1 //===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===//
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 transformation implements the well known scalar replacement of
11 // aggregates transformation. This xform breaks up alloca instructions of
12 // aggregate type (structure or array) into individual alloca instructions for
13 // each member (if possible). Then, if possible, it transforms the individual
14 // alloca instructions into nice clean scalar SSA form.
16 // This combines a simple SRoA algorithm with the Mem2Reg algorithm because
17 // often interact, especially for C++ programs. As such, iterating between
18 // SRoA, then Mem2Reg until we run out of things to promote works well.
20 //===----------------------------------------------------------------------===//
22 #define DEBUG_TYPE "scalarrepl"
23 #include "llvm/Transforms/Scalar.h"
24 #include "llvm/Constants.h"
25 #include "llvm/DerivedTypes.h"
26 #include "llvm/Function.h"
27 #include "llvm/GlobalVariable.h"
28 #include "llvm/Instructions.h"
29 #include "llvm/IntrinsicInst.h"
30 #include "llvm/LLVMContext.h"
31 #include "llvm/Module.h"
32 #include "llvm/Pass.h"
33 #include "llvm/Analysis/DIBuilder.h"
34 #include "llvm/Analysis/Dominators.h"
35 #include "llvm/Analysis/Loads.h"
36 #include "llvm/Analysis/ValueTracking.h"
37 #include "llvm/Target/TargetData.h"
38 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
39 #include "llvm/Transforms/Utils/Local.h"
40 #include "llvm/Transforms/Utils/SSAUpdater.h"
41 #include "llvm/Support/CallSite.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/ErrorHandling.h"
44 #include "llvm/Support/GetElementPtrTypeIterator.h"
45 #include "llvm/Support/IRBuilder.h"
46 #include "llvm/Support/MathExtras.h"
47 #include "llvm/Support/raw_ostream.h"
48 #include "llvm/ADT/SetVector.h"
49 #include "llvm/ADT/SmallVector.h"
50 #include "llvm/ADT/Statistic.h"
53 STATISTIC(NumReplaced, "Number of allocas broken up");
54 STATISTIC(NumPromoted, "Number of allocas promoted");
55 STATISTIC(NumAdjusted, "Number of scalar allocas adjusted to allow promotion");
56 STATISTIC(NumConverted, "Number of aggregates converted to scalar");
57 STATISTIC(NumGlobals, "Number of allocas copied from constant global");
60 struct SROA : public FunctionPass {
61 SROA(int T, bool hasDT, char &ID)
62 : FunctionPass(ID), HasDomTree(hasDT) {
69 bool runOnFunction(Function &F);
71 bool performScalarRepl(Function &F);
72 bool performPromotion(Function &F);
78 /// DeadInsts - Keep track of instructions we have made dead, so that
79 /// we can remove them after we are done working.
80 SmallVector<Value*, 32> DeadInsts;
82 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures
83 /// information about the uses. All these fields are initialized to false
84 /// and set to true when something is learned.
86 /// The alloca to promote.
89 /// CheckedPHIs - This is a set of verified PHI nodes, to prevent infinite
90 /// looping and avoid redundant work.
91 SmallPtrSet<PHINode*, 8> CheckedPHIs;
93 /// isUnsafe - This is set to true if the alloca cannot be SROA'd.
96 /// isMemCpySrc - This is true if this aggregate is memcpy'd from.
99 /// isMemCpyDst - This is true if this aggregate is memcpy'd into.
100 bool isMemCpyDst : 1;
102 /// hasSubelementAccess - This is true if a subelement of the alloca is
103 /// ever accessed, or false if the alloca is only accessed with mem
104 /// intrinsics or load/store that only access the entire alloca at once.
105 bool hasSubelementAccess : 1;
107 /// hasALoadOrStore - This is true if there are any loads or stores to it.
108 /// The alloca may just be accessed with memcpy, for example, which would
110 bool hasALoadOrStore : 1;
112 explicit AllocaInfo(AllocaInst *ai)
113 : AI(ai), isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false),
114 hasSubelementAccess(false), hasALoadOrStore(false) {}
117 unsigned SRThreshold;
119 void MarkUnsafe(AllocaInfo &I, Instruction *User) {
121 DEBUG(dbgs() << " Transformation preventing inst: " << *User << '\n');
124 bool isSafeAllocaToScalarRepl(AllocaInst *AI);
126 void isSafeForScalarRepl(Instruction *I, uint64_t Offset, AllocaInfo &Info);
127 void isSafePHISelectUseForScalarRepl(Instruction *User, uint64_t Offset,
129 void isSafeGEP(GetElementPtrInst *GEPI, uint64_t &Offset, AllocaInfo &Info);
130 void isSafeMemAccess(uint64_t Offset, uint64_t MemSize,
131 const Type *MemOpType, bool isStore, AllocaInfo &Info,
132 Instruction *TheAccess, bool AllowWholeAccess);
133 bool TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size);
134 uint64_t FindElementAndOffset(const Type *&T, uint64_t &Offset,
137 void DoScalarReplacement(AllocaInst *AI,
138 std::vector<AllocaInst*> &WorkList);
139 void DeleteDeadInstructions();
141 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
142 SmallVector<AllocaInst*, 32> &NewElts);
143 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
144 SmallVector<AllocaInst*, 32> &NewElts);
145 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
146 SmallVector<AllocaInst*, 32> &NewElts);
147 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
149 SmallVector<AllocaInst*, 32> &NewElts);
150 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
151 SmallVector<AllocaInst*, 32> &NewElts);
152 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
153 SmallVector<AllocaInst*, 32> &NewElts);
155 static MemTransferInst *isOnlyCopiedFromConstantGlobal(AllocaInst *AI);
158 // SROA_DT - SROA that uses DominatorTree.
159 struct SROA_DT : public SROA {
162 SROA_DT(int T = -1) : SROA(T, true, ID) {
163 initializeSROA_DTPass(*PassRegistry::getPassRegistry());
166 // getAnalysisUsage - This pass does not require any passes, but we know it
167 // will not alter the CFG, so say so.
168 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
169 AU.addRequired<DominatorTree>();
170 AU.setPreservesCFG();
174 // SROA_SSAUp - SROA that uses SSAUpdater.
175 struct SROA_SSAUp : public SROA {
178 SROA_SSAUp(int T = -1) : SROA(T, false, ID) {
179 initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry());
182 // getAnalysisUsage - This pass does not require any passes, but we know it
183 // will not alter the CFG, so say so.
184 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
185 AU.setPreservesCFG();
191 char SROA_DT::ID = 0;
192 char SROA_SSAUp::ID = 0;
194 INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl",
195 "Scalar Replacement of Aggregates (DT)", false, false)
196 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
197 INITIALIZE_PASS_END(SROA_DT, "scalarrepl",
198 "Scalar Replacement of Aggregates (DT)", false, false)
200 INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa",
201 "Scalar Replacement of Aggregates (SSAUp)", false, false)
202 INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa",
203 "Scalar Replacement of Aggregates (SSAUp)", false, false)
205 // Public interface to the ScalarReplAggregates pass
206 FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold,
209 return new SROA_DT(Threshold);
210 return new SROA_SSAUp(Threshold);
214 //===----------------------------------------------------------------------===//
215 // Convert To Scalar Optimization.
216 //===----------------------------------------------------------------------===//
219 /// ConvertToScalarInfo - This class implements the "Convert To Scalar"
220 /// optimization, which scans the uses of an alloca and determines if it can
221 /// rewrite it in terms of a single new alloca that can be mem2reg'd.
222 class ConvertToScalarInfo {
223 /// AllocaSize - The size of the alloca being considered in bytes.
225 const TargetData &TD;
227 /// IsNotTrivial - This is set to true if there is some access to the object
228 /// which means that mem2reg can't promote it.
231 /// VectorTy - This tracks the type that we should promote the vector to if
232 /// it is possible to turn it into a vector. This starts out null, and if it
233 /// isn't possible to turn into a vector type, it gets set to VoidTy.
234 const Type *VectorTy;
236 /// HadAVector - True if there is at least one vector access to the alloca.
237 /// We don't want to turn random arrays into vectors and use vector element
238 /// insert/extract, but if there are element accesses to something that is
239 /// also declared as a vector, we do want to promote to a vector.
242 /// HadNonMemTransferAccess - True if there is at least one access to the
243 /// alloca that is not a MemTransferInst. We don't want to turn structs into
244 /// large integers unless there is some potential for optimization.
245 bool HadNonMemTransferAccess;
248 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td)
249 : AllocaSize(Size), TD(td), IsNotTrivial(false), VectorTy(0),
250 HadAVector(false), HadNonMemTransferAccess(false) { }
252 AllocaInst *TryConvert(AllocaInst *AI);
255 bool CanConvertToScalar(Value *V, uint64_t Offset);
256 void MergeInType(const Type *In, uint64_t Offset, bool IsLoadOrStore);
257 bool MergeInVectorType(const VectorType *VInTy, uint64_t Offset);
258 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset);
260 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType,
261 uint64_t Offset, IRBuilder<> &Builder);
262 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal,
263 uint64_t Offset, IRBuilder<> &Builder);
265 } // end anonymous namespace.
268 /// TryConvert - Analyze the specified alloca, and if it is safe to do so,
269 /// rewrite it to be a new alloca which is mem2reg'able. This returns the new
270 /// alloca if possible or null if not.
271 AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) {
272 // If we can't convert this scalar, or if mem2reg can trivially do it, bail
274 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial)
277 // If we were able to find a vector type that can handle this with
278 // insert/extract elements, and if there was at least one use that had
279 // a vector type, promote this to a vector. We don't want to promote
280 // random stuff that doesn't use vectors (e.g. <9 x double>) because then
281 // we just get a lot of insert/extracts. If at least one vector is
282 // involved, then we probably really do have a union of vector/array.
284 if (VectorTy && VectorTy->isVectorTy() && HadAVector) {
285 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = "
286 << *VectorTy << '\n');
287 NewTy = VectorTy; // Use the vector type.
289 unsigned BitWidth = AllocaSize * 8;
290 if (!HadAVector && !HadNonMemTransferAccess &&
291 !TD.fitsInLegalInteger(BitWidth))
294 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n");
295 // Create and insert the integer alloca.
296 NewTy = IntegerType::get(AI->getContext(), BitWidth);
298 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin());
299 ConvertUsesToScalar(AI, NewAI, 0);
303 /// MergeInType - Add the 'In' type to the accumulated vector type (VectorTy)
304 /// so far at the offset specified by Offset (which is specified in bytes).
306 /// There are three cases we handle here:
307 /// 1) A union of vector types of the same size and potentially its elements.
308 /// Here we turn element accesses into insert/extract element operations.
309 /// This promotes a <4 x float> with a store of float to the third element
310 /// into a <4 x float> that uses insert element.
311 /// 2) A union of vector types with power-of-2 size differences, e.g. a float,
312 /// <2 x float> and <4 x float>. Here we turn element accesses into insert
313 /// and extract element operations, and <2 x float> accesses into a cast to
314 /// <2 x double>, an extract, and a cast back to <2 x float>.
315 /// 3) A fully general blob of memory, which we turn into some (potentially
316 /// large) integer type with extract and insert operations where the loads
317 /// and stores would mutate the memory. We mark this by setting VectorTy
319 void ConvertToScalarInfo::MergeInType(const Type *In, uint64_t Offset,
320 bool IsLoadOrStore) {
321 // If we already decided to turn this into a blob of integer memory, there is
322 // nothing to be done.
323 if (VectorTy && VectorTy->isVoidTy())
326 // If this could be contributing to a vector, analyze it.
328 // If the In type is a vector that is the same size as the alloca, see if it
329 // matches the existing VecTy.
330 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) {
331 if (MergeInVectorType(VInTy, Offset))
333 } else if (In->isFloatTy() || In->isDoubleTy() ||
334 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 &&
335 isPowerOf2_32(In->getPrimitiveSizeInBits()))) {
336 // Full width accesses can be ignored, because they can always be turned
338 unsigned EltSize = In->getPrimitiveSizeInBits()/8;
339 if (IsLoadOrStore && EltSize == AllocaSize)
342 // If we're accessing something that could be an element of a vector, see
343 // if the implied vector agrees with what we already have and if Offset is
344 // compatible with it.
345 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 &&
346 (!VectorTy || Offset * 8 < VectorTy->getPrimitiveSizeInBits())) {
348 VectorTy = VectorType::get(In, AllocaSize/EltSize);
352 unsigned CurrentEltSize = cast<VectorType>(VectorTy)->getElementType()
353 ->getPrimitiveSizeInBits()/8;
354 if (EltSize == CurrentEltSize)
357 if (In->isIntegerTy() && isPowerOf2_32(AllocaSize / EltSize))
362 // Otherwise, we have a case that we can't handle with an optimized vector
363 // form. We can still turn this into a large integer.
364 VectorTy = Type::getVoidTy(In->getContext());
367 /// MergeInVectorType - Handles the vector case of MergeInType, returning true
368 /// if the type was successfully merged and false otherwise.
369 bool ConvertToScalarInfo::MergeInVectorType(const VectorType *VInTy,
371 // Remember if we saw a vector type.
374 // TODO: Support nonzero offsets?
378 // Only allow vectors that are a power-of-2 away from the size of the alloca.
379 if (!isPowerOf2_64(AllocaSize / (VInTy->getBitWidth() / 8)))
382 // If this the first vector we see, remember the type so that we know the
389 unsigned BitWidth = cast<VectorType>(VectorTy)->getBitWidth();
390 unsigned InBitWidth = VInTy->getBitWidth();
392 // Vectors of the same size can be converted using a simple bitcast.
393 if (InBitWidth == BitWidth && AllocaSize == (InBitWidth / 8))
396 const Type *ElementTy = cast<VectorType>(VectorTy)->getElementType();
397 const Type *InElementTy = cast<VectorType>(VInTy)->getElementType();
399 // Do not allow mixed integer and floating-point accesses from vectors of
401 if (ElementTy->isFloatingPointTy() != InElementTy->isFloatingPointTy())
404 if (ElementTy->isFloatingPointTy()) {
405 // Only allow floating-point vectors of different sizes if they have the
406 // same element type.
407 // TODO: This could be loosened a bit, but would anything benefit?
408 if (ElementTy != InElementTy)
411 // There are no arbitrary-precision floating-point types, which limits the
412 // number of legal vector types with larger element types that we can form
413 // to bitcast and extract a subvector.
414 // TODO: We could support some more cases with mixed fp128 and double here.
415 if (!(BitWidth == 64 || BitWidth == 128) ||
416 !(InBitWidth == 64 || InBitWidth == 128))
419 assert(ElementTy->isIntegerTy() && "Vector elements must be either integer "
420 "or floating-point.");
421 unsigned BitWidth = ElementTy->getPrimitiveSizeInBits();
422 unsigned InBitWidth = InElementTy->getPrimitiveSizeInBits();
424 // Do not allow integer types smaller than a byte or types whose widths are
425 // not a multiple of a byte.
426 if (BitWidth < 8 || InBitWidth < 8 ||
427 BitWidth % 8 != 0 || InBitWidth % 8 != 0)
431 // Pick the largest of the two vector types.
432 if (InBitWidth > BitWidth)
438 /// CanConvertToScalar - V is a pointer. If we can convert the pointee and all
439 /// its accesses to a single vector type, return true and set VecTy to
440 /// the new type. If we could convert the alloca into a single promotable
441 /// integer, return true but set VecTy to VoidTy. Further, if the use is not a
442 /// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset
443 /// is the current offset from the base of the alloca being analyzed.
445 /// If we see at least one access to the value that is as a vector type, set the
447 bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) {
448 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
449 Instruction *User = cast<Instruction>(*UI);
451 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
452 // Don't break volatile loads.
453 if (LI->isVolatile())
455 // Don't touch MMX operations.
456 if (LI->getType()->isX86_MMXTy())
458 HadNonMemTransferAccess = true;
459 MergeInType(LI->getType(), Offset, true);
463 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
464 // Storing the pointer, not into the value?
465 if (SI->getOperand(0) == V || SI->isVolatile()) return false;
466 // Don't touch MMX operations.
467 if (SI->getOperand(0)->getType()->isX86_MMXTy())
469 HadNonMemTransferAccess = true;
470 MergeInType(SI->getOperand(0)->getType(), Offset, true);
474 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
475 IsNotTrivial = true; // Can't be mem2reg'd.
476 if (!CanConvertToScalar(BCI, Offset))
481 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
482 // If this is a GEP with a variable indices, we can't handle it.
483 if (!GEP->hasAllConstantIndices())
486 // Compute the offset that this GEP adds to the pointer.
487 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
488 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(),
489 &Indices[0], Indices.size());
490 // See if all uses can be converted.
491 if (!CanConvertToScalar(GEP, Offset+GEPOffset))
493 IsNotTrivial = true; // Can't be mem2reg'd.
494 HadNonMemTransferAccess = true;
498 // If this is a constant sized memset of a constant value (e.g. 0) we can
500 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
501 // Store of constant value and constant size.
502 if (!isa<ConstantInt>(MSI->getValue()) ||
503 !isa<ConstantInt>(MSI->getLength()))
505 IsNotTrivial = true; // Can't be mem2reg'd.
506 HadNonMemTransferAccess = true;
510 // If this is a memcpy or memmove into or out of the whole allocation, we
511 // can handle it like a load or store of the scalar type.
512 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
513 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength());
514 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0)
517 IsNotTrivial = true; // Can't be mem2reg'd.
521 // Otherwise, we cannot handle this!
528 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
529 /// directly. This happens when we are converting an "integer union" to a
530 /// single integer scalar, or when we are converting a "vector union" to a
531 /// vector with insert/extractelement instructions.
533 /// Offset is an offset from the original alloca, in bits that need to be
534 /// shifted to the right. By the end of this, there should be no uses of Ptr.
535 void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI,
537 while (!Ptr->use_empty()) {
538 Instruction *User = cast<Instruction>(Ptr->use_back());
540 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
541 ConvertUsesToScalar(CI, NewAI, Offset);
542 CI->eraseFromParent();
546 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
547 // Compute the offset that this GEP adds to the pointer.
548 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
549 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(),
550 &Indices[0], Indices.size());
551 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8);
552 GEP->eraseFromParent();
556 IRBuilder<> Builder(User);
558 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
559 // The load is a bit extract from NewAI shifted right by Offset bits.
560 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp");
562 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder);
563 LI->replaceAllUsesWith(NewLoadVal);
564 LI->eraseFromParent();
568 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
569 assert(SI->getOperand(0) != Ptr && "Consistency error!");
570 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
571 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset,
573 Builder.CreateStore(New, NewAI);
574 SI->eraseFromParent();
576 // If the load we just inserted is now dead, then the inserted store
577 // overwrote the entire thing.
578 if (Old->use_empty())
579 Old->eraseFromParent();
583 // If this is a constant sized memset of a constant value (e.g. 0) we can
584 // transform it into a store of the expanded constant value.
585 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
586 assert(MSI->getRawDest() == Ptr && "Consistency error!");
587 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
589 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue();
591 // Compute the value replicated the right number of times.
592 APInt APVal(NumBytes*8, Val);
594 // Splat the value if non-zero.
596 for (unsigned i = 1; i != NumBytes; ++i)
599 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
600 Value *New = ConvertScalar_InsertValue(
601 ConstantInt::get(User->getContext(), APVal),
602 Old, Offset, Builder);
603 Builder.CreateStore(New, NewAI);
605 // If the load we just inserted is now dead, then the memset overwrote
607 if (Old->use_empty())
608 Old->eraseFromParent();
610 MSI->eraseFromParent();
614 // If this is a memcpy or memmove into or out of the whole allocation, we
615 // can handle it like a load or store of the scalar type.
616 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
617 assert(Offset == 0 && "must be store to start of alloca");
619 // If the source and destination are both to the same alloca, then this is
620 // a noop copy-to-self, just delete it. Otherwise, emit a load and store
622 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &TD, 0));
624 if (GetUnderlyingObject(MTI->getSource(), &TD, 0) != OrigAI) {
625 // Dest must be OrigAI, change this to be a load from the original
626 // pointer (bitcasted), then a store to our new alloca.
627 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?");
628 Value *SrcPtr = MTI->getSource();
629 const PointerType* SPTy = cast<PointerType>(SrcPtr->getType());
630 const PointerType* AIPTy = cast<PointerType>(NewAI->getType());
631 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
632 AIPTy = PointerType::get(AIPTy->getElementType(),
633 SPTy->getAddressSpace());
635 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy);
637 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval");
638 SrcVal->setAlignment(MTI->getAlignment());
639 Builder.CreateStore(SrcVal, NewAI);
640 } else if (GetUnderlyingObject(MTI->getDest(), &TD, 0) != OrigAI) {
641 // Src must be OrigAI, change this to be a load from NewAI then a store
642 // through the original dest pointer (bitcasted).
643 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?");
644 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval");
646 const PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType());
647 const PointerType* AIPTy = cast<PointerType>(NewAI->getType());
648 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
649 AIPTy = PointerType::get(AIPTy->getElementType(),
650 DPTy->getAddressSpace());
652 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy);
654 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr);
655 NewStore->setAlignment(MTI->getAlignment());
657 // Noop transfer. Src == Dst
660 MTI->eraseFromParent();
664 llvm_unreachable("Unsupported operation!");
668 /// getScaledElementType - Gets a scaled element type for a partial vector
669 /// access of an alloca. The input types must be integer or floating-point
670 /// scalar or vector types, and the resulting type is an integer, float or
672 static const Type *getScaledElementType(const Type *Ty1, const Type *Ty2,
673 unsigned NewBitWidth) {
674 bool IsFP1 = Ty1->isFloatingPointTy() ||
675 (Ty1->isVectorTy() &&
676 cast<VectorType>(Ty1)->getElementType()->isFloatingPointTy());
677 bool IsFP2 = Ty2->isFloatingPointTy() ||
678 (Ty2->isVectorTy() &&
679 cast<VectorType>(Ty2)->getElementType()->isFloatingPointTy());
681 LLVMContext &Context = Ty1->getContext();
683 // Prefer floating-point types over integer types, as integer types may have
684 // been created by earlier scalar replacement.
685 if (IsFP1 || IsFP2) {
686 if (NewBitWidth == 32)
687 return Type::getFloatTy(Context);
688 if (NewBitWidth == 64)
689 return Type::getDoubleTy(Context);
692 return Type::getIntNTy(Context, NewBitWidth);
695 /// CreateShuffleVectorCast - Creates a shuffle vector to convert one vector
696 /// to another vector of the same element type which has the same allocation
697 /// size but different primitive sizes (e.g. <3 x i32> and <4 x i32>).
698 static Value *CreateShuffleVectorCast(Value *FromVal, const Type *ToType,
699 IRBuilder<> &Builder) {
700 const Type *FromType = FromVal->getType();
701 const VectorType *FromVTy = cast<VectorType>(FromType);
702 const VectorType *ToVTy = cast<VectorType>(ToType);
703 assert((ToVTy->getElementType() == FromVTy->getElementType()) &&
704 "Vectors must have the same element type");
705 Value *UnV = UndefValue::get(FromType);
706 unsigned numEltsFrom = FromVTy->getNumElements();
707 unsigned numEltsTo = ToVTy->getNumElements();
709 SmallVector<Constant*, 3> Args;
710 const Type* Int32Ty = Builder.getInt32Ty();
711 unsigned minNumElts = std::min(numEltsFrom, numEltsTo);
713 for (i=0; i != minNumElts; ++i)
714 Args.push_back(ConstantInt::get(Int32Ty, i));
717 Constant* UnC = UndefValue::get(Int32Ty);
718 for (; i != numEltsTo; ++i)
721 Constant *Mask = ConstantVector::get(Args);
722 return Builder.CreateShuffleVector(FromVal, UnV, Mask, "tmpV");
725 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer
726 /// or vector value FromVal, extracting the bits from the offset specified by
727 /// Offset. This returns the value, which is of type ToType.
729 /// This happens when we are converting an "integer union" to a single
730 /// integer scalar, or when we are converting a "vector union" to a vector with
731 /// insert/extractelement instructions.
733 /// Offset is an offset from the original alloca, in bits that need to be
734 /// shifted to the right.
735 Value *ConvertToScalarInfo::
736 ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType,
737 uint64_t Offset, IRBuilder<> &Builder) {
738 // If the load is of the whole new alloca, no conversion is needed.
739 const Type *FromType = FromVal->getType();
740 if (FromType == ToType && Offset == 0)
743 // If the result alloca is a vector type, this is either an element
744 // access or a bitcast to another vector type of the same size.
745 if (const VectorType *VTy = dyn_cast<VectorType>(FromType)) {
746 unsigned FromTypeSize = TD.getTypeAllocSize(FromType);
747 unsigned ToTypeSize = TD.getTypeAllocSize(ToType);
748 if (FromTypeSize == ToTypeSize) {
749 // If the two types have the same primitive size, use a bit cast.
750 // Otherwise, it is two vectors with the same element type that has
751 // the same allocation size but different number of elements so use
753 if (FromType->getPrimitiveSizeInBits() ==
754 ToType->getPrimitiveSizeInBits())
755 return Builder.CreateBitCast(FromVal, ToType, "tmp");
757 return CreateShuffleVectorCast(FromVal, ToType, Builder);
760 if (isPowerOf2_64(FromTypeSize / ToTypeSize)) {
761 assert(!(ToType->isVectorTy() && Offset != 0) && "Can't extract a value "
762 "of a smaller vector type at a nonzero offset.");
764 const Type *CastElementTy = getScaledElementType(FromType, ToType,
766 unsigned NumCastVectorElements = FromTypeSize / ToTypeSize;
768 LLVMContext &Context = FromVal->getContext();
769 const Type *CastTy = VectorType::get(CastElementTy,
770 NumCastVectorElements);
771 Value *Cast = Builder.CreateBitCast(FromVal, CastTy, "tmp");
773 unsigned EltSize = TD.getTypeAllocSizeInBits(CastElementTy);
774 unsigned Elt = Offset/EltSize;
775 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
776 Value *Extract = Builder.CreateExtractElement(Cast, ConstantInt::get(
777 Type::getInt32Ty(Context), Elt), "tmp");
778 return Builder.CreateBitCast(Extract, ToType, "tmp");
781 // Otherwise it must be an element access.
784 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType());
785 Elt = Offset/EltSize;
786 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
788 // Return the element extracted out of it.
789 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get(
790 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp");
791 if (V->getType() != ToType)
792 V = Builder.CreateBitCast(V, ToType, "tmp");
796 // If ToType is a first class aggregate, extract out each of the pieces and
797 // use insertvalue's to form the FCA.
798 if (const StructType *ST = dyn_cast<StructType>(ToType)) {
799 const StructLayout &Layout = *TD.getStructLayout(ST);
800 Value *Res = UndefValue::get(ST);
801 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
802 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i),
803 Offset+Layout.getElementOffsetInBits(i),
805 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
810 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) {
811 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType());
812 Value *Res = UndefValue::get(AT);
813 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
814 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(),
815 Offset+i*EltSize, Builder);
816 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
821 // Otherwise, this must be a union that was converted to an integer value.
822 const IntegerType *NTy = cast<IntegerType>(FromVal->getType());
824 // If this is a big-endian system and the load is narrower than the
825 // full alloca type, we need to do a shift to get the right bits.
827 if (TD.isBigEndian()) {
828 // On big-endian machines, the lowest bit is stored at the bit offset
829 // from the pointer given by getTypeStoreSizeInBits. This matters for
830 // integers with a bitwidth that is not a multiple of 8.
831 ShAmt = TD.getTypeStoreSizeInBits(NTy) -
832 TD.getTypeStoreSizeInBits(ToType) - Offset;
837 // Note: we support negative bitwidths (with shl) which are not defined.
838 // We do this to support (f.e.) loads off the end of a structure where
839 // only some bits are used.
840 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth())
841 FromVal = Builder.CreateLShr(FromVal,
842 ConstantInt::get(FromVal->getType(),
844 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
845 FromVal = Builder.CreateShl(FromVal,
846 ConstantInt::get(FromVal->getType(),
849 // Finally, unconditionally truncate the integer to the right width.
850 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType);
851 if (LIBitWidth < NTy->getBitWidth())
853 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(),
855 else if (LIBitWidth > NTy->getBitWidth())
857 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(),
860 // If the result is an integer, this is a trunc or bitcast.
861 if (ToType->isIntegerTy()) {
863 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) {
864 // Just do a bitcast, we know the sizes match up.
865 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp");
867 // Otherwise must be a pointer.
868 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp");
870 assert(FromVal->getType() == ToType && "Didn't convert right?");
874 /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer
875 /// or vector value "Old" at the offset specified by Offset.
877 /// This happens when we are converting an "integer union" to a
878 /// single integer scalar, or when we are converting a "vector union" to a
879 /// vector with insert/extractelement instructions.
881 /// Offset is an offset from the original alloca, in bits that need to be
882 /// shifted to the right.
883 Value *ConvertToScalarInfo::
884 ConvertScalar_InsertValue(Value *SV, Value *Old,
885 uint64_t Offset, IRBuilder<> &Builder) {
886 // Convert the stored type to the actual type, shift it left to insert
887 // then 'or' into place.
888 const Type *AllocaType = Old->getType();
889 LLVMContext &Context = Old->getContext();
891 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
892 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy);
893 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType());
895 // Changing the whole vector with memset or with an access of a different
897 if (ValSize == VecSize) {
898 // If the two types have the same primitive size, use a bit cast.
899 // Otherwise, it is two vectors with the same element type that has
900 // the same allocation size but different number of elements so use
902 if (VTy->getPrimitiveSizeInBits() ==
903 SV->getType()->getPrimitiveSizeInBits())
904 return Builder.CreateBitCast(SV, AllocaType, "tmp");
906 return CreateShuffleVectorCast(SV, VTy, Builder);
909 if (isPowerOf2_64(VecSize / ValSize)) {
910 assert(!(SV->getType()->isVectorTy() && Offset != 0) && "Can't insert a "
911 "value of a smaller vector type at a nonzero offset.");
913 const Type *CastElementTy = getScaledElementType(VTy, SV->getType(),
915 unsigned NumCastVectorElements = VecSize / ValSize;
917 LLVMContext &Context = SV->getContext();
918 const Type *OldCastTy = VectorType::get(CastElementTy,
919 NumCastVectorElements);
920 Value *OldCast = Builder.CreateBitCast(Old, OldCastTy, "tmp");
922 Value *SVCast = Builder.CreateBitCast(SV, CastElementTy, "tmp");
924 unsigned EltSize = TD.getTypeAllocSizeInBits(CastElementTy);
925 unsigned Elt = Offset/EltSize;
926 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
928 Builder.CreateInsertElement(OldCast, SVCast, ConstantInt::get(
929 Type::getInt32Ty(Context), Elt), "tmp");
930 return Builder.CreateBitCast(Insert, AllocaType, "tmp");
933 // Must be an element insertion.
934 assert(SV->getType() == VTy->getElementType());
935 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType());
936 unsigned Elt = Offset/EltSize;
937 return Builder.CreateInsertElement(Old, SV,
938 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt),
942 // If SV is a first-class aggregate value, insert each value recursively.
943 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) {
944 const StructLayout &Layout = *TD.getStructLayout(ST);
945 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
946 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
947 Old = ConvertScalar_InsertValue(Elt, Old,
948 Offset+Layout.getElementOffsetInBits(i),
954 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) {
955 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType());
956 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
957 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
958 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder);
963 // If SV is a float, convert it to the appropriate integer type.
964 // If it is a pointer, do the same.
965 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType());
966 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType);
967 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType());
968 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType);
969 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy())
970 SV = Builder.CreateBitCast(SV,
971 IntegerType::get(SV->getContext(),SrcWidth), "tmp");
972 else if (SV->getType()->isPointerTy())
973 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp");
975 // Zero extend or truncate the value if needed.
976 if (SV->getType() != AllocaType) {
977 if (SV->getType()->getPrimitiveSizeInBits() <
978 AllocaType->getPrimitiveSizeInBits())
979 SV = Builder.CreateZExt(SV, AllocaType, "tmp");
981 // Truncation may be needed if storing more than the alloca can hold
982 // (undefined behavior).
983 SV = Builder.CreateTrunc(SV, AllocaType, "tmp");
984 SrcWidth = DestWidth;
985 SrcStoreWidth = DestStoreWidth;
989 // If this is a big-endian system and the store is narrower than the
990 // full alloca type, we need to do a shift to get the right bits.
992 if (TD.isBigEndian()) {
993 // On big-endian machines, the lowest bit is stored at the bit offset
994 // from the pointer given by getTypeStoreSizeInBits. This matters for
995 // integers with a bitwidth that is not a multiple of 8.
996 ShAmt = DestStoreWidth - SrcStoreWidth - Offset;
1001 // Note: we support negative bitwidths (with shr) which are not defined.
1002 // We do this to support (f.e.) stores off the end of a structure where
1003 // only some bits in the structure are set.
1004 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
1005 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
1006 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(),
1009 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
1010 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(),
1012 Mask = Mask.lshr(-ShAmt);
1015 // Mask out the bits we are about to insert from the old value, and or
1017 if (SrcWidth != DestWidth) {
1018 assert(DestWidth > SrcWidth);
1019 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask");
1020 SV = Builder.CreateOr(Old, SV, "ins");
1026 //===----------------------------------------------------------------------===//
1028 //===----------------------------------------------------------------------===//
1031 bool SROA::runOnFunction(Function &F) {
1032 TD = getAnalysisIfAvailable<TargetData>();
1034 bool Changed = performPromotion(F);
1036 // FIXME: ScalarRepl currently depends on TargetData more than it
1037 // theoretically needs to. It should be refactored in order to support
1038 // target-independent IR. Until this is done, just skip the actual
1039 // scalar-replacement portion of this pass.
1040 if (!TD) return Changed;
1043 bool LocalChange = performScalarRepl(F);
1044 if (!LocalChange) break; // No need to repromote if no scalarrepl
1046 LocalChange = performPromotion(F);
1047 if (!LocalChange) break; // No need to re-scalarrepl if no promotion
1054 class AllocaPromoter : public LoadAndStorePromoter {
1057 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
1058 DbgDeclareInst *DD, DIBuilder *&DB)
1059 : LoadAndStorePromoter(Insts, S, DD, DB), AI(0) {}
1061 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) {
1062 // Remember which alloca we're promoting (for isInstInList).
1064 LoadAndStorePromoter::run(Insts);
1065 AI->eraseFromParent();
1068 virtual bool isInstInList(Instruction *I,
1069 const SmallVectorImpl<Instruction*> &Insts) const {
1070 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1071 return LI->getOperand(0) == AI;
1072 return cast<StoreInst>(I)->getPointerOperand() == AI;
1075 } // end anon namespace
1077 /// isSafeSelectToSpeculate - Select instructions that use an alloca and are
1078 /// subsequently loaded can be rewritten to load both input pointers and then
1079 /// select between the result, allowing the load of the alloca to be promoted.
1081 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1082 /// %V = load i32* %P2
1084 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1085 /// %V2 = load i32* %Other
1086 /// %V = select i1 %cond, i32 %V1, i32 %V2
1088 /// We can do this to a select if its only uses are loads and if the operand to
1089 /// the select can be loaded unconditionally.
1090 static bool isSafeSelectToSpeculate(SelectInst *SI, const TargetData *TD) {
1091 bool TDerefable = SI->getTrueValue()->isDereferenceablePointer();
1092 bool FDerefable = SI->getFalseValue()->isDereferenceablePointer();
1094 for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end();
1096 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1097 if (LI == 0 || LI->isVolatile()) return false;
1099 // Both operands to the select need to be dereferencable, either absolutely
1100 // (e.g. allocas) or at this point because we can see other accesses to it.
1101 if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI,
1102 LI->getAlignment(), TD))
1104 if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI,
1105 LI->getAlignment(), TD))
1112 /// isSafePHIToSpeculate - PHI instructions that use an alloca and are
1113 /// subsequently loaded can be rewritten to load both input pointers in the pred
1114 /// blocks and then PHI the results, allowing the load of the alloca to be
1117 /// %P2 = phi [i32* %Alloca, i32* %Other]
1118 /// %V = load i32* %P2
1120 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1122 /// %V2 = load i32* %Other
1124 /// %V = phi [i32 %V1, i32 %V2]
1126 /// We can do this to a select if its only uses are loads and if the operand to
1127 /// the select can be loaded unconditionally.
1128 static bool isSafePHIToSpeculate(PHINode *PN, const TargetData *TD) {
1129 // For now, we can only do this promotion if the load is in the same block as
1130 // the PHI, and if there are no stores between the phi and load.
1131 // TODO: Allow recursive phi users.
1132 // TODO: Allow stores.
1133 BasicBlock *BB = PN->getParent();
1134 unsigned MaxAlign = 0;
1135 for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end();
1137 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1138 if (LI == 0 || LI->isVolatile()) return false;
1140 // For now we only allow loads in the same block as the PHI. This is a
1141 // common case that happens when instcombine merges two loads through a PHI.
1142 if (LI->getParent() != BB) return false;
1144 // Ensure that there are no instructions between the PHI and the load that
1146 for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI)
1147 if (BBI->mayWriteToMemory())
1150 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1153 // Okay, we know that we have one or more loads in the same block as the PHI.
1154 // We can transform this if it is safe to push the loads into the predecessor
1155 // blocks. The only thing to watch out for is that we can't put a possibly
1156 // trapping load in the predecessor if it is a critical edge.
1157 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1158 BasicBlock *Pred = PN->getIncomingBlock(i);
1160 // If the predecessor has a single successor, then the edge isn't critical.
1161 if (Pred->getTerminator()->getNumSuccessors() == 1)
1164 Value *InVal = PN->getIncomingValue(i);
1166 // If the InVal is an invoke in the pred, we can't put a load on the edge.
1167 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal))
1168 if (II->getParent() == Pred)
1171 // If this pointer is always safe to load, or if we can prove that there is
1172 // already a load in the block, then we can move the load to the pred block.
1173 if (InVal->isDereferenceablePointer() ||
1174 isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, TD))
1184 /// tryToMakeAllocaBePromotable - This returns true if the alloca only has
1185 /// direct (non-volatile) loads and stores to it. If the alloca is close but
1186 /// not quite there, this will transform the code to allow promotion. As such,
1187 /// it is a non-pure predicate.
1188 static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) {
1189 SetVector<Instruction*, SmallVector<Instruction*, 4>,
1190 SmallPtrSet<Instruction*, 4> > InstsToRewrite;
1192 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
1195 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
1196 if (LI->isVolatile())
1201 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1202 if (SI->getOperand(0) == AI || SI->isVolatile())
1203 return false; // Don't allow a store OF the AI, only INTO the AI.
1207 if (SelectInst *SI = dyn_cast<SelectInst>(U)) {
1208 // If the condition being selected on is a constant, fold the select, yes
1209 // this does (rarely) happen early on.
1210 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) {
1211 Value *Result = SI->getOperand(1+CI->isZero());
1212 SI->replaceAllUsesWith(Result);
1213 SI->eraseFromParent();
1215 // This is very rare and we just scrambled the use list of AI, start
1217 return tryToMakeAllocaBePromotable(AI, TD);
1220 // If it is safe to turn "load (select c, AI, ptr)" into a select of two
1221 // loads, then we can transform this by rewriting the select.
1222 if (!isSafeSelectToSpeculate(SI, TD))
1225 InstsToRewrite.insert(SI);
1229 if (PHINode *PN = dyn_cast<PHINode>(U)) {
1230 if (PN->use_empty()) { // Dead PHIs can be stripped.
1231 InstsToRewrite.insert(PN);
1235 // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads
1236 // in the pred blocks, then we can transform this by rewriting the PHI.
1237 if (!isSafePHIToSpeculate(PN, TD))
1240 InstsToRewrite.insert(PN);
1247 // If there are no instructions to rewrite, then all uses are load/stores and
1249 if (InstsToRewrite.empty())
1252 // If we have instructions that need to be rewritten for this to be promotable
1253 // take care of it now.
1254 for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) {
1255 if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) {
1256 // Selects in InstsToRewrite only have load uses. Rewrite each as two
1257 // loads with a new select.
1258 while (!SI->use_empty()) {
1259 LoadInst *LI = cast<LoadInst>(SI->use_back());
1261 IRBuilder<> Builder(LI);
1262 LoadInst *TrueLoad =
1263 Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t");
1264 LoadInst *FalseLoad =
1265 Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".t");
1267 // Transfer alignment and TBAA info if present.
1268 TrueLoad->setAlignment(LI->getAlignment());
1269 FalseLoad->setAlignment(LI->getAlignment());
1270 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
1271 TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1272 FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1275 Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad);
1277 LI->replaceAllUsesWith(V);
1278 LI->eraseFromParent();
1281 // Now that all the loads are gone, the select is gone too.
1282 SI->eraseFromParent();
1286 // Otherwise, we have a PHI node which allows us to push the loads into the
1288 PHINode *PN = cast<PHINode>(InstsToRewrite[i]);
1289 if (PN->use_empty()) {
1290 PN->eraseFromParent();
1294 const Type *LoadTy = cast<PointerType>(PN->getType())->getElementType();
1295 PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(),
1296 PN->getName()+".ld", PN);
1298 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1299 // matter which one we get and if any differ, it doesn't matter.
1300 LoadInst *SomeLoad = cast<LoadInst>(PN->use_back());
1301 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1302 unsigned Align = SomeLoad->getAlignment();
1304 // Rewrite all loads of the PN to use the new PHI.
1305 while (!PN->use_empty()) {
1306 LoadInst *LI = cast<LoadInst>(PN->use_back());
1307 LI->replaceAllUsesWith(NewPN);
1308 LI->eraseFromParent();
1311 // Inject loads into all of the pred blocks. Keep track of which blocks we
1312 // insert them into in case we have multiple edges from the same block.
1313 DenseMap<BasicBlock*, LoadInst*> InsertedLoads;
1315 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1316 BasicBlock *Pred = PN->getIncomingBlock(i);
1317 LoadInst *&Load = InsertedLoads[Pred];
1319 Load = new LoadInst(PN->getIncomingValue(i),
1320 PN->getName() + "." + Pred->getName(),
1321 Pred->getTerminator());
1322 Load->setAlignment(Align);
1323 if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1326 NewPN->addIncoming(Load, Pred);
1329 PN->eraseFromParent();
1336 bool SROA::performPromotion(Function &F) {
1337 std::vector<AllocaInst*> Allocas;
1338 DominatorTree *DT = 0;
1340 DT = &getAnalysis<DominatorTree>();
1342 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
1344 bool Changed = false;
1345 SmallVector<Instruction*, 64> Insts;
1350 // Find allocas that are safe to promote, by looking at all instructions in
1352 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
1353 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
1354 if (tryToMakeAllocaBePromotable(AI, TD))
1355 Allocas.push_back(AI);
1357 if (Allocas.empty()) break;
1360 PromoteMemToReg(Allocas, *DT);
1363 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
1364 AllocaInst *AI = Allocas[i];
1366 // Build list of instructions to promote.
1367 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
1369 Insts.push_back(cast<Instruction>(*UI));
1371 DbgDeclareInst *DDI = FindAllocaDbgDeclare(AI);
1373 DIB = new DIBuilder(*AI->getParent()->getParent()->getParent());
1374 AllocaPromoter(Insts, SSA, DDI, DIB).run(AI, Insts);
1378 NumPromoted += Allocas.size();
1382 // FIXME: Is there a better way to handle the lazy initialization of DIB
1383 // so that there doesn't need to be an explicit delete?
1390 /// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for
1391 /// SROA. It must be a struct or array type with a small number of elements.
1392 static bool ShouldAttemptScalarRepl(AllocaInst *AI) {
1393 const Type *T = AI->getAllocatedType();
1394 // Do not promote any struct into more than 32 separate vars.
1395 if (const StructType *ST = dyn_cast<StructType>(T))
1396 return ST->getNumElements() <= 32;
1397 // Arrays are much less likely to be safe for SROA; only consider
1398 // them if they are very small.
1399 if (const ArrayType *AT = dyn_cast<ArrayType>(T))
1400 return AT->getNumElements() <= 8;
1405 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
1406 // which runs on all of the malloc/alloca instructions in the function, removing
1407 // them if they are only used by getelementptr instructions.
1409 bool SROA::performScalarRepl(Function &F) {
1410 std::vector<AllocaInst*> WorkList;
1412 // Scan the entry basic block, adding allocas to the worklist.
1413 BasicBlock &BB = F.getEntryBlock();
1414 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
1415 if (AllocaInst *A = dyn_cast<AllocaInst>(I))
1416 WorkList.push_back(A);
1418 // Process the worklist
1419 bool Changed = false;
1420 while (!WorkList.empty()) {
1421 AllocaInst *AI = WorkList.back();
1422 WorkList.pop_back();
1424 // Handle dead allocas trivially. These can be formed by SROA'ing arrays
1425 // with unused elements.
1426 if (AI->use_empty()) {
1427 AI->eraseFromParent();
1432 // If this alloca is impossible for us to promote, reject it early.
1433 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized())
1436 // Check to see if this allocation is only modified by a memcpy/memmove from
1437 // a constant global. If this is the case, we can change all users to use
1438 // the constant global instead. This is commonly produced by the CFE by
1439 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
1440 // is only subsequently read.
1441 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) {
1442 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n');
1443 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n');
1444 Constant *TheSrc = cast<Constant>(TheCopy->getSource());
1445 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType()));
1446 TheCopy->eraseFromParent(); // Don't mutate the global.
1447 AI->eraseFromParent();
1453 // Check to see if we can perform the core SROA transformation. We cannot
1454 // transform the allocation instruction if it is an array allocation
1455 // (allocations OF arrays are ok though), and an allocation of a scalar
1456 // value cannot be decomposed at all.
1457 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType());
1459 // Do not promote [0 x %struct].
1460 if (AllocaSize == 0) continue;
1462 // Do not promote any struct whose size is too big.
1463 if (AllocaSize > SRThreshold) continue;
1465 // If the alloca looks like a good candidate for scalar replacement, and if
1466 // all its users can be transformed, then split up the aggregate into its
1467 // separate elements.
1468 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) {
1469 DoScalarReplacement(AI, WorkList);
1474 // If we can turn this aggregate value (potentially with casts) into a
1475 // simple scalar value that can be mem2reg'd into a register value.
1476 // IsNotTrivial tracks whether this is something that mem2reg could have
1477 // promoted itself. If so, we don't want to transform it needlessly. Note
1478 // that we can't just check based on the type: the alloca may be of an i32
1479 // but that has pointer arithmetic to set byte 3 of it or something.
1480 if (AllocaInst *NewAI =
1481 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) {
1482 NewAI->takeName(AI);
1483 AI->eraseFromParent();
1489 // Otherwise, couldn't process this alloca.
1495 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl
1496 /// predicate, do SROA now.
1497 void SROA::DoScalarReplacement(AllocaInst *AI,
1498 std::vector<AllocaInst*> &WorkList) {
1499 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n');
1500 SmallVector<AllocaInst*, 32> ElementAllocas;
1501 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
1502 ElementAllocas.reserve(ST->getNumContainedTypes());
1503 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
1504 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
1506 AI->getName() + "." + Twine(i), AI);
1507 ElementAllocas.push_back(NA);
1508 WorkList.push_back(NA); // Add to worklist for recursive processing
1511 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
1512 ElementAllocas.reserve(AT->getNumElements());
1513 const Type *ElTy = AT->getElementType();
1514 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1515 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
1516 AI->getName() + "." + Twine(i), AI);
1517 ElementAllocas.push_back(NA);
1518 WorkList.push_back(NA); // Add to worklist for recursive processing
1522 // Now that we have created the new alloca instructions, rewrite all the
1523 // uses of the old alloca.
1524 RewriteForScalarRepl(AI, AI, 0, ElementAllocas);
1526 // Now erase any instructions that were made dead while rewriting the alloca.
1527 DeleteDeadInstructions();
1528 AI->eraseFromParent();
1533 /// DeleteDeadInstructions - Erase instructions on the DeadInstrs list,
1534 /// recursively including all their operands that become trivially dead.
1535 void SROA::DeleteDeadInstructions() {
1536 while (!DeadInsts.empty()) {
1537 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val());
1539 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
1540 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
1541 // Zero out the operand and see if it becomes trivially dead.
1542 // (But, don't add allocas to the dead instruction list -- they are
1543 // already on the worklist and will be deleted separately.)
1545 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U))
1546 DeadInsts.push_back(U);
1549 I->eraseFromParent();
1553 /// isSafeForScalarRepl - Check if instruction I is a safe use with regard to
1554 /// performing scalar replacement of alloca AI. The results are flagged in
1555 /// the Info parameter. Offset indicates the position within AI that is
1556 /// referenced by this instruction.
1557 void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset,
1559 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
1560 Instruction *User = cast<Instruction>(*UI);
1562 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1563 isSafeForScalarRepl(BC, Offset, Info);
1564 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1565 uint64_t GEPOffset = Offset;
1566 isSafeGEP(GEPI, GEPOffset, Info);
1568 isSafeForScalarRepl(GEPI, GEPOffset, Info);
1569 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
1570 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
1572 return MarkUnsafe(Info, User);
1573 isSafeMemAccess(Offset, Length->getZExtValue(), 0,
1574 UI.getOperandNo() == 0, Info, MI,
1575 true /*AllowWholeAccess*/);
1576 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1577 if (LI->isVolatile())
1578 return MarkUnsafe(Info, User);
1579 const Type *LIType = LI->getType();
1580 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType),
1581 LIType, false, Info, LI, true /*AllowWholeAccess*/);
1582 Info.hasALoadOrStore = true;
1584 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1585 // Store is ok if storing INTO the pointer, not storing the pointer
1586 if (SI->isVolatile() || SI->getOperand(0) == I)
1587 return MarkUnsafe(Info, User);
1589 const Type *SIType = SI->getOperand(0)->getType();
1590 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType),
1591 SIType, true, Info, SI, true /*AllowWholeAccess*/);
1592 Info.hasALoadOrStore = true;
1593 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) {
1594 isSafePHISelectUseForScalarRepl(User, Offset, Info);
1596 return MarkUnsafe(Info, User);
1598 if (Info.isUnsafe) return;
1603 /// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer
1604 /// derived from the alloca, we can often still split the alloca into elements.
1605 /// This is useful if we have a large alloca where one element is phi'd
1606 /// together somewhere: we can SRoA and promote all the other elements even if
1607 /// we end up not being able to promote this one.
1609 /// All we require is that the uses of the PHI do not index into other parts of
1610 /// the alloca. The most important use case for this is single load and stores
1611 /// that are PHI'd together, which can happen due to code sinking.
1612 void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset,
1614 // If we've already checked this PHI, don't do it again.
1615 if (PHINode *PN = dyn_cast<PHINode>(I))
1616 if (!Info.CheckedPHIs.insert(PN))
1619 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
1620 Instruction *User = cast<Instruction>(*UI);
1622 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1623 isSafePHISelectUseForScalarRepl(BC, Offset, Info);
1624 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1625 // Only allow "bitcast" GEPs for simplicity. We could generalize this,
1626 // but would have to prove that we're staying inside of an element being
1628 if (!GEPI->hasAllZeroIndices())
1629 return MarkUnsafe(Info, User);
1630 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info);
1631 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1632 if (LI->isVolatile())
1633 return MarkUnsafe(Info, User);
1634 const Type *LIType = LI->getType();
1635 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType),
1636 LIType, false, Info, LI, false /*AllowWholeAccess*/);
1637 Info.hasALoadOrStore = true;
1639 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1640 // Store is ok if storing INTO the pointer, not storing the pointer
1641 if (SI->isVolatile() || SI->getOperand(0) == I)
1642 return MarkUnsafe(Info, User);
1644 const Type *SIType = SI->getOperand(0)->getType();
1645 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType),
1646 SIType, true, Info, SI, false /*AllowWholeAccess*/);
1647 Info.hasALoadOrStore = true;
1648 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) {
1649 isSafePHISelectUseForScalarRepl(User, Offset, Info);
1651 return MarkUnsafe(Info, User);
1653 if (Info.isUnsafe) return;
1657 /// isSafeGEP - Check if a GEP instruction can be handled for scalar
1658 /// replacement. It is safe when all the indices are constant, in-bounds
1659 /// references, and when the resulting offset corresponds to an element within
1660 /// the alloca type. The results are flagged in the Info parameter. Upon
1661 /// return, Offset is adjusted as specified by the GEP indices.
1662 void SROA::isSafeGEP(GetElementPtrInst *GEPI,
1663 uint64_t &Offset, AllocaInfo &Info) {
1664 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI);
1668 // Walk through the GEP type indices, checking the types that this indexes
1670 for (; GEPIt != E; ++GEPIt) {
1671 // Ignore struct elements, no extra checking needed for these.
1672 if ((*GEPIt)->isStructTy())
1675 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand());
1677 return MarkUnsafe(Info, GEPI);
1680 // Compute the offset due to this GEP and check if the alloca has a
1681 // component element at that offset.
1682 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
1683 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(),
1684 &Indices[0], Indices.size());
1685 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0))
1686 MarkUnsafe(Info, GEPI);
1689 /// isHomogeneousAggregate - Check if type T is a struct or array containing
1690 /// elements of the same type (which is always true for arrays). If so,
1691 /// return true with NumElts and EltTy set to the number of elements and the
1692 /// element type, respectively.
1693 static bool isHomogeneousAggregate(const Type *T, unsigned &NumElts,
1694 const Type *&EltTy) {
1695 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) {
1696 NumElts = AT->getNumElements();
1697 EltTy = (NumElts == 0 ? 0 : AT->getElementType());
1700 if (const StructType *ST = dyn_cast<StructType>(T)) {
1701 NumElts = ST->getNumContainedTypes();
1702 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0));
1703 for (unsigned n = 1; n < NumElts; ++n) {
1704 if (ST->getContainedType(n) != EltTy)
1712 /// isCompatibleAggregate - Check if T1 and T2 are either the same type or are
1713 /// "homogeneous" aggregates with the same element type and number of elements.
1714 static bool isCompatibleAggregate(const Type *T1, const Type *T2) {
1718 unsigned NumElts1, NumElts2;
1719 const Type *EltTy1, *EltTy2;
1720 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) &&
1721 isHomogeneousAggregate(T2, NumElts2, EltTy2) &&
1722 NumElts1 == NumElts2 &&
1729 /// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI
1730 /// alloca or has an offset and size that corresponds to a component element
1731 /// within it. The offset checked here may have been formed from a GEP with a
1732 /// pointer bitcasted to a different type.
1734 /// If AllowWholeAccess is true, then this allows uses of the entire alloca as a
1735 /// unit. If false, it only allows accesses known to be in a single element.
1736 void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize,
1737 const Type *MemOpType, bool isStore,
1738 AllocaInfo &Info, Instruction *TheAccess,
1739 bool AllowWholeAccess) {
1740 // Check if this is a load/store of the entire alloca.
1741 if (Offset == 0 && AllowWholeAccess &&
1742 MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) {
1743 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer
1744 // loads/stores (which are essentially the same as the MemIntrinsics with
1745 // regard to copying padding between elements). But, if an alloca is
1746 // flagged as both a source and destination of such operations, we'll need
1747 // to check later for padding between elements.
1748 if (!MemOpType || MemOpType->isIntegerTy()) {
1750 Info.isMemCpyDst = true;
1752 Info.isMemCpySrc = true;
1755 // This is also safe for references using a type that is compatible with
1756 // the type of the alloca, so that loads/stores can be rewritten using
1757 // insertvalue/extractvalue.
1758 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) {
1759 Info.hasSubelementAccess = true;
1763 // Check if the offset/size correspond to a component within the alloca type.
1764 const Type *T = Info.AI->getAllocatedType();
1765 if (TypeHasComponent(T, Offset, MemSize)) {
1766 Info.hasSubelementAccess = true;
1770 return MarkUnsafe(Info, TheAccess);
1773 /// TypeHasComponent - Return true if T has a component type with the
1774 /// specified offset and size. If Size is zero, do not check the size.
1775 bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) {
1778 if (const StructType *ST = dyn_cast<StructType>(T)) {
1779 const StructLayout *Layout = TD->getStructLayout(ST);
1780 unsigned EltIdx = Layout->getElementContainingOffset(Offset);
1781 EltTy = ST->getContainedType(EltIdx);
1782 EltSize = TD->getTypeAllocSize(EltTy);
1783 Offset -= Layout->getElementOffset(EltIdx);
1784 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) {
1785 EltTy = AT->getElementType();
1786 EltSize = TD->getTypeAllocSize(EltTy);
1787 if (Offset >= AT->getNumElements() * EltSize)
1793 if (Offset == 0 && (Size == 0 || EltSize == Size))
1795 // Check if the component spans multiple elements.
1796 if (Offset + Size > EltSize)
1798 return TypeHasComponent(EltTy, Offset, Size);
1801 /// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite
1802 /// the instruction I, which references it, to use the separate elements.
1803 /// Offset indicates the position within AI that is referenced by this
1805 void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
1806 SmallVector<AllocaInst*, 32> &NewElts) {
1807 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) {
1808 Use &TheUse = UI.getUse();
1809 Instruction *User = cast<Instruction>(*UI++);
1811 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1812 RewriteBitCast(BC, AI, Offset, NewElts);
1816 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1817 RewriteGEP(GEPI, AI, Offset, NewElts);
1821 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
1822 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
1823 uint64_t MemSize = Length->getZExtValue();
1825 MemSize == TD->getTypeAllocSize(AI->getAllocatedType()))
1826 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts);
1827 // Otherwise the intrinsic can only touch a single element and the
1828 // address operand will be updated, so nothing else needs to be done.
1832 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1833 const Type *LIType = LI->getType();
1835 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) {
1837 // %res = load { i32, i32 }* %alloc
1839 // %load.0 = load i32* %alloc.0
1840 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0
1841 // %load.1 = load i32* %alloc.1
1842 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1
1843 // (Also works for arrays instead of structs)
1844 Value *Insert = UndefValue::get(LIType);
1845 IRBuilder<> Builder(LI);
1846 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1847 Value *Load = Builder.CreateLoad(NewElts[i], "load");
1848 Insert = Builder.CreateInsertValue(Insert, Load, i, "insert");
1850 LI->replaceAllUsesWith(Insert);
1851 DeadInsts.push_back(LI);
1852 } else if (LIType->isIntegerTy() &&
1853 TD->getTypeAllocSize(LIType) ==
1854 TD->getTypeAllocSize(AI->getAllocatedType())) {
1855 // If this is a load of the entire alloca to an integer, rewrite it.
1856 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts);
1861 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1862 Value *Val = SI->getOperand(0);
1863 const Type *SIType = Val->getType();
1864 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) {
1866 // store { i32, i32 } %val, { i32, i32 }* %alloc
1868 // %val.0 = extractvalue { i32, i32 } %val, 0
1869 // store i32 %val.0, i32* %alloc.0
1870 // %val.1 = extractvalue { i32, i32 } %val, 1
1871 // store i32 %val.1, i32* %alloc.1
1872 // (Also works for arrays instead of structs)
1873 IRBuilder<> Builder(SI);
1874 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1875 Value *Extract = Builder.CreateExtractValue(Val, i, Val->getName());
1876 Builder.CreateStore(Extract, NewElts[i]);
1878 DeadInsts.push_back(SI);
1879 } else if (SIType->isIntegerTy() &&
1880 TD->getTypeAllocSize(SIType) ==
1881 TD->getTypeAllocSize(AI->getAllocatedType())) {
1882 // If this is a store of the entire alloca from an integer, rewrite it.
1883 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts);
1888 if (isa<SelectInst>(User) || isa<PHINode>(User)) {
1889 // If we have a PHI user of the alloca itself (as opposed to a GEP or
1890 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to
1892 if (!isa<AllocaInst>(I)) continue;
1894 assert(Offset == 0 && NewElts[0] &&
1895 "Direct alloca use should have a zero offset");
1897 // If we have a use of the alloca, we know the derived uses will be
1898 // utilizing just the first element of the scalarized result. Insert a
1899 // bitcast of the first alloca before the user as required.
1900 AllocaInst *NewAI = NewElts[0];
1901 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI);
1902 NewAI->moveBefore(BCI);
1909 /// RewriteBitCast - Update a bitcast reference to the alloca being replaced
1910 /// and recursively continue updating all of its uses.
1911 void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
1912 SmallVector<AllocaInst*, 32> &NewElts) {
1913 RewriteForScalarRepl(BC, AI, Offset, NewElts);
1914 if (BC->getOperand(0) != AI)
1917 // The bitcast references the original alloca. Replace its uses with
1918 // references to the first new element alloca.
1919 Instruction *Val = NewElts[0];
1920 if (Val->getType() != BC->getDestTy()) {
1921 Val = new BitCastInst(Val, BC->getDestTy(), "", BC);
1924 BC->replaceAllUsesWith(Val);
1925 DeadInsts.push_back(BC);
1928 /// FindElementAndOffset - Return the index of the element containing Offset
1929 /// within the specified type, which must be either a struct or an array.
1930 /// Sets T to the type of the element and Offset to the offset within that
1931 /// element. IdxTy is set to the type of the index result to be used in a
1932 /// GEP instruction.
1933 uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset,
1934 const Type *&IdxTy) {
1936 if (const StructType *ST = dyn_cast<StructType>(T)) {
1937 const StructLayout *Layout = TD->getStructLayout(ST);
1938 Idx = Layout->getElementContainingOffset(Offset);
1939 T = ST->getContainedType(Idx);
1940 Offset -= Layout->getElementOffset(Idx);
1941 IdxTy = Type::getInt32Ty(T->getContext());
1944 const ArrayType *AT = cast<ArrayType>(T);
1945 T = AT->getElementType();
1946 uint64_t EltSize = TD->getTypeAllocSize(T);
1947 Idx = Offset / EltSize;
1948 Offset -= Idx * EltSize;
1949 IdxTy = Type::getInt64Ty(T->getContext());
1953 /// RewriteGEP - Check if this GEP instruction moves the pointer across
1954 /// elements of the alloca that are being split apart, and if so, rewrite
1955 /// the GEP to be relative to the new element.
1956 void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
1957 SmallVector<AllocaInst*, 32> &NewElts) {
1958 uint64_t OldOffset = Offset;
1959 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
1960 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(),
1961 &Indices[0], Indices.size());
1963 RewriteForScalarRepl(GEPI, AI, Offset, NewElts);
1965 const Type *T = AI->getAllocatedType();
1967 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy);
1968 if (GEPI->getOperand(0) == AI)
1969 OldIdx = ~0ULL; // Force the GEP to be rewritten.
1971 T = AI->getAllocatedType();
1972 uint64_t EltOffset = Offset;
1973 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy);
1975 // If this GEP does not move the pointer across elements of the alloca
1976 // being split, then it does not needs to be rewritten.
1980 const Type *i32Ty = Type::getInt32Ty(AI->getContext());
1981 SmallVector<Value*, 8> NewArgs;
1982 NewArgs.push_back(Constant::getNullValue(i32Ty));
1983 while (EltOffset != 0) {
1984 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy);
1985 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx));
1987 Instruction *Val = NewElts[Idx];
1988 if (NewArgs.size() > 1) {
1989 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(),
1990 NewArgs.end(), "", GEPI);
1991 Val->takeName(GEPI);
1993 if (Val->getType() != GEPI->getType())
1994 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI);
1995 GEPI->replaceAllUsesWith(Val);
1996 DeadInsts.push_back(GEPI);
1999 /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI.
2000 /// Rewrite it to copy or set the elements of the scalarized memory.
2001 void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
2003 SmallVector<AllocaInst*, 32> &NewElts) {
2004 // If this is a memcpy/memmove, construct the other pointer as the
2005 // appropriate type. The "Other" pointer is the pointer that goes to memory
2006 // that doesn't have anything to do with the alloca that we are promoting. For
2007 // memset, this Value* stays null.
2008 Value *OtherPtr = 0;
2009 unsigned MemAlignment = MI->getAlignment();
2010 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy
2011 if (Inst == MTI->getRawDest())
2012 OtherPtr = MTI->getRawSource();
2014 assert(Inst == MTI->getRawSource());
2015 OtherPtr = MTI->getRawDest();
2019 // If there is an other pointer, we want to convert it to the same pointer
2020 // type as AI has, so we can GEP through it safely.
2022 unsigned AddrSpace =
2023 cast<PointerType>(OtherPtr->getType())->getAddressSpace();
2025 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an
2026 // optimization, but it's also required to detect the corner case where
2027 // both pointer operands are referencing the same memory, and where
2028 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This
2029 // function is only called for mem intrinsics that access the whole
2030 // aggregate, so non-zero GEPs are not an issue here.)
2031 OtherPtr = OtherPtr->stripPointerCasts();
2033 // Copying the alloca to itself is a no-op: just delete it.
2034 if (OtherPtr == AI || OtherPtr == NewElts[0]) {
2035 // This code will run twice for a no-op memcpy -- once for each operand.
2036 // Put only one reference to MI on the DeadInsts list.
2037 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(),
2038 E = DeadInsts.end(); I != E; ++I)
2039 if (*I == MI) return;
2040 DeadInsts.push_back(MI);
2044 // If the pointer is not the right type, insert a bitcast to the right
2047 PointerType::get(AI->getType()->getElementType(), AddrSpace);
2049 if (OtherPtr->getType() != NewTy)
2050 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI);
2053 // Process each element of the aggregate.
2054 bool SROADest = MI->getRawDest() == Inst;
2056 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext()));
2058 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2059 // If this is a memcpy/memmove, emit a GEP of the other element address.
2060 Value *OtherElt = 0;
2061 unsigned OtherEltAlign = MemAlignment;
2064 Value *Idx[2] = { Zero,
2065 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) };
2066 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2,
2067 OtherPtr->getName()+"."+Twine(i),
2070 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType());
2071 const Type *OtherTy = OtherPtrTy->getElementType();
2072 if (const StructType *ST = dyn_cast<StructType>(OtherTy)) {
2073 EltOffset = TD->getStructLayout(ST)->getElementOffset(i);
2075 const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType();
2076 EltOffset = TD->getTypeAllocSize(EltTy)*i;
2079 // The alignment of the other pointer is the guaranteed alignment of the
2080 // element, which is affected by both the known alignment of the whole
2081 // mem intrinsic and the alignment of the element. If the alignment of
2082 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the
2083 // known alignment is just 4 bytes.
2084 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset);
2087 Value *EltPtr = NewElts[i];
2088 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType();
2090 // If we got down to a scalar, insert a load or store as appropriate.
2091 if (EltTy->isSingleValueType()) {
2092 if (isa<MemTransferInst>(MI)) {
2094 // From Other to Alloca.
2095 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI);
2096 new StoreInst(Elt, EltPtr, MI);
2098 // From Alloca to Other.
2099 Value *Elt = new LoadInst(EltPtr, "tmp", MI);
2100 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI);
2104 assert(isa<MemSetInst>(MI));
2106 // If the stored element is zero (common case), just store a null
2109 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) {
2111 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0>
2113 // If EltTy is a vector type, get the element type.
2114 const Type *ValTy = EltTy->getScalarType();
2116 // Construct an integer with the right value.
2117 unsigned EltSize = TD->getTypeSizeInBits(ValTy);
2118 APInt OneVal(EltSize, CI->getZExtValue());
2119 APInt TotalVal(OneVal);
2121 for (unsigned i = 0; 8*i < EltSize; ++i) {
2122 TotalVal = TotalVal.shl(8);
2126 // Convert the integer value to the appropriate type.
2127 StoreVal = ConstantInt::get(CI->getContext(), TotalVal);
2128 if (ValTy->isPointerTy())
2129 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy);
2130 else if (ValTy->isFloatingPointTy())
2131 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy);
2132 assert(StoreVal->getType() == ValTy && "Type mismatch!");
2134 // If the requested value was a vector constant, create it.
2135 if (EltTy != ValTy) {
2136 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements();
2137 SmallVector<Constant*, 16> Elts(NumElts, StoreVal);
2138 StoreVal = ConstantVector::get(Elts);
2141 new StoreInst(StoreVal, EltPtr, MI);
2144 // Otherwise, if we're storing a byte variable, use a memset call for
2148 unsigned EltSize = TD->getTypeAllocSize(EltTy);
2150 IRBuilder<> Builder(MI);
2152 // Finally, insert the meminst for this element.
2153 if (isa<MemSetInst>(MI)) {
2154 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize,
2157 assert(isa<MemTransferInst>(MI));
2158 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr
2159 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr
2161 if (isa<MemCpyInst>(MI))
2162 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile());
2164 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile());
2167 DeadInsts.push_back(MI);
2170 /// RewriteStoreUserOfWholeAlloca - We found a store of an integer that
2171 /// overwrites the entire allocation. Extract out the pieces of the stored
2172 /// integer and store them individually.
2173 void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
2174 SmallVector<AllocaInst*, 32> &NewElts){
2175 // Extract each element out of the integer according to its structure offset
2176 // and store the element value to the individual alloca.
2177 Value *SrcVal = SI->getOperand(0);
2178 const Type *AllocaEltTy = AI->getAllocatedType();
2179 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
2181 IRBuilder<> Builder(SI);
2183 // Handle tail padding by extending the operand
2184 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits)
2185 SrcVal = Builder.CreateZExt(SrcVal,
2186 IntegerType::get(SI->getContext(), AllocaSizeBits));
2188 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI
2191 // There are two forms here: AI could be an array or struct. Both cases
2192 // have different ways to compute the element offset.
2193 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
2194 const StructLayout *Layout = TD->getStructLayout(EltSTy);
2196 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2197 // Get the number of bits to shift SrcVal to get the value.
2198 const Type *FieldTy = EltSTy->getElementType(i);
2199 uint64_t Shift = Layout->getElementOffsetInBits(i);
2201 if (TD->isBigEndian())
2202 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy);
2204 Value *EltVal = SrcVal;
2206 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
2207 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt");
2210 // Truncate down to an integer of the right size.
2211 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
2213 // Ignore zero sized fields like {}, they obviously contain no data.
2214 if (FieldSizeBits == 0) continue;
2216 if (FieldSizeBits != AllocaSizeBits)
2217 EltVal = Builder.CreateTrunc(EltVal,
2218 IntegerType::get(SI->getContext(), FieldSizeBits));
2219 Value *DestField = NewElts[i];
2220 if (EltVal->getType() == FieldTy) {
2221 // Storing to an integer field of this size, just do it.
2222 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) {
2223 // Bitcast to the right element type (for fp/vector values).
2224 EltVal = Builder.CreateBitCast(EltVal, FieldTy);
2226 // Otherwise, bitcast the dest pointer (for aggregates).
2227 DestField = Builder.CreateBitCast(DestField,
2228 PointerType::getUnqual(EltVal->getType()));
2230 new StoreInst(EltVal, DestField, SI);
2234 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
2235 const Type *ArrayEltTy = ATy->getElementType();
2236 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
2237 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy);
2241 if (TD->isBigEndian())
2242 Shift = AllocaSizeBits-ElementOffset;
2246 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2247 // Ignore zero sized fields like {}, they obviously contain no data.
2248 if (ElementSizeBits == 0) continue;
2250 Value *EltVal = SrcVal;
2252 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
2253 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt");
2256 // Truncate down to an integer of the right size.
2257 if (ElementSizeBits != AllocaSizeBits)
2258 EltVal = Builder.CreateTrunc(EltVal,
2259 IntegerType::get(SI->getContext(),
2261 Value *DestField = NewElts[i];
2262 if (EltVal->getType() == ArrayEltTy) {
2263 // Storing to an integer field of this size, just do it.
2264 } else if (ArrayEltTy->isFloatingPointTy() ||
2265 ArrayEltTy->isVectorTy()) {
2266 // Bitcast to the right element type (for fp/vector values).
2267 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy);
2269 // Otherwise, bitcast the dest pointer (for aggregates).
2270 DestField = Builder.CreateBitCast(DestField,
2271 PointerType::getUnqual(EltVal->getType()));
2273 new StoreInst(EltVal, DestField, SI);
2275 if (TD->isBigEndian())
2276 Shift -= ElementOffset;
2278 Shift += ElementOffset;
2282 DeadInsts.push_back(SI);
2285 /// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to
2286 /// an integer. Load the individual pieces to form the aggregate value.
2287 void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
2288 SmallVector<AllocaInst*, 32> &NewElts) {
2289 // Extract each element out of the NewElts according to its structure offset
2290 // and form the result value.
2291 const Type *AllocaEltTy = AI->getAllocatedType();
2292 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
2294 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI
2297 // There are two forms here: AI could be an array or struct. Both cases
2298 // have different ways to compute the element offset.
2299 const StructLayout *Layout = 0;
2300 uint64_t ArrayEltBitOffset = 0;
2301 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
2302 Layout = TD->getStructLayout(EltSTy);
2304 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
2305 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
2309 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits));
2311 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2312 // Load the value from the alloca. If the NewElt is an aggregate, cast
2313 // the pointer to an integer of the same size before doing the load.
2314 Value *SrcField = NewElts[i];
2315 const Type *FieldTy =
2316 cast<PointerType>(SrcField->getType())->getElementType();
2317 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
2319 // Ignore zero sized fields like {}, they obviously contain no data.
2320 if (FieldSizeBits == 0) continue;
2322 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(),
2324 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() &&
2325 !FieldTy->isVectorTy())
2326 SrcField = new BitCastInst(SrcField,
2327 PointerType::getUnqual(FieldIntTy),
2329 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI);
2331 // If SrcField is a fp or vector of the right size but that isn't an
2332 // integer type, bitcast to an integer so we can shift it.
2333 if (SrcField->getType() != FieldIntTy)
2334 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI);
2336 // Zero extend the field to be the same size as the final alloca so that
2337 // we can shift and insert it.
2338 if (SrcField->getType() != ResultVal->getType())
2339 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI);
2341 // Determine the number of bits to shift SrcField.
2343 if (Layout) // Struct case.
2344 Shift = Layout->getElementOffsetInBits(i);
2346 Shift = i*ArrayEltBitOffset;
2348 if (TD->isBigEndian())
2349 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth();
2352 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift);
2353 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI);
2356 // Don't create an 'or x, 0' on the first iteration.
2357 if (!isa<Constant>(ResultVal) ||
2358 !cast<Constant>(ResultVal)->isNullValue())
2359 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI);
2361 ResultVal = SrcField;
2364 // Handle tail padding by truncating the result
2365 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits)
2366 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI);
2368 LI->replaceAllUsesWith(ResultVal);
2369 DeadInsts.push_back(LI);
2372 /// HasPadding - Return true if the specified type has any structure or
2373 /// alignment padding in between the elements that would be split apart
2374 /// by SROA; return false otherwise.
2375 static bool HasPadding(const Type *Ty, const TargetData &TD) {
2376 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2377 Ty = ATy->getElementType();
2378 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty);
2381 // SROA currently handles only Arrays and Structs.
2382 const StructType *STy = cast<StructType>(Ty);
2383 const StructLayout *SL = TD.getStructLayout(STy);
2384 unsigned PrevFieldBitOffset = 0;
2385 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2386 unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
2388 // Check to see if there is any padding between this element and the
2391 unsigned PrevFieldEnd =
2392 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1));
2393 if (PrevFieldEnd < FieldBitOffset)
2396 PrevFieldBitOffset = FieldBitOffset;
2398 // Check for tail padding.
2399 if (unsigned EltCount = STy->getNumElements()) {
2400 unsigned PrevFieldEnd = PrevFieldBitOffset +
2401 TD.getTypeSizeInBits(STy->getElementType(EltCount-1));
2402 if (PrevFieldEnd < SL->getSizeInBits())
2408 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
2409 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe,
2410 /// or 1 if safe after canonicalization has been performed.
2411 bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) {
2412 // Loop over the use list of the alloca. We can only transform it if all of
2413 // the users are safe to transform.
2414 AllocaInfo Info(AI);
2416 isSafeForScalarRepl(AI, 0, Info);
2417 if (Info.isUnsafe) {
2418 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n');
2422 // Okay, we know all the users are promotable. If the aggregate is a memcpy
2423 // source and destination, we have to be careful. In particular, the memcpy
2424 // could be moving around elements that live in structure padding of the LLVM
2425 // types, but may actually be used. In these cases, we refuse to promote the
2427 if (Info.isMemCpySrc && Info.isMemCpyDst &&
2428 HasPadding(AI->getAllocatedType(), *TD))
2431 // If the alloca never has an access to just *part* of it, but is accessed
2432 // via loads and stores, then we should use ConvertToScalarInfo to promote
2433 // the alloca instead of promoting each piece at a time and inserting fission
2435 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) {
2436 // If the struct/array just has one element, use basic SRoA.
2437 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
2438 if (ST->getNumElements() > 1) return false;
2440 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1)
2450 /// PointsToConstantGlobal - Return true if V (possibly indirectly) points to
2451 /// some part of a constant global variable. This intentionally only accepts
2452 /// constant expressions because we don't can't rewrite arbitrary instructions.
2453 static bool PointsToConstantGlobal(Value *V) {
2454 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
2455 return GV->isConstant();
2456 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
2457 if (CE->getOpcode() == Instruction::BitCast ||
2458 CE->getOpcode() == Instruction::GetElementPtr)
2459 return PointsToConstantGlobal(CE->getOperand(0));
2463 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
2464 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
2465 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
2466 /// track of whether it moves the pointer (with isOffset) but otherwise traverse
2467 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
2468 /// the alloca, and if the source pointer is a pointer to a constant global, we
2469 /// can optimize this.
2470 static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy,
2472 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
2473 User *U = cast<Instruction>(*UI);
2475 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
2476 // Ignore non-volatile loads, they are always ok.
2477 if (LI->isVolatile()) return false;
2481 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
2482 // If uses of the bitcast are ok, we are ok.
2483 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset))
2487 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
2488 // If the GEP has all zero indices, it doesn't offset the pointer. If it
2489 // doesn't, it does.
2490 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy,
2491 isOffset || !GEP->hasAllZeroIndices()))
2496 if (CallSite CS = U) {
2497 // If this is the function being called then we treat it like a load and
2499 if (CS.isCallee(UI))
2502 // If this is a readonly/readnone call site, then we know it is just a
2503 // load (but one that potentially returns the value itself), so we can
2504 // ignore it if we know that the value isn't captured.
2505 unsigned ArgNo = CS.getArgumentNo(UI);
2506 if (CS.onlyReadsMemory() &&
2507 (CS.getInstruction()->use_empty() ||
2508 CS.paramHasAttr(ArgNo+1, Attribute::NoCapture)))
2511 // If this is being passed as a byval argument, the caller is making a
2512 // copy, so it is only a read of the alloca.
2513 if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal))
2517 // If this is isn't our memcpy/memmove, reject it as something we can't
2519 MemTransferInst *MI = dyn_cast<MemTransferInst>(U);
2523 // If the transfer is using the alloca as a source of the transfer, then
2524 // ignore it since it is a load (unless the transfer is volatile).
2525 if (UI.getOperandNo() == 1) {
2526 if (MI->isVolatile()) return false;
2530 // If we already have seen a copy, reject the second one.
2531 if (TheCopy) return false;
2533 // If the pointer has been offset from the start of the alloca, we can't
2534 // safely handle this.
2535 if (isOffset) return false;
2537 // If the memintrinsic isn't using the alloca as the dest, reject it.
2538 if (UI.getOperandNo() != 0) return false;
2540 // If the source of the memcpy/move is not a constant global, reject it.
2541 if (!PointsToConstantGlobal(MI->getSource()))
2544 // Otherwise, the transform is safe. Remember the copy instruction.
2550 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
2551 /// modified by a copy from a constant global. If we can prove this, we can
2552 /// replace any uses of the alloca with uses of the global directly.
2553 MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) {
2554 MemTransferInst *TheCopy = 0;
2555 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false))