1 //===- InstCombineLoadStoreAlloca.cpp -------------------------------------===//
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 //===----------------------------------------------------------------------===//
10 // This file implements the visit functions for load, store and alloca.
12 //===----------------------------------------------------------------------===//
14 #include "InstCombineInternal.h"
15 #include "llvm/ADT/MapVector.h"
16 #include "llvm/ADT/SmallString.h"
17 #include "llvm/ADT/Statistic.h"
18 #include "llvm/Analysis/Loads.h"
19 #include "llvm/IR/ConstantRange.h"
20 #include "llvm/IR/DataLayout.h"
21 #include "llvm/IR/DebugInfo.h"
22 #include "llvm/IR/IntrinsicInst.h"
23 #include "llvm/IR/LLVMContext.h"
24 #include "llvm/IR/MDBuilder.h"
25 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
26 #include "llvm/Transforms/Utils/Local.h"
29 #define DEBUG_TYPE "instcombine"
31 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
32 STATISTIC(NumGlobalCopies, "Number of allocas copied from constant global");
34 /// pointsToConstantGlobal - Return true if V (possibly indirectly) points to
35 /// some part of a constant global variable. This intentionally only accepts
36 /// constant expressions because we can't rewrite arbitrary instructions.
37 static bool pointsToConstantGlobal(Value *V) {
38 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
39 return GV->isConstant();
41 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
42 if (CE->getOpcode() == Instruction::BitCast ||
43 CE->getOpcode() == Instruction::AddrSpaceCast ||
44 CE->getOpcode() == Instruction::GetElementPtr)
45 return pointsToConstantGlobal(CE->getOperand(0));
50 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
51 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
52 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
53 /// track of whether it moves the pointer (with IsOffset) but otherwise traverse
54 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
55 /// the alloca, and if the source pointer is a pointer to a constant global, we
56 /// can optimize this.
58 isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy,
59 SmallVectorImpl<Instruction *> &ToDelete) {
60 // We track lifetime intrinsics as we encounter them. If we decide to go
61 // ahead and replace the value with the global, this lets the caller quickly
62 // eliminate the markers.
64 SmallVector<std::pair<Value *, bool>, 35> ValuesToInspect;
65 ValuesToInspect.emplace_back(V, false);
66 while (!ValuesToInspect.empty()) {
67 auto ValuePair = ValuesToInspect.pop_back_val();
68 const bool IsOffset = ValuePair.second;
69 for (auto &U : ValuePair.first->uses()) {
70 auto *I = cast<Instruction>(U.getUser());
72 if (auto *LI = dyn_cast<LoadInst>(I)) {
73 // Ignore non-volatile loads, they are always ok.
74 if (!LI->isSimple()) return false;
78 if (isa<BitCastInst>(I) || isa<AddrSpaceCastInst>(I)) {
79 // If uses of the bitcast are ok, we are ok.
80 ValuesToInspect.emplace_back(I, IsOffset);
83 if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) {
84 // If the GEP has all zero indices, it doesn't offset the pointer. If it
86 ValuesToInspect.emplace_back(I, IsOffset || !GEP->hasAllZeroIndices());
90 if (auto CS = CallSite(I)) {
91 // If this is the function being called then we treat it like a load and
96 unsigned DataOpNo = CS.getDataOperandNo(&U);
97 bool IsArgOperand = CS.isArgOperand(&U);
99 // Inalloca arguments are clobbered by the call.
100 if (IsArgOperand && CS.isInAllocaArgument(DataOpNo))
103 // If this is a readonly/readnone call site, then we know it is just a
104 // load (but one that potentially returns the value itself), so we can
105 // ignore it if we know that the value isn't captured.
106 if (CS.onlyReadsMemory() &&
107 (CS.getInstruction()->use_empty() || CS.doesNotCapture(DataOpNo)))
110 // If this is being passed as a byval argument, the caller is making a
111 // copy, so it is only a read of the alloca.
112 if (IsArgOperand && CS.isByValArgument(DataOpNo))
116 // Lifetime intrinsics can be handled by the caller.
117 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
118 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
119 II->getIntrinsicID() == Intrinsic::lifetime_end) {
120 assert(II->use_empty() && "Lifetime markers have no result to use!");
121 ToDelete.push_back(II);
126 // If this is isn't our memcpy/memmove, reject it as something we can't
128 MemTransferInst *MI = dyn_cast<MemTransferInst>(I);
132 // If the transfer is using the alloca as a source of the transfer, then
133 // ignore it since it is a load (unless the transfer is volatile).
134 if (U.getOperandNo() == 1) {
135 if (MI->isVolatile()) return false;
139 // If we already have seen a copy, reject the second one.
140 if (TheCopy) return false;
142 // If the pointer has been offset from the start of the alloca, we can't
143 // safely handle this.
144 if (IsOffset) return false;
146 // If the memintrinsic isn't using the alloca as the dest, reject it.
147 if (U.getOperandNo() != 0) return false;
149 // If the source of the memcpy/move is not a constant global, reject it.
150 if (!pointsToConstantGlobal(MI->getSource()))
153 // Otherwise, the transform is safe. Remember the copy instruction.
160 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
161 /// modified by a copy from a constant global. If we can prove this, we can
162 /// replace any uses of the alloca with uses of the global directly.
163 static MemTransferInst *
164 isOnlyCopiedFromConstantGlobal(AllocaInst *AI,
165 SmallVectorImpl<Instruction *> &ToDelete) {
166 MemTransferInst *TheCopy = nullptr;
167 if (isOnlyCopiedFromConstantGlobal(AI, TheCopy, ToDelete))
172 /// Returns true if V is dereferenceable for size of alloca.
173 static bool isDereferenceableForAllocaSize(const Value *V, const AllocaInst *AI,
174 const DataLayout &DL) {
175 if (AI->isArrayAllocation())
177 uint64_t AllocaSize = DL.getTypeStoreSize(AI->getAllocatedType());
180 return isDereferenceableAndAlignedPointer(V, AI->getAlignment(),
181 APInt(64, AllocaSize), DL);
184 static Instruction *simplifyAllocaArraySize(InstCombiner &IC, AllocaInst &AI) {
185 // Check for array size of 1 (scalar allocation).
186 if (!AI.isArrayAllocation()) {
187 // i32 1 is the canonical array size for scalar allocations.
188 if (AI.getArraySize()->getType()->isIntegerTy(32))
192 Value *V = IC.Builder->getInt32(1);
197 // Convert: alloca Ty, C - where C is a constant != 1 into: alloca [C x Ty], 1
198 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
199 Type *NewTy = ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
200 AllocaInst *New = IC.Builder->CreateAlloca(NewTy, nullptr, AI.getName());
201 New->setAlignment(AI.getAlignment());
203 // Scan to the end of the allocation instructions, to skip over a block of
204 // allocas if possible...also skip interleaved debug info
206 BasicBlock::iterator It(New);
207 while (isa<AllocaInst>(*It) || isa<DbgInfoIntrinsic>(*It))
210 // Now that I is pointing to the first non-allocation-inst in the block,
211 // insert our getelementptr instruction...
213 Type *IdxTy = IC.getDataLayout().getIntPtrType(AI.getType());
214 Value *NullIdx = Constant::getNullValue(IdxTy);
215 Value *Idx[2] = {NullIdx, NullIdx};
217 GetElementPtrInst::CreateInBounds(New, Idx, New->getName() + ".sub");
218 IC.InsertNewInstBefore(GEP, *It);
220 // Now make everything use the getelementptr instead of the original
222 return IC.replaceInstUsesWith(AI, GEP);
225 if (isa<UndefValue>(AI.getArraySize()))
226 return IC.replaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
228 // Ensure that the alloca array size argument has type intptr_t, so that
229 // any casting is exposed early.
230 Type *IntPtrTy = IC.getDataLayout().getIntPtrType(AI.getType());
231 if (AI.getArraySize()->getType() != IntPtrTy) {
232 Value *V = IC.Builder->CreateIntCast(AI.getArraySize(), IntPtrTy, false);
241 // If I and V are pointers in different address space, it is not allowed to
242 // use replaceAllUsesWith since I and V have different types. A
243 // non-target-specific transformation should not use addrspacecast on V since
244 // the two address space may be disjoint depending on target.
246 // This class chases down uses of the old pointer until reaching the load
247 // instructions, then replaces the old pointer in the load instructions with
248 // the new pointer. If during the chasing it sees bitcast or GEP, it will
249 // create new bitcast or GEP with the new pointer and use them in the load
251 class PointerReplacer {
253 PointerReplacer(InstCombiner &IC) : IC(IC) {}
254 void replacePointer(Instruction &I, Value *V);
257 void findLoadAndReplace(Instruction &I);
258 void replace(Instruction *I);
259 Value *getReplacement(Value *I);
261 SmallVector<Instruction *, 4> Path;
262 MapVector<Value *, Value *> WorkMap;
265 } // end anonymous namespace
267 void PointerReplacer::findLoadAndReplace(Instruction &I) {
268 for (auto U : I.users()) {
269 auto *Inst = dyn_cast<Instruction>(&*U);
272 DEBUG(dbgs() << "Found pointer user: " << *U << '\n');
273 if (isa<LoadInst>(Inst)) {
277 } else if (isa<GetElementPtrInst>(Inst) || isa<BitCastInst>(Inst)) {
278 Path.push_back(Inst);
279 findLoadAndReplace(*Inst);
287 Value *PointerReplacer::getReplacement(Value *V) {
288 auto Loc = WorkMap.find(V);
289 if (Loc != WorkMap.end())
294 void PointerReplacer::replace(Instruction *I) {
295 if (getReplacement(I))
298 if (auto *LT = dyn_cast<LoadInst>(I)) {
299 auto *V = getReplacement(LT->getPointerOperand());
300 assert(V && "Operand not replaced");
301 auto *NewI = new LoadInst(V);
303 IC.InsertNewInstWith(NewI, *LT);
304 IC.replaceInstUsesWith(*LT, NewI);
306 } else if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) {
307 auto *V = getReplacement(GEP->getPointerOperand());
308 assert(V && "Operand not replaced");
309 SmallVector<Value *, 8> Indices;
310 Indices.append(GEP->idx_begin(), GEP->idx_end());
311 auto *NewI = GetElementPtrInst::Create(
312 V->getType()->getPointerElementType(), V, Indices);
313 IC.InsertNewInstWith(NewI, *GEP);
316 } else if (auto *BC = dyn_cast<BitCastInst>(I)) {
317 auto *V = getReplacement(BC->getOperand(0));
318 assert(V && "Operand not replaced");
319 auto *NewT = PointerType::get(BC->getType()->getPointerElementType(),
320 V->getType()->getPointerAddressSpace());
321 auto *NewI = new BitCastInst(V, NewT);
322 IC.InsertNewInstWith(NewI, *BC);
326 llvm_unreachable("should never reach here");
330 void PointerReplacer::replacePointer(Instruction &I, Value *V) {
332 auto *PT = cast<PointerType>(I.getType());
333 auto *NT = cast<PointerType>(V->getType());
334 assert(PT != NT && PT->getElementType() == NT->getElementType() &&
338 findLoadAndReplace(I);
341 Instruction *InstCombiner::visitAllocaInst(AllocaInst &AI) {
342 if (auto *I = simplifyAllocaArraySize(*this, AI))
345 if (AI.getAllocatedType()->isSized()) {
346 // If the alignment is 0 (unspecified), assign it the preferred alignment.
347 if (AI.getAlignment() == 0)
348 AI.setAlignment(DL.getPrefTypeAlignment(AI.getAllocatedType()));
350 // Move all alloca's of zero byte objects to the entry block and merge them
351 // together. Note that we only do this for alloca's, because malloc should
352 // allocate and return a unique pointer, even for a zero byte allocation.
353 if (DL.getTypeAllocSize(AI.getAllocatedType()) == 0) {
354 // For a zero sized alloca there is no point in doing an array allocation.
355 // This is helpful if the array size is a complicated expression not used
357 if (AI.isArrayAllocation()) {
358 AI.setOperand(0, ConstantInt::get(AI.getArraySize()->getType(), 1));
362 // Get the first instruction in the entry block.
363 BasicBlock &EntryBlock = AI.getParent()->getParent()->getEntryBlock();
364 Instruction *FirstInst = EntryBlock.getFirstNonPHIOrDbg();
365 if (FirstInst != &AI) {
366 // If the entry block doesn't start with a zero-size alloca then move
367 // this one to the start of the entry block. There is no problem with
368 // dominance as the array size was forced to a constant earlier already.
369 AllocaInst *EntryAI = dyn_cast<AllocaInst>(FirstInst);
370 if (!EntryAI || !EntryAI->getAllocatedType()->isSized() ||
371 DL.getTypeAllocSize(EntryAI->getAllocatedType()) != 0) {
372 AI.moveBefore(FirstInst);
376 // If the alignment of the entry block alloca is 0 (unspecified),
377 // assign it the preferred alignment.
378 if (EntryAI->getAlignment() == 0)
379 EntryAI->setAlignment(
380 DL.getPrefTypeAlignment(EntryAI->getAllocatedType()));
381 // Replace this zero-sized alloca with the one at the start of the entry
382 // block after ensuring that the address will be aligned enough for both
384 unsigned MaxAlign = std::max(EntryAI->getAlignment(),
386 EntryAI->setAlignment(MaxAlign);
387 if (AI.getType() != EntryAI->getType())
388 return new BitCastInst(EntryAI, AI.getType());
389 return replaceInstUsesWith(AI, EntryAI);
394 if (AI.getAlignment()) {
395 // Check to see if this allocation is only modified by a memcpy/memmove from
396 // a constant global whose alignment is equal to or exceeds that of the
397 // allocation. If this is the case, we can change all users to use
398 // the constant global instead. This is commonly produced by the CFE by
399 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
400 // is only subsequently read.
401 SmallVector<Instruction *, 4> ToDelete;
402 if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(&AI, ToDelete)) {
403 unsigned SourceAlign = getOrEnforceKnownAlignment(
404 Copy->getSource(), AI.getAlignment(), DL, &AI, &AC, &DT);
405 if (AI.getAlignment() <= SourceAlign &&
406 isDereferenceableForAllocaSize(Copy->getSource(), &AI, DL)) {
407 DEBUG(dbgs() << "Found alloca equal to global: " << AI << '\n');
408 DEBUG(dbgs() << " memcpy = " << *Copy << '\n');
409 for (unsigned i = 0, e = ToDelete.size(); i != e; ++i)
410 eraseInstFromFunction(*ToDelete[i]);
411 Constant *TheSrc = cast<Constant>(Copy->getSource());
412 auto *SrcTy = TheSrc->getType();
413 auto *DestTy = PointerType::get(AI.getType()->getPointerElementType(),
414 SrcTy->getPointerAddressSpace());
416 ConstantExpr::getPointerBitCastOrAddrSpaceCast(TheSrc, DestTy);
417 if (AI.getType()->getPointerAddressSpace() ==
418 SrcTy->getPointerAddressSpace()) {
419 Instruction *NewI = replaceInstUsesWith(AI, Cast);
420 eraseInstFromFunction(*Copy);
424 PointerReplacer PtrReplacer(*this);
425 PtrReplacer.replacePointer(AI, Cast);
432 // At last, use the generic allocation site handler to aggressively remove
434 return visitAllocSite(AI);
437 // Are we allowed to form a atomic load or store of this type?
438 static bool isSupportedAtomicType(Type *Ty) {
439 return Ty->isIntegerTy() || Ty->isPointerTy() || Ty->isFloatingPointTy();
442 /// \brief Helper to combine a load to a new type.
444 /// This just does the work of combining a load to a new type. It handles
445 /// metadata, etc., and returns the new instruction. The \c NewTy should be the
446 /// loaded *value* type. This will convert it to a pointer, cast the operand to
447 /// that pointer type, load it, etc.
449 /// Note that this will create all of the instructions with whatever insert
450 /// point the \c InstCombiner currently is using.
451 static LoadInst *combineLoadToNewType(InstCombiner &IC, LoadInst &LI, Type *NewTy,
452 const Twine &Suffix = "") {
453 assert((!LI.isAtomic() || isSupportedAtomicType(NewTy)) &&
454 "can't fold an atomic load to requested type");
456 Value *Ptr = LI.getPointerOperand();
457 unsigned AS = LI.getPointerAddressSpace();
458 SmallVector<std::pair<unsigned, MDNode *>, 8> MD;
459 LI.getAllMetadata(MD);
461 LoadInst *NewLoad = IC.Builder->CreateAlignedLoad(
462 IC.Builder->CreateBitCast(Ptr, NewTy->getPointerTo(AS)),
463 LI.getAlignment(), LI.isVolatile(), LI.getName() + Suffix);
464 NewLoad->setAtomic(LI.getOrdering(), LI.getSynchScope());
465 MDBuilder MDB(NewLoad->getContext());
466 for (const auto &MDPair : MD) {
467 unsigned ID = MDPair.first;
468 MDNode *N = MDPair.second;
469 // Note, essentially every kind of metadata should be preserved here! This
470 // routine is supposed to clone a load instruction changing *only its type*.
471 // The only metadata it makes sense to drop is metadata which is invalidated
472 // when the pointer type changes. This should essentially never be the case
473 // in LLVM, but we explicitly switch over only known metadata to be
474 // conservatively correct. If you are adding metadata to LLVM which pertains
475 // to loads, you almost certainly want to add it here.
477 case LLVMContext::MD_dbg:
478 case LLVMContext::MD_tbaa:
479 case LLVMContext::MD_prof:
480 case LLVMContext::MD_fpmath:
481 case LLVMContext::MD_tbaa_struct:
482 case LLVMContext::MD_invariant_load:
483 case LLVMContext::MD_alias_scope:
484 case LLVMContext::MD_noalias:
485 case LLVMContext::MD_nontemporal:
486 case LLVMContext::MD_mem_parallel_loop_access:
487 // All of these directly apply.
488 NewLoad->setMetadata(ID, N);
491 case LLVMContext::MD_nonnull:
492 copyNonnullMetadata(LI, N, *NewLoad);
494 case LLVMContext::MD_align:
495 case LLVMContext::MD_dereferenceable:
496 case LLVMContext::MD_dereferenceable_or_null:
497 // These only directly apply if the new type is also a pointer.
498 if (NewTy->isPointerTy())
499 NewLoad->setMetadata(ID, N);
501 case LLVMContext::MD_range:
502 copyRangeMetadata(IC.getDataLayout(), LI, N, *NewLoad);
509 /// \brief Combine a store to a new type.
511 /// Returns the newly created store instruction.
512 static StoreInst *combineStoreToNewValue(InstCombiner &IC, StoreInst &SI, Value *V) {
513 assert((!SI.isAtomic() || isSupportedAtomicType(V->getType())) &&
514 "can't fold an atomic store of requested type");
516 Value *Ptr = SI.getPointerOperand();
517 unsigned AS = SI.getPointerAddressSpace();
518 SmallVector<std::pair<unsigned, MDNode *>, 8> MD;
519 SI.getAllMetadata(MD);
521 StoreInst *NewStore = IC.Builder->CreateAlignedStore(
522 V, IC.Builder->CreateBitCast(Ptr, V->getType()->getPointerTo(AS)),
523 SI.getAlignment(), SI.isVolatile());
524 NewStore->setAtomic(SI.getOrdering(), SI.getSynchScope());
525 for (const auto &MDPair : MD) {
526 unsigned ID = MDPair.first;
527 MDNode *N = MDPair.second;
528 // Note, essentially every kind of metadata should be preserved here! This
529 // routine is supposed to clone a store instruction changing *only its
530 // type*. The only metadata it makes sense to drop is metadata which is
531 // invalidated when the pointer type changes. This should essentially
532 // never be the case in LLVM, but we explicitly switch over only known
533 // metadata to be conservatively correct. If you are adding metadata to
534 // LLVM which pertains to stores, you almost certainly want to add it
537 case LLVMContext::MD_dbg:
538 case LLVMContext::MD_tbaa:
539 case LLVMContext::MD_prof:
540 case LLVMContext::MD_fpmath:
541 case LLVMContext::MD_tbaa_struct:
542 case LLVMContext::MD_alias_scope:
543 case LLVMContext::MD_noalias:
544 case LLVMContext::MD_nontemporal:
545 case LLVMContext::MD_mem_parallel_loop_access:
546 // All of these directly apply.
547 NewStore->setMetadata(ID, N);
550 case LLVMContext::MD_invariant_load:
551 case LLVMContext::MD_nonnull:
552 case LLVMContext::MD_range:
553 case LLVMContext::MD_align:
554 case LLVMContext::MD_dereferenceable:
555 case LLVMContext::MD_dereferenceable_or_null:
556 // These don't apply for stores.
564 /// \brief Combine loads to match the type of their uses' value after looking
565 /// through intervening bitcasts.
567 /// The core idea here is that if the result of a load is used in an operation,
568 /// we should load the type most conducive to that operation. For example, when
569 /// loading an integer and converting that immediately to a pointer, we should
570 /// instead directly load a pointer.
572 /// However, this routine must never change the width of a load or the number of
573 /// loads as that would introduce a semantic change. This combine is expected to
574 /// be a semantic no-op which just allows loads to more closely model the types
575 /// of their consuming operations.
577 /// Currently, we also refuse to change the precise type used for an atomic load
578 /// or a volatile load. This is debatable, and might be reasonable to change
579 /// later. However, it is risky in case some backend or other part of LLVM is
580 /// relying on the exact type loaded to select appropriate atomic operations.
581 static Instruction *combineLoadToOperationType(InstCombiner &IC, LoadInst &LI) {
582 // FIXME: We could probably with some care handle both volatile and ordered
583 // atomic loads here but it isn't clear that this is important.
584 if (!LI.isUnordered())
590 // swifterror values can't be bitcasted.
591 if (LI.getPointerOperand()->isSwiftError())
594 Type *Ty = LI.getType();
595 const DataLayout &DL = IC.getDataLayout();
597 // Try to canonicalize loads which are only ever stored to operate over
598 // integers instead of any other type. We only do this when the loaded type
599 // is sized and has a size exactly the same as its store size and the store
600 // size is a legal integer type.
601 if (!Ty->isIntegerTy() && Ty->isSized() &&
602 DL.isLegalInteger(DL.getTypeStoreSizeInBits(Ty)) &&
603 DL.getTypeStoreSizeInBits(Ty) == DL.getTypeSizeInBits(Ty) &&
604 !DL.isNonIntegralPointerType(Ty)) {
605 if (all_of(LI.users(), [&LI](User *U) {
606 auto *SI = dyn_cast<StoreInst>(U);
607 return SI && SI->getPointerOperand() != &LI &&
608 !SI->getPointerOperand()->isSwiftError();
610 LoadInst *NewLoad = combineLoadToNewType(
612 Type::getIntNTy(LI.getContext(), DL.getTypeStoreSizeInBits(Ty)));
613 // Replace all the stores with stores of the newly loaded value.
614 for (auto UI = LI.user_begin(), UE = LI.user_end(); UI != UE;) {
615 auto *SI = cast<StoreInst>(*UI++);
616 IC.Builder->SetInsertPoint(SI);
617 combineStoreToNewValue(IC, *SI, NewLoad);
618 IC.eraseInstFromFunction(*SI);
620 assert(LI.use_empty() && "Failed to remove all users of the load!");
621 // Return the old load so the combiner can delete it safely.
626 // Fold away bit casts of the loaded value by loading the desired type.
627 // We can do this for BitCastInsts as well as casts from and to pointer types,
628 // as long as those are noops (i.e., the source or dest type have the same
629 // bitwidth as the target's pointers).
631 if (auto* CI = dyn_cast<CastInst>(LI.user_back()))
632 if (CI->isNoopCast(DL))
633 if (!LI.isAtomic() || isSupportedAtomicType(CI->getDestTy())) {
634 LoadInst *NewLoad = combineLoadToNewType(IC, LI, CI->getDestTy());
635 CI->replaceAllUsesWith(NewLoad);
636 IC.eraseInstFromFunction(*CI);
640 // FIXME: We should also canonicalize loads of vectors when their elements are
641 // cast to other types.
645 static Instruction *unpackLoadToAggregate(InstCombiner &IC, LoadInst &LI) {
646 // FIXME: We could probably with some care handle both volatile and atomic
647 // stores here but it isn't clear that this is important.
651 Type *T = LI.getType();
652 if (!T->isAggregateType())
655 StringRef Name = LI.getName();
656 assert(LI.getAlignment() && "Alignment must be set at this point");
658 if (auto *ST = dyn_cast<StructType>(T)) {
659 // If the struct only have one element, we unpack.
660 auto NumElements = ST->getNumElements();
661 if (NumElements == 1) {
662 LoadInst *NewLoad = combineLoadToNewType(IC, LI, ST->getTypeAtIndex(0U),
664 return IC.replaceInstUsesWith(LI, IC.Builder->CreateInsertValue(
665 UndefValue::get(T), NewLoad, 0, Name));
668 // We don't want to break loads with padding here as we'd loose
669 // the knowledge that padding exists for the rest of the pipeline.
670 const DataLayout &DL = IC.getDataLayout();
671 auto *SL = DL.getStructLayout(ST);
672 if (SL->hasPadding())
675 auto Align = LI.getAlignment();
677 Align = DL.getABITypeAlignment(ST);
679 auto *Addr = LI.getPointerOperand();
680 auto *IdxType = Type::getInt32Ty(T->getContext());
681 auto *Zero = ConstantInt::get(IdxType, 0);
683 Value *V = UndefValue::get(T);
684 for (unsigned i = 0; i < NumElements; i++) {
685 Value *Indices[2] = {
687 ConstantInt::get(IdxType, i),
689 auto *Ptr = IC.Builder->CreateInBoundsGEP(ST, Addr, makeArrayRef(Indices),
691 auto EltAlign = MinAlign(Align, SL->getElementOffset(i));
692 auto *L = IC.Builder->CreateAlignedLoad(Ptr, EltAlign, Name + ".unpack");
693 V = IC.Builder->CreateInsertValue(V, L, i);
697 return IC.replaceInstUsesWith(LI, V);
700 if (auto *AT = dyn_cast<ArrayType>(T)) {
701 auto *ET = AT->getElementType();
702 auto NumElements = AT->getNumElements();
703 if (NumElements == 1) {
704 LoadInst *NewLoad = combineLoadToNewType(IC, LI, ET, ".unpack");
705 return IC.replaceInstUsesWith(LI, IC.Builder->CreateInsertValue(
706 UndefValue::get(T), NewLoad, 0, Name));
709 // Bail out if the array is too large. Ideally we would like to optimize
710 // arrays of arbitrary size but this has a terrible impact on compile time.
711 // The threshold here is chosen arbitrarily, maybe needs a little bit of
713 if (NumElements > IC.MaxArraySizeForCombine)
716 const DataLayout &DL = IC.getDataLayout();
717 auto EltSize = DL.getTypeAllocSize(ET);
718 auto Align = LI.getAlignment();
720 Align = DL.getABITypeAlignment(T);
722 auto *Addr = LI.getPointerOperand();
723 auto *IdxType = Type::getInt64Ty(T->getContext());
724 auto *Zero = ConstantInt::get(IdxType, 0);
726 Value *V = UndefValue::get(T);
728 for (uint64_t i = 0; i < NumElements; i++) {
729 Value *Indices[2] = {
731 ConstantInt::get(IdxType, i),
733 auto *Ptr = IC.Builder->CreateInBoundsGEP(AT, Addr, makeArrayRef(Indices),
735 auto *L = IC.Builder->CreateAlignedLoad(Ptr, MinAlign(Align, Offset),
737 V = IC.Builder->CreateInsertValue(V, L, i);
742 return IC.replaceInstUsesWith(LI, V);
748 // If we can determine that all possible objects pointed to by the provided
749 // pointer value are, not only dereferenceable, but also definitively less than
750 // or equal to the provided maximum size, then return true. Otherwise, return
751 // false (constant global values and allocas fall into this category).
753 // FIXME: This should probably live in ValueTracking (or similar).
754 static bool isObjectSizeLessThanOrEq(Value *V, uint64_t MaxSize,
755 const DataLayout &DL) {
756 SmallPtrSet<Value *, 4> Visited;
757 SmallVector<Value *, 4> Worklist(1, V);
760 Value *P = Worklist.pop_back_val();
761 P = P->stripPointerCasts();
763 if (!Visited.insert(P).second)
766 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
767 Worklist.push_back(SI->getTrueValue());
768 Worklist.push_back(SI->getFalseValue());
772 if (PHINode *PN = dyn_cast<PHINode>(P)) {
773 for (Value *IncValue : PN->incoming_values())
774 Worklist.push_back(IncValue);
778 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(P)) {
779 if (GA->isInterposable())
781 Worklist.push_back(GA->getAliasee());
785 // If we know how big this object is, and it is less than MaxSize, continue
786 // searching. Otherwise, return false.
787 if (AllocaInst *AI = dyn_cast<AllocaInst>(P)) {
788 if (!AI->getAllocatedType()->isSized())
791 ConstantInt *CS = dyn_cast<ConstantInt>(AI->getArraySize());
795 uint64_t TypeSize = DL.getTypeAllocSize(AI->getAllocatedType());
796 // Make sure that, even if the multiplication below would wrap as an
797 // uint64_t, we still do the right thing.
798 if ((CS->getValue().zextOrSelf(128)*APInt(128, TypeSize)).ugt(MaxSize))
803 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
804 if (!GV->hasDefinitiveInitializer() || !GV->isConstant())
807 uint64_t InitSize = DL.getTypeAllocSize(GV->getValueType());
808 if (InitSize > MaxSize)
814 } while (!Worklist.empty());
819 // If we're indexing into an object of a known size, and the outer index is
820 // not a constant, but having any value but zero would lead to undefined
821 // behavior, replace it with zero.
823 // For example, if we have:
824 // @f.a = private unnamed_addr constant [1 x i32] [i32 12], align 4
826 // %arrayidx = getelementptr inbounds [1 x i32]* @f.a, i64 0, i64 %x
827 // ... = load i32* %arrayidx, align 4
828 // Then we know that we can replace %x in the GEP with i64 0.
830 // FIXME: We could fold any GEP index to zero that would cause UB if it were
831 // not zero. Currently, we only handle the first such index. Also, we could
832 // also search through non-zero constant indices if we kept track of the
833 // offsets those indices implied.
834 static bool canReplaceGEPIdxWithZero(InstCombiner &IC, GetElementPtrInst *GEPI,
835 Instruction *MemI, unsigned &Idx) {
836 if (GEPI->getNumOperands() < 2)
839 // Find the first non-zero index of a GEP. If all indices are zero, return
840 // one past the last index.
841 auto FirstNZIdx = [](const GetElementPtrInst *GEPI) {
843 for (unsigned IE = GEPI->getNumOperands(); I != IE; ++I) {
844 Value *V = GEPI->getOperand(I);
845 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V))
855 // Skip through initial 'zero' indices, and find the corresponding pointer
856 // type. See if the next index is not a constant.
857 Idx = FirstNZIdx(GEPI);
858 if (Idx == GEPI->getNumOperands())
860 if (isa<Constant>(GEPI->getOperand(Idx)))
863 SmallVector<Value *, 4> Ops(GEPI->idx_begin(), GEPI->idx_begin() + Idx);
865 GetElementPtrInst::getIndexedType(GEPI->getSourceElementType(), Ops);
866 if (!AllocTy || !AllocTy->isSized())
868 const DataLayout &DL = IC.getDataLayout();
869 uint64_t TyAllocSize = DL.getTypeAllocSize(AllocTy);
871 // If there are more indices after the one we might replace with a zero, make
872 // sure they're all non-negative. If any of them are negative, the overall
873 // address being computed might be before the base address determined by the
874 // first non-zero index.
875 auto IsAllNonNegative = [&]() {
876 for (unsigned i = Idx+1, e = GEPI->getNumOperands(); i != e; ++i) {
877 KnownBits Known = IC.computeKnownBits(GEPI->getOperand(i), 0, MemI);
878 if (Known.isNonNegative())
886 // FIXME: If the GEP is not inbounds, and there are extra indices after the
887 // one we'll replace, those could cause the address computation to wrap
888 // (rendering the IsAllNonNegative() check below insufficient). We can do
889 // better, ignoring zero indices (and other indices we can prove small
890 // enough not to wrap).
891 if (Idx+1 != GEPI->getNumOperands() && !GEPI->isInBounds())
894 // Note that isObjectSizeLessThanOrEq will return true only if the pointer is
895 // also known to be dereferenceable.
896 return isObjectSizeLessThanOrEq(GEPI->getOperand(0), TyAllocSize, DL) &&
900 // If we're indexing into an object with a variable index for the memory
901 // access, but the object has only one element, we can assume that the index
902 // will always be zero. If we replace the GEP, return it.
903 template <typename T>
904 static Instruction *replaceGEPIdxWithZero(InstCombiner &IC, Value *Ptr,
906 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Ptr)) {
908 if (canReplaceGEPIdxWithZero(IC, GEPI, &MemI, Idx)) {
909 Instruction *NewGEPI = GEPI->clone();
910 NewGEPI->setOperand(Idx,
911 ConstantInt::get(GEPI->getOperand(Idx)->getType(), 0));
912 NewGEPI->insertBefore(GEPI);
913 MemI.setOperand(MemI.getPointerOperandIndex(), NewGEPI);
921 static bool canSimplifyNullLoadOrGEP(LoadInst &LI, Value *Op) {
922 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
923 const Value *GEPI0 = GEPI->getOperand(0);
924 if (isa<ConstantPointerNull>(GEPI0) && GEPI->getPointerAddressSpace() == 0)
927 if (isa<UndefValue>(Op) ||
928 (isa<ConstantPointerNull>(Op) && LI.getPointerAddressSpace() == 0))
933 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
934 Value *Op = LI.getOperand(0);
936 // Try to canonicalize the loaded type.
937 if (Instruction *Res = combineLoadToOperationType(*this, LI))
940 // Attempt to improve the alignment.
941 unsigned KnownAlign = getOrEnforceKnownAlignment(
942 Op, DL.getPrefTypeAlignment(LI.getType()), DL, &LI, &AC, &DT);
943 unsigned LoadAlign = LI.getAlignment();
944 unsigned EffectiveLoadAlign =
945 LoadAlign != 0 ? LoadAlign : DL.getABITypeAlignment(LI.getType());
947 if (KnownAlign > EffectiveLoadAlign)
948 LI.setAlignment(KnownAlign);
949 else if (LoadAlign == 0)
950 LI.setAlignment(EffectiveLoadAlign);
952 // Replace GEP indices if possible.
953 if (Instruction *NewGEPI = replaceGEPIdxWithZero(*this, Op, LI)) {
954 Worklist.Add(NewGEPI);
958 if (Instruction *Res = unpackLoadToAggregate(*this, LI))
961 // Do really simple store-to-load forwarding and load CSE, to catch cases
962 // where there are several consecutive memory accesses to the same location,
963 // separated by a few arithmetic operations.
964 BasicBlock::iterator BBI(LI);
965 bool IsLoadCSE = false;
966 if (Value *AvailableVal = FindAvailableLoadedValue(
967 &LI, LI.getParent(), BBI, DefMaxInstsToScan, AA, &IsLoadCSE)) {
969 combineMetadataForCSE(cast<LoadInst>(AvailableVal), &LI);
971 return replaceInstUsesWith(
972 LI, Builder->CreateBitOrPointerCast(AvailableVal, LI.getType(),
973 LI.getName() + ".cast"));
976 // None of the following transforms are legal for volatile/ordered atomic
977 // loads. Most of them do apply for unordered atomics.
978 if (!LI.isUnordered()) return nullptr;
980 // load(gep null, ...) -> unreachable
981 // load null/undef -> unreachable
982 // TODO: Consider a target hook for valid address spaces for this xforms.
983 if (canSimplifyNullLoadOrGEP(LI, Op)) {
984 // Insert a new store to null instruction before the load to indicate
985 // that this code is not reachable. We do this instead of inserting
986 // an unreachable instruction directly because we cannot modify the
988 new StoreInst(UndefValue::get(LI.getType()),
989 Constant::getNullValue(Op->getType()), &LI);
990 return replaceInstUsesWith(LI, UndefValue::get(LI.getType()));
993 if (Op->hasOneUse()) {
994 // Change select and PHI nodes to select values instead of addresses: this
995 // helps alias analysis out a lot, allows many others simplifications, and
996 // exposes redundancy in the code.
998 // Note that we cannot do the transformation unless we know that the
999 // introduced loads cannot trap! Something like this is valid as long as
1000 // the condition is always false: load (select bool %C, int* null, int* %G),
1001 // but it would not be valid if we transformed it to load from null
1004 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
1005 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
1006 unsigned Align = LI.getAlignment();
1007 if (isSafeToLoadUnconditionally(SI->getOperand(1), Align, DL, SI) &&
1008 isSafeToLoadUnconditionally(SI->getOperand(2), Align, DL, SI)) {
1009 LoadInst *V1 = Builder->CreateLoad(SI->getOperand(1),
1010 SI->getOperand(1)->getName()+".val");
1011 LoadInst *V2 = Builder->CreateLoad(SI->getOperand(2),
1012 SI->getOperand(2)->getName()+".val");
1013 assert(LI.isUnordered() && "implied by above");
1014 V1->setAlignment(Align);
1015 V1->setAtomic(LI.getOrdering(), LI.getSynchScope());
1016 V2->setAlignment(Align);
1017 V2->setAtomic(LI.getOrdering(), LI.getSynchScope());
1018 return SelectInst::Create(SI->getCondition(), V1, V2);
1021 // load (select (cond, null, P)) -> load P
1022 if (isa<ConstantPointerNull>(SI->getOperand(1)) &&
1023 LI.getPointerAddressSpace() == 0) {
1024 LI.setOperand(0, SI->getOperand(2));
1028 // load (select (cond, P, null)) -> load P
1029 if (isa<ConstantPointerNull>(SI->getOperand(2)) &&
1030 LI.getPointerAddressSpace() == 0) {
1031 LI.setOperand(0, SI->getOperand(1));
1039 /// \brief Look for extractelement/insertvalue sequence that acts like a bitcast.
1041 /// \returns underlying value that was "cast", or nullptr otherwise.
1043 /// For example, if we have:
1045 /// %E0 = extractelement <2 x double> %U, i32 0
1046 /// %V0 = insertvalue [2 x double] undef, double %E0, 0
1047 /// %E1 = extractelement <2 x double> %U, i32 1
1048 /// %V1 = insertvalue [2 x double] %V0, double %E1, 1
1050 /// and the layout of a <2 x double> is isomorphic to a [2 x double],
1051 /// then %V1 can be safely approximated by a conceptual "bitcast" of %U.
1052 /// Note that %U may contain non-undef values where %V1 has undef.
1053 static Value *likeBitCastFromVector(InstCombiner &IC, Value *V) {
1055 while (auto *IV = dyn_cast<InsertValueInst>(V)) {
1056 auto *E = dyn_cast<ExtractElementInst>(IV->getInsertedValueOperand());
1059 auto *W = E->getVectorOperand();
1064 auto *CI = dyn_cast<ConstantInt>(E->getIndexOperand());
1065 if (!CI || IV->getNumIndices() != 1 || CI->getZExtValue() != *IV->idx_begin())
1067 V = IV->getAggregateOperand();
1069 if (!isa<UndefValue>(V) ||!U)
1072 auto *UT = cast<VectorType>(U->getType());
1073 auto *VT = V->getType();
1074 // Check that types UT and VT are bitwise isomorphic.
1075 const auto &DL = IC.getDataLayout();
1076 if (DL.getTypeStoreSizeInBits(UT) != DL.getTypeStoreSizeInBits(VT)) {
1079 if (auto *AT = dyn_cast<ArrayType>(VT)) {
1080 if (AT->getNumElements() != UT->getNumElements())
1083 auto *ST = cast<StructType>(VT);
1084 if (ST->getNumElements() != UT->getNumElements())
1086 for (const auto *EltT : ST->elements()) {
1087 if (EltT != UT->getElementType())
1094 /// \brief Combine stores to match the type of value being stored.
1096 /// The core idea here is that the memory does not have any intrinsic type and
1097 /// where we can we should match the type of a store to the type of value being
1100 /// However, this routine must never change the width of a store or the number of
1101 /// stores as that would introduce a semantic change. This combine is expected to
1102 /// be a semantic no-op which just allows stores to more closely model the types
1103 /// of their incoming values.
1105 /// Currently, we also refuse to change the precise type used for an atomic or
1106 /// volatile store. This is debatable, and might be reasonable to change later.
1107 /// However, it is risky in case some backend or other part of LLVM is relying
1108 /// on the exact type stored to select appropriate atomic operations.
1110 /// \returns true if the store was successfully combined away. This indicates
1111 /// the caller must erase the store instruction. We have to let the caller erase
1112 /// the store instruction as otherwise there is no way to signal whether it was
1113 /// combined or not: IC.EraseInstFromFunction returns a null pointer.
1114 static bool combineStoreToValueType(InstCombiner &IC, StoreInst &SI) {
1115 // FIXME: We could probably with some care handle both volatile and ordered
1116 // atomic stores here but it isn't clear that this is important.
1117 if (!SI.isUnordered())
1120 // swifterror values can't be bitcasted.
1121 if (SI.getPointerOperand()->isSwiftError())
1124 Value *V = SI.getValueOperand();
1126 // Fold away bit casts of the stored value by storing the original type.
1127 if (auto *BC = dyn_cast<BitCastInst>(V)) {
1128 V = BC->getOperand(0);
1129 if (!SI.isAtomic() || isSupportedAtomicType(V->getType())) {
1130 combineStoreToNewValue(IC, SI, V);
1135 if (Value *U = likeBitCastFromVector(IC, V))
1136 if (!SI.isAtomic() || isSupportedAtomicType(U->getType())) {
1137 combineStoreToNewValue(IC, SI, U);
1141 // FIXME: We should also canonicalize stores of vectors when their elements
1142 // are cast to other types.
1146 static bool unpackStoreToAggregate(InstCombiner &IC, StoreInst &SI) {
1147 // FIXME: We could probably with some care handle both volatile and atomic
1148 // stores here but it isn't clear that this is important.
1152 Value *V = SI.getValueOperand();
1153 Type *T = V->getType();
1155 if (!T->isAggregateType())
1158 if (auto *ST = dyn_cast<StructType>(T)) {
1159 // If the struct only have one element, we unpack.
1160 unsigned Count = ST->getNumElements();
1162 V = IC.Builder->CreateExtractValue(V, 0);
1163 combineStoreToNewValue(IC, SI, V);
1167 // We don't want to break loads with padding here as we'd loose
1168 // the knowledge that padding exists for the rest of the pipeline.
1169 const DataLayout &DL = IC.getDataLayout();
1170 auto *SL = DL.getStructLayout(ST);
1171 if (SL->hasPadding())
1174 auto Align = SI.getAlignment();
1176 Align = DL.getABITypeAlignment(ST);
1178 SmallString<16> EltName = V->getName();
1180 auto *Addr = SI.getPointerOperand();
1181 SmallString<16> AddrName = Addr->getName();
1182 AddrName += ".repack";
1184 auto *IdxType = Type::getInt32Ty(ST->getContext());
1185 auto *Zero = ConstantInt::get(IdxType, 0);
1186 for (unsigned i = 0; i < Count; i++) {
1187 Value *Indices[2] = {
1189 ConstantInt::get(IdxType, i),
1191 auto *Ptr = IC.Builder->CreateInBoundsGEP(ST, Addr, makeArrayRef(Indices),
1193 auto *Val = IC.Builder->CreateExtractValue(V, i, EltName);
1194 auto EltAlign = MinAlign(Align, SL->getElementOffset(i));
1195 IC.Builder->CreateAlignedStore(Val, Ptr, EltAlign);
1201 if (auto *AT = dyn_cast<ArrayType>(T)) {
1202 // If the array only have one element, we unpack.
1203 auto NumElements = AT->getNumElements();
1204 if (NumElements == 1) {
1205 V = IC.Builder->CreateExtractValue(V, 0);
1206 combineStoreToNewValue(IC, SI, V);
1210 // Bail out if the array is too large. Ideally we would like to optimize
1211 // arrays of arbitrary size but this has a terrible impact on compile time.
1212 // The threshold here is chosen arbitrarily, maybe needs a little bit of
1214 if (NumElements > IC.MaxArraySizeForCombine)
1217 const DataLayout &DL = IC.getDataLayout();
1218 auto EltSize = DL.getTypeAllocSize(AT->getElementType());
1219 auto Align = SI.getAlignment();
1221 Align = DL.getABITypeAlignment(T);
1223 SmallString<16> EltName = V->getName();
1225 auto *Addr = SI.getPointerOperand();
1226 SmallString<16> AddrName = Addr->getName();
1227 AddrName += ".repack";
1229 auto *IdxType = Type::getInt64Ty(T->getContext());
1230 auto *Zero = ConstantInt::get(IdxType, 0);
1232 uint64_t Offset = 0;
1233 for (uint64_t i = 0; i < NumElements; i++) {
1234 Value *Indices[2] = {
1236 ConstantInt::get(IdxType, i),
1238 auto *Ptr = IC.Builder->CreateInBoundsGEP(AT, Addr, makeArrayRef(Indices),
1240 auto *Val = IC.Builder->CreateExtractValue(V, i, EltName);
1241 auto EltAlign = MinAlign(Align, Offset);
1242 IC.Builder->CreateAlignedStore(Val, Ptr, EltAlign);
1252 /// equivalentAddressValues - Test if A and B will obviously have the same
1253 /// value. This includes recognizing that %t0 and %t1 will have the same
1254 /// value in code like this:
1255 /// %t0 = getelementptr \@a, 0, 3
1256 /// store i32 0, i32* %t0
1257 /// %t1 = getelementptr \@a, 0, 3
1258 /// %t2 = load i32* %t1
1260 static bool equivalentAddressValues(Value *A, Value *B) {
1261 // Test if the values are trivially equivalent.
1262 if (A == B) return true;
1264 // Test if the values come form identical arithmetic instructions.
1265 // This uses isIdenticalToWhenDefined instead of isIdenticalTo because
1266 // its only used to compare two uses within the same basic block, which
1267 // means that they'll always either have the same value or one of them
1268 // will have an undefined value.
1269 if (isa<BinaryOperator>(A) ||
1272 isa<GetElementPtrInst>(A))
1273 if (Instruction *BI = dyn_cast<Instruction>(B))
1274 if (cast<Instruction>(A)->isIdenticalToWhenDefined(BI))
1277 // Otherwise they may not be equivalent.
1281 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
1282 Value *Val = SI.getOperand(0);
1283 Value *Ptr = SI.getOperand(1);
1285 // Try to canonicalize the stored type.
1286 if (combineStoreToValueType(*this, SI))
1287 return eraseInstFromFunction(SI);
1289 // Attempt to improve the alignment.
1290 unsigned KnownAlign = getOrEnforceKnownAlignment(
1291 Ptr, DL.getPrefTypeAlignment(Val->getType()), DL, &SI, &AC, &DT);
1292 unsigned StoreAlign = SI.getAlignment();
1293 unsigned EffectiveStoreAlign =
1294 StoreAlign != 0 ? StoreAlign : DL.getABITypeAlignment(Val->getType());
1296 if (KnownAlign > EffectiveStoreAlign)
1297 SI.setAlignment(KnownAlign);
1298 else if (StoreAlign == 0)
1299 SI.setAlignment(EffectiveStoreAlign);
1301 // Try to canonicalize the stored type.
1302 if (unpackStoreToAggregate(*this, SI))
1303 return eraseInstFromFunction(SI);
1305 // Replace GEP indices if possible.
1306 if (Instruction *NewGEPI = replaceGEPIdxWithZero(*this, Ptr, SI)) {
1307 Worklist.Add(NewGEPI);
1311 // Don't hack volatile/ordered stores.
1312 // FIXME: Some bits are legal for ordered atomic stores; needs refactoring.
1313 if (!SI.isUnordered()) return nullptr;
1315 // If the RHS is an alloca with a single use, zapify the store, making the
1317 if (Ptr->hasOneUse()) {
1318 if (isa<AllocaInst>(Ptr))
1319 return eraseInstFromFunction(SI);
1320 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
1321 if (isa<AllocaInst>(GEP->getOperand(0))) {
1322 if (GEP->getOperand(0)->hasOneUse())
1323 return eraseInstFromFunction(SI);
1328 // Do really simple DSE, to catch cases where there are several consecutive
1329 // stores to the same location, separated by a few arithmetic operations. This
1330 // situation often occurs with bitfield accesses.
1331 BasicBlock::iterator BBI(SI);
1332 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
1335 // Don't count debug info directives, lest they affect codegen,
1336 // and we skip pointer-to-pointer bitcasts, which are NOPs.
1337 if (isa<DbgInfoIntrinsic>(BBI) ||
1338 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy())) {
1343 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
1344 // Prev store isn't volatile, and stores to the same location?
1345 if (PrevSI->isUnordered() && equivalentAddressValues(PrevSI->getOperand(1),
1346 SI.getOperand(1))) {
1349 eraseInstFromFunction(*PrevSI);
1355 // If this is a load, we have to stop. However, if the loaded value is from
1356 // the pointer we're loading and is producing the pointer we're storing,
1357 // then *this* store is dead (X = load P; store X -> P).
1358 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
1359 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr)) {
1360 assert(SI.isUnordered() && "can't eliminate ordering operation");
1361 return eraseInstFromFunction(SI);
1364 // Otherwise, this is a load from some other location. Stores before it
1369 // Don't skip over loads, throws or things that can modify memory.
1370 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory() || BBI->mayThrow())
1374 // store X, null -> turns into 'unreachable' in SimplifyCFG
1375 if (isa<ConstantPointerNull>(Ptr) && SI.getPointerAddressSpace() == 0) {
1376 if (!isa<UndefValue>(Val)) {
1377 SI.setOperand(0, UndefValue::get(Val->getType()));
1378 if (Instruction *U = dyn_cast<Instruction>(Val))
1379 Worklist.Add(U); // Dropped a use.
1381 return nullptr; // Do not modify these!
1384 // store undef, Ptr -> noop
1385 if (isa<UndefValue>(Val))
1386 return eraseInstFromFunction(SI);
1388 // If this store is the last instruction in the basic block (possibly
1389 // excepting debug info instructions), and if the block ends with an
1390 // unconditional branch, try to move it to the successor block.
1391 BBI = SI.getIterator();
1394 } while (isa<DbgInfoIntrinsic>(BBI) ||
1395 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy()));
1396 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
1397 if (BI->isUnconditional())
1398 if (SimplifyStoreAtEndOfBlock(SI))
1399 return nullptr; // xform done!
1404 /// SimplifyStoreAtEndOfBlock - Turn things like:
1405 /// if () { *P = v1; } else { *P = v2 }
1406 /// into a phi node with a store in the successor.
1408 /// Simplify things like:
1409 /// *P = v1; if () { *P = v2; }
1410 /// into a phi node with a store in the successor.
1412 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
1413 assert(SI.isUnordered() &&
1414 "this code has not been auditted for volatile or ordered store case");
1416 BasicBlock *StoreBB = SI.getParent();
1418 // Check to see if the successor block has exactly two incoming edges. If
1419 // so, see if the other predecessor contains a store to the same location.
1420 // if so, insert a PHI node (if needed) and move the stores down.
1421 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
1423 // Determine whether Dest has exactly two predecessors and, if so, compute
1424 // the other predecessor.
1425 pred_iterator PI = pred_begin(DestBB);
1426 BasicBlock *P = *PI;
1427 BasicBlock *OtherBB = nullptr;
1432 if (++PI == pred_end(DestBB))
1441 if (++PI != pred_end(DestBB))
1444 // Bail out if all the relevant blocks aren't distinct (this can happen,
1445 // for example, if SI is in an infinite loop)
1446 if (StoreBB == DestBB || OtherBB == DestBB)
1449 // Verify that the other block ends in a branch and is not otherwise empty.
1450 BasicBlock::iterator BBI(OtherBB->getTerminator());
1451 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
1452 if (!OtherBr || BBI == OtherBB->begin())
1455 // If the other block ends in an unconditional branch, check for the 'if then
1456 // else' case. there is an instruction before the branch.
1457 StoreInst *OtherStore = nullptr;
1458 if (OtherBr->isUnconditional()) {
1460 // Skip over debugging info.
1461 while (isa<DbgInfoIntrinsic>(BBI) ||
1462 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy())) {
1463 if (BBI==OtherBB->begin())
1467 // If this isn't a store, isn't a store to the same location, or is not the
1468 // right kind of store, bail out.
1469 OtherStore = dyn_cast<StoreInst>(BBI);
1470 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1) ||
1471 !SI.isSameOperationAs(OtherStore))
1474 // Otherwise, the other block ended with a conditional branch. If one of the
1475 // destinations is StoreBB, then we have the if/then case.
1476 if (OtherBr->getSuccessor(0) != StoreBB &&
1477 OtherBr->getSuccessor(1) != StoreBB)
1480 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
1481 // if/then triangle. See if there is a store to the same ptr as SI that
1482 // lives in OtherBB.
1484 // Check to see if we find the matching store.
1485 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
1486 if (OtherStore->getOperand(1) != SI.getOperand(1) ||
1487 !SI.isSameOperationAs(OtherStore))
1491 // If we find something that may be using or overwriting the stored
1492 // value, or if we run out of instructions, we can't do the xform.
1493 if (BBI->mayReadFromMemory() || BBI->mayThrow() ||
1494 BBI->mayWriteToMemory() || BBI == OtherBB->begin())
1498 // In order to eliminate the store in OtherBr, we have to
1499 // make sure nothing reads or overwrites the stored value in
1501 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
1502 // FIXME: This should really be AA driven.
1503 if (I->mayReadFromMemory() || I->mayThrow() || I->mayWriteToMemory())
1508 // Insert a PHI node now if we need it.
1509 Value *MergedVal = OtherStore->getOperand(0);
1510 if (MergedVal != SI.getOperand(0)) {
1511 PHINode *PN = PHINode::Create(MergedVal->getType(), 2, "storemerge");
1512 PN->addIncoming(SI.getOperand(0), SI.getParent());
1513 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
1514 MergedVal = InsertNewInstBefore(PN, DestBB->front());
1517 // Advance to a place where it is safe to insert the new store and
1519 BBI = DestBB->getFirstInsertionPt();
1520 StoreInst *NewSI = new StoreInst(MergedVal, SI.getOperand(1),
1524 SI.getSynchScope());
1525 InsertNewInstBefore(NewSI, *BBI);
1526 // The debug locations of the original instructions might differ; merge them.
1527 NewSI->setDebugLoc(DILocation::getMergedLocation(SI.getDebugLoc(),
1528 OtherStore->getDebugLoc()));
1530 // If the two stores had AA tags, merge them.
1532 SI.getAAMetadata(AATags);
1534 OtherStore->getAAMetadata(AATags, /* Merge = */ true);
1535 NewSI->setAAMetadata(AATags);
1538 // Nuke the old stores.
1539 eraseInstFromFunction(SI);
1540 eraseInstFromFunction(*OtherStore);