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 "InstCombine.h"
15 #include "llvm/ADT/Statistic.h"
16 #include "llvm/Analysis/Loads.h"
17 #include "llvm/IR/DataLayout.h"
18 #include "llvm/IR/LLVMContext.h"
19 #include "llvm/IR/IntrinsicInst.h"
20 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
21 #include "llvm/Transforms/Utils/Local.h"
24 #define DEBUG_TYPE "instcombine"
26 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
27 STATISTIC(NumGlobalCopies, "Number of allocas copied from constant global");
29 /// pointsToConstantGlobal - Return true if V (possibly indirectly) points to
30 /// some part of a constant global variable. This intentionally only accepts
31 /// constant expressions because we can't rewrite arbitrary instructions.
32 static bool pointsToConstantGlobal(Value *V) {
33 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
34 return GV->isConstant();
36 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
37 if (CE->getOpcode() == Instruction::BitCast ||
38 CE->getOpcode() == Instruction::AddrSpaceCast ||
39 CE->getOpcode() == Instruction::GetElementPtr)
40 return pointsToConstantGlobal(CE->getOperand(0));
45 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
46 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
47 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
48 /// track of whether it moves the pointer (with IsOffset) but otherwise traverse
49 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
50 /// the alloca, and if the source pointer is a pointer to a constant global, we
51 /// can optimize this.
53 isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy,
54 SmallVectorImpl<Instruction *> &ToDelete) {
55 // We track lifetime intrinsics as we encounter them. If we decide to go
56 // ahead and replace the value with the global, this lets the caller quickly
57 // eliminate the markers.
59 SmallVector<std::pair<Value *, bool>, 35> ValuesToInspect;
60 ValuesToInspect.push_back(std::make_pair(V, false));
61 while (!ValuesToInspect.empty()) {
62 auto ValuePair = ValuesToInspect.pop_back_val();
63 const bool IsOffset = ValuePair.second;
64 for (auto &U : ValuePair.first->uses()) {
65 Instruction *I = cast<Instruction>(U.getUser());
67 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
68 // Ignore non-volatile loads, they are always ok.
69 if (!LI->isSimple()) return false;
73 if (isa<BitCastInst>(I) || isa<AddrSpaceCastInst>(I)) {
74 // If uses of the bitcast are ok, we are ok.
75 ValuesToInspect.push_back(std::make_pair(I, IsOffset));
78 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
79 // If the GEP has all zero indices, it doesn't offset the pointer. If it
81 ValuesToInspect.push_back(
82 std::make_pair(I, IsOffset || !GEP->hasAllZeroIndices()));
86 if (CallSite CS = I) {
87 // If this is the function being called then we treat it like a load and
92 // Inalloca arguments are clobbered by the call.
93 unsigned ArgNo = CS.getArgumentNo(&U);
94 if (CS.isInAllocaArgument(ArgNo))
97 // If this is a readonly/readnone call site, then we know it is just a
98 // load (but one that potentially returns the value itself), so we can
99 // ignore it if we know that the value isn't captured.
100 if (CS.onlyReadsMemory() &&
101 (CS.getInstruction()->use_empty() || CS.doesNotCapture(ArgNo)))
104 // If this is being passed as a byval argument, the caller is making a
105 // copy, so it is only a read of the alloca.
106 if (CS.isByValArgument(ArgNo))
110 // Lifetime intrinsics can be handled by the caller.
111 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
112 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
113 II->getIntrinsicID() == Intrinsic::lifetime_end) {
114 assert(II->use_empty() && "Lifetime markers have no result to use!");
115 ToDelete.push_back(II);
120 // If this is isn't our memcpy/memmove, reject it as something we can't
122 MemTransferInst *MI = dyn_cast<MemTransferInst>(I);
126 // If the transfer is using the alloca as a source of the transfer, then
127 // ignore it since it is a load (unless the transfer is volatile).
128 if (U.getOperandNo() == 1) {
129 if (MI->isVolatile()) return false;
133 // If we already have seen a copy, reject the second one.
134 if (TheCopy) return false;
136 // If the pointer has been offset from the start of the alloca, we can't
137 // safely handle this.
138 if (IsOffset) return false;
140 // If the memintrinsic isn't using the alloca as the dest, reject it.
141 if (U.getOperandNo() != 0) return false;
143 // If the source of the memcpy/move is not a constant global, reject it.
144 if (!pointsToConstantGlobal(MI->getSource()))
147 // Otherwise, the transform is safe. Remember the copy instruction.
154 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
155 /// modified by a copy from a constant global. If we can prove this, we can
156 /// replace any uses of the alloca with uses of the global directly.
157 static MemTransferInst *
158 isOnlyCopiedFromConstantGlobal(AllocaInst *AI,
159 SmallVectorImpl<Instruction *> &ToDelete) {
160 MemTransferInst *TheCopy = nullptr;
161 if (isOnlyCopiedFromConstantGlobal(AI, TheCopy, ToDelete))
166 Instruction *InstCombiner::visitAllocaInst(AllocaInst &AI) {
167 // Ensure that the alloca array size argument has type intptr_t, so that
168 // any casting is exposed early.
170 Type *IntPtrTy = DL->getIntPtrType(AI.getType());
171 if (AI.getArraySize()->getType() != IntPtrTy) {
172 Value *V = Builder->CreateIntCast(AI.getArraySize(),
179 // Convert: alloca Ty, C - where C is a constant != 1 into: alloca [C x Ty], 1
180 if (AI.isArrayAllocation()) { // Check C != 1
181 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
183 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
184 AllocaInst *New = Builder->CreateAlloca(NewTy, nullptr, AI.getName());
185 New->setAlignment(AI.getAlignment());
187 // Scan to the end of the allocation instructions, to skip over a block of
188 // allocas if possible...also skip interleaved debug info
190 BasicBlock::iterator It = New;
191 while (isa<AllocaInst>(*It) || isa<DbgInfoIntrinsic>(*It)) ++It;
193 // Now that I is pointing to the first non-allocation-inst in the block,
194 // insert our getelementptr instruction...
197 ? DL->getIntPtrType(AI.getType())
198 : Type::getInt64Ty(AI.getContext());
199 Value *NullIdx = Constant::getNullValue(IdxTy);
200 Value *Idx[2] = { NullIdx, NullIdx };
202 GetElementPtrInst::CreateInBounds(New, Idx, New->getName() + ".sub");
203 InsertNewInstBefore(GEP, *It);
205 // Now make everything use the getelementptr instead of the original
207 return ReplaceInstUsesWith(AI, GEP);
208 } else if (isa<UndefValue>(AI.getArraySize())) {
209 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
213 if (DL && AI.getAllocatedType()->isSized()) {
214 // If the alignment is 0 (unspecified), assign it the preferred alignment.
215 if (AI.getAlignment() == 0)
216 AI.setAlignment(DL->getPrefTypeAlignment(AI.getAllocatedType()));
218 // Move all alloca's of zero byte objects to the entry block and merge them
219 // together. Note that we only do this for alloca's, because malloc should
220 // allocate and return a unique pointer, even for a zero byte allocation.
221 if (DL->getTypeAllocSize(AI.getAllocatedType()) == 0) {
222 // For a zero sized alloca there is no point in doing an array allocation.
223 // This is helpful if the array size is a complicated expression not used
225 if (AI.isArrayAllocation()) {
226 AI.setOperand(0, ConstantInt::get(AI.getArraySize()->getType(), 1));
230 // Get the first instruction in the entry block.
231 BasicBlock &EntryBlock = AI.getParent()->getParent()->getEntryBlock();
232 Instruction *FirstInst = EntryBlock.getFirstNonPHIOrDbg();
233 if (FirstInst != &AI) {
234 // If the entry block doesn't start with a zero-size alloca then move
235 // this one to the start of the entry block. There is no problem with
236 // dominance as the array size was forced to a constant earlier already.
237 AllocaInst *EntryAI = dyn_cast<AllocaInst>(FirstInst);
238 if (!EntryAI || !EntryAI->getAllocatedType()->isSized() ||
239 DL->getTypeAllocSize(EntryAI->getAllocatedType()) != 0) {
240 AI.moveBefore(FirstInst);
244 // If the alignment of the entry block alloca is 0 (unspecified),
245 // assign it the preferred alignment.
246 if (EntryAI->getAlignment() == 0)
247 EntryAI->setAlignment(
248 DL->getPrefTypeAlignment(EntryAI->getAllocatedType()));
249 // Replace this zero-sized alloca with the one at the start of the entry
250 // block after ensuring that the address will be aligned enough for both
252 unsigned MaxAlign = std::max(EntryAI->getAlignment(),
254 EntryAI->setAlignment(MaxAlign);
255 if (AI.getType() != EntryAI->getType())
256 return new BitCastInst(EntryAI, AI.getType());
257 return ReplaceInstUsesWith(AI, EntryAI);
262 if (AI.getAlignment()) {
263 // Check to see if this allocation is only modified by a memcpy/memmove from
264 // a constant global whose alignment is equal to or exceeds that of the
265 // allocation. If this is the case, we can change all users to use
266 // the constant global instead. This is commonly produced by the CFE by
267 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
268 // is only subsequently read.
269 SmallVector<Instruction *, 4> ToDelete;
270 if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(&AI, ToDelete)) {
271 unsigned SourceAlign = getOrEnforceKnownAlignment(
272 Copy->getSource(), AI.getAlignment(), DL, AC, &AI, DT);
273 if (AI.getAlignment() <= SourceAlign) {
274 DEBUG(dbgs() << "Found alloca equal to global: " << AI << '\n');
275 DEBUG(dbgs() << " memcpy = " << *Copy << '\n');
276 for (unsigned i = 0, e = ToDelete.size(); i != e; ++i)
277 EraseInstFromFunction(*ToDelete[i]);
278 Constant *TheSrc = cast<Constant>(Copy->getSource());
280 = ConstantExpr::getPointerBitCastOrAddrSpaceCast(TheSrc, AI.getType());
281 Instruction *NewI = ReplaceInstUsesWith(AI, Cast);
282 EraseInstFromFunction(*Copy);
289 // At last, use the generic allocation site handler to aggressively remove
291 return visitAllocSite(AI);
294 /// \brief Helper to combine a load to a new type.
296 /// This just does the work of combining a load to a new type. It handles
297 /// metadata, etc., and returns the new instruction. The \c NewTy should be the
298 /// loaded *value* type. This will convert it to a pointer, cast the operand to
299 /// that pointer type, load it, etc.
301 /// Note that this will create all of the instructions with whatever insert
302 /// point the \c InstCombiner currently is using.
303 static LoadInst *combineLoadToNewType(InstCombiner &IC, LoadInst &LI, Type *NewTy) {
304 Value *Ptr = LI.getPointerOperand();
305 unsigned AS = LI.getPointerAddressSpace();
306 SmallVector<std::pair<unsigned, MDNode *>, 8> MD;
307 LI.getAllMetadata(MD);
309 LoadInst *NewLoad = IC.Builder->CreateAlignedLoad(
310 IC.Builder->CreateBitCast(Ptr, NewTy->getPointerTo(AS)),
311 LI.getAlignment(), LI.getName());
312 for (const auto &MDPair : MD) {
313 unsigned ID = MDPair.first;
314 MDNode *N = MDPair.second;
315 // Note, essentially every kind of metadata should be preserved here! This
316 // routine is supposed to clone a load instruction changing *only its type*.
317 // The only metadata it makes sense to drop is metadata which is invalidated
318 // when the pointer type changes. This should essentially never be the case
319 // in LLVM, but we explicitly switch over only known metadata to be
320 // conservatively correct. If you are adding metadata to LLVM which pertains
321 // to loads, you almost certainly want to add it here.
323 case LLVMContext::MD_dbg:
324 case LLVMContext::MD_tbaa:
325 case LLVMContext::MD_prof:
326 case LLVMContext::MD_fpmath:
327 case LLVMContext::MD_tbaa_struct:
328 case LLVMContext::MD_invariant_load:
329 case LLVMContext::MD_alias_scope:
330 case LLVMContext::MD_noalias:
331 case LLVMContext::MD_nontemporal:
332 case LLVMContext::MD_mem_parallel_loop_access:
333 // All of these directly apply.
334 NewLoad->setMetadata(ID, N);
337 case LLVMContext::MD_nonnull:
338 // FIXME: We should translate this into range metadata for integer types
340 if (NewTy->isPointerTy())
341 NewLoad->setMetadata(ID, N);
344 case LLVMContext::MD_range:
345 // FIXME: It would be nice to propagate this in some way, but the type
346 // conversions make it hard.
353 /// \brief Combine loads to match the type of value their uses after looking
354 /// through intervening bitcasts.
356 /// The core idea here is that if the result of a load is used in an operation,
357 /// we should load the type most conducive to that operation. For example, when
358 /// loading an integer and converting that immediately to a pointer, we should
359 /// instead directly load a pointer.
361 /// However, this routine must never change the width of a load or the number of
362 /// loads as that would introduce a semantic change. This combine is expected to
363 /// be a semantic no-op which just allows loads to more closely model the types
364 /// of their consuming operations.
366 /// Currently, we also refuse to change the precise type used for an atomic load
367 /// or a volatile load. This is debatable, and might be reasonable to change
368 /// later. However, it is risky in case some backend or other part of LLVM is
369 /// relying on the exact type loaded to select appropriate atomic operations.
370 static Instruction *combineLoadToOperationType(InstCombiner &IC, LoadInst &LI) {
371 // FIXME: We could probably with some care handle both volatile and atomic
372 // loads here but it isn't clear that this is important.
380 // Fold away bit casts of the loaded value by loading the desired type.
382 if (auto *BC = dyn_cast<BitCastInst>(LI.user_back())) {
383 LoadInst *NewLoad = combineLoadToNewType(IC, LI, BC->getDestTy());
384 BC->replaceAllUsesWith(NewLoad);
385 IC.EraseInstFromFunction(*BC);
389 // FIXME: We should also canonicalize loads of vectors when their elements are
390 // cast to other types.
394 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
395 Value *Op = LI.getOperand(0);
397 // Try to canonicalize the loaded type.
398 if (Instruction *Res = combineLoadToOperationType(*this, LI))
401 // Attempt to improve the alignment.
403 unsigned KnownAlign = getOrEnforceKnownAlignment(
404 Op, DL->getPrefTypeAlignment(LI.getType()), DL, AC, &LI, DT);
405 unsigned LoadAlign = LI.getAlignment();
406 unsigned EffectiveLoadAlign = LoadAlign != 0 ? LoadAlign :
407 DL->getABITypeAlignment(LI.getType());
409 if (KnownAlign > EffectiveLoadAlign)
410 LI.setAlignment(KnownAlign);
411 else if (LoadAlign == 0)
412 LI.setAlignment(EffectiveLoadAlign);
415 // None of the following transforms are legal for volatile/atomic loads.
416 // FIXME: Some of it is okay for atomic loads; needs refactoring.
417 if (!LI.isSimple()) return nullptr;
419 // Do really simple store-to-load forwarding and load CSE, to catch cases
420 // where there are several consecutive memory accesses to the same location,
421 // separated by a few arithmetic operations.
422 BasicBlock::iterator BBI = &LI;
423 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
424 return ReplaceInstUsesWith(
425 LI, Builder->CreateBitOrPointerCast(AvailableVal, LI.getType(),
426 LI.getName() + ".cast"));
428 // load(gep null, ...) -> unreachable
429 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
430 const Value *GEPI0 = GEPI->getOperand(0);
431 // TODO: Consider a target hook for valid address spaces for this xform.
432 if (isa<ConstantPointerNull>(GEPI0) && GEPI->getPointerAddressSpace() == 0){
433 // Insert a new store to null instruction before the load to indicate
434 // that this code is not reachable. We do this instead of inserting
435 // an unreachable instruction directly because we cannot modify the
437 new StoreInst(UndefValue::get(LI.getType()),
438 Constant::getNullValue(Op->getType()), &LI);
439 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
443 // load null/undef -> unreachable
444 // TODO: Consider a target hook for valid address spaces for this xform.
445 if (isa<UndefValue>(Op) ||
446 (isa<ConstantPointerNull>(Op) && LI.getPointerAddressSpace() == 0)) {
447 // Insert a new store to null instruction before the load to indicate that
448 // this code is not reachable. We do this instead of inserting an
449 // unreachable instruction directly because we cannot modify the CFG.
450 new StoreInst(UndefValue::get(LI.getType()),
451 Constant::getNullValue(Op->getType()), &LI);
452 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
455 if (Op->hasOneUse()) {
456 // Change select and PHI nodes to select values instead of addresses: this
457 // helps alias analysis out a lot, allows many others simplifications, and
458 // exposes redundancy in the code.
460 // Note that we cannot do the transformation unless we know that the
461 // introduced loads cannot trap! Something like this is valid as long as
462 // the condition is always false: load (select bool %C, int* null, int* %G),
463 // but it would not be valid if we transformed it to load from null
466 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
467 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
468 unsigned Align = LI.getAlignment();
469 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI, Align, DL) &&
470 isSafeToLoadUnconditionally(SI->getOperand(2), SI, Align, DL)) {
471 LoadInst *V1 = Builder->CreateLoad(SI->getOperand(1),
472 SI->getOperand(1)->getName()+".val");
473 LoadInst *V2 = Builder->CreateLoad(SI->getOperand(2),
474 SI->getOperand(2)->getName()+".val");
475 V1->setAlignment(Align);
476 V2->setAlignment(Align);
477 return SelectInst::Create(SI->getCondition(), V1, V2);
480 // load (select (cond, null, P)) -> load P
481 if (isa<ConstantPointerNull>(SI->getOperand(1)) &&
482 LI.getPointerAddressSpace() == 0) {
483 LI.setOperand(0, SI->getOperand(2));
487 // load (select (cond, P, null)) -> load P
488 if (isa<ConstantPointerNull>(SI->getOperand(2)) &&
489 LI.getPointerAddressSpace() == 0) {
490 LI.setOperand(0, SI->getOperand(1));
498 /// \brief Combine stores to match the type of value being stored.
500 /// The core idea here is that the memory does not have any intrinsic type and
501 /// where we can we should match the type of a store to the type of value being
504 /// However, this routine must never change the width of a store or the number of
505 /// stores as that would introduce a semantic change. This combine is expected to
506 /// be a semantic no-op which just allows stores to more closely model the types
507 /// of their incoming values.
509 /// Currently, we also refuse to change the precise type used for an atomic or
510 /// volatile store. This is debatable, and might be reasonable to change later.
511 /// However, it is risky in case some backend or other part of LLVM is relying
512 /// on the exact type stored to select appropriate atomic operations.
514 /// \returns true if the store was successfully combined away. This indicates
515 /// the caller must erase the store instruction. We have to let the caller erase
516 /// the store instruction sas otherwise there is no way to signal whether it was
517 /// combined or not: IC.EraseInstFromFunction returns a null pointer.
518 static bool combineStoreToValueType(InstCombiner &IC, StoreInst &SI) {
519 // FIXME: We could probably with some care handle both volatile and atomic
520 // stores here but it isn't clear that this is important.
524 Value *Ptr = SI.getPointerOperand();
525 Value *V = SI.getValueOperand();
526 unsigned AS = SI.getPointerAddressSpace();
527 SmallVector<std::pair<unsigned, MDNode *>, 8> MD;
528 SI.getAllMetadata(MD);
530 // Fold away bit casts of the stored value by storing the original type.
531 if (auto *BC = dyn_cast<BitCastInst>(V)) {
532 V = BC->getOperand(0);
533 StoreInst *NewStore = IC.Builder->CreateAlignedStore(
534 V, IC.Builder->CreateBitCast(Ptr, V->getType()->getPointerTo(AS)),
536 for (const auto &MDPair : MD) {
537 unsigned ID = MDPair.first;
538 MDNode *N = MDPair.second;
539 // Note, essentially every kind of metadata should be preserved here! This
540 // routine is supposed to clone a store instruction changing *only its
541 // type*. The only metadata it makes sense to drop is metadata which is
542 // invalidated when the pointer type changes. This should essentially
543 // never be the case in LLVM, but we explicitly switch over only known
544 // metadata to be conservatively correct. If you are adding metadata to
545 // LLVM which pertains to stores, you almost certainly want to add it
548 case LLVMContext::MD_dbg:
549 case LLVMContext::MD_tbaa:
550 case LLVMContext::MD_prof:
551 case LLVMContext::MD_fpmath:
552 case LLVMContext::MD_tbaa_struct:
553 case LLVMContext::MD_alias_scope:
554 case LLVMContext::MD_noalias:
555 case LLVMContext::MD_nontemporal:
556 case LLVMContext::MD_mem_parallel_loop_access:
557 // All of these directly apply.
558 NewStore->setMetadata(ID, N);
561 case LLVMContext::MD_invariant_load:
562 case LLVMContext::MD_nonnull:
563 case LLVMContext::MD_range:
564 // These don't apply for stores.
571 // FIXME: We should also canonicalize loads of vectors when their elements are
572 // cast to other types.
576 /// equivalentAddressValues - Test if A and B will obviously have the same
577 /// value. This includes recognizing that %t0 and %t1 will have the same
578 /// value in code like this:
579 /// %t0 = getelementptr \@a, 0, 3
580 /// store i32 0, i32* %t0
581 /// %t1 = getelementptr \@a, 0, 3
582 /// %t2 = load i32* %t1
584 static bool equivalentAddressValues(Value *A, Value *B) {
585 // Test if the values are trivially equivalent.
586 if (A == B) return true;
588 // Test if the values come form identical arithmetic instructions.
589 // This uses isIdenticalToWhenDefined instead of isIdenticalTo because
590 // its only used to compare two uses within the same basic block, which
591 // means that they'll always either have the same value or one of them
592 // will have an undefined value.
593 if (isa<BinaryOperator>(A) ||
596 isa<GetElementPtrInst>(A))
597 if (Instruction *BI = dyn_cast<Instruction>(B))
598 if (cast<Instruction>(A)->isIdenticalToWhenDefined(BI))
601 // Otherwise they may not be equivalent.
605 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
606 Value *Val = SI.getOperand(0);
607 Value *Ptr = SI.getOperand(1);
609 // Try to canonicalize the stored type.
610 if (combineStoreToValueType(*this, SI))
611 return EraseInstFromFunction(SI);
613 // Attempt to improve the alignment.
615 unsigned KnownAlign = getOrEnforceKnownAlignment(
616 Ptr, DL->getPrefTypeAlignment(Val->getType()), DL, AC, &SI, DT);
617 unsigned StoreAlign = SI.getAlignment();
618 unsigned EffectiveStoreAlign = StoreAlign != 0 ? StoreAlign :
619 DL->getABITypeAlignment(Val->getType());
621 if (KnownAlign > EffectiveStoreAlign)
622 SI.setAlignment(KnownAlign);
623 else if (StoreAlign == 0)
624 SI.setAlignment(EffectiveStoreAlign);
627 // Don't hack volatile/atomic stores.
628 // FIXME: Some bits are legal for atomic stores; needs refactoring.
629 if (!SI.isSimple()) return nullptr;
631 // If the RHS is an alloca with a single use, zapify the store, making the
633 if (Ptr->hasOneUse()) {
634 if (isa<AllocaInst>(Ptr))
635 return EraseInstFromFunction(SI);
636 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
637 if (isa<AllocaInst>(GEP->getOperand(0))) {
638 if (GEP->getOperand(0)->hasOneUse())
639 return EraseInstFromFunction(SI);
644 // Do really simple DSE, to catch cases where there are several consecutive
645 // stores to the same location, separated by a few arithmetic operations. This
646 // situation often occurs with bitfield accesses.
647 BasicBlock::iterator BBI = &SI;
648 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
651 // Don't count debug info directives, lest they affect codegen,
652 // and we skip pointer-to-pointer bitcasts, which are NOPs.
653 if (isa<DbgInfoIntrinsic>(BBI) ||
654 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy())) {
659 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
660 // Prev store isn't volatile, and stores to the same location?
661 if (PrevSI->isSimple() && equivalentAddressValues(PrevSI->getOperand(1),
665 EraseInstFromFunction(*PrevSI);
671 // If this is a load, we have to stop. However, if the loaded value is from
672 // the pointer we're loading and is producing the pointer we're storing,
673 // then *this* store is dead (X = load P; store X -> P).
674 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
675 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
677 return EraseInstFromFunction(SI);
679 // Otherwise, this is a load from some other location. Stores before it
684 // Don't skip over loads or things that can modify memory.
685 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
689 // store X, null -> turns into 'unreachable' in SimplifyCFG
690 if (isa<ConstantPointerNull>(Ptr) && SI.getPointerAddressSpace() == 0) {
691 if (!isa<UndefValue>(Val)) {
692 SI.setOperand(0, UndefValue::get(Val->getType()));
693 if (Instruction *U = dyn_cast<Instruction>(Val))
694 Worklist.Add(U); // Dropped a use.
696 return nullptr; // Do not modify these!
699 // store undef, Ptr -> noop
700 if (isa<UndefValue>(Val))
701 return EraseInstFromFunction(SI);
703 // If this store is the last instruction in the basic block (possibly
704 // excepting debug info instructions), and if the block ends with an
705 // unconditional branch, try to move it to the successor block.
709 } while (isa<DbgInfoIntrinsic>(BBI) ||
710 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy()));
711 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
712 if (BI->isUnconditional())
713 if (SimplifyStoreAtEndOfBlock(SI))
714 return nullptr; // xform done!
719 /// SimplifyStoreAtEndOfBlock - Turn things like:
720 /// if () { *P = v1; } else { *P = v2 }
721 /// into a phi node with a store in the successor.
723 /// Simplify things like:
724 /// *P = v1; if () { *P = v2; }
725 /// into a phi node with a store in the successor.
727 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
728 BasicBlock *StoreBB = SI.getParent();
730 // Check to see if the successor block has exactly two incoming edges. If
731 // so, see if the other predecessor contains a store to the same location.
732 // if so, insert a PHI node (if needed) and move the stores down.
733 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
735 // Determine whether Dest has exactly two predecessors and, if so, compute
736 // the other predecessor.
737 pred_iterator PI = pred_begin(DestBB);
739 BasicBlock *OtherBB = nullptr;
744 if (++PI == pred_end(DestBB))
753 if (++PI != pred_end(DestBB))
756 // Bail out if all the relevant blocks aren't distinct (this can happen,
757 // for example, if SI is in an infinite loop)
758 if (StoreBB == DestBB || OtherBB == DestBB)
761 // Verify that the other block ends in a branch and is not otherwise empty.
762 BasicBlock::iterator BBI = OtherBB->getTerminator();
763 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
764 if (!OtherBr || BBI == OtherBB->begin())
767 // If the other block ends in an unconditional branch, check for the 'if then
768 // else' case. there is an instruction before the branch.
769 StoreInst *OtherStore = nullptr;
770 if (OtherBr->isUnconditional()) {
772 // Skip over debugging info.
773 while (isa<DbgInfoIntrinsic>(BBI) ||
774 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy())) {
775 if (BBI==OtherBB->begin())
779 // If this isn't a store, isn't a store to the same location, or is not the
780 // right kind of store, bail out.
781 OtherStore = dyn_cast<StoreInst>(BBI);
782 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1) ||
783 !SI.isSameOperationAs(OtherStore))
786 // Otherwise, the other block ended with a conditional branch. If one of the
787 // destinations is StoreBB, then we have the if/then case.
788 if (OtherBr->getSuccessor(0) != StoreBB &&
789 OtherBr->getSuccessor(1) != StoreBB)
792 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
793 // if/then triangle. See if there is a store to the same ptr as SI that
796 // Check to see if we find the matching store.
797 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
798 if (OtherStore->getOperand(1) != SI.getOperand(1) ||
799 !SI.isSameOperationAs(OtherStore))
803 // If we find something that may be using or overwriting the stored
804 // value, or if we run out of instructions, we can't do the xform.
805 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
806 BBI == OtherBB->begin())
810 // In order to eliminate the store in OtherBr, we have to
811 // make sure nothing reads or overwrites the stored value in
813 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
814 // FIXME: This should really be AA driven.
815 if (I->mayReadFromMemory() || I->mayWriteToMemory())
820 // Insert a PHI node now if we need it.
821 Value *MergedVal = OtherStore->getOperand(0);
822 if (MergedVal != SI.getOperand(0)) {
823 PHINode *PN = PHINode::Create(MergedVal->getType(), 2, "storemerge");
824 PN->addIncoming(SI.getOperand(0), SI.getParent());
825 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
826 MergedVal = InsertNewInstBefore(PN, DestBB->front());
829 // Advance to a place where it is safe to insert the new store and
831 BBI = DestBB->getFirstInsertionPt();
832 StoreInst *NewSI = new StoreInst(MergedVal, SI.getOperand(1),
837 InsertNewInstBefore(NewSI, *BBI);
838 NewSI->setDebugLoc(OtherStore->getDebugLoc());
840 // If the two stores had AA tags, merge them.
842 SI.getAAMetadata(AATags);
844 OtherStore->getAAMetadata(AATags, /* Merge = */ true);
845 NewSI->setAAMetadata(AATags);
848 // Nuke the old stores.
849 EraseInstFromFunction(SI);
850 EraseInstFromFunction(*OtherStore);