Add support for bottom-up SLP vectorization infrastructure.
This commit adds the infrastructure for performing bottom-up SLP vectorization (and other optimizations) on parallel computations.
The infrastructure has three potential users:
1. The loop vectorizer needs to be able to vectorize AOS data structures such as (sum += A[i] + A[i+1]).
2. The BB-vectorizer needs this infrastructure for bottom-up SLP vectorization, because bottom-up vectorization is faster to compute.
3. A loop-roller needs to be able to analyze consecutive chains and roll them into a loop, in order to reduce code size. A loop roller does not need to create vector instructions, and this infrastructure separates the chain analysis from the vectorization.
This patch also includes a simple (100 LOC) bottom up SLP vectorizer that uses the infrastructure, and can vectorize this code:
void SAXPY(int *x, int *y, int a, int i) {
x[i] = a * x[i] + y[i];
x[i+1] = a * x[i+1] + y[i+1];
x[i+2] = a * x[i+2] + y[i+2];
x[i+3] = a * x[i+3] + y[i+3];
}
llvm-svn: 179117
diff --git a/llvm/lib/Transforms/Vectorize/VecUtils.cpp b/llvm/lib/Transforms/Vectorize/VecUtils.cpp
new file mode 100644
index 0000000..7e9f12d
--- /dev/null
+++ b/llvm/lib/Transforms/Vectorize/VecUtils.cpp
@@ -0,0 +1,439 @@
+//===- VecUtils.h --- Vectorization Utilities -----------------------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+#define DEBUG_TYPE "VecUtils"
+
+#include "VecUtils.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/ADT/SmallSet.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/Analysis/ScalarEvolution.h"
+#include "llvm/Analysis/ScalarEvolutionExpressions.h"
+#include "llvm/Analysis/TargetTransformInfo.h"
+#include "llvm/Analysis/Verifier.h"
+#include "llvm/IR/Constants.h"
+#include "llvm/IR/DataLayout.h"
+#include "llvm/IR/Function.h"
+#include "llvm/IR/Instructions.h"
+#include "llvm/IR/Module.h"
+#include "llvm/IR/Type.h"
+#include "llvm/IR/Value.h"
+#include "llvm/Pass.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/Target/TargetLibraryInfo.h"
+#include "llvm/Transforms/Scalar.h"
+#include "llvm/Transforms/Utils/Local.h"
+#include <algorithm>
+#include <map>
+
+using namespace llvm;
+
+namespace llvm {
+
+BoUpSLP::BoUpSLP(BasicBlock *Bb, ScalarEvolution *S, DataLayout *Dl,
+ TargetTransformInfo *Tti, AliasAnalysis *Aa) :
+ BB(Bb), SE(S), DL(Dl), TTI(Tti), AA(Aa) {
+ numberInstructions();
+}
+
+void BoUpSLP::numberInstructions() {
+ int Loc = 0;
+ InstrIdx.clear();
+ InstrVec.clear();
+ // Number the instructions in the block.
+ for (BasicBlock::iterator it=BB->begin(), e=BB->end(); it != e; ++it) {
+ InstrIdx[it] = Loc++;
+ InstrVec.push_back(it);
+ assert(InstrVec[InstrIdx[it]] == it && "Invalid allocation");
+ }
+}
+
+Value *BoUpSLP::getPointerOperand(Value *I) {
+ if (LoadInst *LI = dyn_cast<LoadInst>(I)) return LI->getPointerOperand();
+ if (StoreInst *SI = dyn_cast<StoreInst>(I)) return SI->getPointerOperand();
+ return 0;
+}
+
+unsigned BoUpSLP::getAddressSpaceOperand(Value *I) {
+ if (LoadInst *L=dyn_cast<LoadInst>(I)) return L->getPointerAddressSpace();
+ if (StoreInst *S=dyn_cast<StoreInst>(I)) return S->getPointerAddressSpace();
+ return -1;
+}
+
+bool BoUpSLP::isConsecutiveAccess(Value *A, Value *B) {
+ Value *PtrA = getPointerOperand(A);
+ Value *PtrB = getPointerOperand(B);
+ unsigned ASA = getAddressSpaceOperand(A);
+ unsigned ASB = getAddressSpaceOperand(B);
+
+ // Check that the address spaces match and that the pointers are valid.
+ if (!PtrA || !PtrB || (ASA != ASB)) return false;
+
+ // Check that A and B are of the same type.
+ if (PtrA->getType() != PtrB->getType()) return false;
+
+ // Calculate the distance.
+ const SCEV *PtrSCEVA = SE->getSCEV(PtrA);
+ const SCEV *PtrSCEVB = SE->getSCEV(PtrB);
+ const SCEV *OffsetSCEV = SE->getMinusSCEV(PtrSCEVA, PtrSCEVB);
+ const SCEVConstant *ConstOffSCEV = dyn_cast<SCEVConstant>(OffsetSCEV);
+
+ // Non constant distance.
+ if (!ConstOffSCEV) return false;
+
+ unsigned Offset = ConstOffSCEV->getValue()->getSExtValue();
+ Type *Ty = cast<PointerType>(PtrA->getType())->getElementType();
+ // The Instructions are connsecutive if the size of the first load/store is
+ // the same as the offset.
+ unsigned Sz = (DL ? DL->getTypeStoreSize(Ty) : Ty->getScalarSizeInBits()/8);
+ return ((-Offset) == Sz);
+}
+
+bool BoUpSLP::vectorizeStores(StoreList &Stores, int costThreshold) {
+ ValueSet Heads, Tails;
+ SmallDenseMap<Value*, Value*> ConsecutiveChain;
+ bool Changed = false;
+
+ // Do a quadratic search on all of the given stores and find
+ // all of the pairs of loads that follow each other.
+ for (unsigned i = 0, e = Stores.size(); i < e; ++i)
+ for (unsigned j = 0; j < e; ++j) {
+ if (i == j) continue;
+ if (isConsecutiveAccess(Stores[i], Stores[j])) {
+ Tails.insert(Stores[j]);
+ Heads.insert(Stores[i]);
+ ConsecutiveChain[Stores[i]] = Stores[j];
+ }
+ }
+
+ // For stores that start but don't end a link in the chain:
+ for (ValueSet::iterator it = Heads.begin(), e = Heads.end();it != e; ++it) {
+ if (Tails.count(*it)) continue;
+
+ // We found a store instr that starts a chain. Now follow the chain and try
+ // to vectorize it.
+ ValueList Operands;
+ Value *I = *it;
+ int MinCost = 0, MinVF = 0;
+ while (Tails.count(I) || Heads.count(I)) {
+ Operands.push_back(I);
+ unsigned VF = Operands.size();
+ if (isPowerOf2_32(VF) && VF > 1) {
+ int cost = getTreeRollCost(Operands, 0);
+ DEBUG(dbgs() << "Found cost=" << cost << " for VF=" << VF << "\n");
+ if (cost < MinCost) { MinCost = cost; MinVF = VF; }
+ }
+ // Move to the next value in the chain.
+ I = ConsecutiveChain[I];
+ }
+
+ if (MinCost <= costThreshold && MinVF > 1) {
+ DEBUG(dbgs() << "Decided to vectorize cost=" << MinCost << "\n");
+ vectorizeTree(Operands, MinVF);
+ Stores.clear();
+ // The current numbering is invalid because we added and removed instrs.
+ numberInstructions();
+ Changed = true;
+ }
+ }
+
+ return Changed;
+}
+
+int BoUpSLP::getScalarizationCost(Type *Ty) {
+ int Cost = 0;
+ for (unsigned i = 0, e = cast<VectorType>(Ty)->getNumElements(); i < e; ++i)
+ Cost += TTI->getVectorInstrCost(Instruction::InsertElement, Ty, i);
+ return Cost;
+}
+
+AliasAnalysis::Location BoUpSLP::getLocation(Instruction *I) {
+ if (StoreInst *SI = dyn_cast<StoreInst>(I)) return AA->getLocation(SI);
+ if (LoadInst *LI = dyn_cast<LoadInst>(I)) return AA->getLocation(LI);
+ return AliasAnalysis::Location();
+}
+
+Value *BoUpSLP::isUnsafeToSink(Instruction *Src, Instruction *Dst) {
+ assert(Src->getParent() == Dst->getParent() && "Not the same BB");
+ BasicBlock::iterator I = Src, E = Dst;
+ /// Scan all of the instruction from SRC to DST and check if
+ /// the source may alias.
+ for (++I; I != E; ++I) {
+ // Ignore store instructions that are marked as 'ignore'.
+ if (MemBarrierIgnoreList.count(I)) continue;
+ if (Src->mayWriteToMemory()) /* Write */ {
+ if (!I->mayReadOrWriteMemory()) continue;
+ } else /* Read */ {
+ if (!I->mayWriteToMemory()) continue;
+ }
+ AliasAnalysis::Location A = getLocation(&*I);
+ AliasAnalysis::Location B = getLocation(Src);
+
+ if (!A.Ptr || !B.Ptr || AA->alias(A, B))
+ return I;
+ }
+ return 0;
+}
+
+int BoUpSLP::getTreeRollCost(ValueList &VL, unsigned Depth) {
+ if (Depth == 6) return max_cost;
+ Type *ScalarTy = VL[0]->getType();
+
+ if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
+ ScalarTy = SI->getValueOperand()->getType();
+
+ /// Don't mess with vectors.
+ if (ScalarTy->isVectorTy()) return max_cost;
+
+ VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
+
+ // Check if all of the operands are constants.
+ bool AllConst = true;
+ bool AllSameScalar = true;
+ for (unsigned i = 0, e = VL.size(); i < e; ++i) {
+ AllConst &= isa<Constant>(VL[i]);
+ AllSameScalar &= (VL[0] == VL[i]);
+ // Must have a single use.
+ Instruction *I = dyn_cast<Instruction>(VL[i]);
+ // Need to scalarize instructions with multiple users or from other BBs.
+ if (I && ((I->getNumUses() > 1) || (I->getParent() != BB)))
+ return getScalarizationCost(VecTy);
+ }
+
+ // Is this a simple vector constant.
+ if (AllConst) return 0;
+
+ // If all of the operands are identical we can broadcast them.
+ if (AllSameScalar)
+ return TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy, 0);
+
+ // Scalarize unknown structures.
+ Instruction *VL0 = dyn_cast<Instruction>(VL[0]);
+ if (!VL0) return getScalarizationCost(VecTy);
+ assert(VL0->getParent() == BB && "Wrong BB");
+
+ unsigned Opcode = VL0->getOpcode();
+ for (unsigned i = 0, e = VL.size(); i < e; ++i) {
+ Instruction *I = dyn_cast<Instruction>(VL[i]);
+ // If not all of the instructions are identical then we have to scalarize.
+ if (!I || Opcode != I->getOpcode()) return getScalarizationCost(VecTy);
+ }
+
+ // Check if it is safe to sink the loads or the stores.
+ if (Opcode == Instruction::Load || Opcode == Instruction::Store) {
+ int MaxIdx = InstrIdx[VL0];
+ for (unsigned i = 1, e = VL.size(); i < e; ++i )
+ MaxIdx = std::max(MaxIdx, InstrIdx[VL[i]]);
+
+ Instruction *Last = InstrVec[MaxIdx];
+ for (unsigned i = 0, e = VL.size(); i < e; ++i ) {
+ if (VL[i] == Last) continue;
+ Value *Barrier = isUnsafeToSink(cast<Instruction>(VL[i]), Last);
+ if (Barrier) {
+ DEBUG(dbgs() << "LR: Can't sink " << *VL[i] << "\n down to " <<
+ *Last << "\n because of " << *Barrier << "\n");
+ return max_cost;
+ }
+ }
+ }
+
+ switch (Opcode) {
+ case Instruction::Add:
+ case Instruction::FAdd:
+ case Instruction::Sub:
+ case Instruction::FSub:
+ case Instruction::Mul:
+ case Instruction::FMul:
+ case Instruction::UDiv:
+ case Instruction::SDiv:
+ case Instruction::FDiv:
+ case Instruction::URem:
+ case Instruction::SRem:
+ case Instruction::FRem:
+ case Instruction::Shl:
+ case Instruction::LShr:
+ case Instruction::AShr:
+ case Instruction::And:
+ case Instruction::Or:
+ case Instruction::Xor: {
+ ValueList Operands;
+ int Cost = 0;
+ // Calculate the cost of all of the operands.
+ for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
+ // Prepare the operand vector.
+ for (unsigned j = 0; j < VL.size(); ++j)
+ Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
+ Cost += getTreeRollCost(Operands, Depth+1);
+ Operands.clear();
+ }
+
+ // Calculate the cost of this instruction.
+ int ScalarCost = VecTy->getNumElements() *
+ TTI->getArithmeticInstrCost(Opcode, ScalarTy);
+ int VecCost = TTI->getArithmeticInstrCost(Opcode, VecTy);
+ Cost += (VecCost - ScalarCost);
+ return Cost;
+ }
+ case Instruction::Load: {
+ // If we are scalarize the loads, add the cost of forming the vector.
+ for (unsigned i = 0, e = VL.size()-1; i < e; ++i)
+ if (!isConsecutiveAccess(VL[i], VL[i+1]))
+ return getScalarizationCost(VecTy);
+
+ // Cost of wide load - cost of scalar loads.
+ int ScalarLdCost = VecTy->getNumElements() *
+ TTI->getMemoryOpCost(Instruction::Load, ScalarTy, 1, 0);
+ int VecLdCost = TTI->getMemoryOpCost(Instruction::Load, ScalarTy, 1, 0);
+ return VecLdCost - ScalarLdCost;
+ }
+ case Instruction::Store: {
+ // We know that we can merge the stores. Calculate the cost.
+ int ScalarStCost = VecTy->getNumElements() *
+ TTI->getMemoryOpCost(Instruction::Store, ScalarTy, 1, 0);
+ int VecStCost = TTI->getMemoryOpCost(Instruction::Store, ScalarTy, 1,0);
+ int StoreCost = VecStCost - ScalarStCost;
+
+ ValueList Operands;
+ for (unsigned j = 0; j < VL.size(); ++j) {
+ Operands.push_back(cast<Instruction>(VL[j])->getOperand(0));
+ MemBarrierIgnoreList.insert(VL[j]);
+ }
+
+ int TotalCost = StoreCost + getTreeRollCost(Operands, Depth + 1);
+ MemBarrierIgnoreList.clear();
+ return TotalCost;
+ }
+ default:
+ // Unable to vectorize unknown instructions.
+ return getScalarizationCost(VecTy);
+ }
+}
+
+Instruction *BoUpSLP::GetLastInstr(ValueList &VL, unsigned VF) {
+ int MaxIdx = InstrIdx[BB->getFirstNonPHI()];
+ for (unsigned i = 0; i < VF; ++i )
+ MaxIdx = std::max(MaxIdx, InstrIdx[VL[i]]);
+ return InstrVec[MaxIdx + 1];
+}
+
+Value *BoUpSLP::Scalarize(ValueList &VL, VectorType *Ty) {
+ IRBuilder<> Builder(GetLastInstr(VL, Ty->getNumElements()));
+ Value *Vec = UndefValue::get(Ty);
+ for (unsigned i=0; i < Ty->getNumElements(); ++i)
+ Vec = Builder.CreateInsertElement(Vec, VL[i], Builder.getInt32(i));
+ return Vec;
+}
+
+Value *BoUpSLP::vectorizeTree(ValueList &VL, int VF) {
+ Type *ScalarTy = VL[0]->getType();
+ if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
+ ScalarTy = SI->getValueOperand()->getType();
+ VectorType *VecTy = VectorType::get(ScalarTy, VF);
+
+ // Check if all of the operands are constants or identical.
+ bool AllConst = true;
+ bool AllSameScalar = true;
+ for (unsigned i = 0, e = VF; i < e; ++i) {
+ AllConst &= !!dyn_cast<Constant>(VL[i]);
+ AllSameScalar &= (VL[0] == VL[i]);
+ // Must have a single use.
+ Instruction *I = dyn_cast<Instruction>(VL[i]);
+ if (I && (I->getNumUses() > 1 || I->getParent() != BB))
+ return Scalarize(VL, VecTy);
+ }
+
+ // Is this a simple vector constant.
+ if (AllConst || AllSameScalar) return Scalarize(VL, VecTy);
+
+ // Scalarize unknown structures.
+ Instruction *VL0 = dyn_cast<Instruction>(VL[0]);
+ if (!VL0) return Scalarize(VL, VecTy);
+
+ unsigned Opcode = VL0->getOpcode();
+ for (unsigned i = 0, e = VF; i < e; ++i) {
+ Instruction *I = dyn_cast<Instruction>(VL[i]);
+ // If not all of the instructions are identical then we have to scalarize.
+ if (!I || Opcode != I->getOpcode()) return Scalarize(VL, VecTy);
+ }
+
+ switch (Opcode) {
+ case Instruction::Add:
+ case Instruction::FAdd:
+ case Instruction::Sub:
+ case Instruction::FSub:
+ case Instruction::Mul:
+ case Instruction::FMul:
+ case Instruction::UDiv:
+ case Instruction::SDiv:
+ case Instruction::FDiv:
+ case Instruction::URem:
+ case Instruction::SRem:
+ case Instruction::FRem:
+ case Instruction::Shl:
+ case Instruction::LShr:
+ case Instruction::AShr:
+ case Instruction::And:
+ case Instruction::Or:
+ case Instruction::Xor: {
+ ValueList LHSVL, RHSVL;
+ for (int i = 0; i < VF; ++i) {
+ RHSVL.push_back(cast<Instruction>(VL[i])->getOperand(0));
+ LHSVL.push_back(cast<Instruction>(VL[i])->getOperand(1));
+ }
+
+ Value *RHS = vectorizeTree(RHSVL, VF);
+ Value *LHS = vectorizeTree(LHSVL, VF);
+ IRBuilder<> Builder(GetLastInstr(VL, VF));
+ BinaryOperator *BinOp = dyn_cast<BinaryOperator>(VL0);
+ return Builder.CreateBinOp(BinOp->getOpcode(), RHS,LHS);
+ }
+ case Instruction::Load: {
+ LoadInst *LI = dyn_cast<LoadInst>(VL0);
+ unsigned Alignment = LI->getAlignment();
+
+ // Check if all of the loads are consecutive.
+ for (unsigned i = 1, e = VF; i < e; ++i)
+ if (!isConsecutiveAccess(VL[i-1], VL[i]))
+ return Scalarize(VL, VecTy);
+
+ IRBuilder<> Builder(GetLastInstr(VL, VF));
+ Value *VecPtr = Builder.CreateBitCast(LI->getPointerOperand(),
+ VecTy->getPointerTo());
+ LI = Builder.CreateLoad(VecPtr);
+ LI->setAlignment(Alignment);
+ return LI;
+ }
+ case Instruction::Store: {
+ StoreInst *SI = dyn_cast<StoreInst>(VL0);
+ unsigned Alignment = SI->getAlignment();
+
+ ValueList ValueOp;
+ for (int i = 0; i < VF; ++i)
+ ValueOp.push_back(cast<StoreInst>(VL[i])->getValueOperand());
+
+ Value *VecValue = vectorizeTree(ValueOp, VF);
+
+ IRBuilder<> Builder(GetLastInstr(VL, VF));
+ Value *VecPtr = Builder.CreateBitCast(SI->getPointerOperand(),
+ VecTy->getPointerTo());
+ Builder.CreateStore(VecValue, VecPtr)->setAlignment(Alignment);
+
+ for (int i = 0; i < VF; ++i)
+ cast<Instruction>(VL[i])->eraseFromParent();
+ return 0;
+ }
+ default:
+ return Scalarize(VL, VecTy);
+ }
+}
+
+} // end of namespace