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/SmallString.h"
16 #include "llvm/ADT/Statistic.h"
17 #include "llvm/Analysis/Loads.h"
18 #include "llvm/IR/ConstantRange.h"
19 #include "llvm/IR/DataLayout.h"
20 #include "llvm/IR/LLVMContext.h"
21 #include "llvm/IR/IntrinsicInst.h"
22 #include "llvm/IR/MDBuilder.h"
23 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
24 #include "llvm/Transforms/Utils/Local.h"
27 #define DEBUG_TYPE "instcombine"
29 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
30 STATISTIC(NumGlobalCopies, "Number of allocas copied from constant global");
32 /// pointsToConstantGlobal - Return true if V (possibly indirectly) points to
33 /// some part of a constant global variable. This intentionally only accepts
34 /// constant expressions because we can't rewrite arbitrary instructions.
35 static bool pointsToConstantGlobal(Value *V) {
36 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
37 return GV->isConstant();
39 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
40 if (CE->getOpcode() == Instruction::BitCast ||
41 CE->getOpcode() == Instruction::AddrSpaceCast ||
42 CE->getOpcode() == Instruction::GetElementPtr)
43 return pointsToConstantGlobal(CE->getOperand(0));
48 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
49 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
50 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
51 /// track of whether it moves the pointer (with IsOffset) but otherwise traverse
52 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
53 /// the alloca, and if the source pointer is a pointer to a constant global, we
54 /// can optimize this.
56 isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy,
57 SmallVectorImpl<Instruction *> &ToDelete) {
58 // We track lifetime intrinsics as we encounter them. If we decide to go
59 // ahead and replace the value with the global, this lets the caller quickly
60 // eliminate the markers.
62 SmallVector<std::pair<Value *, bool>, 35> ValuesToInspect;
63 ValuesToInspect.emplace_back(V, false);
64 while (!ValuesToInspect.empty()) {
65 auto ValuePair = ValuesToInspect.pop_back_val();
66 const bool IsOffset = ValuePair.second;
67 for (auto &U : ValuePair.first->uses()) {
68 auto *I = cast<Instruction>(U.getUser());
70 if (auto *LI = dyn_cast<LoadInst>(I)) {
71 // Ignore non-volatile loads, they are always ok.
72 if (!LI->isSimple()) return false;
76 if (isa<BitCastInst>(I) || isa<AddrSpaceCastInst>(I)) {
77 // If uses of the bitcast are ok, we are ok.
78 ValuesToInspect.emplace_back(I, IsOffset);
81 if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) {
82 // If the GEP has all zero indices, it doesn't offset the pointer. If it
84 ValuesToInspect.emplace_back(I, IsOffset || !GEP->hasAllZeroIndices());
88 if (auto CS = CallSite(I)) {
89 // If this is the function being called then we treat it like a load and
94 unsigned DataOpNo = CS.getDataOperandNo(&U);
95 bool IsArgOperand = CS.isArgOperand(&U);
97 // Inalloca arguments are clobbered by the call.
98 if (IsArgOperand && CS.isInAllocaArgument(DataOpNo))
101 // If this is a readonly/readnone call site, then we know it is just a
102 // load (but one that potentially returns the value itself), so we can
103 // ignore it if we know that the value isn't captured.
104 if (CS.onlyReadsMemory() &&
105 (CS.getInstruction()->use_empty() || CS.doesNotCapture(DataOpNo)))
108 // If this is being passed as a byval argument, the caller is making a
109 // copy, so it is only a read of the alloca.
110 if (IsArgOperand && CS.isByValArgument(DataOpNo))
114 // Lifetime intrinsics can be handled by the caller.
115 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
116 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
117 II->getIntrinsicID() == Intrinsic::lifetime_end) {
118 assert(II->use_empty() && "Lifetime markers have no result to use!");
119 ToDelete.push_back(II);
124 // If this is isn't our memcpy/memmove, reject it as something we can't
126 MemTransferInst *MI = dyn_cast<MemTransferInst>(I);
130 // If the transfer is using the alloca as a source of the transfer, then
131 // ignore it since it is a load (unless the transfer is volatile).
132 if (U.getOperandNo() == 1) {
133 if (MI->isVolatile()) return false;
137 // If we already have seen a copy, reject the second one.
138 if (TheCopy) return false;
140 // If the pointer has been offset from the start of the alloca, we can't
141 // safely handle this.
142 if (IsOffset) return false;
144 // If the memintrinsic isn't using the alloca as the dest, reject it.
145 if (U.getOperandNo() != 0) return false;
147 // If the source of the memcpy/move is not a constant global, reject it.
148 if (!pointsToConstantGlobal(MI->getSource()))
151 // Otherwise, the transform is safe. Remember the copy instruction.
158 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
159 /// modified by a copy from a constant global. If we can prove this, we can
160 /// replace any uses of the alloca with uses of the global directly.
161 static MemTransferInst *
162 isOnlyCopiedFromConstantGlobal(AllocaInst *AI,
163 SmallVectorImpl<Instruction *> &ToDelete) {
164 MemTransferInst *TheCopy = nullptr;
165 if (isOnlyCopiedFromConstantGlobal(AI, TheCopy, ToDelete))
170 static Instruction *simplifyAllocaArraySize(InstCombiner &IC, AllocaInst &AI) {
171 // Check for array size of 1 (scalar allocation).
172 if (!AI.isArrayAllocation()) {
173 // i32 1 is the canonical array size for scalar allocations.
174 if (AI.getArraySize()->getType()->isIntegerTy(32))
178 Value *V = IC.Builder->getInt32(1);
183 // Convert: alloca Ty, C - where C is a constant != 1 into: alloca [C x Ty], 1
184 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
185 Type *NewTy = ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
186 AllocaInst *New = IC.Builder->CreateAlloca(NewTy, nullptr, AI.getName());
187 New->setAlignment(AI.getAlignment());
189 // Scan to the end of the allocation instructions, to skip over a block of
190 // allocas if possible...also skip interleaved debug info
192 BasicBlock::iterator It(New);
193 while (isa<AllocaInst>(*It) || isa<DbgInfoIntrinsic>(*It))
196 // Now that I is pointing to the first non-allocation-inst in the block,
197 // insert our getelementptr instruction...
199 Type *IdxTy = IC.getDataLayout().getIntPtrType(AI.getType());
200 Value *NullIdx = Constant::getNullValue(IdxTy);
201 Value *Idx[2] = {NullIdx, NullIdx};
203 GetElementPtrInst::CreateInBounds(New, Idx, New->getName() + ".sub");
204 IC.InsertNewInstBefore(GEP, *It);
206 // Now make everything use the getelementptr instead of the original
208 return IC.replaceInstUsesWith(AI, GEP);
211 if (isa<UndefValue>(AI.getArraySize()))
212 return IC.replaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
214 // Ensure that the alloca array size argument has type intptr_t, so that
215 // any casting is exposed early.
216 Type *IntPtrTy = IC.getDataLayout().getIntPtrType(AI.getType());
217 if (AI.getArraySize()->getType() != IntPtrTy) {
218 Value *V = IC.Builder->CreateIntCast(AI.getArraySize(), IntPtrTy, false);
226 Instruction *InstCombiner::visitAllocaInst(AllocaInst &AI) {
227 if (auto *I = simplifyAllocaArraySize(*this, AI))
230 if (AI.getAllocatedType()->isSized()) {
231 // If the alignment is 0 (unspecified), assign it the preferred alignment.
232 if (AI.getAlignment() == 0)
233 AI.setAlignment(DL.getPrefTypeAlignment(AI.getAllocatedType()));
235 // Move all alloca's of zero byte objects to the entry block and merge them
236 // together. Note that we only do this for alloca's, because malloc should
237 // allocate and return a unique pointer, even for a zero byte allocation.
238 if (DL.getTypeAllocSize(AI.getAllocatedType()) == 0) {
239 // For a zero sized alloca there is no point in doing an array allocation.
240 // This is helpful if the array size is a complicated expression not used
242 if (AI.isArrayAllocation()) {
243 AI.setOperand(0, ConstantInt::get(AI.getArraySize()->getType(), 1));
247 // Get the first instruction in the entry block.
248 BasicBlock &EntryBlock = AI.getParent()->getParent()->getEntryBlock();
249 Instruction *FirstInst = EntryBlock.getFirstNonPHIOrDbg();
250 if (FirstInst != &AI) {
251 // If the entry block doesn't start with a zero-size alloca then move
252 // this one to the start of the entry block. There is no problem with
253 // dominance as the array size was forced to a constant earlier already.
254 AllocaInst *EntryAI = dyn_cast<AllocaInst>(FirstInst);
255 if (!EntryAI || !EntryAI->getAllocatedType()->isSized() ||
256 DL.getTypeAllocSize(EntryAI->getAllocatedType()) != 0) {
257 AI.moveBefore(FirstInst);
261 // If the alignment of the entry block alloca is 0 (unspecified),
262 // assign it the preferred alignment.
263 if (EntryAI->getAlignment() == 0)
264 EntryAI->setAlignment(
265 DL.getPrefTypeAlignment(EntryAI->getAllocatedType()));
266 // Replace this zero-sized alloca with the one at the start of the entry
267 // block after ensuring that the address will be aligned enough for both
269 unsigned MaxAlign = std::max(EntryAI->getAlignment(),
271 EntryAI->setAlignment(MaxAlign);
272 if (AI.getType() != EntryAI->getType())
273 return new BitCastInst(EntryAI, AI.getType());
274 return replaceInstUsesWith(AI, EntryAI);
279 if (AI.getAlignment()) {
280 // Check to see if this allocation is only modified by a memcpy/memmove from
281 // a constant global whose alignment is equal to or exceeds that of the
282 // allocation. If this is the case, we can change all users to use
283 // the constant global instead. This is commonly produced by the CFE by
284 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
285 // is only subsequently read.
286 SmallVector<Instruction *, 4> ToDelete;
287 if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(&AI, ToDelete)) {
288 unsigned SourceAlign = getOrEnforceKnownAlignment(
289 Copy->getSource(), AI.getAlignment(), DL, &AI, &AC, &DT);
290 if (AI.getAlignment() <= SourceAlign) {
291 DEBUG(dbgs() << "Found alloca equal to global: " << AI << '\n');
292 DEBUG(dbgs() << " memcpy = " << *Copy << '\n');
293 for (unsigned i = 0, e = ToDelete.size(); i != e; ++i)
294 eraseInstFromFunction(*ToDelete[i]);
295 Constant *TheSrc = cast<Constant>(Copy->getSource());
297 = ConstantExpr::getPointerBitCastOrAddrSpaceCast(TheSrc, AI.getType());
298 Instruction *NewI = replaceInstUsesWith(AI, Cast);
299 eraseInstFromFunction(*Copy);
306 // At last, use the generic allocation site handler to aggressively remove
308 return visitAllocSite(AI);
311 // Are we allowed to form a atomic load or store of this type?
312 static bool isSupportedAtomicType(Type *Ty) {
313 return Ty->isIntegerTy() || Ty->isPointerTy() || Ty->isFloatingPointTy();
316 /// \brief Helper to combine a load to a new type.
318 /// This just does the work of combining a load to a new type. It handles
319 /// metadata, etc., and returns the new instruction. The \c NewTy should be the
320 /// loaded *value* type. This will convert it to a pointer, cast the operand to
321 /// that pointer type, load it, etc.
323 /// Note that this will create all of the instructions with whatever insert
324 /// point the \c InstCombiner currently is using.
325 static LoadInst *combineLoadToNewType(InstCombiner &IC, LoadInst &LI, Type *NewTy,
326 const Twine &Suffix = "") {
327 assert((!LI.isAtomic() || isSupportedAtomicType(NewTy)) &&
328 "can't fold an atomic load to requested type");
330 Value *Ptr = LI.getPointerOperand();
331 unsigned AS = LI.getPointerAddressSpace();
332 SmallVector<std::pair<unsigned, MDNode *>, 8> MD;
333 LI.getAllMetadata(MD);
335 LoadInst *NewLoad = IC.Builder->CreateAlignedLoad(
336 IC.Builder->CreateBitCast(Ptr, NewTy->getPointerTo(AS)),
337 LI.getAlignment(), LI.isVolatile(), LI.getName() + Suffix);
338 NewLoad->setAtomic(LI.getOrdering(), LI.getSynchScope());
339 MDBuilder MDB(NewLoad->getContext());
340 for (const auto &MDPair : MD) {
341 unsigned ID = MDPair.first;
342 MDNode *N = MDPair.second;
343 // Note, essentially every kind of metadata should be preserved here! This
344 // routine is supposed to clone a load instruction changing *only its type*.
345 // The only metadata it makes sense to drop is metadata which is invalidated
346 // when the pointer type changes. This should essentially never be the case
347 // in LLVM, but we explicitly switch over only known metadata to be
348 // conservatively correct. If you are adding metadata to LLVM which pertains
349 // to loads, you almost certainly want to add it here.
351 case LLVMContext::MD_dbg:
352 case LLVMContext::MD_tbaa:
353 case LLVMContext::MD_prof:
354 case LLVMContext::MD_fpmath:
355 case LLVMContext::MD_tbaa_struct:
356 case LLVMContext::MD_invariant_load:
357 case LLVMContext::MD_alias_scope:
358 case LLVMContext::MD_noalias:
359 case LLVMContext::MD_nontemporal:
360 case LLVMContext::MD_mem_parallel_loop_access:
361 // All of these directly apply.
362 NewLoad->setMetadata(ID, N);
365 case LLVMContext::MD_nonnull:
366 // This only directly applies if the new type is also a pointer.
367 if (NewTy->isPointerTy()) {
368 NewLoad->setMetadata(ID, N);
371 // If it's integral now, translate it to !range metadata.
372 if (NewTy->isIntegerTy()) {
373 auto *ITy = cast<IntegerType>(NewTy);
374 auto *NullInt = ConstantExpr::getPtrToInt(
375 ConstantPointerNull::get(cast<PointerType>(Ptr->getType())), ITy);
377 ConstantExpr::getAdd(NullInt, ConstantInt::get(ITy, 1));
378 NewLoad->setMetadata(LLVMContext::MD_range,
379 MDB.createRange(NonNullInt, NullInt));
382 case LLVMContext::MD_align:
383 case LLVMContext::MD_dereferenceable:
384 case LLVMContext::MD_dereferenceable_or_null:
385 // These only directly apply if the new type is also a pointer.
386 if (NewTy->isPointerTy())
387 NewLoad->setMetadata(ID, N);
389 case LLVMContext::MD_range:
390 // FIXME: It would be nice to propagate this in some way, but the type
391 // conversions make it hard.
393 // If it's a pointer now and the range does not contain 0, make it !nonnull.
394 if (NewTy->isPointerTy()) {
395 unsigned BitWidth = IC.getDataLayout().getTypeSizeInBits(NewTy);
396 if (!getConstantRangeFromMetadata(*N).contains(APInt(BitWidth, 0))) {
397 MDNode *NN = MDNode::get(LI.getContext(), None);
398 NewLoad->setMetadata(LLVMContext::MD_nonnull, NN);
407 /// \brief Combine a store to a new type.
409 /// Returns the newly created store instruction.
410 static StoreInst *combineStoreToNewValue(InstCombiner &IC, StoreInst &SI, Value *V) {
411 assert((!SI.isAtomic() || isSupportedAtomicType(V->getType())) &&
412 "can't fold an atomic store of requested type");
414 Value *Ptr = SI.getPointerOperand();
415 unsigned AS = SI.getPointerAddressSpace();
416 SmallVector<std::pair<unsigned, MDNode *>, 8> MD;
417 SI.getAllMetadata(MD);
419 StoreInst *NewStore = IC.Builder->CreateAlignedStore(
420 V, IC.Builder->CreateBitCast(Ptr, V->getType()->getPointerTo(AS)),
421 SI.getAlignment(), SI.isVolatile());
422 NewStore->setAtomic(SI.getOrdering(), SI.getSynchScope());
423 for (const auto &MDPair : MD) {
424 unsigned ID = MDPair.first;
425 MDNode *N = MDPair.second;
426 // Note, essentially every kind of metadata should be preserved here! This
427 // routine is supposed to clone a store instruction changing *only its
428 // type*. The only metadata it makes sense to drop is metadata which is
429 // invalidated when the pointer type changes. This should essentially
430 // never be the case in LLVM, but we explicitly switch over only known
431 // metadata to be conservatively correct. If you are adding metadata to
432 // LLVM which pertains to stores, you almost certainly want to add it
435 case LLVMContext::MD_dbg:
436 case LLVMContext::MD_tbaa:
437 case LLVMContext::MD_prof:
438 case LLVMContext::MD_fpmath:
439 case LLVMContext::MD_tbaa_struct:
440 case LLVMContext::MD_alias_scope:
441 case LLVMContext::MD_noalias:
442 case LLVMContext::MD_nontemporal:
443 case LLVMContext::MD_mem_parallel_loop_access:
444 // All of these directly apply.
445 NewStore->setMetadata(ID, N);
448 case LLVMContext::MD_invariant_load:
449 case LLVMContext::MD_nonnull:
450 case LLVMContext::MD_range:
451 case LLVMContext::MD_align:
452 case LLVMContext::MD_dereferenceable:
453 case LLVMContext::MD_dereferenceable_or_null:
454 // These don't apply for stores.
462 /// \brief Combine loads to match the type of their uses' value after looking
463 /// through intervening bitcasts.
465 /// The core idea here is that if the result of a load is used in an operation,
466 /// we should load the type most conducive to that operation. For example, when
467 /// loading an integer and converting that immediately to a pointer, we should
468 /// instead directly load a pointer.
470 /// However, this routine must never change the width of a load or the number of
471 /// loads as that would introduce a semantic change. This combine is expected to
472 /// be a semantic no-op which just allows loads to more closely model the types
473 /// of their consuming operations.
475 /// Currently, we also refuse to change the precise type used for an atomic load
476 /// or a volatile load. This is debatable, and might be reasonable to change
477 /// later. However, it is risky in case some backend or other part of LLVM is
478 /// relying on the exact type loaded to select appropriate atomic operations.
479 static Instruction *combineLoadToOperationType(InstCombiner &IC, LoadInst &LI) {
480 // FIXME: We could probably with some care handle both volatile and ordered
481 // atomic loads here but it isn't clear that this is important.
482 if (!LI.isUnordered())
488 // swifterror values can't be bitcasted.
489 if (LI.getPointerOperand()->isSwiftError())
492 Type *Ty = LI.getType();
493 const DataLayout &DL = IC.getDataLayout();
495 // Try to canonicalize loads which are only ever stored to operate over
496 // integers instead of any other type. We only do this when the loaded type
497 // is sized and has a size exactly the same as its store size and the store
498 // size is a legal integer type.
499 if (!Ty->isIntegerTy() && Ty->isSized() &&
500 DL.isLegalInteger(DL.getTypeStoreSizeInBits(Ty)) &&
501 DL.getTypeStoreSizeInBits(Ty) == DL.getTypeSizeInBits(Ty) &&
502 !DL.isNonIntegralPointerType(Ty)) {
503 if (all_of(LI.users(), [&LI](User *U) {
504 auto *SI = dyn_cast<StoreInst>(U);
505 return SI && SI->getPointerOperand() != &LI &&
506 !SI->getPointerOperand()->isSwiftError();
508 LoadInst *NewLoad = combineLoadToNewType(
510 Type::getIntNTy(LI.getContext(), DL.getTypeStoreSizeInBits(Ty)));
511 // Replace all the stores with stores of the newly loaded value.
512 for (auto UI = LI.user_begin(), UE = LI.user_end(); UI != UE;) {
513 auto *SI = cast<StoreInst>(*UI++);
514 IC.Builder->SetInsertPoint(SI);
515 combineStoreToNewValue(IC, *SI, NewLoad);
516 IC.eraseInstFromFunction(*SI);
518 assert(LI.use_empty() && "Failed to remove all users of the load!");
519 // Return the old load so the combiner can delete it safely.
524 // Fold away bit casts of the loaded value by loading the desired type.
525 // We can do this for BitCastInsts as well as casts from and to pointer types,
526 // as long as those are noops (i.e., the source or dest type have the same
527 // bitwidth as the target's pointers).
529 if (auto* CI = dyn_cast<CastInst>(LI.user_back()))
530 if (CI->isNoopCast(DL))
531 if (!LI.isAtomic() || isSupportedAtomicType(CI->getDestTy())) {
532 LoadInst *NewLoad = combineLoadToNewType(IC, LI, CI->getDestTy());
533 CI->replaceAllUsesWith(NewLoad);
534 IC.eraseInstFromFunction(*CI);
538 // FIXME: We should also canonicalize loads of vectors when their elements are
539 // cast to other types.
543 static Instruction *unpackLoadToAggregate(InstCombiner &IC, LoadInst &LI) {
544 // FIXME: We could probably with some care handle both volatile and atomic
545 // stores here but it isn't clear that this is important.
549 Type *T = LI.getType();
550 if (!T->isAggregateType())
553 StringRef Name = LI.getName();
554 assert(LI.getAlignment() && "Alignment must be set at this point");
556 if (auto *ST = dyn_cast<StructType>(T)) {
557 // If the struct only have one element, we unpack.
558 auto NumElements = ST->getNumElements();
559 if (NumElements == 1) {
560 LoadInst *NewLoad = combineLoadToNewType(IC, LI, ST->getTypeAtIndex(0U),
562 return IC.replaceInstUsesWith(LI, IC.Builder->CreateInsertValue(
563 UndefValue::get(T), NewLoad, 0, Name));
566 // We don't want to break loads with padding here as we'd loose
567 // the knowledge that padding exists for the rest of the pipeline.
568 const DataLayout &DL = IC.getDataLayout();
569 auto *SL = DL.getStructLayout(ST);
570 if (SL->hasPadding())
573 auto Align = LI.getAlignment();
575 Align = DL.getABITypeAlignment(ST);
577 auto *Addr = LI.getPointerOperand();
578 auto *IdxType = Type::getInt32Ty(T->getContext());
579 auto *Zero = ConstantInt::get(IdxType, 0);
581 Value *V = UndefValue::get(T);
582 for (unsigned i = 0; i < NumElements; i++) {
583 Value *Indices[2] = {
585 ConstantInt::get(IdxType, i),
587 auto *Ptr = IC.Builder->CreateInBoundsGEP(ST, Addr, makeArrayRef(Indices),
589 auto EltAlign = MinAlign(Align, SL->getElementOffset(i));
590 auto *L = IC.Builder->CreateAlignedLoad(Ptr, EltAlign, Name + ".unpack");
591 V = IC.Builder->CreateInsertValue(V, L, i);
595 return IC.replaceInstUsesWith(LI, V);
598 if (auto *AT = dyn_cast<ArrayType>(T)) {
599 auto *ET = AT->getElementType();
600 auto NumElements = AT->getNumElements();
601 if (NumElements == 1) {
602 LoadInst *NewLoad = combineLoadToNewType(IC, LI, ET, ".unpack");
603 return IC.replaceInstUsesWith(LI, IC.Builder->CreateInsertValue(
604 UndefValue::get(T), NewLoad, 0, Name));
607 // Bail out if the array is too large. Ideally we would like to optimize
608 // arrays of arbitrary size but this has a terrible impact on compile time.
609 // The threshold here is chosen arbitrarily, maybe needs a little bit of
611 if (NumElements > 1024)
614 const DataLayout &DL = IC.getDataLayout();
615 auto EltSize = DL.getTypeAllocSize(ET);
616 auto Align = LI.getAlignment();
618 Align = DL.getABITypeAlignment(T);
620 auto *Addr = LI.getPointerOperand();
621 auto *IdxType = Type::getInt64Ty(T->getContext());
622 auto *Zero = ConstantInt::get(IdxType, 0);
624 Value *V = UndefValue::get(T);
626 for (uint64_t i = 0; i < NumElements; i++) {
627 Value *Indices[2] = {
629 ConstantInt::get(IdxType, i),
631 auto *Ptr = IC.Builder->CreateInBoundsGEP(AT, Addr, makeArrayRef(Indices),
633 auto *L = IC.Builder->CreateAlignedLoad(Ptr, MinAlign(Align, Offset),
635 V = IC.Builder->CreateInsertValue(V, L, i);
640 return IC.replaceInstUsesWith(LI, V);
646 // If we can determine that all possible objects pointed to by the provided
647 // pointer value are, not only dereferenceable, but also definitively less than
648 // or equal to the provided maximum size, then return true. Otherwise, return
649 // false (constant global values and allocas fall into this category).
651 // FIXME: This should probably live in ValueTracking (or similar).
652 static bool isObjectSizeLessThanOrEq(Value *V, uint64_t MaxSize,
653 const DataLayout &DL) {
654 SmallPtrSet<Value *, 4> Visited;
655 SmallVector<Value *, 4> Worklist(1, V);
658 Value *P = Worklist.pop_back_val();
659 P = P->stripPointerCasts();
661 if (!Visited.insert(P).second)
664 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
665 Worklist.push_back(SI->getTrueValue());
666 Worklist.push_back(SI->getFalseValue());
670 if (PHINode *PN = dyn_cast<PHINode>(P)) {
671 for (Value *IncValue : PN->incoming_values())
672 Worklist.push_back(IncValue);
676 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(P)) {
677 if (GA->isInterposable())
679 Worklist.push_back(GA->getAliasee());
683 // If we know how big this object is, and it is less than MaxSize, continue
684 // searching. Otherwise, return false.
685 if (AllocaInst *AI = dyn_cast<AllocaInst>(P)) {
686 if (!AI->getAllocatedType()->isSized())
689 ConstantInt *CS = dyn_cast<ConstantInt>(AI->getArraySize());
693 uint64_t TypeSize = DL.getTypeAllocSize(AI->getAllocatedType());
694 // Make sure that, even if the multiplication below would wrap as an
695 // uint64_t, we still do the right thing.
696 if ((CS->getValue().zextOrSelf(128)*APInt(128, TypeSize)).ugt(MaxSize))
701 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
702 if (!GV->hasDefinitiveInitializer() || !GV->isConstant())
705 uint64_t InitSize = DL.getTypeAllocSize(GV->getValueType());
706 if (InitSize > MaxSize)
712 } while (!Worklist.empty());
717 // If we're indexing into an object of a known size, and the outer index is
718 // not a constant, but having any value but zero would lead to undefined
719 // behavior, replace it with zero.
721 // For example, if we have:
722 // @f.a = private unnamed_addr constant [1 x i32] [i32 12], align 4
724 // %arrayidx = getelementptr inbounds [1 x i32]* @f.a, i64 0, i64 %x
725 // ... = load i32* %arrayidx, align 4
726 // Then we know that we can replace %x in the GEP with i64 0.
728 // FIXME: We could fold any GEP index to zero that would cause UB if it were
729 // not zero. Currently, we only handle the first such index. Also, we could
730 // also search through non-zero constant indices if we kept track of the
731 // offsets those indices implied.
732 static bool canReplaceGEPIdxWithZero(InstCombiner &IC, GetElementPtrInst *GEPI,
733 Instruction *MemI, unsigned &Idx) {
734 if (GEPI->getNumOperands() < 2)
737 // Find the first non-zero index of a GEP. If all indices are zero, return
738 // one past the last index.
739 auto FirstNZIdx = [](const GetElementPtrInst *GEPI) {
741 for (unsigned IE = GEPI->getNumOperands(); I != IE; ++I) {
742 Value *V = GEPI->getOperand(I);
743 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V))
753 // Skip through initial 'zero' indices, and find the corresponding pointer
754 // type. See if the next index is not a constant.
755 Idx = FirstNZIdx(GEPI);
756 if (Idx == GEPI->getNumOperands())
758 if (isa<Constant>(GEPI->getOperand(Idx)))
761 SmallVector<Value *, 4> Ops(GEPI->idx_begin(), GEPI->idx_begin() + Idx);
763 GetElementPtrInst::getIndexedType(GEPI->getSourceElementType(), Ops);
764 if (!AllocTy || !AllocTy->isSized())
766 const DataLayout &DL = IC.getDataLayout();
767 uint64_t TyAllocSize = DL.getTypeAllocSize(AllocTy);
769 // If there are more indices after the one we might replace with a zero, make
770 // sure they're all non-negative. If any of them are negative, the overall
771 // address being computed might be before the base address determined by the
772 // first non-zero index.
773 auto IsAllNonNegative = [&]() {
774 for (unsigned i = Idx+1, e = GEPI->getNumOperands(); i != e; ++i) {
775 bool KnownNonNegative, KnownNegative;
776 IC.ComputeSignBit(GEPI->getOperand(i), KnownNonNegative,
777 KnownNegative, 0, MemI);
778 if (KnownNonNegative)
786 // FIXME: If the GEP is not inbounds, and there are extra indices after the
787 // one we'll replace, those could cause the address computation to wrap
788 // (rendering the IsAllNonNegative() check below insufficient). We can do
789 // better, ignoring zero indices (and other indices we can prove small
790 // enough not to wrap).
791 if (Idx+1 != GEPI->getNumOperands() && !GEPI->isInBounds())
794 // Note that isObjectSizeLessThanOrEq will return true only if the pointer is
795 // also known to be dereferenceable.
796 return isObjectSizeLessThanOrEq(GEPI->getOperand(0), TyAllocSize, DL) &&
800 // If we're indexing into an object with a variable index for the memory
801 // access, but the object has only one element, we can assume that the index
802 // will always be zero. If we replace the GEP, return it.
803 template <typename T>
804 static Instruction *replaceGEPIdxWithZero(InstCombiner &IC, Value *Ptr,
806 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Ptr)) {
808 if (canReplaceGEPIdxWithZero(IC, GEPI, &MemI, Idx)) {
809 Instruction *NewGEPI = GEPI->clone();
810 NewGEPI->setOperand(Idx,
811 ConstantInt::get(GEPI->getOperand(Idx)->getType(), 0));
812 NewGEPI->insertBefore(GEPI);
813 MemI.setOperand(MemI.getPointerOperandIndex(), NewGEPI);
821 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
822 Value *Op = LI.getOperand(0);
824 // Try to canonicalize the loaded type.
825 if (Instruction *Res = combineLoadToOperationType(*this, LI))
828 // Attempt to improve the alignment.
829 unsigned KnownAlign = getOrEnforceKnownAlignment(
830 Op, DL.getPrefTypeAlignment(LI.getType()), DL, &LI, &AC, &DT);
831 unsigned LoadAlign = LI.getAlignment();
832 unsigned EffectiveLoadAlign =
833 LoadAlign != 0 ? LoadAlign : DL.getABITypeAlignment(LI.getType());
835 if (KnownAlign > EffectiveLoadAlign)
836 LI.setAlignment(KnownAlign);
837 else if (LoadAlign == 0)
838 LI.setAlignment(EffectiveLoadAlign);
840 // Replace GEP indices if possible.
841 if (Instruction *NewGEPI = replaceGEPIdxWithZero(*this, Op, LI)) {
842 Worklist.Add(NewGEPI);
846 if (Instruction *Res = unpackLoadToAggregate(*this, LI))
849 // Do really simple store-to-load forwarding and load CSE, to catch cases
850 // where there are several consecutive memory accesses to the same location,
851 // separated by a few arithmetic operations.
852 BasicBlock::iterator BBI(LI);
853 bool IsLoadCSE = false;
854 if (Value *AvailableVal = FindAvailableLoadedValue(
855 &LI, LI.getParent(), BBI, DefMaxInstsToScan, AA, &IsLoadCSE)) {
857 combineMetadataForCSE(cast<LoadInst>(AvailableVal), &LI);
859 return replaceInstUsesWith(
860 LI, Builder->CreateBitOrPointerCast(AvailableVal, LI.getType(),
861 LI.getName() + ".cast"));
864 // None of the following transforms are legal for volatile/ordered atomic
865 // loads. Most of them do apply for unordered atomics.
866 if (!LI.isUnordered()) return nullptr;
868 // load(gep null, ...) -> unreachable
869 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
870 const Value *GEPI0 = GEPI->getOperand(0);
871 // TODO: Consider a target hook for valid address spaces for this xform.
872 if (isa<ConstantPointerNull>(GEPI0) && GEPI->getPointerAddressSpace() == 0){
873 // Insert a new store to null instruction before the load to indicate
874 // that this code is not reachable. We do this instead of inserting
875 // an unreachable instruction directly because we cannot modify the
877 new StoreInst(UndefValue::get(LI.getType()),
878 Constant::getNullValue(Op->getType()), &LI);
879 return replaceInstUsesWith(LI, UndefValue::get(LI.getType()));
883 // load null/undef -> unreachable
884 // TODO: Consider a target hook for valid address spaces for this xform.
885 if (isa<UndefValue>(Op) ||
886 (isa<ConstantPointerNull>(Op) && LI.getPointerAddressSpace() == 0)) {
887 // Insert a new store to null instruction before the load to indicate that
888 // this code is not reachable. We do this instead of inserting an
889 // unreachable instruction directly because we cannot modify the CFG.
890 new StoreInst(UndefValue::get(LI.getType()),
891 Constant::getNullValue(Op->getType()), &LI);
892 return replaceInstUsesWith(LI, UndefValue::get(LI.getType()));
895 if (Op->hasOneUse()) {
896 // Change select and PHI nodes to select values instead of addresses: this
897 // helps alias analysis out a lot, allows many others simplifications, and
898 // exposes redundancy in the code.
900 // Note that we cannot do the transformation unless we know that the
901 // introduced loads cannot trap! Something like this is valid as long as
902 // the condition is always false: load (select bool %C, int* null, int* %G),
903 // but it would not be valid if we transformed it to load from null
906 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
907 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
908 unsigned Align = LI.getAlignment();
909 if (isSafeToLoadUnconditionally(SI->getOperand(1), Align, DL, SI) &&
910 isSafeToLoadUnconditionally(SI->getOperand(2), Align, DL, SI)) {
911 LoadInst *V1 = Builder->CreateLoad(SI->getOperand(1),
912 SI->getOperand(1)->getName()+".val");
913 LoadInst *V2 = Builder->CreateLoad(SI->getOperand(2),
914 SI->getOperand(2)->getName()+".val");
915 assert(LI.isUnordered() && "implied by above");
916 V1->setAlignment(Align);
917 V1->setAtomic(LI.getOrdering(), LI.getSynchScope());
918 V2->setAlignment(Align);
919 V2->setAtomic(LI.getOrdering(), LI.getSynchScope());
920 return SelectInst::Create(SI->getCondition(), V1, V2);
923 // load (select (cond, null, P)) -> load P
924 if (isa<ConstantPointerNull>(SI->getOperand(1)) &&
925 LI.getPointerAddressSpace() == 0) {
926 LI.setOperand(0, SI->getOperand(2));
930 // load (select (cond, P, null)) -> load P
931 if (isa<ConstantPointerNull>(SI->getOperand(2)) &&
932 LI.getPointerAddressSpace() == 0) {
933 LI.setOperand(0, SI->getOperand(1));
941 /// \brief Look for extractelement/insertvalue sequence that acts like a bitcast.
943 /// \returns underlying value that was "cast", or nullptr otherwise.
945 /// For example, if we have:
947 /// %E0 = extractelement <2 x double> %U, i32 0
948 /// %V0 = insertvalue [2 x double] undef, double %E0, 0
949 /// %E1 = extractelement <2 x double> %U, i32 1
950 /// %V1 = insertvalue [2 x double] %V0, double %E1, 1
952 /// and the layout of a <2 x double> is isomorphic to a [2 x double],
953 /// then %V1 can be safely approximated by a conceptual "bitcast" of %U.
954 /// Note that %U may contain non-undef values where %V1 has undef.
955 static Value *likeBitCastFromVector(InstCombiner &IC, Value *V) {
957 while (auto *IV = dyn_cast<InsertValueInst>(V)) {
958 auto *E = dyn_cast<ExtractElementInst>(IV->getInsertedValueOperand());
961 auto *W = E->getVectorOperand();
966 auto *CI = dyn_cast<ConstantInt>(E->getIndexOperand());
967 if (!CI || IV->getNumIndices() != 1 || CI->getZExtValue() != *IV->idx_begin())
969 V = IV->getAggregateOperand();
971 if (!isa<UndefValue>(V) ||!U)
974 auto *UT = cast<VectorType>(U->getType());
975 auto *VT = V->getType();
976 // Check that types UT and VT are bitwise isomorphic.
977 const auto &DL = IC.getDataLayout();
978 if (DL.getTypeStoreSizeInBits(UT) != DL.getTypeStoreSizeInBits(VT)) {
981 if (auto *AT = dyn_cast<ArrayType>(VT)) {
982 if (AT->getNumElements() != UT->getNumElements())
985 auto *ST = cast<StructType>(VT);
986 if (ST->getNumElements() != UT->getNumElements())
988 for (const auto *EltT : ST->elements()) {
989 if (EltT != UT->getElementType())
996 /// \brief Combine stores to match the type of value being stored.
998 /// The core idea here is that the memory does not have any intrinsic type and
999 /// where we can we should match the type of a store to the type of value being
1002 /// However, this routine must never change the width of a store or the number of
1003 /// stores as that would introduce a semantic change. This combine is expected to
1004 /// be a semantic no-op which just allows stores to more closely model the types
1005 /// of their incoming values.
1007 /// Currently, we also refuse to change the precise type used for an atomic or
1008 /// volatile store. This is debatable, and might be reasonable to change later.
1009 /// However, it is risky in case some backend or other part of LLVM is relying
1010 /// on the exact type stored to select appropriate atomic operations.
1012 /// \returns true if the store was successfully combined away. This indicates
1013 /// the caller must erase the store instruction. We have to let the caller erase
1014 /// the store instruction as otherwise there is no way to signal whether it was
1015 /// combined or not: IC.EraseInstFromFunction returns a null pointer.
1016 static bool combineStoreToValueType(InstCombiner &IC, StoreInst &SI) {
1017 // FIXME: We could probably with some care handle both volatile and ordered
1018 // atomic stores here but it isn't clear that this is important.
1019 if (!SI.isUnordered())
1022 // swifterror values can't be bitcasted.
1023 if (SI.getPointerOperand()->isSwiftError())
1026 Value *V = SI.getValueOperand();
1028 // Fold away bit casts of the stored value by storing the original type.
1029 if (auto *BC = dyn_cast<BitCastInst>(V)) {
1030 V = BC->getOperand(0);
1031 if (!SI.isAtomic() || isSupportedAtomicType(V->getType())) {
1032 combineStoreToNewValue(IC, SI, V);
1037 if (Value *U = likeBitCastFromVector(IC, V))
1038 if (!SI.isAtomic() || isSupportedAtomicType(U->getType())) {
1039 combineStoreToNewValue(IC, SI, U);
1043 // FIXME: We should also canonicalize stores of vectors when their elements
1044 // are cast to other types.
1048 static bool unpackStoreToAggregate(InstCombiner &IC, StoreInst &SI) {
1049 // FIXME: We could probably with some care handle both volatile and atomic
1050 // stores here but it isn't clear that this is important.
1054 Value *V = SI.getValueOperand();
1055 Type *T = V->getType();
1057 if (!T->isAggregateType())
1060 if (auto *ST = dyn_cast<StructType>(T)) {
1061 // If the struct only have one element, we unpack.
1062 unsigned Count = ST->getNumElements();
1064 V = IC.Builder->CreateExtractValue(V, 0);
1065 combineStoreToNewValue(IC, SI, V);
1069 // We don't want to break loads with padding here as we'd loose
1070 // the knowledge that padding exists for the rest of the pipeline.
1071 const DataLayout &DL = IC.getDataLayout();
1072 auto *SL = DL.getStructLayout(ST);
1073 if (SL->hasPadding())
1076 auto Align = SI.getAlignment();
1078 Align = DL.getABITypeAlignment(ST);
1080 SmallString<16> EltName = V->getName();
1082 auto *Addr = SI.getPointerOperand();
1083 SmallString<16> AddrName = Addr->getName();
1084 AddrName += ".repack";
1086 auto *IdxType = Type::getInt32Ty(ST->getContext());
1087 auto *Zero = ConstantInt::get(IdxType, 0);
1088 for (unsigned i = 0; i < Count; i++) {
1089 Value *Indices[2] = {
1091 ConstantInt::get(IdxType, i),
1093 auto *Ptr = IC.Builder->CreateInBoundsGEP(ST, Addr, makeArrayRef(Indices),
1095 auto *Val = IC.Builder->CreateExtractValue(V, i, EltName);
1096 auto EltAlign = MinAlign(Align, SL->getElementOffset(i));
1097 IC.Builder->CreateAlignedStore(Val, Ptr, EltAlign);
1103 if (auto *AT = dyn_cast<ArrayType>(T)) {
1104 // If the array only have one element, we unpack.
1105 auto NumElements = AT->getNumElements();
1106 if (NumElements == 1) {
1107 V = IC.Builder->CreateExtractValue(V, 0);
1108 combineStoreToNewValue(IC, SI, V);
1112 // Bail out if the array is too large. Ideally we would like to optimize
1113 // arrays of arbitrary size but this has a terrible impact on compile time.
1114 // The threshold here is chosen arbitrarily, maybe needs a little bit of
1116 if (NumElements > 1024)
1119 const DataLayout &DL = IC.getDataLayout();
1120 auto EltSize = DL.getTypeAllocSize(AT->getElementType());
1121 auto Align = SI.getAlignment();
1123 Align = DL.getABITypeAlignment(T);
1125 SmallString<16> EltName = V->getName();
1127 auto *Addr = SI.getPointerOperand();
1128 SmallString<16> AddrName = Addr->getName();
1129 AddrName += ".repack";
1131 auto *IdxType = Type::getInt64Ty(T->getContext());
1132 auto *Zero = ConstantInt::get(IdxType, 0);
1134 uint64_t Offset = 0;
1135 for (uint64_t i = 0; i < NumElements; i++) {
1136 Value *Indices[2] = {
1138 ConstantInt::get(IdxType, i),
1140 auto *Ptr = IC.Builder->CreateInBoundsGEP(AT, Addr, makeArrayRef(Indices),
1142 auto *Val = IC.Builder->CreateExtractValue(V, i, EltName);
1143 auto EltAlign = MinAlign(Align, Offset);
1144 IC.Builder->CreateAlignedStore(Val, Ptr, EltAlign);
1154 /// equivalentAddressValues - Test if A and B will obviously have the same
1155 /// value. This includes recognizing that %t0 and %t1 will have the same
1156 /// value in code like this:
1157 /// %t0 = getelementptr \@a, 0, 3
1158 /// store i32 0, i32* %t0
1159 /// %t1 = getelementptr \@a, 0, 3
1160 /// %t2 = load i32* %t1
1162 static bool equivalentAddressValues(Value *A, Value *B) {
1163 // Test if the values are trivially equivalent.
1164 if (A == B) return true;
1166 // Test if the values come form identical arithmetic instructions.
1167 // This uses isIdenticalToWhenDefined instead of isIdenticalTo because
1168 // its only used to compare two uses within the same basic block, which
1169 // means that they'll always either have the same value or one of them
1170 // will have an undefined value.
1171 if (isa<BinaryOperator>(A) ||
1174 isa<GetElementPtrInst>(A))
1175 if (Instruction *BI = dyn_cast<Instruction>(B))
1176 if (cast<Instruction>(A)->isIdenticalToWhenDefined(BI))
1179 // Otherwise they may not be equivalent.
1183 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
1184 Value *Val = SI.getOperand(0);
1185 Value *Ptr = SI.getOperand(1);
1187 // Try to canonicalize the stored type.
1188 if (combineStoreToValueType(*this, SI))
1189 return eraseInstFromFunction(SI);
1191 // Attempt to improve the alignment.
1192 unsigned KnownAlign = getOrEnforceKnownAlignment(
1193 Ptr, DL.getPrefTypeAlignment(Val->getType()), DL, &SI, &AC, &DT);
1194 unsigned StoreAlign = SI.getAlignment();
1195 unsigned EffectiveStoreAlign =
1196 StoreAlign != 0 ? StoreAlign : DL.getABITypeAlignment(Val->getType());
1198 if (KnownAlign > EffectiveStoreAlign)
1199 SI.setAlignment(KnownAlign);
1200 else if (StoreAlign == 0)
1201 SI.setAlignment(EffectiveStoreAlign);
1203 // Try to canonicalize the stored type.
1204 if (unpackStoreToAggregate(*this, SI))
1205 return eraseInstFromFunction(SI);
1207 // Replace GEP indices if possible.
1208 if (Instruction *NewGEPI = replaceGEPIdxWithZero(*this, Ptr, SI)) {
1209 Worklist.Add(NewGEPI);
1213 // Don't hack volatile/ordered stores.
1214 // FIXME: Some bits are legal for ordered atomic stores; needs refactoring.
1215 if (!SI.isUnordered()) return nullptr;
1217 // If the RHS is an alloca with a single use, zapify the store, making the
1219 if (Ptr->hasOneUse()) {
1220 if (isa<AllocaInst>(Ptr))
1221 return eraseInstFromFunction(SI);
1222 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
1223 if (isa<AllocaInst>(GEP->getOperand(0))) {
1224 if (GEP->getOperand(0)->hasOneUse())
1225 return eraseInstFromFunction(SI);
1230 // Do really simple DSE, to catch cases where there are several consecutive
1231 // stores to the same location, separated by a few arithmetic operations. This
1232 // situation often occurs with bitfield accesses.
1233 BasicBlock::iterator BBI(SI);
1234 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
1237 // Don't count debug info directives, lest they affect codegen,
1238 // and we skip pointer-to-pointer bitcasts, which are NOPs.
1239 if (isa<DbgInfoIntrinsic>(BBI) ||
1240 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy())) {
1245 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
1246 // Prev store isn't volatile, and stores to the same location?
1247 if (PrevSI->isUnordered() && equivalentAddressValues(PrevSI->getOperand(1),
1248 SI.getOperand(1))) {
1251 eraseInstFromFunction(*PrevSI);
1257 // If this is a load, we have to stop. However, if the loaded value is from
1258 // the pointer we're loading and is producing the pointer we're storing,
1259 // then *this* store is dead (X = load P; store X -> P).
1260 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
1261 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr)) {
1262 assert(SI.isUnordered() && "can't eliminate ordering operation");
1263 return eraseInstFromFunction(SI);
1266 // Otherwise, this is a load from some other location. Stores before it
1271 // Don't skip over loads or things that can modify memory.
1272 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
1276 // store X, null -> turns into 'unreachable' in SimplifyCFG
1277 if (isa<ConstantPointerNull>(Ptr) && SI.getPointerAddressSpace() == 0) {
1278 if (!isa<UndefValue>(Val)) {
1279 SI.setOperand(0, UndefValue::get(Val->getType()));
1280 if (Instruction *U = dyn_cast<Instruction>(Val))
1281 Worklist.Add(U); // Dropped a use.
1283 return nullptr; // Do not modify these!
1286 // store undef, Ptr -> noop
1287 if (isa<UndefValue>(Val))
1288 return eraseInstFromFunction(SI);
1290 // If this store is the last instruction in the basic block (possibly
1291 // excepting debug info instructions), and if the block ends with an
1292 // unconditional branch, try to move it to the successor block.
1293 BBI = SI.getIterator();
1296 } while (isa<DbgInfoIntrinsic>(BBI) ||
1297 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy()));
1298 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
1299 if (BI->isUnconditional())
1300 if (SimplifyStoreAtEndOfBlock(SI))
1301 return nullptr; // xform done!
1306 /// SimplifyStoreAtEndOfBlock - Turn things like:
1307 /// if () { *P = v1; } else { *P = v2 }
1308 /// into a phi node with a store in the successor.
1310 /// Simplify things like:
1311 /// *P = v1; if () { *P = v2; }
1312 /// into a phi node with a store in the successor.
1314 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
1315 assert(SI.isUnordered() &&
1316 "this code has not been auditted for volatile or ordered store case");
1318 BasicBlock *StoreBB = SI.getParent();
1320 // Check to see if the successor block has exactly two incoming edges. If
1321 // so, see if the other predecessor contains a store to the same location.
1322 // if so, insert a PHI node (if needed) and move the stores down.
1323 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
1325 // Determine whether Dest has exactly two predecessors and, if so, compute
1326 // the other predecessor.
1327 pred_iterator PI = pred_begin(DestBB);
1328 BasicBlock *P = *PI;
1329 BasicBlock *OtherBB = nullptr;
1334 if (++PI == pred_end(DestBB))
1343 if (++PI != pred_end(DestBB))
1346 // Bail out if all the relevant blocks aren't distinct (this can happen,
1347 // for example, if SI is in an infinite loop)
1348 if (StoreBB == DestBB || OtherBB == DestBB)
1351 // Verify that the other block ends in a branch and is not otherwise empty.
1352 BasicBlock::iterator BBI(OtherBB->getTerminator());
1353 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
1354 if (!OtherBr || BBI == OtherBB->begin())
1357 // If the other block ends in an unconditional branch, check for the 'if then
1358 // else' case. there is an instruction before the branch.
1359 StoreInst *OtherStore = nullptr;
1360 if (OtherBr->isUnconditional()) {
1362 // Skip over debugging info.
1363 while (isa<DbgInfoIntrinsic>(BBI) ||
1364 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy())) {
1365 if (BBI==OtherBB->begin())
1369 // If this isn't a store, isn't a store to the same location, or is not the
1370 // right kind of store, bail out.
1371 OtherStore = dyn_cast<StoreInst>(BBI);
1372 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1) ||
1373 !SI.isSameOperationAs(OtherStore))
1376 // Otherwise, the other block ended with a conditional branch. If one of the
1377 // destinations is StoreBB, then we have the if/then case.
1378 if (OtherBr->getSuccessor(0) != StoreBB &&
1379 OtherBr->getSuccessor(1) != StoreBB)
1382 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
1383 // if/then triangle. See if there is a store to the same ptr as SI that
1384 // lives in OtherBB.
1386 // Check to see if we find the matching store.
1387 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
1388 if (OtherStore->getOperand(1) != SI.getOperand(1) ||
1389 !SI.isSameOperationAs(OtherStore))
1393 // If we find something that may be using or overwriting the stored
1394 // value, or if we run out of instructions, we can't do the xform.
1395 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
1396 BBI == OtherBB->begin())
1400 // In order to eliminate the store in OtherBr, we have to
1401 // make sure nothing reads or overwrites the stored value in
1403 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
1404 // FIXME: This should really be AA driven.
1405 if (I->mayReadFromMemory() || I->mayWriteToMemory())
1410 // Insert a PHI node now if we need it.
1411 Value *MergedVal = OtherStore->getOperand(0);
1412 if (MergedVal != SI.getOperand(0)) {
1413 PHINode *PN = PHINode::Create(MergedVal->getType(), 2, "storemerge");
1414 PN->addIncoming(SI.getOperand(0), SI.getParent());
1415 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
1416 MergedVal = InsertNewInstBefore(PN, DestBB->front());
1419 // Advance to a place where it is safe to insert the new store and
1421 BBI = DestBB->getFirstInsertionPt();
1422 StoreInst *NewSI = new StoreInst(MergedVal, SI.getOperand(1),
1426 SI.getSynchScope());
1427 InsertNewInstBefore(NewSI, *BBI);
1428 NewSI->setDebugLoc(OtherStore->getDebugLoc());
1430 // If the two stores had AA tags, merge them.
1432 SI.getAAMetadata(AATags);
1434 OtherStore->getAAMetadata(AATags, /* Merge = */ true);
1435 NewSI->setAAMetadata(AATags);
1438 // Nuke the old stores.
1439 eraseInstFromFunction(SI);
1440 eraseInstFromFunction(*OtherStore);