Gitiles
Code Review Sign In
gerrit-public.fairphone.software / fp2-dev / platform / external / llvm / c0f376dba2b51aacf06cfefbc1c46120bb8b725d / . / lib / Target / SparcV9 / SparcV9BurgISel.cpp
blob: 2cf6a64d643b2291bf3eb7a1fae6af9dc6cab273 [file] [log] [blame]
//===- SparcV9BurgISel.cpp - SparcV9 BURG-based Instruction Selector ------===//
//
// The LLVM Compiler Infrastructure
//
// This file was developed by the LLVM research group and is distributed under
// the University of Illinois Open Source License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// SparcV9 BURG-based instruction selector. It uses the SSA graph to
// construct a forest of BURG instruction trees (class InstrForest) and then
// uses the BURG-generated tree grammar (BURM) to find the optimal instruction
// sequences for the SparcV9.
//
//===----------------------------------------------------------------------===//
#include "MachineInstrAnnot.h"
#include "SparcV9BurgISel.h"
#include "SparcV9InstrForest.h"
#include "SparcV9Internals.h"
#include "SparcV9TmpInstr.h"
#include "SparcV9FrameInfo.h"
#include "SparcV9RegisterInfo.h"
#include "MachineFunctionInfo.h"
#include "llvm/CodeGen/IntrinsicLowering.h"
#include "llvm/CodeGen/MachineConstantPool.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/Constants.h"
#include "llvm/DerivedTypes.h"
#include "llvm/Instructions.h"
#include "llvm/Intrinsics.h"
#include "llvm/Module.h"
#include "llvm/Pass.h"
#include "llvm/Support/CFG.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Type.h"
#include "llvm/Config/alloca.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/LeakDetector.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/hash_map"
#include <algorithm>
#include <cmath>
#include <iostream>
using namespace llvm;
//==------------------------------------------------------------------------==//
// InstrForest (V9ISel BURG instruction trees) implementation
//==------------------------------------------------------------------------==//
namespace llvm {
class InstructionNode : public InstrTreeNode {
bool codeIsFoldedIntoParent;
public:
InstructionNode(Instruction *_instr);
Instruction *getInstruction() const {
assert(treeNodeType == NTInstructionNode);
return cast<Instruction>(val);
}
void markFoldedIntoParent() { codeIsFoldedIntoParent = true; }
bool isFoldedIntoParent() { return codeIsFoldedIntoParent; }
// Methods to support type inquiry through isa, cast, and dyn_cast:
static inline bool classof(const InstructionNode *N) { return true; }
static inline bool classof(const InstrTreeNode *N) {
return N->getNodeType() == InstrTreeNode::NTInstructionNode;
}
protected:
virtual void dumpNode(int indent) const;
};
class VRegListNode : public InstrTreeNode {
public:
VRegListNode() : InstrTreeNode(NTVRegListNode, 0) { opLabel = VRegListOp; }
// Methods to support type inquiry through isa, cast, and dyn_cast:
static inline bool classof(const VRegListNode *N) { return true; }
static inline bool classof(const InstrTreeNode *N) {
return N->getNodeType() == InstrTreeNode::NTVRegListNode;
}
protected:
virtual void dumpNode(int indent) const;
};
class VRegNode : public InstrTreeNode {
public:
VRegNode(Value* _val) : InstrTreeNode(NTVRegNode, _val) {
opLabel = VRegNodeOp;
}
// Methods to support type inquiry through isa, cast, and dyn_cast:
static inline bool classof(const VRegNode *N) { return true; }
static inline bool classof(const InstrTreeNode *N) {
return N->getNodeType() == InstrTreeNode::NTVRegNode;
}
protected:
virtual void dumpNode(int indent) const;
};
class ConstantNode : public InstrTreeNode {
public:
ConstantNode(Constant *constVal)
: InstrTreeNode(NTConstNode, (Value*)constVal) {
opLabel = ConstantNodeOp;
}
Constant *getConstVal() const { return (Constant*) val;}
// Methods to support type inquiry through isa, cast, and dyn_cast:
static inline bool classof(const ConstantNode *N) { return true; }
static inline bool classof(const InstrTreeNode *N) {
return N->getNodeType() == InstrTreeNode::NTConstNode;
}
protected:
virtual void dumpNode(int indent) const;
};
class LabelNode : public InstrTreeNode {
public:
LabelNode(BasicBlock* BB) : InstrTreeNode(NTLabelNode, (Value*)BB) {
opLabel = LabelNodeOp;
}
BasicBlock *getBasicBlock() const { return (BasicBlock*)val;}
// Methods to support type inquiry through isa, cast, and dyn_cast:
static inline bool classof(const LabelNode *N) { return true; }
static inline bool classof(const InstrTreeNode *N) {
return N->getNodeType() == InstrTreeNode::NTLabelNode;
}
protected:
virtual void dumpNode(int indent) const;
};
/// InstrForest - A forest of instruction trees for a single function.
/// The goal of InstrForest is to group instructions into a single
/// tree if one or more of them might be potentially combined into a
/// single complex instruction in the target machine. We group two
/// instructions O and I if: (1) Instruction O computes an operand used
/// by instruction I, and (2) O and I are part of the same basic block,
/// and (3) O has only a single use, viz., I.
///
class InstrForest : private hash_map<const Instruction *, InstructionNode*> {
public:
// Use a vector for the root set to get a deterministic iterator
// for stable code generation. Even though we need to erase nodes
// during forest construction, a vector should still be efficient
// because the elements to erase are nearly always near the end.
typedef std::vector<InstructionNode*> RootSet;
typedef RootSet:: iterator root_iterator;
typedef RootSet::const_iterator const_root_iterator;
private:
RootSet treeRoots;
public:
/*ctor*/ InstrForest (Function *F);
/*dtor*/ ~InstrForest ();
/// getTreeNodeForInstr - Returns the tree node for an Instruction.
///
inline InstructionNode *getTreeNodeForInstr(Instruction* instr) {
return (*this)[instr];
}
/// Iterators for the root nodes for all the trees.
///
const_root_iterator roots_begin() const { return treeRoots.begin(); }
root_iterator roots_begin() { return treeRoots.begin(); }
const_root_iterator roots_end () const { return treeRoots.end(); }
root_iterator roots_end () { return treeRoots.end(); }
void dump() const;
private:
// Methods used to build the instruction forest.
void eraseRoot (InstructionNode* node);
void setLeftChild (InstrTreeNode* parent, InstrTreeNode* child);
void setRightChild(InstrTreeNode* parent, InstrTreeNode* child);
void setParent (InstrTreeNode* child, InstrTreeNode* parent);
void noteTreeNodeForInstr(Instruction* instr, InstructionNode* treeNode);
InstructionNode* buildTreeForInstruction(Instruction* instr);
};
void InstrTreeNode::dump(int dumpChildren, int indent) const {
dumpNode(indent);
if (dumpChildren) {
if (LeftChild)
LeftChild->dump(dumpChildren, indent+1);
if (RightChild)
RightChild->dump(dumpChildren, indent+1);
}
}
InstructionNode::InstructionNode(Instruction* I)
: InstrTreeNode(NTInstructionNode, I), codeIsFoldedIntoParent(false) {
opLabel = I->getOpcode();
// Distinguish special cases of some instructions such as Ret and Br
//
if (opLabel == Instruction::Ret && cast<ReturnInst>(I)->getReturnValue()) {
opLabel = RetValueOp; // ret(value) operation
}
else if (opLabel ==Instruction::Br && !cast<BranchInst>(I)->isUnconditional())
{
opLabel = BrCondOp; // br(cond) operation
} else if (opLabel >= Instruction::SetEQ && opLabel <= Instruction::SetGT) {
opLabel = SetCCOp; // common label for all SetCC ops
} else if (opLabel == Instruction::Alloca && I->getNumOperands() > 0) {
opLabel = AllocaN; // Alloca(ptr, N) operation
} else if (opLabel == Instruction::GetElementPtr &&
cast<GetElementPtrInst>(I)->hasIndices()) {
opLabel = opLabel + 100; // getElem with index vector
} else if (opLabel == Instruction::Xor &&
BinaryOperator::isNot(I)) {
opLabel = (I->getType() == Type::BoolTy)? NotOp // boolean Not operator
: BNotOp; // bitwise Not operator
} else if (opLabel == Instruction::And || opLabel == Instruction::Or ||
opLabel == Instruction::Xor) {
// Distinguish bitwise operators from logical operators!
if (I->getType() != Type::BoolTy)
opLabel = opLabel + 100; // bitwise operator
} else if (opLabel == Instruction::Cast) {
const Type *ITy = I->getType();
switch(ITy->getTypeID())
{
case Type::BoolTyID: opLabel = ToBoolTy; break;
case Type::UByteTyID: opLabel = ToUByteTy; break;
case Type::SByteTyID: opLabel = ToSByteTy; break;
case Type::UShortTyID: opLabel = ToUShortTy; break;
case Type::ShortTyID: opLabel = ToShortTy; break;
case Type::UIntTyID: opLabel = ToUIntTy; break;
case Type::IntTyID: opLabel = ToIntTy; break;
case Type::ULongTyID: opLabel = ToULongTy; break;
case Type::LongTyID: opLabel = ToLongTy; break;
case Type::FloatTyID: opLabel = ToFloatTy; break;
case Type::DoubleTyID: opLabel = ToDoubleTy; break;
case Type::ArrayTyID: opLabel = ToArrayTy; break;
case Type::PointerTyID: opLabel = ToPointerTy; break;
default:
// Just use `Cast' opcode otherwise. It's probably ignored.
break;
}
}
}
void InstructionNode::dumpNode(int indent) const {
for (int i=0; i < indent; i++)
std::cerr << " ";
std::cerr << getInstruction()->getOpcodeName()
<< " [label " << getOpLabel() << "]" << "\n";
}
void VRegListNode::dumpNode(int indent) const {
for (int i=0; i < indent; i++)
std::cerr << " ";
std::cerr << "List" << "\n";
}
void VRegNode::dumpNode(int indent) const {
for (int i=0; i < indent; i++)
std::cerr << " ";
std::cerr << "VReg " << *getValue() << "\n";
}
void ConstantNode::dumpNode(int indent) const {
for (int i=0; i < indent; i++)
std::cerr << " ";
std::cerr << "Constant " << *getValue() << "\n";
}
void LabelNode::dumpNode(int indent) const {
for (int i=0; i < indent; i++)
std::cerr << " ";
std::cerr << "Label " << *getValue() << "\n";
}
/// InstrForest ctor - Create a forest of instruction trees for a
/// single function.
///
InstrForest::InstrForest(Function *F) {
for (Function::iterator BB = F->begin(), FE = F->end(); BB != FE; ++BB) {
for(BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
buildTreeForInstruction(I);
}
}
InstrForest::~InstrForest() {
for_each(treeRoots.begin(), treeRoots.end(), deleter<InstructionNode>);
}
void InstrForest::dump() const {
for (const_root_iterator I = roots_begin(); I != roots_end(); ++I)
(*I)->dump(/*dumpChildren*/ 1, /*indent*/ 0);
}
inline void InstrForest::eraseRoot(InstructionNode* node) {
for (RootSet::reverse_iterator RI=treeRoots.rbegin(), RE=treeRoots.rend();
RI != RE; ++RI)
if (*RI == node)
treeRoots.erase(RI.base()-1);
}
inline void InstrForest::noteTreeNodeForInstr(Instruction *instr,
InstructionNode *treeNode) {
(*this)[instr] = treeNode;
treeRoots.push_back(treeNode); // mark node as root of a new tree
}
inline void InstrForest::setLeftChild(InstrTreeNode *parent,
InstrTreeNode *child) {
parent->LeftChild = child;
child->Parent = parent;
if (InstructionNode* instrNode = dyn_cast<InstructionNode>(child))
eraseRoot(instrNode); // no longer a tree root
}
inline void InstrForest::setRightChild(InstrTreeNode *parent,
InstrTreeNode *child) {
parent->RightChild = child;
child->Parent = parent;
if (InstructionNode* instrNode = dyn_cast<InstructionNode>(child))
eraseRoot(instrNode); // no longer a tree root
}
InstructionNode* InstrForest::buildTreeForInstruction(Instruction *instr) {
InstructionNode *treeNode = getTreeNodeForInstr(instr);
if (treeNode) {
// treeNode has already been constructed for this instruction
assert(treeNode->getInstruction() == instr);
return treeNode;
}
// Otherwise, create a new tree node for this instruction.
treeNode = new InstructionNode(instr);
noteTreeNodeForInstr(instr, treeNode);
if (instr->getOpcode() == Instruction::Call) {
// Operands of call instruction
return treeNode;
}
// If the instruction has more than 2 instruction operands,
// then we need to create artificial list nodes to hold them.
// (Note that we only count operands that get tree nodes, and not
// others such as branch labels for a branch or switch instruction.)
// To do this efficiently, we'll walk all operands, build treeNodes
// for all appropriate operands and save them in an array. We then
// insert children at the end, creating list nodes where needed.
// As a performance optimization, allocate a child array only
// if a fixed array is too small.
int numChildren = 0;
InstrTreeNode** childArray = new InstrTreeNode*[instr->getNumOperands()];
// Walk the operands of the instruction
for (Instruction::op_iterator O = instr->op_begin(); O!=instr->op_end();
++O) {
Value* operand = *O;
// Check if the operand is a data value, not an branch label, type,
// method or module. If the operand is an address type (i.e., label
// or method) that is used in an non-branching operation, e.g., `add'.
// that should be considered a data value.
// Check latter condition here just to simplify the next IF.
bool includeAddressOperand =
(isa<BasicBlock>(operand) || isa<Function>(operand))
&& !instr->isTerminator();
if (includeAddressOperand || isa<Instruction>(operand) ||
isa<Constant>(operand) || isa<Argument>(operand)) {
// This operand is a data value.
// An instruction that computes the incoming value is added as a
// child of the current instruction if:
// the value has only a single use
// AND both instructions are in the same basic block.
// AND the current instruction is not a PHI (because the incoming
// value is conceptually in a predecessor block,
// even though it may be in the same static block)
// (Note that if the value has only a single use (viz., `instr'),
// the def of the value can be safely moved just before instr
// and therefore it is safe to combine these two instructions.)
// In all other cases, the virtual register holding the value
// is used directly, i.e., made a child of the instruction node.
InstrTreeNode* opTreeNode;
if (isa<Instruction>(operand) && operand->hasOneUse() &&
cast<Instruction>(operand)->getParent() == instr->getParent() &&
instr->getOpcode() != Instruction::PHI &&
instr->getOpcode() != Instruction::Call) {
// Recursively create a treeNode for it.
opTreeNode = buildTreeForInstruction((Instruction*)operand);
} else if (Constant *CPV = dyn_cast<Constant>(operand)) {
if (isa<GlobalValue>(CPV))
opTreeNode = new VRegNode(operand);
else if (isa<UndefValue>(CPV)) {
opTreeNode = new
ConstantNode(Constant::getNullValue(CPV->getType()));
} else {
// Create a leaf node for a constant
opTreeNode = new ConstantNode(CPV);
}
} else {
// Create a leaf node for the virtual register
opTreeNode = new VRegNode(operand);
}
childArray[numChildren++] = opTreeNode;
}
}
// Add any selected operands as children in the tree.
// Certain instructions can have more than 2 in some instances (viz.,
// a CALL or a memory access -- LOAD, STORE, and GetElemPtr -- to an
// array or struct). Make the operands of every such instruction into
// a right-leaning binary tree with the operand nodes at the leaves
// and VRegList nodes as internal nodes.
InstrTreeNode *parent = treeNode;
if (numChildren > 2) {
unsigned instrOpcode = treeNode->getInstruction()->getOpcode();
assert(instrOpcode == Instruction::PHI ||
instrOpcode == Instruction::Call ||
instrOpcode == Instruction::Load ||
instrOpcode == Instruction::Store ||
instrOpcode == Instruction::GetElementPtr);
}
// Insert the first child as a direct child
if (numChildren >= 1)
setLeftChild(parent, childArray[0]);
int n;
// Create a list node for children 2 .. N-1, if any
for (n = numChildren-1; n >= 2; n--) {
// We have more than two children
InstrTreeNode *listNode = new VRegListNode();
setRightChild(parent, listNode);
setLeftChild(listNode, childArray[numChildren - n]);
parent = listNode;
}
// Now insert the last remaining child (if any).
if (numChildren >= 2) {
assert(n == 1);
setRightChild(parent, childArray[numChildren - 1]);
}
delete [] childArray;
return treeNode;
}
//==------------------------------------------------------------------------==//
// V9ISel Command-line options and declarations
//==------------------------------------------------------------------------==//
namespace {
/// Allow the user to select the amount of debugging information printed
/// out by V9ISel.
///
enum SelectDebugLevel_t {
Select_NoDebugInfo,
Select_PrintMachineCode,
Select_DebugInstTrees,
Select_DebugBurgTrees,
};
cl::opt<SelectDebugLevel_t>
SelectDebugLevel("dselect", cl::Hidden,
cl::desc("enable instruction selection debug information"),
cl::values(
clEnumValN(Select_NoDebugInfo, "n", "disable debug output"),
clEnumValN(Select_PrintMachineCode, "y", "print generated machine code"),
clEnumValN(Select_DebugInstTrees, "i",
"print debugging info for instruction selection"),
clEnumValN(Select_DebugBurgTrees, "b", "print burg trees"),
clEnumValEnd));
/// V9ISel - This is the FunctionPass that drives the instruction selection
/// process on the SparcV9 target.
///
class V9ISel : public FunctionPass {
TargetMachine &Target;
void InsertCodeForPhis(Function &F);
void InsertPhiElimInstructions(BasicBlock *BB,
const std::vector<MachineInstr*>& CpVec);
void SelectInstructionsForTree(InstrTreeNode* treeRoot, int goalnt);
void PostprocessMachineCodeForTree(InstructionNode* instrNode,
int ruleForNode, short* nts);
public:
V9ISel(TargetMachine &TM) : Target(TM) {}
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
AU.setPreservesCFG();
}
bool runOnFunction(Function &F);
virtual const char *getPassName() const {
return "SparcV9 BURG Instruction Selector";
}
};
}
//==------------------------------------------------------------------------==//
// Various V9ISel helper functions
//==------------------------------------------------------------------------==//
static const uint32_t MAXLO = (1 << 10) - 1; // set bits set by %lo(*)
static const uint32_t MAXSIMM = (1 << 12) - 1; // set bits in simm13 field of OR
/// ConvertConstantToIntType - Function to get the value of an integral
/// constant in the form that must be put into the machine register. The
/// specified constant is interpreted as (i.e., converted if necessary to) the
/// specified destination type. The result is always returned as an uint64_t,
/// since the representation of int64_t and uint64_t are identical. The
/// argument can be any known const. isValidConstant is set to true if a valid
/// constant was found.
///
uint64_t ConvertConstantToIntType(const TargetMachine &target, const Value *V,
const Type *destType, bool &isValidConstant) {
isValidConstant = false;
uint64_t C = 0;
if (! destType->isIntegral() && ! isa<PointerType>(destType))
return C;
if (! isa<Constant>(V) || isa<GlobalValue>(V))
return C;
// GlobalValue: no conversions needed: get value and return it
if (const GlobalValue* GV = dyn_cast<GlobalValue>(V)) {
isValidConstant = true; // may be overwritten by recursive call
return ConvertConstantToIntType(target, GV, destType, isValidConstant);
}
// ConstantBool: no conversions needed: get value and return it
if (const ConstantBool *CB = dyn_cast<ConstantBool>(V)) {
isValidConstant = true;
return (uint64_t) CB->getValue();
}
// ConstantPointerNull: it's really just a big, shiny version of zero.
if (isa<ConstantPointerNull>(V)) {
isValidConstant = true;
return 0;
}
// For other types of constants, some conversion may be needed.
// First, extract the constant operand according to its own type
if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
switch(CE->getOpcode()) {
case Instruction::Cast: // recursively get the value as cast
C = ConvertConstantToIntType(target, CE->getOperand(0), CE->getType(),
isValidConstant);
break;
default: // not simplifying other ConstantExprs
break;
}
else if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
isValidConstant = true;
C = CI->getRawValue();
} else if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
isValidConstant = true;
double fC = CFP->getValue();
C = (destType->isSigned()? (uint64_t) (int64_t) fC
: (uint64_t) fC);
} else if (isa<UndefValue>(V)) {
isValidConstant = true;
C = 0;
}
// Now if a valid value was found, convert it to destType.
if (isValidConstant) {
unsigned opSize = target.getTargetData().getTypeSize(V->getType());
unsigned destSize = target.getTargetData().getTypeSize(destType);
uint64_t maskHi = (destSize < 8)? (1U << 8*destSize) - 1 : ~0;
assert(opSize <= 8 && destSize <= 8 && ">8-byte int type unexpected");
if (destType->isSigned()) {
if (opSize > destSize) // operand is larger than dest:
C = C & maskHi; // mask high bits
if (opSize > destSize ||
(opSize == destSize && ! V->getType()->isSigned()))
if (C & (1U << (8*destSize - 1)))
C = C | ~maskHi; // sign-extend from destSize to 64 bits
}
else {
if (opSize > destSize || (V->getType()->isSigned() && destSize < 8)) {
// operand is larger than dest,
// OR both are equal but smaller than the full register size
// AND operand is signed, so it may have extra sign bits:
// mask high bits
C = C & maskHi;
}
}
}
return C;
}
/// CreateSETUWConst - Copy a 32-bit unsigned constant into the register
/// `dest', using SETHI, OR in the worst case. This function correctly emulates
/// the SETUW pseudo-op for SPARC v9 (if argument isSigned == false). The
/// isSigned=true case is used to implement SETSW without duplicating code. It
/// optimizes some common cases:
/// (1) Small value that fits in simm13 field of OR: don't need SETHI.
/// (2) isSigned = true and C is a small negative signed value, i.e.,
/// high bits are 1, and the remaining bits fit in simm13(OR).
static inline void
CreateSETUWConst(uint32_t C,
Instruction* dest, std::vector<MachineInstr*>& mvec,
bool isSigned = false) {
MachineInstr *miSETHI = NULL, *miOR = NULL;
// In order to get efficient code, we should not generate the SETHI if
// all high bits are 1 (i.e., this is a small signed value that fits in
// the simm13 field of OR). So we check for and handle that case specially.
// NOTE: The value C = 0x80000000 is bad: sC < 0 *and* -sC < 0.
// In fact, sC == -sC, so we have to check for this explicitly.
int32_t sC = (int32_t) C;
bool smallNegValue =isSigned && sC < 0 && sC != -sC && -sC < (int32_t)MAXSIMM;
// Set the high 22 bits in dest if non-zero and simm13 field of OR not enough
if (!smallNegValue && (C & ~MAXLO) && C > MAXSIMM) {
miSETHI = BuildMI(V9::SETHI, 2).addZImm(C).addRegDef(dest);
miSETHI->getOperand(0).markHi32();
mvec.push_back(miSETHI);
}
// Set the low 10 or 12 bits in dest. This is necessary if no SETHI
// was generated, or if the low 10 bits are non-zero.
if (miSETHI==NULL || C & MAXLO) {
if (miSETHI) {
// unsigned value with high-order bits set using SETHI
miOR = BuildMI(V9::ORi,3).addReg(dest).addZImm(C).addRegDef(dest);
miOR->getOperand(1).markLo32();
} else {
// unsigned or small signed value that fits in simm13 field of OR
assert(smallNegValue || (C & ~MAXSIMM) == 0);
miOR = BuildMI(V9::ORi, 3).addMReg(SparcV9::g0)
.addSImm(sC).addRegDef(dest);
}
mvec.push_back(miOR);
}
assert((miSETHI || miOR) && "Oops, no code was generated!");
}
/// CreateSETSWConst - Set a 32-bit signed constant in the register `dest',
/// with sign-extension to 64 bits. This uses SETHI, OR, SRA in the worst case.
/// This function correctly emulates the SETSW pseudo-op for SPARC v9. It
/// optimizes the same cases as SETUWConst, plus:
/// (1) SRA is not needed for positive or small negative values.
///
static inline void
CreateSETSWConst(int32_t C,
Instruction* dest, std::vector<MachineInstr*>& mvec) {
// Set the low 32 bits of dest
CreateSETUWConst((uint32_t) C, dest, mvec, /*isSigned*/true);
// Sign-extend to the high 32 bits if needed.
// NOTE: The value C = 0x80000000 is bad: -C == C and so -C is < MAXSIMM
if (C < 0 && (C == -C || -C > (int32_t) MAXSIMM))
mvec.push_back(BuildMI(V9::SRAi5,3).addReg(dest).addZImm(0).addRegDef(dest));
}
/// CreateSETXConst - Set a 64-bit signed or unsigned constant in the
/// register `dest'. Use SETUWConst for each 32 bit word, plus a
/// left-shift-by-32 in between. This function correctly emulates the SETX
/// pseudo-op for SPARC v9. It optimizes the same cases as SETUWConst for each
/// 32 bit word.
///
static inline void
CreateSETXConst(uint64_t C,
Instruction* tmpReg, Instruction* dest,
std::vector<MachineInstr*>& mvec) {
assert(C > (unsigned int) ~0 && "Use SETUW/SETSW for 32-bit values!");
MachineInstr* MI;
// Code to set the upper 32 bits of the value in register `tmpReg'
CreateSETUWConst((C >> 32), tmpReg, mvec);
// Shift tmpReg left by 32 bits
mvec.push_back(BuildMI(V9::SLLXi6, 3).addReg(tmpReg).addZImm(32)
.addRegDef(tmpReg));
// Code to set the low 32 bits of the value in register `dest'
CreateSETUWConst(C, dest, mvec);
// dest = OR(tmpReg, dest)
mvec.push_back(BuildMI(V9::ORr,3).addReg(dest).addReg(tmpReg).addRegDef(dest));
}
/// CreateSETUWLabel - Set a 32-bit constant (given by a symbolic label) in
/// the register `dest'.
///
static inline void
CreateSETUWLabel(Value* val,
Instruction* dest, std::vector<MachineInstr*>& mvec) {
MachineInstr* MI;
// Set the high 22 bits in dest
MI = BuildMI(V9::SETHI, 2).addReg(val).addRegDef(dest);
MI->getOperand(0).markHi32();
mvec.push_back(MI);
// Set the low 10 bits in dest
MI = BuildMI(V9::ORr, 3).addReg(dest).addReg(val).addRegDef(dest);
MI->getOperand(1).markLo32();
mvec.push_back(MI);
}
/// CreateSETXLabel - Set a 64-bit constant (given by a symbolic label) in the
/// register `dest'.
///
static inline void
CreateSETXLabel(Value* val, Instruction* tmpReg,
Instruction* dest, std::vector<MachineInstr*>& mvec) {
assert(isa<Constant>(val) &&
"I only know about constant values and global addresses");
MachineInstr* MI;
MI = BuildMI(V9::SETHI, 2).addPCDisp(val).addRegDef(tmpReg);
MI->getOperand(0).markHi64();
mvec.push_back(MI);
MI = BuildMI(V9::ORi, 3).addReg(tmpReg).addPCDisp(val).addRegDef(tmpReg);
MI->getOperand(1).markLo64();
mvec.push_back(MI);
mvec.push_back(BuildMI(V9::SLLXi6, 3).addReg(tmpReg).addZImm(32)
.addRegDef(tmpReg));
MI = BuildMI(V9::SETHI, 2).addPCDisp(val).addRegDef(dest);
MI->getOperand(0).markHi32();
mvec.push_back(MI);
MI = BuildMI(V9::ORr, 3).addReg(dest).addReg(tmpReg).addRegDef(dest);
mvec.push_back(MI);
MI = BuildMI(V9::ORi, 3).addReg(dest).addPCDisp(val).addRegDef(dest);
MI->getOperand(1).markLo32();
mvec.push_back(MI);
}
/// CreateUIntSetInstruction - Create code to Set an unsigned constant in the
/// register `dest'. Uses CreateSETUWConst, CreateSETSWConst or CreateSETXConst
/// as needed. CreateSETSWConst is an optimization for the case that the
/// unsigned value has all ones in the 33 high bits (so that sign-extension sets
/// them all).
///
static inline void
CreateUIntSetInstruction(uint64_t C, Instruction* dest,
std::vector<MachineInstr*>& mvec,
MachineCodeForInstruction& mcfi) {
static const uint64_t lo32 = (uint32_t) ~0;
if (C <= lo32) // High 32 bits are 0. Set low 32 bits.
CreateSETUWConst((uint32_t) C, dest, mvec);
else if ((C & ~lo32) == ~lo32 && (C & (1U << 31))) {
// All high 33 (not 32) bits are 1s: sign-extension will take care
// of high 32 bits, so use the sequence for signed int
CreateSETSWConst((int32_t) C, dest, mvec);
} else if (C > lo32) {
// C does not fit in 32 bits
TmpInstruction* tmpReg = new TmpInstruction(mcfi, Type::IntTy);
CreateSETXConst(C, tmpReg, dest, mvec);
}
}
/// CreateIntSetInstruction - Create code to Set a signed constant in the
/// register `dest'. Really the same as CreateUIntSetInstruction.
///
static inline void
CreateIntSetInstruction(int64_t C, Instruction* dest,
std::vector<MachineInstr*>& mvec,
MachineCodeForInstruction& mcfi) {
CreateUIntSetInstruction((uint64_t) C, dest, mvec, mcfi);
}
/// MaxConstantsTableTy - Table mapping LLVM opcodes to the max. immediate
/// constant usable for that operation in the SparcV9 backend. Used by
/// ConstantMayNotFitInImmedField().
///
struct MaxConstantsTableTy {
// Entry == 0 ==> no immediate constant field exists at all.
// Entry > 0 ==> abs(immediate constant) <= Entry
std::vector<int> tbl;
int getMaxConstantForInstr (unsigned llvmOpCode);
MaxConstantsTableTy ();
unsigned size() const { return tbl.size (); }
int &operator[] (unsigned index) { return tbl[index]; }
};
int MaxConstantsTableTy::getMaxConstantForInstr(unsigned llvmOpCode) {
int modelOpCode = -1;
if (llvmOpCode >= Instruction::BinaryOpsBegin &&
llvmOpCode < Instruction::BinaryOpsEnd)
modelOpCode = V9::ADDi;
else
switch(llvmOpCode) {
case Instruction::Ret: modelOpCode = V9::JMPLCALLi; break;
case Instruction::Malloc:
case Instruction::Alloca:
case Instruction::GetElementPtr:
case Instruction::PHI:
case Instruction::Cast:
case Instruction::Call: modelOpCode = V9::ADDi; break;
case Instruction::Shl:
case Instruction::Shr: modelOpCode = V9::SLLXi6; break;
default: break;
};
return (modelOpCode < 0)? 0: SparcV9MachineInstrDesc[modelOpCode].maxImmedConst;
}
MaxConstantsTableTy::MaxConstantsTableTy () : tbl (Instruction::OtherOpsEnd) {
unsigned op;
assert(tbl.size() == Instruction::OtherOpsEnd &&
"assignments below will be illegal!");
for (op = Instruction::TermOpsBegin; op < Instruction::TermOpsEnd; ++op)
tbl[op] = getMaxConstantForInstr(op);
for (op = Instruction::BinaryOpsBegin; op < Instruction::BinaryOpsEnd; ++op)
tbl[op] = getMaxConstantForInstr(op);
for (op = Instruction::MemoryOpsBegin; op < Instruction::MemoryOpsEnd; ++op)
tbl[op] = getMaxConstantForInstr(op);
for (op = Instruction::OtherOpsBegin; op < Instruction::OtherOpsEnd; ++op)
tbl[op] = getMaxConstantForInstr(op);
}
bool ConstantMayNotFitInImmedField(const Constant* CV, const Instruction* I) {
// The one and only MaxConstantsTable, used only by this function.
static MaxConstantsTableTy MaxConstantsTable;
if (I->getOpcode() >= MaxConstantsTable.size()) // user-defined op (or bug!)
return true;
// can always use %g0
if (isa<ConstantPointerNull>(CV) || isa<UndefValue>(CV))
return false;
if (isa<SwitchInst>(I)) // Switch instructions will be lowered!
return false;
if (const ConstantInt* CI = dyn_cast<ConstantInt>(CV))
return labs((int64_t)CI->getRawValue()) > MaxConstantsTable[I->getOpcode()];
if (isa<ConstantBool>(CV))
return 1 > MaxConstantsTable[I->getOpcode()];
return true;
}
/// ChooseLoadInstruction - Return the appropriate load instruction opcode
/// based on the given LLVM value type.
///
static inline MachineOpCode ChooseLoadInstruction(const Type *DestTy) {
switch (DestTy->getTypeID()) {
case Type::BoolTyID:
case Type::UByteTyID: return V9::LDUBr;
case Type::SByteTyID: return V9::LDSBr;
case Type::UShortTyID: return V9::LDUHr;
case Type::ShortTyID: return V9::LDSHr;
case Type::UIntTyID: return V9::LDUWr;
case Type::IntTyID: return V9::LDSWr;
case Type::PointerTyID:
case Type::ULongTyID:
case Type::LongTyID: return V9::LDXr;
case Type::FloatTyID: return V9::LDFr;
case Type::DoubleTyID: return V9::LDDFr;
default: assert(0 && "Invalid type for Load instruction");
}
return 0;
}
/// ChooseStoreInstruction - Return the appropriate store instruction opcode
/// based on the given LLVM value type.
///
static inline MachineOpCode ChooseStoreInstruction(const Type *DestTy) {
switch (DestTy->getTypeID()) {
case Type::BoolTyID:
case Type::UByteTyID:
case Type::SByteTyID: return V9::STBr;
case Type::UShortTyID:
case Type::ShortTyID: return V9::STHr;
case Type::UIntTyID:
case Type::IntTyID: return V9::STWr;
case Type::PointerTyID:
case Type::ULongTyID:
case Type::LongTyID: return V9::STXr;
case Type::FloatTyID: return V9::STFr;
case Type::DoubleTyID: return V9::STDFr;
default: assert(0 && "Invalid type for Store instruction");
}
return 0;
}
static inline MachineOpCode ChooseAddInstructionByType(const Type* resultType) {
MachineOpCode opCode = V9::INVALID_OPCODE;
if (resultType->isIntegral() || isa<PointerType>(resultType)
|| isa<FunctionType>(resultType) || resultType == Type::LabelTy) {
opCode = V9::ADDr;
} else
switch(resultType->getTypeID()) {
case Type::FloatTyID: opCode = V9::FADDS; break;
case Type::DoubleTyID: opCode = V9::FADDD; break;
default: assert(0 && "Invalid type for ADD instruction"); break;
}
return opCode;
}
/// convertOpcodeFromRegToImm - Because the SparcV9 instruction selector likes
/// to re-write operands to instructions, making them change from a Value*
/// (virtual register) to a Constant* (making an immediate field), we need to
/// change the opcode from a register-based instruction to an immediate-based
/// instruction, hence this mapping.
///
static unsigned convertOpcodeFromRegToImm(unsigned Opcode) {
switch (Opcode) {
/* arithmetic */
case V9::ADDr: return V9::ADDi;
case V9::ADDccr: return V9::ADDcci;
case V9::ADDCr: return V9::ADDCi;
case V9::ADDCccr: return V9::ADDCcci;
case V9::SUBr: return V9::SUBi;
case V9::SUBccr: return V9::SUBcci;
case V9::SUBCr: return V9::SUBCi;
case V9::SUBCccr: return V9::SUBCcci;
case V9::MULXr: return V9::MULXi;
case V9::SDIVXr: return V9::SDIVXi;
case V9::UDIVXr: return V9::UDIVXi;
/* logical */
case V9::ANDr: return V9::ANDi;
case V9::ANDccr: return V9::ANDcci;
case V9::ANDNr: return V9::ANDNi;
case V9::ANDNccr: return V9::ANDNcci;
case V9::ORr: return V9::ORi;
case V9::ORccr: return V9::ORcci;
case V9::ORNr: return V9::ORNi;
case V9::ORNccr: return V9::ORNcci;
case V9::XORr: return V9::XORi;
case V9::XORccr: return V9::XORcci;
case V9::XNORr: return V9::XNORi;
case V9::XNORccr: return V9::XNORcci;
/* shift */
case V9::SLLr5: return V9::SLLi5;
case V9::SRLr5: return V9::SRLi5;
case V9::SRAr5: return V9::SRAi5;
case V9::SLLXr6: return V9::SLLXi6;
case V9::SRLXr6: return V9::SRLXi6;
case V9::SRAXr6: return V9::SRAXi6;
/* Conditional move on int comparison with zero */
case V9::MOVRZr: return V9::MOVRZi;
case V9::MOVRLEZr: return V9::MOVRLEZi;
case V9::MOVRLZr: return V9::MOVRLZi;
case V9::MOVRNZr: return V9::MOVRNZi;
case V9::MOVRGZr: return V9::MOVRGZi;
case V9::MOVRGEZr: return V9::MOVRGEZi;
/* Conditional move on int condition code */
case V9::MOVAr: return V9::MOVAi;
case V9::MOVNr: return V9::MOVNi;
case V9::MOVNEr: return V9::MOVNEi;
case V9::MOVEr: return V9::MOVEi;
case V9::MOVGr: return V9::MOVGi;
case V9::MOVLEr: return V9::MOVLEi;
case V9::MOVGEr: return V9::MOVGEi;
case V9::MOVLr: return V9::MOVLi;
case V9::MOVGUr: return V9::MOVGUi;
case V9::MOVLEUr: return V9::MOVLEUi;
case V9::MOVCCr: return V9::MOVCCi;
case V9::MOVCSr: return V9::MOVCSi;
case V9::MOVPOSr: return V9::MOVPOSi;
case V9::MOVNEGr: return V9::MOVNEGi;
case V9::MOVVCr: return V9::MOVVCi;
case V9::MOVVSr: return V9::MOVVSi;
/* Conditional move of int reg on fp condition code */
case V9::MOVFAr: return V9::MOVFAi;
case V9::MOVFNr: return V9::MOVFNi;
case V9::MOVFUr: return V9::MOVFUi;
case V9::MOVFGr: return V9::MOVFGi;
case V9::MOVFUGr: return V9::MOVFUGi;
case V9::MOVFLr: return V9::MOVFLi;
case V9::MOVFULr: return V9::MOVFULi;
case V9::MOVFLGr: return V9::MOVFLGi;
case V9::MOVFNEr: return V9::MOVFNEi;
case V9::MOVFEr: return V9::MOVFEi;
case V9::MOVFUEr: return V9::MOVFUEi;
case V9::MOVFGEr: return V9::MOVFGEi;
case V9::MOVFUGEr: return V9::MOVFUGEi;
case V9::MOVFLEr: return V9::MOVFLEi;
case V9::MOVFULEr: return V9::MOVFULEi;
case V9::MOVFOr: return V9::MOVFOi;
/* load */
case V9::LDSBr: return V9::LDSBi;
case V9::LDSHr: return V9::LDSHi;
case V9::LDSWr: return V9::LDSWi;
case V9::LDUBr: return V9::LDUBi;
case V9::LDUHr: return V9::LDUHi;
case V9::LDUWr: return V9::LDUWi;
case V9::LDXr: return V9::LDXi;
case V9::LDFr: return V9::LDFi;
case V9::LDDFr: return V9::LDDFi;
case V9::LDQFr: return V9::LDQFi;
case V9::LDFSRr: return V9::LDFSRi;
case V9::LDXFSRr: return V9::LDXFSRi;
/* store */
case V9::STBr: return V9::STBi;
case V9::STHr: return V9::STHi;
case V9::STWr: return V9::STWi;
case V9::STXr: return V9::STXi;
case V9::STFr: return V9::STFi;
case V9::STDFr: return V9::STDFi;
case V9::STFSRr: return V9::STFSRi;
case V9::STXFSRr: return V9::STXFSRi;
/* jump & return */
case V9::JMPLCALLr: return V9::JMPLCALLi;
case V9::JMPLRETr: return V9::JMPLRETi;
/* save and restore */
case V9::SAVEr: return V9::SAVEi;
case V9::RESTOREr: return V9::RESTOREi;
default:
// It's already in correct format
// Or, it's just not handled yet, but an assert() would break LLC
#if 0
std::cerr << "Unhandled opcode in convertOpcodeFromRegToImm(): " << Opcode
<< "\n";
#endif
return Opcode;
}
}
/// CreateCodeToLoadConst - Create an instruction sequence to put the
/// constant `val' into the virtual register `dest'. `val' may be a Constant or
/// a GlobalValue, viz., the constant address of a global variable or function.
/// The generated instructions are returned in `mvec'. Any temp. registers
/// (TmpInstruction) created are recorded in mcfi. Any stack space required is
/// allocated via MachineFunction.
///
void CreateCodeToLoadConst(const TargetMachine& target, Function* F,
Value* val, Instruction* dest,
std::vector<MachineInstr*>& mvec,
MachineCodeForInstruction& mcfi) {
assert(isa<Constant>(val) &&
"I only know about constant values and global addresses");
// Use a "set" instruction for known constants or symbolic constants (labels)
// that can go in an integer reg.
// We have to use a "load" instruction for all other constants,
// in particular, floating point constants.
const Type* valType = val->getType();
if (isa<GlobalValue>(val)) {
TmpInstruction* tmpReg =
new TmpInstruction(mcfi, PointerType::get(val->getType()), val);
CreateSETXLabel(val, tmpReg, dest, mvec);
return;
}
bool isValid;
uint64_t C = ConvertConstantToIntType(target, val, dest->getType(), isValid);
if (isValid) {
if (dest->getType()->isSigned())
CreateUIntSetInstruction(C, dest, mvec, mcfi);
else
CreateIntSetInstruction((int64_t) C, dest, mvec, mcfi);
} else {
// Make an instruction sequence to load the constant, viz:
// SETX <addr-of-constant>, tmpReg, addrReg
// LOAD /*addr*/ addrReg, /*offset*/ 0, dest
// First, create a tmp register to be used by the SETX sequence.
TmpInstruction* tmpReg =
new TmpInstruction(mcfi, PointerType::get(val->getType()));
// Create another TmpInstruction for the address register
TmpInstruction* addrReg =
new TmpInstruction(mcfi, PointerType::get(val->getType()));
// Get the constant pool index for this constant
MachineConstantPool *CP = MachineFunction::get(F).getConstantPool();
Constant *C = cast<Constant>(val);
unsigned CPI = CP->getConstantPoolIndex(C);
// Put the address of the constant into a register
MachineInstr* MI;
MI = BuildMI(V9::SETHI, 2).addConstantPoolIndex(CPI).addRegDef(tmpReg);
MI->getOperand(0).markHi64();
mvec.push_back(MI);
//Create another tmp register for the SETX sequence to preserve SSA
TmpInstruction* tmpReg2 =
new TmpInstruction(mcfi, PointerType::get(val->getType()));
MI = BuildMI(V9::ORi, 3).addReg(tmpReg).addConstantPoolIndex(CPI)
.addRegDef(tmpReg2);
MI->getOperand(1).markLo64();
mvec.push_back(MI);
//Create another tmp register for the SETX sequence to preserve SSA
TmpInstruction* tmpReg3 =
new TmpInstruction(mcfi, PointerType::get(val->getType()));
mvec.push_back(BuildMI(V9::SLLXi6, 3).addReg(tmpReg2).addZImm(32)
.addRegDef(tmpReg3));
MI = BuildMI(V9::SETHI, 2).addConstantPoolIndex(CPI).addRegDef(addrReg);
MI->getOperand(0).markHi32();
mvec.push_back(MI);
// Create another TmpInstruction for the address register
TmpInstruction* addrReg2 =
new TmpInstruction(mcfi, PointerType::get(val->getType()));
MI = BuildMI(V9::ORr, 3).addReg(addrReg).addReg(tmpReg3).addRegDef(addrReg2);
mvec.push_back(MI);
// Create another TmpInstruction for the address register
TmpInstruction* addrReg3 =
new TmpInstruction(mcfi, PointerType::get(val->getType()));
MI = BuildMI(V9::ORi, 3).addReg(addrReg2).addConstantPoolIndex(CPI)
.addRegDef(addrReg3);
MI->getOperand(1).markLo32();
mvec.push_back(MI);
// Now load the constant from out ConstantPool label
unsigned Opcode = ChooseLoadInstruction(val->getType());
Opcode = convertOpcodeFromRegToImm(Opcode);
mvec.push_back(BuildMI(Opcode, 3)
.addReg(addrReg3).addSImm((int64_t)0).addRegDef(dest));
}
}
/// CreateCodeToCopyFloatToInt - Similarly, create an instruction sequence
/// to copy an FP register `val' to an integer register `dest' by copying to
/// memory and back. The generated instructions are returned in `mvec'. Any
/// temp. virtual registers (TmpInstruction) created are recorded in mcfi.
/// Temporary stack space required is allocated via MachineFunction.
///
void CreateCodeToCopyFloatToInt(const TargetMachine& target, Function* F,
Value* val, Instruction* dest,
std::vector<MachineInstr*>& mvec,
MachineCodeForInstruction& mcfi) {
const Type* opTy = val->getType();
const Type* destTy = dest->getType();
assert(opTy->isFloatingPoint() && "Source type must be float/double");
assert((destTy->isIntegral() || isa<PointerType>(destTy))
&& "Dest type must be integer, bool or pointer");
// FIXME: For now, we allocate permanent space because the stack frame
// manager does not allow locals to be allocated (e.g., for alloca) after
// a temp is allocated!
int offset = MachineFunction::get(F).getInfo<SparcV9FunctionInfo>()->allocateLocalVar(val);
unsigned FPReg = target.getRegInfo()->getFramePointer();
// Store instruction stores `val' to [%fp+offset].
// The store opCode is based only the source value being copied.
unsigned StoreOpcode = ChooseStoreInstruction(opTy);
StoreOpcode = convertOpcodeFromRegToImm(StoreOpcode);
mvec.push_back(BuildMI(StoreOpcode, 3)
.addReg(val).addMReg(FPReg).addSImm(offset));
// Load instruction loads [%fp+offset] to `dest'.
// The type of the load opCode is the integer type that matches the
// source type in size:
// On SparcV9: int for float, long for double.
// Note that we *must* use signed loads even for unsigned dest types, to
// ensure correct sign-extension for UByte, UShort or UInt:
const Type* loadTy = (opTy == Type::FloatTy)? Type::IntTy : Type::LongTy;
unsigned LoadOpcode = ChooseLoadInstruction(loadTy);
LoadOpcode = convertOpcodeFromRegToImm(LoadOpcode);
mvec.push_back(BuildMI(LoadOpcode, 3).addMReg(FPReg)
.addSImm(offset).addRegDef(dest));
}
/// CreateBitExtensionInstructions - Helper function for sign-extension and
/// zero-extension. For SPARC v9, we sign-extend the given operand using SLL;
/// SRA/SRL.
///
inline void
CreateBitExtensionInstructions(bool signExtend, const TargetMachine& target,
Function* F, Value* srcVal, Value* destVal,
unsigned int numLowBits,
std::vector<MachineInstr*>& mvec,
MachineCodeForInstruction& mcfi) {
MachineInstr* M;
assert(numLowBits <= 32 && "Otherwise, nothing should be done here!");
if (numLowBits < 32) {
// SLL is needed since operand size is < 32 bits.
TmpInstruction *tmpI = new TmpInstruction(mcfi, destVal->getType(),
srcVal, destVal, "make32");
mvec.push_back(BuildMI(V9::SLLXi6, 3).addReg(srcVal)
.addZImm(32-numLowBits).addRegDef(tmpI));
srcVal = tmpI;
}
mvec.push_back(BuildMI(signExtend? V9::SRAi5 : V9::SRLi5, 3)
.addReg(srcVal).addZImm(32-numLowBits).addRegDef(destVal));
}
/// CreateSignExtensionInstructions - Create instruction sequence to produce
/// a sign-extended register value from an arbitrary-sized integer value (sized
/// in bits, not bytes). The generated instructions are returned in `mvec'. Any
/// temp. registers (TmpInstruction) created are recorded in mcfi. Any stack
/// space required is allocated via MachineFunction.
///
void CreateSignExtensionInstructions(const TargetMachine& target,
Function* F, Value* srcVal, Value* destVal,
unsigned int numLowBits,
std::vector<MachineInstr*>& mvec,
MachineCodeForInstruction& mcfi) {
CreateBitExtensionInstructions(/*signExtend*/ true, target, F, srcVal,
destVal, numLowBits, mvec, mcfi);
}
/// CreateZeroExtensionInstructions - Create instruction sequence to produce
/// a zero-extended register value from an arbitrary-sized integer value (sized
/// in bits, not bytes). For SPARC v9, we sign-extend the given operand using
/// SLL; SRL. The generated instructions are returned in `mvec'. Any temp.
/// registers (TmpInstruction) created are recorded in mcfi. Any stack space
/// required is allocated via MachineFunction.
///
void CreateZeroExtensionInstructions(const TargetMachine& target,
Function* F, Value* srcVal, Value* destVal,
unsigned int numLowBits,
std::vector<MachineInstr*>& mvec,
MachineCodeForInstruction& mcfi) {
CreateBitExtensionInstructions(/*signExtend*/ false, target, F, srcVal,
destVal, numLowBits, mvec, mcfi);
}
/// CreateCodeToCopyIntToFloat - Create an instruction sequence to copy an
/// integer register `val' to a floating point register `dest' by copying to
/// memory and back. val must be an integral type. dest must be a Float or
/// Double. The generated instructions are returned in `mvec'. Any temp.
/// registers (TmpInstruction) created are recorded in mcfi. Any stack space
/// required is allocated via MachineFunction.
///
void CreateCodeToCopyIntToFloat(const TargetMachine& target,
Function* F, Value* val, Instruction* dest,
std::vector<MachineInstr*>& mvec,
MachineCodeForInstruction& mcfi) {
assert((val->getType()->isIntegral() || isa<PointerType>(val->getType()))
&& "Source type must be integral (integer or bool) or pointer");
assert(dest->getType()->isFloatingPoint()
&& "Dest type must be float/double");
// Get a stack slot to use for the copy
int offset = MachineFunction::get(F).getInfo<SparcV9FunctionInfo>()->allocateLocalVar(val);
// Get the size of the source value being copied.
size_t srcSize = target.getTargetData().getTypeSize(val->getType());
// Store instruction stores `val' to [%fp+offset].
// The store and load opCodes are based on the size of the source value.
// If the value is smaller than 32 bits, we must sign- or zero-extend it
// to 32 bits since the load-float will load 32 bits.
// Note that the store instruction is the same for signed and unsigned ints.
const Type* storeType = (srcSize <= 4)? Type::IntTy : Type::LongTy;
Value* storeVal = val;
if (srcSize < target.getTargetData().getTypeSize(Type::FloatTy)) {
// sign- or zero-extend respectively
storeVal = new TmpInstruction(mcfi, storeType, val);
if (val->getType()->isSigned())
CreateSignExtensionInstructions(target, F, val, storeVal, 8*srcSize,
mvec, mcfi);
else
CreateZeroExtensionInstructions(target, F, val, storeVal, 8*srcSize,
mvec, mcfi);
}
unsigned FPReg = target.getRegInfo()->getFramePointer();
unsigned StoreOpcode = ChooseStoreInstruction(storeType);
StoreOpcode = convertOpcodeFromRegToImm(StoreOpcode);
mvec.push_back(BuildMI(StoreOpcode, 3)
.addReg(storeVal).addMReg(FPReg).addSImm(offset));
// Load instruction loads [%fp+offset] to `dest'.
// The type of the load opCode is the floating point type that matches the
// stored type in size:
// On SparcV9: float for int or smaller, double for long.
const Type* loadType = (srcSize <= 4)? Type::FloatTy : Type::DoubleTy;
unsigned LoadOpcode = ChooseLoadInstruction(loadType);
LoadOpcode = convertOpcodeFromRegToImm(LoadOpcode);
mvec.push_back(BuildMI(LoadOpcode, 3)
.addMReg(FPReg).addSImm(offset).addRegDef(dest));
}
/// InsertCodeToLoadConstant - Generates code to load the constant
/// into a TmpInstruction (virtual reg) and returns the virtual register.
///
static TmpInstruction*
InsertCodeToLoadConstant(Function *F, Value* opValue, Instruction* vmInstr,
std::vector<MachineInstr*>& loadConstVec,
TargetMachine& target) {
// Create a tmp virtual register to hold the constant.
MachineCodeForInstruction &mcfi = MachineCodeForInstruction::get(vmInstr);
TmpInstruction* tmpReg = new TmpInstruction(mcfi, opValue);
CreateCodeToLoadConst(target, F, opValue, tmpReg, loadConstVec, mcfi);
// Record the mapping from the tmp VM instruction to machine instruction.
// Do this for all machine instructions that were not mapped to any
// other temp values created by
// tmpReg->addMachineInstruction(loadConstVec.back());
return tmpReg;
}
MachineOperand::MachineOperandType
ChooseRegOrImmed(int64_t intValue, bool isSigned,
MachineOpCode opCode, const TargetMachine& target,
bool canUseImmed, unsigned int& getMachineRegNum,
int64_t& getImmedValue) {
MachineOperand::MachineOperandType opType=MachineOperand::MO_VirtualRegister;
getMachineRegNum = 0;
getImmedValue = 0;
if (canUseImmed &&
target.getInstrInfo()->constantFitsInImmedField(opCode, intValue)) {
opType = isSigned? MachineOperand::MO_SignExtendedImmed
: MachineOperand::MO_UnextendedImmed;
getImmedValue = intValue;
} else if (intValue == 0 &&
target.getRegInfo()->getZeroRegNum() != (unsigned)-1) {
opType = MachineOperand::MO_MachineRegister;
getMachineRegNum = target.getRegInfo()->getZeroRegNum();
}
return opType;
}
MachineOperand::MachineOperandType
ChooseRegOrImmed(Value* val,
MachineOpCode opCode, const TargetMachine& target,
bool canUseImmed, unsigned int& getMachineRegNum,
int64_t& getImmedValue) {
getMachineRegNum = 0;
getImmedValue = 0;
// To use reg or immed, constant needs to be integer, bool, or a NULL pointer.
// ConvertConstantToIntType() does the right conversions.
bool isValidConstant;
uint64_t valueToUse =
ConvertConstantToIntType(target, val, val->getType(), isValidConstant);
if (! isValidConstant)
return MachineOperand::MO_VirtualRegister;
// Now check if the constant value fits in the IMMED field.
return ChooseRegOrImmed((int64_t) valueToUse, val->getType()->isSigned(),
opCode, target, canUseImmed,
getMachineRegNum, getImmedValue);
}
/// CreateCopyInstructionsByType - Create instruction(s) to copy src to dest,
/// for arbitrary types. The generated instructions are returned in `mvec'. Any
/// temp. registers (TmpInstruction) created are recorded in mcfi. Any stack
/// space required is allocated via MachineFunction.
///
void CreateCopyInstructionsByType(const TargetMachine& target,
Function *F, Value* src, Instruction* dest,
std::vector<MachineInstr*>& mvec,
MachineCodeForInstruction& mcfi) {
bool loadConstantToReg = false;
const Type* resultType = dest->getType();
MachineOpCode opCode = ChooseAddInstructionByType(resultType);
assert (opCode != V9::INVALID_OPCODE
&& "Unsupported result type in CreateCopyInstructionsByType()");
// If `src' is a constant that doesn't fit in the immed field or if it is
// a global variable (i.e., a constant address), generate a load
// instruction instead of an add.
if (isa<GlobalValue>(src))
loadConstantToReg = true;
else if (isa<Constant>(src)) {
unsigned int machineRegNum;
int64_t immedValue;
MachineOperand::MachineOperandType opType =
ChooseRegOrImmed(src, opCode, target, /*canUseImmed*/ true,
machineRegNum, immedValue);
if (opType == MachineOperand::MO_VirtualRegister)
loadConstantToReg = true;
}
if (loadConstantToReg) {
// `src' is constant and cannot fit in immed field for the ADD.
// Insert instructions to "load" the constant into a register.
CreateCodeToLoadConst(target, F, src, dest, mvec, mcfi);
} else {
// Create a reg-to-reg copy instruction for the given type:
// -- For FP values, create a FMOVS or FMOVD instruction
// -- For non-FP values, create an add-with-0 instruction (opCode as above)
// Make `src' the second operand, in case it is a small constant!
MachineInstr* MI;
if (resultType->isFloatingPoint())
MI = (BuildMI(resultType == Type::FloatTy? V9::FMOVS : V9::FMOVD, 2)
.addReg(src).addRegDef(dest));
else {
const Type* Ty =isa<PointerType>(resultType)? Type::ULongTy :resultType;
MI = (BuildMI(opCode, 3)
.addSImm((int64_t) 0).addReg(src).addRegDef(dest));
}
mvec.push_back(MI);
}
}
/// FixConstantOperandsForInstr - Make a machine instruction use its constant
/// operands more efficiently. If the constant is 0, then use the hardwired 0
/// register, if any. Else, if the constant fits in the IMMEDIATE field, then
/// use that field. Otherwise, else create instructions to put the constant
/// into a register, either directly or by loading explicitly from the constant
/// pool. In the first 2 cases, the operand of `minstr' is modified in place.
/// Returns a vector of machine instructions generated for operands that fall
/// under case 3; these must be inserted before `minstr'.
///
std::vector<MachineInstr*>
FixConstantOperandsForInstr(Instruction* vmInstr, MachineInstr* minstr,
TargetMachine& target) {
std::vector<MachineInstr*> MVec;
MachineOpCode opCode = minstr->getOpcode();
const TargetInstrInfo& instrInfo = *target.getInstrInfo();
int resultPos = instrInfo.get(opCode).resultPos;
int immedPos = instrInfo.getImmedConstantPos(opCode);
Function *F = vmInstr->getParent()->getParent();
for (unsigned op=0; op < minstr->getNumOperands(); op++) {
const MachineOperand& mop = minstr->getOperand(op);
// Skip the result position, preallocated machine registers, or operands
// that cannot be constants (CC regs or PC-relative displacements)
if (resultPos == (int)op ||
mop.getType() == MachineOperand::MO_MachineRegister ||
mop.getType() == MachineOperand::MO_CCRegister ||
mop.getType() == MachineOperand::MO_PCRelativeDisp)
continue;
bool constantThatMustBeLoaded = false;
unsigned int machineRegNum = 0;
int64_t immedValue = 0;
Value* opValue = NULL;
MachineOperand::MachineOperandType opType =
MachineOperand::MO_VirtualRegister;
// Operand may be a virtual register or a compile-time constant
if (mop.getType() == MachineOperand::MO_VirtualRegister) {
assert(mop.getVRegValue() != NULL);
opValue = mop.getVRegValue();
if (Constant *opConst = dyn_cast<Constant>(opValue))
if (!isa<GlobalValue>(opConst)) {
opType = ChooseRegOrImmed(opConst, opCode, target,
(immedPos == (int)op), machineRegNum,
immedValue);
if (opType == MachineOperand::MO_VirtualRegister)
constantThatMustBeLoaded = true;
}
} else {
// If the operand is from the constant pool, don't try to change it.
if (mop.getType() == MachineOperand::MO_ConstantPoolIndex) {
continue;
}
assert(mop.isImmediate());
bool isSigned = mop.getType() == MachineOperand::MO_SignExtendedImmed;
// Bit-selection flags indicate an instruction that is extracting
// bits from its operand so ignore this even if it is a big constant.
if (mop.isHiBits32() || mop.isLoBits32() ||
mop.isHiBits64() || mop.isLoBits64())
continue;
opType = ChooseRegOrImmed(mop.getImmedValue(), isSigned,
opCode, target, (immedPos == (int)op),
machineRegNum, immedValue);
if (opType == MachineOperand::MO_SignExtendedImmed ||
opType == MachineOperand::MO_UnextendedImmed) {
// The optype is an immediate value
// This means we need to change the opcode, e.g. ADDr -> ADDi
unsigned newOpcode = convertOpcodeFromRegToImm(opCode);
minstr->setOpcode(newOpcode);
}
if (opType == mop.getType())
continue; // no change: this is the most common case
if (opType == MachineOperand::MO_VirtualRegister) {
constantThatMustBeLoaded = true;
opValue = isSigned
? (Value*)ConstantSInt::get(Type::LongTy, immedValue)
: (Value*)ConstantUInt::get(Type::ULongTy,(uint64_t)immedValue);
}
}
if (opType == MachineOperand::MO_MachineRegister)
minstr->SetMachineOperandReg(op, machineRegNum);
else if (opType == MachineOperand::MO_SignExtendedImmed ||
opType == MachineOperand::MO_UnextendedImmed) {
minstr->SetMachineOperandConst(op, opType, immedValue);
// The optype is or has become an immediate
// This means we need to change the opcode, e.g. ADDr -> ADDi
unsigned newOpcode = convertOpcodeFromRegToImm(opCode);
minstr->setOpcode(newOpcode);
} else if (constantThatMustBeLoaded ||
(opValue && isa<GlobalValue>(opValue)))
{ // opValue is a constant that must be explicitly loaded into a reg
assert(opValue);
TmpInstruction* tmpReg = InsertCodeToLoadConstant(F, opValue, vmInstr,
MVec, target);
minstr->SetMachineOperandVal(op, MachineOperand::MO_VirtualRegister,
tmpReg);
}
}
// Also, check for implicit operands used by the machine instruction
// (no need to check those defined since they cannot be constants).
// These include:
// -- arguments to a Call
// -- return value of a Return
// Any such operand that is a constant value needs to be fixed also.
// The current instructions with implicit refs (viz., Call and Return)
// have no immediate fields, so the constant always needs to be loaded
// into a register.
bool isCall = instrInfo.isCall(opCode);
unsigned lastCallArgNum = 0; // unused if not a call
CallArgsDescriptor* argDesc = NULL; // unused if not a call
if (isCall)
argDesc = CallArgsDescriptor::get(minstr);
for (unsigned i=0, N=minstr->getNumImplicitRefs(); i < N; ++i)
if (isa<Constant>(minstr->getImplicitRef(i))) {
Value* oldVal = minstr->getImplicitRef(i);
TmpInstruction* tmpReg =
InsertCodeToLoadConstant(F, oldVal, vmInstr, MVec, target);
minstr->setImplicitRef(i, tmpReg);
if (isCall) {
// find and replace the argument in the CallArgsDescriptor
unsigned i=lastCallArgNum;
while (argDesc->getArgInfo(i).getArgVal() != oldVal)
++i;
assert(i < argDesc->getNumArgs() &&
"Constant operands to a call *must* be in the arg list");
lastCallArgNum = i;
argDesc->getArgInfo(i).replaceArgVal(tmpReg);
}
}
return MVec;
}
static inline void Add3OperandInstr(unsigned Opcode, InstructionNode* Node,
std::vector<MachineInstr*>& mvec) {
mvec.push_back(BuildMI(Opcode, 3).addReg(Node->leftChild()->getValue())
.addReg(Node->rightChild()->getValue())
.addRegDef(Node->getValue()));
}
/// IsZero - Check for a constant 0.
///
static inline bool IsZero(Value* idx) {
return (isa<Constant>(idx) && cast<Constant>(idx)->isNullValue()) ||
isa<UndefValue>(idx);
}
/// FoldGetElemChain - Fold a chain of GetElementPtr instructions containing
/// only constant offsets into an equivalent (Pointer, IndexVector) pair.
/// Returns the pointer Value, and stores the resulting IndexVector in argument
/// chainIdxVec. This is a helper function for FoldConstantIndices that does the
/// actual folding.
//
static Value*
FoldGetElemChain(InstrTreeNode* ptrNode, std::vector<Value*>& chainIdxVec,
bool lastInstHasLeadingNonZero) {
InstructionNode* gepNode = dyn_cast<InstructionNode>(ptrNode);
GetElementPtrInst* gepInst =
dyn_cast_or_null<GetElementPtrInst>(gepNode ? gepNode->getInstruction() :0);
// ptr value is not computed in this tree or ptr value does not come from GEP
// instruction
if (gepInst == NULL)
return NULL;
// Return NULL if we don't fold any instructions in.
Value* ptrVal = NULL;
// Now chase the chain of getElementInstr instructions, if any.
// Check for any non-constant indices and stop there.
// Also, stop if the first index of child is a non-zero array index
// and the last index of the current node is a non-array index:
// in that case, a non-array declared type is being accessed as an array
// which is not type-safe, but could be legal.
InstructionNode* ptrChild = gepNode;
while (ptrChild && (ptrChild->getOpLabel() == Instruction::GetElementPtr ||
ptrChild->getOpLabel() == GetElemPtrIdx)) {
// Child is a GetElemPtr instruction
gepInst = cast<GetElementPtrInst>(ptrChild->getValue());
User::op_iterator OI, firstIdx = gepInst->idx_begin();
User::op_iterator lastIdx = gepInst->idx_end();
bool allConstantOffsets = true;
// The first index of every GEP must be an array index.
assert((*firstIdx)->getType() == Type::LongTy &&
"INTERNAL ERROR: Structure index for a pointer type!");
// If the last instruction had a leading non-zero index, check if the
// current one references a sequential (i.e., indexable) type.
// If not, the code is not type-safe and we would create an illegal GEP
// by folding them, so don't fold any more instructions.
if (lastInstHasLeadingNonZero)
if (! isa<SequentialType>(gepInst->getType()->getElementType()))
break; // cannot fold in any preceding getElementPtr instrs.
// Check that all offsets are constant for this instruction
for (OI = firstIdx; allConstantOffsets && OI != lastIdx; ++OI)
allConstantOffsets = isa<ConstantInt>(*OI);
if (allConstantOffsets) {
// Get pointer value out of ptrChild.
ptrVal = gepInst->getPointerOperand();
// Insert its index vector at the start, skipping any leading [0]
// Remember the old size to check if anything was inserted.
unsigned oldSize = chainIdxVec.size();
int firstIsZero = IsZero(*firstIdx);
chainIdxVec.insert(chainIdxVec.begin(), firstIdx + firstIsZero, lastIdx);
// Remember if it has leading zero index: it will be discarded later.
if (oldSize < chainIdxVec.size())
lastInstHasLeadingNonZero = !firstIsZero;
// Mark the folded node so no code is generated for it.
((InstructionNode*) ptrChild)->markFoldedIntoParent();
// Get the previous GEP instruction and continue trying to fold
ptrChild = dyn_cast<InstructionNode>(ptrChild->leftChild());
} else // cannot fold this getElementPtr instr. or any preceding ones
break;
}
// If the first getElementPtr instruction had a leading [0], add it back.
// Note that this instruction is the *last* one that was successfully
// folded *and* contributed any indices, in the loop above.
if (ptrVal && ! lastInstHasLeadingNonZero)
chainIdxVec.insert(chainIdxVec.begin(), ConstantSInt::get(Type::LongTy,0));
return ptrVal;
}
/// GetGEPInstArgs - Helper function for GetMemInstArgs that handles the
/// final getElementPtr instruction used by (or same as) the memory operation.
/// Extracts the indices of the current instruction and tries to fold in
/// preceding ones if all indices of the current one are constant.
///
static Value *GetGEPInstArgs(InstructionNode *gepNode,
std::vector<Value *> &idxVec,
bool &allConstantIndices) {
allConstantIndices = true;
GetElementPtrInst* gepI = cast<GetElementPtrInst>(gepNode->getInstruction());
// Default pointer is the one from the current instruction.
Value* ptrVal = gepI->getPointerOperand();
InstrTreeNode* ptrChild = gepNode->leftChild();
// Extract the index vector of the GEP instruction.
// If all indices are constant and first index is zero, try to fold
// in preceding GEPs with all constant indices.
for (User::op_iterator OI=gepI->idx_begin(), OE=gepI->idx_end();
allConstantIndices && OI != OE; ++OI)
if (! isa<Constant>(*OI))
allConstantIndices = false; // note: this also terminates loop!
// If we have only constant indices, fold chains of constant indices
// in this and any preceding GetElemPtr instructions.
bool foldedGEPs = false;
bool leadingNonZeroIdx = gepI && ! IsZero(*gepI->idx_begin());
if (allConstantIndices)
if (Value* newPtr = FoldGetElemChain(ptrChild, idxVec, leadingNonZeroIdx)) {
ptrVal = newPtr;
foldedGEPs = true;
}
// Append the index vector of the current instruction.
// Skip the leading [0] index if preceding GEPs were folded into this.
idxVec.insert(idxVec.end(),
gepI->idx_begin() + (foldedGEPs && !leadingNonZeroIdx),
gepI->idx_end());
return ptrVal;
}
/// GetMemInstArgs - Get the pointer value and the index vector for a memory
/// operation (GetElementPtr, Load, or Store). If all indices of the given
/// memory operation are constant, fold in constant indices in a chain of
/// preceding GetElementPtr instructions (if any), and return the pointer value
/// of the first instruction in the chain. All folded instructions are marked so
/// no code is generated for them. Returns the pointer Value to use, and
/// returns the resulting IndexVector in idxVec. Sets allConstantIndices
/// to true/false if all indices are/aren't const.
///
static Value *GetMemInstArgs(InstructionNode *memInstrNode,
std::vector<Value*> &idxVec,
bool& allConstantIndices) {
allConstantIndices = false;
Instruction* memInst = memInstrNode->getInstruction();
assert(idxVec.size() == 0 && "Need empty vector to return indices");
// If there is a GetElemPtr instruction to fold in to this instr,
// it must be in the left child for Load and GetElemPtr, and in the
// right child for Store instructions.
InstrTreeNode* ptrChild = (memInst->getOpcode() == Instruction::Store
? memInstrNode->rightChild()
: memInstrNode->leftChild());
// Default pointer is the one from the current instruction.
Value* ptrVal = ptrChild->getValue();
// Find the "last" GetElemPtr instruction: this one or the immediate child.
// There will be none if this is a load or a store from a scalar pointer.
InstructionNode* gepNode = NULL;
if (isa<GetElementPtrInst>(memInst))
gepNode = memInstrNode;
else if (isa<InstructionNode>(ptrChild) && isa<GetElementPtrInst>(ptrVal)) {
// Child of load/store is a GEP and memInst is its only use.
// Use its indices and mark it as folded.
gepNode = cast<InstructionNode>(ptrChild);
gepNode->markFoldedIntoParent();
}
// If there are no indices, return the current pointer.
// Else extract the pointer from the GEP and fold the indices.
return gepNode ? GetGEPInstArgs(gepNode, idxVec, allConstantIndices)
: ptrVal;
}
static inline MachineOpCode
ChooseBprInstruction(const InstructionNode* instrNode) {
MachineOpCode opCode;
Instruction* setCCInstr =
((InstructionNode*) instrNode->leftChild())->getInstruction();
switch(setCCInstr->getOpcode()) {
case Instruction::SetEQ: opCode = V9::BRZ; break;
case Instruction::SetNE: opCode = V9::BRNZ; break;
case Instruction::SetLE: opCode = V9::BRLEZ; break;
case Instruction::SetGE: opCode = V9::BRGEZ; break;
case Instruction::SetLT: opCode = V9::BRLZ; break;
case Instruction::SetGT: opCode = V9::BRGZ; break;
default:
assert(0 && "Unrecognized VM instruction!");
opCode = V9::INVALID_OPCODE;
break;
}
return opCode;
}
static inline MachineOpCode
ChooseBpccInstruction(const InstructionNode* instrNode,
const BinaryOperator* setCCInstr) {
MachineOpCode opCode = V9::INVALID_OPCODE;
bool isSigned = setCCInstr->getOperand(0)->getType()->isSigned();
if (isSigned) {
switch(setCCInstr->getOpcode()) {
case Instruction::SetEQ: opCode = V9::BE; break;
case Instruction::SetNE: opCode = V9::BNE; break;
case Instruction::SetLE: opCode = V9::BLE; break;
case Instruction::SetGE: opCode = V9::BGE; break;
case Instruction::SetLT: opCode = V9::BL; break;
case Instruction::SetGT: opCode = V9::BG; break;
default:
assert(0 && "Unrecognized VM instruction!");
break;
}
} else {
switch(setCCInstr->getOpcode()) {
case Instruction::SetEQ: opCode = V9::BE; break;
case Instruction::SetNE: opCode = V9::BNE; break;
case Instruction::SetLE: opCode = V9::BLEU; break;
case Instruction::SetGE: opCode = V9::BCC; break;
case Instruction::SetLT: opCode = V9::BCS; break;
case Instruction::SetGT: opCode = V9::BGU; break;
default:
assert(0 && "Unrecognized VM instruction!");
break;
}
}
return opCode;
}
static inline MachineOpCode
ChooseBFpccInstruction(const InstructionNode* instrNode,
const BinaryOperator* setCCInstr) {
MachineOpCode opCode = V9::INVALID_OPCODE;
switch(setCCInstr->getOpcode()) {
case Instruction::SetEQ: opCode = V9::FBE; break;
case Instruction::SetNE: opCode = V9::FBNE; break;
case Instruction::SetLE: opCode = V9::FBLE; break;
case Instruction::SetGE: opCode = V9::FBGE; break;
case Instruction::SetLT: opCode = V9::FBL; break;
case Instruction::SetGT: opCode = V9::FBG; break;
default:
assert(0 && "Unrecognized VM instruction!");
break;
}
return opCode;
}
// GetTmpForCC - Create a unique TmpInstruction for a boolean value,
// representing the CC register used by a branch on that value.
// For now, hack this using a little static cache of TmpInstructions.
// Eventually the entire BURG instruction selection should be put
// into a separate class that can hold such information.
// The static cache is not too bad because the memory for these
// TmpInstructions will be freed along with the rest of the Function anyway.
//
static TmpInstruction *GetTmpForCC (Value* boolVal, const Function *F,
const Type* ccType,
MachineCodeForInstruction& mcfi) {
typedef hash_map<const Value*, TmpInstruction*> BoolTmpCache;
static BoolTmpCache boolToTmpCache; // Map boolVal -> TmpInstruction*
static const Function *lastFunction = 0;// Use to flush cache between funcs
assert(boolVal->getType() == Type::BoolTy && "Weird but ok! Delete assert");
if (lastFunction != F) {
lastFunction = F;
boolToTmpCache.clear();
}
// Look for tmpI and create a new one otherwise. The new value is
// directly written to map using the ref returned by operator[].
TmpInstruction*& tmpI = boolToTmpCache[boolVal];
if (tmpI == NULL)
tmpI = new TmpInstruction(mcfi, ccType, boolVal);
return tmpI;
}
static inline MachineOpCode
ChooseBccInstruction(const InstructionNode* instrNode, const Type*& setCCType) {
InstructionNode* setCCNode = (InstructionNode*) instrNode->leftChild();
assert(setCCNode->getOpLabel() == SetCCOp);
BinaryOperator* setCCInstr =cast<BinaryOperator>(setCCNode->getInstruction());
setCCType = setCCInstr->getOperand(0)->getType();
if (setCCType->isFloatingPoint())
return ChooseBFpccInstruction(instrNode, setCCInstr);
else
return ChooseBpccInstruction(instrNode, setCCInstr);
}
/// ChooseMovFpcciInstruction - WARNING: since this function has only one
/// caller, it always returns the opcode that expects an immediate and a
/// register. If this function is ever used in cases where an opcode that takes
/// two registers is required, then modify this function and use
/// convertOpcodeFromRegToImm() where required. It will be necessary to expand
/// convertOpcodeFromRegToImm() to handle the new cases of opcodes.
///
static inline MachineOpCode
ChooseMovFpcciInstruction(const InstructionNode* instrNode) {
MachineOpCode opCode = V9::INVALID_OPCODE;
switch(instrNode->getInstruction()->getOpcode()) {
case Instruction::SetEQ: opCode = V9::MOVFEi; break;
case Instruction::SetNE: opCode = V9::MOVFNEi; break;
case Instruction::SetLE: opCode = V9::MOVFLEi; break;
case Instruction::SetGE: opCode = V9::MOVFGEi; break;
case Instruction::SetLT: opCode = V9::MOVFLi; break;
case Instruction::SetGT: opCode = V9::MOVFGi; break;
default:
assert(0 && "Unrecognized VM instruction!");
break;
}
return opCode;
}
/// ChooseMovpcciForSetCC -- Choose a conditional-move instruction
/// based on the type of SetCC operation.
///
/// WARNING: like the previous function, this function always returns
/// the opcode that expects an immediate and a register. See above.
///
static MachineOpCode ChooseMovpcciForSetCC(const InstructionNode* instrNode) {
MachineOpCode opCode = V9::INVALID_OPCODE;
const Type* opType = instrNode->leftChild()->getValue()->getType();
assert(opType->isIntegral() || isa<PointerType>(opType));
bool noSign = opType->isUnsigned() || isa<PointerType>(opType);
switch(instrNode->getInstruction()->getOpcode()) {
case Instruction::SetEQ: opCode = V9::MOVEi; break;
case Instruction::SetLE: opCode = noSign? V9::MOVLEUi : V9::MOVLEi; break;
case Instruction::SetGE: opCode = noSign? V9::MOVCCi : V9::MOVGEi; break;
case Instruction::SetLT: opCode = noSign? V9::MOVCSi : V9::MOVLi; break;
case Instruction::SetGT: opCode = noSign? V9::MOVGUi : V9::MOVGi; break;
case Instruction::SetNE: opCode = V9::MOVNEi; break;
default: assert(0 && "Unrecognized LLVM instr!"); break;
}
return opCode;
}
/// ChooseMovpregiForSetCC -- Choose a conditional-move-on-register-value
/// instruction based on the type of SetCC operation. These instructions
/// compare a register with 0 and perform the move is the comparison is true.
///
/// WARNING: like the previous function, this function it always returns
/// the opcode that expects an immediate and a register. See above.
///
static MachineOpCode ChooseMovpregiForSetCC(const InstructionNode* instrNode) {
MachineOpCode opCode = V9::INVALID_OPCODE;
switch(instrNode->getInstruction()->getOpcode()) {
case Instruction::SetEQ: opCode = V9::MOVRZi; break;
case Instruction::SetLE: opCode = V9::MOVRLEZi; break;
case Instruction::SetGE: opCode = V9::MOVRGEZi; break;
case Instruction::SetLT: opCode = V9::MOVRLZi; break;
case Instruction::SetGT: opCode = V9::MOVRGZi; break;
case Instruction::SetNE: opCode = V9::MOVRNZi; break;
default: assert(0 && "Unrecognized VM instr!"); break;
}
return opCode;
}
static inline MachineOpCode
ChooseConvertToFloatInstr(const TargetMachine& target,
OpLabel vopCode, const Type* opType) {
assert((vopCode == ToFloatTy || vopCode == ToDoubleTy) &&
"Unrecognized convert-to-float opcode!");
assert((opType->isIntegral() || opType->isFloatingPoint() ||
isa<PointerType>(opType))
&& "Trying to convert a non-scalar type to FLOAT/DOUBLE?");
MachineOpCode opCode = V9::INVALID_OPCODE;
unsigned opSize = target.getTargetData().getTypeSize(opType);
if (opType == Type::FloatTy)
opCode = (vopCode == ToFloatTy? V9::NOP : V9::FSTOD);
else if (opType == Type::DoubleTy)
opCode = (vopCode == ToFloatTy? V9::FDTOS : V9::NOP);
else if (opSize <= 4)
opCode = (vopCode == ToFloatTy? V9::FITOS : V9::FITOD);
else {
assert(opSize == 8 && "Unrecognized type size > 4 and < 8!");
opCode = (vopCode == ToFloatTy? V9::FXTOS : V9::FXTOD);
}
return opCode;
}
static inline MachineOpCode
ChooseConvertFPToIntInstr(const TargetMachine& target,
const Type* destType, const Type* opType) {
assert((opType == Type::FloatTy || opType == Type::DoubleTy)
&& "This function should only be called for FLOAT or DOUBLE");
assert((destType->isIntegral() || isa<PointerType>(destType))
&& "Trying to convert FLOAT/DOUBLE to a non-scalar type?");
MachineOpCode opCode = V9::INVALID_OPCODE;
unsigned destSize = target.getTargetData().getTypeSize(destType);
if (destType == Type::UIntTy)
assert(destType != Type::UIntTy && "Expand FP-to-uint beforehand.");
else if (destSize <= 4)
opCode = (opType == Type::FloatTy)? V9::FSTOI : V9::FDTOI;
else {
assert(destSize == 8 && "Unrecognized type size > 4 and < 8!");
opCode = (opType == Type::FloatTy)? V9::FSTOX : V9::FDTOX;
}
return opCode;
}
static MachineInstr*
CreateConvertFPToIntInstr(const TargetMachine& target, Value* srcVal,
Value* destVal, const Type* destType) {
MachineOpCode opCode = ChooseConvertFPToIntInstr(target, destType,
srcVal->getType());
assert(opCode != V9::INVALID_OPCODE && "Expected to need conversion!");
return BuildMI(opCode, 2).addReg(srcVal).addRegDef(destVal);
}
/// CreateCodeToConvertFloatToInt: Convert FP value to signed or unsigned
/// integer. The FP value must be converted to the dest type in an FP register,
/// and the result is then copied from FP to int register via memory. SPARC
/// does not have a float-to-uint conversion, only a float-to-int (fdtoi).
/// Since fdtoi converts to signed integers, any FP value V between MAXINT+1 and
/// MAXUNSIGNED (i.e., 2^31 <= V <= 2^32-1) would be converted incorrectly.
/// Therefore, for converting an FP value to uint32_t, we first need to convert
/// to uint64_t and then to uint32_t.
///
static void
CreateCodeToConvertFloatToInt(const TargetMachine& target,
Value* opVal, Instruction* destI,
std::vector<MachineInstr*>& mvec,
MachineCodeForInstruction& mcfi) {
Function* F = destI->getParent()->getParent();
// Create a temporary to represent the FP register into which the
// int value will placed after conversion. The type of this temporary
// depends on the type of FP register to use: single-prec for a 32-bit
// int or smaller; double-prec for a 64-bit int.
size_t destSize = target.getTargetData().getTypeSize(destI->getType());
const Type* castDestType = destI->getType(); // type for the cast instr result
const Type* castDestRegType; // type for cast instruction result reg
TmpInstruction* destForCast; // dest for cast instruction
Instruction* fpToIntCopyDest = destI; // dest for fp-reg-to-int-reg copy instr
// For converting an FP value to uint32_t, we first need to convert to
// uint64_t and then to uint32_t, as explained above.
if (destI->getType() == Type::UIntTy) {
castDestType = Type::ULongTy; // use this instead of type of destI
castDestRegType = Type::DoubleTy; // uint64_t needs 64-bit FP register.
destForCast = new TmpInstruction(mcfi, castDestRegType, opVal);
fpToIntCopyDest = new TmpInstruction(mcfi, castDestType, destForCast);
} else {
castDestRegType = (destSize > 4)? Type::DoubleTy : Type::FloatTy;
destForCast = new TmpInstruction(mcfi, castDestRegType, opVal);
}
// Create the fp-to-int conversion instruction (src and dest regs are FP regs)
mvec.push_back(CreateConvertFPToIntInstr(target, opVal, destForCast,
castDestType));
// Create the fpreg-to-intreg copy code
CreateCodeToCopyFloatToInt(target, F, destForCast, fpToIntCopyDest, mvec,
mcfi);
// Create the uint64_t to uint32_t conversion, if needed
if (destI->getType() == Type::UIntTy)
CreateZeroExtensionInstructions(target, F, fpToIntCopyDest, destI,
/*numLowBits*/ 32, mvec, mcfi);
}
static inline MachineOpCode
ChooseAddInstruction(const InstructionNode* instrNode) {
return ChooseAddInstructionByType(instrNode->getInstruction()->getType());
}
static inline MachineInstr*
CreateMovFloatInstruction(const InstructionNode* instrNode,
const Type* resultType) {
return BuildMI((resultType == Type::FloatTy) ? V9::FMOVS : V9::FMOVD, 2)
.addReg(instrNode->leftChild()->getValue())
.addRegDef(instrNode->getValue());
}
static inline MachineInstr*
CreateAddConstInstruction(const InstructionNode* instrNode) {
MachineInstr* minstr = NULL;
Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue();
assert(isa<Constant>(constOp));
// Cases worth optimizing are:
// (1) Add with 0 for float or double: use an FMOV of appropriate type,
// instead of an FADD (1 vs 3 cycles). There is no integer MOV.
if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
double dval = FPC->getValue();
if (dval == 0.0)
minstr = CreateMovFloatInstruction(instrNode,
instrNode->getInstruction()->getType());
}
return minstr;
}
static inline MachineOpCode ChooseSubInstructionByType(const Type* resultType) {
MachineOpCode opCode = V9::INVALID_OPCODE;
if (resultType->isInteger() || isa<PointerType>(resultType)) {
opCode = V9::SUBr;
} else {
switch(resultType->getTypeID()) {
case Type::FloatTyID: opCode = V9::FSUBS; break;
case Type::DoubleTyID: opCode = V9::FSUBD; break;
default: assert(0 && "Invalid type for SUB instruction"); break;
}
}
return opCode;
}
static inline MachineInstr*
CreateSubConstInstruction(const InstructionNode* instrNode) {
MachineInstr* minstr = NULL;
Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue();
assert(isa<Constant>(constOp));
// Cases worth optimizing are:
// (1) Sub with 0 for float or double: use an FMOV of appropriate type,
// instead of an FSUB (1 vs 3 cycles). There is no integer MOV.
if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
double dval = FPC->getValue();
if (dval == 0.0)
minstr = CreateMovFloatInstruction(instrNode,
instrNode->getInstruction()->getType());
}
return minstr;
}
static inline MachineOpCode
ChooseFcmpInstruction(const InstructionNode* instrNode) {
MachineOpCode opCode = V9::INVALID_OPCODE;
Value* operand = ((InstrTreeNode*) instrNode->leftChild())->getValue();
switch(operand->getType()->getTypeID()) {
case Type::FloatTyID: opCode = V9::FCMPS; break;
case Type::DoubleTyID: opCode = V9::FCMPD; break;
default: assert(0 && "Invalid type for FCMP instruction"); break;
}
return opCode;
}
/// BothFloatToDouble - Assumes that leftArg and rightArg of instrNode are both
/// cast instructions. Returns true if both are floats cast to double.
///
static inline bool BothFloatToDouble(const InstructionNode* instrNode) {
InstrTreeNode* leftArg = instrNode->leftChild();
InstrTreeNode* rightArg = instrNode->rightChild();
InstrTreeNode* leftArgArg = leftArg->leftChild();
InstrTreeNode* rightArgArg = rightArg->leftChild();
assert(leftArg->getValue()->getType() == rightArg->getValue()->getType());
return (leftArg->getValue()->getType() == Type::DoubleTy &&
leftArgArg->getValue()->getType() == Type::FloatTy &&
rightArgArg->getValue()->getType() == Type::FloatTy);
}
static inline MachineOpCode ChooseMulInstructionByType(const Type* resultType) {
MachineOpCode opCode = V9::INVALID_OPCODE;
if (resultType->isInteger())
opCode = V9::MULXr;
else
switch(resultType->getTypeID()) {
case Type::FloatTyID: opCode = V9::FMULS; break;
case Type::DoubleTyID: opCode = V9::FMULD; break;
default: assert(0 && "Invalid type for MUL instruction"); break;
}
return opCode;
}
static inline MachineInstr*
CreateIntNegInstruction(const TargetMachine& target, Value* vreg) {
return BuildMI(V9::SUBr, 3).addMReg(target.getRegInfo()->getZeroRegNum())
.addReg(vreg).addRegDef(vreg);
}
/// CreateShiftInstructions - Create instruction sequence for any shift
/// operation. SLL or SLLX on an operand smaller than the integer reg. size
/// (64bits) requires a second instruction for explicit sign-extension. Note
/// that we only have to worry about a sign-bit appearing in the most
/// significant bit of the operand after shifting (e.g., bit 32 of Int or bit 16
/// of Short), so we do not have to worry about results that are as large as a
/// normal integer register.
///
static inline void
CreateShiftInstructions(const TargetMachine& target, Function* F,
MachineOpCode shiftOpCode, Value* argVal1,
Value* optArgVal2, /* Use optArgVal2 if not NULL */
unsigned optShiftNum, /* else use optShiftNum */
Instruction* destVal, std::vector<MachineInstr*>& mvec,
MachineCodeForInstruction& mcfi) {
assert((optArgVal2 != NULL || optShiftNum <= 64) &&
"Large shift sizes unexpected, but can be handled below: "
"You need to check whether or not it fits in immed field below");
// If this is a logical left shift of a type smaller than the standard
// integer reg. size, we have to extend the sign-bit into upper bits
// of dest, so we need to put the result of the SLL into a temporary.
Value* shiftDest = destVal;
unsigned opSize = target.getTargetData().getTypeSize(argVal1->getType());
if ((shiftOpCode == V9::SLLr5 || shiftOpCode == V9::SLLXr6) && opSize < 8) {
// put SLL result into a temporary
shiftDest = new TmpInstruction(mcfi, argVal1, optArgVal2, "sllTmp");
}
MachineInstr* M = (optArgVal2 != NULL)
? BuildMI(shiftOpCode, 3).addReg(argVal1).addReg(optArgVal2)
.addReg(shiftDest, MachineOperand::Def)
: BuildMI(shiftOpCode, 3).addReg(argVal1).addZImm(optShiftNum)
.addReg(shiftDest, MachineOperand::Def);
mvec.push_back(M);
if (shiftDest != destVal) {
// extend the sign-bit of the result into all upper bits of dest
assert(8*opSize <= 32 && "Unexpected type size > 4 and < IntRegSize?");
CreateSignExtensionInstructions(target, F, shiftDest, destVal, 8*opSize,
mvec, mcfi);
}
}
/// CreateMulConstInstruction - Does not create any instructions if we
/// cannot exploit constant to create a cheaper instruction. This returns the
/// approximate cost of the instructions generated, which is used to pick the
/// cheapest when both operands are constant.
///
static unsigned
CreateMulConstInstruction(const TargetMachine &target, Function* F,
Value* lval, Value* rval, Instruction* destVal,
std::vector<MachineInstr*>& mvec,
MachineCodeForInstruction& mcfi) {
// Use max. multiply cost, viz., cost of MULX
unsigned cost = target.getInstrInfo()->minLatency(V9::MULXr);
unsigned firstNewInstr = mvec.size();
Value* constOp = rval;
if (! isa<Constant>(constOp))
return cost;
// Cases worth optimizing are:
// (1) Multiply by 0 or 1 for any type: replace with copy (ADD or FMOV)
// (2) Multiply by 2^x for integer types: replace with Shift
const Type* resultType = destVal->getType();
if (resultType->isInteger() || isa<PointerType>(resultType)) {
bool isValidConst;
int64_t C = (int64_t) ConvertConstantToIntType(target, constOp,
constOp->getType(),
isValidConst);
if (isValidConst) {
unsigned pow;
bool needNeg = false;
if (C < 0) {
needNeg = true;
C = -C;
}
if (C == 0 || C == 1) {
cost = target.getInstrInfo()->minLatency(V9::ADDr);
unsigned Zero = target.getRegInfo()->getZeroRegNum();
MachineInstr* M;
if (C == 0)
M =BuildMI(V9::ADDr,3).addMReg(Zero).addMReg(Zero).addRegDef(destVal);
else
M = BuildMI(V9::ADDr,3).addReg(lval).addMReg(Zero).addRegDef(destVal);
mvec.push_back(M);
} else if (isPowerOf2(C, pow)) {
unsigned opSize = target.getTargetData().getTypeSize(resultType);
MachineOpCode opCode = (opSize <= 32)? V9::SLLr5 : V9::SLLXr6;
CreateShiftInstructions(target, F, opCode, lval, NULL, pow,
destVal, mvec, mcfi);
}
if (mvec.size() > 0 && needNeg) {
// insert <reg = SUB 0, reg> after the instr to flip the sign
MachineInstr* M = CreateIntNegInstruction(target, destVal);
mvec.push_back(M);
}
}
} else {
if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
double dval = FPC->getValue();
if (fabs(dval) == 1) {
MachineOpCode opCode = (dval < 0)
? (resultType == Type::FloatTy? V9::FNEGS : V9::FNEGD)
: (resultType == Type::FloatTy? V9::FMOVS : V9::FMOVD);
mvec.push_back(BuildMI(opCode,2).addReg(lval).addRegDef(destVal));
}
}
}
if (firstNewInstr < mvec.size()) {
cost = 0;
for (unsigned i=firstNewInstr; i < mvec.size(); ++i)
cost += target.getInstrInfo()->minLatency(mvec[i]->getOpcode());
}
return cost;
}
/// CreateCheapestMulConstInstruction - Does not create any instructions
/// if we cannot exploit constant to create a cheaper instruction.
///
static inline void
CreateCheapestMulConstInstruction(const TargetMachine &target, Function* F,
Value* lval, Value* rval,
Instruction* destVal,
std::vector<MachineInstr*>& mvec,
MachineCodeForInstruction& mcfi) {
Value* constOp;
if (isa<Constant>(lval) && isa<Constant>(rval)) {
// both operands are constant: evaluate and "set" in dest
Constant* P = ConstantExpr::get(Instruction::Mul,
cast<Constant>(lval),
cast<Constant>(rval));
CreateCodeToLoadConst (target, F, P, destVal, mvec, mcfi);
}
else if (isa<Constant>(rval)) // rval is constant, but not lval
CreateMulConstInstruction(target, F, lval, rval, destVal, mvec, mcfi);
else if (isa<Constant>(lval)) // lval is constant, but not rval
CreateMulConstInstruction(target, F, lval, rval, destVal, mvec, mcfi);
// else neither is constant
return;
}
/// CreateMulInstruction - Returns NULL if we cannot exploit constant
/// to create a cheaper instruction.
///
static inline void
CreateMulInstruction(const TargetMachine &target, Function* F,
Value* lval, Value* rval, Instruction* destVal,
std::vector<MachineInstr*>& mvec,
MachineCodeForInstruction& mcfi,
MachineOpCode forceMulOp = -1) {
unsigned L = mvec.size();
CreateCheapestMulConstInstruction(target,F, lval, rval, destVal, mvec, mcfi);
if (mvec.size() == L) {
// no instructions were added so create MUL reg, reg, reg.
// Use FSMULD if both operands are actually floats cast to doubles.
// Otherwise, use the default opcode for the appropriate type.
MachineOpCode mulOp = ((forceMulOp != -1)
? forceMulOp
: ChooseMulInstructionByType(destVal->getType()));
mvec.push_back(BuildMI(mulOp, 3).addReg(lval).addReg(rval)
.addRegDef(destVal));
}
}
/// ChooseDivInstruction - Generate a divide instruction for Div or Rem.
/// For Rem, this assumes that the operand type will be signed if the result
/// type is signed. This is correct because they must have the same sign.
///
static inline MachineOpCode
ChooseDivInstruction(TargetMachine &target, const InstructionNode* instrNode) {
MachineOpCode opCode = V9::INVALID_OPCODE;
const Type* resultType = instrNode->getInstruction()->getType();
if (resultType->isInteger())
opCode = resultType->isSigned()? V9::SDIVXr : V9::UDIVXr;
else
switch(resultType->getTypeID()) {
case Type::FloatTyID: opCode = V9::FDIVS; break;
case Type::DoubleTyID: opCode = V9::FDIVD; break;
default: assert(0 && "Invalid type for DIV instruction"); break;
}
return opCode;
}
/// CreateDivConstInstruction - Return if we cannot exploit constant to create
/// a cheaper instruction.
///
static void CreateDivConstInstruction(TargetMachine &target,
const InstructionNode* instrNode,
std::vector<MachineInstr*>& mvec) {
Value* LHS = instrNode->leftChild()->getValue();
Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue();
if (!isa<Constant>(constOp))
return;
Instruction* destVal = instrNode->getInstruction();
unsigned ZeroReg = target.getRegInfo()->getZeroRegNum();
// Cases worth optimizing are:
// (1) Divide by 1 for any type: replace with copy (ADD or FMOV)
// (2) Divide by 2^x for integer types: replace with SR[L or A]{X}
const Type* resultType = instrNode->getInstruction()->getType();
if (resultType->isInteger()) {
unsigned pow;
bool isValidConst;
int64_t C = (int64_t) ConvertConstantToIntType(target, constOp,
constOp->getType(),
isValidConst);
if (isValidConst) {
bool needNeg = false;
if (C < 0) {
needNeg = true;
C = -C;
}
if (C == 1) {
mvec.push_back(BuildMI(V9::ADDr, 3).addReg(LHS).addMReg(ZeroReg)
.addRegDef(destVal));
} else if (isPowerOf2(C, pow)) {
unsigned opCode;
Value* shiftOperand;
unsigned opSize = target.getTargetData().getTypeSize(resultType);
if (resultType->isSigned()) {
// For N / 2^k, if the operand N is negative,
// we need to add (2^k - 1) before right-shifting by k, i.e.,
//
// (N / 2^k) = N >> k, if N >= 0;
// (N + 2^k - 1) >> k, if N < 0
//
// If N is <= 32 bits, use:
// sra N, 31, t1 // t1 = ~0, if N < 0, 0 else
// srl t1, 32-k, t2 // t2 = 2^k - 1, if N < 0, 0 else
// add t2, N, t3 // t3 = N + 2^k -1, if N < 0, N else
// sra t3, k, result // result = N / 2^k
//
// If N is 64 bits, use:
// srax N, k-1, t1 // t1 = sign bit in high k positions
// srlx t1, 64-k, t2 // t2 = 2^k - 1, if N < 0, 0 else
// add t2, N, t3 // t3 = N + 2^k -1, if N < 0, N else
// sra t3, k, result // result = N / 2^k
TmpInstruction *sraTmp, *srlTmp, *addTmp;
MachineCodeForInstruction& mcfi
= MachineCodeForInstruction::get(destVal);
sraTmp = new TmpInstruction(mcfi, resultType, LHS, 0, "getSign");
srlTmp = new TmpInstruction(mcfi, resultType, LHS, 0, "getPlus2km1");
addTmp = new TmpInstruction(mcfi, resultType, LHS, srlTmp,"incIfNeg");
// Create the SRA or SRAX instruction to get the sign bit
mvec.push_back(BuildMI((opSize > 4)? V9::SRAXi6 : V9::SRAi5, 3)
.addReg(LHS)
.addSImm((resultType==Type::LongTy)? pow-1 : 31)
.addRegDef(sraTmp));
// Create the SRL or SRLX instruction to get the sign bit
mvec.push_back(BuildMI((opSize > 4)? V9::SRLXi6 : V9::SRLi5, 3)
.addReg(sraTmp)
.addSImm((resultType==Type::LongTy)? 64-pow : 32-pow)
.addRegDef(srlTmp));
// Create the ADD instruction to add 2^pow-1 for negative values
mvec.push_back(BuildMI(V9::ADDr, 3).addReg(LHS).addReg(srlTmp)
.addRegDef(addTmp));
// Get the shift operand and "right-shift" opcode to do the divide
shiftOperand = addTmp;
opCode = (opSize > 4)? V9::SRAXi6 : V9::SRAi5;
} else {
// Get the shift operand and "right-shift" opcode to do the divide
shiftOperand = LHS;
opCode = (opSize > 4)? V9::SRLXi6 : V9::SRLi5;
}
// Now do the actual shift!
mvec.push_back(BuildMI(opCode, 3).addReg(shiftOperand).addZImm(pow)
.addRegDef(destVal));
}
if (needNeg && (C == 1 || isPowerOf2(C, pow))) {
// insert <reg = SUB 0, reg> after the instr to flip the sign
mvec.push_back(CreateIntNegInstruction(target, destVal));
}
}
} else {
if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
double dval = FPC->getValue();
if (fabs(dval) == 1) {
unsigned opCode =
(dval < 0) ? (resultType == Type::FloatTy? V9::FNEGS : V9::FNEGD)
: (resultType == Type::FloatTy? V9::FMOVS : V9::FMOVD);
mvec.push_back(BuildMI(opCode, 2).addReg(LHS).addRegDef(destVal));
}
}
}
}
static void CreateCodeForVariableSizeAlloca(const TargetMachine& target,
Instruction* result, unsigned tsize,
Value* numElementsVal,
std::vector<MachineInstr*>& getMvec)
{
Value* totalSizeVal;
MachineInstr* M;
MachineCodeForInstruction& mcfi = MachineCodeForInstruction::get(result);
Function *F = result->getParent()->getParent();
// Enforce the alignment constraints on the stack pointer at
// compile time if the total size is a known constant.
if (isa<Constant>(numElementsVal)) {
bool isValid;
int64_t numElem = (int64_t)
ConvertConstantToIntType(target, numElementsVal,
numElementsVal->getType(), isValid);
assert(isValid && "Unexpectedly large array dimension in alloca!");
int64_t total = numElem * tsize;
if (int extra= total % SparcV9FrameInfo::StackFrameSizeAlignment)
total += SparcV9FrameInfo::StackFrameSizeAlignment - extra;
totalSizeVal = ConstantSInt::get(Type::IntTy, total);
} else {
// The size is not a constant. Generate code to compute it and
// code to pad the size for stack alignment.
// Create a Value to hold the (constant) element size
Value* tsizeVal = ConstantSInt::get(Type::IntTy, tsize);
// Create temporary values to hold the result of MUL, SLL, SRL
// To pad `size' to next smallest multiple of 16:
// size = (size + 15) & (-16 = 0xfffffffffffffff0)
TmpInstruction* tmpProd = new TmpInstruction(mcfi,numElementsVal, tsizeVal);
TmpInstruction* tmpAdd15= new TmpInstruction(mcfi,numElementsVal, tmpProd);
TmpInstruction* tmpAndf0= new TmpInstruction(mcfi,numElementsVal, tmpAdd15);
// Instruction 1: mul numElements, typeSize -> tmpProd
// This will optimize the MUL as far as possible.
CreateMulInstruction(target, F, numElementsVal, tsizeVal, tmpProd, getMvec,
mcfi, -1);
// Instruction 2: andn tmpProd, 0x0f -> tmpAndn
getMvec.push_back(BuildMI(V9::ADDi, 3).addReg(tmpProd).addSImm(15)
.addReg(tmpAdd15, MachineOperand::Def));
// Instruction 3: add tmpAndn, 0x10 -> tmpAdd16
getMvec.push_back(BuildMI(V9::ANDi, 3).addReg(tmpAdd15).addSImm(-16)
.addReg(tmpAndf0, MachineOperand::Def));
totalSizeVal = tmpAndf0;
}
// Get the constant offset from SP for dynamically allocated storage
// and create a temporary Value to hold it.
MachineFunction& mcInfo = MachineFunction::get(F);
bool growUp;
ConstantSInt* dynamicAreaOffset =
ConstantSInt::get(Type::IntTy,
target.getFrameInfo()->getDynamicAreaOffset(mcInfo,growUp));
assert(! growUp && "Has SPARC v9 stack frame convention changed?");
unsigned SPReg = target.getRegInfo()->getStackPointer();
// Instruction 2: sub %sp, totalSizeVal -> %sp
getMvec.push_back(BuildMI(V9::SUBr, 3).addMReg(SPReg).addReg(totalSizeVal)
.addMReg(SPReg,MachineOperand::Def));
// Instruction 3: add %sp, frameSizeBelowDynamicArea -> result
getMvec.push_back(BuildMI(V9::ADDr,3).addMReg(SPReg).addReg(dynamicAreaOffset)
.addRegDef(result));
}
static void
CreateCodeForFixedSizeAlloca(const TargetMachine& target,
Instruction* result, unsigned tsize,
unsigned numElements,
std::vector<MachineInstr*>& getMvec) {
assert(result && result->getParent() &&
"Result value is not part of a function?");
Function *F = result->getParent()->getParent();
MachineFunction &mcInfo = MachineFunction::get(F);
// If the alloca is of zero bytes (which is perfectly legal) we bump it up to
// one byte. This is unnecessary, but I really don't want to break any
// fragile logic in this code. FIXME.
if (tsize == 0)
tsize = 1;
// Put the variable in the dynamically sized area of the frame if either:
// (a) The offset is too large to use as an immediate in load/stores
// (check LDX because all load/stores have the same-size immed. field).
// (b) The object is "large", so it could cause many other locals,
// spills, and temporaries to have large offsets.
// NOTE: We use LARGE = 8 * argSlotSize = 64 bytes.
// You've gotta love having only 13 bits for constant offset values :-|.
//
unsigned paddedSize;
int offsetFromFP = mcInfo.getInfo<SparcV9FunctionInfo>()->computeOffsetforLocalVar(result,
paddedSize,
tsize * numElements);
if (((int)paddedSize) > 8 * SparcV9FrameInfo::SizeOfEachArgOnStack ||
!target.getInstrInfo()->constantFitsInImmedField(V9::LDXi,offsetFromFP)) {
CreateCodeForVariableSizeAlloca(target, result, tsize,
ConstantSInt::get(Type::IntTy,numElements),
getMvec);
return;
}
// else offset fits in immediate field so go ahead and allocate it.
offsetFromFP = mcInfo.getInfo<SparcV9FunctionInfo>()->allocateLocalVar(result, tsize *numElements);
// Create a temporary Value to hold the constant offset.
// This is needed because it may not fit in the immediate field.
ConstantSInt* offsetVal = ConstantSInt::get(Type::IntTy, offsetFromFP);
// Instruction 1: add %fp, offsetFromFP -> result
unsigned FPReg = target.getRegInfo()->getFramePointer();
getMvec.push_back(BuildMI(V9::ADDr, 3).addMReg(FPReg).addReg(offsetVal)
.addRegDef(result));
}
/// SetOperandsForMemInstr - Choose addressing mode for the given load or store
/// instruction. Use [reg+reg] if it is an indexed reference, and the index
/// offset is not a constant or if it cannot fit in the offset field. Use
/// [reg+offset] in all other cases. This assumes that all array refs are
/// "lowered" to one of these forms:
/// %x = load (subarray*) ptr, constant ; single constant offset
/// %x = load (subarray*) ptr, offsetVal ; single non-constant offset
/// Generally, this should happen via strength reduction + LICM. Also, strength
/// reduction should take care of using the same register for the loop index
/// variable and an array index, when that is profitable.
///
static void SetOperandsForMemInstr(unsigned Opcode,
std::vector<MachineInstr*>& mvec,
InstructionNode* vmInstrNode,
const TargetMachine& target) {
Instruction* memInst = vmInstrNode->getInstruction();
// Index vector, ptr value, and flag if all indices are const.
std::vector<Value*> idxVec;
bool allConstantIndices;
Value* ptrVal = GetMemInstArgs(vmInstrNode, idxVec, allConstantIndices);
// Now create the appropriate operands for the machine instruction.
// First, initialize so we default to storing the offset in a register.
int64_t smallConstOffset = 0;
Value* valueForRegOffset = NULL;
MachineOperand::MachineOperandType offsetOpType =
MachineOperand::MO_VirtualRegister;
// Check if there is an index vector and if so, compute the
// right offset for structures and for arrays
if (!idxVec.empty()) {
const PointerType* ptrType = cast<PointerType>(ptrVal->getType());
// If all indices are constant, compute the combined offset directly.
if (allConstantIndices) {
// Compute the offset value using the index vector. Create a
// virtual reg. for it since it may not fit in the immed field.
uint64_t offset = target.getTargetData().getIndexedOffset(ptrType,idxVec);
valueForRegOffset = ConstantSInt::get(Type::LongTy, offset);
} else {
// There is at least one non-constant offset. Therefore, this must
// be an array ref, and must have been lowered to a single non-zero
// offset. (An extra leading zero offset, if any, can be ignored.)
// Generate code sequence to compute address from index.
bool firstIdxIsZero = IsZero(idxVec[0]);
assert(idxVec.size() == 1U + firstIdxIsZero
&& "Array refs must be lowered before Instruction Selection");
Value* idxVal = idxVec[firstIdxIsZero];
std::vector<MachineInstr*> mulVec;
Instruction* addr =
new TmpInstruction(MachineCodeForInstruction::get(memInst),
Type::ULongTy, memInst);
// Get the array type indexed by idxVal, and compute its element size.
// The call to getTypeSize() will fail if size is not constant.
const Type* vecType = (firstIdxIsZero
? GetElementPtrInst::getIndexedType(ptrType,
std::vector<Value*>(1U, idxVec[0]),
/*AllowCompositeLeaf*/ true)
: ptrType);
const Type* eltType = cast<SequentialType>(vecType)->getElementType();
ConstantUInt* eltSizeVal = ConstantUInt::get(Type::ULongTy,
target.getTargetData().getTypeSize(eltType));
// CreateMulInstruction() folds constants intelligently enough.
CreateMulInstruction(target, memInst->getParent()->getParent(),
idxVal, /* lval, not likely to be const*/
eltSizeVal, /* rval, likely to be constant */
addr, /* result */
mulVec, MachineCodeForInstruction::get(memInst),
-1);
assert(mulVec.size() > 0 && "No multiply code created?");
mvec.insert(mvec.end(), mulVec.begin(), mulVec.end());
valueForRegOffset = addr;
}
} else {
offsetOpType = MachineOperand::MO_SignExtendedImmed;
smallConstOffset = 0;
}
// For STORE:
// Operand 0 is value, operand 1 is ptr, operand 2 is offset
// For LOAD or GET_ELEMENT_PTR,
// Operand 0 is ptr, operand 1 is offset, operand 2 is result.
unsigned offsetOpNum, ptrOpNum;
MachineInstr *MI;
if (memInst->getOpcode() == Instruction::Store) {
if (offsetOpType == MachineOperand::MO_VirtualRegister) {
MI = BuildMI(Opcode, 3).addReg(vmInstrNode->leftChild()->getValue())
.addReg(ptrVal).addReg(valueForRegOffset);
} else {
Opcode = convertOpcodeFromRegToImm(Opcode);
MI = BuildMI(Opcode, 3).addReg(vmInstrNode->leftChild()->getValue())
.addReg(ptrVal).addSImm(smallConstOffset);
}
} else {
if (offsetOpType == MachineOperand::MO_VirtualRegister) {
MI = BuildMI(Opcode, 3).addReg(ptrVal).addReg(valueForRegOffset)
.addRegDef(memInst);
} else {
Opcode = convertOpcodeFromRegToImm(Opcode);
MI = BuildMI(Opcode, 3).addReg(ptrVal).addSImm(smallConstOffset)
.addRegDef(memInst);
}
}
mvec.push_back(MI);
}
/// ForwardOperand - Substitute operand `operandNum' of the instruction in
/// node `treeNode' in place of the use(s) of that instruction in node `parent'.
/// Check both explicit and implicit operands! Also make sure to skip over a
/// parent who: (1) is a list node in the Burg tree, or (2) itself had its
/// results forwarded to its parent.
///
static void ForwardOperand (InstructionNode *treeNode, InstrTreeNode *parent,
int operandNum) {
assert(treeNode && parent && "Invalid invocation of ForwardOperand");
Instruction* unusedOp = treeNode->getInstruction();
Value* fwdOp = unusedOp->getOperand(operandNum);
// The parent itself may be a list node, so find the real parent instruction
while (parent->getNodeType() != InstrTreeNode::NTInstructionNode) {
parent = parent->parent();
assert(parent && "ERROR: Non-instruction node has no parent in tree.");
}
InstructionNode* parentInstrNode = (InstructionNode*) parent;
Instruction* userInstr = parentInstrNode->getInstruction();
MachineCodeForInstruction &mvec = MachineCodeForInstruction::get(userInstr);
// The parent's mvec would be empty if it was itself forwarded.
// Recursively call ForwardOperand in that case...
//
if (mvec.size() == 0) {
assert(parent->parent() != NULL &&
"Parent could not have been forwarded, yet has no instructions?");
ForwardOperand(treeNode, parent->parent(), operandNum);
} else {
for (unsigned i=0, N=mvec.size(); i < N; i++) {
MachineInstr* minstr = mvec[i];
for (unsigned i=0, numOps=minstr->getNumOperands(); i < numOps; ++i) {
const MachineOperand& mop = minstr->getOperand(i);
if (mop.getType() == MachineOperand::MO_VirtualRegister &&
mop.getVRegValue() == unusedOp) {
minstr->SetMachineOperandVal(i, MachineOperand::MO_VirtualRegister,
fwdOp);
}
}
for (unsigned i=0,numOps=minstr->getNumImplicitRefs(); i<numOps; ++i)
if (minstr->getImplicitRef(i) == unusedOp)
minstr->setImplicitRef(i, fwdOp);
}
}
}
/// AllUsesAreBranches - Returns true if all the uses of I are
/// Branch instructions, false otherwise.
///
inline bool AllUsesAreBranches(const Instruction* I) {
for (Value::use_const_iterator UI=I->use_begin(), UE=I->use_end();
UI != UE; ++UI)
if (! isa<TmpInstruction>(*UI) // ignore tmp instructions here
&& cast<Instruction>(*UI)->getOpcode() != Instruction::Br)
return false;
return true;
}
/// CodeGenIntrinsic - Generate code for any intrinsic that needs a special
/// code sequence instead of a regular call. If not that kind of intrinsic, do
/// nothing. Returns true if code was generated, otherwise false.
///
static bool CodeGenIntrinsic(Intrinsic::ID iid, CallInst &callInstr,
TargetMachine &target,
std::vector<MachineInstr*>& mvec) {
switch (iid) {
default:
assert(0 && "Unknown intrinsic function call should have been lowered!");
case Intrinsic::vastart: {
// Get the address of the first incoming vararg argument on the stack
Function* func = cast<Function>(callInstr.getParent()->getParent());
int numFixedArgs = func->getFunctionType()->getNumParams();
int fpReg = SparcV9::i6;
int firstVarArgOff = numFixedArgs * 8 +
SparcV9FrameInfo::FirstIncomingArgOffsetFromFP;
mvec.push_back(BuildMI(V9::ADDi, 3).addMReg(fpReg).addSImm(firstVarArgOff).
addRegDef(&callInstr));
return true;
}
case Intrinsic::vaend:
return true; // no-op on SparcV9
case Intrinsic::vacopy:
// Simple copy of current va_list (arg1) to new va_list (result)
mvec.push_back(BuildMI(V9::ORr, 3).
addMReg(target.getRegInfo()->getZeroRegNum()).
addReg(callInstr.getOperand(1)).
addRegDef(&callInstr));
return true;
}
}
/// ThisIsAChainRule - returns true if the given BURG rule is a chain rule.
///
extern bool ThisIsAChainRule(int eruleno) {
switch(eruleno) {
case 111: // stmt: reg
case 123:
case 124:
case 125:
case 126:
case 127:
case 128:
case 129:
case 130:
case 131:
case 132:
case 133:
case 155:
case 221:
case 222:
case 241:
case 242:
case 243:
case 244:
case 245:
case 321:
return true; break;
default:
return false; break;
}
}
/// GetInstructionsByRule - Choose machine instructions for the
/// SPARC V9 according to the patterns chosen by the BURG-generated parser.
/// This is where most of the work in the V9 instruction selector gets done.
///
void GetInstructionsByRule(InstructionNode* subtreeRoot, int ruleForNode,
short* nts, TargetMachine &target,
std::vector<MachineInstr*>& mvec) {
bool checkCast = false; // initialize here to use fall-through
bool maskUnsignedResult = false;
int nextRule;
int forwardOperandNum = -1;
unsigned allocaSize = 0;
MachineInstr* M, *M2;
unsigned L;
bool foldCase = false;
mvec.clear();
// If the code for this instruction was folded into the parent (user),
// then do nothing!
if (subtreeRoot->isFoldedIntoParent())
return;
// Let's check for chain rules outside the switch so that we don't have
// to duplicate the list of chain rule production numbers here again
if (ThisIsAChainRule(ruleForNode)) {
// Chain rules have a single nonterminal on the RHS.
// Get the rule that matches the RHS non-terminal and use that instead.
assert(nts[0] && ! nts[1]
&& "A chain rule should have only one RHS non-terminal!");
nextRule = burm_rule(subtreeRoot->state, nts[0]);
nts = burm_nts[nextRule];
GetInstructionsByRule(subtreeRoot, nextRule, nts, target, mvec);
} else {
switch(ruleForNode) {
case 1: // stmt: Ret
case 2: // stmt: RetValue(reg)
{ // NOTE: Prepass of register allocation is responsible
// for moving return value to appropriate register.
// Copy the return value to the required return register.
// Mark the return Value as an implicit ref of the RET instr..
// Mark the return-address register as a hidden virtual reg.
// Finally put a NOP in the delay slot.
ReturnInst *returnInstr=cast<ReturnInst>(subtreeRoot->getInstruction());
Value* retVal = returnInstr->getReturnValue();
MachineCodeForInstruction& mcfi =
MachineCodeForInstruction::get(returnInstr);
// Create a hidden virtual reg to represent the return address register
// used by the machine instruction but not represented in LLVM.
Instruction* returnAddrTmp = new TmpInstruction(mcfi, returnInstr);
MachineInstr* retMI =
BuildMI(V9::JMPLRETi, 3).addReg(returnAddrTmp).addSImm(8)
.addMReg(target.getRegInfo()->getZeroRegNum(), MachineOperand::Def);
// If there is a value to return, we need to:
// (a) Sign-extend the value if it is smaller than 8 bytes (reg size)
// (b) Insert a copy to copy the return value to the appropriate reg.
// -- For FP values, create a FMOVS or FMOVD instruction
// -- For non-FP values, create an add-with-0 instruction
if (retVal != NULL) {
const SparcV9RegInfo& regInfo =
(SparcV9RegInfo&) *target.getRegInfo();
const Type* retType = retVal->getType();
unsigned regClassID = regInfo.getRegClassIDOfType(retType);
unsigned retRegNum = (retType->isFloatingPoint()
? (unsigned) SparcV9FloatRegClass::f0
: (unsigned) SparcV9IntRegClass::i0);
retRegNum = regInfo.getUnifiedRegNum(regClassID, retRegNum);
// Insert sign-extension instructions for small signed values.
Value* retValToUse = retVal;
if (retType->isIntegral() && retType->isSigned()) {
unsigned retSize = target.getTargetData().getTypeSize(retType);
if (retSize <= 4) {
// Create a temporary virtual reg. to hold the sign-extension.
retValToUse = new TmpInstruction(mcfi, retVal);
// Sign-extend retVal and put the result in the temporary reg.
CreateSignExtensionInstructions
(target, returnInstr->getParent()->getParent(),
retVal, retValToUse, 8*retSize, mvec, mcfi);
}
}
// (b) Now, insert a copy to to the appropriate register:
// -- For FP values, create a FMOVS or FMOVD instruction
// -- For non-FP values, create an add-with-0 instruction
// First, create a virtual register to represent the register and
// mark this vreg as being an implicit operand of the ret MI.
TmpInstruction* retVReg =
new TmpInstruction(mcfi, retValToUse, NULL, "argReg");
retMI->addImplicitRef(retVReg);
if (retType->isFloatingPoint())
M = (BuildMI(retType==Type::FloatTy? V9::FMOVS : V9::FMOVD, 2)
.addReg(retValToUse).addReg(retVReg, MachineOperand::Def));
else
M = (BuildMI(ChooseAddInstructionByType(retType), 3)
.addReg(retValToUse).addSImm((int64_t) 0)
.addReg(retVReg, MachineOperand::Def));
// Mark the operand with the register it should be assigned
M->SetRegForOperand(M->getNumOperands()-1, retRegNum);
retMI->SetRegForImplicitRef(retMI->getNumImplicitRefs()-1, retRegNum);
mvec.push_back(M);
}
// Now insert the RET instruction and a NOP for the delay slot
mvec.push_back(retMI);
mvec.push_back(BuildMI(V9::NOP, 0));
break;
}
case 3: // stmt: Store(reg,reg)
case 4: // stmt: Store(reg,ptrreg)
SetOperandsForMemInstr(ChooseStoreInstruction(
subtreeRoot->leftChild()->getValue()->getType()),
mvec, subtreeRoot, target);
break;
case 5: // stmt: BrUncond
{
BranchInst *BI = cast<BranchInst>(subtreeRoot->getInstruction());
mvec.push_back(BuildMI(V9::BA, 1).addPCDisp(BI->getSuccessor(0)));
// delay slot
mvec.push_back(BuildMI(V9::NOP, 0));
break;
}
case 206: // stmt: BrCond(setCCconst)
{ // setCCconst => boolean was computed with `%b = setCC type reg1 const'
// If the constant is ZERO, we can use the branch-on-integer-register
// instructions and avoid the SUBcc instruction entirely.
// Otherwise this is just the same as case 5, so just fall through.
//
InstrTreeNode* constNode = subtreeRoot->leftChild()->rightChild();
assert(constNode &&
constNode->getNodeType() ==InstrTreeNode::NTConstNode);
Constant *constVal = cast<Constant>(constNode->getValue());
bool isValidConst;
if ((constVal->getType()->isInteger()
|| isa<PointerType>(constVal->getType()))
&& ConvertConstantToIntType(target,
constVal, constVal->getType(), isValidConst) == 0
&& isValidConst)
{
// That constant is a zero after all...
// Use the left child of setCC as the first argument!
// Mark the setCC node so that no code is generated for it.
InstructionNode* setCCNode = (InstructionNode*)
subtreeRoot->leftChild();
assert(setCCNode->getOpLabel() == SetCCOp);
setCCNode->markFoldedIntoParent();
BranchInst* brInst=cast<BranchInst>(subtreeRoot->getInstruction());
M = BuildMI(ChooseBprInstruction(subtreeRoot), 2)
.addReg(setCCNode->leftChild()->getValue())
.addPCDisp(brInst->getSuccessor(0));
mvec.push_back(M);
// delay slot
mvec.push_back(BuildMI(V9::NOP, 0));
// false branch
mvec.push_back(BuildMI(V9::BA, 1)
.addPCDisp(brInst->getSuccessor(1)));
// delay slot
mvec.push_back(BuildMI(V9::NOP, 0));
break;
}
// ELSE FALL THROUGH
}
case 6: // stmt: BrCond(setCC)
{ // bool => boolean was computed with SetCC.
// The branch to use depends on whether it is FP, signed, or unsigned.
// If it is an integer CC, we also need to find the unique
// TmpInstruction representing that CC.
//
BranchInst* brInst = cast<BranchInst>(subtreeRoot->getInstruction());
const Type* setCCType;
unsigned Opcode = ChooseBccInstruction(subtreeRoot, setCCType);
Value* ccValue = GetTmpForCC(subtreeRoot->leftChild()->getValue(),
brInst->getParent()->getParent(),
setCCType,
MachineCodeForInstruction::get(brInst));
M = BuildMI(Opcode, 2).addCCReg(ccValue)
.addPCDisp(brInst->getSuccessor(0));
mvec.push_back(M);
// delay slot
mvec.push_back(BuildMI(V9::NOP, 0));
// false branch
mvec.push_back(BuildMI(V9::BA, 1).addPCDisp(brInst->getSuccessor(1)));
// delay slot
mvec.push_back(BuildMI(V9::NOP, 0));
break;
}
case 208: // stmt: BrCond(boolconst)
{
// boolconst => boolean is a constant; use BA to first or second label
Constant* constVal =
cast<Constant>(subtreeRoot->leftChild()->getValue());
unsigned dest = cast<ConstantBool>(constVal)->getValue()? 0 : 1;
M = BuildMI(V9::BA, 1).addPCDisp(
cast<BranchInst>(subtreeRoot->getInstruction())->getSuccessor(dest));
mvec.push_back(M);
// delay slot
mvec.push_back(BuildMI(V9::NOP, 0));
break;
}
case 8: // stmt: BrCond(boolreg)
{ // boolreg => boolean is recorded in an integer register.
// Use branch-on-integer-register instruction.
//
BranchInst *BI = cast<BranchInst>(subtreeRoot->getInstruction());
M = BuildMI(V9::BRNZ, 2).addReg(subtreeRoot->leftChild()->getValue())
.addPCDisp(BI->getSuccessor(0));
mvec.push_back(M);
// delay slot
mvec.push_back(BuildMI(V9::NOP, 0));
// false branch
mvec.push_back(BuildMI(V9::BA, 1).addPCDisp(BI->getSuccessor(1)));
// delay slot
mvec.push_back(BuildMI(V9::NOP, 0));
break;
}
case 9: // stmt: Switch(reg)
assert(0 && "*** SWITCH instruction is not implemented yet.");
break;
case 10: // reg: VRegList(reg, reg)
assert(0 && "VRegList should never be the topmost non-chain rule");
break;
case 21: // bool: Not(bool,reg): Compute with a conditional-move-on-reg
{ // First find the unary operand. It may be left or right, usually right.
Instruction* notI = subtreeRoot->getInstruction();
Value* notArg = BinaryOperator::getNotArgument(
cast<BinaryOperator>(subtreeRoot->getInstruction()));
unsigned ZeroReg = target.getRegInfo()->getZeroRegNum();
// Unconditionally set register to 0
mvec.push_back(BuildMI(V9::SETHI, 2).addZImm(0).addRegDef(notI));
// Now conditionally move 1 into the register.
// Mark the register as a use (as well as a def) because the old
// value will be retained if the condition is false.
mvec.push_back(BuildMI(V9::MOVRZi, 3).addReg(notArg).addZImm(1)
.addReg(notI, MachineOperand::UseAndDef));
break;
}
case 421: // reg: BNot(reg,reg): Compute as reg = reg XOR-NOT 0
{ // First find the unary operand. It may be left or right, usually right.
Value* notArg = BinaryOperator::getNotArgument(
cast<BinaryOperator>(subtreeRoot->getInstruction()));
unsigned ZeroReg = target.getRegInfo()->getZeroRegNum();
mvec.push_back(BuildMI(V9::XNORr, 3).addReg(notArg).addMReg(ZeroReg)
.addRegDef(subtreeRoot->getValue()));
break;
}
case 322: // reg: Not(tobool, reg):
// Fold CAST-TO-BOOL with NOT by inverting the sense of cast-to-bool
foldCase = true;
// Just fall through!
case 22: // reg: ToBoolTy(reg):
{
Instruction* castI = subtreeRoot->getInstruction();
Value* opVal = subtreeRoot->leftChild()->getValue();
assert(opVal->getType()->isIntegral() ||
isa<PointerType>(opVal->getType()));
// Unconditionally set register to 0
mvec.push_back(BuildMI(V9::SETHI, 2).addZImm(0).addRegDef(castI));
// Now conditionally move 1 into the register.
// Mark the register as a use (as well as a def) because the old
// value will be retained if the condition is false.
MachineOpCode opCode = foldCase? V9::MOVRZi : V9::MOVRNZi;
mvec.push_back(BuildMI(opCode, 3).addReg(opVal).addZImm(1)
.addReg(castI, MachineOperand::UseAndDef));
break;
}
case 23: // reg: ToUByteTy(reg)
case 24: // reg: ToSByteTy(reg)
case 25: // reg: ToUShortTy(reg)
case 26: // reg: ToShortTy(reg)
case 27: // reg: ToUIntTy(reg)
case 28: // reg: ToIntTy(reg)
case 29: // reg: ToULongTy(reg)
case 30: // reg: ToLongTy(reg)
{
//======================================================================
// Rules for integer conversions:
//
//--------
// From ISO 1998 C++ Standard, Sec. 4.7:
//
// 2. If the destination type is unsigned, the resulting value is
// the least unsigned integer congruent to the source integer
// (modulo 2n where n is the number of bits used to represent the
// unsigned type). [Note: In a two s complement representation,
// this conversion is conceptual and there is no change in the
// bit pattern (if there is no truncation). ]
//
// 3. If the destination type is signed, the value is unchanged if
// it can be represented in the destination type (and bitfield width);
// otherwise, the value is implementation-defined.
//--------
//
// Since we assume 2s complement representations, this implies:
//
// -- If operand is smaller than destination, zero-extend or sign-extend
// according to the signedness of the *operand*: source decides:
// (1) If operand is signed, sign-extend it.
// If dest is unsigned, zero-ext the result!
// (2) If operand is unsigned, our current invariant is that
// it's high bits are correct, so zero-extension is not needed.
//
// -- If operand is same size as or larger than destination,
// zero-extend or sign-extend according to the signedness of
// the *destination*: destination decides:
// (1) If destination is signed, sign-extend (truncating if needed)
// This choice is implementation defined. We sign-extend the
// operand, which matches both Sun's cc and gcc3.2.
// (2) If destination is unsigned, zero-extend (truncating if needed)
//======================================================================
Instruction* destI = subtreeRoot->getInstruction();
Function* currentFunc = destI->getParent()->getParent();
MachineCodeForInstruction& mcfi=MachineCodeForInstruction::get(destI);
Value* opVal = subtreeRoot->leftChild()->getValue();
const Type* opType = opVal->getType();
const Type* destType = destI->getType();
unsigned opSize = target.getTargetData().getTypeSize(opType);
unsigned destSize = target.getTargetData().getTypeSize(destType);
bool isIntegral = opType->isIntegral() || isa<PointerType>(opType);
if (opType == Type::BoolTy ||
opType == destType ||
isIntegral && opSize == destSize && opSize == 8) {
// nothing to do in all these cases
forwardOperandNum = 0; // forward first operand to user
} else if (opType->isFloatingPoint()) {
CreateCodeToConvertFloatToInt(target, opVal, destI, mvec, mcfi);
if (destI->getType()->isUnsigned() && destI->getType() !=Type::UIntTy)
maskUnsignedResult = true; // not handled by fp->int code
} else if (isIntegral) {
bool opSigned = opType->isSigned();
bool destSigned = destType->isSigned();
unsigned extSourceInBits = 8 * std::min<unsigned>(opSize, destSize);
assert(! (opSize == destSize && opSigned == destSigned) &&
"How can different int types have same size and signedness?");
bool signExtend = (opSize < destSize && opSigned ||
opSize >= destSize && destSigned);
bool signAndZeroExtend = (opSize < destSize && destSize < 8u &&
opSigned && !destSigned);
assert(!signAndZeroExtend || signExtend);
bool zeroExtendOnly = opSize >= destSize && !destSigned;
assert(!zeroExtendOnly || !signExtend);
if (signExtend) {
Value* signExtDest = (signAndZeroExtend
? new TmpInstruction(mcfi, destType, opVal)
: destI);
CreateSignExtensionInstructions
(target, currentFunc,opVal,signExtDest,extSourceInBits,mvec,mcfi);
if (signAndZeroExtend)
CreateZeroExtensionInstructions
(target, currentFunc, signExtDest, destI, 8*destSize, mvec, mcfi);
}
else if (zeroExtendOnly) {
CreateZeroExtensionInstructions
(target, currentFunc, opVal, destI, extSourceInBits, mvec, mcfi);
}
else
forwardOperandNum = 0; // forward first operand to user
} else
assert(0 && "Unrecognized operand type for convert-to-integer");
break;
}
case 31: // reg: ToFloatTy(reg):
case 32: // reg: ToDoubleTy(reg):
case 232: // reg: ToDoubleTy(Constant):
// If this instruction has a parent (a user) in the tree
// and the user is translated as an FsMULd instruction,
// then the cast is unnecessary. So check that first.
// In the future, we'll want to do the same for the FdMULq instruction,
// so do the check here instead of only for ToFloatTy(reg).
//
if (subtreeRoot->parent() != NULL) {
const MachineCodeForInstruction& mcfi =
MachineCodeForInstruction::get(
cast<InstructionNode>(subtreeRoot->parent())->getInstruction());
if (mcfi.size() == 0 || mcfi.front()->getOpcode() == V9::FSMULD)
forwardOperandNum = 0; // forward first operand to user
}
if (forwardOperandNum != 0) { // we do need the cast
Value* leftVal = subtreeRoot->leftChild()->getValue();
const Type* opType = leftVal->getType();
MachineOpCode opCode=ChooseConvertToFloatInstr(target,
subtreeRoot->getOpLabel(), opType);
if (opCode == V9::NOP) { // no conversion needed
forwardOperandNum = 0; // forward first operand to user
} else {
// If the source operand is a non-FP type it must be
// first copied from int to float register via memory!
Instruction *dest = subtreeRoot->getInstruction();
Value* srcForCast;
int n = 0;
if (! opType->isFloatingPoint()) {
// Create a temporary to represent the FP register
// into which the integer will be copied via memory.
// The type of this temporary will determine the FP
// register used: single-prec for a 32-bit int or smaller,
// double-prec for a 64-bit int.
//
uint64_t srcSize =
target.getTargetData().getTypeSize(leftVal->getType());
Type* tmpTypeToUse =
(srcSize <= 4)? Type::FloatTy : Type::DoubleTy;
MachineCodeForInstruction &destMCFI =
MachineCodeForInstruction::get(dest);
srcForCast = new TmpInstruction(destMCFI, tmpTypeToUse, dest);
CreateCodeToCopyIntToFloat(target,
dest->getParent()->getParent(),
leftVal, cast<Instruction>(srcForCast),
mvec, destMCFI);
} else
srcForCast = leftVal;
M = BuildMI(opCode, 2).addReg(srcForCast).addRegDef(dest);
mvec.push_back(M);
}
}
break;
case 19: // reg: ToArrayTy(reg):
case 20: // reg: ToPointerTy(reg):
forwardOperandNum = 0; // forward first operand to user
break;
case 233: // reg: Add(reg, Constant)
maskUnsignedResult = true;
M = CreateAddConstInstruction(subtreeRoot);
if (M != NULL) {
mvec.push_back(M);
break;
}
// ELSE FALL THROUGH
case 33: // reg: Add(reg, reg)
maskUnsignedResult = true;
Add3OperandInstr(ChooseAddInstruction(subtreeRoot), subtreeRoot, mvec);
break;
case 234: // reg: Sub(reg, Constant)
maskUnsignedResult = true;
M = CreateSubConstInstruction(subtreeRoot);
if (M != NULL) {
mvec.push_back(M);
break;
}
// ELSE FALL THROUGH
case 34: // reg: Sub(reg, reg)
maskUnsignedResult = true;
Add3OperandInstr(ChooseSubInstructionByType(
subtreeRoot->getInstruction()->getType()),
subtreeRoot, mvec);
break;
case 135: // reg: Mul(todouble, todouble)
checkCast = true;
// FALL THROUGH
case 35: // reg: Mul(reg, reg)
{
maskUnsignedResult = true;
MachineOpCode forceOp = ((checkCast && BothFloatToDouble(subtreeRoot))
? (MachineOpCode)V9::FSMULD
: -1);
Instruction* mulInstr = subtreeRoot->getInstruction();
CreateMulInstruction(target, mulInstr->getParent()->getParent(),
subtreeRoot->leftChild()->getValue(),
subtreeRoot->rightChild()->getValue(),
mulInstr, mvec,
MachineCodeForInstruction::get(mulInstr),forceOp);
break;
}
case 335: // reg: Mul(todouble, todoubleConst)
checkCast = true;
// FALL THROUGH
case 235: // reg: Mul(reg, Constant)
{
maskUnsignedResult = true;
MachineOpCode forceOp = ((checkCast && BothFloatToDouble(subtreeRoot))
? (MachineOpCode)V9::FSMULD
: -1);
Instruction* mulInstr = subtreeRoot->getInstruction();
CreateMulInstruction(target, mulInstr->getParent()->getParent(),
subtreeRoot->leftChild()->getValue(),
subtreeRoot->rightChild()->getValue(),
mulInstr, mvec,
MachineCodeForInstruction::get(mulInstr),
forceOp);
break;
}
case 236: // reg: Div(reg, Constant)
maskUnsignedResult = true;
L = mvec.size();
CreateDivConstInstruction(target, subtreeRoot, mvec);
if (mvec.size() > L)
break;
// ELSE FALL THROUGH
case 36: // reg: Div(reg, reg)
{
maskUnsignedResult = true;
// If either operand of divide is smaller than 64 bits, we have
// to make sure the unused top bits are correct because they affect
// the result. These bits are already correct for unsigned values.
// They may be incorrect for signed values, so sign extend to fill in.
Instruction* divI = subtreeRoot->getInstruction();
Value* divOp1 = subtreeRoot->leftChild()->getValue();
Value* divOp2 = subtreeRoot->rightChild()->getValue();
Value* divOp1ToUse = divOp1;
Value* divOp2ToUse = divOp2;
if (divI->getType()->isSigned()) {
unsigned opSize=target.getTargetData().getTypeSize(divI->getType());
if (opSize < 8) {
MachineCodeForInstruction& mcfi=MachineCodeForInstruction::get(divI);
divOp1ToUse = new TmpInstruction(mcfi, divOp1);
divOp2ToUse = new TmpInstruction(mcfi, divOp2);
CreateSignExtensionInstructions(target,
divI->getParent()->getParent(),
divOp1, divOp1ToUse,
8*opSize, mvec, mcfi);
CreateSignExtensionInstructions(target,
divI->getParent()->getParent(),
divOp2, divOp2ToUse,
8*opSize, mvec, mcfi);
}
}
mvec.push_back(BuildMI(ChooseDivInstruction(target, subtreeRoot), 3)
.addReg(divOp1ToUse)
.addReg(divOp2ToUse)
.addRegDef(divI));
break;
}
case 37: // reg: Rem(reg, reg)
case 237: // reg: Rem(reg, Constant)
{
maskUnsignedResult = true;
Instruction* remI = subtreeRoot->getInstruction();
Value* divOp1 = subtreeRoot->leftChild()->getValue();
Value* divOp2 = subtreeRoot->rightChild()->getValue();
MachineCodeForInstruction& mcfi = MachineCodeForInstruction::get(remI);
// If second operand of divide is smaller than 64 bits, we have
// to make sure the unused top bits are correct because they affect
// the result. These bits are already correct for unsigned values.
// They may be incorrect for signed values, so sign extend to fill in.
//
Value* divOpToUse = divOp2;
if (divOp2->getType()->isSigned()) {
unsigned opSize=target.getTargetData().getTypeSize(divOp2->getType());
if (opSize < 8) {
divOpToUse = new TmpInstruction(mcfi, divOp2);
CreateSignExtensionInstructions(target,
remI->getParent()->getParent(),
divOp2, divOpToUse,
8*opSize, mvec, mcfi);
}
}
// Now compute: result = rem V1, V2 as:
// result = V1 - (V1 / signExtend(V2)) * signExtend(V2)
//
TmpInstruction* quot = new TmpInstruction(mcfi, divOp1, divOpToUse);
TmpInstruction* prod = new TmpInstruction(mcfi, quot, divOpToUse);
mvec.push_back(BuildMI(ChooseDivInstruction(target, subtreeRoot), 3)
.addReg(divOp1).addReg(divOpToUse).addRegDef(quot));
mvec.push_back(BuildMI(ChooseMulInstructionByType(remI->getType()), 3)
.addReg(quot).addReg(divOpToUse).addRegDef(prod));
mvec.push_back(BuildMI(ChooseSubInstructionByType(remI->getType()), 3)
.addReg(divOp1).addReg(prod).addRegDef(remI));
break;
}
case 38: // bool: And(bool, bool)
case 138: // bool: And(bool, not)
case 238: // bool: And(bool, boolconst)
case 338: // reg : BAnd(reg, reg)
case 538: // reg : BAnd(reg, Constant)
Add3OperandInstr(V9::ANDr, subtreeRoot, mvec);
break;
case 438: // bool: BAnd(bool, bnot)
{ // Use the argument of NOT as the second argument!
// Mark the NOT node so that no code is generated for it.
// If the type is boolean, set 1 or 0 in the result register.
InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
Value* notArg = BinaryOperator::getNotArgument(
cast<BinaryOperator>(notNode->getInstruction()));
notNode->markFoldedIntoParent();
Value *lhs = subtreeRoot->leftChild()->getValue();
Value *dest = subtreeRoot->getValue();
mvec.push_back(BuildMI(V9::ANDNr, 3).addReg(lhs).addReg(notArg)
.addReg(dest, MachineOperand::Def));
if (notArg->getType() == Type::BoolTy) {
// set 1 in result register if result of above is non-zero
mvec.push_back(BuildMI(V9::MOVRNZi, 3).addReg(dest).addZImm(1)
.addReg(dest, MachineOperand::UseAndDef));
}
break;
}
case 39: // bool: Or(bool, bool)
case 139: // bool: Or(bool, not)
case 239: // bool: Or(bool, boolconst)
case 339: // reg : BOr(reg, reg)
case 539: // reg : BOr(reg, Constant)
Add3OperandInstr(V9::ORr, subtreeRoot, mvec);
break;
case 439: // bool: BOr(bool, bnot)
{ // Use the argument of NOT as the second argument!
// Mark the NOT node so that no code is generated for it.
// If the type is boolean, set 1 or 0 in the result register.
InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
Value* notArg = BinaryOperator::getNotArgument(
cast<BinaryOperator>(notNode->getInstruction()));
notNode->markFoldedIntoParent();
Value *lhs = subtreeRoot->leftChild()->getValue();
Value *dest = subtreeRoot->getValue();
mvec.push_back(BuildMI(V9::ORNr, 3).addReg(lhs).addReg(notArg)
.addReg(dest, MachineOperand::Def));
if (notArg->getType() == Type::BoolTy) {
// set 1 in result register if result of above is non-zero
mvec.push_back(BuildMI(V9::MOVRNZi, 3).addReg(dest).addZImm(1)
.addReg(dest, MachineOperand::UseAndDef));
}
break;
}
case 40: // bool: Xor(bool, bool)
case 140: // bool: Xor(bool, not)
case 240: // bool: Xor(bool, boolconst)
case 340: // reg : BXor(reg, reg)
case 540: // reg : BXor(reg, Constant)
Add3OperandInstr(V9::XORr, subtreeRoot, mvec);
break;
case 440: // bool: BXor(bool, bnot)
{ // Use the argument of NOT as the second argument!
// Mark the NOT node so that no code is generated for it.
// If the type is boolean, set 1 or 0 in the result register.
InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
Value* notArg = BinaryOperator::getNotArgument(
cast<BinaryOperator>(notNode->getInstruction()));
notNode->markFoldedIntoParent();
Value *lhs = subtreeRoot->leftChild()->getValue();
Value *dest = subtreeRoot->getValue();
mvec.push_back(BuildMI(V9::XNORr, 3).addReg(lhs).addReg(notArg)
.addReg(dest, MachineOperand::Def));
if (notArg->getType() == Type::BoolTy) {
// set 1 in result register if result of above is non-zero
mvec.push_back(BuildMI(V9::MOVRNZi, 3).addReg(dest).addZImm(1)
.addReg(dest, MachineOperand::UseAndDef));
}
break;
}
case 41: // setCCconst: SetCC(reg, Constant)
{ // Comparison is with a constant:
//
// If the bool result must be computed into a register (see below),
// and the constant is int ZERO, we can use the MOVR[op] instructions
// and avoid the SUBcc instruction entirely.
// Otherwise this is just the same as case 42, so just fall through.
//
// The result of the SetCC must be computed and stored in a register if
// it is used outside the current basic block (so it must be computed
// as a boolreg) or it is used by anything other than a branch.
// We will use a conditional move to do this.
//
Instruction* setCCInstr = subtreeRoot->getInstruction();
bool computeBoolVal = (subtreeRoot->parent() == NULL ||
! AllUsesAreBranches(setCCInstr));
if (computeBoolVal) {
InstrTreeNode* constNode = subtreeRoot->rightChild();
assert(constNode &&
constNode->getNodeType() ==InstrTreeNode::NTConstNode);
Constant *constVal = cast<Constant>(constNode->getValue());
bool isValidConst;
if ((constVal->getType()->isInteger()
|| isa<PointerType>(constVal->getType()))
&& ConvertConstantToIntType(target,
constVal, constVal->getType(), isValidConst) == 0
&& isValidConst)
{
// That constant is an integer zero after all...
// Use a MOVR[op] to compute the boolean result
// Unconditionally set register to 0
mvec.push_back(BuildMI(V9::SETHI, 2).addZImm(0)
.addRegDef(setCCInstr));
// Now conditionally move 1 into the register.
// Mark the register as a use (as well as a def) because the old
// value will be retained if the condition is false.
MachineOpCode movOpCode = ChooseMovpregiForSetCC(subtreeRoot);
mvec.push_back(BuildMI(movOpCode, 3)
.addReg(subtreeRoot->leftChild()->getValue())
.addZImm(1)
.addReg(setCCInstr, MachineOperand::UseAndDef));
break;
}
}
// ELSE FALL THROUGH
}
case 42: // bool: SetCC(reg, reg):
{
// This generates a SUBCC instruction, putting the difference in a
// result reg. if needed, and/or setting a condition code if needed.
//
Instruction* setCCInstr = subtreeRoot->getInstruction();
Value* leftVal = subtreeRoot->leftChild()->getValue();
Value* rightVal = subtreeRoot->rightChild()->getValue();
const Type* opType = leftVal->getType();
bool isFPCompare = opType->isFloatingPoint();
// If the boolean result of the SetCC is used outside the current basic
// block (so it must be computed as a boolreg) or is used by anything
// other than a branch, the boolean must be computed and stored
// in a result register. We will use a conditional move to do this.
//
bool computeBoolVal = (subtreeRoot->parent() == NULL ||
! AllUsesAreBranches(setCCInstr));
// A TmpInstruction is created to represent the CC "result".
// Unlike other instances of TmpInstruction, this one is used
// by machine code of multiple LLVM instructions, viz.,
// the SetCC and the branch. Make sure to get the same one!
// Note that we do this even for FP CC registers even though they
// are explicit operands, because the type of the operand
// needs to be a floating point condition code, not an integer
// condition code. Think of this as casting the bool result to
// a FP condition code register.
// Later, we mark the 4th operand as being a CC register, and as a def.
//
TmpInstruction* tmpForCC = GetTmpForCC(setCCInstr,
setCCInstr->getParent()->getParent(),
leftVal->getType(),
MachineCodeForInstruction::get(setCCInstr));
// If the operands are signed values smaller than 4 bytes, then they
// must be sign-extended in order to do a valid 32-bit comparison
// and get the right result in the 32-bit CC register (%icc).
//
Value* leftOpToUse = leftVal;
Value* rightOpToUse = rightVal;
if (opType->isIntegral() && opType->isSigned()) {
unsigned opSize = target.getTargetData().getTypeSize(opType);
if (opSize < 4) {
MachineCodeForInstruction& mcfi =
MachineCodeForInstruction::get(setCCInstr);
// create temporary virtual regs. to hold the sign-extensions
leftOpToUse = new TmpInstruction(mcfi, leftVal);
rightOpToUse = new TmpInstruction(mcfi, rightVal);
// sign-extend each operand and put the result in the temporary reg.
CreateSignExtensionInstructions
(target, setCCInstr->getParent()->getParent(),
leftVal, leftOpToUse, 8*opSize, mvec, mcfi);
CreateSignExtensionInstructions
(target, setCCInstr->getParent()->getParent(),
rightVal, rightOpToUse, 8*opSize, mvec, mcfi);
}
}
if (! isFPCompare) {
// Integer condition: set CC and discard result.
mvec.push_back(BuildMI(V9::SUBccr, 4)
.addReg(leftOpToUse)
.addReg(rightOpToUse)
.addMReg(target.getRegInfo()->
getZeroRegNum(), MachineOperand::Def)
.addCCReg(tmpForCC, MachineOperand::Def));
} else {
// FP condition: dest of FCMP should be some FCCn register
mvec.push_back(BuildMI(ChooseFcmpInstruction(subtreeRoot), 3)
.addCCReg(tmpForCC, MachineOperand::Def)
.addReg(leftOpToUse)
.addReg(rightOpToUse));
}
if (computeBoolVal) {
MachineOpCode movOpCode = (isFPCompare
? ChooseMovFpcciInstruction(subtreeRoot)
: ChooseMovpcciForSetCC(subtreeRoot));
// Unconditionally set register to 0
M = BuildMI(V9::SETHI, 2).addZImm(0).addRegDef(setCCInstr);
mvec.push_back(M);
// Now conditionally move 1 into the register.
// Mark the register as a use (as well as a def) because the old
// value will be retained if the condition is false.
M = (BuildMI(movOpCode, 3).addCCReg(tmpForCC).addZImm(1)
.addReg(setCCInstr, MachineOperand::UseAndDef));
mvec.push_back(M);
}
break;
}
case 51: // reg: Load(reg)
case 52: // reg: Load(ptrreg)
SetOperandsForMemInstr(ChooseLoadInstruction(
subtreeRoot->getValue()->getType()),
mvec, subtreeRoot, target);
break;
case 55: // reg: GetElemPtr(reg)
case 56: // reg: GetElemPtrIdx(reg,reg)
// If the GetElemPtr was folded into the user (parent), it will be
// caught above. For other cases, we have to compute the address.
SetOperandsForMemInstr(V9::ADDr, mvec, subtreeRoot, target);
break;
case 57: // reg: Alloca: Implement as 1 instruction:
{ // add %fp, offsetFromFP -> result
AllocationInst* instr =
cast<AllocationInst>(subtreeRoot->getInstruction());
unsigned tsize =
target.getTargetData().getTypeSize(instr->getAllocatedType());
assert(tsize != 0);
CreateCodeForFixedSizeAlloca(target, instr, tsize, 1, mvec);
break;
}
case 58: // reg: Alloca(reg): Implement as 3 instructions:
// mul num, typeSz -> tmp
// sub %sp, tmp -> %sp
{ // add %sp, frameSizeBelowDynamicArea -> result
AllocationInst* instr =
cast<AllocationInst>(subtreeRoot->getInstruction());
const Type* eltType = instr->getAllocatedType();
// If #elements is constant, use simpler code for fixed-size allocas
int tsize = (int) target.getTargetData().getTypeSize(eltType);
Value* numElementsVal = NULL;
bool isArray = instr->isArrayAllocation();
if (!isArray || isa<Constant>(numElementsVal = instr->getArraySize())) {
// total size is constant: generate code for fixed-size alloca
unsigned numElements = isArray?
cast<ConstantUInt>(numElementsVal)->getValue() : 1;
CreateCodeForFixedSizeAlloca(target, instr, tsize,
numElements, mvec);
} else {
// total size is not constant.
CreateCodeForVariableSizeAlloca(target, instr, tsize,
numElementsVal, mvec);
}
break;
}
case 61: // reg: Call
{ // Generate a direct (CALL) or indirect (JMPL) call.
// Mark the return-address register, the indirection
// register (for indirect calls), the operands of the Call,
// and the return value (if any) as implicit operands
// of the machine instruction.
//
// If this is a varargs function, floating point arguments
// have to passed in integer registers so insert
// copy-float-to-int instructions for each float operand.
//
CallInst *callInstr = cast<CallInst>(subtreeRoot->getInstruction());
Value *callee = callInstr->getCalledValue();
Function* calledFunc = dyn_cast<Function>(callee);
// Check if this is an intrinsic function that needs a special code
// sequence (e.g., va_start). Indirect calls cannot be special.
//
bool specialIntrinsic = false;
Intrinsic::ID iid;
if (calledFunc && (iid=(Intrinsic::ID)calledFunc->getIntrinsicID()))
specialIntrinsic = CodeGenIntrinsic(iid, *callInstr, target, mvec);
// If not, generate the normal call sequence for the function.
// This can also handle any intrinsics that are just function calls.
//
if (! specialIntrinsic) {
Function* currentFunc = callInstr->getParent()->getParent();
MachineFunction& MF = MachineFunction::get(currentFunc);
MachineCodeForInstruction& mcfi =
MachineCodeForInstruction::get(callInstr);
const SparcV9RegInfo& regInfo =
(SparcV9RegInfo&) *target.getRegInfo();
const TargetFrameInfo& frameInfo = *target.getFrameInfo();
// Create hidden virtual register for return address with type void*
TmpInstruction* retAddrReg =
new TmpInstruction(mcfi, PointerType::get(Type::VoidTy), callInstr);
// Generate the machine instruction and its operands.
// Use CALL for direct function calls; this optimistically assumes
// the PC-relative address fits in the CALL address field (22 bits).
// Use JMPL for indirect calls.
// This will be added to mvec later, after operand copies.
//
MachineInstr* callMI;
if (calledFunc) // direct function call
callMI = BuildMI(V9::CALL, 1).addPCDisp(callee);
else // indirect function call
callMI = (BuildMI(V9::JMPLCALLi,3).addReg(callee)
.addSImm((int64_t)0).addRegDef(retAddrReg));
const FunctionType* funcType =
cast<FunctionType>(cast<PointerType>(callee->getType())
->getElementType());
bool isVarArgs = funcType->isVarArg();
bool noPrototype = isVarArgs && funcType->getNumParams() == 0;
// Use a descriptor to pass information about call arguments
// to the register allocator. This descriptor will be "owned"
// and freed automatically when the MachineCodeForInstruction
// object for the callInstr goes away.
CallArgsDescriptor* argDesc =
new CallArgsDescriptor(callInstr, retAddrReg,isVarArgs,noPrototype);
assert(callInstr->getOperand(0) == callee
&& "This is assumed in the loop below!");
// Insert sign-extension instructions for small signed values,
// if this is an unknown function (i.e., called via a funcptr)
// or an external one (i.e., which may not be compiled by llc).
//
if (calledFunc == NULL || calledFunc->isExternal()) {
for (unsigned i=1, N=callInstr->getNumOperands(); i < N; ++i) {
Value* argVal = callInstr->getOperand(i);
const Type* argType = argVal->getType();
if (argType->isIntegral() && argType->isSigned()) {
unsigned argSize = target.getTargetData().getTypeSize(argType);
if (argSize <= 4) {
// create a temporary virtual reg. to hold the sign-extension
TmpInstruction* argExtend = new TmpInstruction(mcfi, argVal);
// sign-extend argVal and put the result in the temporary reg.
CreateSignExtensionInstructions
(target, currentFunc, argVal, argExtend,
8*argSize, mvec, mcfi);
// replace argVal with argExtend in CallArgsDescriptor
argDesc->getArgInfo(i-1).replaceArgVal(argExtend);
}
}
}
}
// Insert copy instructions to get all the arguments into
// all the places that they need to be.
//
for (unsigned i=1, N=callInstr->getNumOperands(); i < N; ++i) {
int argNo = i-1;
CallArgInfo& argInfo = argDesc->getArgInfo(argNo);
Value* argVal = argInfo.getArgVal(); // don't use callInstr arg here
const Type* argType = argVal->getType();
unsigned regType = regInfo.getRegTypeForDataType(argType);
unsigned argSize = target.getTargetData().getTypeSize(argType);
int regNumForArg = SparcV9RegInfo::getInvalidRegNum();
unsigned regClassIDOfArgReg;
// Check for FP arguments to varargs functions.
// Any such argument in the first $K$ args must be passed in an
// integer register. If there is no prototype, it must also
// be passed as an FP register.
// K = #integer argument registers.
bool isFPArg = argVal->getType()->isFloatingPoint();
if (isVarArgs && isFPArg) {
if (noPrototype) {
// It is a function with no prototype: pass value
// as an FP value as well as a varargs value. The FP value
// may go in a register or on the stack. The copy instruction
// to the outgoing reg/stack is created by the normal argument
// handling code since this is the "normal" passing mode.
//
regNumForArg = regInfo.regNumForFPArg(regType,
false, false, argNo,
regClassIDOfArgReg);
if (regNumForArg == regInfo.getInvalidRegNum())
argInfo.setUseStackSlot();
else
argInfo.setUseFPArgReg();
}
// If this arg. is in the first $K$ regs, add special copy-
// float-to-int instructions to pass the value as an int.
// To check if it is in the first $K$, get the register
// number for the arg #i. These copy instructions are
// generated here because they are extra cases and not needed
// for the normal argument handling (some code reuse is
// possible though -- later).
//
int copyRegNum = regInfo.regNumForIntArg(false, false, argNo,
regClassIDOfArgReg);
if (copyRegNum != regInfo.getInvalidRegNum()) {
// Create a virtual register to represent copyReg. Mark
// this vreg as being an implicit operand of the call MI
const Type* loadTy = (argType == Type::FloatTy
? Type::IntTy : Type::LongTy);
TmpInstruction* argVReg = new TmpInstruction(mcfi, loadTy,
argVal, NULL,
"argRegCopy");
callMI->addImplicitRef(argVReg);
// Get a temp stack location to use to copy
// float-to-int via the stack.
//
// FIXME: For now, we allocate permanent space because
// the stack frame manager does not allow locals to be
// allocated (e.g., for alloca) after a temp is
// allocated!
//
// int tmpOffset = MF.getInfo<SparcV9FunctionInfo>()->pushTempValue(argSize);
int tmpOffset = MF.getInfo<SparcV9FunctionInfo>()->allocateLocalVar(argVReg);
// Generate the store from FP reg to stack
unsigned StoreOpcode = ChooseStoreInstruction(argType);
M = BuildMI(convertOpcodeFromRegToImm(StoreOpcode), 3)
.addReg(argVal).addMReg(regInfo.getFramePointer())
.addSImm(tmpOffset);
mvec.push_back(M);
// Generate the load from stack to int arg reg
unsigned LoadOpcode = ChooseLoadInstruction(loadTy);
M = BuildMI(convertOpcodeFromRegToImm(LoadOpcode), 3)
.addMReg(regInfo.getFramePointer()).addSImm(tmpOffset)
.addReg(argVReg, MachineOperand::Def);
// Mark operand with register it should be assigned
// both for copy and for the callMI
M->SetRegForOperand(M->getNumOperands()-1, copyRegNum);
callMI->SetRegForImplicitRef(callMI->getNumImplicitRefs()-1,
copyRegNum);
mvec.push_back(M);
// Add info about the argument to the CallArgsDescriptor
argInfo.setUseIntArgReg();
argInfo.setArgCopy(copyRegNum);
} else {
// Cannot fit in first $K$ regs so pass arg on stack
argInfo.setUseStackSlot();
}
} else if (isFPArg) {
// Get the outgoing arg reg to see if there is one.
regNumForArg = regInfo.regNumForFPArg(regType, false, false,
argNo, regClassIDOfArgReg);
if (regNumForArg == regInfo.getInvalidRegNum())
argInfo.setUseStackSlot();
else {
argInfo.setUseFPArgReg();
regNumForArg =regInfo.getUnifiedRegNum(regClassIDOfArgReg,
regNumForArg);
}
} else {
// Get the outgoing arg reg to see if there is one.
regNumForArg = regInfo.regNumForIntArg(false,false,
argNo, regClassIDOfArgReg);
if (regNumForArg == regInfo.getInvalidRegNum())
argInfo.setUseStackSlot();
else {
argInfo.setUseIntArgReg();
regNumForArg =regInfo.getUnifiedRegNum(regClassIDOfArgReg,
regNumForArg);
}
}
//
// Now insert copy instructions to stack slot or arg. register
//
if (argInfo.usesStackSlot()) {
// Get the stack offset for this argument slot.
// FP args on stack are right justified so adjust offset!
// int arguments are also right justified but they are
// always loaded as a full double-word so the offset does
// not need to be adjusted.
int argOffset = frameInfo.getOutgoingArgOffset(MF, argNo);
if (argType->isFloatingPoint()) {
unsigned slotSize = SparcV9FrameInfo::SizeOfEachArgOnStack;
assert(argSize <= slotSize && "Insufficient slot size!");
argOffset += slotSize - argSize;
}
// Now generate instruction to copy argument to stack
MachineOpCode storeOpCode =
(argType->isFloatingPoint()
? ((argSize == 4)? V9::STFi : V9::STDFi) : V9::STXi);
M = BuildMI(storeOpCode, 3).addReg(argVal)
.addMReg(regInfo.getStackPointer()).addSImm(argOffset);
mvec.push_back(M);
}
else if (regNumForArg != regInfo.getInvalidRegNum()) {
// Create a virtual register to represent the arg reg. Mark
// this vreg as being an implicit operand of the call MI.
TmpInstruction* argVReg =
new TmpInstruction(mcfi, argVal, NULL, "argReg");
callMI->addImplicitRef(argVReg);
// Generate the reg-to-reg copy into the outgoing arg reg.
// -- For FP values, create a FMOVS or FMOVD instruction
// -- For non-FP values, create an add-with-0 instruction
if (argType->isFloatingPoint())
M=(BuildMI(argType==Type::FloatTy? V9::FMOVS :V9::FMOVD,2)
.addReg(argVal).addReg(argVReg, MachineOperand::Def));
else
M = (BuildMI(ChooseAddInstructionByType(argType), 3)
.addReg(argVal).addSImm((int64_t) 0)
.addReg(argVReg, MachineOperand::Def));
// Mark the operand with the register it should be assigned
M->SetRegForOperand(M->getNumOperands()-1, regNumForArg);
callMI->SetRegForImplicitRef(callMI->getNumImplicitRefs()-1,
regNumForArg);
mvec.push_back(M);
}
else
assert(argInfo.getArgCopy() != regInfo.getInvalidRegNum() &&
"Arg. not in stack slot, primary or secondary register?");
}
// add call instruction and delay slot before copying return value
mvec.push_back(callMI);
mvec.push_back(BuildMI(V9::NOP, 0));
// Add the return value as an implicit ref. The call operands
// were added above. Also, add code to copy out the return value.
// This is always register-to-register for int or FP return values.
//
if (callInstr->getType() != Type::VoidTy) {
// Get the return value reg.
const Type* retType = callInstr->getType();
int regNum = (retType->isFloatingPoint()
? (unsigned) SparcV9FloatRegClass::f0
: (unsigned) SparcV9IntRegClass::o0);
unsigned regClassID = regInfo.getRegClassIDOfType(retType);
regNum = regInfo.getUnifiedRegNum(regClassID, regNum);
// Create a virtual register to represent it and mark
// this vreg as being an implicit operand of the call MI
TmpInstruction* retVReg =
new TmpInstruction(mcfi, callInstr, NULL, "argReg");
callMI->addImplicitRef(retVReg, /*isDef*/ true);
// Generate the reg-to-reg copy from the return value reg.
// -- For FP values, create a FMOVS or FMOVD instruction
// -- For non-FP values, create an add-with-0 instruction
if (retType->isFloatingPoint())
M = (BuildMI(retType==Type::FloatTy? V9::FMOVS : V9::FMOVD, 2)
.addReg(retVReg).addReg(callInstr, MachineOperand::Def));
else
M = (BuildMI(ChooseAddInstructionByType(retType), 3)
.addReg(retVReg).addSImm((int64_t) 0)
.addReg(callInstr, MachineOperand::Def));
// Mark the operand with the register it should be assigned
// Also mark the implicit ref of the call defining this operand
M->SetRegForOperand(0, regNum);
callMI->SetRegForImplicitRef(callMI->getNumImplicitRefs()-1,regNum);
mvec.push_back(M);
}
// For the CALL instruction, the ret. addr. reg. is also implicit
if (isa<Function>(callee))
callMI->addImplicitRef(retAddrReg, /*isDef*/ true);
MF.getInfo<SparcV9FunctionInfo>()->popAllTempValues(); // free temps used for this inst
}
break;
}
case 62: // reg: Shl(reg, reg)
{
Value* argVal1 = subtreeRoot->leftChild()->getValue();
Value* argVal2 = subtreeRoot->rightChild()->getValue();
Instruction* shlInstr = subtreeRoot->getInstruction();
const Type* opType = argVal1->getType();
assert((opType->isInteger() || isa<PointerType>(opType)) &&
"Shl unsupported for other types");
unsigned opSize = target.getTargetData().getTypeSize(opType);
CreateShiftInstructions(target, shlInstr->getParent()->getParent(),
(opSize > 4)? V9::SLLXr6:V9::SLLr5,
argVal1, argVal2, 0, shlInstr, mvec,
MachineCodeForInstruction::get(shlInstr));
break;
}
case 63: // reg: Shr(reg, reg)
{
const Type* opType = subtreeRoot->leftChild()->getValue()->getType();
assert((opType->isInteger() || isa<PointerType>(opType)) &&
"Shr unsupported for other types");
unsigned opSize = target.getTargetData().getTypeSize(opType);
Add3OperandInstr(opType->isSigned()
? (opSize > 4? V9::SRAXr6 : V9::SRAr5)
: (opSize > 4? V9::SRLXr6 : V9::SRLr5),
subtreeRoot, mvec);
break;
}
case 64: // reg: Phi(reg,reg)
break; // don't forward the value
case 65: // reg: VANext(reg): the va_next(va_list, type) instruction
{ // Increment the va_list pointer register according to the type.
// All LLVM argument types are <= 64 bits, so use one doubleword.
Instruction* vaNextI = subtreeRoot->getInstruction();
assert(target.getTargetData().getTypeSize(vaNextI->getType()) <= 8 &&
"We assumed that all LLVM parameter types <= 8 bytes!");
unsigned argSize = SparcV9FrameInfo::SizeOfEachArgOnStack;
mvec.push_back(BuildMI(V9::ADDi, 3).addReg(vaNextI->getOperand(0)).
addSImm(argSize).addRegDef(vaNextI));
break;
}
case 66: // reg: VAArg (reg): the va_arg instruction
{ // Load argument from stack using current va_list pointer value.
// Use 64-bit load for all non-FP args, and LDDF or double for FP.
Instruction* vaArgI = subtreeRoot->getInstruction();
MachineOpCode loadOp = (vaArgI->getType()->isFloatingPoint()
? (vaArgI->getType() == Type::FloatTy
? V9::LDFi : V9::LDDFi)
: V9::LDXi);
mvec.push_back(BuildMI(loadOp, 3).addReg(vaArgI->getOperand(0)).
addSImm(0).addRegDef(vaArgI));
break;
}
case 71: // reg: VReg
case 72: // reg: Constant
break; // don't forward the value
default:
assert(0 && "Unrecognized BURG rule");
break;
}
}
if (forwardOperandNum >= 0) {
// We did not generate a machine instruction but need to use operand.
// If user is in the same tree, replace Value in its machine operand.
// If not, insert a copy instruction which should get coalesced away
// by register allocation.
if (subtreeRoot->parent() != NULL)
ForwardOperand(subtreeRoot, subtreeRoot->parent(), forwardOperandNum);
else {
std::vector<MachineInstr*> minstrVec;
Instruction* instr = subtreeRoot->getInstruction();
CreateCopyInstructionsByType(target,
instr->getParent()->getParent(),
instr->getOperand(forwardOperandNum),
instr, minstrVec,
MachineCodeForInstruction::get(instr));
assert(minstrVec.size() > 0);
mvec.insert(mvec.end(), minstrVec.begin(), minstrVec.end());
}
}
if (maskUnsignedResult) {
// If result is unsigned and smaller than int reg size,
// we need to clear high bits of result value.
assert(forwardOperandNum < 0 && "Need mask but no instruction generated");
Instruction* dest = subtreeRoot->getInstruction();
if (dest->getType()->isUnsigned()) {
unsigned destSize=target.getTargetData().getTypeSize(dest->getType());
if (destSize <= 4) {
// Mask high 64 - N bits, where N = 4*destSize.
// Use a TmpInstruction to represent the
// intermediate result before masking. Since those instructions
// have already been generated, go back and substitute tmpI
// for dest in the result position of each one of them.
//
MachineCodeForInstruction& mcfi = MachineCodeForInstruction::get(dest);
TmpInstruction *tmpI = new TmpInstruction(mcfi, dest->getType(),
dest, NULL, "maskHi");
Value* srlArgToUse = tmpI;
unsigned numSubst = 0;
for (unsigned i=0, N=mvec.size(); i < N; ++i) {
// Make sure we substitute all occurrences of dest in these instrs.
// Otherwise, we will have bogus code.
bool someArgsWereIgnored = false;
// Make sure not to substitute an upwards-exposed use -- that would
// introduce a use of `tmpI' with no preceding def. Therefore,
// substitute a use or def-and-use operand only if a previous def
// operand has already been substituted (i.e., numSubst > 0).
//
numSubst += mvec[i]->substituteValue(dest, tmpI,
/*defsOnly*/ numSubst == 0,
/*notDefsAndUses*/ numSubst > 0,
someArgsWereIgnored);
assert(!someArgsWereIgnored &&
"Operand `dest' exists but not replaced: probably bogus!");
}
assert(numSubst > 0 && "Operand `dest' not replaced: probably bogus!");
// Left shift 32-N if size (N) is less than 32 bits.
// Use another tmp. virtual register to represent this result.
if (destSize < 4) {
srlArgToUse = new TmpInstruction(mcfi, dest->getType(),
tmpI, NULL, "maskHi2");
mvec.push_back(BuildMI(V9::SLLXi6, 3).addReg(tmpI)
.addZImm(8*(4-destSize))
.addReg(srlArgToUse, MachineOperand::Def));
}
// Logical right shift 32-N to get zero extension in top 64-N bits.
mvec.push_back(BuildMI(V9::SRLi5, 3).addReg(srlArgToUse)
.addZImm(8*(4-destSize))
.addReg(dest, MachineOperand::Def));
} else if (destSize < 8) {
assert(0 && "Unsupported type size: 32 < size < 64 bits");
}
}
}
}
} // End llvm namespace
//==------------------------------------------------------------------------==//
// Class V9ISel Implementation
//==------------------------------------------------------------------------==//
bool V9ISel::runOnFunction(Function &F) {
// First pass - Walk the function, lowering any calls to intrinsic functions
// which the instruction selector cannot handle.
for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; )
if (CallInst *CI = dyn_cast<CallInst>(I++))
if (Function *F = CI->getCalledFunction())
switch (F->getIntrinsicID()) {
case Intrinsic::not_intrinsic:
case Intrinsic::vastart:
case Intrinsic::vacopy:
case Intrinsic::vaend:
// We directly implement these intrinsics. Note that this knowledge
// is incestuously entangled with the code in
// SparcInstrSelection.cpp and must be updated when it is updated.
// Since ALL of the code in this library is incestuously intertwined
// with it already and sparc specific, we will live with this.
break;
default:
// All other intrinsic calls we must lower.
Instruction *Before = CI->getPrev();
Target.getIntrinsicLowering().LowerIntrinsicCall(CI);
if (Before) { // Move iterator to instruction after call
I = Before; ++I;
} else {
I = BB->begin();
}
}
// Build the instruction trees to be given as inputs to BURG.
InstrForest instrForest(&F);
if (SelectDebugLevel >= Select_DebugInstTrees) {
std::cerr << "\n\n*** Input to instruction selection for function "
<< F.getName() << "\n\n" << F
<< "\n\n*** Instruction trees for function "
<< F.getName() << "\n\n";
instrForest.dump();
}
// Invoke BURG instruction selection for each tree
for (InstrForest::const_root_iterator RI = instrForest.roots_begin();
RI != instrForest.roots_end(); ++RI) {
InstructionNode* basicNode = *RI;
assert(basicNode->parent() == NULL && "A `root' node has a parent?");
// Invoke BURM to label each tree node with a state
burm_label(basicNode);
if (SelectDebugLevel >= Select_DebugBurgTrees) {
printcover(basicNode, 1, 0);
std::cerr << "\nCover cost == " << treecost(basicNode, 1, 0) <<"\n\n";
printMatches(basicNode);
}
// Then recursively walk the tree to select instructions
SelectInstructionsForTree(basicNode, /*goalnt*/1);
}
// Create the MachineBasicBlocks and add all of the MachineInstrs
// defined in the MachineCodeForInstruction objects to the MachineBasicBlocks.
MachineFunction &MF = MachineFunction::get(&F);
std::map<const BasicBlock *, MachineBasicBlock *> MBBMap;
for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
MachineBasicBlock *MBB = new MachineBasicBlock(BI);
MF.getBasicBlockList().push_back(MBB);
MBBMap[BI] = MBB;
for (BasicBlock::iterator II = BI->begin(); II != BI->end(); ++II) {
MachineCodeForInstruction &mvec = MachineCodeForInstruction::get(II);
MBB->insert(MBB->end(), mvec.begin(), mvec.end());
}
}
// Initialize Machine-CFG for the function.
for (MachineFunction::iterator i = MF.begin (), e = MF.end (); i != e; ++i) {
MachineBasicBlock &MBB = *i;
const BasicBlock *BB = MBB.getBasicBlock ();
// for each successor S of BB, add MBBMap[S] as a successor of MBB.
for (succ_const_iterator si = succ_begin(BB), se = succ_end(BB); si != se;
++si) {
MachineBasicBlock *succMBB = MBBMap[*si];
assert (succMBB && "Can't find MachineBasicBlock for this successor");
MBB.addSuccessor (succMBB);
}
}
// Insert phi elimination code
InsertCodeForPhis(F);
if (SelectDebugLevel >= Select_PrintMachineCode) {
std::cerr << "\n*** Machine instructions after INSTRUCTION SELECTION\n";
MachineFunction::get(&F).dump();
}
return true;
}
/// InsertCodeForPhis - This method inserts Phi elimination code for
/// all Phi nodes in the given function. After this method is called,
/// the Phi nodes still exist in the LLVM code, but copies are added to the
/// machine code.
///
void V9ISel::InsertCodeForPhis(Function &F) {
// Iterate over every Phi node PN in F:
MachineFunction &MF = MachineFunction::get(&F);
for (MachineFunction::iterator BB = MF.begin(); BB != MF.end(); ++BB) {
for (BasicBlock::const_iterator IIt = BB->getBasicBlock()->begin();
const PHINode *PN = dyn_cast<PHINode>(IIt); ++IIt) {
// Create a new temporary register to hold the result of the Phi copy.
// The leak detector shouldn't track these nodes. They are not garbage,
// even though their parent field is never filled in.
Value *PhiCpRes = new PHINode(PN->getType(), PN->getName() + ":PhiCp");
LeakDetector::removeGarbageObject(PhiCpRes);
// For each of PN's incoming values, insert a copy in the corresponding
// predecessor block.
MachineCodeForInstruction &MCforPN = MachineCodeForInstruction::get (PN);
for (unsigned i = 0; i < PN->getNumIncomingValues(); ++i) {
std::vector<MachineInstr*> mvec, CpVec;
Target.getRegInfo()->cpValue2Value(PN->getIncomingValue(i),
PhiCpRes, mvec);
for (std::vector<MachineInstr*>::iterator MI=mvec.begin();
MI != mvec.end(); ++MI) {
std::vector<MachineInstr*> CpVec2 =
FixConstantOperandsForInstr(const_cast<PHINode*>(PN), *MI, Target);
CpVec2.push_back(*MI);
CpVec.insert(CpVec.end(), CpVec2.begin(), CpVec2.end());
}
// Insert the copy instructions into the predecessor BB.
InsertPhiElimInstructions(PN->getIncomingBlock(i), CpVec);
MCforPN.insert (MCforPN.end (), CpVec.begin (), CpVec.end ());
}
// Insert a copy instruction from PhiCpRes to PN.
std::vector<MachineInstr*> mvec;
Target.getRegInfo()->cpValue2Value(PhiCpRes, const_cast<PHINode*>(PN),
mvec);
BB->insert(BB->begin(), mvec.begin(), mvec.end());
MCforPN.insert (MCforPN.end (), mvec.begin (), mvec.end ());
} // for each Phi Instr in BB
} // for all BBs in function
}
/// InsertPhiElimInstructions - Inserts the instructions in CpVec into the
/// MachineBasicBlock corresponding to BB, just before its terminator
/// instruction. This is used by InsertCodeForPhis() to insert copies, above.
///
void V9ISel::InsertPhiElimInstructions(BasicBlock *BB,
const std::vector<MachineInstr*>& CpVec)
{
Instruction *TermInst = (Instruction*)BB->getTerminator();
MachineCodeForInstruction &MC4Term = MachineCodeForInstruction::get(TermInst);
MachineInstr *FirstMIOfTerm = MC4Term.front();
assert (FirstMIOfTerm && "No Machine Instrs for terminator");
MachineBasicBlock *MBB = FirstMIOfTerm->getParent();
assert(MBB && "Machine BB for predecessor's terminator not found");
MachineBasicBlock::iterator MCIt = FirstMIOfTerm;
assert(MCIt != MBB->end() && "Start inst of terminator not found");
// Insert the copy instructions just before the first machine instruction
// generated for the terminator.
MBB->insert(MCIt, CpVec.begin(), CpVec.end());
}
/// SelectInstructionsForTree - Recursively walk the tree to select
/// instructions. Do this top-down so that child instructions can exploit
/// decisions made at the child instructions.
///
/// E.g., if br(setle(reg,const)) decides the constant is 0 and uses
/// a branch-on-integer-register instruction, then the setle node
/// can use that information to avoid generating the SUBcc instruction.
///
/// Note that this cannot be done bottom-up because setle must do this
/// only if it is a child of the branch (otherwise, the result of setle
/// may be used by multiple instructions).
///
void V9ISel::SelectInstructionsForTree(InstrTreeNode* treeRoot, int goalnt) {
// Get the rule that matches this node.
int ruleForNode = burm_rule(treeRoot->state, goalnt);
if (ruleForNode == 0) {
std::cerr << "Could not match instruction tree for instr selection\n";
abort();
}
// Get this rule's non-terminals and the corresponding child nodes (if any)
short *nts = burm_nts[ruleForNode];
// First, select instructions for the current node and rule.
// (If this is a list node, not an instruction, then skip this step).
// This function is specific to the target architecture.
if (treeRoot->opLabel != VRegListOp) {
std::vector<MachineInstr*> minstrVec;
InstructionNode* instrNode = (InstructionNode*)treeRoot;
assert(instrNode->getNodeType() == InstrTreeNode::NTInstructionNode);
GetInstructionsByRule(instrNode, ruleForNode, nts, Target, minstrVec);
MachineCodeForInstruction &mvec =
MachineCodeForInstruction::get(instrNode->getInstruction());
mvec.insert(mvec.end(), minstrVec.begin(), minstrVec.end());
}
// Then, recursively compile the child nodes, if any.
//
if (nts[0]) {
// i.e., there is at least one kid
InstrTreeNode* kids[2];
int currentRule = ruleForNode;
burm_kids(treeRoot, currentRule, kids);
// First skip over any chain rules so that we don't visit
// the current node again.
while (ThisIsAChainRule(currentRule)) {
currentRule = burm_rule(treeRoot->state, nts[0]);
nts = burm_nts[currentRule];
burm_kids(treeRoot, currentRule, kids);
}
// Now we have the first non-chain rule so we have found
// the actual child nodes. Recursively compile them.
for (unsigned i = 0; nts[i]; i++) {
assert(i < 2);
InstrTreeNode::InstrTreeNodeType nodeType = kids[i]->getNodeType();
if (nodeType == InstrTreeNode::NTVRegListNode ||
nodeType == InstrTreeNode::NTInstructionNode)
SelectInstructionsForTree(kids[i], nts[i]);
}
}
// Finally, do any post-processing on this node after its children
// have been translated.
if (treeRoot->opLabel != VRegListOp)
PostprocessMachineCodeForTree((InstructionNode*)treeRoot, ruleForNode, nts);
}
/// PostprocessMachineCodeForTree - Apply any final cleanups to
/// machine code for the root of a subtree after selection for all its
/// children has been completed.
///
void V9ISel::PostprocessMachineCodeForTree(InstructionNode *instrNode,
int ruleForNode, short *nts) {
// Fix up any constant operands in the machine instructions to either
// use an immediate field or to load the constant into a register.
// Walk backwards and use direct indexes to allow insertion before current.
Instruction* vmInstr = instrNode->getInstruction();
MachineCodeForInstruction &mvec = MachineCodeForInstruction::get(vmInstr);
for (unsigned i = mvec.size(); i != 0; --i) {
std::vector<MachineInstr*> loadConstVec =
FixConstantOperandsForInstr(vmInstr, mvec[i-1], Target);
mvec.insert(mvec.begin()+i-1, loadConstVec.begin(), loadConstVec.end());
}
}
/// createSparcV9BurgInstSelector - Creates and returns a new SparcV9
/// BURG-based instruction selection pass.
///
FunctionPass *llvm::createSparcV9BurgInstSelector(TargetMachine &TM) {
return new V9ISel(TM);
}