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/Pass.h"
32 #include "llvm/Analysis/Dominators.h"
33 #include "llvm/Target/TargetData.h"
34 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
35 #include "llvm/Transforms/Utils/Local.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/ErrorHandling.h"
38 #include "llvm/Support/GetElementPtrTypeIterator.h"
39 #include "llvm/Support/IRBuilder.h"
40 #include "llvm/Support/MathExtras.h"
41 #include "llvm/Support/raw_ostream.h"
42 #include "llvm/ADT/SmallVector.h"
43 #include "llvm/ADT/Statistic.h"
46 STATISTIC(NumReplaced, "Number of allocas broken up");
47 STATISTIC(NumPromoted, "Number of allocas promoted");
48 STATISTIC(NumConverted, "Number of aggregates converted to scalar");
49 STATISTIC(NumGlobals, "Number of allocas copied from constant global");
52 struct SROA : public FunctionPass {
53 static char ID; // Pass identification, replacement for typeid
54 explicit SROA(signed T = -1) : FunctionPass(&ID) {
61 bool runOnFunction(Function &F);
63 bool performScalarRepl(Function &F);
64 bool performPromotion(Function &F);
66 // getAnalysisUsage - This pass does not require any passes, but we know it
67 // will not alter the CFG, so say so.
68 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
69 AU.addRequired<DominatorTree>();
70 AU.addRequired<DominanceFrontier>();
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 /// isUnsafe - This is set to true if the alloca cannot be SROA'd.
88 /// isMemCpySrc - This is true if this aggregate is memcpy'd from.
91 /// isMemCpyDst - This is true if this aggregate is memcpy'd into.
95 : isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false) {}
100 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; }
102 bool isSafeAllocaToScalarRepl(AllocaInst *AI);
104 void isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
106 void isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t &Offset,
108 void isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize,
109 const Type *MemOpType, bool isStore, AllocaInfo &Info);
110 bool TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size);
111 uint64_t FindElementAndOffset(const Type *&T, uint64_t &Offset,
114 void DoScalarReplacement(AllocaInst *AI,
115 std::vector<AllocaInst*> &WorkList);
116 void DeleteDeadInstructions();
117 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocaInst *Base);
119 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
120 SmallVector<AllocaInst*, 32> &NewElts);
121 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
122 SmallVector<AllocaInst*, 32> &NewElts);
123 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
124 SmallVector<AllocaInst*, 32> &NewElts);
125 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
127 SmallVector<AllocaInst*, 32> &NewElts);
128 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
129 SmallVector<AllocaInst*, 32> &NewElts);
130 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
131 SmallVector<AllocaInst*, 32> &NewElts);
133 static MemTransferInst *isOnlyCopiedFromConstantGlobal(AllocaInst *AI);
138 INITIALIZE_PASS(SROA, "scalarrepl",
139 "Scalar Replacement of Aggregates", false, false);
141 // Public interface to the ScalarReplAggregates pass
142 FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) {
143 return new SROA(Threshold);
147 //===----------------------------------------------------------------------===//
148 // Convert To Scalar Optimization.
149 //===----------------------------------------------------------------------===//
152 /// ConvertToScalarInfo - This class implements the "Convert To Scalar"
153 /// optimization, which scans the uses of an alloca and determines if it can
154 /// rewrite it in terms of a single new alloca that can be mem2reg'd.
155 class ConvertToScalarInfo {
156 /// AllocaSize - The size of the alloca being considered.
158 const TargetData &TD;
160 /// IsNotTrivial - This is set to true if there is some access to the object
161 /// which means that mem2reg can't promote it.
164 /// VectorTy - This tracks the type that we should promote the vector to if
165 /// it is possible to turn it into a vector. This starts out null, and if it
166 /// isn't possible to turn into a vector type, it gets set to VoidTy.
167 const Type *VectorTy;
169 /// HadAVector - True if there is at least one vector access to the alloca.
170 /// We don't want to turn random arrays into vectors and use vector element
171 /// insert/extract, but if there are element accesses to something that is
172 /// also declared as a vector, we do want to promote to a vector.
176 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td)
177 : AllocaSize(Size), TD(td) {
178 IsNotTrivial = false;
183 AllocaInst *TryConvert(AllocaInst *AI);
186 bool CanConvertToScalar(Value *V, uint64_t Offset);
187 void MergeInType(const Type *In, uint64_t Offset);
188 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset);
190 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType,
191 uint64_t Offset, IRBuilder<> &Builder);
192 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal,
193 uint64_t Offset, IRBuilder<> &Builder);
195 } // end anonymous namespace.
197 /// TryConvert - Analyze the specified alloca, and if it is safe to do so,
198 /// rewrite it to be a new alloca which is mem2reg'able. This returns the new
199 /// alloca if possible or null if not.
200 AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) {
201 // If we can't convert this scalar, or if mem2reg can trivially do it, bail
203 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial)
206 // If we were able to find a vector type that can handle this with
207 // insert/extract elements, and if there was at least one use that had
208 // a vector type, promote this to a vector. We don't want to promote
209 // random stuff that doesn't use vectors (e.g. <9 x double>) because then
210 // we just get a lot of insert/extracts. If at least one vector is
211 // involved, then we probably really do have a union of vector/array.
213 if (VectorTy && VectorTy->isVectorTy() && HadAVector) {
214 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = "
215 << *VectorTy << '\n');
216 NewTy = VectorTy; // Use the vector type.
218 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n");
219 // Create and insert the integer alloca.
220 NewTy = IntegerType::get(AI->getContext(), AllocaSize*8);
222 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin());
223 ConvertUsesToScalar(AI, NewAI, 0);
227 /// MergeInType - Add the 'In' type to the accumulated vector type (VectorTy)
228 /// so far at the offset specified by Offset (which is specified in bytes).
230 /// There are two cases we handle here:
231 /// 1) A union of vector types of the same size and potentially its elements.
232 /// Here we turn element accesses into insert/extract element operations.
233 /// This promotes a <4 x float> with a store of float to the third element
234 /// into a <4 x float> that uses insert element.
235 /// 2) A fully general blob of memory, which we turn into some (potentially
236 /// large) integer type with extract and insert operations where the loads
237 /// and stores would mutate the memory. We mark this by setting VectorTy
239 void ConvertToScalarInfo::MergeInType(const Type *In, uint64_t Offset) {
240 // If we already decided to turn this into a blob of integer memory, there is
241 // nothing to be done.
242 if (VectorTy && VectorTy->isVoidTy())
245 // If this could be contributing to a vector, analyze it.
247 // If the In type is a vector that is the same size as the alloca, see if it
248 // matches the existing VecTy.
249 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) {
250 // Remember if we saw a vector type.
253 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) {
254 // If we're storing/loading a vector of the right size, allow it as a
255 // vector. If this the first vector we see, remember the type so that
256 // we know the element size. If this is a subsequent access, ignore it
257 // even if it is a differing type but the same size. Worst case we can
258 // bitcast the resultant vectors.
263 } else if (In->isFloatTy() || In->isDoubleTy() ||
264 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 &&
265 isPowerOf2_32(In->getPrimitiveSizeInBits()))) {
266 // If we're accessing something that could be an element of a vector, see
267 // if the implied vector agrees with what we already have and if Offset is
268 // compatible with it.
269 unsigned EltSize = In->getPrimitiveSizeInBits()/8;
270 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 &&
272 cast<VectorType>(VectorTy)->getElementType()
273 ->getPrimitiveSizeInBits()/8 == EltSize)) {
275 VectorTy = VectorType::get(In, AllocaSize/EltSize);
280 // Otherwise, we have a case that we can't handle with an optimized vector
281 // form. We can still turn this into a large integer.
282 VectorTy = Type::getVoidTy(In->getContext());
285 /// CanConvertToScalar - V is a pointer. If we can convert the pointee and all
286 /// its accesses to a single vector type, return true and set VecTy to
287 /// the new type. If we could convert the alloca into a single promotable
288 /// integer, return true but set VecTy to VoidTy. Further, if the use is not a
289 /// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset
290 /// is the current offset from the base of the alloca being analyzed.
292 /// If we see at least one access to the value that is as a vector type, set the
294 bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) {
295 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
296 Instruction *User = cast<Instruction>(*UI);
298 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
299 // Don't break volatile loads.
300 if (LI->isVolatile())
302 MergeInType(LI->getType(), Offset);
306 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
307 // Storing the pointer, not into the value?
308 if (SI->getOperand(0) == V || SI->isVolatile()) return false;
309 MergeInType(SI->getOperand(0)->getType(), Offset);
313 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
314 IsNotTrivial = true; // Can't be mem2reg'd.
315 if (!CanConvertToScalar(BCI, Offset))
320 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
321 // If this is a GEP with a variable indices, we can't handle it.
322 if (!GEP->hasAllConstantIndices())
325 // Compute the offset that this GEP adds to the pointer.
326 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
327 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(),
328 &Indices[0], Indices.size());
329 // See if all uses can be converted.
330 if (!CanConvertToScalar(GEP, Offset+GEPOffset))
332 IsNotTrivial = true; // Can't be mem2reg'd.
336 // If this is a constant sized memset of a constant value (e.g. 0) we can
338 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
339 // Store of constant value and constant size.
340 if (!isa<ConstantInt>(MSI->getValue()) ||
341 !isa<ConstantInt>(MSI->getLength()))
343 IsNotTrivial = true; // Can't be mem2reg'd.
347 // If this is a memcpy or memmove into or out of the whole allocation, we
348 // can handle it like a load or store of the scalar type.
349 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
350 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength());
351 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0)
354 IsNotTrivial = true; // Can't be mem2reg'd.
358 // Otherwise, we cannot handle this!
365 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
366 /// directly. This happens when we are converting an "integer union" to a
367 /// single integer scalar, or when we are converting a "vector union" to a
368 /// vector with insert/extractelement instructions.
370 /// Offset is an offset from the original alloca, in bits that need to be
371 /// shifted to the right. By the end of this, there should be no uses of Ptr.
372 void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI,
374 while (!Ptr->use_empty()) {
375 Instruction *User = cast<Instruction>(Ptr->use_back());
377 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
378 ConvertUsesToScalar(CI, NewAI, Offset);
379 CI->eraseFromParent();
383 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
384 // Compute the offset that this GEP adds to the pointer.
385 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
386 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(),
387 &Indices[0], Indices.size());
388 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8);
389 GEP->eraseFromParent();
393 IRBuilder<> Builder(User->getParent(), User);
395 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
396 // The load is a bit extract from NewAI shifted right by Offset bits.
397 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp");
399 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder);
400 LI->replaceAllUsesWith(NewLoadVal);
401 LI->eraseFromParent();
405 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
406 assert(SI->getOperand(0) != Ptr && "Consistency error!");
407 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
408 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset,
410 Builder.CreateStore(New, NewAI);
411 SI->eraseFromParent();
413 // If the load we just inserted is now dead, then the inserted store
414 // overwrote the entire thing.
415 if (Old->use_empty())
416 Old->eraseFromParent();
420 // If this is a constant sized memset of a constant value (e.g. 0) we can
421 // transform it into a store of the expanded constant value.
422 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
423 assert(MSI->getRawDest() == Ptr && "Consistency error!");
424 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
426 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue();
428 // Compute the value replicated the right number of times.
429 APInt APVal(NumBytes*8, Val);
431 // Splat the value if non-zero.
433 for (unsigned i = 1; i != NumBytes; ++i)
436 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
437 Value *New = ConvertScalar_InsertValue(
438 ConstantInt::get(User->getContext(), APVal),
439 Old, Offset, Builder);
440 Builder.CreateStore(New, NewAI);
442 // If the load we just inserted is now dead, then the memset overwrote
444 if (Old->use_empty())
445 Old->eraseFromParent();
447 MSI->eraseFromParent();
451 // If this is a memcpy or memmove into or out of the whole allocation, we
452 // can handle it like a load or store of the scalar type.
453 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
454 assert(Offset == 0 && "must be store to start of alloca");
456 // If the source and destination are both to the same alloca, then this is
457 // a noop copy-to-self, just delete it. Otherwise, emit a load and store
459 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject(0));
461 if (MTI->getSource()->getUnderlyingObject(0) != OrigAI) {
462 // Dest must be OrigAI, change this to be a load from the original
463 // pointer (bitcasted), then a store to our new alloca.
464 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?");
465 Value *SrcPtr = MTI->getSource();
466 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType());
468 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval");
469 SrcVal->setAlignment(MTI->getAlignment());
470 Builder.CreateStore(SrcVal, NewAI);
471 } else if (MTI->getDest()->getUnderlyingObject(0) != OrigAI) {
472 // Src must be OrigAI, change this to be a load from NewAI then a store
473 // through the original dest pointer (bitcasted).
474 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?");
475 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval");
477 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType());
478 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr);
479 NewStore->setAlignment(MTI->getAlignment());
481 // Noop transfer. Src == Dst
484 MTI->eraseFromParent();
488 llvm_unreachable("Unsupported operation!");
492 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer
493 /// or vector value FromVal, extracting the bits from the offset specified by
494 /// Offset. This returns the value, which is of type ToType.
496 /// This happens when we are converting an "integer union" to a single
497 /// integer scalar, or when we are converting a "vector union" to a vector with
498 /// insert/extractelement instructions.
500 /// Offset is an offset from the original alloca, in bits that need to be
501 /// shifted to the right.
502 Value *ConvertToScalarInfo::
503 ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType,
504 uint64_t Offset, IRBuilder<> &Builder) {
505 // If the load is of the whole new alloca, no conversion is needed.
506 if (FromVal->getType() == ToType && Offset == 0)
509 // If the result alloca is a vector type, this is either an element
510 // access or a bitcast to another vector type of the same size.
511 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) {
512 if (ToType->isVectorTy())
513 return Builder.CreateBitCast(FromVal, ToType, "tmp");
515 // Otherwise it must be an element access.
518 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType());
519 Elt = Offset/EltSize;
520 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
522 // Return the element extracted out of it.
523 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get(
524 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp");
525 if (V->getType() != ToType)
526 V = Builder.CreateBitCast(V, ToType, "tmp");
530 // If ToType is a first class aggregate, extract out each of the pieces and
531 // use insertvalue's to form the FCA.
532 if (const StructType *ST = dyn_cast<StructType>(ToType)) {
533 const StructLayout &Layout = *TD.getStructLayout(ST);
534 Value *Res = UndefValue::get(ST);
535 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
536 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i),
537 Offset+Layout.getElementOffsetInBits(i),
539 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
544 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) {
545 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType());
546 Value *Res = UndefValue::get(AT);
547 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
548 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(),
549 Offset+i*EltSize, Builder);
550 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
555 // Otherwise, this must be a union that was converted to an integer value.
556 const IntegerType *NTy = cast<IntegerType>(FromVal->getType());
558 // If this is a big-endian system and the load is narrower than the
559 // full alloca type, we need to do a shift to get the right bits.
561 if (TD.isBigEndian()) {
562 // On big-endian machines, the lowest bit is stored at the bit offset
563 // from the pointer given by getTypeStoreSizeInBits. This matters for
564 // integers with a bitwidth that is not a multiple of 8.
565 ShAmt = TD.getTypeStoreSizeInBits(NTy) -
566 TD.getTypeStoreSizeInBits(ToType) - Offset;
571 // Note: we support negative bitwidths (with shl) which are not defined.
572 // We do this to support (f.e.) loads off the end of a structure where
573 // only some bits are used.
574 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth())
575 FromVal = Builder.CreateLShr(FromVal,
576 ConstantInt::get(FromVal->getType(),
578 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
579 FromVal = Builder.CreateShl(FromVal,
580 ConstantInt::get(FromVal->getType(),
583 // Finally, unconditionally truncate the integer to the right width.
584 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType);
585 if (LIBitWidth < NTy->getBitWidth())
587 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(),
589 else if (LIBitWidth > NTy->getBitWidth())
591 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(),
594 // If the result is an integer, this is a trunc or bitcast.
595 if (ToType->isIntegerTy()) {
597 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) {
598 // Just do a bitcast, we know the sizes match up.
599 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp");
601 // Otherwise must be a pointer.
602 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp");
604 assert(FromVal->getType() == ToType && "Didn't convert right?");
608 /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer
609 /// or vector value "Old" at the offset specified by Offset.
611 /// This happens when we are converting an "integer union" to a
612 /// single integer scalar, or when we are converting a "vector union" to a
613 /// vector with insert/extractelement instructions.
615 /// Offset is an offset from the original alloca, in bits that need to be
616 /// shifted to the right.
617 Value *ConvertToScalarInfo::
618 ConvertScalar_InsertValue(Value *SV, Value *Old,
619 uint64_t Offset, IRBuilder<> &Builder) {
620 // Convert the stored type to the actual type, shift it left to insert
621 // then 'or' into place.
622 const Type *AllocaType = Old->getType();
623 LLVMContext &Context = Old->getContext();
625 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
626 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy);
627 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType());
629 // Changing the whole vector with memset or with an access of a different
631 if (ValSize == VecSize)
632 return Builder.CreateBitCast(SV, AllocaType, "tmp");
634 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType());
636 // Must be an element insertion.
637 unsigned Elt = Offset/EltSize;
639 if (SV->getType() != VTy->getElementType())
640 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp");
642 SV = Builder.CreateInsertElement(Old, SV,
643 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt),
648 // If SV is a first-class aggregate value, insert each value recursively.
649 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) {
650 const StructLayout &Layout = *TD.getStructLayout(ST);
651 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
652 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
653 Old = ConvertScalar_InsertValue(Elt, Old,
654 Offset+Layout.getElementOffsetInBits(i),
660 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) {
661 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType());
662 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
663 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
664 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder);
669 // If SV is a float, convert it to the appropriate integer type.
670 // If it is a pointer, do the same.
671 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType());
672 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType);
673 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType());
674 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType);
675 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy())
676 SV = Builder.CreateBitCast(SV,
677 IntegerType::get(SV->getContext(),SrcWidth), "tmp");
678 else if (SV->getType()->isPointerTy())
679 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp");
681 // Zero extend or truncate the value if needed.
682 if (SV->getType() != AllocaType) {
683 if (SV->getType()->getPrimitiveSizeInBits() <
684 AllocaType->getPrimitiveSizeInBits())
685 SV = Builder.CreateZExt(SV, AllocaType, "tmp");
687 // Truncation may be needed if storing more than the alloca can hold
688 // (undefined behavior).
689 SV = Builder.CreateTrunc(SV, AllocaType, "tmp");
690 SrcWidth = DestWidth;
691 SrcStoreWidth = DestStoreWidth;
695 // If this is a big-endian system and the store is narrower than the
696 // full alloca type, we need to do a shift to get the right bits.
698 if (TD.isBigEndian()) {
699 // On big-endian machines, the lowest bit is stored at the bit offset
700 // from the pointer given by getTypeStoreSizeInBits. This matters for
701 // integers with a bitwidth that is not a multiple of 8.
702 ShAmt = DestStoreWidth - SrcStoreWidth - Offset;
707 // Note: we support negative bitwidths (with shr) which are not defined.
708 // We do this to support (f.e.) stores off the end of a structure where
709 // only some bits in the structure are set.
710 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
711 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
712 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(),
715 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
716 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(),
718 Mask = Mask.lshr(-ShAmt);
721 // Mask out the bits we are about to insert from the old value, and or
723 if (SrcWidth != DestWidth) {
724 assert(DestWidth > SrcWidth);
725 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask");
726 SV = Builder.CreateOr(Old, SV, "ins");
732 //===----------------------------------------------------------------------===//
734 //===----------------------------------------------------------------------===//
737 bool SROA::runOnFunction(Function &F) {
738 TD = getAnalysisIfAvailable<TargetData>();
740 bool Changed = performPromotion(F);
742 // FIXME: ScalarRepl currently depends on TargetData more than it
743 // theoretically needs to. It should be refactored in order to support
744 // target-independent IR. Until this is done, just skip the actual
745 // scalar-replacement portion of this pass.
746 if (!TD) return Changed;
749 bool LocalChange = performScalarRepl(F);
750 if (!LocalChange) break; // No need to repromote if no scalarrepl
752 LocalChange = performPromotion(F);
753 if (!LocalChange) break; // No need to re-scalarrepl if no promotion
760 bool SROA::performPromotion(Function &F) {
761 std::vector<AllocaInst*> Allocas;
762 DominatorTree &DT = getAnalysis<DominatorTree>();
763 DominanceFrontier &DF = getAnalysis<DominanceFrontier>();
765 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
767 bool Changed = false;
772 // Find allocas that are safe to promote, by looking at all instructions in
774 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
775 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
776 if (isAllocaPromotable(AI))
777 Allocas.push_back(AI);
779 if (Allocas.empty()) break;
781 PromoteMemToReg(Allocas, DT, DF);
782 NumPromoted += Allocas.size();
790 /// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for
791 /// SROA. It must be a struct or array type with a small number of elements.
792 static bool ShouldAttemptScalarRepl(AllocaInst *AI) {
793 const Type *T = AI->getAllocatedType();
794 // Do not promote any struct into more than 32 separate vars.
795 if (const StructType *ST = dyn_cast<StructType>(T))
796 return ST->getNumElements() <= 32;
797 // Arrays are much less likely to be safe for SROA; only consider
798 // them if they are very small.
799 if (const ArrayType *AT = dyn_cast<ArrayType>(T))
800 return AT->getNumElements() <= 8;
805 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
806 // which runs on all of the malloc/alloca instructions in the function, removing
807 // them if they are only used by getelementptr instructions.
809 bool SROA::performScalarRepl(Function &F) {
810 std::vector<AllocaInst*> WorkList;
812 // Scan the entry basic block, adding allocas to the worklist.
813 BasicBlock &BB = F.getEntryBlock();
814 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
815 if (AllocaInst *A = dyn_cast<AllocaInst>(I))
816 WorkList.push_back(A);
818 // Process the worklist
819 bool Changed = false;
820 while (!WorkList.empty()) {
821 AllocaInst *AI = WorkList.back();
824 // Handle dead allocas trivially. These can be formed by SROA'ing arrays
825 // with unused elements.
826 if (AI->use_empty()) {
827 AI->eraseFromParent();
832 // If this alloca is impossible for us to promote, reject it early.
833 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized())
836 // Check to see if this allocation is only modified by a memcpy/memmove from
837 // a constant global. If this is the case, we can change all users to use
838 // the constant global instead. This is commonly produced by the CFE by
839 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
840 // is only subsequently read.
841 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) {
842 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n');
843 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n');
844 Constant *TheSrc = cast<Constant>(TheCopy->getSource());
845 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType()));
846 TheCopy->eraseFromParent(); // Don't mutate the global.
847 AI->eraseFromParent();
853 // Check to see if we can perform the core SROA transformation. We cannot
854 // transform the allocation instruction if it is an array allocation
855 // (allocations OF arrays are ok though), and an allocation of a scalar
856 // value cannot be decomposed at all.
857 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType());
859 // Do not promote [0 x %struct].
860 if (AllocaSize == 0) continue;
862 // Do not promote any struct whose size is too big.
863 if (AllocaSize > SRThreshold) continue;
865 // If the alloca looks like a good candidate for scalar replacement, and if
866 // all its users can be transformed, then split up the aggregate into its
867 // separate elements.
868 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) {
869 DoScalarReplacement(AI, WorkList);
874 // If we can turn this aggregate value (potentially with casts) into a
875 // simple scalar value that can be mem2reg'd into a register value.
876 // IsNotTrivial tracks whether this is something that mem2reg could have
877 // promoted itself. If so, we don't want to transform it needlessly. Note
878 // that we can't just check based on the type: the alloca may be of an i32
879 // but that has pointer arithmetic to set byte 3 of it or something.
880 if (AllocaInst *NewAI =
881 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) {
883 AI->eraseFromParent();
889 // Otherwise, couldn't process this alloca.
895 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl
896 /// predicate, do SROA now.
897 void SROA::DoScalarReplacement(AllocaInst *AI,
898 std::vector<AllocaInst*> &WorkList) {
899 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n');
900 SmallVector<AllocaInst*, 32> ElementAllocas;
901 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
902 ElementAllocas.reserve(ST->getNumContainedTypes());
903 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
904 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
906 AI->getName() + "." + Twine(i), AI);
907 ElementAllocas.push_back(NA);
908 WorkList.push_back(NA); // Add to worklist for recursive processing
911 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
912 ElementAllocas.reserve(AT->getNumElements());
913 const Type *ElTy = AT->getElementType();
914 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
915 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
916 AI->getName() + "." + Twine(i), AI);
917 ElementAllocas.push_back(NA);
918 WorkList.push_back(NA); // Add to worklist for recursive processing
922 // Now that we have created the new alloca instructions, rewrite all the
923 // uses of the old alloca.
924 RewriteForScalarRepl(AI, AI, 0, ElementAllocas);
926 // Now erase any instructions that were made dead while rewriting the alloca.
927 DeleteDeadInstructions();
928 AI->eraseFromParent();
933 /// DeleteDeadInstructions - Erase instructions on the DeadInstrs list,
934 /// recursively including all their operands that become trivially dead.
935 void SROA::DeleteDeadInstructions() {
936 while (!DeadInsts.empty()) {
937 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val());
939 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
940 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
941 // Zero out the operand and see if it becomes trivially dead.
942 // (But, don't add allocas to the dead instruction list -- they are
943 // already on the worklist and will be deleted separately.)
945 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U))
946 DeadInsts.push_back(U);
949 I->eraseFromParent();
953 /// isSafeForScalarRepl - Check if instruction I is a safe use with regard to
954 /// performing scalar replacement of alloca AI. The results are flagged in
955 /// the Info parameter. Offset indicates the position within AI that is
956 /// referenced by this instruction.
957 void SROA::isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
959 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
960 Instruction *User = cast<Instruction>(*UI);
962 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
963 isSafeForScalarRepl(BC, AI, Offset, Info);
964 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
965 uint64_t GEPOffset = Offset;
966 isSafeGEP(GEPI, AI, GEPOffset, Info);
968 isSafeForScalarRepl(GEPI, AI, GEPOffset, Info);
969 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
970 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
972 isSafeMemAccess(AI, Offset, Length->getZExtValue(), 0,
973 UI.getOperandNo() == 0, Info);
976 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
977 if (!LI->isVolatile()) {
978 const Type *LIType = LI->getType();
979 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(LIType),
980 LIType, false, Info);
983 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
984 // Store is ok if storing INTO the pointer, not storing the pointer
985 if (!SI->isVolatile() && SI->getOperand(0) != I) {
986 const Type *SIType = SI->getOperand(0)->getType();
987 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(SIType),
992 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n');
995 if (Info.isUnsafe) return;
999 /// isSafeGEP - Check if a GEP instruction can be handled for scalar
1000 /// replacement. It is safe when all the indices are constant, in-bounds
1001 /// references, and when the resulting offset corresponds to an element within
1002 /// the alloca type. The results are flagged in the Info parameter. Upon
1003 /// return, Offset is adjusted as specified by the GEP indices.
1004 void SROA::isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI,
1005 uint64_t &Offset, AllocaInfo &Info) {
1006 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI);
1010 // Walk through the GEP type indices, checking the types that this indexes
1012 for (; GEPIt != E; ++GEPIt) {
1013 // Ignore struct elements, no extra checking needed for these.
1014 if ((*GEPIt)->isStructTy())
1017 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand());
1019 return MarkUnsafe(Info);
1022 // Compute the offset due to this GEP and check if the alloca has a
1023 // component element at that offset.
1024 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
1025 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(),
1026 &Indices[0], Indices.size());
1027 if (!TypeHasComponent(AI->getAllocatedType(), Offset, 0))
1031 /// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI
1032 /// alloca or has an offset and size that corresponds to a component element
1033 /// within it. The offset checked here may have been formed from a GEP with a
1034 /// pointer bitcasted to a different type.
1035 void SROA::isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize,
1036 const Type *MemOpType, bool isStore,
1038 // Check if this is a load/store of the entire alloca.
1039 if (Offset == 0 && MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) {
1040 bool UsesAggregateType = (MemOpType == AI->getAllocatedType());
1041 // This is safe for MemIntrinsics (where MemOpType is 0), integer types
1042 // (which are essentially the same as the MemIntrinsics, especially with
1043 // regard to copying padding between elements), or references using the
1044 // aggregate type of the alloca.
1045 if (!MemOpType || MemOpType->isIntegerTy() || UsesAggregateType) {
1046 if (!UsesAggregateType) {
1048 Info.isMemCpyDst = true;
1050 Info.isMemCpySrc = true;
1055 // Check if the offset/size correspond to a component within the alloca type.
1056 const Type *T = AI->getAllocatedType();
1057 if (TypeHasComponent(T, Offset, MemSize))
1060 return MarkUnsafe(Info);
1063 /// TypeHasComponent - Return true if T has a component type with the
1064 /// specified offset and size. If Size is zero, do not check the size.
1065 bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) {
1068 if (const StructType *ST = dyn_cast<StructType>(T)) {
1069 const StructLayout *Layout = TD->getStructLayout(ST);
1070 unsigned EltIdx = Layout->getElementContainingOffset(Offset);
1071 EltTy = ST->getContainedType(EltIdx);
1072 EltSize = TD->getTypeAllocSize(EltTy);
1073 Offset -= Layout->getElementOffset(EltIdx);
1074 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) {
1075 EltTy = AT->getElementType();
1076 EltSize = TD->getTypeAllocSize(EltTy);
1077 if (Offset >= AT->getNumElements() * EltSize)
1083 if (Offset == 0 && (Size == 0 || EltSize == Size))
1085 // Check if the component spans multiple elements.
1086 if (Offset + Size > EltSize)
1088 return TypeHasComponent(EltTy, Offset, Size);
1091 /// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite
1092 /// the instruction I, which references it, to use the separate elements.
1093 /// Offset indicates the position within AI that is referenced by this
1095 void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
1096 SmallVector<AllocaInst*, 32> &NewElts) {
1097 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
1098 Instruction *User = cast<Instruction>(*UI);
1100 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1101 RewriteBitCast(BC, AI, Offset, NewElts);
1102 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1103 RewriteGEP(GEPI, AI, Offset, NewElts);
1104 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
1105 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
1106 uint64_t MemSize = Length->getZExtValue();
1108 MemSize == TD->getTypeAllocSize(AI->getAllocatedType()))
1109 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts);
1110 // Otherwise the intrinsic can only touch a single element and the
1111 // address operand will be updated, so nothing else needs to be done.
1112 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1113 const Type *LIType = LI->getType();
1114 if (LIType == AI->getAllocatedType()) {
1116 // %res = load { i32, i32 }* %alloc
1118 // %load.0 = load i32* %alloc.0
1119 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0
1120 // %load.1 = load i32* %alloc.1
1121 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1
1122 // (Also works for arrays instead of structs)
1123 Value *Insert = UndefValue::get(LIType);
1124 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1125 Value *Load = new LoadInst(NewElts[i], "load", LI);
1126 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI);
1128 LI->replaceAllUsesWith(Insert);
1129 DeadInsts.push_back(LI);
1130 } else if (LIType->isIntegerTy() &&
1131 TD->getTypeAllocSize(LIType) ==
1132 TD->getTypeAllocSize(AI->getAllocatedType())) {
1133 // If this is a load of the entire alloca to an integer, rewrite it.
1134 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts);
1136 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1137 Value *Val = SI->getOperand(0);
1138 const Type *SIType = Val->getType();
1139 if (SIType == AI->getAllocatedType()) {
1141 // store { i32, i32 } %val, { i32, i32 }* %alloc
1143 // %val.0 = extractvalue { i32, i32 } %val, 0
1144 // store i32 %val.0, i32* %alloc.0
1145 // %val.1 = extractvalue { i32, i32 } %val, 1
1146 // store i32 %val.1, i32* %alloc.1
1147 // (Also works for arrays instead of structs)
1148 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1149 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI);
1150 new StoreInst(Extract, NewElts[i], SI);
1152 DeadInsts.push_back(SI);
1153 } else if (SIType->isIntegerTy() &&
1154 TD->getTypeAllocSize(SIType) ==
1155 TD->getTypeAllocSize(AI->getAllocatedType())) {
1156 // If this is a store of the entire alloca from an integer, rewrite it.
1157 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts);
1163 /// RewriteBitCast - Update a bitcast reference to the alloca being replaced
1164 /// and recursively continue updating all of its uses.
1165 void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
1166 SmallVector<AllocaInst*, 32> &NewElts) {
1167 RewriteForScalarRepl(BC, AI, Offset, NewElts);
1168 if (BC->getOperand(0) != AI)
1171 // The bitcast references the original alloca. Replace its uses with
1172 // references to the first new element alloca.
1173 Instruction *Val = NewElts[0];
1174 if (Val->getType() != BC->getDestTy()) {
1175 Val = new BitCastInst(Val, BC->getDestTy(), "", BC);
1178 BC->replaceAllUsesWith(Val);
1179 DeadInsts.push_back(BC);
1182 /// FindElementAndOffset - Return the index of the element containing Offset
1183 /// within the specified type, which must be either a struct or an array.
1184 /// Sets T to the type of the element and Offset to the offset within that
1185 /// element. IdxTy is set to the type of the index result to be used in a
1186 /// GEP instruction.
1187 uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset,
1188 const Type *&IdxTy) {
1190 if (const StructType *ST = dyn_cast<StructType>(T)) {
1191 const StructLayout *Layout = TD->getStructLayout(ST);
1192 Idx = Layout->getElementContainingOffset(Offset);
1193 T = ST->getContainedType(Idx);
1194 Offset -= Layout->getElementOffset(Idx);
1195 IdxTy = Type::getInt32Ty(T->getContext());
1198 const ArrayType *AT = cast<ArrayType>(T);
1199 T = AT->getElementType();
1200 uint64_t EltSize = TD->getTypeAllocSize(T);
1201 Idx = Offset / EltSize;
1202 Offset -= Idx * EltSize;
1203 IdxTy = Type::getInt64Ty(T->getContext());
1207 /// RewriteGEP - Check if this GEP instruction moves the pointer across
1208 /// elements of the alloca that are being split apart, and if so, rewrite
1209 /// the GEP to be relative to the new element.
1210 void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
1211 SmallVector<AllocaInst*, 32> &NewElts) {
1212 uint64_t OldOffset = Offset;
1213 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
1214 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(),
1215 &Indices[0], Indices.size());
1217 RewriteForScalarRepl(GEPI, AI, Offset, NewElts);
1219 const Type *T = AI->getAllocatedType();
1221 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy);
1222 if (GEPI->getOperand(0) == AI)
1223 OldIdx = ~0ULL; // Force the GEP to be rewritten.
1225 T = AI->getAllocatedType();
1226 uint64_t EltOffset = Offset;
1227 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy);
1229 // If this GEP does not move the pointer across elements of the alloca
1230 // being split, then it does not needs to be rewritten.
1234 const Type *i32Ty = Type::getInt32Ty(AI->getContext());
1235 SmallVector<Value*, 8> NewArgs;
1236 NewArgs.push_back(Constant::getNullValue(i32Ty));
1237 while (EltOffset != 0) {
1238 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy);
1239 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx));
1241 Instruction *Val = NewElts[Idx];
1242 if (NewArgs.size() > 1) {
1243 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(),
1244 NewArgs.end(), "", GEPI);
1245 Val->takeName(GEPI);
1247 if (Val->getType() != GEPI->getType())
1248 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI);
1249 GEPI->replaceAllUsesWith(Val);
1250 DeadInsts.push_back(GEPI);
1253 /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI.
1254 /// Rewrite it to copy or set the elements of the scalarized memory.
1255 void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
1257 SmallVector<AllocaInst*, 32> &NewElts) {
1258 // If this is a memcpy/memmove, construct the other pointer as the
1259 // appropriate type. The "Other" pointer is the pointer that goes to memory
1260 // that doesn't have anything to do with the alloca that we are promoting. For
1261 // memset, this Value* stays null.
1262 Value *OtherPtr = 0;
1263 unsigned MemAlignment = MI->getAlignment();
1264 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy
1265 if (Inst == MTI->getRawDest())
1266 OtherPtr = MTI->getRawSource();
1268 assert(Inst == MTI->getRawSource());
1269 OtherPtr = MTI->getRawDest();
1273 // If there is an other pointer, we want to convert it to the same pointer
1274 // type as AI has, so we can GEP through it safely.
1276 unsigned AddrSpace =
1277 cast<PointerType>(OtherPtr->getType())->getAddressSpace();
1279 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an
1280 // optimization, but it's also required to detect the corner case where
1281 // both pointer operands are referencing the same memory, and where
1282 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This
1283 // function is only called for mem intrinsics that access the whole
1284 // aggregate, so non-zero GEPs are not an issue here.)
1285 OtherPtr = OtherPtr->stripPointerCasts();
1287 // Copying the alloca to itself is a no-op: just delete it.
1288 if (OtherPtr == AI || OtherPtr == NewElts[0]) {
1289 // This code will run twice for a no-op memcpy -- once for each operand.
1290 // Put only one reference to MI on the DeadInsts list.
1291 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(),
1292 E = DeadInsts.end(); I != E; ++I)
1293 if (*I == MI) return;
1294 DeadInsts.push_back(MI);
1298 // If the pointer is not the right type, insert a bitcast to the right
1301 PointerType::get(AI->getType()->getElementType(), AddrSpace);
1303 if (OtherPtr->getType() != NewTy)
1304 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI);
1307 // Process each element of the aggregate.
1308 Value *TheFn = MI->getCalledValue();
1309 const Type *BytePtrTy = MI->getRawDest()->getType();
1310 bool SROADest = MI->getRawDest() == Inst;
1312 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext()));
1314 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1315 // If this is a memcpy/memmove, emit a GEP of the other element address.
1316 Value *OtherElt = 0;
1317 unsigned OtherEltAlign = MemAlignment;
1320 Value *Idx[2] = { Zero,
1321 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) };
1322 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2,
1323 OtherPtr->getName()+"."+Twine(i),
1326 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType());
1327 const Type *OtherTy = OtherPtrTy->getElementType();
1328 if (const StructType *ST = dyn_cast<StructType>(OtherTy)) {
1329 EltOffset = TD->getStructLayout(ST)->getElementOffset(i);
1331 const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType();
1332 EltOffset = TD->getTypeAllocSize(EltTy)*i;
1335 // The alignment of the other pointer is the guaranteed alignment of the
1336 // element, which is affected by both the known alignment of the whole
1337 // mem intrinsic and the alignment of the element. If the alignment of
1338 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the
1339 // known alignment is just 4 bytes.
1340 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset);
1343 Value *EltPtr = NewElts[i];
1344 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType();
1346 // If we got down to a scalar, insert a load or store as appropriate.
1347 if (EltTy->isSingleValueType()) {
1348 if (isa<MemTransferInst>(MI)) {
1350 // From Other to Alloca.
1351 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI);
1352 new StoreInst(Elt, EltPtr, MI);
1354 // From Alloca to Other.
1355 Value *Elt = new LoadInst(EltPtr, "tmp", MI);
1356 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI);
1360 assert(isa<MemSetInst>(MI));
1362 // If the stored element is zero (common case), just store a null
1365 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) {
1367 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0>
1369 // If EltTy is a vector type, get the element type.
1370 const Type *ValTy = EltTy->getScalarType();
1372 // Construct an integer with the right value.
1373 unsigned EltSize = TD->getTypeSizeInBits(ValTy);
1374 APInt OneVal(EltSize, CI->getZExtValue());
1375 APInt TotalVal(OneVal);
1377 for (unsigned i = 0; 8*i < EltSize; ++i) {
1378 TotalVal = TotalVal.shl(8);
1382 // Convert the integer value to the appropriate type.
1383 StoreVal = ConstantInt::get(CI->getContext(), TotalVal);
1384 if (ValTy->isPointerTy())
1385 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy);
1386 else if (ValTy->isFloatingPointTy())
1387 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy);
1388 assert(StoreVal->getType() == ValTy && "Type mismatch!");
1390 // If the requested value was a vector constant, create it.
1391 if (EltTy != ValTy) {
1392 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements();
1393 SmallVector<Constant*, 16> Elts(NumElts, StoreVal);
1394 StoreVal = ConstantVector::get(&Elts[0], NumElts);
1397 new StoreInst(StoreVal, EltPtr, MI);
1400 // Otherwise, if we're storing a byte variable, use a memset call for
1404 // Cast the element pointer to BytePtrTy.
1405 if (EltPtr->getType() != BytePtrTy)
1406 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getName(), MI);
1408 // Cast the other pointer (if we have one) to BytePtrTy.
1409 if (OtherElt && OtherElt->getType() != BytePtrTy) {
1410 // Preserve address space of OtherElt
1411 const PointerType* OtherPTy = cast<PointerType>(OtherElt->getType());
1412 const PointerType* PTy = cast<PointerType>(BytePtrTy);
1413 if (OtherPTy->getElementType() != PTy->getElementType()) {
1414 Type *NewOtherPTy = PointerType::get(PTy->getElementType(),
1415 OtherPTy->getAddressSpace());
1416 OtherElt = new BitCastInst(OtherElt, NewOtherPTy,
1417 OtherElt->getNameStr(), MI);
1421 unsigned EltSize = TD->getTypeAllocSize(EltTy);
1423 // Finally, insert the meminst for this element.
1424 if (isa<MemTransferInst>(MI)) {
1426 SROADest ? EltPtr : OtherElt, // Dest ptr
1427 SROADest ? OtherElt : EltPtr, // Src ptr
1428 ConstantInt::get(MI->getArgOperand(2)->getType(), EltSize), // Size
1430 ConstantInt::get(Type::getInt32Ty(MI->getContext()), OtherEltAlign),
1431 MI->getVolatileCst()
1433 // In case we fold the address space overloaded memcpy of A to B
1434 // with memcpy of B to C, change the function to be a memcpy of A to C.
1435 const Type *Tys[] = { Ops[0]->getType(), Ops[1]->getType(),
1436 Ops[2]->getType() };
1437 Module *M = MI->getParent()->getParent()->getParent();
1438 TheFn = Intrinsic::getDeclaration(M, MI->getIntrinsicID(), Tys, 3);
1439 CallInst::Create(TheFn, Ops, Ops + 5, "", MI);
1441 assert(isa<MemSetInst>(MI));
1443 EltPtr, MI->getArgOperand(1), // Dest, Value,
1444 ConstantInt::get(MI->getArgOperand(2)->getType(), EltSize), // Size
1446 ConstantInt::get(Type::getInt1Ty(MI->getContext()), 0) // isVolatile
1448 const Type *Tys[] = { Ops[0]->getType(), Ops[2]->getType() };
1449 Module *M = MI->getParent()->getParent()->getParent();
1450 TheFn = Intrinsic::getDeclaration(M, Intrinsic::memset, Tys, 2);
1451 CallInst::Create(TheFn, Ops, Ops + 5, "", MI);
1454 DeadInsts.push_back(MI);
1457 /// RewriteStoreUserOfWholeAlloca - We found a store of an integer that
1458 /// overwrites the entire allocation. Extract out the pieces of the stored
1459 /// integer and store them individually.
1460 void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
1461 SmallVector<AllocaInst*, 32> &NewElts){
1462 // Extract each element out of the integer according to its structure offset
1463 // and store the element value to the individual alloca.
1464 Value *SrcVal = SI->getOperand(0);
1465 const Type *AllocaEltTy = AI->getAllocatedType();
1466 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
1468 // Handle tail padding by extending the operand
1469 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits)
1470 SrcVal = new ZExtInst(SrcVal,
1471 IntegerType::get(SI->getContext(), AllocaSizeBits),
1474 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI
1477 // There are two forms here: AI could be an array or struct. Both cases
1478 // have different ways to compute the element offset.
1479 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
1480 const StructLayout *Layout = TD->getStructLayout(EltSTy);
1482 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1483 // Get the number of bits to shift SrcVal to get the value.
1484 const Type *FieldTy = EltSTy->getElementType(i);
1485 uint64_t Shift = Layout->getElementOffsetInBits(i);
1487 if (TD->isBigEndian())
1488 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy);
1490 Value *EltVal = SrcVal;
1492 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
1493 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
1494 "sroa.store.elt", SI);
1497 // Truncate down to an integer of the right size.
1498 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
1500 // Ignore zero sized fields like {}, they obviously contain no data.
1501 if (FieldSizeBits == 0) continue;
1503 if (FieldSizeBits != AllocaSizeBits)
1504 EltVal = new TruncInst(EltVal,
1505 IntegerType::get(SI->getContext(), FieldSizeBits),
1507 Value *DestField = NewElts[i];
1508 if (EltVal->getType() == FieldTy) {
1509 // Storing to an integer field of this size, just do it.
1510 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) {
1511 // Bitcast to the right element type (for fp/vector values).
1512 EltVal = new BitCastInst(EltVal, FieldTy, "", SI);
1514 // Otherwise, bitcast the dest pointer (for aggregates).
1515 DestField = new BitCastInst(DestField,
1516 PointerType::getUnqual(EltVal->getType()),
1519 new StoreInst(EltVal, DestField, SI);
1523 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
1524 const Type *ArrayEltTy = ATy->getElementType();
1525 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
1526 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy);
1530 if (TD->isBigEndian())
1531 Shift = AllocaSizeBits-ElementOffset;
1535 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1536 // Ignore zero sized fields like {}, they obviously contain no data.
1537 if (ElementSizeBits == 0) continue;
1539 Value *EltVal = SrcVal;
1541 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
1542 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
1543 "sroa.store.elt", SI);
1546 // Truncate down to an integer of the right size.
1547 if (ElementSizeBits != AllocaSizeBits)
1548 EltVal = new TruncInst(EltVal,
1549 IntegerType::get(SI->getContext(),
1550 ElementSizeBits),"",SI);
1551 Value *DestField = NewElts[i];
1552 if (EltVal->getType() == ArrayEltTy) {
1553 // Storing to an integer field of this size, just do it.
1554 } else if (ArrayEltTy->isFloatingPointTy() ||
1555 ArrayEltTy->isVectorTy()) {
1556 // Bitcast to the right element type (for fp/vector values).
1557 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI);
1559 // Otherwise, bitcast the dest pointer (for aggregates).
1560 DestField = new BitCastInst(DestField,
1561 PointerType::getUnqual(EltVal->getType()),
1564 new StoreInst(EltVal, DestField, SI);
1566 if (TD->isBigEndian())
1567 Shift -= ElementOffset;
1569 Shift += ElementOffset;
1573 DeadInsts.push_back(SI);
1576 /// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to
1577 /// an integer. Load the individual pieces to form the aggregate value.
1578 void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
1579 SmallVector<AllocaInst*, 32> &NewElts) {
1580 // Extract each element out of the NewElts according to its structure offset
1581 // and form the result value.
1582 const Type *AllocaEltTy = AI->getAllocatedType();
1583 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
1585 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI
1588 // There are two forms here: AI could be an array or struct. Both cases
1589 // have different ways to compute the element offset.
1590 const StructLayout *Layout = 0;
1591 uint64_t ArrayEltBitOffset = 0;
1592 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
1593 Layout = TD->getStructLayout(EltSTy);
1595 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
1596 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
1600 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits));
1602 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1603 // Load the value from the alloca. If the NewElt is an aggregate, cast
1604 // the pointer to an integer of the same size before doing the load.
1605 Value *SrcField = NewElts[i];
1606 const Type *FieldTy =
1607 cast<PointerType>(SrcField->getType())->getElementType();
1608 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
1610 // Ignore zero sized fields like {}, they obviously contain no data.
1611 if (FieldSizeBits == 0) continue;
1613 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(),
1615 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() &&
1616 !FieldTy->isVectorTy())
1617 SrcField = new BitCastInst(SrcField,
1618 PointerType::getUnqual(FieldIntTy),
1620 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI);
1622 // If SrcField is a fp or vector of the right size but that isn't an
1623 // integer type, bitcast to an integer so we can shift it.
1624 if (SrcField->getType() != FieldIntTy)
1625 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI);
1627 // Zero extend the field to be the same size as the final alloca so that
1628 // we can shift and insert it.
1629 if (SrcField->getType() != ResultVal->getType())
1630 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI);
1632 // Determine the number of bits to shift SrcField.
1634 if (Layout) // Struct case.
1635 Shift = Layout->getElementOffsetInBits(i);
1637 Shift = i*ArrayEltBitOffset;
1639 if (TD->isBigEndian())
1640 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth();
1643 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift);
1644 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI);
1647 // Don't create an 'or x, 0' on the first iteration.
1648 if (!isa<Constant>(ResultVal) ||
1649 !cast<Constant>(ResultVal)->isNullValue())
1650 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI);
1652 ResultVal = SrcField;
1655 // Handle tail padding by truncating the result
1656 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits)
1657 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI);
1659 LI->replaceAllUsesWith(ResultVal);
1660 DeadInsts.push_back(LI);
1663 /// HasPadding - Return true if the specified type has any structure or
1664 /// alignment padding, false otherwise.
1665 static bool HasPadding(const Type *Ty, const TargetData &TD) {
1666 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1667 const StructLayout *SL = TD.getStructLayout(STy);
1668 unsigned PrevFieldBitOffset = 0;
1669 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1670 unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
1672 // Padding in sub-elements?
1673 if (HasPadding(STy->getElementType(i), TD))
1676 // Check to see if there is any padding between this element and the
1679 unsigned PrevFieldEnd =
1680 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1));
1681 if (PrevFieldEnd < FieldBitOffset)
1685 PrevFieldBitOffset = FieldBitOffset;
1688 // Check for tail padding.
1689 if (unsigned EltCount = STy->getNumElements()) {
1690 unsigned PrevFieldEnd = PrevFieldBitOffset +
1691 TD.getTypeSizeInBits(STy->getElementType(EltCount-1));
1692 if (PrevFieldEnd < SL->getSizeInBits())
1696 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1697 return HasPadding(ATy->getElementType(), TD);
1698 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1699 return HasPadding(VTy->getElementType(), TD);
1701 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty);
1704 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
1705 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe,
1706 /// or 1 if safe after canonicalization has been performed.
1707 bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) {
1708 // Loop over the use list of the alloca. We can only transform it if all of
1709 // the users are safe to transform.
1712 isSafeForScalarRepl(AI, AI, 0, Info);
1713 if (Info.isUnsafe) {
1714 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n');
1718 // Okay, we know all the users are promotable. If the aggregate is a memcpy
1719 // source and destination, we have to be careful. In particular, the memcpy
1720 // could be moving around elements that live in structure padding of the LLVM
1721 // types, but may actually be used. In these cases, we refuse to promote the
1723 if (Info.isMemCpySrc && Info.isMemCpyDst &&
1724 HasPadding(AI->getAllocatedType(), *TD))
1732 /// PointsToConstantGlobal - Return true if V (possibly indirectly) points to
1733 /// some part of a constant global variable. This intentionally only accepts
1734 /// constant expressions because we don't can't rewrite arbitrary instructions.
1735 static bool PointsToConstantGlobal(Value *V) {
1736 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
1737 return GV->isConstant();
1738 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
1739 if (CE->getOpcode() == Instruction::BitCast ||
1740 CE->getOpcode() == Instruction::GetElementPtr)
1741 return PointsToConstantGlobal(CE->getOperand(0));
1745 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
1746 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
1747 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
1748 /// track of whether it moves the pointer (with isOffset) but otherwise traverse
1749 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
1750 /// the alloca, and if the source pointer is a pointer to a constant global, we
1751 /// can optimize this.
1752 static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy,
1754 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
1755 User *U = cast<Instruction>(*UI);
1757 if (LoadInst *LI = dyn_cast<LoadInst>(U))
1758 // Ignore non-volatile loads, they are always ok.
1759 if (!LI->isVolatile())
1762 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
1763 // If uses of the bitcast are ok, we are ok.
1764 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset))
1768 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
1769 // If the GEP has all zero indices, it doesn't offset the pointer. If it
1770 // doesn't, it does.
1771 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy,
1772 isOffset || !GEP->hasAllZeroIndices()))
1777 // If this is isn't our memcpy/memmove, reject it as something we can't
1779 MemTransferInst *MI = dyn_cast<MemTransferInst>(U);
1783 // If we already have seen a copy, reject the second one.
1784 if (TheCopy) return false;
1786 // If the pointer has been offset from the start of the alloca, we can't
1787 // safely handle this.
1788 if (isOffset) return false;
1790 // If the memintrinsic isn't using the alloca as the dest, reject it.
1791 if (UI.getOperandNo() != 0) return false;
1793 // If the source of the memcpy/move is not a constant global, reject it.
1794 if (!PointsToConstantGlobal(MI->getSource()))
1797 // Otherwise, the transform is safe. Remember the copy instruction.
1803 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
1804 /// modified by a copy from a constant global. If we can prove this, we can
1805 /// replace any uses of the alloca with uses of the global directly.
1806 MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) {
1807 MemTransferInst *TheCopy = 0;
1808 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false))