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/Dominators.h"
34 #include "llvm/Analysis/Loads.h"
35 #include "llvm/Analysis/ValueTracking.h"
36 #include "llvm/Target/TargetData.h"
37 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
38 #include "llvm/Transforms/Utils/Local.h"
39 #include "llvm/Transforms/Utils/SSAUpdater.h"
40 #include "llvm/Support/CallSite.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/ErrorHandling.h"
43 #include "llvm/Support/GetElementPtrTypeIterator.h"
44 #include "llvm/Support/IRBuilder.h"
45 #include "llvm/Support/MathExtras.h"
46 #include "llvm/Support/raw_ostream.h"
47 #include "llvm/ADT/SetVector.h"
48 #include "llvm/ADT/SmallVector.h"
49 #include "llvm/ADT/Statistic.h"
52 STATISTIC(NumReplaced, "Number of allocas broken up");
53 STATISTIC(NumPromoted, "Number of allocas promoted");
54 STATISTIC(NumAdjusted, "Number of scalar allocas adjusted to allow promotion");
55 STATISTIC(NumConverted, "Number of aggregates converted to scalar");
56 STATISTIC(NumGlobals, "Number of allocas copied from constant global");
59 struct SROA : public FunctionPass {
60 SROA(int T, bool hasDT, char &ID)
61 : FunctionPass(ID), HasDomTree(hasDT) {
68 bool runOnFunction(Function &F);
70 bool performScalarRepl(Function &F);
71 bool performPromotion(Function &F);
77 /// DeadInsts - Keep track of instructions we have made dead, so that
78 /// we can remove them after we are done working.
79 SmallVector<Value*, 32> DeadInsts;
81 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures
82 /// information about the uses. All these fields are initialized to false
83 /// and set to true when something is learned.
85 /// The alloca to promote.
88 /// CheckedPHIs - This is a set of verified PHI nodes, to prevent infinite
89 /// looping and avoid redundant work.
90 SmallPtrSet<PHINode*, 8> CheckedPHIs;
92 /// isUnsafe - This is set to true if the alloca cannot be SROA'd.
95 /// isMemCpySrc - This is true if this aggregate is memcpy'd from.
98 /// isMemCpyDst - This is true if this aggregate is memcpy'd into.
101 /// hasSubelementAccess - This is true if a subelement of the alloca is
102 /// ever accessed, or false if the alloca is only accessed with mem
103 /// intrinsics or load/store that only access the entire alloca at once.
104 bool hasSubelementAccess : 1;
106 /// hasALoadOrStore - This is true if there are any loads or stores to it.
107 /// The alloca may just be accessed with memcpy, for example, which would
109 bool hasALoadOrStore : 1;
111 explicit AllocaInfo(AllocaInst *ai)
112 : AI(ai), isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false),
113 hasSubelementAccess(false), hasALoadOrStore(false) {}
116 unsigned SRThreshold;
118 void MarkUnsafe(AllocaInfo &I, Instruction *User) {
120 DEBUG(dbgs() << " Transformation preventing inst: " << *User << '\n');
123 bool isSafeAllocaToScalarRepl(AllocaInst *AI);
125 void isSafeForScalarRepl(Instruction *I, uint64_t Offset, AllocaInfo &Info);
126 void isSafePHISelectUseForScalarRepl(Instruction *User, uint64_t Offset,
128 void isSafeGEP(GetElementPtrInst *GEPI, uint64_t &Offset, AllocaInfo &Info);
129 void isSafeMemAccess(uint64_t Offset, uint64_t MemSize,
130 const Type *MemOpType, bool isStore, AllocaInfo &Info,
131 Instruction *TheAccess, bool AllowWholeAccess);
132 bool TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size);
133 uint64_t FindElementAndOffset(const Type *&T, uint64_t &Offset,
136 void DoScalarReplacement(AllocaInst *AI,
137 std::vector<AllocaInst*> &WorkList);
138 void DeleteDeadInstructions();
140 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
141 SmallVector<AllocaInst*, 32> &NewElts);
142 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
143 SmallVector<AllocaInst*, 32> &NewElts);
144 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
145 SmallVector<AllocaInst*, 32> &NewElts);
146 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
148 SmallVector<AllocaInst*, 32> &NewElts);
149 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
150 SmallVector<AllocaInst*, 32> &NewElts);
151 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
152 SmallVector<AllocaInst*, 32> &NewElts);
154 static MemTransferInst *isOnlyCopiedFromConstantGlobal(AllocaInst *AI);
157 // SROA_DT - SROA that uses DominatorTree.
158 struct SROA_DT : public SROA {
161 SROA_DT(int T = -1) : SROA(T, true, ID) {
162 initializeSROA_DTPass(*PassRegistry::getPassRegistry());
165 // getAnalysisUsage - This pass does not require any passes, but we know it
166 // will not alter the CFG, so say so.
167 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
168 AU.addRequired<DominatorTree>();
169 AU.setPreservesCFG();
173 // SROA_SSAUp - SROA that uses SSAUpdater.
174 struct SROA_SSAUp : public SROA {
177 SROA_SSAUp(int T = -1) : SROA(T, false, ID) {
178 initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry());
181 // getAnalysisUsage - This pass does not require any passes, but we know it
182 // will not alter the CFG, so say so.
183 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
184 AU.setPreservesCFG();
190 char SROA_DT::ID = 0;
191 char SROA_SSAUp::ID = 0;
193 INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl",
194 "Scalar Replacement of Aggregates (DT)", false, false)
195 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
196 INITIALIZE_PASS_END(SROA_DT, "scalarrepl",
197 "Scalar Replacement of Aggregates (DT)", false, false)
199 INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa",
200 "Scalar Replacement of Aggregates (SSAUp)", false, false)
201 INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa",
202 "Scalar Replacement of Aggregates (SSAUp)", false, false)
204 // Public interface to the ScalarReplAggregates pass
205 FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold,
208 return new SROA_DT(Threshold);
209 return new SROA_SSAUp(Threshold);
213 //===----------------------------------------------------------------------===//
214 // Convert To Scalar Optimization.
215 //===----------------------------------------------------------------------===//
218 /// ConvertToScalarInfo - This class implements the "Convert To Scalar"
219 /// optimization, which scans the uses of an alloca and determines if it can
220 /// rewrite it in terms of a single new alloca that can be mem2reg'd.
221 class ConvertToScalarInfo {
222 /// AllocaSize - The size of the alloca being considered.
224 const TargetData &TD;
226 /// IsNotTrivial - This is set to true if there is some access to the object
227 /// which means that mem2reg can't promote it.
230 /// VectorTy - This tracks the type that we should promote the vector to if
231 /// it is possible to turn it into a vector. This starts out null, and if it
232 /// isn't possible to turn into a vector type, it gets set to VoidTy.
233 const Type *VectorTy;
235 /// HadAVector - True if there is at least one vector access to the alloca.
236 /// We don't want to turn random arrays into vectors and use vector element
237 /// insert/extract, but if there are element accesses to something that is
238 /// also declared as a vector, we do want to promote to a vector.
242 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td)
243 : AllocaSize(Size), TD(td) {
244 IsNotTrivial = false;
249 AllocaInst *TryConvert(AllocaInst *AI);
252 bool CanConvertToScalar(Value *V, uint64_t Offset);
253 void MergeInType(const Type *In, uint64_t Offset);
254 bool MergeInVectorType(const VectorType *VInTy, uint64_t Offset);
255 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset);
257 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType,
258 uint64_t Offset, IRBuilder<> &Builder);
259 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal,
260 uint64_t Offset, IRBuilder<> &Builder);
262 } // end anonymous namespace.
265 /// TryConvert - Analyze the specified alloca, and if it is safe to do so,
266 /// rewrite it to be a new alloca which is mem2reg'able. This returns the new
267 /// alloca if possible or null if not.
268 AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) {
269 // If we can't convert this scalar, or if mem2reg can trivially do it, bail
271 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial)
274 // If we were able to find a vector type that can handle this with
275 // insert/extract elements, and if there was at least one use that had
276 // a vector type, promote this to a vector. We don't want to promote
277 // random stuff that doesn't use vectors (e.g. <9 x double>) because then
278 // we just get a lot of insert/extracts. If at least one vector is
279 // involved, then we probably really do have a union of vector/array.
281 if (VectorTy && VectorTy->isVectorTy() && HadAVector) {
282 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = "
283 << *VectorTy << '\n');
284 NewTy = VectorTy; // Use the vector type.
286 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n");
287 // Create and insert the integer alloca.
288 NewTy = IntegerType::get(AI->getContext(), AllocaSize*8);
290 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin());
291 ConvertUsesToScalar(AI, NewAI, 0);
295 /// MergeInType - Add the 'In' type to the accumulated vector type (VectorTy)
296 /// so far at the offset specified by Offset (which is specified in bytes).
298 /// There are two cases we handle here:
299 /// 1) A union of vector types of the same size and potentially its elements.
300 /// Here we turn element accesses into insert/extract element operations.
301 /// This promotes a <4 x float> with a store of float to the third element
302 /// into a <4 x float> that uses insert element.
303 /// 2) A fully general blob of memory, which we turn into some (potentially
304 /// large) integer type with extract and insert operations where the loads
305 /// and stores would mutate the memory. We mark this by setting VectorTy
307 void ConvertToScalarInfo::MergeInType(const Type *In, uint64_t Offset) {
308 // If we already decided to turn this into a blob of integer memory, there is
309 // nothing to be done.
310 if (VectorTy && VectorTy->isVoidTy())
313 // If this could be contributing to a vector, analyze it.
315 // If the In type is a vector that is the same size as the alloca, see if it
316 // matches the existing VecTy.
317 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) {
318 if (MergeInVectorType(VInTy, Offset))
320 } else if (In->isFloatTy() || In->isDoubleTy() ||
321 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 &&
322 isPowerOf2_32(In->getPrimitiveSizeInBits()))) {
323 // If we're accessing something that could be an element of a vector, see
324 // if the implied vector agrees with what we already have and if Offset is
325 // compatible with it.
326 unsigned EltSize = In->getPrimitiveSizeInBits()/8;
327 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 &&
329 cast<VectorType>(VectorTy)->getElementType()
330 ->getPrimitiveSizeInBits()/8 == EltSize)) {
332 VectorTy = VectorType::get(In, AllocaSize/EltSize);
337 // Otherwise, we have a case that we can't handle with an optimized vector
338 // form. We can still turn this into a large integer.
339 VectorTy = Type::getVoidTy(In->getContext());
342 /// MergeInVectorType - Handles the vector case of MergeInType, returning true
343 /// if the type was successfully merged and false otherwise.
344 bool ConvertToScalarInfo::MergeInVectorType(const VectorType *VInTy,
346 // Remember if we saw a vector type.
349 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) {
350 // If we're storing/loading a vector of the right size, allow it as a
351 // vector. If this the first vector we see, remember the type so that
352 // we know the element size. If this is a subsequent access, ignore it
353 // even if it is a differing type but the same size. Worst case we can
354 // bitcast the resultant vectors.
363 /// CanConvertToScalar - V is a pointer. If we can convert the pointee and all
364 /// its accesses to a single vector type, return true and set VecTy to
365 /// the new type. If we could convert the alloca into a single promotable
366 /// integer, return true but set VecTy to VoidTy. Further, if the use is not a
367 /// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset
368 /// is the current offset from the base of the alloca being analyzed.
370 /// If we see at least one access to the value that is as a vector type, set the
372 bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) {
373 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
374 Instruction *User = cast<Instruction>(*UI);
376 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
377 // Don't break volatile loads.
378 if (LI->isVolatile())
380 // Don't touch MMX operations.
381 if (LI->getType()->isX86_MMXTy())
383 MergeInType(LI->getType(), Offset);
387 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
388 // Storing the pointer, not into the value?
389 if (SI->getOperand(0) == V || SI->isVolatile()) return false;
390 // Don't touch MMX operations.
391 if (SI->getOperand(0)->getType()->isX86_MMXTy())
393 MergeInType(SI->getOperand(0)->getType(), Offset);
397 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
398 IsNotTrivial = true; // Can't be mem2reg'd.
399 if (!CanConvertToScalar(BCI, Offset))
404 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
405 // If this is a GEP with a variable indices, we can't handle it.
406 if (!GEP->hasAllConstantIndices())
409 // Compute the offset that this GEP adds to the pointer.
410 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
411 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(),
412 &Indices[0], Indices.size());
413 // See if all uses can be converted.
414 if (!CanConvertToScalar(GEP, Offset+GEPOffset))
416 IsNotTrivial = true; // Can't be mem2reg'd.
420 // If this is a constant sized memset of a constant value (e.g. 0) we can
422 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
423 // Store of constant value and constant size.
424 if (!isa<ConstantInt>(MSI->getValue()) ||
425 !isa<ConstantInt>(MSI->getLength()))
427 IsNotTrivial = true; // Can't be mem2reg'd.
431 // If this is a memcpy or memmove into or out of the whole allocation, we
432 // can handle it like a load or store of the scalar type.
433 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
434 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength());
435 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0)
438 IsNotTrivial = true; // Can't be mem2reg'd.
442 // Otherwise, we cannot handle this!
449 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
450 /// directly. This happens when we are converting an "integer union" to a
451 /// single integer scalar, or when we are converting a "vector union" to a
452 /// vector with insert/extractelement instructions.
454 /// Offset is an offset from the original alloca, in bits that need to be
455 /// shifted to the right. By the end of this, there should be no uses of Ptr.
456 void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI,
458 while (!Ptr->use_empty()) {
459 Instruction *User = cast<Instruction>(Ptr->use_back());
461 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
462 ConvertUsesToScalar(CI, NewAI, Offset);
463 CI->eraseFromParent();
467 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
468 // Compute the offset that this GEP adds to the pointer.
469 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
470 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(),
471 &Indices[0], Indices.size());
472 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8);
473 GEP->eraseFromParent();
477 IRBuilder<> Builder(User);
479 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
480 // The load is a bit extract from NewAI shifted right by Offset bits.
481 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp");
483 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder);
484 LI->replaceAllUsesWith(NewLoadVal);
485 LI->eraseFromParent();
489 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
490 assert(SI->getOperand(0) != Ptr && "Consistency error!");
491 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
492 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset,
494 Builder.CreateStore(New, NewAI);
495 SI->eraseFromParent();
497 // If the load we just inserted is now dead, then the inserted store
498 // overwrote the entire thing.
499 if (Old->use_empty())
500 Old->eraseFromParent();
504 // If this is a constant sized memset of a constant value (e.g. 0) we can
505 // transform it into a store of the expanded constant value.
506 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
507 assert(MSI->getRawDest() == Ptr && "Consistency error!");
508 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
510 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue();
512 // Compute the value replicated the right number of times.
513 APInt APVal(NumBytes*8, Val);
515 // Splat the value if non-zero.
517 for (unsigned i = 1; i != NumBytes; ++i)
520 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
521 Value *New = ConvertScalar_InsertValue(
522 ConstantInt::get(User->getContext(), APVal),
523 Old, Offset, Builder);
524 Builder.CreateStore(New, NewAI);
526 // If the load we just inserted is now dead, then the memset overwrote
528 if (Old->use_empty())
529 Old->eraseFromParent();
531 MSI->eraseFromParent();
535 // If this is a memcpy or memmove into or out of the whole allocation, we
536 // can handle it like a load or store of the scalar type.
537 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
538 assert(Offset == 0 && "must be store to start of alloca");
540 // If the source and destination are both to the same alloca, then this is
541 // a noop copy-to-self, just delete it. Otherwise, emit a load and store
543 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &TD, 0));
545 if (GetUnderlyingObject(MTI->getSource(), &TD, 0) != OrigAI) {
546 // Dest must be OrigAI, change this to be a load from the original
547 // pointer (bitcasted), then a store to our new alloca.
548 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?");
549 Value *SrcPtr = MTI->getSource();
550 const PointerType* SPTy = cast<PointerType>(SrcPtr->getType());
551 const PointerType* AIPTy = cast<PointerType>(NewAI->getType());
552 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
553 AIPTy = PointerType::get(AIPTy->getElementType(),
554 SPTy->getAddressSpace());
556 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy);
558 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval");
559 SrcVal->setAlignment(MTI->getAlignment());
560 Builder.CreateStore(SrcVal, NewAI);
561 } else if (GetUnderlyingObject(MTI->getDest(), &TD, 0) != OrigAI) {
562 // Src must be OrigAI, change this to be a load from NewAI then a store
563 // through the original dest pointer (bitcasted).
564 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?");
565 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval");
567 const PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType());
568 const PointerType* AIPTy = cast<PointerType>(NewAI->getType());
569 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
570 AIPTy = PointerType::get(AIPTy->getElementType(),
571 DPTy->getAddressSpace());
573 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy);
575 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr);
576 NewStore->setAlignment(MTI->getAlignment());
578 // Noop transfer. Src == Dst
581 MTI->eraseFromParent();
585 llvm_unreachable("Unsupported operation!");
589 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer
590 /// or vector value FromVal, extracting the bits from the offset specified by
591 /// Offset. This returns the value, which is of type ToType.
593 /// This happens when we are converting an "integer union" to a single
594 /// integer scalar, or when we are converting a "vector union" to a vector with
595 /// insert/extractelement instructions.
597 /// Offset is an offset from the original alloca, in bits that need to be
598 /// shifted to the right.
599 Value *ConvertToScalarInfo::
600 ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType,
601 uint64_t Offset, IRBuilder<> &Builder) {
602 // If the load is of the whole new alloca, no conversion is needed.
603 if (FromVal->getType() == ToType && Offset == 0)
606 // If the result alloca is a vector type, this is either an element
607 // access or a bitcast to another vector type of the same size.
608 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) {
609 if (ToType->isVectorTy())
610 return Builder.CreateBitCast(FromVal, ToType, "tmp");
612 // Otherwise it must be an element access.
615 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType());
616 Elt = Offset/EltSize;
617 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
619 // Return the element extracted out of it.
620 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get(
621 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp");
622 if (V->getType() != ToType)
623 V = Builder.CreateBitCast(V, ToType, "tmp");
627 // If ToType is a first class aggregate, extract out each of the pieces and
628 // use insertvalue's to form the FCA.
629 if (const StructType *ST = dyn_cast<StructType>(ToType)) {
630 const StructLayout &Layout = *TD.getStructLayout(ST);
631 Value *Res = UndefValue::get(ST);
632 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
633 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i),
634 Offset+Layout.getElementOffsetInBits(i),
636 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
641 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) {
642 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType());
643 Value *Res = UndefValue::get(AT);
644 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
645 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(),
646 Offset+i*EltSize, Builder);
647 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
652 // Otherwise, this must be a union that was converted to an integer value.
653 const IntegerType *NTy = cast<IntegerType>(FromVal->getType());
655 // If this is a big-endian system and the load is narrower than the
656 // full alloca type, we need to do a shift to get the right bits.
658 if (TD.isBigEndian()) {
659 // On big-endian machines, the lowest bit is stored at the bit offset
660 // from the pointer given by getTypeStoreSizeInBits. This matters for
661 // integers with a bitwidth that is not a multiple of 8.
662 ShAmt = TD.getTypeStoreSizeInBits(NTy) -
663 TD.getTypeStoreSizeInBits(ToType) - Offset;
668 // Note: we support negative bitwidths (with shl) which are not defined.
669 // We do this to support (f.e.) loads off the end of a structure where
670 // only some bits are used.
671 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth())
672 FromVal = Builder.CreateLShr(FromVal,
673 ConstantInt::get(FromVal->getType(),
675 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
676 FromVal = Builder.CreateShl(FromVal,
677 ConstantInt::get(FromVal->getType(),
680 // Finally, unconditionally truncate the integer to the right width.
681 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType);
682 if (LIBitWidth < NTy->getBitWidth())
684 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(),
686 else if (LIBitWidth > NTy->getBitWidth())
688 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(),
691 // If the result is an integer, this is a trunc or bitcast.
692 if (ToType->isIntegerTy()) {
694 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) {
695 // Just do a bitcast, we know the sizes match up.
696 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp");
698 // Otherwise must be a pointer.
699 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp");
701 assert(FromVal->getType() == ToType && "Didn't convert right?");
705 /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer
706 /// or vector value "Old" at the offset specified by Offset.
708 /// This happens when we are converting an "integer union" to a
709 /// single integer scalar, or when we are converting a "vector union" to a
710 /// vector with insert/extractelement instructions.
712 /// Offset is an offset from the original alloca, in bits that need to be
713 /// shifted to the right.
714 Value *ConvertToScalarInfo::
715 ConvertScalar_InsertValue(Value *SV, Value *Old,
716 uint64_t Offset, IRBuilder<> &Builder) {
717 // Convert the stored type to the actual type, shift it left to insert
718 // then 'or' into place.
719 const Type *AllocaType = Old->getType();
720 LLVMContext &Context = Old->getContext();
722 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
723 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy);
724 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType());
726 // Changing the whole vector with memset or with an access of a different
728 if (ValSize == VecSize)
729 return Builder.CreateBitCast(SV, AllocaType, "tmp");
731 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType());
733 // Must be an element insertion.
734 unsigned Elt = Offset/EltSize;
736 if (SV->getType() != VTy->getElementType())
737 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp");
739 SV = Builder.CreateInsertElement(Old, SV,
740 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt),
745 // If SV is a first-class aggregate value, insert each value recursively.
746 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) {
747 const StructLayout &Layout = *TD.getStructLayout(ST);
748 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
749 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
750 Old = ConvertScalar_InsertValue(Elt, Old,
751 Offset+Layout.getElementOffsetInBits(i),
757 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) {
758 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType());
759 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
760 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
761 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder);
766 // If SV is a float, convert it to the appropriate integer type.
767 // If it is a pointer, do the same.
768 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType());
769 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType);
770 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType());
771 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType);
772 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy())
773 SV = Builder.CreateBitCast(SV,
774 IntegerType::get(SV->getContext(),SrcWidth), "tmp");
775 else if (SV->getType()->isPointerTy())
776 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp");
778 // Zero extend or truncate the value if needed.
779 if (SV->getType() != AllocaType) {
780 if (SV->getType()->getPrimitiveSizeInBits() <
781 AllocaType->getPrimitiveSizeInBits())
782 SV = Builder.CreateZExt(SV, AllocaType, "tmp");
784 // Truncation may be needed if storing more than the alloca can hold
785 // (undefined behavior).
786 SV = Builder.CreateTrunc(SV, AllocaType, "tmp");
787 SrcWidth = DestWidth;
788 SrcStoreWidth = DestStoreWidth;
792 // If this is a big-endian system and the store is narrower than the
793 // full alloca type, we need to do a shift to get the right bits.
795 if (TD.isBigEndian()) {
796 // On big-endian machines, the lowest bit is stored at the bit offset
797 // from the pointer given by getTypeStoreSizeInBits. This matters for
798 // integers with a bitwidth that is not a multiple of 8.
799 ShAmt = DestStoreWidth - SrcStoreWidth - Offset;
804 // Note: we support negative bitwidths (with shr) which are not defined.
805 // We do this to support (f.e.) stores off the end of a structure where
806 // only some bits in the structure are set.
807 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
808 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
809 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(),
812 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
813 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(),
815 Mask = Mask.lshr(-ShAmt);
818 // Mask out the bits we are about to insert from the old value, and or
820 if (SrcWidth != DestWidth) {
821 assert(DestWidth > SrcWidth);
822 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask");
823 SV = Builder.CreateOr(Old, SV, "ins");
829 //===----------------------------------------------------------------------===//
831 //===----------------------------------------------------------------------===//
834 bool SROA::runOnFunction(Function &F) {
835 TD = getAnalysisIfAvailable<TargetData>();
837 bool Changed = performPromotion(F);
839 // FIXME: ScalarRepl currently depends on TargetData more than it
840 // theoretically needs to. It should be refactored in order to support
841 // target-independent IR. Until this is done, just skip the actual
842 // scalar-replacement portion of this pass.
843 if (!TD) return Changed;
846 bool LocalChange = performScalarRepl(F);
847 if (!LocalChange) break; // No need to repromote if no scalarrepl
849 LocalChange = performPromotion(F);
850 if (!LocalChange) break; // No need to re-scalarrepl if no promotion
857 class AllocaPromoter : public LoadAndStorePromoter {
860 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S)
861 : LoadAndStorePromoter(Insts, S), AI(0) {}
863 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) {
864 // Remember which alloca we're promoting (for isInstInList).
866 LoadAndStorePromoter::run(Insts);
867 AI->eraseFromParent();
870 virtual bool isInstInList(Instruction *I,
871 const SmallVectorImpl<Instruction*> &Insts) const {
872 if (LoadInst *LI = dyn_cast<LoadInst>(I))
873 return LI->getOperand(0) == AI;
874 return cast<StoreInst>(I)->getPointerOperand() == AI;
877 } // end anon namespace
879 /// isSafeSelectToSpeculate - Select instructions that use an alloca and are
880 /// subsequently loaded can be rewritten to load both input pointers and then
881 /// select between the result, allowing the load of the alloca to be promoted.
883 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
884 /// %V = load i32* %P2
886 /// %V1 = load i32* %Alloca -> will be mem2reg'd
887 /// %V2 = load i32* %Other
888 /// %V = select i1 %cond, i32 %V1, i32 %V2
890 /// We can do this to a select if its only uses are loads and if the operand to
891 /// the select can be loaded unconditionally.
892 static bool isSafeSelectToSpeculate(SelectInst *SI, const TargetData *TD) {
893 bool TDerefable = SI->getTrueValue()->isDereferenceablePointer();
894 bool FDerefable = SI->getFalseValue()->isDereferenceablePointer();
896 for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end();
898 LoadInst *LI = dyn_cast<LoadInst>(*UI);
899 if (LI == 0 || LI->isVolatile()) return false;
901 // Both operands to the select need to be dereferencable, either absolutely
902 // (e.g. allocas) or at this point because we can see other accesses to it.
903 if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI,
904 LI->getAlignment(), TD))
906 if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI,
907 LI->getAlignment(), TD))
914 /// isSafePHIToSpeculate - PHI instructions that use an alloca and are
915 /// subsequently loaded can be rewritten to load both input pointers in the pred
916 /// blocks and then PHI the results, allowing the load of the alloca to be
919 /// %P2 = phi [i32* %Alloca, i32* %Other]
920 /// %V = load i32* %P2
922 /// %V1 = load i32* %Alloca -> will be mem2reg'd
924 /// %V2 = load i32* %Other
926 /// %V = phi [i32 %V1, i32 %V2]
928 /// We can do this to a select if its only uses are loads and if the operand to
929 /// the select can be loaded unconditionally.
930 static bool isSafePHIToSpeculate(PHINode *PN, const TargetData *TD) {
931 // For now, we can only do this promotion if the load is in the same block as
932 // the PHI, and if there are no stores between the phi and load.
933 // TODO: Allow recursive phi users.
934 // TODO: Allow stores.
935 BasicBlock *BB = PN->getParent();
936 unsigned MaxAlign = 0;
937 for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end();
939 LoadInst *LI = dyn_cast<LoadInst>(*UI);
940 if (LI == 0 || LI->isVolatile()) return false;
942 // For now we only allow loads in the same block as the PHI. This is a
943 // common case that happens when instcombine merges two loads through a PHI.
944 if (LI->getParent() != BB) return false;
946 // Ensure that there are no instructions between the PHI and the load that
948 for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI)
949 if (BBI->mayWriteToMemory())
952 MaxAlign = std::max(MaxAlign, LI->getAlignment());
955 // Okay, we know that we have one or more loads in the same block as the PHI.
956 // We can transform this if it is safe to push the loads into the predecessor
957 // blocks. The only thing to watch out for is that we can't put a possibly
958 // trapping load in the predecessor if it is a critical edge.
959 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
960 BasicBlock *Pred = PN->getIncomingBlock(i);
962 // If the predecessor has a single successor, then the edge isn't critical.
963 if (Pred->getTerminator()->getNumSuccessors() == 1)
966 Value *InVal = PN->getIncomingValue(i);
968 // If the InVal is an invoke in the pred, we can't put a load on the edge.
969 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal))
970 if (II->getParent() == Pred)
973 // If this pointer is always safe to load, or if we can prove that there is
974 // already a load in the block, then we can move the load to the pred block.
975 if (InVal->isDereferenceablePointer() ||
976 isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, TD))
986 /// tryToMakeAllocaBePromotable - This returns true if the alloca only has
987 /// direct (non-volatile) loads and stores to it. If the alloca is close but
988 /// not quite there, this will transform the code to allow promotion. As such,
989 /// it is a non-pure predicate.
990 static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) {
991 SetVector<Instruction*, SmallVector<Instruction*, 4>,
992 SmallPtrSet<Instruction*, 4> > InstsToRewrite;
994 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
997 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
998 if (LI->isVolatile())
1003 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1004 if (SI->getOperand(0) == AI || SI->isVolatile())
1005 return false; // Don't allow a store OF the AI, only INTO the AI.
1009 if (SelectInst *SI = dyn_cast<SelectInst>(U)) {
1010 // If the condition being selected on is a constant, fold the select, yes
1011 // this does (rarely) happen early on.
1012 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) {
1013 Value *Result = SI->getOperand(1+CI->isZero());
1014 SI->replaceAllUsesWith(Result);
1015 SI->eraseFromParent();
1017 // This is very rare and we just scrambled the use list of AI, start
1019 return tryToMakeAllocaBePromotable(AI, TD);
1022 // If it is safe to turn "load (select c, AI, ptr)" into a select of two
1023 // loads, then we can transform this by rewriting the select.
1024 if (!isSafeSelectToSpeculate(SI, TD))
1027 InstsToRewrite.insert(SI);
1031 if (PHINode *PN = dyn_cast<PHINode>(U)) {
1032 if (PN->use_empty()) { // Dead PHIs can be stripped.
1033 InstsToRewrite.insert(PN);
1037 // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads
1038 // in the pred blocks, then we can transform this by rewriting the PHI.
1039 if (!isSafePHIToSpeculate(PN, TD))
1042 InstsToRewrite.insert(PN);
1049 // If there are no instructions to rewrite, then all uses are load/stores and
1051 if (InstsToRewrite.empty())
1054 // If we have instructions that need to be rewritten for this to be promotable
1055 // take care of it now.
1056 for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) {
1057 if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) {
1058 // Selects in InstsToRewrite only have load uses. Rewrite each as two
1059 // loads with a new select.
1060 while (!SI->use_empty()) {
1061 LoadInst *LI = cast<LoadInst>(SI->use_back());
1063 IRBuilder<> Builder(LI);
1064 LoadInst *TrueLoad =
1065 Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t");
1066 LoadInst *FalseLoad =
1067 Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".t");
1069 // Transfer alignment and TBAA info if present.
1070 TrueLoad->setAlignment(LI->getAlignment());
1071 FalseLoad->setAlignment(LI->getAlignment());
1072 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
1073 TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1074 FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1077 Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad);
1079 LI->replaceAllUsesWith(V);
1080 LI->eraseFromParent();
1083 // Now that all the loads are gone, the select is gone too.
1084 SI->eraseFromParent();
1088 // Otherwise, we have a PHI node which allows us to push the loads into the
1090 PHINode *PN = cast<PHINode>(InstsToRewrite[i]);
1091 if (PN->use_empty()) {
1092 PN->eraseFromParent();
1096 const Type *LoadTy = cast<PointerType>(PN->getType())->getElementType();
1097 PHINode *NewPN = PHINode::Create(LoadTy, PN->getName()+".ld", PN);
1099 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1100 // matter which one we get and if any differ, it doesn't matter.
1101 LoadInst *SomeLoad = cast<LoadInst>(PN->use_back());
1102 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1103 unsigned Align = SomeLoad->getAlignment();
1105 // Rewrite all loads of the PN to use the new PHI.
1106 while (!PN->use_empty()) {
1107 LoadInst *LI = cast<LoadInst>(PN->use_back());
1108 LI->replaceAllUsesWith(NewPN);
1109 LI->eraseFromParent();
1112 // Inject loads into all of the pred blocks. Keep track of which blocks we
1113 // insert them into in case we have multiple edges from the same block.
1114 DenseMap<BasicBlock*, LoadInst*> InsertedLoads;
1116 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1117 BasicBlock *Pred = PN->getIncomingBlock(i);
1118 LoadInst *&Load = InsertedLoads[Pred];
1120 Load = new LoadInst(PN->getIncomingValue(i),
1121 PN->getName() + "." + Pred->getName(),
1122 Pred->getTerminator());
1123 Load->setAlignment(Align);
1124 if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1127 NewPN->addIncoming(Load, Pred);
1130 PN->eraseFromParent();
1138 bool SROA::performPromotion(Function &F) {
1139 std::vector<AllocaInst*> Allocas;
1140 DominatorTree *DT = 0;
1142 DT = &getAnalysis<DominatorTree>();
1144 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
1146 bool Changed = false;
1147 SmallVector<Instruction*, 64> Insts;
1151 // Find allocas that are safe to promote, by looking at all instructions in
1153 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
1154 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
1155 if (tryToMakeAllocaBePromotable(AI, TD))
1156 Allocas.push_back(AI);
1158 if (Allocas.empty()) break;
1161 PromoteMemToReg(Allocas, *DT);
1164 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
1165 AllocaInst *AI = Allocas[i];
1167 // Build list of instructions to promote.
1168 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
1170 Insts.push_back(cast<Instruction>(*UI));
1172 AllocaPromoter(Insts, SSA).run(AI, Insts);
1176 NumPromoted += Allocas.size();
1184 /// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for
1185 /// SROA. It must be a struct or array type with a small number of elements.
1186 static bool ShouldAttemptScalarRepl(AllocaInst *AI) {
1187 const Type *T = AI->getAllocatedType();
1188 // Do not promote any struct into more than 32 separate vars.
1189 if (const StructType *ST = dyn_cast<StructType>(T))
1190 return ST->getNumElements() <= 32;
1191 // Arrays are much less likely to be safe for SROA; only consider
1192 // them if they are very small.
1193 if (const ArrayType *AT = dyn_cast<ArrayType>(T))
1194 return AT->getNumElements() <= 8;
1199 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
1200 // which runs on all of the malloc/alloca instructions in the function, removing
1201 // them if they are only used by getelementptr instructions.
1203 bool SROA::performScalarRepl(Function &F) {
1204 std::vector<AllocaInst*> WorkList;
1206 // Scan the entry basic block, adding allocas to the worklist.
1207 BasicBlock &BB = F.getEntryBlock();
1208 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
1209 if (AllocaInst *A = dyn_cast<AllocaInst>(I))
1210 WorkList.push_back(A);
1212 // Process the worklist
1213 bool Changed = false;
1214 while (!WorkList.empty()) {
1215 AllocaInst *AI = WorkList.back();
1216 WorkList.pop_back();
1218 // Handle dead allocas trivially. These can be formed by SROA'ing arrays
1219 // with unused elements.
1220 if (AI->use_empty()) {
1221 AI->eraseFromParent();
1226 // If this alloca is impossible for us to promote, reject it early.
1227 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized())
1230 // Check to see if this allocation is only modified by a memcpy/memmove from
1231 // a constant global. If this is the case, we can change all users to use
1232 // the constant global instead. This is commonly produced by the CFE by
1233 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
1234 // is only subsequently read.
1235 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) {
1236 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n');
1237 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n');
1238 Constant *TheSrc = cast<Constant>(TheCopy->getSource());
1239 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType()));
1240 TheCopy->eraseFromParent(); // Don't mutate the global.
1241 AI->eraseFromParent();
1247 // Check to see if we can perform the core SROA transformation. We cannot
1248 // transform the allocation instruction if it is an array allocation
1249 // (allocations OF arrays are ok though), and an allocation of a scalar
1250 // value cannot be decomposed at all.
1251 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType());
1253 // Do not promote [0 x %struct].
1254 if (AllocaSize == 0) continue;
1256 // Do not promote any struct whose size is too big.
1257 if (AllocaSize > SRThreshold) continue;
1259 // If the alloca looks like a good candidate for scalar replacement, and if
1260 // all its users can be transformed, then split up the aggregate into its
1261 // separate elements.
1262 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) {
1263 DoScalarReplacement(AI, WorkList);
1268 // If we can turn this aggregate value (potentially with casts) into a
1269 // simple scalar value that can be mem2reg'd into a register value.
1270 // IsNotTrivial tracks whether this is something that mem2reg could have
1271 // promoted itself. If so, we don't want to transform it needlessly. Note
1272 // that we can't just check based on the type: the alloca may be of an i32
1273 // but that has pointer arithmetic to set byte 3 of it or something.
1274 if (AllocaInst *NewAI =
1275 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) {
1276 NewAI->takeName(AI);
1277 AI->eraseFromParent();
1283 // Otherwise, couldn't process this alloca.
1289 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl
1290 /// predicate, do SROA now.
1291 void SROA::DoScalarReplacement(AllocaInst *AI,
1292 std::vector<AllocaInst*> &WorkList) {
1293 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n');
1294 SmallVector<AllocaInst*, 32> ElementAllocas;
1295 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
1296 ElementAllocas.reserve(ST->getNumContainedTypes());
1297 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
1298 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
1300 AI->getName() + "." + Twine(i), AI);
1301 ElementAllocas.push_back(NA);
1302 WorkList.push_back(NA); // Add to worklist for recursive processing
1305 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
1306 ElementAllocas.reserve(AT->getNumElements());
1307 const Type *ElTy = AT->getElementType();
1308 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1309 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
1310 AI->getName() + "." + Twine(i), AI);
1311 ElementAllocas.push_back(NA);
1312 WorkList.push_back(NA); // Add to worklist for recursive processing
1316 // Now that we have created the new alloca instructions, rewrite all the
1317 // uses of the old alloca.
1318 RewriteForScalarRepl(AI, AI, 0, ElementAllocas);
1320 // Now erase any instructions that were made dead while rewriting the alloca.
1321 DeleteDeadInstructions();
1322 AI->eraseFromParent();
1327 /// DeleteDeadInstructions - Erase instructions on the DeadInstrs list,
1328 /// recursively including all their operands that become trivially dead.
1329 void SROA::DeleteDeadInstructions() {
1330 while (!DeadInsts.empty()) {
1331 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val());
1333 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
1334 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
1335 // Zero out the operand and see if it becomes trivially dead.
1336 // (But, don't add allocas to the dead instruction list -- they are
1337 // already on the worklist and will be deleted separately.)
1339 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U))
1340 DeadInsts.push_back(U);
1343 I->eraseFromParent();
1347 /// isSafeForScalarRepl - Check if instruction I is a safe use with regard to
1348 /// performing scalar replacement of alloca AI. The results are flagged in
1349 /// the Info parameter. Offset indicates the position within AI that is
1350 /// referenced by this instruction.
1351 void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset,
1353 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
1354 Instruction *User = cast<Instruction>(*UI);
1356 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1357 isSafeForScalarRepl(BC, Offset, Info);
1358 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1359 uint64_t GEPOffset = Offset;
1360 isSafeGEP(GEPI, GEPOffset, Info);
1362 isSafeForScalarRepl(GEPI, GEPOffset, Info);
1363 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
1364 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
1366 return MarkUnsafe(Info, User);
1367 isSafeMemAccess(Offset, Length->getZExtValue(), 0,
1368 UI.getOperandNo() == 0, Info, MI,
1369 true /*AllowWholeAccess*/);
1370 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1371 if (LI->isVolatile())
1372 return MarkUnsafe(Info, User);
1373 const Type *LIType = LI->getType();
1374 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType),
1375 LIType, false, Info, LI, true /*AllowWholeAccess*/);
1376 Info.hasALoadOrStore = true;
1378 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1379 // Store is ok if storing INTO the pointer, not storing the pointer
1380 if (SI->isVolatile() || SI->getOperand(0) == I)
1381 return MarkUnsafe(Info, User);
1383 const Type *SIType = SI->getOperand(0)->getType();
1384 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType),
1385 SIType, true, Info, SI, true /*AllowWholeAccess*/);
1386 Info.hasALoadOrStore = true;
1387 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) {
1388 isSafePHISelectUseForScalarRepl(User, Offset, Info);
1390 return MarkUnsafe(Info, User);
1392 if (Info.isUnsafe) return;
1397 /// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer
1398 /// derived from the alloca, we can often still split the alloca into elements.
1399 /// This is useful if we have a large alloca where one element is phi'd
1400 /// together somewhere: we can SRoA and promote all the other elements even if
1401 /// we end up not being able to promote this one.
1403 /// All we require is that the uses of the PHI do not index into other parts of
1404 /// the alloca. The most important use case for this is single load and stores
1405 /// that are PHI'd together, which can happen due to code sinking.
1406 void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset,
1408 // If we've already checked this PHI, don't do it again.
1409 if (PHINode *PN = dyn_cast<PHINode>(I))
1410 if (!Info.CheckedPHIs.insert(PN))
1413 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
1414 Instruction *User = cast<Instruction>(*UI);
1416 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1417 isSafePHISelectUseForScalarRepl(BC, Offset, Info);
1418 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1419 // Only allow "bitcast" GEPs for simplicity. We could generalize this,
1420 // but would have to prove that we're staying inside of an element being
1422 if (!GEPI->hasAllZeroIndices())
1423 return MarkUnsafe(Info, User);
1424 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info);
1425 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1426 if (LI->isVolatile())
1427 return MarkUnsafe(Info, User);
1428 const Type *LIType = LI->getType();
1429 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType),
1430 LIType, false, Info, LI, false /*AllowWholeAccess*/);
1431 Info.hasALoadOrStore = true;
1433 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1434 // Store is ok if storing INTO the pointer, not storing the pointer
1435 if (SI->isVolatile() || SI->getOperand(0) == I)
1436 return MarkUnsafe(Info, User);
1438 const Type *SIType = SI->getOperand(0)->getType();
1439 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType),
1440 SIType, true, Info, SI, false /*AllowWholeAccess*/);
1441 Info.hasALoadOrStore = true;
1442 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) {
1443 isSafePHISelectUseForScalarRepl(User, Offset, Info);
1445 return MarkUnsafe(Info, User);
1447 if (Info.isUnsafe) return;
1451 /// isSafeGEP - Check if a GEP instruction can be handled for scalar
1452 /// replacement. It is safe when all the indices are constant, in-bounds
1453 /// references, and when the resulting offset corresponds to an element within
1454 /// the alloca type. The results are flagged in the Info parameter. Upon
1455 /// return, Offset is adjusted as specified by the GEP indices.
1456 void SROA::isSafeGEP(GetElementPtrInst *GEPI,
1457 uint64_t &Offset, AllocaInfo &Info) {
1458 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI);
1462 // Walk through the GEP type indices, checking the types that this indexes
1464 for (; GEPIt != E; ++GEPIt) {
1465 // Ignore struct elements, no extra checking needed for these.
1466 if ((*GEPIt)->isStructTy())
1469 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand());
1471 return MarkUnsafe(Info, GEPI);
1474 // Compute the offset due to this GEP and check if the alloca has a
1475 // component element at that offset.
1476 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
1477 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(),
1478 &Indices[0], Indices.size());
1479 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0))
1480 MarkUnsafe(Info, GEPI);
1483 /// isHomogeneousAggregate - Check if type T is a struct or array containing
1484 /// elements of the same type (which is always true for arrays). If so,
1485 /// return true with NumElts and EltTy set to the number of elements and the
1486 /// element type, respectively.
1487 static bool isHomogeneousAggregate(const Type *T, unsigned &NumElts,
1488 const Type *&EltTy) {
1489 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) {
1490 NumElts = AT->getNumElements();
1491 EltTy = (NumElts == 0 ? 0 : AT->getElementType());
1494 if (const StructType *ST = dyn_cast<StructType>(T)) {
1495 NumElts = ST->getNumContainedTypes();
1496 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0));
1497 for (unsigned n = 1; n < NumElts; ++n) {
1498 if (ST->getContainedType(n) != EltTy)
1506 /// isCompatibleAggregate - Check if T1 and T2 are either the same type or are
1507 /// "homogeneous" aggregates with the same element type and number of elements.
1508 static bool isCompatibleAggregate(const Type *T1, const Type *T2) {
1512 unsigned NumElts1, NumElts2;
1513 const Type *EltTy1, *EltTy2;
1514 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) &&
1515 isHomogeneousAggregate(T2, NumElts2, EltTy2) &&
1516 NumElts1 == NumElts2 &&
1523 /// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI
1524 /// alloca or has an offset and size that corresponds to a component element
1525 /// within it. The offset checked here may have been formed from a GEP with a
1526 /// pointer bitcasted to a different type.
1528 /// If AllowWholeAccess is true, then this allows uses of the entire alloca as a
1529 /// unit. If false, it only allows accesses known to be in a single element.
1530 void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize,
1531 const Type *MemOpType, bool isStore,
1532 AllocaInfo &Info, Instruction *TheAccess,
1533 bool AllowWholeAccess) {
1534 // Check if this is a load/store of the entire alloca.
1535 if (Offset == 0 && AllowWholeAccess &&
1536 MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) {
1537 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer
1538 // loads/stores (which are essentially the same as the MemIntrinsics with
1539 // regard to copying padding between elements). But, if an alloca is
1540 // flagged as both a source and destination of such operations, we'll need
1541 // to check later for padding between elements.
1542 if (!MemOpType || MemOpType->isIntegerTy()) {
1544 Info.isMemCpyDst = true;
1546 Info.isMemCpySrc = true;
1549 // This is also safe for references using a type that is compatible with
1550 // the type of the alloca, so that loads/stores can be rewritten using
1551 // insertvalue/extractvalue.
1552 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) {
1553 Info.hasSubelementAccess = true;
1557 // Check if the offset/size correspond to a component within the alloca type.
1558 const Type *T = Info.AI->getAllocatedType();
1559 if (TypeHasComponent(T, Offset, MemSize)) {
1560 Info.hasSubelementAccess = true;
1564 return MarkUnsafe(Info, TheAccess);
1567 /// TypeHasComponent - Return true if T has a component type with the
1568 /// specified offset and size. If Size is zero, do not check the size.
1569 bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) {
1572 if (const StructType *ST = dyn_cast<StructType>(T)) {
1573 const StructLayout *Layout = TD->getStructLayout(ST);
1574 unsigned EltIdx = Layout->getElementContainingOffset(Offset);
1575 EltTy = ST->getContainedType(EltIdx);
1576 EltSize = TD->getTypeAllocSize(EltTy);
1577 Offset -= Layout->getElementOffset(EltIdx);
1578 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) {
1579 EltTy = AT->getElementType();
1580 EltSize = TD->getTypeAllocSize(EltTy);
1581 if (Offset >= AT->getNumElements() * EltSize)
1587 if (Offset == 0 && (Size == 0 || EltSize == Size))
1589 // Check if the component spans multiple elements.
1590 if (Offset + Size > EltSize)
1592 return TypeHasComponent(EltTy, Offset, Size);
1595 /// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite
1596 /// the instruction I, which references it, to use the separate elements.
1597 /// Offset indicates the position within AI that is referenced by this
1599 void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
1600 SmallVector<AllocaInst*, 32> &NewElts) {
1601 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) {
1602 Use &TheUse = UI.getUse();
1603 Instruction *User = cast<Instruction>(*UI++);
1605 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1606 RewriteBitCast(BC, AI, Offset, NewElts);
1610 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1611 RewriteGEP(GEPI, AI, Offset, NewElts);
1615 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
1616 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
1617 uint64_t MemSize = Length->getZExtValue();
1619 MemSize == TD->getTypeAllocSize(AI->getAllocatedType()))
1620 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts);
1621 // Otherwise the intrinsic can only touch a single element and the
1622 // address operand will be updated, so nothing else needs to be done.
1626 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1627 const Type *LIType = LI->getType();
1629 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) {
1631 // %res = load { i32, i32 }* %alloc
1633 // %load.0 = load i32* %alloc.0
1634 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0
1635 // %load.1 = load i32* %alloc.1
1636 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1
1637 // (Also works for arrays instead of structs)
1638 Value *Insert = UndefValue::get(LIType);
1639 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1640 Value *Load = new LoadInst(NewElts[i], "load", LI);
1641 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI);
1643 LI->replaceAllUsesWith(Insert);
1644 DeadInsts.push_back(LI);
1645 } else if (LIType->isIntegerTy() &&
1646 TD->getTypeAllocSize(LIType) ==
1647 TD->getTypeAllocSize(AI->getAllocatedType())) {
1648 // If this is a load of the entire alloca to an integer, rewrite it.
1649 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts);
1654 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1655 Value *Val = SI->getOperand(0);
1656 const Type *SIType = Val->getType();
1657 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) {
1659 // store { i32, i32 } %val, { i32, i32 }* %alloc
1661 // %val.0 = extractvalue { i32, i32 } %val, 0
1662 // store i32 %val.0, i32* %alloc.0
1663 // %val.1 = extractvalue { i32, i32 } %val, 1
1664 // store i32 %val.1, i32* %alloc.1
1665 // (Also works for arrays instead of structs)
1666 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1667 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI);
1668 new StoreInst(Extract, NewElts[i], SI);
1670 DeadInsts.push_back(SI);
1671 } else if (SIType->isIntegerTy() &&
1672 TD->getTypeAllocSize(SIType) ==
1673 TD->getTypeAllocSize(AI->getAllocatedType())) {
1674 // If this is a store of the entire alloca from an integer, rewrite it.
1675 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts);
1680 if (isa<SelectInst>(User) || isa<PHINode>(User)) {
1681 // If we have a PHI user of the alloca itself (as opposed to a GEP or
1682 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to
1684 if (!isa<AllocaInst>(I)) continue;
1686 assert(Offset == 0 && NewElts[0] &&
1687 "Direct alloca use should have a zero offset");
1689 // If we have a use of the alloca, we know the derived uses will be
1690 // utilizing just the first element of the scalarized result. Insert a
1691 // bitcast of the first alloca before the user as required.
1692 AllocaInst *NewAI = NewElts[0];
1693 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI);
1694 NewAI->moveBefore(BCI);
1701 /// RewriteBitCast - Update a bitcast reference to the alloca being replaced
1702 /// and recursively continue updating all of its uses.
1703 void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
1704 SmallVector<AllocaInst*, 32> &NewElts) {
1705 RewriteForScalarRepl(BC, AI, Offset, NewElts);
1706 if (BC->getOperand(0) != AI)
1709 // The bitcast references the original alloca. Replace its uses with
1710 // references to the first new element alloca.
1711 Instruction *Val = NewElts[0];
1712 if (Val->getType() != BC->getDestTy()) {
1713 Val = new BitCastInst(Val, BC->getDestTy(), "", BC);
1716 BC->replaceAllUsesWith(Val);
1717 DeadInsts.push_back(BC);
1720 /// FindElementAndOffset - Return the index of the element containing Offset
1721 /// within the specified type, which must be either a struct or an array.
1722 /// Sets T to the type of the element and Offset to the offset within that
1723 /// element. IdxTy is set to the type of the index result to be used in a
1724 /// GEP instruction.
1725 uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset,
1726 const Type *&IdxTy) {
1728 if (const StructType *ST = dyn_cast<StructType>(T)) {
1729 const StructLayout *Layout = TD->getStructLayout(ST);
1730 Idx = Layout->getElementContainingOffset(Offset);
1731 T = ST->getContainedType(Idx);
1732 Offset -= Layout->getElementOffset(Idx);
1733 IdxTy = Type::getInt32Ty(T->getContext());
1736 const ArrayType *AT = cast<ArrayType>(T);
1737 T = AT->getElementType();
1738 uint64_t EltSize = TD->getTypeAllocSize(T);
1739 Idx = Offset / EltSize;
1740 Offset -= Idx * EltSize;
1741 IdxTy = Type::getInt64Ty(T->getContext());
1745 /// RewriteGEP - Check if this GEP instruction moves the pointer across
1746 /// elements of the alloca that are being split apart, and if so, rewrite
1747 /// the GEP to be relative to the new element.
1748 void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
1749 SmallVector<AllocaInst*, 32> &NewElts) {
1750 uint64_t OldOffset = Offset;
1751 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
1752 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(),
1753 &Indices[0], Indices.size());
1755 RewriteForScalarRepl(GEPI, AI, Offset, NewElts);
1757 const Type *T = AI->getAllocatedType();
1759 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy);
1760 if (GEPI->getOperand(0) == AI)
1761 OldIdx = ~0ULL; // Force the GEP to be rewritten.
1763 T = AI->getAllocatedType();
1764 uint64_t EltOffset = Offset;
1765 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy);
1767 // If this GEP does not move the pointer across elements of the alloca
1768 // being split, then it does not needs to be rewritten.
1772 const Type *i32Ty = Type::getInt32Ty(AI->getContext());
1773 SmallVector<Value*, 8> NewArgs;
1774 NewArgs.push_back(Constant::getNullValue(i32Ty));
1775 while (EltOffset != 0) {
1776 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy);
1777 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx));
1779 Instruction *Val = NewElts[Idx];
1780 if (NewArgs.size() > 1) {
1781 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(),
1782 NewArgs.end(), "", GEPI);
1783 Val->takeName(GEPI);
1785 if (Val->getType() != GEPI->getType())
1786 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI);
1787 GEPI->replaceAllUsesWith(Val);
1788 DeadInsts.push_back(GEPI);
1791 /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI.
1792 /// Rewrite it to copy or set the elements of the scalarized memory.
1793 void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
1795 SmallVector<AllocaInst*, 32> &NewElts) {
1796 // If this is a memcpy/memmove, construct the other pointer as the
1797 // appropriate type. The "Other" pointer is the pointer that goes to memory
1798 // that doesn't have anything to do with the alloca that we are promoting. For
1799 // memset, this Value* stays null.
1800 Value *OtherPtr = 0;
1801 unsigned MemAlignment = MI->getAlignment();
1802 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy
1803 if (Inst == MTI->getRawDest())
1804 OtherPtr = MTI->getRawSource();
1806 assert(Inst == MTI->getRawSource());
1807 OtherPtr = MTI->getRawDest();
1811 // If there is an other pointer, we want to convert it to the same pointer
1812 // type as AI has, so we can GEP through it safely.
1814 unsigned AddrSpace =
1815 cast<PointerType>(OtherPtr->getType())->getAddressSpace();
1817 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an
1818 // optimization, but it's also required to detect the corner case where
1819 // both pointer operands are referencing the same memory, and where
1820 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This
1821 // function is only called for mem intrinsics that access the whole
1822 // aggregate, so non-zero GEPs are not an issue here.)
1823 OtherPtr = OtherPtr->stripPointerCasts();
1825 // Copying the alloca to itself is a no-op: just delete it.
1826 if (OtherPtr == AI || OtherPtr == NewElts[0]) {
1827 // This code will run twice for a no-op memcpy -- once for each operand.
1828 // Put only one reference to MI on the DeadInsts list.
1829 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(),
1830 E = DeadInsts.end(); I != E; ++I)
1831 if (*I == MI) return;
1832 DeadInsts.push_back(MI);
1836 // If the pointer is not the right type, insert a bitcast to the right
1839 PointerType::get(AI->getType()->getElementType(), AddrSpace);
1841 if (OtherPtr->getType() != NewTy)
1842 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI);
1845 // Process each element of the aggregate.
1846 bool SROADest = MI->getRawDest() == Inst;
1848 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext()));
1850 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1851 // If this is a memcpy/memmove, emit a GEP of the other element address.
1852 Value *OtherElt = 0;
1853 unsigned OtherEltAlign = MemAlignment;
1856 Value *Idx[2] = { Zero,
1857 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) };
1858 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2,
1859 OtherPtr->getName()+"."+Twine(i),
1862 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType());
1863 const Type *OtherTy = OtherPtrTy->getElementType();
1864 if (const StructType *ST = dyn_cast<StructType>(OtherTy)) {
1865 EltOffset = TD->getStructLayout(ST)->getElementOffset(i);
1867 const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType();
1868 EltOffset = TD->getTypeAllocSize(EltTy)*i;
1871 // The alignment of the other pointer is the guaranteed alignment of the
1872 // element, which is affected by both the known alignment of the whole
1873 // mem intrinsic and the alignment of the element. If the alignment of
1874 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the
1875 // known alignment is just 4 bytes.
1876 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset);
1879 Value *EltPtr = NewElts[i];
1880 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType();
1882 // If we got down to a scalar, insert a load or store as appropriate.
1883 if (EltTy->isSingleValueType()) {
1884 if (isa<MemTransferInst>(MI)) {
1886 // From Other to Alloca.
1887 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI);
1888 new StoreInst(Elt, EltPtr, MI);
1890 // From Alloca to Other.
1891 Value *Elt = new LoadInst(EltPtr, "tmp", MI);
1892 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI);
1896 assert(isa<MemSetInst>(MI));
1898 // If the stored element is zero (common case), just store a null
1901 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) {
1903 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0>
1905 // If EltTy is a vector type, get the element type.
1906 const Type *ValTy = EltTy->getScalarType();
1908 // Construct an integer with the right value.
1909 unsigned EltSize = TD->getTypeSizeInBits(ValTy);
1910 APInt OneVal(EltSize, CI->getZExtValue());
1911 APInt TotalVal(OneVal);
1913 for (unsigned i = 0; 8*i < EltSize; ++i) {
1914 TotalVal = TotalVal.shl(8);
1918 // Convert the integer value to the appropriate type.
1919 StoreVal = ConstantInt::get(CI->getContext(), TotalVal);
1920 if (ValTy->isPointerTy())
1921 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy);
1922 else if (ValTy->isFloatingPointTy())
1923 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy);
1924 assert(StoreVal->getType() == ValTy && "Type mismatch!");
1926 // If the requested value was a vector constant, create it.
1927 if (EltTy != ValTy) {
1928 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements();
1929 SmallVector<Constant*, 16> Elts(NumElts, StoreVal);
1930 StoreVal = ConstantVector::get(Elts);
1933 new StoreInst(StoreVal, EltPtr, MI);
1936 // Otherwise, if we're storing a byte variable, use a memset call for
1940 unsigned EltSize = TD->getTypeAllocSize(EltTy);
1942 IRBuilder<> Builder(MI);
1944 // Finally, insert the meminst for this element.
1945 if (isa<MemSetInst>(MI)) {
1946 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize,
1949 assert(isa<MemTransferInst>(MI));
1950 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr
1951 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr
1953 if (isa<MemCpyInst>(MI))
1954 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile());
1956 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile());
1959 DeadInsts.push_back(MI);
1962 /// RewriteStoreUserOfWholeAlloca - We found a store of an integer that
1963 /// overwrites the entire allocation. Extract out the pieces of the stored
1964 /// integer and store them individually.
1965 void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
1966 SmallVector<AllocaInst*, 32> &NewElts){
1967 // Extract each element out of the integer according to its structure offset
1968 // and store the element value to the individual alloca.
1969 Value *SrcVal = SI->getOperand(0);
1970 const Type *AllocaEltTy = AI->getAllocatedType();
1971 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
1973 IRBuilder<> Builder(SI);
1975 // Handle tail padding by extending the operand
1976 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits)
1977 SrcVal = Builder.CreateZExt(SrcVal,
1978 IntegerType::get(SI->getContext(), AllocaSizeBits));
1980 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI
1983 // There are two forms here: AI could be an array or struct. Both cases
1984 // have different ways to compute the element offset.
1985 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
1986 const StructLayout *Layout = TD->getStructLayout(EltSTy);
1988 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1989 // Get the number of bits to shift SrcVal to get the value.
1990 const Type *FieldTy = EltSTy->getElementType(i);
1991 uint64_t Shift = Layout->getElementOffsetInBits(i);
1993 if (TD->isBigEndian())
1994 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy);
1996 Value *EltVal = SrcVal;
1998 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
1999 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt");
2002 // Truncate down to an integer of the right size.
2003 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
2005 // Ignore zero sized fields like {}, they obviously contain no data.
2006 if (FieldSizeBits == 0) continue;
2008 if (FieldSizeBits != AllocaSizeBits)
2009 EltVal = Builder.CreateTrunc(EltVal,
2010 IntegerType::get(SI->getContext(), FieldSizeBits));
2011 Value *DestField = NewElts[i];
2012 if (EltVal->getType() == FieldTy) {
2013 // Storing to an integer field of this size, just do it.
2014 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) {
2015 // Bitcast to the right element type (for fp/vector values).
2016 EltVal = Builder.CreateBitCast(EltVal, FieldTy);
2018 // Otherwise, bitcast the dest pointer (for aggregates).
2019 DestField = Builder.CreateBitCast(DestField,
2020 PointerType::getUnqual(EltVal->getType()));
2022 new StoreInst(EltVal, DestField, SI);
2026 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
2027 const Type *ArrayEltTy = ATy->getElementType();
2028 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
2029 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy);
2033 if (TD->isBigEndian())
2034 Shift = AllocaSizeBits-ElementOffset;
2038 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2039 // Ignore zero sized fields like {}, they obviously contain no data.
2040 if (ElementSizeBits == 0) continue;
2042 Value *EltVal = SrcVal;
2044 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
2045 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt");
2048 // Truncate down to an integer of the right size.
2049 if (ElementSizeBits != AllocaSizeBits)
2050 EltVal = Builder.CreateTrunc(EltVal,
2051 IntegerType::get(SI->getContext(),
2053 Value *DestField = NewElts[i];
2054 if (EltVal->getType() == ArrayEltTy) {
2055 // Storing to an integer field of this size, just do it.
2056 } else if (ArrayEltTy->isFloatingPointTy() ||
2057 ArrayEltTy->isVectorTy()) {
2058 // Bitcast to the right element type (for fp/vector values).
2059 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy);
2061 // Otherwise, bitcast the dest pointer (for aggregates).
2062 DestField = Builder.CreateBitCast(DestField,
2063 PointerType::getUnqual(EltVal->getType()));
2065 new StoreInst(EltVal, DestField, SI);
2067 if (TD->isBigEndian())
2068 Shift -= ElementOffset;
2070 Shift += ElementOffset;
2074 DeadInsts.push_back(SI);
2077 /// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to
2078 /// an integer. Load the individual pieces to form the aggregate value.
2079 void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
2080 SmallVector<AllocaInst*, 32> &NewElts) {
2081 // Extract each element out of the NewElts according to its structure offset
2082 // and form the result value.
2083 const Type *AllocaEltTy = AI->getAllocatedType();
2084 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
2086 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI
2089 // There are two forms here: AI could be an array or struct. Both cases
2090 // have different ways to compute the element offset.
2091 const StructLayout *Layout = 0;
2092 uint64_t ArrayEltBitOffset = 0;
2093 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
2094 Layout = TD->getStructLayout(EltSTy);
2096 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
2097 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
2101 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits));
2103 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2104 // Load the value from the alloca. If the NewElt is an aggregate, cast
2105 // the pointer to an integer of the same size before doing the load.
2106 Value *SrcField = NewElts[i];
2107 const Type *FieldTy =
2108 cast<PointerType>(SrcField->getType())->getElementType();
2109 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
2111 // Ignore zero sized fields like {}, they obviously contain no data.
2112 if (FieldSizeBits == 0) continue;
2114 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(),
2116 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() &&
2117 !FieldTy->isVectorTy())
2118 SrcField = new BitCastInst(SrcField,
2119 PointerType::getUnqual(FieldIntTy),
2121 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI);
2123 // If SrcField is a fp or vector of the right size but that isn't an
2124 // integer type, bitcast to an integer so we can shift it.
2125 if (SrcField->getType() != FieldIntTy)
2126 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI);
2128 // Zero extend the field to be the same size as the final alloca so that
2129 // we can shift and insert it.
2130 if (SrcField->getType() != ResultVal->getType())
2131 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI);
2133 // Determine the number of bits to shift SrcField.
2135 if (Layout) // Struct case.
2136 Shift = Layout->getElementOffsetInBits(i);
2138 Shift = i*ArrayEltBitOffset;
2140 if (TD->isBigEndian())
2141 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth();
2144 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift);
2145 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI);
2148 // Don't create an 'or x, 0' on the first iteration.
2149 if (!isa<Constant>(ResultVal) ||
2150 !cast<Constant>(ResultVal)->isNullValue())
2151 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI);
2153 ResultVal = SrcField;
2156 // Handle tail padding by truncating the result
2157 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits)
2158 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI);
2160 LI->replaceAllUsesWith(ResultVal);
2161 DeadInsts.push_back(LI);
2164 /// HasPadding - Return true if the specified type has any structure or
2165 /// alignment padding in between the elements that would be split apart
2166 /// by SROA; return false otherwise.
2167 static bool HasPadding(const Type *Ty, const TargetData &TD) {
2168 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2169 Ty = ATy->getElementType();
2170 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty);
2173 // SROA currently handles only Arrays and Structs.
2174 const StructType *STy = cast<StructType>(Ty);
2175 const StructLayout *SL = TD.getStructLayout(STy);
2176 unsigned PrevFieldBitOffset = 0;
2177 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2178 unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
2180 // Check to see if there is any padding between this element and the
2183 unsigned PrevFieldEnd =
2184 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1));
2185 if (PrevFieldEnd < FieldBitOffset)
2188 PrevFieldBitOffset = FieldBitOffset;
2190 // Check for tail padding.
2191 if (unsigned EltCount = STy->getNumElements()) {
2192 unsigned PrevFieldEnd = PrevFieldBitOffset +
2193 TD.getTypeSizeInBits(STy->getElementType(EltCount-1));
2194 if (PrevFieldEnd < SL->getSizeInBits())
2200 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
2201 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe,
2202 /// or 1 if safe after canonicalization has been performed.
2203 bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) {
2204 // Loop over the use list of the alloca. We can only transform it if all of
2205 // the users are safe to transform.
2206 AllocaInfo Info(AI);
2208 isSafeForScalarRepl(AI, 0, Info);
2209 if (Info.isUnsafe) {
2210 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n');
2214 // Okay, we know all the users are promotable. If the aggregate is a memcpy
2215 // source and destination, we have to be careful. In particular, the memcpy
2216 // could be moving around elements that live in structure padding of the LLVM
2217 // types, but may actually be used. In these cases, we refuse to promote the
2219 if (Info.isMemCpySrc && Info.isMemCpyDst &&
2220 HasPadding(AI->getAllocatedType(), *TD))
2223 // If the alloca never has an access to just *part* of it, but is accessed
2224 // via loads and stores, then we should use ConvertToScalarInfo to promote
2225 // the alloca instead of promoting each piece at a time and inserting fission
2227 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) {
2228 // If the struct/array just has one element, use basic SRoA.
2229 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
2230 if (ST->getNumElements() > 1) return false;
2232 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1)
2242 /// PointsToConstantGlobal - Return true if V (possibly indirectly) points to
2243 /// some part of a constant global variable. This intentionally only accepts
2244 /// constant expressions because we don't can't rewrite arbitrary instructions.
2245 static bool PointsToConstantGlobal(Value *V) {
2246 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
2247 return GV->isConstant();
2248 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
2249 if (CE->getOpcode() == Instruction::BitCast ||
2250 CE->getOpcode() == Instruction::GetElementPtr)
2251 return PointsToConstantGlobal(CE->getOperand(0));
2255 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
2256 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
2257 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
2258 /// track of whether it moves the pointer (with isOffset) but otherwise traverse
2259 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
2260 /// the alloca, and if the source pointer is a pointer to a constant global, we
2261 /// can optimize this.
2262 static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy,
2264 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
2265 User *U = cast<Instruction>(*UI);
2267 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
2268 // Ignore non-volatile loads, they are always ok.
2269 if (LI->isVolatile()) return false;
2273 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
2274 // If uses of the bitcast are ok, we are ok.
2275 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset))
2279 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
2280 // If the GEP has all zero indices, it doesn't offset the pointer. If it
2281 // doesn't, it does.
2282 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy,
2283 isOffset || !GEP->hasAllZeroIndices()))
2288 if (CallSite CS = U) {
2289 // If this is a readonly/readnone call site, then we know it is just a
2290 // load and we can ignore it.
2291 if (CS.onlyReadsMemory())
2294 // If this is the function being called then we treat it like a load and
2296 if (CS.isCallee(UI))
2299 // If this is being passed as a byval argument, the caller is making a
2300 // copy, so it is only a read of the alloca.
2301 unsigned ArgNo = CS.getArgumentNo(UI);
2302 if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal))
2306 // If this is isn't our memcpy/memmove, reject it as something we can't
2308 MemTransferInst *MI = dyn_cast<MemTransferInst>(U);
2312 // If the transfer is using the alloca as a source of the transfer, then
2313 // ignore it since it is a load (unless the transfer is volatile).
2314 if (UI.getOperandNo() == 1) {
2315 if (MI->isVolatile()) return false;
2319 // If we already have seen a copy, reject the second one.
2320 if (TheCopy) return false;
2322 // If the pointer has been offset from the start of the alloca, we can't
2323 // safely handle this.
2324 if (isOffset) return false;
2326 // If the memintrinsic isn't using the alloca as the dest, reject it.
2327 if (UI.getOperandNo() != 0) return false;
2329 // If the source of the memcpy/move is not a constant global, reject it.
2330 if (!PointsToConstantGlobal(MI->getSource()))
2333 // Otherwise, the transform is safe. Remember the copy instruction.
2339 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
2340 /// modified by a copy from a constant global. If we can prove this, we can
2341 /// replace any uses of the alloca with uses of the global directly.
2342 MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) {
2343 MemTransferInst *TheCopy = 0;
2344 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false))