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 static Instruction *simplifyAllocaArraySize(InstCombiner &IC, AllocaInst &AI) {
173 // Check for array size of 1 (scalar allocation).
174 if (!AI.isArrayAllocation()) {
175 // i32 1 is the canonical array size for scalar allocations.
176 if (AI.getArraySize()->getType()->isIntegerTy(32))
180 Value *V = IC.Builder->getInt32(1);
185 // Convert: alloca Ty, C - where C is a constant != 1 into: alloca [C x Ty], 1
186 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
187 Type *NewTy = ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
188 AllocaInst *New = IC.Builder->CreateAlloca(NewTy, nullptr, AI.getName());
189 New->setAlignment(AI.getAlignment());
191 // Scan to the end of the allocation instructions, to skip over a block of
192 // allocas if possible...also skip interleaved debug info
194 BasicBlock::iterator It(New);
195 while (isa<AllocaInst>(*It) || isa<DbgInfoIntrinsic>(*It))
198 // Now that I is pointing to the first non-allocation-inst in the block,
199 // insert our getelementptr instruction...
201 Type *IdxTy = IC.getDataLayout().getIntPtrType(AI.getType());
202 Value *NullIdx = Constant::getNullValue(IdxTy);
203 Value *Idx[2] = {NullIdx, NullIdx};
205 GetElementPtrInst::CreateInBounds(New, Idx, New->getName() + ".sub");
206 IC.InsertNewInstBefore(GEP, *It);
208 // Now make everything use the getelementptr instead of the original
210 return IC.replaceInstUsesWith(AI, GEP);
213 if (isa<UndefValue>(AI.getArraySize()))
214 return IC.replaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
216 // Ensure that the alloca array size argument has type intptr_t, so that
217 // any casting is exposed early.
218 Type *IntPtrTy = IC.getDataLayout().getIntPtrType(AI.getType());
219 if (AI.getArraySize()->getType() != IntPtrTy) {
220 Value *V = IC.Builder->CreateIntCast(AI.getArraySize(), IntPtrTy, false);
229 // If I and V are pointers in different address space, it is not allowed to
230 // use replaceAllUsesWith since I and V have different types. A
231 // non-target-specific transformation should not use addrspacecast on V since
232 // the two address space may be disjoint depending on target.
234 // This class chases down uses of the old pointer until reaching the load
235 // instructions, then replaces the old pointer in the load instructions with
236 // the new pointer. If during the chasing it sees bitcast or GEP, it will
237 // create new bitcast or GEP with the new pointer and use them in the load
239 class PointerReplacer {
241 PointerReplacer(InstCombiner &IC) : IC(IC) {}
242 void replacePointer(Instruction &I, Value *V);
245 void findLoadAndReplace(Instruction &I);
246 void replace(Instruction *I);
247 Value *getReplacement(Value *I);
249 SmallVector<Instruction *, 4> Path;
250 MapVector<Value *, Value *> WorkMap;
253 } // end anonymous namespace
255 void PointerReplacer::findLoadAndReplace(Instruction &I) {
256 for (auto U : I.users()) {
257 auto *Inst = dyn_cast<Instruction>(&*U);
260 DEBUG(dbgs() << "Found pointer user: " << *U << '\n');
261 if (isa<LoadInst>(Inst)) {
265 } else if (isa<GetElementPtrInst>(Inst) || isa<BitCastInst>(Inst)) {
266 Path.push_back(Inst);
267 findLoadAndReplace(*Inst);
275 Value *PointerReplacer::getReplacement(Value *V) {
276 auto Loc = WorkMap.find(V);
277 if (Loc != WorkMap.end())
282 void PointerReplacer::replace(Instruction *I) {
283 if (getReplacement(I))
286 if (auto *LT = dyn_cast<LoadInst>(I)) {
287 auto *V = getReplacement(LT->getPointerOperand());
288 assert(V && "Operand not replaced");
289 auto *NewI = new LoadInst(V);
291 IC.InsertNewInstWith(NewI, *LT);
292 IC.replaceInstUsesWith(*LT, NewI);
294 } else if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) {
295 auto *V = getReplacement(GEP->getPointerOperand());
296 assert(V && "Operand not replaced");
297 SmallVector<Value *, 8> Indices;
298 Indices.append(GEP->idx_begin(), GEP->idx_end());
299 auto *NewI = GetElementPtrInst::Create(
300 V->getType()->getPointerElementType(), V, Indices);
301 IC.InsertNewInstWith(NewI, *GEP);
304 } else if (auto *BC = dyn_cast<BitCastInst>(I)) {
305 auto *V = getReplacement(BC->getOperand(0));
306 assert(V && "Operand not replaced");
307 auto *NewT = PointerType::get(BC->getType()->getPointerElementType(),
308 V->getType()->getPointerAddressSpace());
309 auto *NewI = new BitCastInst(V, NewT);
310 IC.InsertNewInstWith(NewI, *BC);
314 llvm_unreachable("should never reach here");
318 void PointerReplacer::replacePointer(Instruction &I, Value *V) {
320 auto *PT = cast<PointerType>(I.getType());
321 auto *NT = cast<PointerType>(V->getType());
322 assert(PT != NT && PT->getElementType() == NT->getElementType() &&
326 findLoadAndReplace(I);
329 Instruction *InstCombiner::visitAllocaInst(AllocaInst &AI) {
330 if (auto *I = simplifyAllocaArraySize(*this, AI))
333 if (AI.getAllocatedType()->isSized()) {
334 // If the alignment is 0 (unspecified), assign it the preferred alignment.
335 if (AI.getAlignment() == 0)
336 AI.setAlignment(DL.getPrefTypeAlignment(AI.getAllocatedType()));
338 // Move all alloca's of zero byte objects to the entry block and merge them
339 // together. Note that we only do this for alloca's, because malloc should
340 // allocate and return a unique pointer, even for a zero byte allocation.
341 if (DL.getTypeAllocSize(AI.getAllocatedType()) == 0) {
342 // For a zero sized alloca there is no point in doing an array allocation.
343 // This is helpful if the array size is a complicated expression not used
345 if (AI.isArrayAllocation()) {
346 AI.setOperand(0, ConstantInt::get(AI.getArraySize()->getType(), 1));
350 // Get the first instruction in the entry block.
351 BasicBlock &EntryBlock = AI.getParent()->getParent()->getEntryBlock();
352 Instruction *FirstInst = EntryBlock.getFirstNonPHIOrDbg();
353 if (FirstInst != &AI) {
354 // If the entry block doesn't start with a zero-size alloca then move
355 // this one to the start of the entry block. There is no problem with
356 // dominance as the array size was forced to a constant earlier already.
357 AllocaInst *EntryAI = dyn_cast<AllocaInst>(FirstInst);
358 if (!EntryAI || !EntryAI->getAllocatedType()->isSized() ||
359 DL.getTypeAllocSize(EntryAI->getAllocatedType()) != 0) {
360 AI.moveBefore(FirstInst);
364 // If the alignment of the entry block alloca is 0 (unspecified),
365 // assign it the preferred alignment.
366 if (EntryAI->getAlignment() == 0)
367 EntryAI->setAlignment(
368 DL.getPrefTypeAlignment(EntryAI->getAllocatedType()));
369 // Replace this zero-sized alloca with the one at the start of the entry
370 // block after ensuring that the address will be aligned enough for both
372 unsigned MaxAlign = std::max(EntryAI->getAlignment(),
374 EntryAI->setAlignment(MaxAlign);
375 if (AI.getType() != EntryAI->getType())
376 return new BitCastInst(EntryAI, AI.getType());
377 return replaceInstUsesWith(AI, EntryAI);
382 if (AI.getAlignment()) {
383 // Check to see if this allocation is only modified by a memcpy/memmove from
384 // a constant global whose alignment is equal to or exceeds that of the
385 // allocation. If this is the case, we can change all users to use
386 // the constant global instead. This is commonly produced by the CFE by
387 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
388 // is only subsequently read.
389 SmallVector<Instruction *, 4> ToDelete;
390 if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(&AI, ToDelete)) {
391 unsigned SourceAlign = getOrEnforceKnownAlignment(
392 Copy->getSource(), AI.getAlignment(), DL, &AI, &AC, &DT);
393 if (AI.getAlignment() <= SourceAlign) {
394 DEBUG(dbgs() << "Found alloca equal to global: " << AI << '\n');
395 DEBUG(dbgs() << " memcpy = " << *Copy << '\n');
396 for (unsigned i = 0, e = ToDelete.size(); i != e; ++i)
397 eraseInstFromFunction(*ToDelete[i]);
398 Constant *TheSrc = cast<Constant>(Copy->getSource());
399 auto *SrcTy = TheSrc->getType();
400 auto *DestTy = PointerType::get(AI.getType()->getPointerElementType(),
401 SrcTy->getPointerAddressSpace());
403 ConstantExpr::getPointerBitCastOrAddrSpaceCast(TheSrc, DestTy);
404 if (AI.getType()->getPointerAddressSpace() ==
405 SrcTy->getPointerAddressSpace()) {
406 Instruction *NewI = replaceInstUsesWith(AI, Cast);
407 eraseInstFromFunction(*Copy);
411 PointerReplacer PtrReplacer(*this);
412 PtrReplacer.replacePointer(AI, Cast);
419 // At last, use the generic allocation site handler to aggressively remove
421 return visitAllocSite(AI);
424 // Are we allowed to form a atomic load or store of this type?
425 static bool isSupportedAtomicType(Type *Ty) {
426 return Ty->isIntegerTy() || Ty->isPointerTy() || Ty->isFloatingPointTy();
429 /// \brief Helper to combine a load to a new type.
431 /// This just does the work of combining a load to a new type. It handles
432 /// metadata, etc., and returns the new instruction. The \c NewTy should be the
433 /// loaded *value* type. This will convert it to a pointer, cast the operand to
434 /// that pointer type, load it, etc.
436 /// Note that this will create all of the instructions with whatever insert
437 /// point the \c InstCombiner currently is using.
438 static LoadInst *combineLoadToNewType(InstCombiner &IC, LoadInst &LI, Type *NewTy,
439 const Twine &Suffix = "") {
440 assert((!LI.isAtomic() || isSupportedAtomicType(NewTy)) &&
441 "can't fold an atomic load to requested type");
443 Value *Ptr = LI.getPointerOperand();
444 unsigned AS = LI.getPointerAddressSpace();
445 SmallVector<std::pair<unsigned, MDNode *>, 8> MD;
446 LI.getAllMetadata(MD);
448 LoadInst *NewLoad = IC.Builder->CreateAlignedLoad(
449 IC.Builder->CreateBitCast(Ptr, NewTy->getPointerTo(AS)),
450 LI.getAlignment(), LI.isVolatile(), LI.getName() + Suffix);
451 NewLoad->setAtomic(LI.getOrdering(), LI.getSynchScope());
452 MDBuilder MDB(NewLoad->getContext());
453 for (const auto &MDPair : MD) {
454 unsigned ID = MDPair.first;
455 MDNode *N = MDPair.second;
456 // Note, essentially every kind of metadata should be preserved here! This
457 // routine is supposed to clone a load instruction changing *only its type*.
458 // The only metadata it makes sense to drop is metadata which is invalidated
459 // when the pointer type changes. This should essentially never be the case
460 // in LLVM, but we explicitly switch over only known metadata to be
461 // conservatively correct. If you are adding metadata to LLVM which pertains
462 // to loads, you almost certainly want to add it here.
464 case LLVMContext::MD_dbg:
465 case LLVMContext::MD_tbaa:
466 case LLVMContext::MD_prof:
467 case LLVMContext::MD_fpmath:
468 case LLVMContext::MD_tbaa_struct:
469 case LLVMContext::MD_invariant_load:
470 case LLVMContext::MD_alias_scope:
471 case LLVMContext::MD_noalias:
472 case LLVMContext::MD_nontemporal:
473 case LLVMContext::MD_mem_parallel_loop_access:
474 // All of these directly apply.
475 NewLoad->setMetadata(ID, N);
478 case LLVMContext::MD_nonnull:
479 // This only directly applies if the new type is also a pointer.
480 if (NewTy->isPointerTy()) {
481 NewLoad->setMetadata(ID, N);
484 // If it's integral now, translate it to !range metadata.
485 if (NewTy->isIntegerTy()) {
486 auto *ITy = cast<IntegerType>(NewTy);
487 auto *NullInt = ConstantExpr::getPtrToInt(
488 ConstantPointerNull::get(cast<PointerType>(Ptr->getType())), ITy);
490 ConstantExpr::getAdd(NullInt, ConstantInt::get(ITy, 1));
491 NewLoad->setMetadata(LLVMContext::MD_range,
492 MDB.createRange(NonNullInt, NullInt));
495 case LLVMContext::MD_align:
496 case LLVMContext::MD_dereferenceable:
497 case LLVMContext::MD_dereferenceable_or_null:
498 // These only directly apply if the new type is also a pointer.
499 if (NewTy->isPointerTy())
500 NewLoad->setMetadata(ID, N);
502 case LLVMContext::MD_range:
503 // FIXME: It would be nice to propagate this in some way, but the type
504 // conversions make it hard.
506 // If it's a pointer now and the range does not contain 0, make it !nonnull.
507 if (NewTy->isPointerTy()) {
508 unsigned BitWidth = IC.getDataLayout().getTypeSizeInBits(NewTy);
509 if (!getConstantRangeFromMetadata(*N).contains(APInt(BitWidth, 0))) {
510 MDNode *NN = MDNode::get(LI.getContext(), None);
511 NewLoad->setMetadata(LLVMContext::MD_nonnull, NN);
520 /// \brief Combine a store to a new type.
522 /// Returns the newly created store instruction.
523 static StoreInst *combineStoreToNewValue(InstCombiner &IC, StoreInst &SI, Value *V) {
524 assert((!SI.isAtomic() || isSupportedAtomicType(V->getType())) &&
525 "can't fold an atomic store of requested type");
527 Value *Ptr = SI.getPointerOperand();
528 unsigned AS = SI.getPointerAddressSpace();
529 SmallVector<std::pair<unsigned, MDNode *>, 8> MD;
530 SI.getAllMetadata(MD);
532 StoreInst *NewStore = IC.Builder->CreateAlignedStore(
533 V, IC.Builder->CreateBitCast(Ptr, V->getType()->getPointerTo(AS)),
534 SI.getAlignment(), SI.isVolatile());
535 NewStore->setAtomic(SI.getOrdering(), SI.getSynchScope());
536 for (const auto &MDPair : MD) {
537 unsigned ID = MDPair.first;
538 MDNode *N = MDPair.second;
539 // Note, essentially every kind of metadata should be preserved here! This
540 // routine is supposed to clone a store instruction changing *only its
541 // type*. The only metadata it makes sense to drop is metadata which is
542 // invalidated when the pointer type changes. This should essentially
543 // never be the case in LLVM, but we explicitly switch over only known
544 // metadata to be conservatively correct. If you are adding metadata to
545 // LLVM which pertains to stores, you almost certainly want to add it
548 case LLVMContext::MD_dbg:
549 case LLVMContext::MD_tbaa:
550 case LLVMContext::MD_prof:
551 case LLVMContext::MD_fpmath:
552 case LLVMContext::MD_tbaa_struct:
553 case LLVMContext::MD_alias_scope:
554 case LLVMContext::MD_noalias:
555 case LLVMContext::MD_nontemporal:
556 case LLVMContext::MD_mem_parallel_loop_access:
557 // All of these directly apply.
558 NewStore->setMetadata(ID, N);
561 case LLVMContext::MD_invariant_load:
562 case LLVMContext::MD_nonnull:
563 case LLVMContext::MD_range:
564 case LLVMContext::MD_align:
565 case LLVMContext::MD_dereferenceable:
566 case LLVMContext::MD_dereferenceable_or_null:
567 // These don't apply for stores.
575 /// \brief Combine loads to match the type of their uses' value after looking
576 /// through intervening bitcasts.
578 /// The core idea here is that if the result of a load is used in an operation,
579 /// we should load the type most conducive to that operation. For example, when
580 /// loading an integer and converting that immediately to a pointer, we should
581 /// instead directly load a pointer.
583 /// However, this routine must never change the width of a load or the number of
584 /// loads as that would introduce a semantic change. This combine is expected to
585 /// be a semantic no-op which just allows loads to more closely model the types
586 /// of their consuming operations.
588 /// Currently, we also refuse to change the precise type used for an atomic load
589 /// or a volatile load. This is debatable, and might be reasonable to change
590 /// later. However, it is risky in case some backend or other part of LLVM is
591 /// relying on the exact type loaded to select appropriate atomic operations.
592 static Instruction *combineLoadToOperationType(InstCombiner &IC, LoadInst &LI) {
593 // FIXME: We could probably with some care handle both volatile and ordered
594 // atomic loads here but it isn't clear that this is important.
595 if (!LI.isUnordered())
601 // swifterror values can't be bitcasted.
602 if (LI.getPointerOperand()->isSwiftError())
605 Type *Ty = LI.getType();
606 const DataLayout &DL = IC.getDataLayout();
608 // Try to canonicalize loads which are only ever stored to operate over
609 // integers instead of any other type. We only do this when the loaded type
610 // is sized and has a size exactly the same as its store size and the store
611 // size is a legal integer type.
612 if (!Ty->isIntegerTy() && Ty->isSized() &&
613 DL.isLegalInteger(DL.getTypeStoreSizeInBits(Ty)) &&
614 DL.getTypeStoreSizeInBits(Ty) == DL.getTypeSizeInBits(Ty) &&
615 !DL.isNonIntegralPointerType(Ty)) {
616 if (all_of(LI.users(), [&LI](User *U) {
617 auto *SI = dyn_cast<StoreInst>(U);
618 return SI && SI->getPointerOperand() != &LI &&
619 !SI->getPointerOperand()->isSwiftError();
621 LoadInst *NewLoad = combineLoadToNewType(
623 Type::getIntNTy(LI.getContext(), DL.getTypeStoreSizeInBits(Ty)));
624 // Replace all the stores with stores of the newly loaded value.
625 for (auto UI = LI.user_begin(), UE = LI.user_end(); UI != UE;) {
626 auto *SI = cast<StoreInst>(*UI++);
627 IC.Builder->SetInsertPoint(SI);
628 combineStoreToNewValue(IC, *SI, NewLoad);
629 IC.eraseInstFromFunction(*SI);
631 assert(LI.use_empty() && "Failed to remove all users of the load!");
632 // Return the old load so the combiner can delete it safely.
637 // Fold away bit casts of the loaded value by loading the desired type.
638 // We can do this for BitCastInsts as well as casts from and to pointer types,
639 // as long as those are noops (i.e., the source or dest type have the same
640 // bitwidth as the target's pointers).
642 if (auto* CI = dyn_cast<CastInst>(LI.user_back()))
643 if (CI->isNoopCast(DL))
644 if (!LI.isAtomic() || isSupportedAtomicType(CI->getDestTy())) {
645 LoadInst *NewLoad = combineLoadToNewType(IC, LI, CI->getDestTy());
646 CI->replaceAllUsesWith(NewLoad);
647 IC.eraseInstFromFunction(*CI);
651 // FIXME: We should also canonicalize loads of vectors when their elements are
652 // cast to other types.
656 static Instruction *unpackLoadToAggregate(InstCombiner &IC, LoadInst &LI) {
657 // FIXME: We could probably with some care handle both volatile and atomic
658 // stores here but it isn't clear that this is important.
662 Type *T = LI.getType();
663 if (!T->isAggregateType())
666 StringRef Name = LI.getName();
667 assert(LI.getAlignment() && "Alignment must be set at this point");
669 if (auto *ST = dyn_cast<StructType>(T)) {
670 // If the struct only have one element, we unpack.
671 auto NumElements = ST->getNumElements();
672 if (NumElements == 1) {
673 LoadInst *NewLoad = combineLoadToNewType(IC, LI, ST->getTypeAtIndex(0U),
675 return IC.replaceInstUsesWith(LI, IC.Builder->CreateInsertValue(
676 UndefValue::get(T), NewLoad, 0, Name));
679 // We don't want to break loads with padding here as we'd loose
680 // the knowledge that padding exists for the rest of the pipeline.
681 const DataLayout &DL = IC.getDataLayout();
682 auto *SL = DL.getStructLayout(ST);
683 if (SL->hasPadding())
686 auto Align = LI.getAlignment();
688 Align = DL.getABITypeAlignment(ST);
690 auto *Addr = LI.getPointerOperand();
691 auto *IdxType = Type::getInt32Ty(T->getContext());
692 auto *Zero = ConstantInt::get(IdxType, 0);
694 Value *V = UndefValue::get(T);
695 for (unsigned i = 0; i < NumElements; i++) {
696 Value *Indices[2] = {
698 ConstantInt::get(IdxType, i),
700 auto *Ptr = IC.Builder->CreateInBoundsGEP(ST, Addr, makeArrayRef(Indices),
702 auto EltAlign = MinAlign(Align, SL->getElementOffset(i));
703 auto *L = IC.Builder->CreateAlignedLoad(Ptr, EltAlign, Name + ".unpack");
704 V = IC.Builder->CreateInsertValue(V, L, i);
708 return IC.replaceInstUsesWith(LI, V);
711 if (auto *AT = dyn_cast<ArrayType>(T)) {
712 auto *ET = AT->getElementType();
713 auto NumElements = AT->getNumElements();
714 if (NumElements == 1) {
715 LoadInst *NewLoad = combineLoadToNewType(IC, LI, ET, ".unpack");
716 return IC.replaceInstUsesWith(LI, IC.Builder->CreateInsertValue(
717 UndefValue::get(T), NewLoad, 0, Name));
720 // Bail out if the array is too large. Ideally we would like to optimize
721 // arrays of arbitrary size but this has a terrible impact on compile time.
722 // The threshold here is chosen arbitrarily, maybe needs a little bit of
724 if (NumElements > IC.MaxArraySizeForCombine)
727 const DataLayout &DL = IC.getDataLayout();
728 auto EltSize = DL.getTypeAllocSize(ET);
729 auto Align = LI.getAlignment();
731 Align = DL.getABITypeAlignment(T);
733 auto *Addr = LI.getPointerOperand();
734 auto *IdxType = Type::getInt64Ty(T->getContext());
735 auto *Zero = ConstantInt::get(IdxType, 0);
737 Value *V = UndefValue::get(T);
739 for (uint64_t i = 0; i < NumElements; i++) {
740 Value *Indices[2] = {
742 ConstantInt::get(IdxType, i),
744 auto *Ptr = IC.Builder->CreateInBoundsGEP(AT, Addr, makeArrayRef(Indices),
746 auto *L = IC.Builder->CreateAlignedLoad(Ptr, MinAlign(Align, Offset),
748 V = IC.Builder->CreateInsertValue(V, L, i);
753 return IC.replaceInstUsesWith(LI, V);
759 // If we can determine that all possible objects pointed to by the provided
760 // pointer value are, not only dereferenceable, but also definitively less than
761 // or equal to the provided maximum size, then return true. Otherwise, return
762 // false (constant global values and allocas fall into this category).
764 // FIXME: This should probably live in ValueTracking (or similar).
765 static bool isObjectSizeLessThanOrEq(Value *V, uint64_t MaxSize,
766 const DataLayout &DL) {
767 SmallPtrSet<Value *, 4> Visited;
768 SmallVector<Value *, 4> Worklist(1, V);
771 Value *P = Worklist.pop_back_val();
772 P = P->stripPointerCasts();
774 if (!Visited.insert(P).second)
777 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
778 Worklist.push_back(SI->getTrueValue());
779 Worklist.push_back(SI->getFalseValue());
783 if (PHINode *PN = dyn_cast<PHINode>(P)) {
784 for (Value *IncValue : PN->incoming_values())
785 Worklist.push_back(IncValue);
789 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(P)) {
790 if (GA->isInterposable())
792 Worklist.push_back(GA->getAliasee());
796 // If we know how big this object is, and it is less than MaxSize, continue
797 // searching. Otherwise, return false.
798 if (AllocaInst *AI = dyn_cast<AllocaInst>(P)) {
799 if (!AI->getAllocatedType()->isSized())
802 ConstantInt *CS = dyn_cast<ConstantInt>(AI->getArraySize());
806 uint64_t TypeSize = DL.getTypeAllocSize(AI->getAllocatedType());
807 // Make sure that, even if the multiplication below would wrap as an
808 // uint64_t, we still do the right thing.
809 if ((CS->getValue().zextOrSelf(128)*APInt(128, TypeSize)).ugt(MaxSize))
814 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
815 if (!GV->hasDefinitiveInitializer() || !GV->isConstant())
818 uint64_t InitSize = DL.getTypeAllocSize(GV->getValueType());
819 if (InitSize > MaxSize)
825 } while (!Worklist.empty());
830 // If we're indexing into an object of a known size, and the outer index is
831 // not a constant, but having any value but zero would lead to undefined
832 // behavior, replace it with zero.
834 // For example, if we have:
835 // @f.a = private unnamed_addr constant [1 x i32] [i32 12], align 4
837 // %arrayidx = getelementptr inbounds [1 x i32]* @f.a, i64 0, i64 %x
838 // ... = load i32* %arrayidx, align 4
839 // Then we know that we can replace %x in the GEP with i64 0.
841 // FIXME: We could fold any GEP index to zero that would cause UB if it were
842 // not zero. Currently, we only handle the first such index. Also, we could
843 // also search through non-zero constant indices if we kept track of the
844 // offsets those indices implied.
845 static bool canReplaceGEPIdxWithZero(InstCombiner &IC, GetElementPtrInst *GEPI,
846 Instruction *MemI, unsigned &Idx) {
847 if (GEPI->getNumOperands() < 2)
850 // Find the first non-zero index of a GEP. If all indices are zero, return
851 // one past the last index.
852 auto FirstNZIdx = [](const GetElementPtrInst *GEPI) {
854 for (unsigned IE = GEPI->getNumOperands(); I != IE; ++I) {
855 Value *V = GEPI->getOperand(I);
856 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V))
866 // Skip through initial 'zero' indices, and find the corresponding pointer
867 // type. See if the next index is not a constant.
868 Idx = FirstNZIdx(GEPI);
869 if (Idx == GEPI->getNumOperands())
871 if (isa<Constant>(GEPI->getOperand(Idx)))
874 SmallVector<Value *, 4> Ops(GEPI->idx_begin(), GEPI->idx_begin() + Idx);
876 GetElementPtrInst::getIndexedType(GEPI->getSourceElementType(), Ops);
877 if (!AllocTy || !AllocTy->isSized())
879 const DataLayout &DL = IC.getDataLayout();
880 uint64_t TyAllocSize = DL.getTypeAllocSize(AllocTy);
882 // If there are more indices after the one we might replace with a zero, make
883 // sure they're all non-negative. If any of them are negative, the overall
884 // address being computed might be before the base address determined by the
885 // first non-zero index.
886 auto IsAllNonNegative = [&]() {
887 for (unsigned i = Idx+1, e = GEPI->getNumOperands(); i != e; ++i) {
888 KnownBits Known = IC.computeKnownBits(GEPI->getOperand(i), 0, MemI);
889 if (Known.isNonNegative())
897 // FIXME: If the GEP is not inbounds, and there are extra indices after the
898 // one we'll replace, those could cause the address computation to wrap
899 // (rendering the IsAllNonNegative() check below insufficient). We can do
900 // better, ignoring zero indices (and other indices we can prove small
901 // enough not to wrap).
902 if (Idx+1 != GEPI->getNumOperands() && !GEPI->isInBounds())
905 // Note that isObjectSizeLessThanOrEq will return true only if the pointer is
906 // also known to be dereferenceable.
907 return isObjectSizeLessThanOrEq(GEPI->getOperand(0), TyAllocSize, DL) &&
911 // If we're indexing into an object with a variable index for the memory
912 // access, but the object has only one element, we can assume that the index
913 // will always be zero. If we replace the GEP, return it.
914 template <typename T>
915 static Instruction *replaceGEPIdxWithZero(InstCombiner &IC, Value *Ptr,
917 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Ptr)) {
919 if (canReplaceGEPIdxWithZero(IC, GEPI, &MemI, Idx)) {
920 Instruction *NewGEPI = GEPI->clone();
921 NewGEPI->setOperand(Idx,
922 ConstantInt::get(GEPI->getOperand(Idx)->getType(), 0));
923 NewGEPI->insertBefore(GEPI);
924 MemI.setOperand(MemI.getPointerOperandIndex(), NewGEPI);
932 static bool canSimplifyNullLoadOrGEP(LoadInst &LI, Value *Op) {
933 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
934 const Value *GEPI0 = GEPI->getOperand(0);
935 if (isa<ConstantPointerNull>(GEPI0) && GEPI->getPointerAddressSpace() == 0)
938 if (isa<UndefValue>(Op) ||
939 (isa<ConstantPointerNull>(Op) && LI.getPointerAddressSpace() == 0))
944 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
945 Value *Op = LI.getOperand(0);
947 // Try to canonicalize the loaded type.
948 if (Instruction *Res = combineLoadToOperationType(*this, LI))
951 // Attempt to improve the alignment.
952 unsigned KnownAlign = getOrEnforceKnownAlignment(
953 Op, DL.getPrefTypeAlignment(LI.getType()), DL, &LI, &AC, &DT);
954 unsigned LoadAlign = LI.getAlignment();
955 unsigned EffectiveLoadAlign =
956 LoadAlign != 0 ? LoadAlign : DL.getABITypeAlignment(LI.getType());
958 if (KnownAlign > EffectiveLoadAlign)
959 LI.setAlignment(KnownAlign);
960 else if (LoadAlign == 0)
961 LI.setAlignment(EffectiveLoadAlign);
963 // Replace GEP indices if possible.
964 if (Instruction *NewGEPI = replaceGEPIdxWithZero(*this, Op, LI)) {
965 Worklist.Add(NewGEPI);
969 if (Instruction *Res = unpackLoadToAggregate(*this, LI))
972 // Do really simple store-to-load forwarding and load CSE, to catch cases
973 // where there are several consecutive memory accesses to the same location,
974 // separated by a few arithmetic operations.
975 BasicBlock::iterator BBI(LI);
976 bool IsLoadCSE = false;
977 if (Value *AvailableVal = FindAvailableLoadedValue(
978 &LI, LI.getParent(), BBI, DefMaxInstsToScan, AA, &IsLoadCSE)) {
980 combineMetadataForCSE(cast<LoadInst>(AvailableVal), &LI);
982 return replaceInstUsesWith(
983 LI, Builder->CreateBitOrPointerCast(AvailableVal, LI.getType(),
984 LI.getName() + ".cast"));
987 // None of the following transforms are legal for volatile/ordered atomic
988 // loads. Most of them do apply for unordered atomics.
989 if (!LI.isUnordered()) return nullptr;
991 // load(gep null, ...) -> unreachable
992 // load null/undef -> unreachable
993 // TODO: Consider a target hook for valid address spaces for this xforms.
994 if (canSimplifyNullLoadOrGEP(LI, Op)) {
995 // Insert a new store to null instruction before the load to indicate
996 // that this code is not reachable. We do this instead of inserting
997 // an unreachable instruction directly because we cannot modify the
999 new StoreInst(UndefValue::get(LI.getType()),
1000 Constant::getNullValue(Op->getType()), &LI);
1001 return replaceInstUsesWith(LI, UndefValue::get(LI.getType()));
1004 if (Op->hasOneUse()) {
1005 // Change select and PHI nodes to select values instead of addresses: this
1006 // helps alias analysis out a lot, allows many others simplifications, and
1007 // exposes redundancy in the code.
1009 // Note that we cannot do the transformation unless we know that the
1010 // introduced loads cannot trap! Something like this is valid as long as
1011 // the condition is always false: load (select bool %C, int* null, int* %G),
1012 // but it would not be valid if we transformed it to load from null
1015 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
1016 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
1017 unsigned Align = LI.getAlignment();
1018 if (isSafeToLoadUnconditionally(SI->getOperand(1), Align, DL, SI) &&
1019 isSafeToLoadUnconditionally(SI->getOperand(2), Align, DL, SI)) {
1020 LoadInst *V1 = Builder->CreateLoad(SI->getOperand(1),
1021 SI->getOperand(1)->getName()+".val");
1022 LoadInst *V2 = Builder->CreateLoad(SI->getOperand(2),
1023 SI->getOperand(2)->getName()+".val");
1024 assert(LI.isUnordered() && "implied by above");
1025 V1->setAlignment(Align);
1026 V1->setAtomic(LI.getOrdering(), LI.getSynchScope());
1027 V2->setAlignment(Align);
1028 V2->setAtomic(LI.getOrdering(), LI.getSynchScope());
1029 return SelectInst::Create(SI->getCondition(), V1, V2);
1032 // load (select (cond, null, P)) -> load P
1033 if (isa<ConstantPointerNull>(SI->getOperand(1)) &&
1034 LI.getPointerAddressSpace() == 0) {
1035 LI.setOperand(0, SI->getOperand(2));
1039 // load (select (cond, P, null)) -> load P
1040 if (isa<ConstantPointerNull>(SI->getOperand(2)) &&
1041 LI.getPointerAddressSpace() == 0) {
1042 LI.setOperand(0, SI->getOperand(1));
1050 /// \brief Look for extractelement/insertvalue sequence that acts like a bitcast.
1052 /// \returns underlying value that was "cast", or nullptr otherwise.
1054 /// For example, if we have:
1056 /// %E0 = extractelement <2 x double> %U, i32 0
1057 /// %V0 = insertvalue [2 x double] undef, double %E0, 0
1058 /// %E1 = extractelement <2 x double> %U, i32 1
1059 /// %V1 = insertvalue [2 x double] %V0, double %E1, 1
1061 /// and the layout of a <2 x double> is isomorphic to a [2 x double],
1062 /// then %V1 can be safely approximated by a conceptual "bitcast" of %U.
1063 /// Note that %U may contain non-undef values where %V1 has undef.
1064 static Value *likeBitCastFromVector(InstCombiner &IC, Value *V) {
1066 while (auto *IV = dyn_cast<InsertValueInst>(V)) {
1067 auto *E = dyn_cast<ExtractElementInst>(IV->getInsertedValueOperand());
1070 auto *W = E->getVectorOperand();
1075 auto *CI = dyn_cast<ConstantInt>(E->getIndexOperand());
1076 if (!CI || IV->getNumIndices() != 1 || CI->getZExtValue() != *IV->idx_begin())
1078 V = IV->getAggregateOperand();
1080 if (!isa<UndefValue>(V) ||!U)
1083 auto *UT = cast<VectorType>(U->getType());
1084 auto *VT = V->getType();
1085 // Check that types UT and VT are bitwise isomorphic.
1086 const auto &DL = IC.getDataLayout();
1087 if (DL.getTypeStoreSizeInBits(UT) != DL.getTypeStoreSizeInBits(VT)) {
1090 if (auto *AT = dyn_cast<ArrayType>(VT)) {
1091 if (AT->getNumElements() != UT->getNumElements())
1094 auto *ST = cast<StructType>(VT);
1095 if (ST->getNumElements() != UT->getNumElements())
1097 for (const auto *EltT : ST->elements()) {
1098 if (EltT != UT->getElementType())
1105 /// \brief Combine stores to match the type of value being stored.
1107 /// The core idea here is that the memory does not have any intrinsic type and
1108 /// where we can we should match the type of a store to the type of value being
1111 /// However, this routine must never change the width of a store or the number of
1112 /// stores as that would introduce a semantic change. This combine is expected to
1113 /// be a semantic no-op which just allows stores to more closely model the types
1114 /// of their incoming values.
1116 /// Currently, we also refuse to change the precise type used for an atomic or
1117 /// volatile store. This is debatable, and might be reasonable to change later.
1118 /// However, it is risky in case some backend or other part of LLVM is relying
1119 /// on the exact type stored to select appropriate atomic operations.
1121 /// \returns true if the store was successfully combined away. This indicates
1122 /// the caller must erase the store instruction. We have to let the caller erase
1123 /// the store instruction as otherwise there is no way to signal whether it was
1124 /// combined or not: IC.EraseInstFromFunction returns a null pointer.
1125 static bool combineStoreToValueType(InstCombiner &IC, StoreInst &SI) {
1126 // FIXME: We could probably with some care handle both volatile and ordered
1127 // atomic stores here but it isn't clear that this is important.
1128 if (!SI.isUnordered())
1131 // swifterror values can't be bitcasted.
1132 if (SI.getPointerOperand()->isSwiftError())
1135 Value *V = SI.getValueOperand();
1137 // Fold away bit casts of the stored value by storing the original type.
1138 if (auto *BC = dyn_cast<BitCastInst>(V)) {
1139 V = BC->getOperand(0);
1140 if (!SI.isAtomic() || isSupportedAtomicType(V->getType())) {
1141 combineStoreToNewValue(IC, SI, V);
1146 if (Value *U = likeBitCastFromVector(IC, V))
1147 if (!SI.isAtomic() || isSupportedAtomicType(U->getType())) {
1148 combineStoreToNewValue(IC, SI, U);
1152 // FIXME: We should also canonicalize stores of vectors when their elements
1153 // are cast to other types.
1157 static bool unpackStoreToAggregate(InstCombiner &IC, StoreInst &SI) {
1158 // FIXME: We could probably with some care handle both volatile and atomic
1159 // stores here but it isn't clear that this is important.
1163 Value *V = SI.getValueOperand();
1164 Type *T = V->getType();
1166 if (!T->isAggregateType())
1169 if (auto *ST = dyn_cast<StructType>(T)) {
1170 // If the struct only have one element, we unpack.
1171 unsigned Count = ST->getNumElements();
1173 V = IC.Builder->CreateExtractValue(V, 0);
1174 combineStoreToNewValue(IC, SI, V);
1178 // We don't want to break loads with padding here as we'd loose
1179 // the knowledge that padding exists for the rest of the pipeline.
1180 const DataLayout &DL = IC.getDataLayout();
1181 auto *SL = DL.getStructLayout(ST);
1182 if (SL->hasPadding())
1185 auto Align = SI.getAlignment();
1187 Align = DL.getABITypeAlignment(ST);
1189 SmallString<16> EltName = V->getName();
1191 auto *Addr = SI.getPointerOperand();
1192 SmallString<16> AddrName = Addr->getName();
1193 AddrName += ".repack";
1195 auto *IdxType = Type::getInt32Ty(ST->getContext());
1196 auto *Zero = ConstantInt::get(IdxType, 0);
1197 for (unsigned i = 0; i < Count; i++) {
1198 Value *Indices[2] = {
1200 ConstantInt::get(IdxType, i),
1202 auto *Ptr = IC.Builder->CreateInBoundsGEP(ST, Addr, makeArrayRef(Indices),
1204 auto *Val = IC.Builder->CreateExtractValue(V, i, EltName);
1205 auto EltAlign = MinAlign(Align, SL->getElementOffset(i));
1206 IC.Builder->CreateAlignedStore(Val, Ptr, EltAlign);
1212 if (auto *AT = dyn_cast<ArrayType>(T)) {
1213 // If the array only have one element, we unpack.
1214 auto NumElements = AT->getNumElements();
1215 if (NumElements == 1) {
1216 V = IC.Builder->CreateExtractValue(V, 0);
1217 combineStoreToNewValue(IC, SI, V);
1221 // Bail out if the array is too large. Ideally we would like to optimize
1222 // arrays of arbitrary size but this has a terrible impact on compile time.
1223 // The threshold here is chosen arbitrarily, maybe needs a little bit of
1225 if (NumElements > IC.MaxArraySizeForCombine)
1228 const DataLayout &DL = IC.getDataLayout();
1229 auto EltSize = DL.getTypeAllocSize(AT->getElementType());
1230 auto Align = SI.getAlignment();
1232 Align = DL.getABITypeAlignment(T);
1234 SmallString<16> EltName = V->getName();
1236 auto *Addr = SI.getPointerOperand();
1237 SmallString<16> AddrName = Addr->getName();
1238 AddrName += ".repack";
1240 auto *IdxType = Type::getInt64Ty(T->getContext());
1241 auto *Zero = ConstantInt::get(IdxType, 0);
1243 uint64_t Offset = 0;
1244 for (uint64_t i = 0; i < NumElements; i++) {
1245 Value *Indices[2] = {
1247 ConstantInt::get(IdxType, i),
1249 auto *Ptr = IC.Builder->CreateInBoundsGEP(AT, Addr, makeArrayRef(Indices),
1251 auto *Val = IC.Builder->CreateExtractValue(V, i, EltName);
1252 auto EltAlign = MinAlign(Align, Offset);
1253 IC.Builder->CreateAlignedStore(Val, Ptr, EltAlign);
1263 /// equivalentAddressValues - Test if A and B will obviously have the same
1264 /// value. This includes recognizing that %t0 and %t1 will have the same
1265 /// value in code like this:
1266 /// %t0 = getelementptr \@a, 0, 3
1267 /// store i32 0, i32* %t0
1268 /// %t1 = getelementptr \@a, 0, 3
1269 /// %t2 = load i32* %t1
1271 static bool equivalentAddressValues(Value *A, Value *B) {
1272 // Test if the values are trivially equivalent.
1273 if (A == B) return true;
1275 // Test if the values come form identical arithmetic instructions.
1276 // This uses isIdenticalToWhenDefined instead of isIdenticalTo because
1277 // its only used to compare two uses within the same basic block, which
1278 // means that they'll always either have the same value or one of them
1279 // will have an undefined value.
1280 if (isa<BinaryOperator>(A) ||
1283 isa<GetElementPtrInst>(A))
1284 if (Instruction *BI = dyn_cast<Instruction>(B))
1285 if (cast<Instruction>(A)->isIdenticalToWhenDefined(BI))
1288 // Otherwise they may not be equivalent.
1292 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
1293 Value *Val = SI.getOperand(0);
1294 Value *Ptr = SI.getOperand(1);
1296 // Try to canonicalize the stored type.
1297 if (combineStoreToValueType(*this, SI))
1298 return eraseInstFromFunction(SI);
1300 // Attempt to improve the alignment.
1301 unsigned KnownAlign = getOrEnforceKnownAlignment(
1302 Ptr, DL.getPrefTypeAlignment(Val->getType()), DL, &SI, &AC, &DT);
1303 unsigned StoreAlign = SI.getAlignment();
1304 unsigned EffectiveStoreAlign =
1305 StoreAlign != 0 ? StoreAlign : DL.getABITypeAlignment(Val->getType());
1307 if (KnownAlign > EffectiveStoreAlign)
1308 SI.setAlignment(KnownAlign);
1309 else if (StoreAlign == 0)
1310 SI.setAlignment(EffectiveStoreAlign);
1312 // Try to canonicalize the stored type.
1313 if (unpackStoreToAggregate(*this, SI))
1314 return eraseInstFromFunction(SI);
1316 // Replace GEP indices if possible.
1317 if (Instruction *NewGEPI = replaceGEPIdxWithZero(*this, Ptr, SI)) {
1318 Worklist.Add(NewGEPI);
1322 // Don't hack volatile/ordered stores.
1323 // FIXME: Some bits are legal for ordered atomic stores; needs refactoring.
1324 if (!SI.isUnordered()) return nullptr;
1326 // If the RHS is an alloca with a single use, zapify the store, making the
1328 if (Ptr->hasOneUse()) {
1329 if (isa<AllocaInst>(Ptr))
1330 return eraseInstFromFunction(SI);
1331 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
1332 if (isa<AllocaInst>(GEP->getOperand(0))) {
1333 if (GEP->getOperand(0)->hasOneUse())
1334 return eraseInstFromFunction(SI);
1339 // Do really simple DSE, to catch cases where there are several consecutive
1340 // stores to the same location, separated by a few arithmetic operations. This
1341 // situation often occurs with bitfield accesses.
1342 BasicBlock::iterator BBI(SI);
1343 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
1346 // Don't count debug info directives, lest they affect codegen,
1347 // and we skip pointer-to-pointer bitcasts, which are NOPs.
1348 if (isa<DbgInfoIntrinsic>(BBI) ||
1349 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy())) {
1354 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
1355 // Prev store isn't volatile, and stores to the same location?
1356 if (PrevSI->isUnordered() && equivalentAddressValues(PrevSI->getOperand(1),
1357 SI.getOperand(1))) {
1360 eraseInstFromFunction(*PrevSI);
1366 // If this is a load, we have to stop. However, if the loaded value is from
1367 // the pointer we're loading and is producing the pointer we're storing,
1368 // then *this* store is dead (X = load P; store X -> P).
1369 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
1370 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr)) {
1371 assert(SI.isUnordered() && "can't eliminate ordering operation");
1372 return eraseInstFromFunction(SI);
1375 // Otherwise, this is a load from some other location. Stores before it
1380 // Don't skip over loads, throws or things that can modify memory.
1381 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory() || BBI->mayThrow())
1385 // store X, null -> turns into 'unreachable' in SimplifyCFG
1386 if (isa<ConstantPointerNull>(Ptr) && SI.getPointerAddressSpace() == 0) {
1387 if (!isa<UndefValue>(Val)) {
1388 SI.setOperand(0, UndefValue::get(Val->getType()));
1389 if (Instruction *U = dyn_cast<Instruction>(Val))
1390 Worklist.Add(U); // Dropped a use.
1392 return nullptr; // Do not modify these!
1395 // store undef, Ptr -> noop
1396 if (isa<UndefValue>(Val))
1397 return eraseInstFromFunction(SI);
1399 // If this store is the last instruction in the basic block (possibly
1400 // excepting debug info instructions), and if the block ends with an
1401 // unconditional branch, try to move it to the successor block.
1402 BBI = SI.getIterator();
1405 } while (isa<DbgInfoIntrinsic>(BBI) ||
1406 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy()));
1407 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
1408 if (BI->isUnconditional())
1409 if (SimplifyStoreAtEndOfBlock(SI))
1410 return nullptr; // xform done!
1415 /// SimplifyStoreAtEndOfBlock - Turn things like:
1416 /// if () { *P = v1; } else { *P = v2 }
1417 /// into a phi node with a store in the successor.
1419 /// Simplify things like:
1420 /// *P = v1; if () { *P = v2; }
1421 /// into a phi node with a store in the successor.
1423 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
1424 assert(SI.isUnordered() &&
1425 "this code has not been auditted for volatile or ordered store case");
1427 BasicBlock *StoreBB = SI.getParent();
1429 // Check to see if the successor block has exactly two incoming edges. If
1430 // so, see if the other predecessor contains a store to the same location.
1431 // if so, insert a PHI node (if needed) and move the stores down.
1432 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
1434 // Determine whether Dest has exactly two predecessors and, if so, compute
1435 // the other predecessor.
1436 pred_iterator PI = pred_begin(DestBB);
1437 BasicBlock *P = *PI;
1438 BasicBlock *OtherBB = nullptr;
1443 if (++PI == pred_end(DestBB))
1452 if (++PI != pred_end(DestBB))
1455 // Bail out if all the relevant blocks aren't distinct (this can happen,
1456 // for example, if SI is in an infinite loop)
1457 if (StoreBB == DestBB || OtherBB == DestBB)
1460 // Verify that the other block ends in a branch and is not otherwise empty.
1461 BasicBlock::iterator BBI(OtherBB->getTerminator());
1462 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
1463 if (!OtherBr || BBI == OtherBB->begin())
1466 // If the other block ends in an unconditional branch, check for the 'if then
1467 // else' case. there is an instruction before the branch.
1468 StoreInst *OtherStore = nullptr;
1469 if (OtherBr->isUnconditional()) {
1471 // Skip over debugging info.
1472 while (isa<DbgInfoIntrinsic>(BBI) ||
1473 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy())) {
1474 if (BBI==OtherBB->begin())
1478 // If this isn't a store, isn't a store to the same location, or is not the
1479 // right kind of store, bail out.
1480 OtherStore = dyn_cast<StoreInst>(BBI);
1481 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1) ||
1482 !SI.isSameOperationAs(OtherStore))
1485 // Otherwise, the other block ended with a conditional branch. If one of the
1486 // destinations is StoreBB, then we have the if/then case.
1487 if (OtherBr->getSuccessor(0) != StoreBB &&
1488 OtherBr->getSuccessor(1) != StoreBB)
1491 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
1492 // if/then triangle. See if there is a store to the same ptr as SI that
1493 // lives in OtherBB.
1495 // Check to see if we find the matching store.
1496 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
1497 if (OtherStore->getOperand(1) != SI.getOperand(1) ||
1498 !SI.isSameOperationAs(OtherStore))
1502 // If we find something that may be using or overwriting the stored
1503 // value, or if we run out of instructions, we can't do the xform.
1504 if (BBI->mayReadFromMemory() || BBI->mayThrow() ||
1505 BBI->mayWriteToMemory() || BBI == OtherBB->begin())
1509 // In order to eliminate the store in OtherBr, we have to
1510 // make sure nothing reads or overwrites the stored value in
1512 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
1513 // FIXME: This should really be AA driven.
1514 if (I->mayReadFromMemory() || I->mayThrow() || I->mayWriteToMemory())
1519 // Insert a PHI node now if we need it.
1520 Value *MergedVal = OtherStore->getOperand(0);
1521 if (MergedVal != SI.getOperand(0)) {
1522 PHINode *PN = PHINode::Create(MergedVal->getType(), 2, "storemerge");
1523 PN->addIncoming(SI.getOperand(0), SI.getParent());
1524 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
1525 MergedVal = InsertNewInstBefore(PN, DestBB->front());
1528 // Advance to a place where it is safe to insert the new store and
1530 BBI = DestBB->getFirstInsertionPt();
1531 StoreInst *NewSI = new StoreInst(MergedVal, SI.getOperand(1),
1535 SI.getSynchScope());
1536 InsertNewInstBefore(NewSI, *BBI);
1537 // The debug locations of the original instructions might differ; merge them.
1538 NewSI->setDebugLoc(DILocation::getMergedLocation(SI.getDebugLoc(),
1539 OtherStore->getDebugLoc()));
1541 // If the two stores had AA tags, merge them.
1543 SI.getAAMetadata(AATags);
1545 OtherStore->getAAMetadata(AATags, /* Merge = */ true);
1546 NewSI->setAAMetadata(AATags);
1549 // Nuke the old stores.
1550 eraseInstFromFunction(SI);
1551 eraseInstFromFunction(*OtherStore);