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/DataLayout.h"
19 #include "llvm/IR/LLVMContext.h"
20 #include "llvm/IR/IntrinsicInst.h"
21 #include "llvm/IR/MDBuilder.h"
22 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
23 #include "llvm/Transforms/Utils/Local.h"
26 #define DEBUG_TYPE "instcombine"
28 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
29 STATISTIC(NumGlobalCopies, "Number of allocas copied from constant global");
31 /// pointsToConstantGlobal - Return true if V (possibly indirectly) points to
32 /// some part of a constant global variable. This intentionally only accepts
33 /// constant expressions because we can't rewrite arbitrary instructions.
34 static bool pointsToConstantGlobal(Value *V) {
35 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
36 return GV->isConstant();
38 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
39 if (CE->getOpcode() == Instruction::BitCast ||
40 CE->getOpcode() == Instruction::AddrSpaceCast ||
41 CE->getOpcode() == Instruction::GetElementPtr)
42 return pointsToConstantGlobal(CE->getOperand(0));
47 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
48 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
49 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
50 /// track of whether it moves the pointer (with IsOffset) but otherwise traverse
51 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
52 /// the alloca, and if the source pointer is a pointer to a constant global, we
53 /// can optimize this.
55 isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy,
56 SmallVectorImpl<Instruction *> &ToDelete) {
57 // We track lifetime intrinsics as we encounter them. If we decide to go
58 // ahead and replace the value with the global, this lets the caller quickly
59 // eliminate the markers.
61 SmallVector<std::pair<Value *, bool>, 35> ValuesToInspect;
62 ValuesToInspect.push_back(std::make_pair(V, false));
63 while (!ValuesToInspect.empty()) {
64 auto ValuePair = ValuesToInspect.pop_back_val();
65 const bool IsOffset = ValuePair.second;
66 for (auto &U : ValuePair.first->uses()) {
67 Instruction *I = cast<Instruction>(U.getUser());
69 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
70 // Ignore non-volatile loads, they are always ok.
71 if (!LI->isSimple()) return false;
75 if (isa<BitCastInst>(I) || isa<AddrSpaceCastInst>(I)) {
76 // If uses of the bitcast are ok, we are ok.
77 ValuesToInspect.push_back(std::make_pair(I, IsOffset));
80 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
81 // If the GEP has all zero indices, it doesn't offset the pointer. If it
83 ValuesToInspect.push_back(
84 std::make_pair(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 /// \brief Helper to combine a load to a new type.
313 /// This just does the work of combining a load to a new type. It handles
314 /// metadata, etc., and returns the new instruction. The \c NewTy should be the
315 /// loaded *value* type. This will convert it to a pointer, cast the operand to
316 /// that pointer type, load it, etc.
318 /// Note that this will create all of the instructions with whatever insert
319 /// point the \c InstCombiner currently is using.
320 static LoadInst *combineLoadToNewType(InstCombiner &IC, LoadInst &LI, Type *NewTy,
321 const Twine &Suffix = "") {
322 Value *Ptr = LI.getPointerOperand();
323 unsigned AS = LI.getPointerAddressSpace();
324 SmallVector<std::pair<unsigned, MDNode *>, 8> MD;
325 LI.getAllMetadata(MD);
327 LoadInst *NewLoad = IC.Builder->CreateAlignedLoad(
328 IC.Builder->CreateBitCast(Ptr, NewTy->getPointerTo(AS)),
329 LI.getAlignment(), LI.getName() + Suffix);
330 MDBuilder MDB(NewLoad->getContext());
331 for (const auto &MDPair : MD) {
332 unsigned ID = MDPair.first;
333 MDNode *N = MDPair.second;
334 // Note, essentially every kind of metadata should be preserved here! This
335 // routine is supposed to clone a load instruction changing *only its type*.
336 // The only metadata it makes sense to drop is metadata which is invalidated
337 // when the pointer type changes. This should essentially never be the case
338 // in LLVM, but we explicitly switch over only known metadata to be
339 // conservatively correct. If you are adding metadata to LLVM which pertains
340 // to loads, you almost certainly want to add it here.
342 case LLVMContext::MD_dbg:
343 case LLVMContext::MD_tbaa:
344 case LLVMContext::MD_prof:
345 case LLVMContext::MD_fpmath:
346 case LLVMContext::MD_tbaa_struct:
347 case LLVMContext::MD_invariant_load:
348 case LLVMContext::MD_alias_scope:
349 case LLVMContext::MD_noalias:
350 case LLVMContext::MD_nontemporal:
351 case LLVMContext::MD_mem_parallel_loop_access:
352 // All of these directly apply.
353 NewLoad->setMetadata(ID, N);
356 case LLVMContext::MD_nonnull:
357 // This only directly applies if the new type is also a pointer.
358 if (NewTy->isPointerTy()) {
359 NewLoad->setMetadata(ID, N);
362 // If it's integral now, translate it to !range metadata.
363 if (NewTy->isIntegerTy()) {
364 auto *ITy = cast<IntegerType>(NewTy);
365 auto *NullInt = ConstantExpr::getPtrToInt(
366 ConstantPointerNull::get(cast<PointerType>(Ptr->getType())), ITy);
368 ConstantExpr::getAdd(NullInt, ConstantInt::get(ITy, 1));
369 NewLoad->setMetadata(LLVMContext::MD_range,
370 MDB.createRange(NonNullInt, NullInt));
373 case LLVMContext::MD_align:
374 case LLVMContext::MD_dereferenceable:
375 case LLVMContext::MD_dereferenceable_or_null:
376 // These only directly apply if the new type is also a pointer.
377 if (NewTy->isPointerTy())
378 NewLoad->setMetadata(ID, N);
380 case LLVMContext::MD_range:
381 // FIXME: It would be nice to propagate this in some way, but the type
382 // conversions make it hard. If the new type is a pointer, we could
383 // translate it to !nonnull metadata.
390 /// \brief Combine a store to a new type.
392 /// Returns the newly created store instruction.
393 static StoreInst *combineStoreToNewValue(InstCombiner &IC, StoreInst &SI, Value *V) {
394 Value *Ptr = SI.getPointerOperand();
395 unsigned AS = SI.getPointerAddressSpace();
396 SmallVector<std::pair<unsigned, MDNode *>, 8> MD;
397 SI.getAllMetadata(MD);
399 StoreInst *NewStore = IC.Builder->CreateAlignedStore(
400 V, IC.Builder->CreateBitCast(Ptr, V->getType()->getPointerTo(AS)),
402 for (const auto &MDPair : MD) {
403 unsigned ID = MDPair.first;
404 MDNode *N = MDPair.second;
405 // Note, essentially every kind of metadata should be preserved here! This
406 // routine is supposed to clone a store instruction changing *only its
407 // type*. The only metadata it makes sense to drop is metadata which is
408 // invalidated when the pointer type changes. This should essentially
409 // never be the case in LLVM, but we explicitly switch over only known
410 // metadata to be conservatively correct. If you are adding metadata to
411 // LLVM which pertains to stores, you almost certainly want to add it
414 case LLVMContext::MD_dbg:
415 case LLVMContext::MD_tbaa:
416 case LLVMContext::MD_prof:
417 case LLVMContext::MD_fpmath:
418 case LLVMContext::MD_tbaa_struct:
419 case LLVMContext::MD_alias_scope:
420 case LLVMContext::MD_noalias:
421 case LLVMContext::MD_nontemporal:
422 case LLVMContext::MD_mem_parallel_loop_access:
423 // All of these directly apply.
424 NewStore->setMetadata(ID, N);
427 case LLVMContext::MD_invariant_load:
428 case LLVMContext::MD_nonnull:
429 case LLVMContext::MD_range:
430 case LLVMContext::MD_align:
431 case LLVMContext::MD_dereferenceable:
432 case LLVMContext::MD_dereferenceable_or_null:
433 // These don't apply for stores.
441 /// \brief Combine loads to match the type of value their uses after looking
442 /// through intervening bitcasts.
444 /// The core idea here is that if the result of a load is used in an operation,
445 /// we should load the type most conducive to that operation. For example, when
446 /// loading an integer and converting that immediately to a pointer, we should
447 /// instead directly load a pointer.
449 /// However, this routine must never change the width of a load or the number of
450 /// loads as that would introduce a semantic change. This combine is expected to
451 /// be a semantic no-op which just allows loads to more closely model the types
452 /// of their consuming operations.
454 /// Currently, we also refuse to change the precise type used for an atomic load
455 /// or a volatile load. This is debatable, and might be reasonable to change
456 /// later. However, it is risky in case some backend or other part of LLVM is
457 /// relying on the exact type loaded to select appropriate atomic operations.
458 static Instruction *combineLoadToOperationType(InstCombiner &IC, LoadInst &LI) {
459 // FIXME: We could probably with some care handle both volatile and atomic
460 // loads here but it isn't clear that this is important.
467 Type *Ty = LI.getType();
468 const DataLayout &DL = IC.getDataLayout();
470 // Try to canonicalize loads which are only ever stored to operate over
471 // integers instead of any other type. We only do this when the loaded type
472 // is sized and has a size exactly the same as its store size and the store
473 // size is a legal integer type.
474 if (!Ty->isIntegerTy() && Ty->isSized() &&
475 DL.isLegalInteger(DL.getTypeStoreSizeInBits(Ty)) &&
476 DL.getTypeStoreSizeInBits(Ty) == DL.getTypeSizeInBits(Ty)) {
477 if (std::all_of(LI.user_begin(), LI.user_end(), [&LI](User *U) {
478 auto *SI = dyn_cast<StoreInst>(U);
479 return SI && SI->getPointerOperand() != &LI;
481 LoadInst *NewLoad = combineLoadToNewType(
483 Type::getIntNTy(LI.getContext(), DL.getTypeStoreSizeInBits(Ty)));
484 // Replace all the stores with stores of the newly loaded value.
485 for (auto UI = LI.user_begin(), UE = LI.user_end(); UI != UE;) {
486 auto *SI = cast<StoreInst>(*UI++);
487 IC.Builder->SetInsertPoint(SI);
488 combineStoreToNewValue(IC, *SI, NewLoad);
489 IC.EraseInstFromFunction(*SI);
491 assert(LI.use_empty() && "Failed to remove all users of the load!");
492 // Return the old load so the combiner can delete it safely.
497 // Fold away bit casts of the loaded value by loading the desired type.
498 // We can do this for BitCastInsts as well as casts from and to pointer types,
499 // as long as those are noops (i.e., the source or dest type have the same
500 // bitwidth as the target's pointers).
502 if (auto* CI = dyn_cast<CastInst>(LI.user_back())) {
503 if (CI->isNoopCast(DL)) {
504 LoadInst *NewLoad = combineLoadToNewType(IC, LI, CI->getDestTy());
505 CI->replaceAllUsesWith(NewLoad);
506 IC.EraseInstFromFunction(*CI);
511 // FIXME: We should also canonicalize loads of vectors when their elements are
512 // cast to other types.
516 static Instruction *unpackLoadToAggregate(InstCombiner &IC, LoadInst &LI) {
517 // FIXME: We could probably with some care handle both volatile and atomic
518 // stores here but it isn't clear that this is important.
522 Type *T = LI.getType();
523 if (!T->isAggregateType())
526 assert(LI.getAlignment() && "Alignment must be set at this point");
528 if (auto *ST = dyn_cast<StructType>(T)) {
529 // If the struct only have one element, we unpack.
530 unsigned Count = ST->getNumElements();
532 LoadInst *NewLoad = combineLoadToNewType(IC, LI, ST->getTypeAtIndex(0U),
534 return IC.ReplaceInstUsesWith(LI, IC.Builder->CreateInsertValue(
535 UndefValue::get(T), NewLoad, 0, LI.getName()));
538 // We don't want to break loads with padding here as we'd loose
539 // the knowledge that padding exists for the rest of the pipeline.
540 const DataLayout &DL = IC.getDataLayout();
541 auto *SL = DL.getStructLayout(ST);
542 if (SL->hasPadding())
545 auto Name = LI.getName();
546 SmallString<16> LoadName = Name;
547 LoadName += ".unpack";
548 SmallString<16> EltName = Name;
550 auto *Addr = LI.getPointerOperand();
551 Value *V = UndefValue::get(T);
552 auto *IdxType = Type::getInt32Ty(ST->getContext());
553 auto *Zero = ConstantInt::get(IdxType, 0);
554 for (unsigned i = 0; i < Count; i++) {
555 Value *Indices[2] = {
557 ConstantInt::get(IdxType, i),
559 auto *Ptr = IC.Builder->CreateInBoundsGEP(ST, Addr, makeArrayRef(Indices), EltName);
560 auto *L = IC.Builder->CreateAlignedLoad(Ptr, LI.getAlignment(),
562 V = IC.Builder->CreateInsertValue(V, L, i);
566 return IC.ReplaceInstUsesWith(LI, V);
569 if (auto *AT = dyn_cast<ArrayType>(T)) {
570 // If the array only have one element, we unpack.
571 if (AT->getNumElements() == 1) {
572 LoadInst *NewLoad = combineLoadToNewType(IC, LI, AT->getElementType(),
574 return IC.ReplaceInstUsesWith(LI, IC.Builder->CreateInsertValue(
575 UndefValue::get(T), NewLoad, 0, LI.getName()));
582 // If we can determine that all possible objects pointed to by the provided
583 // pointer value are, not only dereferenceable, but also definitively less than
584 // or equal to the provided maximum size, then return true. Otherwise, return
585 // false (constant global values and allocas fall into this category).
587 // FIXME: This should probably live in ValueTracking (or similar).
588 static bool isObjectSizeLessThanOrEq(Value *V, uint64_t MaxSize,
589 const DataLayout &DL) {
590 SmallPtrSet<Value *, 4> Visited;
591 SmallVector<Value *, 4> Worklist(1, V);
594 Value *P = Worklist.pop_back_val();
595 P = P->stripPointerCasts();
597 if (!Visited.insert(P).second)
600 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
601 Worklist.push_back(SI->getTrueValue());
602 Worklist.push_back(SI->getFalseValue());
606 if (PHINode *PN = dyn_cast<PHINode>(P)) {
607 for (Value *IncValue : PN->incoming_values())
608 Worklist.push_back(IncValue);
612 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(P)) {
613 if (GA->mayBeOverridden())
615 Worklist.push_back(GA->getAliasee());
619 // If we know how big this object is, and it is less than MaxSize, continue
620 // searching. Otherwise, return false.
621 if (AllocaInst *AI = dyn_cast<AllocaInst>(P)) {
622 if (!AI->getAllocatedType()->isSized())
625 ConstantInt *CS = dyn_cast<ConstantInt>(AI->getArraySize());
629 uint64_t TypeSize = DL.getTypeAllocSize(AI->getAllocatedType());
630 // Make sure that, even if the multiplication below would wrap as an
631 // uint64_t, we still do the right thing.
632 if ((CS->getValue().zextOrSelf(128)*APInt(128, TypeSize)).ugt(MaxSize))
637 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
638 if (!GV->hasDefinitiveInitializer() || !GV->isConstant())
641 uint64_t InitSize = DL.getTypeAllocSize(GV->getType()->getElementType());
642 if (InitSize > MaxSize)
648 } while (!Worklist.empty());
653 // If we're indexing into an object of a known size, and the outer index is
654 // not a constant, but having any value but zero would lead to undefined
655 // behavior, replace it with zero.
657 // For example, if we have:
658 // @f.a = private unnamed_addr constant [1 x i32] [i32 12], align 4
660 // %arrayidx = getelementptr inbounds [1 x i32]* @f.a, i64 0, i64 %x
661 // ... = load i32* %arrayidx, align 4
662 // Then we know that we can replace %x in the GEP with i64 0.
664 // FIXME: We could fold any GEP index to zero that would cause UB if it were
665 // not zero. Currently, we only handle the first such index. Also, we could
666 // also search through non-zero constant indices if we kept track of the
667 // offsets those indices implied.
668 static bool canReplaceGEPIdxWithZero(InstCombiner &IC, GetElementPtrInst *GEPI,
669 Instruction *MemI, unsigned &Idx) {
670 if (GEPI->getNumOperands() < 2)
673 // Find the first non-zero index of a GEP. If all indices are zero, return
674 // one past the last index.
675 auto FirstNZIdx = [](const GetElementPtrInst *GEPI) {
677 for (unsigned IE = GEPI->getNumOperands(); I != IE; ++I) {
678 Value *V = GEPI->getOperand(I);
679 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V))
689 // Skip through initial 'zero' indices, and find the corresponding pointer
690 // type. See if the next index is not a constant.
691 Idx = FirstNZIdx(GEPI);
692 if (Idx == GEPI->getNumOperands())
694 if (isa<Constant>(GEPI->getOperand(Idx)))
697 SmallVector<Value *, 4> Ops(GEPI->idx_begin(), GEPI->idx_begin() + Idx);
698 Type *AllocTy = GetElementPtrInst::getIndexedType(
699 cast<PointerType>(GEPI->getOperand(0)->getType()->getScalarType())
702 if (!AllocTy || !AllocTy->isSized())
704 const DataLayout &DL = IC.getDataLayout();
705 uint64_t TyAllocSize = DL.getTypeAllocSize(AllocTy);
707 // If there are more indices after the one we might replace with a zero, make
708 // sure they're all non-negative. If any of them are negative, the overall
709 // address being computed might be before the base address determined by the
710 // first non-zero index.
711 auto IsAllNonNegative = [&]() {
712 for (unsigned i = Idx+1, e = GEPI->getNumOperands(); i != e; ++i) {
713 bool KnownNonNegative, KnownNegative;
714 IC.ComputeSignBit(GEPI->getOperand(i), KnownNonNegative,
715 KnownNegative, 0, MemI);
716 if (KnownNonNegative)
724 // FIXME: If the GEP is not inbounds, and there are extra indices after the
725 // one we'll replace, those could cause the address computation to wrap
726 // (rendering the IsAllNonNegative() check below insufficient). We can do
727 // better, ignoring zero indices (and other indices we can prove small
728 // enough not to wrap).
729 if (Idx+1 != GEPI->getNumOperands() && !GEPI->isInBounds())
732 // Note that isObjectSizeLessThanOrEq will return true only if the pointer is
733 // also known to be dereferenceable.
734 return isObjectSizeLessThanOrEq(GEPI->getOperand(0), TyAllocSize, DL) &&
738 // If we're indexing into an object with a variable index for the memory
739 // access, but the object has only one element, we can assume that the index
740 // will always be zero. If we replace the GEP, return it.
741 template <typename T>
742 static Instruction *replaceGEPIdxWithZero(InstCombiner &IC, Value *Ptr,
744 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Ptr)) {
746 if (canReplaceGEPIdxWithZero(IC, GEPI, &MemI, Idx)) {
747 Instruction *NewGEPI = GEPI->clone();
748 NewGEPI->setOperand(Idx,
749 ConstantInt::get(GEPI->getOperand(Idx)->getType(), 0));
750 NewGEPI->insertBefore(GEPI);
751 MemI.setOperand(MemI.getPointerOperandIndex(), NewGEPI);
759 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
760 Value *Op = LI.getOperand(0);
762 // Try to canonicalize the loaded type.
763 if (Instruction *Res = combineLoadToOperationType(*this, LI))
766 // Attempt to improve the alignment.
767 unsigned KnownAlign = getOrEnforceKnownAlignment(
768 Op, DL.getPrefTypeAlignment(LI.getType()), DL, &LI, AC, DT);
769 unsigned LoadAlign = LI.getAlignment();
770 unsigned EffectiveLoadAlign =
771 LoadAlign != 0 ? LoadAlign : DL.getABITypeAlignment(LI.getType());
773 if (KnownAlign > EffectiveLoadAlign)
774 LI.setAlignment(KnownAlign);
775 else if (LoadAlign == 0)
776 LI.setAlignment(EffectiveLoadAlign);
778 // Replace GEP indices if possible.
779 if (Instruction *NewGEPI = replaceGEPIdxWithZero(*this, Op, LI)) {
780 Worklist.Add(NewGEPI);
784 // None of the following transforms are legal for volatile/atomic loads.
785 // FIXME: Some of it is okay for atomic loads; needs refactoring.
786 if (!LI.isSimple()) return nullptr;
788 if (Instruction *Res = unpackLoadToAggregate(*this, LI))
791 // Do really simple store-to-load forwarding and load CSE, to catch cases
792 // where there are several consecutive memory accesses to the same location,
793 // separated by a few arithmetic operations.
794 BasicBlock::iterator BBI(LI);
796 if (Value *AvailableVal =
797 FindAvailableLoadedValue(Op, LI.getParent(), BBI,
798 DefMaxInstsToScan, AA, &AATags)) {
799 if (LoadInst *NLI = dyn_cast<LoadInst>(AvailableVal)) {
800 unsigned KnownIDs[] = {
801 LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope,
802 LLVMContext::MD_noalias, LLVMContext::MD_range,
803 LLVMContext::MD_invariant_load, LLVMContext::MD_nonnull,
804 LLVMContext::MD_invariant_group, LLVMContext::MD_align,
805 LLVMContext::MD_dereferenceable,
806 LLVMContext::MD_dereferenceable_or_null};
807 combineMetadata(NLI, &LI, KnownIDs);
810 return ReplaceInstUsesWith(
811 LI, Builder->CreateBitOrPointerCast(AvailableVal, LI.getType(),
812 LI.getName() + ".cast"));
815 // load(gep null, ...) -> unreachable
816 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
817 const Value *GEPI0 = GEPI->getOperand(0);
818 // TODO: Consider a target hook for valid address spaces for this xform.
819 if (isa<ConstantPointerNull>(GEPI0) && GEPI->getPointerAddressSpace() == 0){
820 // Insert a new store to null instruction before the load to indicate
821 // that this code is not reachable. We do this instead of inserting
822 // an unreachable instruction directly because we cannot modify the
824 new StoreInst(UndefValue::get(LI.getType()),
825 Constant::getNullValue(Op->getType()), &LI);
826 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
830 // load null/undef -> unreachable
831 // TODO: Consider a target hook for valid address spaces for this xform.
832 if (isa<UndefValue>(Op) ||
833 (isa<ConstantPointerNull>(Op) && LI.getPointerAddressSpace() == 0)) {
834 // Insert a new store to null instruction before the load to indicate that
835 // this code is not reachable. We do this instead of inserting an
836 // unreachable instruction directly because we cannot modify the CFG.
837 new StoreInst(UndefValue::get(LI.getType()),
838 Constant::getNullValue(Op->getType()), &LI);
839 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
842 if (Op->hasOneUse()) {
843 // Change select and PHI nodes to select values instead of addresses: this
844 // helps alias analysis out a lot, allows many others simplifications, and
845 // exposes redundancy in the code.
847 // Note that we cannot do the transformation unless we know that the
848 // introduced loads cannot trap! Something like this is valid as long as
849 // the condition is always false: load (select bool %C, int* null, int* %G),
850 // but it would not be valid if we transformed it to load from null
853 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
854 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
855 unsigned Align = LI.getAlignment();
856 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI, Align) &&
857 isSafeToLoadUnconditionally(SI->getOperand(2), SI, Align)) {
858 LoadInst *V1 = Builder->CreateLoad(SI->getOperand(1),
859 SI->getOperand(1)->getName()+".val");
860 LoadInst *V2 = Builder->CreateLoad(SI->getOperand(2),
861 SI->getOperand(2)->getName()+".val");
862 V1->setAlignment(Align);
863 V2->setAlignment(Align);
864 return SelectInst::Create(SI->getCondition(), V1, V2);
867 // load (select (cond, null, P)) -> load P
868 if (isa<ConstantPointerNull>(SI->getOperand(1)) &&
869 LI.getPointerAddressSpace() == 0) {
870 LI.setOperand(0, SI->getOperand(2));
874 // load (select (cond, P, null)) -> load P
875 if (isa<ConstantPointerNull>(SI->getOperand(2)) &&
876 LI.getPointerAddressSpace() == 0) {
877 LI.setOperand(0, SI->getOperand(1));
885 /// \brief Combine stores to match the type of value being stored.
887 /// The core idea here is that the memory does not have any intrinsic type and
888 /// where we can we should match the type of a store to the type of value being
891 /// However, this routine must never change the width of a store or the number of
892 /// stores as that would introduce a semantic change. This combine is expected to
893 /// be a semantic no-op which just allows stores to more closely model the types
894 /// of their incoming values.
896 /// Currently, we also refuse to change the precise type used for an atomic or
897 /// volatile store. This is debatable, and might be reasonable to change later.
898 /// However, it is risky in case some backend or other part of LLVM is relying
899 /// on the exact type stored to select appropriate atomic operations.
901 /// \returns true if the store was successfully combined away. This indicates
902 /// the caller must erase the store instruction. We have to let the caller erase
903 /// the store instruction as otherwise there is no way to signal whether it was
904 /// combined or not: IC.EraseInstFromFunction returns a null pointer.
905 static bool combineStoreToValueType(InstCombiner &IC, StoreInst &SI) {
906 // FIXME: We could probably with some care handle both volatile and atomic
907 // stores here but it isn't clear that this is important.
911 Value *V = SI.getValueOperand();
913 // Fold away bit casts of the stored value by storing the original type.
914 if (auto *BC = dyn_cast<BitCastInst>(V)) {
915 V = BC->getOperand(0);
916 combineStoreToNewValue(IC, SI, V);
920 // FIXME: We should also canonicalize loads of vectors when their elements are
921 // cast to other types.
925 static bool unpackStoreToAggregate(InstCombiner &IC, StoreInst &SI) {
926 // FIXME: We could probably with some care handle both volatile and atomic
927 // stores here but it isn't clear that this is important.
931 Value *V = SI.getValueOperand();
932 Type *T = V->getType();
934 if (!T->isAggregateType())
937 if (auto *ST = dyn_cast<StructType>(T)) {
938 // If the struct only have one element, we unpack.
939 unsigned Count = ST->getNumElements();
941 V = IC.Builder->CreateExtractValue(V, 0);
942 combineStoreToNewValue(IC, SI, V);
946 // We don't want to break loads with padding here as we'd loose
947 // the knowledge that padding exists for the rest of the pipeline.
948 const DataLayout &DL = IC.getDataLayout();
949 auto *SL = DL.getStructLayout(ST);
950 if (SL->hasPadding())
953 SmallString<16> EltName = V->getName();
955 auto *Addr = SI.getPointerOperand();
956 SmallString<16> AddrName = Addr->getName();
957 AddrName += ".repack";
958 auto *IdxType = Type::getInt32Ty(ST->getContext());
959 auto *Zero = ConstantInt::get(IdxType, 0);
960 for (unsigned i = 0; i < Count; i++) {
961 Value *Indices[2] = {
963 ConstantInt::get(IdxType, i),
965 auto *Ptr = IC.Builder->CreateInBoundsGEP(ST, Addr, makeArrayRef(Indices), AddrName);
966 auto *Val = IC.Builder->CreateExtractValue(V, i, EltName);
967 IC.Builder->CreateStore(Val, Ptr);
973 if (auto *AT = dyn_cast<ArrayType>(T)) {
974 // If the array only have one element, we unpack.
975 if (AT->getNumElements() == 1) {
976 V = IC.Builder->CreateExtractValue(V, 0);
977 combineStoreToNewValue(IC, SI, V);
985 /// equivalentAddressValues - Test if A and B will obviously have the same
986 /// value. This includes recognizing that %t0 and %t1 will have the same
987 /// value in code like this:
988 /// %t0 = getelementptr \@a, 0, 3
989 /// store i32 0, i32* %t0
990 /// %t1 = getelementptr \@a, 0, 3
991 /// %t2 = load i32* %t1
993 static bool equivalentAddressValues(Value *A, Value *B) {
994 // Test if the values are trivially equivalent.
995 if (A == B) return true;
997 // Test if the values come form identical arithmetic instructions.
998 // This uses isIdenticalToWhenDefined instead of isIdenticalTo because
999 // its only used to compare two uses within the same basic block, which
1000 // means that they'll always either have the same value or one of them
1001 // will have an undefined value.
1002 if (isa<BinaryOperator>(A) ||
1005 isa<GetElementPtrInst>(A))
1006 if (Instruction *BI = dyn_cast<Instruction>(B))
1007 if (cast<Instruction>(A)->isIdenticalToWhenDefined(BI))
1010 // Otherwise they may not be equivalent.
1014 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
1015 Value *Val = SI.getOperand(0);
1016 Value *Ptr = SI.getOperand(1);
1018 // Try to canonicalize the stored type.
1019 if (combineStoreToValueType(*this, SI))
1020 return EraseInstFromFunction(SI);
1022 // Attempt to improve the alignment.
1023 unsigned KnownAlign = getOrEnforceKnownAlignment(
1024 Ptr, DL.getPrefTypeAlignment(Val->getType()), DL, &SI, AC, DT);
1025 unsigned StoreAlign = SI.getAlignment();
1026 unsigned EffectiveStoreAlign =
1027 StoreAlign != 0 ? StoreAlign : DL.getABITypeAlignment(Val->getType());
1029 if (KnownAlign > EffectiveStoreAlign)
1030 SI.setAlignment(KnownAlign);
1031 else if (StoreAlign == 0)
1032 SI.setAlignment(EffectiveStoreAlign);
1034 // Try to canonicalize the stored type.
1035 if (unpackStoreToAggregate(*this, SI))
1036 return EraseInstFromFunction(SI);
1038 // Replace GEP indices if possible.
1039 if (Instruction *NewGEPI = replaceGEPIdxWithZero(*this, Ptr, SI)) {
1040 Worklist.Add(NewGEPI);
1044 // Don't hack volatile/ordered stores.
1045 // FIXME: Some bits are legal for ordered atomic stores; needs refactoring.
1046 if (!SI.isUnordered()) return nullptr;
1048 // If the RHS is an alloca with a single use, zapify the store, making the
1050 if (Ptr->hasOneUse()) {
1051 if (isa<AllocaInst>(Ptr))
1052 return EraseInstFromFunction(SI);
1053 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
1054 if (isa<AllocaInst>(GEP->getOperand(0))) {
1055 if (GEP->getOperand(0)->hasOneUse())
1056 return EraseInstFromFunction(SI);
1061 // Do really simple DSE, to catch cases where there are several consecutive
1062 // stores to the same location, separated by a few arithmetic operations. This
1063 // situation often occurs with bitfield accesses.
1064 BasicBlock::iterator BBI(SI);
1065 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
1068 // Don't count debug info directives, lest they affect codegen,
1069 // and we skip pointer-to-pointer bitcasts, which are NOPs.
1070 if (isa<DbgInfoIntrinsic>(BBI) ||
1071 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy())) {
1076 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
1077 // Prev store isn't volatile, and stores to the same location?
1078 if (PrevSI->isUnordered() && equivalentAddressValues(PrevSI->getOperand(1),
1079 SI.getOperand(1))) {
1082 EraseInstFromFunction(*PrevSI);
1088 // If this is a load, we have to stop. However, if the loaded value is from
1089 // the pointer we're loading and is producing the pointer we're storing,
1090 // then *this* store is dead (X = load P; store X -> P).
1091 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
1092 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr)) {
1093 assert(SI.isUnordered() && "can't eliminate ordering operation");
1094 return EraseInstFromFunction(SI);
1097 // Otherwise, this is a load from some other location. Stores before it
1102 // Don't skip over loads or things that can modify memory.
1103 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
1107 // store X, null -> turns into 'unreachable' in SimplifyCFG
1108 if (isa<ConstantPointerNull>(Ptr) && SI.getPointerAddressSpace() == 0) {
1109 if (!isa<UndefValue>(Val)) {
1110 SI.setOperand(0, UndefValue::get(Val->getType()));
1111 if (Instruction *U = dyn_cast<Instruction>(Val))
1112 Worklist.Add(U); // Dropped a use.
1114 return nullptr; // Do not modify these!
1117 // store undef, Ptr -> noop
1118 if (isa<UndefValue>(Val))
1119 return EraseInstFromFunction(SI);
1121 // The code below needs to be audited and adjusted for unordered atomics
1125 // If this store is the last instruction in the basic block (possibly
1126 // excepting debug info instructions), and if the block ends with an
1127 // unconditional branch, try to move it to the successor block.
1128 BBI = SI.getIterator();
1131 } while (isa<DbgInfoIntrinsic>(BBI) ||
1132 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy()));
1133 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
1134 if (BI->isUnconditional())
1135 if (SimplifyStoreAtEndOfBlock(SI))
1136 return nullptr; // xform done!
1141 /// SimplifyStoreAtEndOfBlock - Turn things like:
1142 /// if () { *P = v1; } else { *P = v2 }
1143 /// into a phi node with a store in the successor.
1145 /// Simplify things like:
1146 /// *P = v1; if () { *P = v2; }
1147 /// into a phi node with a store in the successor.
1149 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
1150 BasicBlock *StoreBB = SI.getParent();
1152 // Check to see if the successor block has exactly two incoming edges. If
1153 // so, see if the other predecessor contains a store to the same location.
1154 // if so, insert a PHI node (if needed) and move the stores down.
1155 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
1157 // Determine whether Dest has exactly two predecessors and, if so, compute
1158 // the other predecessor.
1159 pred_iterator PI = pred_begin(DestBB);
1160 BasicBlock *P = *PI;
1161 BasicBlock *OtherBB = nullptr;
1166 if (++PI == pred_end(DestBB))
1175 if (++PI != pred_end(DestBB))
1178 // Bail out if all the relevant blocks aren't distinct (this can happen,
1179 // for example, if SI is in an infinite loop)
1180 if (StoreBB == DestBB || OtherBB == DestBB)
1183 // Verify that the other block ends in a branch and is not otherwise empty.
1184 BasicBlock::iterator BBI(OtherBB->getTerminator());
1185 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
1186 if (!OtherBr || BBI == OtherBB->begin())
1189 // If the other block ends in an unconditional branch, check for the 'if then
1190 // else' case. there is an instruction before the branch.
1191 StoreInst *OtherStore = nullptr;
1192 if (OtherBr->isUnconditional()) {
1194 // Skip over debugging info.
1195 while (isa<DbgInfoIntrinsic>(BBI) ||
1196 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy())) {
1197 if (BBI==OtherBB->begin())
1201 // If this isn't a store, isn't a store to the same location, or is not the
1202 // right kind of store, bail out.
1203 OtherStore = dyn_cast<StoreInst>(BBI);
1204 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1) ||
1205 !SI.isSameOperationAs(OtherStore))
1208 // Otherwise, the other block ended with a conditional branch. If one of the
1209 // destinations is StoreBB, then we have the if/then case.
1210 if (OtherBr->getSuccessor(0) != StoreBB &&
1211 OtherBr->getSuccessor(1) != StoreBB)
1214 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
1215 // if/then triangle. See if there is a store to the same ptr as SI that
1216 // lives in OtherBB.
1218 // Check to see if we find the matching store.
1219 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
1220 if (OtherStore->getOperand(1) != SI.getOperand(1) ||
1221 !SI.isSameOperationAs(OtherStore))
1225 // If we find something that may be using or overwriting the stored
1226 // value, or if we run out of instructions, we can't do the xform.
1227 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
1228 BBI == OtherBB->begin())
1232 // In order to eliminate the store in OtherBr, we have to
1233 // make sure nothing reads or overwrites the stored value in
1235 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
1236 // FIXME: This should really be AA driven.
1237 if (I->mayReadFromMemory() || I->mayWriteToMemory())
1242 // Insert a PHI node now if we need it.
1243 Value *MergedVal = OtherStore->getOperand(0);
1244 if (MergedVal != SI.getOperand(0)) {
1245 PHINode *PN = PHINode::Create(MergedVal->getType(), 2, "storemerge");
1246 PN->addIncoming(SI.getOperand(0), SI.getParent());
1247 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
1248 MergedVal = InsertNewInstBefore(PN, DestBB->front());
1251 // Advance to a place where it is safe to insert the new store and
1253 BBI = DestBB->getFirstInsertionPt();
1254 StoreInst *NewSI = new StoreInst(MergedVal, SI.getOperand(1),
1258 SI.getSynchScope());
1259 InsertNewInstBefore(NewSI, *BBI);
1260 NewSI->setDebugLoc(OtherStore->getDebugLoc());
1262 // If the two stores had AA tags, merge them.
1264 SI.getAAMetadata(AATags);
1266 OtherStore->getAAMetadata(AATags, /* Merge = */ true);
1267 NewSI->setAAMetadata(AATags);
1270 // Nuke the old stores.
1271 EraseInstFromFunction(SI);
1272 EraseInstFromFunction(*OtherStore);