create(Op, V, Ty, V->getName(),
+ A->getParent()->getEntryBlock().begin());
+ rememberInstruction(I);
+ return I;
+ }
+
+ Instruction *I = cast<Instruction>(V);
+
+ // Check to see if there is already a cast. If there is, use it.
+ for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
+ UI != E; ++UI) {
+ if ((*UI)->getType() == Ty)
+ if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI)))
+ if (CI->getOpcode() == Op) {
+ BasicBlock::iterator It = I; ++It;
+ if (isa<InvokeInst>(I))
+ It = cast<InvokeInst>(I)->getNormalDest()->begin();
+ while (isa<PHINode>(It)) ++It;
+ if (It != BasicBlock::iterator(CI)) {
+ // Recreate the cast after the user.
+ // The old cast is left in place in case it is being used
+ // as an insert point.
+ Instruction *NewCI = CastInst::Create(Op, V, Ty, "", It);
+ NewCI->takeName(CI);
+ CI->replaceAllUsesWith(NewCI);
+ rememberInstruction(NewCI);
+ return NewCI;
+ }
+ rememberInstruction(CI);
+ return CI;
+ }
+ }
+ BasicBlock::iterator IP = I; ++IP;
+ if (InvokeInst *II = dyn_cast<InvokeInst>(I))
+ IP = II->getNormalDest()->begin();
+ while (isa<PHINode>(IP)) ++IP;
+ Instruction *CI = CastInst::Create(Op, V, Ty, V->getName(), IP);
+ rememberInstruction(CI);
+ return CI;
+}
+
+/// InsertBinop - Insert the specified binary operator, doing a small amount
+/// of work to avoid inserting an obviously redundant operation.
+Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode,
+ Value *LHS, Value *RHS) {
+ // Fold a binop with constant operands.
+ if (Constant *CLHS = dyn_cast<Constant>(LHS))
+ if (Constant *CRHS = dyn_cast<Constant>(RHS))
+ return ConstantExpr::get(Opcode, CLHS, CRHS);
+
+ // Do a quick scan to see if we have this binop nearby. If so, reuse it.
+ unsigned ScanLimit = 6;
+ BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
+ // Scanning starts from the last instruction before the insertion point.
+ BasicBlock::iterator IP = Builder.GetInsertPoint();
+ if (IP != BlockBegin) {
+ --IP;
+ for (; ScanLimit; --IP, --ScanLimit) {
+ if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
+ IP->getOperand(1) == RHS)
+ return IP;
+ if (IP == BlockBegin) break;
+ }
+ }
+
+ // If we haven't found this binop, insert it.
+ Value *BO = Builder.CreateBinOp(Opcode, LHS, RHS, "tmp");
+ rememberInstruction(BO);
+ return BO;
+}
+
+/// FactorOutConstant - Test if S is divisible by Factor, using signed
+/// division. If so, update S with Factor divided out and return true.
+/// S need not be evenly divisble if a reasonable remainder can be
+/// computed.
+/// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
+/// unnecessary; in its place, just signed-divide Ops[i] by the scale and
+/// check to see if the divide was folded.
+static bool FactorOutConstant(const SCEV *&S,
+ const SCEV *&Remainder,
+ const SCEV *Factor,
+ ScalarEvolution &SE,
+ const TargetData *TD) {
+ // Everything is divisible by one.
+ if (Factor->isOne())
+ return true;
+
+ // x/x == 1.
+ if (S == Factor) {
+ S = SE.getIntegerSCEV(1, S->getType());
+ return true;
+ }
+
+ // For a Constant, check for a multiple of the given factor.
+ if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
+ // 0/x == 0.
+ if (C->isZero())
+ return true;
+ // Check for divisibility.
+ if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) {
+ ConstantInt *CI =
+ ConstantInt::get(SE.getContext(),
+ C->getValue()->getValue().sdiv(
+ FC->getValue()->getValue()));
+ // If the quotient is zero and the remainder is non-zero, reject
+ // the value at this scale. It will be considered for subsequent
+ // smaller scales.
+ if (!CI->isZero()) {
+ const SCEV *Div = SE.getConstant(CI);
+ S = Div;
+ Remainder =
+ SE.getAddExpr(Remainder,
+ SE.getConstant(C->getValue()->getValue().srem(
+ FC->getValue()->getValue())));
+ return true;
+ }
+ }
+ }
+
+ // In a Mul, check if there is a constant operand which is a multiple
+ // of the given factor.
+ if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
+ if (TD) {
+ // With TargetData, the size is known. Check if there is a constant
+ // operand which is a multiple of the given factor. If so, we can
+ // factor it.
+ const SCEVConstant *FC = cast<SCEVConstant>(Factor);
+ if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
+ if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) {
+ const SmallVectorImpl<const SCEV *> &MOperands = M->getOperands();
+ SmallVector<const SCEV *, 4> NewMulOps(MOperands.begin(),
+ MOperands.end());
+ NewMulOps[0] =
+ SE.getConstant(C->getValue()->getValue().sdiv(
+ FC->getValue()->getValue()));
+ S = SE.getMulExpr(NewMulOps);
+ return true;
+ }
+ } else {
+ // Without TargetData, check if Factor can be factored out of any of the
+ // Mul's operands. If so, we can just remove it.
+ for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
+ const SCEV *SOp = M->getOperand(i);
+ const SCEV *Remainder = SE.getIntegerSCEV(0, SOp->getType());
+ if (FactorOutConstant(SOp, Remainder, Factor, SE, TD) &&
+ Remainder->isZero()) {
+ const SmallVectorImpl<const SCEV *> &MOperands = M->getOperands();
+ SmallVector<const SCEV *, 4> NewMulOps(MOperands.begin(),
+ MOperands.end());
+ NewMulOps[i] = SOp;
+ S = SE.getMulExpr(NewMulOps);
+ return true;
+ }
+ }
+ }
+ }
+
+ // In an AddRec, check if both start and step are divisible.
+ if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
+ const SCEV *Step = A->getStepRecurrence(SE);
+ const SCEV *StepRem = SE.getIntegerSCEV(0, Step->getType());
+ if (!FactorOutConstant(Step, StepRem, Factor, SE, TD))
+ return false;
+ if (!StepRem->isZero())
+ return false;
+ const SCEV *Start = A->getStart();
+ if (!FactorOutConstant(Start, Remainder, Factor, SE, TD))
+ return false;
+ S = SE.getAddRecExpr(Start, Step, A->getLoop());
+ return true;
+ }
+
+ return false;
+}
+
+/// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs
+/// is the number of SCEVAddRecExprs present, which are kept at the end of
+/// the list.
+///
+static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops,
+ const Type *Ty,
+ ScalarEvolution &SE) {
+ unsigned NumAddRecs = 0;
+ for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i)
+ ++NumAddRecs;
+ // Group Ops into non-addrecs and addrecs.
+ SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs);
+ SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end());
+ // Let ScalarEvolution sort and simplify the non-addrecs list.
+ const SCEV *Sum = NoAddRecs.empty() ?
+ SE.getIntegerSCEV(0, Ty) :
+ SE.getAddExpr(NoAddRecs);
+ // If it returned an add, use the operands. Otherwise it simplified
+ // the sum into a single value, so just use that.
+ if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum))
+ Ops = Add->getOperands();
+ else {
+ Ops.clear();
+ if (!Sum->isZero())
+ Ops.push_back(Sum);
+ }
+ // Then append the addrecs.
+ Ops.insert(Ops.end(), AddRecs.begin(), AddRecs.end());
+}
+
+/// SplitAddRecs - Flatten a list of add operands, moving addrec start values
+/// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}.
+/// This helps expose more opportunities for folding parts of the expressions
+/// into GEP indices.
+///
+static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops,
+ const Type *Ty,
+ ScalarEvolution &SE) {
+ // Find the addrecs.
+ SmallVector<const SCEV *, 8> AddRecs;
+ for (unsigned i = 0, e = Ops.size(); i != e; ++i)
+ while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) {
+ const SCEV *Start = A->getStart();
+ if (Start->isZero()) break;
+ const SCEV *Zero = SE.getIntegerSCEV(0, Ty);
+ AddRecs.push_back(SE.getAddRecExpr(Zero,
+ A->getStepRecurrence(SE),
+ A->getLoop()));
+ if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) {
+ Ops[i] = Zero;
+ Ops.insert(Ops.end(), Add->op_begin(), Add->op_end());
+ e += Add->getNumOperands();
+ } else {
+ Ops[i] = Start;
+ }
+ }
+ if (!AddRecs.empty()) {
+ // Add the addrecs onto the end of the list.
+ Ops.insert(Ops.end(), AddRecs.begin(), AddRecs.end());
+ // Resort the operand list, moving any constants to the front.
+ SimplifyAddOperands(Ops, Ty, SE);
+ }
+}
+
+/// expandAddToGEP - Expand an addition expression with a pointer type into
+/// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps
+/// BasicAliasAnalysis and other passes analyze the result. See the rules
+/// for getelementptr vs. inttoptr in
+/// http://llvm.org/docs/LangRef.html#pointeraliasing
+/// for details.
+///
+/// Design note: The correctness of using getelementptr here depends on
+/// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as
+/// they may introduce pointer arithmetic which may not be safely converted
+/// into getelementptr.
+///
+/// Design note: It might seem desirable for this function to be more
+/// loop-aware. If some of the indices are loop-invariant while others
+/// aren't, it might seem desirable to emit multiple GEPs, keeping the
+/// loop-invariant portions of the overall computation outside the loop.
+/// However, there are a few reasons this is not done here. Hoisting simple
+/// arithmetic is a low-level optimization that often isn't very
+/// important until late in the optimization process. In fact, passes
+/// like InstructionCombining will combine GEPs, even if it means
+/// pushing loop-invariant computation down into loops, so even if the
+/// GEPs were split here, the work would quickly be undone. The
+/// LoopStrengthReduction pass, which is usually run quite late (and
+/// after the last InstructionCombining pass), takes care of hoisting
+/// loop-invariant portions of expressions, after considering what
+/// can be folded using target addressing modes.
+///
+Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin,
+ const SCEV *const *op_end,
+ const PointerType *PTy,
+ const Type *Ty,
+ Value *V) {
+ const Type *ElTy = PTy->getElementType();
+ SmallVector<Value *, 4> GepIndices;
+ SmallVector<const SCEV *, 8> Ops(op_begin, op_end);
+ bool AnyNonZeroIndices = false;
+
+ // Split AddRecs up into parts as either of the parts may be usable
+ // without the other.
+ SplitAddRecs(Ops, Ty, SE);
+
+ // Descend down the pointer's type and attempt to convert the other
+ // operands into GEP indices, at each level. The first index in a GEP
+ // indexes into the array implied by the pointer operand; the rest of
+ // the indices index into the element or field type selected by the
+ // preceding index.
+ for (;;) {
+ // If the scale size is not 0, attempt to factor out a scale for
+ // array indexing.
+ SmallVector<const SCEV *, 8> ScaledOps;
+ if (ElTy->isSized()) {
+ const SCEV *ElSize = SE.getSizeOfExpr(ElTy);
+ if (!ElSize->isZero()) {
+ SmallVector<const SCEV *, 8> NewOps;
+ for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
+ const SCEV *Op = Ops[i];
+ const SCEV *Remainder = SE.getIntegerSCEV(0, Ty);
+ if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.TD)) {
+ // Op now has ElSize factored out.
+ ScaledOps.push_back(Op);
+ if (!Remainder->isZero())
+ NewOps.push_back(Remainder);
+ AnyNonZeroIndices = true;
+ } else {
+ // The operand was not divisible, so add it to the list of operands
+ // we'll scan next iteration.
+ NewOps.push_back(Ops[i]);
+ }
+ }
+ // If we made any changes, update Ops.
+ if (!ScaledOps.empty()) {
+ Ops = NewOps;
+ SimplifyAddOperands(Ops, Ty, SE);
+ }
+ }
+ }
+
+ // Record the scaled array index for this level of the type. If
+ // we didn't find any operands that could be factored, tentatively
+ // assume that element zero was selected (since the zero offset
+ // would obviously be folded away).
+ Value *Scaled = ScaledOps.empty() ?
+ Constant::getNullValue(Ty) :
+ expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
+ GepIndices.push_back(Scaled);
+
+ // Collect struct field index operands.
+ while (const StructType *STy = dyn_cast<StructType>(ElTy)) {
+ bool FoundFieldNo = false;
+ // An empty struct has no fields.
+ if (STy->getNumElements() == 0) break;
+ if (SE.TD) {
+ // With TargetData, field offsets are known. See if a constant offset
+ // falls within any of the struct fields.
+ if (Ops.empty()) break;
+ if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
+ if (SE.getTypeSizeInBits(C->getType()) <= 64) {
+ const StructLayout &SL = *SE.TD->getStructLayout(STy);
+ uint64_t FullOffset = C->getValue()->getZExtValue();
+ if (FullOffset < SL.getSizeInBytes()) {
+ unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
+ GepIndices.push_back(
+ ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx));
+ ElTy = STy->getTypeAtIndex(ElIdx);
+ Ops[0] =
+ SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
+ AnyNonZeroIndices = true;
+ FoundFieldNo = true;
+ }
+ }
+ } else {
+ // Without TargetData, just check for an offsetof expression of the
+ // appropriate struct type.
+ for (unsigned i = 0, e = Ops.size(); i != e; ++i)
+ if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Ops[i])) {
+ const Type *CTy;
+ Constant *FieldNo;
+ if (U->isOffsetOf(CTy, FieldNo) && CTy == STy) {
+ GepIndices.push_back(FieldNo);
+ ElTy =
+ STy->getTypeAtIndex(cast<ConstantInt>(FieldNo)->getZExtValue());
+ Ops[i] = SE.getConstant(Ty, 0);
+ AnyNonZeroIndices = true;
+ FoundFieldNo = true;
+ break;
+ }
+ }
+ }
+ // If no struct field offsets were found, tentatively assume that
+ // field zero was selected (since the zero offset would obviously
+ // be folded away).
+ if (!FoundFieldNo) {
+ ElTy = STy->getTypeAtIndex(0u);
+ GepIndices.push_back(
+ Constant::getNullValue(Type::getInt32Ty(Ty->getContext())));
+ }
+ }
+
+ if (const ArrayType *ATy = dyn_cast<ArrayType>(ElTy))
+ ElTy = ATy->getElementType();
+ else
+ break;
+ }
+
+ // If none of the operands were convertable to proper GEP indices, cast
+ // the base to i8* and do an ugly getelementptr with that. It's still
+ // better than ptrtoint+arithmetic+inttoptr at least.
+ if (!AnyNonZeroIndices) {
+ // Cast the base to i8*.
+ V = InsertNoopCastOfTo(V,
+ Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace()));
+
+ // Expand the operands for a plain byte offset.
+ Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
+
+ // Fold a GEP with constant operands.
+ if (Constant *CLHS = dyn_cast<Constant>(V))
+ if (Constant *CRHS = dyn_cast<Constant>(Idx))
+ return ConstantExpr::getGetElementPtr(CLHS, &CRHS, 1);
+
+ // Do a quick scan to see if we have this GEP nearby. If so, reuse it.
+ unsigned ScanLimit = 6;
+ BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
+ // Scanning starts from the last instruction before the insertion point.
+ BasicBlock::iterator IP = Builder.GetInsertPoint();
+ if (IP != BlockBegin) {
+ --IP;
+ for (; ScanLimit; --IP, --ScanLimit) {
+ if (IP->getOpcode() == Instruction::GetElementPtr &&
+ IP->getOperand(0) == V && IP->getOperand(1) == Idx)
+ return IP;
+ if (IP == BlockBegin) break;
+ }
+ }
+
+ // Emit a GEP.
+ Value *GEP = Builder.CreateGEP(V, Idx, "uglygep");
+ rememberInstruction(GEP);
+ return GEP;
+ }
+
+ // Insert a pretty getelementptr. Note that this GEP is not marked inbounds,
+ // because ScalarEvolution may have changed the address arithmetic to
+ // compute a value which is beyond the end of the allocated object.
+ Value *Casted = V;
+ if (V->getType() != PTy)
+ Casted = InsertNoopCastOfTo(Casted, PTy);
+ Value *GEP = Builder.CreateGEP(Casted,
+ GepIndices.begin(),
+ GepIndices.end(),
+ "scevgep");
+ Ops.push_back(SE.getUnknown(GEP));
+ rememberInstruction(GEP);
+ return expand(SE.getAddExpr(Ops));
+}
+
+/// isNonConstantNegative - Return true if the specified scev is negated, but
+/// not a constant.
+static bool isNonConstantNegative(const SCEV *F) {
+ const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(F);
+ if (!Mul) return false;
+
+ // If there is a constant factor, it will be first.
+ const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
+ if (!SC) return false;
+
+ // Return true if the value is negative, this matches things like (-42 * V).
+ return SC->getValue()->getValue().isNegative();
+}
+
+Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
+ int NumOperands = S->getNumOperands();
+ const Type *Ty = SE.getEffectiveSCEVType(S->getType());
+
+ // Find the index of an operand to start with. Choose the operand with
+ // pointer type, if there is one, or the last operand otherwise.
+ int PIdx = 0;
+ for (; PIdx != NumOperands - 1; ++PIdx)
+ if (isa<PointerType>(S->getOperand(PIdx)->getType())) break;
+
+ // Expand code for the operand that we chose.
+ Value *V = expand(S->getOperand(PIdx));
+
+ // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
+ // comments on expandAddToGEP for details.
+ if (const PointerType *PTy = dyn_cast<PointerType>(V->getType())) {
+ // Take the operand at PIdx out of the list.
+ const SmallVectorImpl<const SCEV *> &Ops = S->getOperands();
+ SmallVector<const SCEV *, 8> NewOps;
+ NewOps.insert(NewOps.end(), Ops.begin(), Ops.begin() + PIdx);
+ NewOps.insert(NewOps.end(), Ops.begin() + PIdx + 1, Ops.end());
+ // Make a GEP.
+ return expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, V);
+ }
+
+ // Otherwise, we'll expand the rest of the SCEVAddExpr as plain integer
+ // arithmetic.
+ V = InsertNoopCastOfTo(V, Ty);
+
+ // Emit a bunch of add instructions
+ for (int i = NumOperands-1; i >= 0; --i) {
+ if (i == PIdx) continue;
+ const SCEV *Op = S->getOperand(i);
+ if (isNonConstantNegative(Op)) {
+ Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty);
+ V = InsertBinop(Instruction::Sub, V, W);
+ } else {
+ Value *W = expandCodeFor(Op, Ty);
+ V = InsertBinop(Instruction::Add, V, W);
+ }
+ }
+ return V;
+}
+
+Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
+ const Type *Ty = SE.getEffectiveSCEVType(S->getType());
+ int FirstOp = 0; // Set if we should emit a subtract.
+ if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getOperand(0)))
+ if (SC->getValue()->isAllOnesValue())
+ FirstOp = 1;
+
+ int i = S->getNumOperands()-2;
+ Value *V = expandCodeFor(S->getOperand(i+1), Ty);
+
+ // Emit a bunch of multiply instructions
+ for (; i >= FirstOp; --i) {
+ Value *W = expandCodeFor(S->getOperand(i), Ty);
+ V = InsertBinop(Instruction::Mul, V, W);
+ }
+
+ // -1 * ... ---> 0 - ...
+ if (FirstOp == 1)
+ V = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), V);
+ return V;
+}
+
+Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
+ const Type *Ty = SE.getEffectiveSCEVType(S->getType());
+
+ Value *LHS = expandCodeFor(S->getLHS(), Ty);
+ if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
+ const APInt &RHS = SC->getValue()->getValue();
+ if (RHS.isPowerOf2())
+ return InsertBinop(Instruction::LShr, LHS,
+ ConstantInt::get(Ty, RHS.logBase2()));
+ }
+
+ Value *RHS = expandCodeFor(S->getRHS(), Ty);
+ return InsertBinop(Instruction::UDiv, LHS, RHS);
+}
+
+/// Move parts of Base into Rest to leave Base with the minimal
+/// expression that provides a pointer operand suitable for a
+/// GEP expansion.
+static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest,
+ ScalarEvolution &SE) {
+ while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
+ Base = A->getStart();
+ Rest = SE.getAddExpr(Rest,
+ SE.getAddRecExpr(SE.getIntegerSCEV(0, A->getType()),
+ A->getStepRecurrence(SE),
+ A->getLoop()));
+ }
+ if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
+ Base = A->getOperand(A->getNumOperands()-1);
+ SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end());
+ NewAddOps.back() = Rest;
+ Rest = SE.getAddExpr(NewAddOps);
+ ExposePointerBase(Base, Rest, SE);
+ }
+}
+
+/// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand
+/// the base addrec, which is the addrec without any non-loop-dominating
+/// values, and return the PHI.
+PHINode *
+SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized,
+ const Loop *L,
+ const Type *ExpandTy,
+ const Type *IntTy) {
+ // Reuse a previously-inserted PHI, if present.
+ for (BasicBlock::iterator I = L->getHeader()->begin();
+ PHINode *PN = dyn_cast<PHINode>(I); ++I)
+ if (isInsertedInstruction(PN) && SE.getSCEV(PN) == Normalized)
+ return PN;
+
+ // Save the original insertion point so we can restore it when we're done.
+ BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
+ BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
+
+ // Expand code for the start value.
+ Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy,
+ L->getHeader()->begin());
+
+ // Expand code for the step value. Insert instructions right before the
+ // terminator corresponding to the back-edge. Do this before creating the PHI
+ // so that PHI reuse code doesn't see an incomplete PHI. If the stride is
+ // negative, insert a sub instead of an add for the increment (unless it's a
+ // constant, because subtracts of constants are canonicalized to adds).
+ const SCEV *Step = Normalized->getStepRecurrence(SE);
+ bool isPointer = isa<PointerType>(ExpandTy);
+ bool isNegative = !isPointer && isNonConstantNegative(Step);
+ if (isNegative)
+ Step = SE.getNegativeSCEV(Step);
+ Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
+
+ // Create the PHI.
+ Builder.SetInsertPoint(L->getHeader(), L->getHeader()->begin());
+ PHINode *PN = Builder.CreatePHI(ExpandTy, "lsr.iv");
+ rememberInstruction(PN);
+
+ // Create the step instructions and populate the PHI.
+ BasicBlock *Header = L->getHeader();
+ for (pred_iterator HPI = pred_begin(Header), HPE = pred_end(Header);
+ HPI != HPE; ++HPI) {
+ BasicBlock *Pred = *HPI;
+
+ // Add a start value.
+ if (!L->contains(Pred)) {
+ PN->addIncoming(StartV, Pred);
+ continue;
+ }
+
+ // Create a step value and add it to the PHI. If IVIncInsertLoop is
+ // non-null and equal to the addrec's loop, insert the instructions
+ // at IVIncInsertPos.
+ Instruction *InsertPos = L == IVIncInsertLoop ?
+ IVIncInsertPos : Pred->getTerminator();
+ Builder.SetInsertPoint(InsertPos->getParent(), InsertPos);
+ Value *IncV;
+ // If the PHI is a pointer, use a GEP, otherwise use an add or sub.
+ if (isPointer) {
+ const PointerType *GEPPtrTy = cast<PointerType>(ExpandTy);
+ // If the step isn't constant, don't use an implicitly scaled GEP, because
+ // that would require a multiply inside the loop.
+ if (!isa<ConstantInt>(StepV))
+ GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()),
+ GEPPtrTy->getAddressSpace());
+ const SCEV *const StepArray[1] = { SE.getSCEV(StepV) };
+ IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN);
+ if (IncV->getType() != PN->getType()) {
+ IncV = Builder.CreateBitCast(IncV, PN->getType(), "tmp");
+ rememberInstruction(IncV);
+ }
+ } else {
+ IncV = isNegative ?
+ Builder.CreateSub(PN, StepV, "lsr.iv.next") :
+ Builder.CreateAdd(PN, StepV, "lsr.iv.next");
+ rememberInstruction(IncV);
+ }
+ PN->addIncoming(IncV, Pred);
+ }
+
+ // Restore the original insert point.
+ if (SaveInsertBB)
+ Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
+
+ // Remember this PHI, even in post-inc mode.
+ InsertedValues.insert(PN);
+
+ return PN;
+}
+
+Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) {
+ const Type *STy = S->getType();
+ const Type *IntTy = SE.getEffectiveSCEVType(STy);
+ const Loop *L = S->getLoop();
+
+ // Determine a normalized form of this expression, which is the expression
+ // before any post-inc adjustment is made.
+ const SCEVAddRecExpr *Normalized = S;
+ if (L == PostIncLoop) {
+ const SCEV *Step = S->getStepRecurrence(SE);
+ Normalized = cast<SCEVAddRecExpr>(SE.getMinusSCEV(S, Step));
+ }
+
+ // Strip off any non-loop-dominating component from the addrec start.
+ const SCEV *Start = Normalized->getStart();
+ const SCEV *PostLoopOffset = 0;
+ if (!Start->properlyDominates(L->getHeader(), SE.DT)) {
+ PostLoopOffset = Start;
+ Start = SE.getIntegerSCEV(0, Normalized->getType());
+ Normalized =
+ cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start,
+ Normalized->getStepRecurrence(SE),
+ Normalized->getLoop()));
+ }
+
+ // Strip off any non-loop-dominating component from the addrec step.
+ const SCEV *Step = Normalized->getStepRecurrence(SE);
+ const SCEV *PostLoopScale = 0;
+ if (!Step->hasComputableLoopEvolution(L) &&
+ !Step->dominates(L->getHeader(), SE.DT)) {
+ PostLoopScale = Step;
+ Step = SE.getIntegerSCEV(1, Normalized->getType());
+ Normalized =
+ cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start, Step,
+ Normalized->getLoop()));
+ }
+
+ // Expand the core addrec. If we need post-loop scaling, force it to
+ // expand to an integer type to avoid the need for additional casting.
+ const Type *ExpandTy = PostLoopScale ? IntTy : STy;
+ PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy);
+
+ // Accomodate post-inc mode, if necessary.
+ Value *Result;
+ if (L != PostIncLoop)
+ Result = PN;
+ else {
+ // In PostInc mode, use the post-incremented value.
+ BasicBlock *LatchBlock = L->getLoopLatch();
+ assert(LatchBlock && "PostInc mode requires a unique loop latch!");
+ Result = PN->getIncomingValueForBlock(LatchBlock);
+ }
+
+ // Re-apply any non-loop-dominating scale.
+ if (PostLoopScale) {
+ Result = Builder.CreateMul(Result,
+ expandCodeFor(PostLoopScale, IntTy));
+ rememberInstruction(Result);
+ }
+
+ // Re-apply any non-loop-dominating offset.
+ if (PostLoopOffset) {
+ if (const PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) {
+ const SCEV *const OffsetArray[1] = { PostLoopOffset };
+ Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result);
+ } else {
+ Result = Builder.CreateAdd(Result,
+ expandCodeFor(PostLoopOffset, IntTy));
+ rememberInstruction(Result);
+ }
+ }
+
+ return Result;
+}
+
+Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
+ if (!CanonicalMode) return expandAddRecExprLiterally(S);
+
+ const Type *Ty = SE.getEffectiveSCEVType(S->getType());
+ const Loop *L = S->getLoop();
+
+ // First check for an existing canonical IV in a suitable type.
+ PHINode *CanonicalIV = 0;
+ if (PHINode *PN = L->getCanonicalInductionVariable())
+ if (SE.isSCEVable(PN->getType()) &&
+ isa<IntegerType>(SE.getEffectiveSCEVType(PN->getType())) &&
+ SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
+ CanonicalIV = PN;
+
+ // Rewrite an AddRec in terms of the canonical induction variable, if
+ // its type is more narrow.
+ if (CanonicalIV &&
+ SE.getTypeSizeInBits(CanonicalIV->getType()) >
+ SE.getTypeSizeInBits(Ty)) {
+ const SmallVectorImpl<const SCEV *> &Ops = S->getOperands();
+ SmallVector<const SCEV *, 4> NewOps(Ops.size());
+ for (unsigned i = 0, e = Ops.size(); i != e; ++i)
+ NewOps[i] = SE.getAnyExtendExpr(Ops[i], CanonicalIV->getType());
+ Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop()));
+ BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
+ BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
+ BasicBlock::iterator NewInsertPt =
+ llvm::next(BasicBlock::iterator(cast<Instruction>(V)));
+ while (isa<PHINode>(NewInsertPt)) ++NewInsertPt;
+ V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0,
+ NewInsertPt);
+ Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
+ return V;
+ }
+
+ // {X,+,F} --> X + {0,+,F}
+ if (!S->getStart()->isZero()) {
+ const SmallVectorImpl<const SCEV *> &SOperands = S->getOperands();
+ SmallVector<const SCEV *, 4> NewOps(SOperands.begin(), SOperands.end());
+ NewOps[0] = SE.getIntegerSCEV(0, Ty);
+ const SCEV *Rest = SE.getAddRecExpr(NewOps, L);
+
+ // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
+ // comments on expandAddToGEP for details.
+ const SCEV *Base = S->getStart();
+ const SCEV *RestArray[1] = { Rest };
+ // Dig into the expression to find the pointer base for a GEP.
+ ExposePointerBase(Base, RestArray[0], SE);
+ // If we found a pointer, expand the AddRec with a GEP.
+ if (const PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
+ // Make sure the Base isn't something exotic, such as a multiplied
+ // or divided pointer value. In those cases, the result type isn't
+ // actually a pointer type.
+ if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
+ Value *StartV = expand(Base);
+ assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
+ return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
+ }
+ }
+
+ // Just do a normal add. Pre-expand the operands to suppress folding.
+ return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())),
+ SE.getUnknown(expand(Rest))));
+ }
+
+ // {0,+,1} --> Insert a canonical induction variable into the loop!
+ if (S->isAffine() &&
+ S->getOperand(1) == SE.getIntegerSCEV(1, Ty)) {
+ // If there's a canonical IV, just use it.
+ if (CanonicalIV) {
+ assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
+ "IVs with types different from the canonical IV should "
+ "already have been handled!");
+ return CanonicalIV;
+ }
+
+ // Create and insert the PHI node for the induction variable in the
+ // specified loop.
+ BasicBlock *Header = L->getHeader();
+ PHINode *PN = PHINode::Create(Ty, "indvar", Header->begin());
+ rememberInstruction(PN);
+
+ Constant *One = ConstantInt::get(Ty, 1);
+ for (pred_iterator HPI = pred_begin(Header), HPE = pred_end(Header);
+ HPI != HPE; ++HPI)
+ if (L->contains(*HPI)) {
+ // Insert a unit add instruction right before the terminator
+ // corresponding to the back-edge.
+ Instruction *Add = BinaryOperator::CreateAdd(PN, One, "indvar.next",
+ (*HPI)->getTerminator());
+ rememberInstruction(Add);
+ PN->addIncoming(Add, *HPI);
+ } else {
+ PN->addIncoming(Constant::getNullValue(Ty), *HPI);
+ }
+ }
+
+ // {0,+,F} --> {0,+,1} * F
+ // Get the canonical induction variable I for this loop.
+ Value *I = CanonicalIV ?
+ CanonicalIV :
+ getOrInsertCanonicalInductionVariable(L, Ty);
+
+ // If this is a simple linear addrec, emit it now as a special case.
+ if (S->isAffine()) // {0,+,F} --> i*F
+ return
+ expand(SE.getTruncateOrNoop(
+ SE.getMulExpr(SE.getUnknown(I),
+ SE.getNoopOrAnyExtend(S->getOperand(1),
+ I->getType())),
+ Ty));
+
+ // If this is a chain of recurrences, turn it into a closed form, using the
+ // folders, then expandCodeFor the closed form. This allows the folders to
+ // simplify the expression without having to build a bunch of special code
+ // into this folder.
+ const SCEV *IH = SE.getUnknown(I); // Get I as a "symbolic" SCEV.
+
+ // Promote S up to the canonical IV type, if the cast is foldable.
+ const SCEV *NewS = S;
+ const SCEV *Ext = SE.getNoopOrAnyExtend(S, I->getType());
+ if (isa<SCEVAddRecExpr>(Ext))
+ NewS = Ext;
+
+ const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
+ //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
+
+ // Truncate the result down to the original type, if needed.
+ const SCEV *T = SE.getTruncateOrNoop(V, Ty);
+ return expand(T);
+}
+
+Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
+ const Type *Ty = SE.getEffectiveSCEVType(S->getType());
+ Value *V = expandCodeFor(S->getOperand(),
+ SE.getEffectiveSCEVType(S->getOperand()->getType()));
+ Value *I = Builder.CreateTrunc(V, Ty, "tmp");
+ rememberInstruction(I);
+ return I;
+}
+
+Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
+ const Type *Ty = SE.getEffectiveSCEVType(S->getType());
+ Value *V = expandCodeFor(S->getOperand(),
+ SE.getEffectiveSCEVType(S->getOperand()->getType()));
+ Value *I = Builder.CreateZExt(V, Ty, "tmp");
+ rememberInstruction(I);
+ return I;
+}
+
+Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
+ const Type *Ty = SE.getEffectiveSCEVType(S->getType());
+ Value *V = expandCodeFor(S->getOperand(),
+ SE.getEffectiveSCEVType(S->getOperand()->getType()));
+ Value *I = Builder.CreateSExt(V, Ty, "tmp");
+ rememberInstruction(I);
+ return I;
+}
+
+Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
+ Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
+ const Type *Ty = LHS->getType();
+ for (int i = S->getNumOperands()-2; i >= 0; --i) {
+ // In the case of mixed integer and pointer types, do the
+ // rest of the comparisons as integer.
+ if (S->getOperand(i)->getType() != Ty) {
+ Ty = SE.getEffectiveSCEVType(Ty);
+ LHS = InsertNoopCastOfTo(LHS, Ty);
+ }
+ Value *RHS = expandCodeFor(S->getOperand(i), Ty);
+ Value *ICmp = Builder.CreateICmpSGT(LHS, RHS, "tmp");
+ rememberInstruction(ICmp);
+ Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
+ rememberInstruction(Sel);
+ LHS = Sel;
+ }
+ // In the case of mixed integer and pointer types, cast the
+ // final result back to the pointer type.
+ if (LHS->getType() != S->getType())
+ LHS = InsertNoopCastOfTo(LHS, S->getType());
+ return LHS;
+}
+
+Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
+ Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
+ const Type *Ty = LHS->getType();
+ for (int i = S->getNumOperands()-2; i >= 0; --i) {
+ // In the case of mixed integer and pointer types, do the
+ // rest of the comparisons as integer.
+ if (S->getOperand(i)->getType() != Ty) {
+ Ty = SE.getEffectiveSCEVType(Ty);
+ LHS = InsertNoopCastOfTo(LHS, Ty);
+ }
+ Value *RHS = expandCodeFor(S->getOperand(i), Ty);
+ Value *ICmp = Builder.CreateICmpUGT(LHS, RHS, "tmp");
+ rememberInstruction(ICmp);
+ Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
+ rememberInstruction(Sel);
+ LHS = Sel;
+ }
+ // In the case of mixed integer and pointer types, cast the
+ // final result back to the pointer type.
+ if (LHS->getType() != S->getType())
+ LHS = InsertNoopCastOfTo(LHS, S->getType());
+ return LHS;
+}
+
+Value *SCEVExpander::expandCodeFor(const SCEV *SH, const Type *Ty) {
+ // Expand the code for this SCEV.
+ Value *V = expand(SH);
+ if (Ty) {
+ assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
+ "non-trivial casts should be done with the SCEVs directly!");
+ V = InsertNoopCastOfTo(V, Ty);
+ }
+ return V;
+}
+
+Value *SCEVExpander::expand(const SCEV *S) {
+ // Compute an insertion point for this SCEV object. Hoist the instructions
+ // as far out in the loop nest as possible.
+ Instruction *InsertPt = Builder.GetInsertPoint();
+ for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ;
+ L = L->getParentLoop())
+ if (S->isLoopInvariant(L)) {
+ if (!L) break;
+ if (BasicBlock *Preheader = L->getLoopPreheader())
+ InsertPt = Preheader->getTerminator();
+ } else {
+ // If the SCEV is computable at this level, insert it into the header
+ // after the PHIs (and after any other instructions that we've inserted
+ // there) so that it is guaranteed to dominate any user inside the loop.
+ if (L && S->hasComputableLoopEvolution(L))
+ InsertPt = L->getHeader()->getFirstNonPHI();
+ while (isInsertedInstruction(InsertPt))
+ InsertPt = llvm::next(BasicBlock::iterator(InsertPt));
+ break;
+ }
+
+ // Check to see if we already expanded this here.
+ std::map<std::pair<const SCEV *, Instruction *>,
+ AssertingVH<Value> >::iterator I =
+ InsertedExpressions.find(std::make_pair(S, InsertPt));
+ if (I != InsertedExpressions.end())
+ return I->second;
+
+ BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
+ BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
+ Builder.SetInsertPoint(InsertPt->getParent(), InsertPt);
+
+ // Expand the expression into instructions.
+ Value *V = visit(S);
+
+ // Remember the expanded value for this SCEV at this location.
+ if (!PostIncLoop)
+ InsertedExpressions[std::make_pair(S, InsertPt)] = V;
+
+ Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
+ return V;
+}
+
+/// getOrInsertCanonicalInductionVariable - This method returns the
+/// canonical induction variable of the specified type for the specified
+/// loop (inserting one if there is none). A canonical induction variable
+/// starts at zero and steps by one on each iteration.
+Value *
+SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
+ const Type *Ty) {
+ assert(Ty->isInteger() && "Can only insert integer induction variables!");
+ const SCEV *H = SE.getAddRecExpr(SE.getIntegerSCEV(0, Ty),
+ SE.getIntegerSCEV(1, Ty), L);
+ BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
+ BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
+ Value *V = expandCodeFor(H, 0, L->getHeader()->begin());
+ if (SaveInsertBB)
+ Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
+ return V;
+}
diff --git a/lib/Analysis/SparsePropagation.cpp b/lib/Analysis/SparsePropagation.cpp
new file mode 100644
index 0000000..d8c207b
--- /dev/null
+++ b/lib/Analysis/SparsePropagation.cpp
@@ -0,0 +1,347 @@
+//===- SparsePropagation.cpp - Sparse Conditional Property Propagation ----===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements an abstract sparse conditional propagation algorithm,
+// modeled after SCCP, but with a customizable lattice function.
+//
+//===----------------------------------------------------------------------===//
+
+#define DEBUG_TYPE "sparseprop"
+#include "llvm/Analysis/SparsePropagation.h"
+#include "llvm/Constants.h"
+#include "llvm/Function.h"
+#include "llvm/Instructions.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+using namespace llvm;
+
+//===----------------------------------------------------------------------===//
+// AbstractLatticeFunction Implementation
+//===----------------------------------------------------------------------===//
+
+AbstractLatticeFunction::~AbstractLatticeFunction() {}
+
+/// PrintValue - Render the specified lattice value to the specified stream.
+void AbstractLatticeFunction::PrintValue(LatticeVal V, raw_ostream &OS) {
+ if (V == UndefVal)
+ OS << "undefined";
+ else if (V == OverdefinedVal)
+ OS << "overdefined";
+ else if (V == UntrackedVal)
+ OS << "untracked";
+ else
+ OS << "unknown lattice value";
+}
+
+//===----------------------------------------------------------------------===//
+// SparseSolver Implementation
+//===----------------------------------------------------------------------===//
+
+/// getOrInitValueState - Return the LatticeVal object that corresponds to the
+/// value, initializing the value's state if it hasn't been entered into the
+/// map yet. This function is necessary because not all values should start
+/// out in the underdefined state... Arguments should be overdefined, and
+/// constants should be marked as constants.
+///
+SparseSolver::LatticeVal SparseSolver::getOrInitValueState(Value *V) {
+ DenseMap<Value*, LatticeVal>::iterator I = ValueState.find(V);
+ if (I != ValueState.end()) return I->second; // Common case, in the map
+
+ LatticeVal LV;
+ if (LatticeFunc->IsUntrackedValue(V))
+ return LatticeFunc->getUntrackedVal();
+ else if (Constant *C = dyn_cast<Constant>(V))
+ LV = LatticeFunc->ComputeConstant(C);
+ else if (Argument *A = dyn_cast<Argument>(V))
+ LV = LatticeFunc->ComputeArgument(A);
+ else if (!isa<Instruction>(V))
+ // All other non-instructions are overdefined.
+ LV = LatticeFunc->getOverdefinedVal();
+ else
+ // All instructions are underdefined by default.
+ LV = LatticeFunc->getUndefVal();
+
+ // If this value is untracked, don't add it to the map.
+ if (LV == LatticeFunc->getUntrackedVal())
+ return LV;
+ return ValueState[V] = LV;
+}
+
+/// UpdateState - When the state for some instruction is potentially updated,
+/// this function notices and adds I to the worklist if needed.
+void SparseSolver::UpdateState(Instruction &Inst, LatticeVal V) {
+ DenseMap<Value*, LatticeVal>::iterator I = ValueState.find(&Inst);
+ if (I != ValueState.end() && I->second == V)
+ return; // No change.
+
+ // An update. Visit uses of I.
+ ValueState[&Inst] = V;
+ InstWorkList.push_back(&Inst);
+}
+
+/// MarkBlockExecutable - This method can be used by clients to mark all of
+/// the blocks that are known to be intrinsically live in the processed unit.
+void SparseSolver::MarkBlockExecutable(BasicBlock *BB) {
+ DEBUG(dbgs() << "Marking Block Executable: " << BB->getName() << "\n");
+ BBExecutable.insert(BB); // Basic block is executable!
+ BBWorkList.push_back(BB); // Add the block to the work list!
+}
+
+/// markEdgeExecutable - Mark a basic block as executable, adding it to the BB
+/// work list if it is not already executable...
+void SparseSolver::markEdgeExecutable(BasicBlock *Source, BasicBlock *Dest) {
+ if (!KnownFeasibleEdges.insert(Edge(Source, Dest)).second)
+ return; // This edge is already known to be executable!
+
+ DEBUG(dbgs() << "Marking Edge Executable: " << Source->getName()
+ << " -> " << Dest->getName() << "\n");
+
+ if (BBExecutable.count(Dest)) {
+ // The destination is already executable, but we just made an edge
+ // feasible that wasn't before. Revisit the PHI nodes in the block
+ // because they have potentially new operands.
+ for (BasicBlock::iterator I = Dest->begin(); isa<PHINode>(I); ++I)
+ visitPHINode(*cast<PHINode>(I));
+
+ } else {
+ MarkBlockExecutable(Dest);
+ }
+}
+
+
+/// getFeasibleSuccessors - Return a vector of booleans to indicate which
+/// successors are reachable from a given terminator instruction.
+void SparseSolver::getFeasibleSuccessors(TerminatorInst &TI,
+ SmallVectorImpl<bool> &Succs,
+ bool AggressiveUndef) {
+ Succs.resize(TI.getNumSuccessors());
+ if (TI.getNumSuccessors() == 0) return;
+
+ if (BranchInst *BI = dyn_cast<BranchInst>(&TI)) {
+ if (BI->isUnconditional()) {
+ Succs[0] = true;
+ return;
+ }
+
+ LatticeVal BCValue;
+ if (AggressiveUndef)
+ BCValue = getOrInitValueState(BI->getCondition());
+ else
+ BCValue = getLatticeState(BI->getCondition());
+
+ if (BCValue == LatticeFunc->getOverdefinedVal() ||
+ BCValue == LatticeFunc->getUntrackedVal()) {
+ // Overdefined condition variables can branch either way.
+ Succs[0] = Succs[1] = true;
+ return;
+ }
+
+ // If undefined, neither is feasible yet.
+ if (BCValue == LatticeFunc->getUndefVal())
+ return;
+
+ Constant *C = LatticeFunc->GetConstant(BCValue, BI->getCondition(), *this);
+ if (C == 0 || !isa<ConstantInt>(C)) {
+ // Non-constant values can go either way.
+ Succs[0] = Succs[1] = true;
+ return;
+ }
+
+ // Constant condition variables mean the branch can only go a single way
+ Succs[C->isNullValue()] = true;
+ return;
+ }
+
+ if (isa<InvokeInst>(TI)) {
+ // Invoke instructions successors are always executable.
+ // TODO: Could ask the lattice function if the value can throw.
+ Succs[0] = Succs[1] = true;
+ return;
+ }
+
+ if (isa<IndirectBrInst>(TI)) {
+ Succs.assign(Succs.size(), true);
+ return;
+ }
+
+ SwitchInst &SI = cast<SwitchInst>(TI);
+ LatticeVal SCValue;
+ if (AggressiveUndef)
+ SCValue = getOrInitValueState(SI.getCondition());
+ else
+ SCValue = getLatticeState(SI.getCondition());
+
+ if (SCValue == LatticeFunc->getOverdefinedVal() ||
+ SCValue == LatticeFunc->getUntrackedVal()) {
+ // All destinations are executable!
+ Succs.assign(TI.getNumSuccessors(), true);
+ return;
+ }
+
+ // If undefined, neither is feasible yet.
+ if (SCValue == LatticeFunc->getUndefVal())
+ return;
+
+ Constant *C = LatticeFunc->GetConstant(SCValue, SI.getCondition(), *this);
+ if (C == 0 || !isa<ConstantInt>(C)) {
+ // All destinations are executable!
+ Succs.assign(TI.getNumSuccessors(), true);
+ return;
+ }
+
+ Succs[SI.findCaseValue(cast<ConstantInt>(C))] = true;
+}
+
+
+/// isEdgeFeasible - Return true if the control flow edge from the 'From'
+/// basic block to the 'To' basic block is currently feasible...
+bool SparseSolver::isEdgeFeasible(BasicBlock *From, BasicBlock *To,
+ bool AggressiveUndef) {
+ SmallVector<bool, 16> SuccFeasible;
+ TerminatorInst *TI = From->getTerminator();
+ getFeasibleSuccessors(*TI, SuccFeasible, AggressiveUndef);
+
+ for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
+ if (TI->getSuccessor(i) == To && SuccFeasible[i])
+ return true;
+
+ return false;
+}
+
+void SparseSolver::visitTerminatorInst(TerminatorInst &TI) {
+ SmallVector<bool, 16> SuccFeasible;
+ getFeasibleSuccessors(TI, SuccFeasible, true);
+
+ BasicBlock *BB = TI.getParent();
+
+ // Mark all feasible successors executable...
+ for (unsigned i = 0, e = SuccFeasible.size(); i != e; ++i)
+ if (SuccFeasible[i])
+ markEdgeExecutable(BB, TI.getSuccessor(i));
+}
+
+void SparseSolver::visitPHINode(PHINode &PN) {
+ // The lattice function may store more information on a PHINode than could be
+ // computed from its incoming values. For example, SSI form stores its sigma
+ // functions as PHINodes with a single incoming value.
+ if (LatticeFunc->IsSpecialCasedPHI(&PN)) {
+ LatticeVal IV = LatticeFunc->ComputeInstructionState(PN, *this);
+ if (IV != LatticeFunc->getUntrackedVal())
+ UpdateState(PN, IV);
+ return;
+ }
+
+ LatticeVal PNIV = getOrInitValueState(&PN);
+ LatticeVal Overdefined = LatticeFunc->getOverdefinedVal();
+
+ // If this value is already overdefined (common) just return.
+ if (PNIV == Overdefined || PNIV == LatticeFunc->getUntrackedVal())
+ return; // Quick exit
+
+ // Super-extra-high-degree PHI nodes are unlikely to ever be interesting,
+ // and slow us down a lot. Just mark them overdefined.
+ if (PN.getNumIncomingValues() > 64) {
+ UpdateState(PN, Overdefined);
+ return;
+ }
+
+ // Look at all of the executable operands of the PHI node. If any of them
+ // are overdefined, the PHI becomes overdefined as well. Otherwise, ask the
+ // transfer function to give us the merge of the incoming values.
+ for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
+ // If the edge is not yet known to be feasible, it doesn't impact the PHI.
+ if (!isEdgeFeasible(PN.getIncomingBlock(i), PN.getParent(), true))
+ continue;
+
+ // Merge in this value.
+ LatticeVal OpVal = getOrInitValueState(PN.getIncomingValue(i));
+ if (OpVal != PNIV)
+ PNIV = LatticeFunc->MergeValues(PNIV, OpVal);
+
+ if (PNIV == Overdefined)
+ break; // Rest of input values don't matter.
+ }
+
+ // Update the PHI with the compute value, which is the merge of the inputs.
+ UpdateState(PN, PNIV);
+}
+
+
+void SparseSolver::visitInst(Instruction &I) {
+ // PHIs are handled by the propagation logic, they are never passed into the
+ // transfer functions.
+ if (PHINode *PN = dyn_cast<PHINode>(&I))
+ return visitPHINode(*PN);
+
+ // Otherwise, ask the transfer function what the result is. If this is
+ // something that we care about, remember it.
+ LatticeVal IV = LatticeFunc->ComputeInstructionState(I, *this);
+ if (IV != LatticeFunc->getUntrackedVal())
+ UpdateState(I, IV);
+
+ if (TerminatorInst *TI = dyn_cast<TerminatorInst>(&I))
+ visitTerminatorInst(*TI);
+}
+
+void SparseSolver::Solve(Function &F) {
+ MarkBlockExecutable(&F.getEntryBlock());
+
+ // Process the work lists until they are empty!
+ while (!BBWorkList.empty() || !InstWorkList.empty()) {
+ // Process the instruction work list.
+ while (!InstWorkList.empty()) {
+ Instruction *I = InstWorkList.back();
+ InstWorkList.pop_back();
+
+ DEBUG(dbgs() << "\nPopped off I-WL: " << *I << "\n");
+
+ // "I" got into the work list because it made a transition. See if any
+ // users are both live and in need of updating.
+ for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
+ UI != E; ++UI) {
+ Instruction *U = cast<Instruction>(*UI);
+ if (BBExecutable.count(U->getParent())) // Inst is executable?
+ visitInst(*U);
+ }
+ }
+
+ // Process the basic block work list.
+ while (!BBWorkList.empty()) {
+ BasicBlock *BB = BBWorkList.back();
+ BBWorkList.pop_back();
+
+ DEBUG(dbgs() << "\nPopped off BBWL: " << *BB);
+
+ // Notify all instructions in this basic block that they are newly
+ // executable.
+ for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
+ visitInst(*I);
+ }
+ }
+}
+
+void SparseSolver::Print(Function &F, raw_ostream &OS) const {
+ OS << "\nFUNCTION: " << F.getNameStr() << "\n";
+ for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
+ if (!BBExecutable.count(BB))
+ OS << "INFEASIBLE: ";
+ OS << "\t";
+ if (BB->hasName())
+ OS << BB->getNameStr() << ":\n";
+ else
+ OS << "; anon bb\n";
+ for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
+ LatticeFunc->PrintValue(getLatticeState(I), OS);
+ OS << *I << "\n";
+ }
+
+ OS << "\n";
+ }
+}
+
diff --git a/lib/Analysis/Trace.cpp b/lib/Analysis/Trace.cpp
new file mode 100644
index 0000000..68a39cd
--- /dev/null
+++ b/lib/Analysis/Trace.cpp
@@ -0,0 +1,51 @@
+//===- Trace.cpp - Implementation of Trace class --------------------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This class represents a single trace of LLVM basic blocks. A trace is a
+// single entry, multiple exit, region of code that is often hot. Trace-based
+// optimizations treat traces almost like they are a large, strange, basic
+// block: because the trace path is assumed to be hot, optimizations for the
+// fall-through path are made at the expense of the non-fall-through paths.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/Trace.h"
+#include "llvm/Function.h"
+#include "llvm/Assembly/Writer.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+using namespace llvm;
+
+Function *Trace::getFunction() const {
+ return getEntryBasicBlock()->getParent();
+}
+
+Module *Trace::getModule() const {
+ return getFunction()->getParent();
+}
+
+/// print - Write trace to output stream.
+///
+void Trace::print(raw_ostream &O) const {
+ Function *F = getFunction();
+ O << "; Trace from function " << F->getNameStr() << ", blocks:\n";
+ for (const_iterator i = begin(), e = end(); i != e; ++i) {
+ O << "; ";
+ WriteAsOperand(O, *i, true, getModule());
+ O << "\n";
+ }
+ O << "; Trace parent function: \n" << *F;
+}
+
+/// dump - Debugger convenience method; writes trace to standard error
+/// output stream.
+///
+void Trace::dump() const {
+ print(dbgs());
+}
diff --git a/lib/Analysis/ValueTracking.cpp b/lib/Analysis/ValueTracking.cpp
new file mode 100644
index 0000000..f9331e7
--- /dev/null
+++ b/lib/Analysis/ValueTracking.cpp
@@ -0,0 +1,1438 @@
+//===- ValueTracking.cpp - Walk computations to compute properties --------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file contains routines that help analyze properties that chains of
+// computations have.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/ValueTracking.h"
+#include "llvm/Constants.h"
+#include "llvm/Instructions.h"
+#include "llvm/GlobalVariable.h"
+#include "llvm/GlobalAlias.h"
+#include "llvm/IntrinsicInst.h"
+#include "llvm/LLVMContext.h"
+#include "llvm/Operator.h"
+#include "llvm/Target/TargetData.h"
+#include "llvm/Support/GetElementPtrTypeIterator.h"
+#include "llvm/Support/MathExtras.h"
+#include <cstring>
+using namespace llvm;
+
+/// ComputeMaskedBits - Determine which of the bits specified in Mask are
+/// known to be either zero or one and return them in the KnownZero/KnownOne
+/// bit sets. This code only analyzes bits in Mask, in order to short-circuit
+/// processing.
+/// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
+/// we cannot optimize based on the assumption that it is zero without changing
+/// it to be an explicit zero. If we don't change it to zero, other code could
+/// optimized based on the contradictory assumption that it is non-zero.
+/// Because instcombine aggressively folds operations with undef args anyway,
+/// this won't lose us code quality.
+///
+/// This function is defined on values with integer type, values with pointer
+/// type (but only if TD is non-null), and vectors of integers. In the case
+/// where V is a vector, the mask, known zero, and known one values are the
+/// same width as the vector element, and the bit is set only if it is true
+/// for all of the elements in the vector.
+void llvm::ComputeMaskedBits(Value *V, const APInt &Mask,
+ APInt &KnownZero, APInt &KnownOne,
+ const TargetData *TD, unsigned Depth) {
+ const unsigned MaxDepth = 6;
+ assert(V && "No Value?");
+ assert(Depth <= MaxDepth && "Limit Search Depth");
+ unsigned BitWidth = Mask.getBitWidth();
+ assert((V->getType()->isIntOrIntVector() || isa<PointerType>(V->getType())) &&
+ "Not integer or pointer type!");
+ assert((!TD ||
+ TD->getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) &&
+ (!V->getType()->isIntOrIntVector() ||
+ V->getType()->getScalarSizeInBits() == BitWidth) &&
+ KnownZero.getBitWidth() == BitWidth &&
+ KnownOne.getBitWidth() == BitWidth &&
+ "V, Mask, KnownOne and KnownZero should have same BitWidth");
+
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
+ // We know all of the bits for a constant!
+ KnownOne = CI->getValue() & Mask;
+ KnownZero = ~KnownOne & Mask;
+ return;
+ }
+ // Null and aggregate-zero are all-zeros.
+ if (isa<ConstantPointerNull>(V) ||
+ isa<ConstantAggregateZero>(V)) {
+ KnownOne.clear();
+ KnownZero = Mask;
+ return;
+ }
+ // Handle a constant vector by taking the intersection of the known bits of
+ // each element.
+ if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) {
+ KnownZero.set(); KnownOne.set();
+ for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
+ APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
+ ComputeMaskedBits(CV->getOperand(i), Mask, KnownZero2, KnownOne2,
+ TD, Depth);
+ KnownZero &= KnownZero2;
+ KnownOne &= KnownOne2;
+ }
+ return;
+ }
+ // The address of an aligned GlobalValue has trailing zeros.
+ if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
+ unsigned Align = GV->getAlignment();
+ if (Align == 0 && TD && GV->getType()->getElementType()->isSized()) {
+ const Type *ObjectType = GV->getType()->getElementType();
+ // If the object is defined in the current Module, we'll be giving
+ // it the preferred alignment. Otherwise, we have to assume that it
+ // may only have the minimum ABI alignment.
+ if (!GV->isDeclaration() && !GV->mayBeOverridden())
+ Align = TD->getPrefTypeAlignment(ObjectType);
+ else
+ Align = TD->getABITypeAlignment(ObjectType);
+ }
+ if (Align > 0)
+ KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
+ CountTrailingZeros_32(Align));
+ else
+ KnownZero.clear();
+ KnownOne.clear();
+ return;
+ }
+ // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
+ // the bits of its aliasee.
+ if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
+ if (GA->mayBeOverridden()) {
+ KnownZero.clear(); KnownOne.clear();
+ } else {
+ ComputeMaskedBits(GA->getAliasee(), Mask, KnownZero, KnownOne,
+ TD, Depth+1);
+ }
+ return;
+ }
+
+ KnownZero.clear(); KnownOne.clear(); // Start out not knowing anything.
+
+ if (Depth == MaxDepth || Mask == 0)
+ return; // Limit search depth.
+
+ Operator *I = dyn_cast<Operator>(V);
+ if (!I) return;
+
+ APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
+ switch (I->getOpcode()) {
+ default: break;
+ case Instruction::And: {
+ // If either the LHS or the RHS are Zero, the result is zero.
+ ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
+ APInt Mask2(Mask & ~KnownZero);
+ ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
+ Depth+1);
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
+
+ // Output known-1 bits are only known if set in both the LHS & RHS.
+ KnownOne &= KnownOne2;
+ // Output known-0 are known to be clear if zero in either the LHS | RHS.
+ KnownZero |= KnownZero2;
+ return;
+ }
+ case Instruction::Or: {
+ ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
+ APInt Mask2(Mask & ~KnownOne);
+ ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
+ Depth+1);
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
+
+ // Output known-0 bits are only known if clear in both the LHS & RHS.
+ KnownZero &= KnownZero2;
+ // Output known-1 are known to be set if set in either the LHS | RHS.
+ KnownOne |= KnownOne2;
+ return;
+ }
+ case Instruction::Xor: {
+ ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, TD, Depth+1);
+ ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, TD,
+ Depth+1);
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
+
+ // Output known-0 bits are known if clear or set in both the LHS & RHS.
+ APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
+ // Output known-1 are known to be set if set in only one of the LHS, RHS.
+ KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
+ KnownZero = KnownZeroOut;
+ return;
+ }
+ case Instruction::Mul: {
+ APInt Mask2 = APInt::getAllOnesValue(BitWidth);
+ ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero, KnownOne, TD,Depth+1);
+ ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
+ Depth+1);
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
+
+ // If low bits are zero in either operand, output low known-0 bits.
+ // Also compute a conserative estimate for high known-0 bits.
+ // More trickiness is possible, but this is sufficient for the
+ // interesting case of alignment computation.
+ KnownOne.clear();
+ unsigned TrailZ = KnownZero.countTrailingOnes() +
+ KnownZero2.countTrailingOnes();
+ unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
+ KnownZero2.countLeadingOnes(),
+ BitWidth) - BitWidth;
+
+ TrailZ = std::min(TrailZ, BitWidth);
+ LeadZ = std::min(LeadZ, BitWidth);
+ KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
+ APInt::getHighBitsSet(BitWidth, LeadZ);
+ KnownZero &= Mask;
+ return;
+ }
+ case Instruction::UDiv: {
+ // For the purposes of computing leading zeros we can conservatively
+ // treat a udiv as a logical right shift by the power of 2 known to
+ // be less than the denominator.
+ APInt AllOnes = APInt::getAllOnesValue(BitWidth);
+ ComputeMaskedBits(I->getOperand(0),
+ AllOnes, KnownZero2, KnownOne2, TD, Depth+1);
+ unsigned LeadZ = KnownZero2.countLeadingOnes();
+
+ KnownOne2.clear();
+ KnownZero2.clear();
+ ComputeMaskedBits(I->getOperand(1),
+ AllOnes, KnownZero2, KnownOne2, TD, Depth+1);
+ unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
+ if (RHSUnknownLeadingOnes != BitWidth)
+ LeadZ = std::min(BitWidth,
+ LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
+
+ KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ) & Mask;
+ return;
+ }
+ case Instruction::Select:
+ ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, TD, Depth+1);
+ ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, TD,
+ Depth+1);
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
+
+ // Only known if known in both the LHS and RHS.
+ KnownOne &= KnownOne2;
+ KnownZero &= KnownZero2;
+ return;
+ case Instruction::FPTrunc:
+ case Instruction::FPExt:
+ case Instruction::FPToUI:
+ case Instruction::FPToSI:
+ case Instruction::SIToFP:
+ case Instruction::UIToFP:
+ return; // Can't work with floating point.
+ case Instruction::PtrToInt:
+ case Instruction::IntToPtr:
+ // We can't handle these if we don't know the pointer size.
+ if (!TD) return;
+ // FALL THROUGH and handle them the same as zext/trunc.
+ case Instruction::ZExt:
+ case Instruction::Trunc: {
+ const Type *SrcTy = I->getOperand(0)->getType();
+
+ unsigned SrcBitWidth;
+ // Note that we handle pointer operands here because of inttoptr/ptrtoint
+ // which fall through here.
+ if (isa<PointerType>(SrcTy))
+ SrcBitWidth = TD->getTypeSizeInBits(SrcTy);
+ else
+ SrcBitWidth = SrcTy->getScalarSizeInBits();
+
+ APInt MaskIn(Mask);
+ MaskIn.zextOrTrunc(SrcBitWidth);
+ KnownZero.zextOrTrunc(SrcBitWidth);
+ KnownOne.zextOrTrunc(SrcBitWidth);
+ ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, TD,
+ Depth+1);
+ KnownZero.zextOrTrunc(BitWidth);
+ KnownOne.zextOrTrunc(BitWidth);
+ // Any top bits are known to be zero.
+ if (BitWidth > SrcBitWidth)
+ KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
+ return;
+ }
+ case Instruction::BitCast: {
+ const Type *SrcTy = I->getOperand(0)->getType();
+ if ((SrcTy->isInteger() || isa<PointerType>(SrcTy)) &&
+ // TODO: For now, not handling conversions like:
+ // (bitcast i64 %x to <2 x i32>)
+ !isa<VectorType>(I->getType())) {
+ ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, TD,
+ Depth+1);
+ return;
+ }
+ break;
+ }
+ case Instruction::SExt: {
+ // Compute the bits in the result that are not present in the input.
+ unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
+
+ APInt MaskIn(Mask);
+ MaskIn.trunc(SrcBitWidth);
+ KnownZero.trunc(SrcBitWidth);
+ KnownOne.trunc(SrcBitWidth);
+ ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, TD,
+ Depth+1);
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ KnownZero.zext(BitWidth);
+ KnownOne.zext(BitWidth);
+
+ // If the sign bit of the input is known set or clear, then we know the
+ // top bits of the result.
+ if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
+ KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
+ else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
+ KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
+ return;
+ }
+ case Instruction::Shl:
+ // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
+ if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
+ uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
+ APInt Mask2(Mask.lshr(ShiftAmt));
+ ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
+ Depth+1);
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ KnownZero <<= ShiftAmt;
+ KnownOne <<= ShiftAmt;
+ KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
+ return;
+ }
+ break;
+ case Instruction::LShr:
+ // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
+ if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
+ // Compute the new bits that are at the top now.
+ uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
+
+ // Unsigned shift right.
+ APInt Mask2(Mask.shl(ShiftAmt));
+ ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne, TD,
+ Depth+1);
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
+ KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
+ // high bits known zero.
+ KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
+ return;
+ }
+ break;
+ case Instruction::AShr:
+ // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
+ if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
+ // Compute the new bits that are at the top now.
+ uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
+
+ // Signed shift right.
+ APInt Mask2(Mask.shl(ShiftAmt));
+ ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
+ Depth+1);
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
+ KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
+
+ APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
+ if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
+ KnownZero |= HighBits;
+ else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
+ KnownOne |= HighBits;
+ return;
+ }
+ break;
+ case Instruction::Sub: {
+ if (ConstantInt *CLHS = dyn_cast<ConstantInt>(I->getOperand(0))) {
+ // We know that the top bits of C-X are clear if X contains less bits
+ // than C (i.e. no wrap-around can happen). For example, 20-X is
+ // positive if we can prove that X is >= 0 and < 16.
+ if (!CLHS->getValue().isNegative()) {
+ unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
+ // NLZ can't be BitWidth with no sign bit
+ APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
+ ComputeMaskedBits(I->getOperand(1), MaskV, KnownZero2, KnownOne2,
+ TD, Depth+1);
+
+ // If all of the MaskV bits are known to be zero, then we know the
+ // output top bits are zero, because we now know that the output is
+ // from [0-C].
+ if ((KnownZero2 & MaskV) == MaskV) {
+ unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
+ // Top bits known zero.
+ KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2) & Mask;
+ }
+ }
+ }
+ }
+ // fall through
+ case Instruction::Add: {
+ // If one of the operands has trailing zeros, then the bits that the
+ // other operand has in those bit positions will be preserved in the
+ // result. For an add, this works with either operand. For a subtract,
+ // this only works if the known zeros are in the right operand.
+ APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
+ APInt Mask2 = APInt::getLowBitsSet(BitWidth,
+ BitWidth - Mask.countLeadingZeros());
+ ComputeMaskedBits(I->getOperand(0), Mask2, LHSKnownZero, LHSKnownOne, TD,
+ Depth+1);
+ assert((LHSKnownZero & LHSKnownOne) == 0 &&
+ "Bits known to be one AND zero?");
+ unsigned LHSKnownZeroOut = LHSKnownZero.countTrailingOnes();
+
+ ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero2, KnownOne2, TD,
+ Depth+1);
+ assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
+ unsigned RHSKnownZeroOut = KnownZero2.countTrailingOnes();
+
+ // Determine which operand has more trailing zeros, and use that
+ // many bits from the other operand.
+ if (LHSKnownZeroOut > RHSKnownZeroOut) {
+ if (I->getOpcode() == Instruction::Add) {
+ APInt Mask = APInt::getLowBitsSet(BitWidth, LHSKnownZeroOut);
+ KnownZero |= KnownZero2 & Mask;
+ KnownOne |= KnownOne2 & Mask;
+ } else {
+ // If the known zeros are in the left operand for a subtract,
+ // fall back to the minimum known zeros in both operands.
+ KnownZero |= APInt::getLowBitsSet(BitWidth,
+ std::min(LHSKnownZeroOut,
+ RHSKnownZeroOut));
+ }
+ } else if (RHSKnownZeroOut >= LHSKnownZeroOut) {
+ APInt Mask = APInt::getLowBitsSet(BitWidth, RHSKnownZeroOut);
+ KnownZero |= LHSKnownZero & Mask;
+ KnownOne |= LHSKnownOne & Mask;
+ }
+ return;
+ }
+ case Instruction::SRem:
+ if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
+ APInt RA = Rem->getValue().abs();
+ if (RA.isPowerOf2()) {
+ APInt LowBits = RA - 1;
+ APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
+ ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
+ Depth+1);
+
+ // The low bits of the first operand are unchanged by the srem.
+ KnownZero = KnownZero2 & LowBits;
+ KnownOne = KnownOne2 & LowBits;
+
+ // If the first operand is non-negative or has all low bits zero, then
+ // the upper bits are all zero.
+ if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
+ KnownZero |= ~LowBits;
+
+ // If the first operand is negative and not all low bits are zero, then
+ // the upper bits are all one.
+ if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0))
+ KnownOne |= ~LowBits;
+
+ KnownZero &= Mask;
+ KnownOne &= Mask;
+
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ }
+ }
+ break;
+ case Instruction::URem: {
+ if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
+ APInt RA = Rem->getValue();
+ if (RA.isPowerOf2()) {
+ APInt LowBits = (RA - 1);
+ APInt Mask2 = LowBits & Mask;
+ KnownZero |= ~LowBits & Mask;
+ ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, TD,
+ Depth+1);
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ break;
+ }
+ }
+
+ // Since the result is less than or equal to either operand, any leading
+ // zero bits in either operand must also exist in the result.
+ APInt AllOnes = APInt::getAllOnesValue(BitWidth);
+ ComputeMaskedBits(I->getOperand(0), AllOnes, KnownZero, KnownOne,
+ TD, Depth+1);
+ ComputeMaskedBits(I->getOperand(1), AllOnes, KnownZero2, KnownOne2,
+ TD, Depth+1);
+
+ unsigned Leaders = std::max(KnownZero.countLeadingOnes(),
+ KnownZero2.countLeadingOnes());
+ KnownOne.clear();
+ KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & Mask;
+ break;
+ }
+
+ case Instruction::Alloca: {
+ AllocaInst *AI = cast<AllocaInst>(V);
+ unsigned Align = AI->getAlignment();
+ if (Align == 0 && TD)
+ Align = TD->getABITypeAlignment(AI->getType()->getElementType());
+
+ if (Align > 0)
+ KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
+ CountTrailingZeros_32(Align));
+ break;
+ }
+ case Instruction::GetElementPtr: {
+ // Analyze all of the subscripts of this getelementptr instruction
+ // to determine if we can prove known low zero bits.
+ APInt LocalMask = APInt::getAllOnesValue(BitWidth);
+ APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
+ ComputeMaskedBits(I->getOperand(0), LocalMask,
+ LocalKnownZero, LocalKnownOne, TD, Depth+1);
+ unsigned TrailZ = LocalKnownZero.countTrailingOnes();
+
+ gep_type_iterator GTI = gep_type_begin(I);
+ for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
+ Value *Index = I->getOperand(i);
+ if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
+ // Handle struct member offset arithmetic.
+ if (!TD) return;
+ const StructLayout *SL = TD->getStructLayout(STy);
+ unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
+ uint64_t Offset = SL->getElementOffset(Idx);
+ TrailZ = std::min(TrailZ,
+ CountTrailingZeros_64(Offset));
+ } else {
+ // Handle array index arithmetic.
+ const Type *IndexedTy = GTI.getIndexedType();
+ if (!IndexedTy->isSized()) return;
+ unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
+ uint64_t TypeSize = TD ? TD->getTypeAllocSize(IndexedTy) : 1;
+ LocalMask = APInt::getAllOnesValue(GEPOpiBits);
+ LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
+ ComputeMaskedBits(Index, LocalMask,
+ LocalKnownZero, LocalKnownOne, TD, Depth+1);
+ TrailZ = std::min(TrailZ,
+ unsigned(CountTrailingZeros_64(TypeSize) +
+ LocalKnownZero.countTrailingOnes()));
+ }
+ }
+
+ KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) & Mask;
+ break;
+ }
+ case Instruction::PHI: {
+ PHINode *P = cast<PHINode>(I);
+ // Handle the case of a simple two-predecessor recurrence PHI.
+ // There's a lot more that could theoretically be done here, but
+ // this is sufficient to catch some interesting cases.
+ if (P->getNumIncomingValues() == 2) {
+ for (unsigned i = 0; i != 2; ++i) {
+ Value *L = P->getIncomingValue(i);
+ Value *R = P->getIncomingValue(!i);
+ Operator *LU = dyn_cast<Operator>(L);
+ if (!LU)
+ continue;
+ unsigned Opcode = LU->getOpcode();
+ // Check for operations that have the property that if
+ // both their operands have low zero bits, the result
+ // will have low zero bits.
+ if (Opcode == Instruction::Add ||
+ Opcode == Instruction::Sub ||
+ Opcode == Instruction::And ||
+ Opcode == Instruction::Or ||
+ Opcode == Instruction::Mul) {
+ Value *LL = LU->getOperand(0);
+ Value *LR = LU->getOperand(1);
+ // Find a recurrence.
+ if (LL == I)
+ L = LR;
+ else if (LR == I)
+ L = LL;
+ else
+ break;
+ // Ok, we have a PHI of the form L op= R. Check for low
+ // zero bits.
+ APInt Mask2 = APInt::getAllOnesValue(BitWidth);
+ ComputeMaskedBits(R, Mask2, KnownZero2, KnownOne2, TD, Depth+1);
+ Mask2 = APInt::getLowBitsSet(BitWidth,
+ KnownZero2.countTrailingOnes());
+
+ // We need to take the minimum number of known bits
+ APInt KnownZero3(KnownZero), KnownOne3(KnownOne);
+ ComputeMaskedBits(L, Mask2, KnownZero3, KnownOne3, TD, Depth+1);
+
+ KnownZero = Mask &
+ APInt::getLowBitsSet(BitWidth,
+ std::min(KnownZero2.countTrailingOnes(),
+ KnownZero3.countTrailingOnes()));
+ break;
+ }
+ }
+ }
+
+ // Otherwise take the unions of the known bit sets of the operands,
+ // taking conservative care to avoid excessive recursion.
+ if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) {
+ KnownZero = APInt::getAllOnesValue(BitWidth);
+ KnownOne = APInt::getAllOnesValue(BitWidth);
+ for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
+ // Skip direct self references.
+ if (P->getIncomingValue(i) == P) continue;
+
+ KnownZero2 = APInt(BitWidth, 0);
+ KnownOne2 = APInt(BitWidth, 0);
+ // Recurse, but cap the recursion to one level, because we don't
+ // want to waste time spinning around in loops.
+ ComputeMaskedBits(P->getIncomingValue(i), KnownZero | KnownOne,
+ KnownZero2, KnownOne2, TD, MaxDepth-1);
+ KnownZero &= KnownZero2;
+ KnownOne &= KnownOne2;
+ // If all bits have been ruled out, there's no need to check
+ // more operands.
+ if (!KnownZero && !KnownOne)
+ break;
+ }
+ }
+ break;
+ }
+ case Instruction::Call:
+ if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
+ switch (II->getIntrinsicID()) {
+ default: break;
+ case Intrinsic::ctpop:
+ case Intrinsic::ctlz:
+ case Intrinsic::cttz: {
+ unsigned LowBits = Log2_32(BitWidth)+1;
+ KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
+ break;
+ }
+ }
+ }
+ break;
+ }
+}
+
+/// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
+/// this predicate to simplify operations downstream. Mask is known to be zero
+/// for bits that V cannot have.
+///
+/// This function is defined on values with integer type, values with pointer
+/// type (but only if TD is non-null), and vectors of integers. In the case
+/// where V is a vector, the mask, known zero, and known one values are the
+/// same width as the vector element, and the bit is set only if it is true
+/// for all of the elements in the vector.
+bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask,
+ const TargetData *TD, unsigned Depth) {
+ APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
+ ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ return (KnownZero & Mask) == Mask;
+}
+
+
+
+/// ComputeNumSignBits - Return the number of times the sign bit of the
+/// register is replicated into the other bits. We know that at least 1 bit
+/// is always equal to the sign bit (itself), but other cases can give us
+/// information. For example, immediately after an "ashr X, 2", we know that
+/// the top 3 bits are all equal to each other, so we return 3.
+///
+/// 'Op' must have a scalar integer type.
+///
+unsigned llvm::ComputeNumSignBits(Value *V, const TargetData *TD,
+ unsigned Depth) {
+ assert((TD || V->getType()->isIntOrIntVector()) &&
+ "ComputeNumSignBits requires a TargetData object to operate "
+ "on non-integer values!");
+ const Type *Ty = V->getType();
+ unsigned TyBits = TD ? TD->getTypeSizeInBits(V->getType()->getScalarType()) :
+ Ty->getScalarSizeInBits();
+ unsigned Tmp, Tmp2;
+ unsigned FirstAnswer = 1;
+
+ // Note that ConstantInt is handled by the general ComputeMaskedBits case
+ // below.
+
+ if (Depth == 6)
+ return 1; // Limit search depth.
+
+ Operator *U = dyn_cast<Operator>(V);
+ switch (Operator::getOpcode(V)) {
+ default: break;
+ case Instruction::SExt:
+ Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
+ return ComputeNumSignBits(U->getOperand(0), TD, Depth+1) + Tmp;
+
+ case Instruction::AShr:
+ Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
+ // ashr X, C -> adds C sign bits.
+ if (ConstantInt *C = dyn_cast<ConstantInt>(U->getOperand(1))) {
+ Tmp += C->getZExtValue();
+ if (Tmp > TyBits) Tmp = TyBits;
+ }
+ return Tmp;
+ case Instruction::Shl:
+ if (ConstantInt *C = dyn_cast<ConstantInt>(U->getOperand(1))) {
+ // shl destroys sign bits.
+ Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
+ if (C->getZExtValue() >= TyBits || // Bad shift.
+ C->getZExtValue() >= Tmp) break; // Shifted all sign bits out.
+ return Tmp - C->getZExtValue();
+ }
+ break;
+ case Instruction::And:
+ case Instruction::Or:
+ case Instruction::Xor: // NOT is handled here.
+ // Logical binary ops preserve the number of sign bits at the worst.
+ Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
+ if (Tmp != 1) {
+ Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
+ FirstAnswer = std::min(Tmp, Tmp2);
+ // We computed what we know about the sign bits as our first
+ // answer. Now proceed to the generic code that uses
+ // ComputeMaskedBits, and pick whichever answer is better.
+ }
+ break;
+
+ case Instruction::Select:
+ Tmp = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
+ if (Tmp == 1) return 1; // Early out.
+ Tmp2 = ComputeNumSignBits(U->getOperand(2), TD, Depth+1);
+ return std::min(Tmp, Tmp2);
+
+ case Instruction::Add:
+ // Add can have at most one carry bit. Thus we know that the output
+ // is, at worst, one more bit than the inputs.
+ Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
+ if (Tmp == 1) return 1; // Early out.
+
+ // Special case decrementing a value (ADD X, -1):
+ if (ConstantInt *CRHS = dyn_cast<ConstantInt>(U->getOperand(1)))
+ if (CRHS->isAllOnesValue()) {
+ APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
+ APInt Mask = APInt::getAllOnesValue(TyBits);
+ ComputeMaskedBits(U->getOperand(0), Mask, KnownZero, KnownOne, TD,
+ Depth+1);
+
+ // If the input is known to be 0 or 1, the output is 0/-1, which is all
+ // sign bits set.
+ if ((KnownZero | APInt(TyBits, 1)) == Mask)
+ return TyBits;
+
+ // If we are subtracting one from a positive number, there is no carry
+ // out of the result.
+ if (KnownZero.isNegative())
+ return Tmp;
+ }
+
+ Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
+ if (Tmp2 == 1) return 1;
+ return std::min(Tmp, Tmp2)-1;
+
+ case Instruction::Sub:
+ Tmp2 = ComputeNumSignBits(U->getOperand(1), TD, Depth+1);
+ if (Tmp2 == 1) return 1;
+
+ // Handle NEG.
+ if (ConstantInt *CLHS = dyn_cast<ConstantInt>(U->getOperand(0)))
+ if (CLHS->isNullValue()) {
+ APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
+ APInt Mask = APInt::getAllOnesValue(TyBits);
+ ComputeMaskedBits(U->getOperand(1), Mask, KnownZero, KnownOne,
+ TD, Depth+1);
+ // If the input is known to be 0 or 1, the output is 0/-1, which is all
+ // sign bits set.
+ if ((KnownZero | APInt(TyBits, 1)) == Mask)
+ return TyBits;
+
+ // If the input is known to be positive (the sign bit is known clear),
+ // the output of the NEG has the same number of sign bits as the input.
+ if (KnownZero.isNegative())
+ return Tmp2;
+
+ // Otherwise, we treat this like a SUB.
+ }
+
+ // Sub can have at most one carry bit. Thus we know that the output
+ // is, at worst, one more bit than the inputs.
+ Tmp = ComputeNumSignBits(U->getOperand(0), TD, Depth+1);
+ if (Tmp == 1) return 1; // Early out.
+ return std::min(Tmp, Tmp2)-1;
+
+ case Instruction::PHI: {
+ PHINode *PN = cast<PHINode>(U);
+ // Don't analyze large in-degree PHIs.
+ if (PN->getNumIncomingValues() > 4) break;
+
+ // Take the minimum of all incoming values. This can't infinitely loop
+ // because of our depth threshold.
+ Tmp = ComputeNumSignBits(PN->getIncomingValue(0), TD, Depth+1);
+ for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
+ if (Tmp == 1) return Tmp;
+ Tmp = std::min(Tmp,
+ ComputeNumSignBits(PN->getIncomingValue(1), TD, Depth+1));
+ }
+ return Tmp;
+ }
+
+ case Instruction::Trunc:
+ // FIXME: it's tricky to do anything useful for this, but it is an important
+ // case for targets like X86.
+ break;
+ }
+
+ // Finally, if we can prove that the top bits of the result are 0's or 1's,
+ // use this information.
+ APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0);
+ APInt Mask = APInt::getAllOnesValue(TyBits);
+ ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
+
+ if (KnownZero.isNegative()) { // sign bit is 0
+ Mask = KnownZero;
+ } else if (KnownOne.isNegative()) { // sign bit is 1;
+ Mask = KnownOne;
+ } else {
+ // Nothing known.
+ return FirstAnswer;
+ }
+
+ // Okay, we know that the sign bit in Mask is set. Use CLZ to determine
+ // the number of identical bits in the top of the input value.
+ Mask = ~Mask;
+ Mask <<= Mask.getBitWidth()-TyBits;
+ // Return # leading zeros. We use 'min' here in case Val was zero before
+ // shifting. We don't want to return '64' as for an i32 "0".
+ return std::max(FirstAnswer, std::min(TyBits, Mask.countLeadingZeros()));
+}
+
+/// ComputeMultiple - This function computes the integer multiple of Base that
+/// equals V. If successful, it returns true and returns the multiple in
+/// Multiple. If unsuccessful, it returns false. It looks
+/// through SExt instructions only if LookThroughSExt is true.
+bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
+ bool LookThroughSExt, unsigned Depth) {
+ const unsigned MaxDepth = 6;
+
+ assert(V && "No Value?");
+ assert(Depth <= MaxDepth && "Limit Search Depth");
+ assert(V->getType()->isInteger() && "Not integer or pointer type!");
+
+ const Type *T = V->getType();
+
+ ConstantInt *CI = dyn_cast<ConstantInt>(V);
+
+ if (Base == 0)
+ return false;
+
+ if (Base == 1) {
+ Multiple = V;
+ return true;
+ }
+
+ ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
+ Constant *BaseVal = ConstantInt::get(T, Base);
+ if (CO && CO == BaseVal) {
+ // Multiple is 1.
+ Multiple = ConstantInt::get(T, 1);
+ return true;
+ }
+
+ if (CI && CI->getZExtValue() % Base == 0) {
+ Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
+ return true;
+ }
+
+ if (Depth == MaxDepth) return false; // Limit search depth.
+
+ Operator *I = dyn_cast<Operator>(V);
+ if (!I) return false;
+
+ switch (I->getOpcode()) {
+ default: break;
+ case Instruction::SExt:
+ if (!LookThroughSExt) return false;
+ // otherwise fall through to ZExt
+ case Instruction::ZExt:
+ return ComputeMultiple(I->getOperand(0), Base, Multiple,
+ LookThroughSExt, Depth+1);
+ case Instruction::Shl:
+ case Instruction::Mul: {
+ Value *Op0 = I->getOperand(0);
+ Value *Op1 = I->getOperand(1);
+
+ if (I->getOpcode() == Instruction::Shl) {
+ ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
+ if (!Op1CI) return false;
+ // Turn Op0 << Op1 into Op0 * 2^Op1
+ APInt Op1Int = Op1CI->getValue();
+ uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
+ Op1 = ConstantInt::get(V->getContext(),
+ APInt(Op1Int.getBitWidth(), 0).set(BitToSet));
+ }
+
+ Value *Mul0 = NULL;
+ Value *Mul1 = NULL;
+ bool M0 = ComputeMultiple(Op0, Base, Mul0,
+ LookThroughSExt, Depth+1);
+ bool M1 = ComputeMultiple(Op1, Base, Mul1,
+ LookThroughSExt, Depth+1);
+
+ if (M0) {
+ if (isa<Constant>(Op1) && isa<Constant>(Mul0)) {
+ // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
+ Multiple = ConstantExpr::getMul(cast<Constant>(Mul0),
+ cast<Constant>(Op1));
+ return true;
+ }
+
+ if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
+ if (Mul0CI->getValue() == 1) {
+ // V == Base * Op1, so return Op1
+ Multiple = Op1;
+ return true;
+ }
+ }
+
+ if (M1) {
+ if (isa<Constant>(Op0) && isa<Constant>(Mul1)) {
+ // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
+ Multiple = ConstantExpr::getMul(cast<Constant>(Mul1),
+ cast<Constant>(Op0));
+ return true;
+ }
+
+ if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
+ if (Mul1CI->getValue() == 1) {
+ // V == Base * Op0, so return Op0
+ Multiple = Op0;
+ return true;
+ }
+ }
+ }
+ }
+
+ // We could not determine if V is a multiple of Base.
+ return false;
+}
+
+/// CannotBeNegativeZero - Return true if we can prove that the specified FP
+/// value is never equal to -0.0.
+///
+/// NOTE: this function will need to be revisited when we support non-default
+/// rounding modes!
+///
+bool llvm::CannotBeNegativeZero(const Value *V, unsigned Depth) {
+ if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
+ return !CFP->getValueAPF().isNegZero();
+
+ if (Depth == 6)
+ return 1; // Limit search depth.
+
+ const Operator *I = dyn_cast<Operator>(V);
+ if (I == 0) return false;
+
+ // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
+ if (I->getOpcode() == Instruction::FAdd &&
+ isa<ConstantFP>(I->getOperand(1)) &&
+ cast<ConstantFP>(I->getOperand(1))->isNullValue())
+ return true;
+
+ // sitofp and uitofp turn into +0.0 for zero.
+ if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I))
+ return true;
+
+ if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
+ // sqrt(-0.0) = -0.0, no other negative results are possible.
+ if (II->getIntrinsicID() == Intrinsic::sqrt)
+ return CannotBeNegativeZero(II->getOperand(1), Depth+1);
+
+ if (const CallInst *CI = dyn_cast<CallInst>(I))
+ if (const Function *F = CI->getCalledFunction()) {
+ if (F->isDeclaration()) {
+ // abs(x) != -0.0
+ if (F->getName() == "abs") return true;
+ // fabs[lf](x) != -0.0
+ if (F->getName() == "fabs") return true;
+ if (F->getName() == "fabsf") return true;
+ if (F->getName() == "fabsl") return true;
+ if (F->getName() == "sqrt" || F->getName() == "sqrtf" ||
+ F->getName() == "sqrtl")
+ return CannotBeNegativeZero(CI->getOperand(1), Depth+1);
+ }
+ }
+
+ return false;
+}
+
+
+/// GetLinearExpression - Analyze the specified value as a linear expression:
+/// "A*V + B", where A and B are constant integers. Return the scale and offset
+/// values as APInts and return V as a Value*. The incoming Value is known to
+/// have IntegerType. Note that this looks through extends, so the high bits
+/// may not be represented in the result.
+static Value *GetLinearExpression(Value *V, APInt &Scale, APInt &Offset,
+ const TargetData *TD, unsigned Depth) {
+ assert(isa<IntegerType>(V->getType()) && "Not an integer value");
+
+ // Limit our recursion depth.
+ if (Depth == 6) {
+ Scale = 1;
+ Offset = 0;
+ return V;
+ }
+
+ if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(V)) {
+ if (ConstantInt *RHSC = dyn_cast<ConstantInt>(BOp->getOperand(1))) {
+ switch (BOp->getOpcode()) {
+ default: break;
+ case Instruction::Or:
+ // X|C == X+C if all the bits in C are unset in X. Otherwise we can't
+ // analyze it.
+ if (!MaskedValueIsZero(BOp->getOperand(0), RHSC->getValue(), TD))
+ break;
+ // FALL THROUGH.
+ case Instruction::Add:
+ V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, TD, Depth+1);
+ Offset += RHSC->getValue();
+ return V;
+ case Instruction::Mul:
+ V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, TD, Depth+1);
+ Offset *= RHSC->getValue();
+ Scale *= RHSC->getValue();
+ return V;
+ case Instruction::Shl:
+ V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, TD, Depth+1);
+ Offset <<= RHSC->getValue().getLimitedValue();
+ Scale <<= RHSC->getValue().getLimitedValue();
+ return V;
+ }
+ }
+ }
+
+ // Since clients don't care about the high bits of the value, just scales and
+ // offsets, we can look through extensions.
+ if (isa<SExtInst>(V) || isa<ZExtInst>(V)) {
+ Value *CastOp = cast<CastInst>(V)->getOperand(0);
+ unsigned OldWidth = Scale.getBitWidth();
+ unsigned SmallWidth = CastOp->getType()->getPrimitiveSizeInBits();
+ Scale.trunc(SmallWidth);
+ Offset.trunc(SmallWidth);
+ Value *Result = GetLinearExpression(CastOp, Scale, Offset, TD, Depth+1);
+ Scale.zext(OldWidth);
+ Offset.zext(OldWidth);
+ return Result;
+ }
+
+ Scale = 1;
+ Offset = 0;
+ return V;
+}
+
+/// DecomposeGEPExpression - If V is a symbolic pointer expression, decompose it
+/// into a base pointer with a constant offset and a number of scaled symbolic
+/// offsets.
+///
+/// The scaled symbolic offsets (represented by pairs of a Value* and a scale in
+/// the VarIndices vector) are Value*'s that are known to be scaled by the
+/// specified amount, but which may have other unrepresented high bits. As such,
+/// the gep cannot necessarily be reconstructed from its decomposed form.
+///
+/// When TargetData is around, this function is capable of analyzing everything
+/// that Value::getUnderlyingObject() can look through. When not, it just looks
+/// through pointer casts.
+///
+const Value *llvm::DecomposeGEPExpression(const Value *V, int64_t &BaseOffs,
+ SmallVectorImpl<std::pair<const Value*, int64_t> > &VarIndices,
+ const TargetData *TD) {
+ // Limit recursion depth to limit compile time in crazy cases.
+ unsigned MaxLookup = 6;
+
+ BaseOffs = 0;
+ do {
+ // See if this is a bitcast or GEP.
+ const Operator *Op = dyn_cast<Operator>(V);
+ if (Op == 0) {
+ // The only non-operator case we can handle are GlobalAliases.
+ if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
+ if (!GA->mayBeOverridden()) {
+ V = GA->getAliasee();
+ continue;
+ }
+ }
+ return V;
+ }
+
+ if (Op->getOpcode() == Instruction::BitCast) {
+ V = Op->getOperand(0);
+ continue;
+ }
+
+ const GEPOperator *GEPOp = dyn_cast<GEPOperator>(Op);
+ if (GEPOp == 0)
+ return V;
+
+ // Don't attempt to analyze GEPs over unsized objects.
+ if (!cast<PointerType>(GEPOp->getOperand(0)->getType())
+ ->getElementType()->isSized())
+ return V;
+
+ // If we are lacking TargetData information, we can't compute the offets of
+ // elements computed by GEPs. However, we can handle bitcast equivalent
+ // GEPs.
+ if (!TD) {
+ if (!GEPOp->hasAllZeroIndices())
+ return V;
+ V = GEPOp->getOperand(0);
+ continue;
+ }
+
+ // Walk the indices of the GEP, accumulating them into BaseOff/VarIndices.
+ gep_type_iterator GTI = gep_type_begin(GEPOp);
+ for (User::const_op_iterator I = GEPOp->op_begin()+1,
+ E = GEPOp->op_end(); I != E; ++I) {
+ Value *Index = *I;
+ // Compute the (potentially symbolic) offset in bytes for this index.
+ if (const StructType *STy = dyn_cast<StructType>(*GTI++)) {
+ // For a struct, add the member offset.
+ unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
+ if (FieldNo == 0) continue;
+
+ BaseOffs += TD->getStructLayout(STy)->getElementOffset(FieldNo);
+ continue;
+ }
+
+ // For an array/pointer, add the element offset, explicitly scaled.
+ if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Index)) {
+ if (CIdx->isZero()) continue;
+ BaseOffs += TD->getTypeAllocSize(*GTI)*CIdx->getSExtValue();
+ continue;
+ }
+
+ uint64_t Scale = TD->getTypeAllocSize(*GTI);
+
+ // Use GetLinearExpression to decompose the index into a C1*V+C2 form.
+ unsigned Width = cast<IntegerType>(Index->getType())->getBitWidth();
+ APInt IndexScale(Width, 0), IndexOffset(Width, 0);
+ Index = GetLinearExpression(Index, IndexScale, IndexOffset, TD, 0);
+
+ // The GEP index scale ("Scale") scales C1*V+C2, yielding (C1*V+C2)*Scale.
+ // This gives us an aggregate computation of (C1*Scale)*V + C2*Scale.
+ BaseOffs += IndexOffset.getZExtValue()*Scale;
+ Scale *= IndexScale.getZExtValue();
+
+
+ // If we already had an occurrance of this index variable, merge this
+ // scale into it. For example, we want to handle:
+ // A[x][x] -> x*16 + x*4 -> x*20
+ // This also ensures that 'x' only appears in the index list once.
+ for (unsigned i = 0, e = VarIndices.size(); i != e; ++i) {
+ if (VarIndices[i].first == Index) {
+ Scale += VarIndices[i].second;
+ VarIndices.erase(VarIndices.begin()+i);
+ break;
+ }
+ }
+
+ // Make sure that we have a scale that makes sense for this target's
+ // pointer size.
+ if (unsigned ShiftBits = 64-TD->getPointerSizeInBits()) {
+ Scale <<= ShiftBits;
+ Scale >>= ShiftBits;
+ }
+
+ if (Scale)
+ VarIndices.push_back(std::make_pair(Index, Scale));
+ }
+
+ // Analyze the base pointer next.
+ V = GEPOp->getOperand(0);
+ } while (--MaxLookup);
+
+ // If the chain of expressions is too deep, just return early.
+ return V;
+}
+
+
+// This is the recursive version of BuildSubAggregate. It takes a few different
+// arguments. Idxs is the index within the nested struct From that we are
+// looking at now (which is of type IndexedType). IdxSkip is the number of
+// indices from Idxs that should be left out when inserting into the resulting
+// struct. To is the result struct built so far, new insertvalue instructions
+// build on that.
+static Value *BuildSubAggregate(Value *From, Value* To, const Type *IndexedType,
+ SmallVector<unsigned, 10> &Idxs,
+ unsigned IdxSkip,
+ Instruction *InsertBefore) {
+ const llvm::StructType *STy = llvm::dyn_cast<llvm::StructType>(IndexedType);
+ if (STy) {
+ // Save the original To argument so we can modify it
+ Value *OrigTo = To;
+ // General case, the type indexed by Idxs is a struct
+ for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
+ // Process each struct element recursively
+ Idxs.push_back(i);
+ Value *PrevTo = To;
+ To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
+ InsertBefore);
+ Idxs.pop_back();
+ if (!To) {
+ // Couldn't find any inserted value for this index? Cleanup
+ while (PrevTo != OrigTo) {
+ InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
+ PrevTo = Del->getAggregateOperand();
+ Del->eraseFromParent();
+ }
+ // Stop processing elements
+ break;
+ }
+ }
+ // If we succesfully found a value for each of our subaggregates
+ if (To)
+ return To;
+ }
+ // Base case, the type indexed by SourceIdxs is not a struct, or not all of
+ // the struct's elements had a value that was inserted directly. In the latter
+ // case, perhaps we can't determine each of the subelements individually, but
+ // we might be able to find the complete struct somewhere.
+
+ // Find the value that is at that particular spot
+ Value *V = FindInsertedValue(From, Idxs.begin(), Idxs.end());
+
+ if (!V)
+ return NULL;
+
+ // Insert the value in the new (sub) aggregrate
+ return llvm::InsertValueInst::Create(To, V, Idxs.begin() + IdxSkip,
+ Idxs.end(), "tmp", InsertBefore);
+}
+
+// This helper takes a nested struct and extracts a part of it (which is again a
+// struct) into a new value. For example, given the struct:
+// { a, { b, { c, d }, e } }
+// and the indices "1, 1" this returns
+// { c, d }.
+//
+// It does this by inserting an insertvalue for each element in the resulting
+// struct, as opposed to just inserting a single struct. This will only work if
+// each of the elements of the substruct are known (ie, inserted into From by an
+// insertvalue instruction somewhere).
+//
+// All inserted insertvalue instructions are inserted before InsertBefore
+static Value *BuildSubAggregate(Value *From, const unsigned *idx_begin,
+ const unsigned *idx_end,
+ Instruction *InsertBefore) {
+ assert(InsertBefore && "Must have someplace to insert!");
+ const Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
+ idx_begin,
+ idx_end);
+ Value *To = UndefValue::get(IndexedType);
+ SmallVector<unsigned, 10> Idxs(idx_begin, idx_end);
+ unsigned IdxSkip = Idxs.size();
+
+ return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
+}
+
+/// FindInsertedValue - Given an aggregrate and an sequence of indices, see if
+/// the scalar value indexed is already around as a register, for example if it
+/// were inserted directly into the aggregrate.
+///
+/// If InsertBefore is not null, this function will duplicate (modified)
+/// insertvalues when a part of a nested struct is extracted.
+Value *llvm::FindInsertedValue(Value *V, const unsigned *idx_begin,
+ const unsigned *idx_end, Instruction *InsertBefore) {
+ // Nothing to index? Just return V then (this is useful at the end of our
+ // recursion)
+ if (idx_begin == idx_end)
+ return V;
+ // We have indices, so V should have an indexable type
+ assert((isa<StructType>(V->getType()) || isa<ArrayType>(V->getType()))
+ && "Not looking at a struct or array?");
+ assert(ExtractValueInst::getIndexedType(V->getType(), idx_begin, idx_end)
+ && "Invalid indices for type?");
+ const CompositeType *PTy = cast<CompositeType>(V->getType());
+
+ if (isa<UndefValue>(V))
+ return UndefValue::get(ExtractValueInst::getIndexedType(PTy,
+ idx_begin,
+ idx_end));
+ else if (isa<ConstantAggregateZero>(V))
+ return Constant::getNullValue(ExtractValueInst::getIndexedType(PTy,
+ idx_begin,
+ idx_end));
+ else if (Constant *C = dyn_cast<Constant>(V)) {
+ if (isa<ConstantArray>(C) || isa<ConstantStruct>(C))
+ // Recursively process this constant
+ return FindInsertedValue(C->getOperand(*idx_begin), idx_begin + 1,
+ idx_end, InsertBefore);
+ } else if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
+ // Loop the indices for the insertvalue instruction in parallel with the
+ // requested indices
+ const unsigned *req_idx = idx_begin;
+ for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
+ i != e; ++i, ++req_idx) {
+ if (req_idx == idx_end) {
+ if (InsertBefore)
+ // The requested index identifies a part of a nested aggregate. Handle
+ // this specially. For example,
+ // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
+ // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
+ // %C = extractvalue {i32, { i32, i32 } } %B, 1
+ // This can be changed into
+ // %A = insertvalue {i32, i32 } undef, i32 10, 0
+ // %C = insertvalue {i32, i32 } %A, i32 11, 1
+ // which allows the unused 0,0 element from the nested struct to be
+ // removed.
+ return BuildSubAggregate(V, idx_begin, req_idx, InsertBefore);
+ else
+ // We can't handle this without inserting insertvalues
+ return 0;
+ }
+
+ // This insert value inserts something else than what we are looking for.
+ // See if the (aggregrate) value inserted into has the value we are
+ // looking for, then.
+ if (*req_idx != *i)
+ return FindInsertedValue(I->getAggregateOperand(), idx_begin, idx_end,
+ InsertBefore);
+ }
+ // If we end up here, the indices of the insertvalue match with those
+ // requested (though possibly only partially). Now we recursively look at
+ // the inserted value, passing any remaining indices.
+ return FindInsertedValue(I->getInsertedValueOperand(), req_idx, idx_end,
+ InsertBefore);
+ } else if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
+ // If we're extracting a value from an aggregrate that was extracted from
+ // something else, we can extract from that something else directly instead.
+ // However, we will need to chain I's indices with the requested indices.
+
+ // Calculate the number of indices required
+ unsigned size = I->getNumIndices() + (idx_end - idx_begin);
+ // Allocate some space to put the new indices in
+ SmallVector<unsigned, 5> Idxs;
+ Idxs.reserve(size);
+ // Add indices from the extract value instruction
+ for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
+ i != e; ++i)
+ Idxs.push_back(*i);
+
+ // Add requested indices
+ for (const unsigned *i = idx_begin, *e = idx_end; i != e; ++i)
+ Idxs.push_back(*i);
+
+ assert(Idxs.size() == size
+ && "Number of indices added not correct?");
+
+ return FindInsertedValue(I->getAggregateOperand(), Idxs.begin(), Idxs.end(),
+ InsertBefore);
+ }
+ // Otherwise, we don't know (such as, extracting from a function return value
+ // or load instruction)
+ return 0;
+}
+
+/// GetConstantStringInfo - This function computes the length of a
+/// null-terminated C string pointed to by V. If successful, it returns true
+/// and returns the string in Str. If unsuccessful, it returns false.
+bool llvm::GetConstantStringInfo(Value *V, std::string &Str, uint64_t Offset,
+ bool StopAtNul) {
+ // If V is NULL then return false;
+ if (V == NULL) return false;
+
+ // Look through bitcast instructions.
+ if (BitCastInst *BCI = dyn_cast<BitCastInst>(V))
+ return GetConstantStringInfo(BCI->getOperand(0), Str, Offset, StopAtNul);
+
+ // If the value is not a GEP instruction nor a constant expression with a
+ // GEP instruction, then return false because ConstantArray can't occur
+ // any other way
+ User *GEP = 0;
+ if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(V)) {
+ GEP = GEPI;
+ } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
+ if (CE->getOpcode() == Instruction::BitCast)
+ return GetConstantStringInfo(CE->getOperand(0), Str, Offset, StopAtNul);
+ if (CE->getOpcode() != Instruction::GetElementPtr)
+ return false;
+ GEP = CE;
+ }
+
+ if (GEP) {
+ // Make sure the GEP has exactly three arguments.
+ if (GEP->getNumOperands() != 3)
+ return false;
+
+ // Make sure the index-ee is a pointer to array of i8.
+ const PointerType *PT = cast<PointerType>(GEP->getOperand(0)->getType());
+ const ArrayType *AT = dyn_cast<ArrayType>(PT->getElementType());
+ if (AT == 0 || !AT->getElementType()->isInteger(8))
+ return false;
+
+ // Check to make sure that the first operand of the GEP is an integer and
+ // has value 0 so that we are sure we're indexing into the initializer.
+ ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
+ if (FirstIdx == 0 || !FirstIdx->isZero())
+ return false;
+
+ // If the second index isn't a ConstantInt, then this is a variable index
+ // into the array. If this occurs, we can't say anything meaningful about
+ // the string.
+ uint64_t StartIdx = 0;
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
+ StartIdx = CI->getZExtValue();
+ else
+ return false;
+ return GetConstantStringInfo(GEP->getOperand(0), Str, StartIdx+Offset,
+ StopAtNul);
+ }
+
+ // The GEP instruction, constant or instruction, must reference a global
+ // variable that is a constant and is initialized. The referenced constant
+ // initializer is the array that we'll use for optimization.
+ GlobalVariable* GV = dyn_cast<GlobalVariable>(V);
+ if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
+ return false;
+ Constant *GlobalInit = GV->getInitializer();
+
+ // Handle the ConstantAggregateZero case
+ if (isa<ConstantAggregateZero>(GlobalInit)) {
+ // This is a degenerate case. The initializer is constant zero so the
+ // length of the string must be zero.
+ Str.clear();
+ return true;
+ }
+
+ // Must be a Constant Array
+ ConstantArray *Array = dyn_cast<ConstantArray>(GlobalInit);
+ if (Array == 0 || !Array->getType()->getElementType()->isInteger(8))
+ return false;
+
+ // Get the number of elements in the array
+ uint64_t NumElts = Array->getType()->getNumElements();
+
+ if (Offset > NumElts)
+ return false;
+
+ // Traverse the constant array from 'Offset' which is the place the GEP refers
+ // to in the array.
+ Str.reserve(NumElts-Offset);
+ for (unsigned i = Offset; i != NumElts; ++i) {
+ Constant *Elt = Array->getOperand(i);
+ ConstantInt *CI = dyn_cast<ConstantInt>(Elt);
+ if (!CI) // This array isn't suitable, non-int initializer.
+ return false;
+ if (StopAtNul && CI->isZero())
+ return true; // we found end of string, success!
+ Str += (char)CI->getZExtValue();
+ }
+
+ // The array isn't null terminated, but maybe this is a memcpy, not a strcpy.
+ return true;
+}