1 //===- VecUtils.h --- Vectorization Utilities -----------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
9 #define DEBUG_TYPE "SLP"
12 #include "llvm/ADT/DenseMap.h"
13 #include "llvm/ADT/SmallPtrSet.h"
14 #include "llvm/ADT/SmallSet.h"
15 #include "llvm/ADT/SmallVector.h"
16 #include "llvm/Analysis/AliasAnalysis.h"
17 #include "llvm/Analysis/ScalarEvolution.h"
18 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
19 #include "llvm/Analysis/TargetTransformInfo.h"
20 #include "llvm/Analysis/Verifier.h"
21 #include "llvm/IR/Constants.h"
22 #include "llvm/IR/DataLayout.h"
23 #include "llvm/IR/Function.h"
24 #include "llvm/IR/Instructions.h"
25 #include "llvm/IR/Module.h"
26 #include "llvm/IR/Type.h"
27 #include "llvm/IR/Value.h"
28 #include "llvm/Pass.h"
29 #include "llvm/Support/CommandLine.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/raw_ostream.h"
32 #include "llvm/Target/TargetLibraryInfo.h"
33 #include "llvm/Transforms/Scalar.h"
34 #include "llvm/Transforms/Utils/Local.h"
40 static const unsigned MinVecRegSize = 128;
42 static const unsigned RecursionMaxDepth = 6;
46 BoUpSLP::BoUpSLP(BasicBlock *Bb, ScalarEvolution *S, DataLayout *Dl,
47 TargetTransformInfo *Tti, AliasAnalysis *Aa) :
48 BB(Bb), SE(S), DL(Dl), TTI(Tti), AA(Aa) {
52 void BoUpSLP::numberInstructions() {
56 // Number the instructions in the block.
57 for (BasicBlock::iterator it=BB->begin(), e=BB->end(); it != e; ++it) {
59 InstrVec.push_back(it);
60 assert(InstrVec[InstrIdx[it]] == it && "Invalid allocation");
64 Value *BoUpSLP::getPointerOperand(Value *I) {
65 if (LoadInst *LI = dyn_cast<LoadInst>(I)) return LI->getPointerOperand();
66 if (StoreInst *SI = dyn_cast<StoreInst>(I)) return SI->getPointerOperand();
70 unsigned BoUpSLP::getAddressSpaceOperand(Value *I) {
71 if (LoadInst *L=dyn_cast<LoadInst>(I)) return L->getPointerAddressSpace();
72 if (StoreInst *S=dyn_cast<StoreInst>(I)) return S->getPointerAddressSpace();
76 bool BoUpSLP::isConsecutiveAccess(Value *A, Value *B) {
77 Value *PtrA = getPointerOperand(A);
78 Value *PtrB = getPointerOperand(B);
79 unsigned ASA = getAddressSpaceOperand(A);
80 unsigned ASB = getAddressSpaceOperand(B);
82 // Check that the address spaces match and that the pointers are valid.
83 if (!PtrA || !PtrB || (ASA != ASB)) return false;
85 // Check that A and B are of the same type.
86 if (PtrA->getType() != PtrB->getType()) return false;
88 // Calculate the distance.
89 const SCEV *PtrSCEVA = SE->getSCEV(PtrA);
90 const SCEV *PtrSCEVB = SE->getSCEV(PtrB);
91 const SCEV *OffsetSCEV = SE->getMinusSCEV(PtrSCEVA, PtrSCEVB);
92 const SCEVConstant *ConstOffSCEV = dyn_cast<SCEVConstant>(OffsetSCEV);
94 // Non constant distance.
95 if (!ConstOffSCEV) return false;
97 unsigned Offset = ConstOffSCEV->getValue()->getSExtValue();
98 Type *Ty = cast<PointerType>(PtrA->getType())->getElementType();
99 // The Instructions are connsecutive if the size of the first load/store is
100 // the same as the offset.
101 unsigned Sz = DL->getTypeStoreSize(Ty);
102 return ((-Offset) == Sz);
105 bool BoUpSLP::vectorizeStoreChain(ValueList &Chain, int CostThreshold) {
106 Type *StoreTy = cast<StoreInst>(Chain[0])->getValueOperand()->getType();
107 unsigned Sz = DL->getTypeSizeInBits(StoreTy);
108 unsigned VF = MinVecRegSize / Sz;
110 if (!isPowerOf2_32(Sz) || VF < 2) return false;
112 bool Changed = false;
113 // Look for profitable vectorizable trees at all offsets, starting at zero.
114 for (unsigned i = 0, e = Chain.size(); i < e; ++i) {
115 if (i + VF > e) return Changed;
116 DEBUG(dbgs()<<"SLP: Analyzing " << VF << " stores at offset "<< i << "\n");
117 ValueList Operands(&Chain[i], &Chain[i] + VF);
119 int Cost = getTreeCost(Operands);
120 DEBUG(dbgs() << "SLP: Found cost=" << Cost << " for VF=" << VF << "\n");
121 if (Cost < CostThreshold) {
122 DEBUG(dbgs() << "SLP: Decided to vectorize cost=" << Cost << "\n");
123 vectorizeTree(Operands, VF);
132 bool BoUpSLP::vectorizeStores(StoreList &Stores, int costThreshold) {
133 ValueSet Heads, Tails;
134 SmallDenseMap<Value*, Value*> ConsecutiveChain;
136 // We may run into multiple chains that merge into a single chain. We mark the
137 // stores that we vectorized so that we don't visit the same store twice.
138 ValueSet VectorizedStores;
139 bool Changed = false;
141 // Do a quadratic search on all of the given stores and find
142 // all of the pairs of loads that follow each other.
143 for (unsigned i = 0, e = Stores.size(); i < e; ++i)
144 for (unsigned j = 0; j < e; ++j) {
145 if (i == j) continue;
146 if (isConsecutiveAccess(Stores[i], Stores[j])) {
147 Tails.insert(Stores[j]);
148 Heads.insert(Stores[i]);
149 ConsecutiveChain[Stores[i]] = Stores[j];
153 // For stores that start but don't end a link in the chain:
154 for (ValueSet::iterator it = Heads.begin(), e = Heads.end();it != e; ++it) {
155 if (Tails.count(*it)) continue;
157 // We found a store instr that starts a chain. Now follow the chain and try
161 // Collect the chain into a list.
162 while (Tails.count(I) || Heads.count(I)) {
163 if (VectorizedStores.count(I)) break;
164 Operands.push_back(I);
165 // Move to the next value in the chain.
166 I = ConsecutiveChain[I];
169 bool Vectorized = vectorizeStoreChain(Operands, costThreshold);
171 // Mark the vectorized stores so that we don't vectorize them again.
173 VectorizedStores.insert(Operands.begin(), Operands.end());
174 Changed |= Vectorized;
180 int BoUpSLP::getScalarizationCost(ValueList &VL) {
181 // Find the type of the operands in VL.
182 Type *ScalarTy = VL[0]->getType();
183 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
184 ScalarTy = SI->getValueOperand()->getType();
185 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
186 // Find the cost of inserting/extracting values from the vector.
187 return getScalarizationCost(VecTy);
190 int BoUpSLP::getScalarizationCost(Type *Ty) {
192 for (unsigned i = 0, e = cast<VectorType>(Ty)->getNumElements(); i < e; ++i)
193 Cost += TTI->getVectorInstrCost(Instruction::InsertElement, Ty, i);
197 AliasAnalysis::Location BoUpSLP::getLocation(Instruction *I) {
198 if (StoreInst *SI = dyn_cast<StoreInst>(I)) return AA->getLocation(SI);
199 if (LoadInst *LI = dyn_cast<LoadInst>(I)) return AA->getLocation(LI);
200 return AliasAnalysis::Location();
203 Value *BoUpSLP::isUnsafeToSink(Instruction *Src, Instruction *Dst) {
204 assert(Src->getParent() == Dst->getParent() && "Not the same BB");
205 BasicBlock::iterator I = Src, E = Dst;
206 /// Scan all of the instruction from SRC to DST and check if
207 /// the source may alias.
208 for (++I; I != E; ++I) {
209 // Ignore store instructions that are marked as 'ignore'.
210 if (MemBarrierIgnoreList.count(I)) continue;
211 if (Src->mayWriteToMemory()) /* Write */ {
212 if (!I->mayReadOrWriteMemory()) continue;
214 if (!I->mayWriteToMemory()) continue;
216 AliasAnalysis::Location A = getLocation(&*I);
217 AliasAnalysis::Location B = getLocation(Src);
219 if (!A.Ptr || !B.Ptr || AA->alias(A, B))
225 void BoUpSLP::vectorizeArith(ValueList &Operands) {
226 Value *Vec = vectorizeTree(Operands, Operands.size());
227 BasicBlock::iterator Loc = cast<Instruction>(Vec);
228 IRBuilder<> Builder(++Loc);
229 // After vectorizing the operands we need to generate extractelement
230 // instructions and replace all of the uses of the scalar values with
231 // the values that we extracted from the vectorized tree.
232 for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
233 Value *S = Builder.CreateExtractElement(Vec, Builder.getInt32(i));
234 Operands[i]->replaceAllUsesWith(S);
238 int BoUpSLP::getTreeCost(ValueList &VL) {
239 // Get rid of the list of stores that were removed, and from the
240 // lists of instructions with multiple users.
241 MemBarrierIgnoreList.clear();
243 MultiUserVals.clear();
244 MustScalarize.clear();
246 // Scan the tree and find which value is used by which lane, and which values
247 // must be scalarized.
248 getTreeUses_rec(VL, 0);
250 // Check that instructions with multiple users can be vectorized. Mark unsafe
252 for (ValueSet::iterator it = MultiUserVals.begin(),
253 e = MultiUserVals.end(); it != e; ++it) {
254 // Check that all of the users of this instr are within the tree
255 // and that they are all from the same lane.
257 for (Value::use_iterator I = (*it)->use_begin(), E = (*it)->use_end();
259 if (LaneMap.find(*I) == LaneMap.end()) {
260 MustScalarize.insert(*it);
261 DEBUG(dbgs()<<"SLP: Adding " << **it <<
262 " to MustScalarize because of an out of tree usage.\n");
265 if (Lane == -1) Lane = LaneMap[*I];
266 if (Lane != LaneMap[*I]) {
267 MustScalarize.insert(*it);
268 DEBUG(dbgs()<<"Adding " << **it <<
269 " to MustScalarize because multiple lane use it: "
270 << Lane << " and " << LaneMap[*I] << ".\n");
276 // Now calculate the cost of vectorizing the tree.
277 return getTreeCost_rec(VL, 0);
280 void BoUpSLP::getTreeUses_rec(ValueList &VL, unsigned Depth) {
281 if (Depth == RecursionMaxDepth) return;
283 // Don't handle vectors.
284 if (VL[0]->getType()->isVectorTy()) return;
285 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
286 if (SI->getValueOperand()->getType()->isVectorTy()) return;
288 // Check if all of the operands are constants.
289 bool AllConst = true;
290 bool AllSameScalar = true;
291 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
292 AllConst &= isa<Constant>(VL[i]);
293 AllSameScalar &= (VL[0] == VL[i]);
294 Instruction *I = dyn_cast<Instruction>(VL[i]);
295 // If one of the instructions is out of this BB, we need to scalarize all.
296 if (I && I->getParent() != BB) return;
299 // If all of the operands are identical or constant we have a simple solution.
300 if (AllConst || AllSameScalar) return;
302 // Scalarize unknown structures.
303 Instruction *VL0 = dyn_cast<Instruction>(VL[0]);
306 unsigned Opcode = VL0->getOpcode();
307 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
308 Instruction *I = dyn_cast<Instruction>(VL[i]);
309 // If not all of the instructions are identical then we have to scalarize.
310 if (!I || Opcode != I->getOpcode()) return;
313 // Mark instructions with multiple users.
314 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
315 Instruction *I = dyn_cast<Instruction>(VL[i]);
316 // Remember to check if all of the users of this instr are vectorized
318 if (I && I->getNumUses() > 1) MultiUserVals.insert(I);
321 for (int i = 0, e = VL.size(); i < e; ++i) {
322 // Check that the instruction is only used within
324 if (LaneMap.count(VL[i]) && LaneMap[VL[i]] != i) return;
325 // Make this instruction as 'seen' and remember the lane.
330 case Instruction::Add:
331 case Instruction::FAdd:
332 case Instruction::Sub:
333 case Instruction::FSub:
334 case Instruction::Mul:
335 case Instruction::FMul:
336 case Instruction::UDiv:
337 case Instruction::SDiv:
338 case Instruction::FDiv:
339 case Instruction::URem:
340 case Instruction::SRem:
341 case Instruction::FRem:
342 case Instruction::Shl:
343 case Instruction::LShr:
344 case Instruction::AShr:
345 case Instruction::And:
346 case Instruction::Or:
347 case Instruction::Xor: {
348 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
350 // Prepare the operand vector.
351 for (unsigned j = 0; j < VL.size(); ++j)
352 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
354 getTreeUses_rec(Operands, Depth+1);
357 case Instruction::Store: {
359 for (unsigned j = 0; j < VL.size(); ++j)
360 Operands.push_back(cast<Instruction>(VL[j])->getOperand(0));
361 getTreeUses_rec(Operands, Depth+1);
369 int BoUpSLP::getTreeCost_rec(ValueList &VL, unsigned Depth) {
370 Type *ScalarTy = VL[0]->getType();
372 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
373 ScalarTy = SI->getValueOperand()->getType();
375 /// Don't mess with vectors.
376 if (ScalarTy->isVectorTy()) return max_cost;
377 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
379 if (Depth == RecursionMaxDepth) return getScalarizationCost(VecTy);
381 // Check if all of the operands are constants.
382 bool AllConst = true;
383 bool AllSameScalar = true;
384 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
385 AllConst &= isa<Constant>(VL[i]);
386 AllSameScalar &= (VL[0] == VL[i]);
387 // Must have a single use.
388 Instruction *I = dyn_cast<Instruction>(VL[i]);
389 // This instruction is outside the basic block or if it is a known hazard.
390 if (MustScalarize.count(VL[i]) || (I && I->getParent() != BB))
391 return getScalarizationCost(VecTy);
394 // Is this a simple vector constant.
395 if (AllConst) return 0;
397 // If all of the operands are identical we can broadcast them.
399 return TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy, 0);
401 // Scalarize unknown structures.
402 Instruction *VL0 = dyn_cast<Instruction>(VL[0]);
403 if (!VL0) return getScalarizationCost(VecTy);
404 assert(VL0->getParent() == BB && "Wrong BB");
406 unsigned Opcode = VL0->getOpcode();
407 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
408 Instruction *I = dyn_cast<Instruction>(VL[i]);
409 // If not all of the instructions are identical then we have to scalarize.
410 if (!I || Opcode != I->getOpcode()) return getScalarizationCost(VecTy);
413 // Check if it is safe to sink the loads or the stores.
414 if (Opcode == Instruction::Load || Opcode == Instruction::Store) {
415 int MaxIdx = InstrIdx[VL0];
416 for (unsigned i = 1, e = VL.size(); i < e; ++i )
417 MaxIdx = std::max(MaxIdx, InstrIdx[VL[i]]);
419 Instruction *Last = InstrVec[MaxIdx];
420 for (unsigned i = 0, e = VL.size(); i < e; ++i ) {
421 if (VL[i] == Last) continue;
422 Value *Barrier = isUnsafeToSink(cast<Instruction>(VL[i]), Last);
424 DEBUG(dbgs() << "SLP: Can't sink " << *VL[i] << "\n down to " <<
425 *Last << "\n because of " << *Barrier << "\n");
432 case Instruction::Add:
433 case Instruction::FAdd:
434 case Instruction::Sub:
435 case Instruction::FSub:
436 case Instruction::Mul:
437 case Instruction::FMul:
438 case Instruction::UDiv:
439 case Instruction::SDiv:
440 case Instruction::FDiv:
441 case Instruction::URem:
442 case Instruction::SRem:
443 case Instruction::FRem:
444 case Instruction::Shl:
445 case Instruction::LShr:
446 case Instruction::AShr:
447 case Instruction::And:
448 case Instruction::Or:
449 case Instruction::Xor: {
451 // Calculate the cost of all of the operands.
452 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
454 // Prepare the operand vector.
455 for (unsigned j = 0; j < VL.size(); ++j)
456 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
458 Cost += getTreeCost_rec(Operands, Depth+1);
459 if (Cost >= max_cost) return max_cost;
462 // Calculate the cost of this instruction.
463 int ScalarCost = VecTy->getNumElements() *
464 TTI->getArithmeticInstrCost(Opcode, ScalarTy);
466 int VecCost = TTI->getArithmeticInstrCost(Opcode, VecTy);
467 Cost += (VecCost - ScalarCost);
470 case Instruction::Load: {
471 // If we are scalarize the loads, add the cost of forming the vector.
472 for (unsigned i = 0, e = VL.size()-1; i < e; ++i)
473 if (!isConsecutiveAccess(VL[i], VL[i+1]))
474 return getScalarizationCost(VecTy);
476 // Cost of wide load - cost of scalar loads.
477 int ScalarLdCost = VecTy->getNumElements() *
478 TTI->getMemoryOpCost(Instruction::Load, ScalarTy, 1, 0);
479 int VecLdCost = TTI->getMemoryOpCost(Instruction::Load, ScalarTy, 1, 0);
480 return VecLdCost - ScalarLdCost;
482 case Instruction::Store: {
483 // We know that we can merge the stores. Calculate the cost.
484 int ScalarStCost = VecTy->getNumElements() *
485 TTI->getMemoryOpCost(Instruction::Store, ScalarTy, 1, 0);
486 int VecStCost = TTI->getMemoryOpCost(Instruction::Store, ScalarTy, 1,0);
487 int StoreCost = VecStCost - ScalarStCost;
490 for (unsigned j = 0; j < VL.size(); ++j) {
491 Operands.push_back(cast<Instruction>(VL[j])->getOperand(0));
492 MemBarrierIgnoreList.insert(VL[j]);
495 int TotalCost = StoreCost + getTreeCost_rec(Operands, Depth + 1);
499 // Unable to vectorize unknown instructions.
500 return getScalarizationCost(VecTy);
504 Instruction *BoUpSLP::GetLastInstr(ValueList &VL, unsigned VF) {
505 int MaxIdx = InstrIdx[BB->getFirstNonPHI()];
506 for (unsigned i = 0; i < VF; ++i )
507 MaxIdx = std::max(MaxIdx, InstrIdx[VL[i]]);
508 return InstrVec[MaxIdx + 1];
511 Value *BoUpSLP::Scalarize(ValueList &VL, VectorType *Ty) {
512 IRBuilder<> Builder(GetLastInstr(VL, Ty->getNumElements()));
513 Value *Vec = UndefValue::get(Ty);
514 for (unsigned i=0; i < Ty->getNumElements(); ++i) {
515 // Generate the 'InsertElement' instruction.
516 Vec = Builder.CreateInsertElement(Vec, VL[i], Builder.getInt32(i));
517 // Remember that this instruction is used as part of a 'gather' sequence.
518 // The caller of the bottom-up slp vectorizer can try to hoist the sequence
519 // if the users are outside of the basic block.
520 GatherInstructions.push_back(Vec);
526 Value *BoUpSLP::vectorizeTree(ValueList &VL, int VF) {
527 Value *V = vectorizeTree_rec(VL, VF);
528 // We moved some instructions around. We have to number them again
529 // before we can do any analysis.
530 numberInstructions();
531 MustScalarize.clear();
535 Value *BoUpSLP::vectorizeTree_rec(ValueList &VL, int VF) {
536 Type *ScalarTy = VL[0]->getType();
537 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
538 ScalarTy = SI->getValueOperand()->getType();
539 VectorType *VecTy = VectorType::get(ScalarTy, VF);
541 // Check if all of the operands are constants or identical.
542 bool AllConst = true;
543 bool AllSameScalar = true;
544 for (unsigned i = 0, e = VF; i < e; ++i) {
545 AllConst &= !!dyn_cast<Constant>(VL[i]);
546 AllSameScalar &= (VL[0] == VL[i]);
547 // The instruction must be in the same BB, and it must be vectorizable.
548 Instruction *I = dyn_cast<Instruction>(VL[i]);
549 if (MustScalarize.count(VL[i]) || (I && I->getParent() != BB))
550 return Scalarize(VL, VecTy);
553 // Check that this is a simple vector constant.
554 if (AllConst || AllSameScalar) return Scalarize(VL, VecTy);
556 // Scalarize unknown structures.
557 Instruction *VL0 = dyn_cast<Instruction>(VL[0]);
558 if (!VL0) return Scalarize(VL, VecTy);
560 if (VectorizedValues.count(VL0)) return VectorizedValues[VL0];
562 unsigned Opcode = VL0->getOpcode();
563 for (unsigned i = 0, e = VF; i < e; ++i) {
564 Instruction *I = dyn_cast<Instruction>(VL[i]);
565 // If not all of the instructions are identical then we have to scalarize.
566 if (!I || Opcode != I->getOpcode()) return Scalarize(VL, VecTy);
570 case Instruction::Add:
571 case Instruction::FAdd:
572 case Instruction::Sub:
573 case Instruction::FSub:
574 case Instruction::Mul:
575 case Instruction::FMul:
576 case Instruction::UDiv:
577 case Instruction::SDiv:
578 case Instruction::FDiv:
579 case Instruction::URem:
580 case Instruction::SRem:
581 case Instruction::FRem:
582 case Instruction::Shl:
583 case Instruction::LShr:
584 case Instruction::AShr:
585 case Instruction::And:
586 case Instruction::Or:
587 case Instruction::Xor: {
588 ValueList LHSVL, RHSVL;
589 for (int i = 0; i < VF; ++i) {
590 RHSVL.push_back(cast<Instruction>(VL[i])->getOperand(0));
591 LHSVL.push_back(cast<Instruction>(VL[i])->getOperand(1));
594 Value *RHS = vectorizeTree_rec(RHSVL, VF);
595 Value *LHS = vectorizeTree_rec(LHSVL, VF);
596 IRBuilder<> Builder(GetLastInstr(VL, VF));
597 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(VL0);
598 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), RHS,LHS);
599 VectorizedValues[VL0] = V;
602 case Instruction::Load: {
603 LoadInst *LI = dyn_cast<LoadInst>(VL0);
604 unsigned Alignment = LI->getAlignment();
606 // Check if all of the loads are consecutive.
607 for (unsigned i = 1, e = VF; i < e; ++i)
608 if (!isConsecutiveAccess(VL[i-1], VL[i]))
609 return Scalarize(VL, VecTy);
611 IRBuilder<> Builder(GetLastInstr(VL, VF));
612 Value *VecPtr = Builder.CreateBitCast(LI->getPointerOperand(),
613 VecTy->getPointerTo());
614 LI = Builder.CreateLoad(VecPtr);
615 LI->setAlignment(Alignment);
616 VectorizedValues[VL0] = LI;
619 case Instruction::Store: {
620 StoreInst *SI = dyn_cast<StoreInst>(VL0);
621 unsigned Alignment = SI->getAlignment();
624 for (int i = 0; i < VF; ++i)
625 ValueOp.push_back(cast<StoreInst>(VL[i])->getValueOperand());
627 Value *VecValue = vectorizeTree_rec(ValueOp, VF);
629 IRBuilder<> Builder(GetLastInstr(VL, VF));
630 Value *VecPtr = Builder.CreateBitCast(SI->getPointerOperand(),
631 VecTy->getPointerTo());
632 Builder.CreateStore(VecValue, VecPtr)->setAlignment(Alignment);
634 for (int i = 0; i < VF; ++i)
635 cast<Instruction>(VL[i])->eraseFromParent();
639 Value *S = Scalarize(VL, VecTy);
640 VectorizedValues[VL0] = S;
645 } // end of namespace