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 they
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/Operator.h"
33 #include "llvm/Pass.h"
34 #include "llvm/Analysis/DebugInfo.h"
35 #include "llvm/Analysis/DIBuilder.h"
36 #include "llvm/Analysis/Dominators.h"
37 #include "llvm/Analysis/Loads.h"
38 #include "llvm/Analysis/ValueTracking.h"
39 #include "llvm/Target/TargetData.h"
40 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
41 #include "llvm/Transforms/Utils/Local.h"
42 #include "llvm/Transforms/Utils/SSAUpdater.h"
43 #include "llvm/Support/CallSite.h"
44 #include "llvm/Support/Debug.h"
45 #include "llvm/Support/ErrorHandling.h"
46 #include "llvm/Support/GetElementPtrTypeIterator.h"
47 #include "llvm/Support/IRBuilder.h"
48 #include "llvm/Support/MathExtras.h"
49 #include "llvm/Support/raw_ostream.h"
50 #include "llvm/ADT/SetVector.h"
51 #include "llvm/ADT/SmallVector.h"
52 #include "llvm/ADT/Statistic.h"
55 STATISTIC(NumReplaced, "Number of allocas broken up");
56 STATISTIC(NumPromoted, "Number of allocas promoted");
57 STATISTIC(NumAdjusted, "Number of scalar allocas adjusted to allow promotion");
58 STATISTIC(NumConverted, "Number of aggregates converted to scalar");
59 STATISTIC(NumGlobals, "Number of allocas copied from constant global");
62 struct SROA : public FunctionPass {
63 SROA(int T, bool hasDT, char &ID)
64 : FunctionPass(ID), HasDomTree(hasDT) {
71 bool runOnFunction(Function &F);
73 bool performScalarRepl(Function &F);
74 bool performPromotion(Function &F);
80 /// DeadInsts - Keep track of instructions we have made dead, so that
81 /// we can remove them after we are done working.
82 SmallVector<Value*, 32> DeadInsts;
84 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures
85 /// information about the uses. All these fields are initialized to false
86 /// and set to true when something is learned.
88 /// The alloca to promote.
91 /// CheckedPHIs - This is a set of verified PHI nodes, to prevent infinite
92 /// looping and avoid redundant work.
93 SmallPtrSet<PHINode*, 8> CheckedPHIs;
95 /// isUnsafe - This is set to true if the alloca cannot be SROA'd.
98 /// isMemCpySrc - This is true if this aggregate is memcpy'd from.
101 /// isMemCpyDst - This is true if this aggregate is memcpy'd into.
102 bool isMemCpyDst : 1;
104 /// hasSubelementAccess - This is true if a subelement of the alloca is
105 /// ever accessed, or false if the alloca is only accessed with mem
106 /// intrinsics or load/store that only access the entire alloca at once.
107 bool hasSubelementAccess : 1;
109 /// hasALoadOrStore - This is true if there are any loads or stores to it.
110 /// The alloca may just be accessed with memcpy, for example, which would
112 bool hasALoadOrStore : 1;
114 explicit AllocaInfo(AllocaInst *ai)
115 : AI(ai), isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false),
116 hasSubelementAccess(false), hasALoadOrStore(false) {}
119 unsigned SRThreshold;
121 void MarkUnsafe(AllocaInfo &I, Instruction *User) {
123 DEBUG(dbgs() << " Transformation preventing inst: " << *User << '\n');
126 bool isSafeAllocaToScalarRepl(AllocaInst *AI);
128 void isSafeForScalarRepl(Instruction *I, uint64_t Offset, AllocaInfo &Info);
129 void isSafePHISelectUseForScalarRepl(Instruction *User, uint64_t Offset,
131 void isSafeGEP(GetElementPtrInst *GEPI, uint64_t &Offset, AllocaInfo &Info);
132 void isSafeMemAccess(uint64_t Offset, uint64_t MemSize,
133 Type *MemOpType, bool isStore, AllocaInfo &Info,
134 Instruction *TheAccess, bool AllowWholeAccess);
135 bool TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size);
136 uint64_t FindElementAndOffset(Type *&T, uint64_t &Offset,
139 void DoScalarReplacement(AllocaInst *AI,
140 std::vector<AllocaInst*> &WorkList);
141 void DeleteDeadInstructions();
143 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
144 SmallVector<AllocaInst*, 32> &NewElts);
145 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
146 SmallVector<AllocaInst*, 32> &NewElts);
147 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
148 SmallVector<AllocaInst*, 32> &NewElts);
149 void RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI,
151 SmallVector<AllocaInst*, 32> &NewElts);
152 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
154 SmallVector<AllocaInst*, 32> &NewElts);
155 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
156 SmallVector<AllocaInst*, 32> &NewElts);
157 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
158 SmallVector<AllocaInst*, 32> &NewElts);
160 static MemTransferInst *isOnlyCopiedFromConstantGlobal(
161 AllocaInst *AI, SmallVector<Instruction*, 4> &ToDelete);
164 // SROA_DT - SROA that uses DominatorTree.
165 struct SROA_DT : public SROA {
168 SROA_DT(int T = -1) : SROA(T, true, ID) {
169 initializeSROA_DTPass(*PassRegistry::getPassRegistry());
172 // getAnalysisUsage - This pass does not require any passes, but we know it
173 // will not alter the CFG, so say so.
174 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
175 AU.addRequired<DominatorTree>();
176 AU.setPreservesCFG();
180 // SROA_SSAUp - SROA that uses SSAUpdater.
181 struct SROA_SSAUp : public SROA {
184 SROA_SSAUp(int T = -1) : SROA(T, false, ID) {
185 initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry());
188 // getAnalysisUsage - This pass does not require any passes, but we know it
189 // will not alter the CFG, so say so.
190 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
191 AU.setPreservesCFG();
197 char SROA_DT::ID = 0;
198 char SROA_SSAUp::ID = 0;
200 INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl",
201 "Scalar Replacement of Aggregates (DT)", false, false)
202 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
203 INITIALIZE_PASS_END(SROA_DT, "scalarrepl",
204 "Scalar Replacement of Aggregates (DT)", false, false)
206 INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa",
207 "Scalar Replacement of Aggregates (SSAUp)", false, false)
208 INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa",
209 "Scalar Replacement of Aggregates (SSAUp)", false, false)
211 // Public interface to the ScalarReplAggregates pass
212 FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold,
215 return new SROA_DT(Threshold);
216 return new SROA_SSAUp(Threshold);
220 //===----------------------------------------------------------------------===//
221 // Convert To Scalar Optimization.
222 //===----------------------------------------------------------------------===//
225 /// ConvertToScalarInfo - This class implements the "Convert To Scalar"
226 /// optimization, which scans the uses of an alloca and determines if it can
227 /// rewrite it in terms of a single new alloca that can be mem2reg'd.
228 class ConvertToScalarInfo {
229 /// AllocaSize - The size of the alloca being considered in bytes.
231 const TargetData &TD;
233 /// IsNotTrivial - This is set to true if there is some access to the object
234 /// which means that mem2reg can't promote it.
237 /// ScalarKind - Tracks the kind of alloca being considered for promotion,
238 /// computed based on the uses of the alloca rather than the LLVM type system.
242 // Accesses via GEPs that are consistent with element access of a vector
243 // type. This will not be converted into a vector unless there is a later
244 // access using an actual vector type.
247 // Accesses via vector operations and GEPs that are consistent with the
248 // layout of a vector type.
251 // An integer bag-of-bits with bitwise operations for insertion and
252 // extraction. Any combination of types can be converted into this kind
257 /// VectorTy - This tracks the type that we should promote the vector to if
258 /// it is possible to turn it into a vector. This starts out null, and if it
259 /// isn't possible to turn into a vector type, it gets set to VoidTy.
260 VectorType *VectorTy;
262 /// HadNonMemTransferAccess - True if there is at least one access to the
263 /// alloca that is not a MemTransferInst. We don't want to turn structs into
264 /// large integers unless there is some potential for optimization.
265 bool HadNonMemTransferAccess;
268 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td)
269 : AllocaSize(Size), TD(td), IsNotTrivial(false), ScalarKind(Unknown),
270 VectorTy(0), HadNonMemTransferAccess(false) { }
272 AllocaInst *TryConvert(AllocaInst *AI);
275 bool CanConvertToScalar(Value *V, uint64_t Offset);
276 void MergeInTypeForLoadOrStore(Type *In, uint64_t Offset);
277 bool MergeInVectorType(VectorType *VInTy, uint64_t Offset);
278 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset);
280 Value *ConvertScalar_ExtractValue(Value *NV, Type *ToType,
281 uint64_t Offset, IRBuilder<> &Builder);
282 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal,
283 uint64_t Offset, IRBuilder<> &Builder);
285 } // end anonymous namespace.
288 /// TryConvert - Analyze the specified alloca, and if it is safe to do so,
289 /// rewrite it to be a new alloca which is mem2reg'able. This returns the new
290 /// alloca if possible or null if not.
291 AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) {
292 // If we can't convert this scalar, or if mem2reg can trivially do it, bail
294 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial)
297 // If an alloca has only memset / memcpy uses, it may still have an Unknown
298 // ScalarKind. Treat it as an Integer below.
299 if (ScalarKind == Unknown)
300 ScalarKind = Integer;
302 if (ScalarKind == Vector && VectorTy->getBitWidth() != AllocaSize * 8)
303 ScalarKind = Integer;
305 // If we were able to find a vector type that can handle this with
306 // insert/extract elements, and if there was at least one use that had
307 // a vector type, promote this to a vector. We don't want to promote
308 // random stuff that doesn't use vectors (e.g. <9 x double>) because then
309 // we just get a lot of insert/extracts. If at least one vector is
310 // involved, then we probably really do have a union of vector/array.
312 if (ScalarKind == Vector) {
313 assert(VectorTy && "Missing type for vector scalar.");
314 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = "
315 << *VectorTy << '\n');
316 NewTy = VectorTy; // Use the vector type.
318 unsigned BitWidth = AllocaSize * 8;
319 if ((ScalarKind == ImplicitVector || ScalarKind == Integer) &&
320 !HadNonMemTransferAccess && !TD.fitsInLegalInteger(BitWidth))
323 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n");
324 // Create and insert the integer alloca.
325 NewTy = IntegerType::get(AI->getContext(), BitWidth);
327 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin());
328 ConvertUsesToScalar(AI, NewAI, 0);
332 /// MergeInTypeForLoadOrStore - Add the 'In' type to the accumulated vector type
333 /// (VectorTy) so far at the offset specified by Offset (which is specified in
336 /// There are two cases we handle here:
337 /// 1) A union of vector types of the same size and potentially its elements.
338 /// Here we turn element accesses into insert/extract element operations.
339 /// This promotes a <4 x float> with a store of float to the third element
340 /// into a <4 x float> that uses insert element.
341 /// 2) A fully general blob of memory, which we turn into some (potentially
342 /// large) integer type with extract and insert operations where the loads
343 /// and stores would mutate the memory. We mark this by setting VectorTy
345 void ConvertToScalarInfo::MergeInTypeForLoadOrStore(Type *In,
347 // If we already decided to turn this into a blob of integer memory, there is
348 // nothing to be done.
349 if (ScalarKind == Integer)
352 // If this could be contributing to a vector, analyze it.
354 // If the In type is a vector that is the same size as the alloca, see if it
355 // matches the existing VecTy.
356 if (VectorType *VInTy = dyn_cast<VectorType>(In)) {
357 if (MergeInVectorType(VInTy, Offset))
359 } else if (In->isFloatTy() || In->isDoubleTy() ||
360 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 &&
361 isPowerOf2_32(In->getPrimitiveSizeInBits()))) {
362 // Full width accesses can be ignored, because they can always be turned
364 unsigned EltSize = In->getPrimitiveSizeInBits()/8;
365 if (EltSize == AllocaSize)
368 // If we're accessing something that could be an element of a vector, see
369 // if the implied vector agrees with what we already have and if Offset is
370 // compatible with it.
371 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 &&
372 (!VectorTy || EltSize == VectorTy->getElementType()
373 ->getPrimitiveSizeInBits()/8)) {
375 ScalarKind = ImplicitVector;
376 VectorTy = VectorType::get(In, AllocaSize/EltSize);
382 // Otherwise, we have a case that we can't handle with an optimized vector
383 // form. We can still turn this into a large integer.
384 ScalarKind = Integer;
387 /// MergeInVectorType - Handles the vector case of MergeInTypeForLoadOrStore,
388 /// returning true if the type was successfully merged and false otherwise.
389 bool ConvertToScalarInfo::MergeInVectorType(VectorType *VInTy,
391 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) {
392 // If we're storing/loading a vector of the right size, allow it as a
393 // vector. If this the first vector we see, remember the type so that
394 // we know the element size. If this is a subsequent access, ignore it
395 // even if it is a differing type but the same size. Worst case we can
396 // bitcast the resultant vectors.
406 /// CanConvertToScalar - V is a pointer. If we can convert the pointee and all
407 /// its accesses to a single vector type, return true and set VecTy to
408 /// the new type. If we could convert the alloca into a single promotable
409 /// integer, return true but set VecTy to VoidTy. Further, if the use is not a
410 /// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset
411 /// is the current offset from the base of the alloca being analyzed.
413 /// If we see at least one access to the value that is as a vector type, set the
415 bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) {
416 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
417 Instruction *User = cast<Instruction>(*UI);
419 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
420 // Don't break volatile loads.
423 // Don't touch MMX operations.
424 if (LI->getType()->isX86_MMXTy())
426 HadNonMemTransferAccess = true;
427 MergeInTypeForLoadOrStore(LI->getType(), Offset);
431 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
432 // Storing the pointer, not into the value?
433 if (SI->getOperand(0) == V || !SI->isSimple()) return false;
434 // Don't touch MMX operations.
435 if (SI->getOperand(0)->getType()->isX86_MMXTy())
437 HadNonMemTransferAccess = true;
438 MergeInTypeForLoadOrStore(SI->getOperand(0)->getType(), Offset);
442 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
443 if (!onlyUsedByLifetimeMarkers(BCI))
444 IsNotTrivial = true; // Can't be mem2reg'd.
445 if (!CanConvertToScalar(BCI, Offset))
450 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
451 // If this is a GEP with a variable indices, we can't handle it.
452 if (!GEP->hasAllConstantIndices())
455 // Compute the offset that this GEP adds to the pointer.
456 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
457 if (!GEP->getPointerOperandType()->isPointerTy())
459 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(),
461 // See if all uses can be converted.
462 if (!CanConvertToScalar(GEP, Offset+GEPOffset))
464 IsNotTrivial = true; // Can't be mem2reg'd.
465 HadNonMemTransferAccess = true;
469 // If this is a constant sized memset of a constant value (e.g. 0) we can
471 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
472 // Store of constant value.
473 if (!isa<ConstantInt>(MSI->getValue()))
476 // Store of constant size.
477 ConstantInt *Len = dyn_cast<ConstantInt>(MSI->getLength());
481 // If the size differs from the alloca, we can only convert the alloca to
482 // an integer bag-of-bits.
483 // FIXME: This should handle all of the cases that are currently accepted
484 // as vector element insertions.
485 if (Len->getZExtValue() != AllocaSize || Offset != 0)
486 ScalarKind = Integer;
488 IsNotTrivial = true; // Can't be mem2reg'd.
489 HadNonMemTransferAccess = true;
493 // If this is a memcpy or memmove into or out of the whole allocation, we
494 // can handle it like a load or store of the scalar type.
495 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
496 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength());
497 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0)
500 IsNotTrivial = true; // Can't be mem2reg'd.
504 // If this is a lifetime intrinsic, we can handle it.
505 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
506 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
507 II->getIntrinsicID() == Intrinsic::lifetime_end) {
512 // Otherwise, we cannot handle this!
519 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
520 /// directly. This happens when we are converting an "integer union" to a
521 /// single integer scalar, or when we are converting a "vector union" to a
522 /// vector with insert/extractelement instructions.
524 /// Offset is an offset from the original alloca, in bits that need to be
525 /// shifted to the right. By the end of this, there should be no uses of Ptr.
526 void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI,
528 while (!Ptr->use_empty()) {
529 Instruction *User = cast<Instruction>(Ptr->use_back());
531 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
532 ConvertUsesToScalar(CI, NewAI, Offset);
533 CI->eraseFromParent();
537 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
538 // Compute the offset that this GEP adds to the pointer.
539 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
540 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(),
542 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8);
543 GEP->eraseFromParent();
547 IRBuilder<> Builder(User);
549 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
550 // The load is a bit extract from NewAI shifted right by Offset bits.
551 Value *LoadedVal = Builder.CreateLoad(NewAI);
553 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder);
554 LI->replaceAllUsesWith(NewLoadVal);
555 LI->eraseFromParent();
559 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
560 assert(SI->getOperand(0) != Ptr && "Consistency error!");
561 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
562 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset,
564 Builder.CreateStore(New, NewAI);
565 SI->eraseFromParent();
567 // If the load we just inserted is now dead, then the inserted store
568 // overwrote the entire thing.
569 if (Old->use_empty())
570 Old->eraseFromParent();
574 // If this is a constant sized memset of a constant value (e.g. 0) we can
575 // transform it into a store of the expanded constant value.
576 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
577 assert(MSI->getRawDest() == Ptr && "Consistency error!");
578 int64_t SNumBytes = cast<ConstantInt>(MSI->getLength())->getSExtValue();
579 if (SNumBytes > 0 && (SNumBytes >> 32) == 0) {
580 unsigned NumBytes = static_cast<unsigned>(SNumBytes);
581 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue();
583 // Compute the value replicated the right number of times.
584 APInt APVal(NumBytes*8, Val);
586 // Splat the value if non-zero.
588 for (unsigned i = 1; i != NumBytes; ++i)
591 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
592 Value *New = ConvertScalar_InsertValue(
593 ConstantInt::get(User->getContext(), APVal),
594 Old, Offset, Builder);
595 Builder.CreateStore(New, NewAI);
597 // If the load we just inserted is now dead, then the memset overwrote
599 if (Old->use_empty())
600 Old->eraseFromParent();
602 MSI->eraseFromParent();
606 // If this is a memcpy or memmove into or out of the whole allocation, we
607 // can handle it like a load or store of the scalar type.
608 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
609 assert(Offset == 0 && "must be store to start of alloca");
611 // If the source and destination are both to the same alloca, then this is
612 // a noop copy-to-self, just delete it. Otherwise, emit a load and store
614 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &TD, 0));
616 if (GetUnderlyingObject(MTI->getSource(), &TD, 0) != OrigAI) {
617 // Dest must be OrigAI, change this to be a load from the original
618 // pointer (bitcasted), then a store to our new alloca.
619 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?");
620 Value *SrcPtr = MTI->getSource();
621 PointerType* SPTy = cast<PointerType>(SrcPtr->getType());
622 PointerType* AIPTy = cast<PointerType>(NewAI->getType());
623 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
624 AIPTy = PointerType::get(AIPTy->getElementType(),
625 SPTy->getAddressSpace());
627 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy);
629 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval");
630 SrcVal->setAlignment(MTI->getAlignment());
631 Builder.CreateStore(SrcVal, NewAI);
632 } else if (GetUnderlyingObject(MTI->getDest(), &TD, 0) != OrigAI) {
633 // Src must be OrigAI, change this to be a load from NewAI then a store
634 // through the original dest pointer (bitcasted).
635 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?");
636 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval");
638 PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType());
639 PointerType* AIPTy = cast<PointerType>(NewAI->getType());
640 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
641 AIPTy = PointerType::get(AIPTy->getElementType(),
642 DPTy->getAddressSpace());
644 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy);
646 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr);
647 NewStore->setAlignment(MTI->getAlignment());
649 // Noop transfer. Src == Dst
652 MTI->eraseFromParent();
656 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
657 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
658 II->getIntrinsicID() == Intrinsic::lifetime_end) {
659 // There's no need to preserve these, as the resulting alloca will be
660 // converted to a register anyways.
661 II->eraseFromParent();
666 llvm_unreachable("Unsupported operation!");
670 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer
671 /// or vector value FromVal, extracting the bits from the offset specified by
672 /// Offset. This returns the value, which is of type ToType.
674 /// This happens when we are converting an "integer union" to a single
675 /// integer scalar, or when we are converting a "vector union" to a vector with
676 /// insert/extractelement instructions.
678 /// Offset is an offset from the original alloca, in bits that need to be
679 /// shifted to the right.
680 Value *ConvertToScalarInfo::
681 ConvertScalar_ExtractValue(Value *FromVal, Type *ToType,
682 uint64_t Offset, IRBuilder<> &Builder) {
683 // If the load is of the whole new alloca, no conversion is needed.
684 Type *FromType = FromVal->getType();
685 if (FromType == ToType && Offset == 0)
688 // If the result alloca is a vector type, this is either an element
689 // access or a bitcast to another vector type of the same size.
690 if (VectorType *VTy = dyn_cast<VectorType>(FromType)) {
691 unsigned FromTypeSize = TD.getTypeAllocSize(FromType);
692 unsigned ToTypeSize = TD.getTypeAllocSize(ToType);
693 if (FromTypeSize == ToTypeSize)
694 return Builder.CreateBitCast(FromVal, ToType);
696 // Otherwise it must be an element access.
699 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType());
700 Elt = Offset/EltSize;
701 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
703 // Return the element extracted out of it.
704 Value *V = Builder.CreateExtractElement(FromVal, Builder.getInt32(Elt));
705 if (V->getType() != ToType)
706 V = Builder.CreateBitCast(V, ToType);
710 // If ToType is a first class aggregate, extract out each of the pieces and
711 // use insertvalue's to form the FCA.
712 if (StructType *ST = dyn_cast<StructType>(ToType)) {
713 const StructLayout &Layout = *TD.getStructLayout(ST);
714 Value *Res = UndefValue::get(ST);
715 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
716 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i),
717 Offset+Layout.getElementOffsetInBits(i),
719 Res = Builder.CreateInsertValue(Res, Elt, i);
724 if (ArrayType *AT = dyn_cast<ArrayType>(ToType)) {
725 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType());
726 Value *Res = UndefValue::get(AT);
727 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
728 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(),
729 Offset+i*EltSize, Builder);
730 Res = Builder.CreateInsertValue(Res, Elt, i);
735 // Otherwise, this must be a union that was converted to an integer value.
736 IntegerType *NTy = cast<IntegerType>(FromVal->getType());
738 // If this is a big-endian system and the load is narrower than the
739 // full alloca type, we need to do a shift to get the right bits.
741 if (TD.isBigEndian()) {
742 // On big-endian machines, the lowest bit is stored at the bit offset
743 // from the pointer given by getTypeStoreSizeInBits. This matters for
744 // integers with a bitwidth that is not a multiple of 8.
745 ShAmt = TD.getTypeStoreSizeInBits(NTy) -
746 TD.getTypeStoreSizeInBits(ToType) - Offset;
751 // Note: we support negative bitwidths (with shl) which are not defined.
752 // We do this to support (f.e.) loads off the end of a structure where
753 // only some bits are used.
754 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth())
755 FromVal = Builder.CreateLShr(FromVal,
756 ConstantInt::get(FromVal->getType(), ShAmt));
757 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
758 FromVal = Builder.CreateShl(FromVal,
759 ConstantInt::get(FromVal->getType(), -ShAmt));
761 // Finally, unconditionally truncate the integer to the right width.
762 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType);
763 if (LIBitWidth < NTy->getBitWidth())
765 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(),
767 else if (LIBitWidth > NTy->getBitWidth())
769 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(),
772 // If the result is an integer, this is a trunc or bitcast.
773 if (ToType->isIntegerTy()) {
775 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) {
776 // Just do a bitcast, we know the sizes match up.
777 FromVal = Builder.CreateBitCast(FromVal, ToType);
779 // Otherwise must be a pointer.
780 FromVal = Builder.CreateIntToPtr(FromVal, ToType);
782 assert(FromVal->getType() == ToType && "Didn't convert right?");
786 /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer
787 /// or vector value "Old" at the offset specified by Offset.
789 /// This happens when we are converting an "integer union" to a
790 /// single integer scalar, or when we are converting a "vector union" to a
791 /// vector with insert/extractelement instructions.
793 /// Offset is an offset from the original alloca, in bits that need to be
794 /// shifted to the right.
795 Value *ConvertToScalarInfo::
796 ConvertScalar_InsertValue(Value *SV, Value *Old,
797 uint64_t Offset, IRBuilder<> &Builder) {
798 // Convert the stored type to the actual type, shift it left to insert
799 // then 'or' into place.
800 Type *AllocaType = Old->getType();
801 LLVMContext &Context = Old->getContext();
803 if (VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
804 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy);
805 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType());
807 // Changing the whole vector with memset or with an access of a different
809 if (ValSize == VecSize)
810 return Builder.CreateBitCast(SV, AllocaType);
812 // Must be an element insertion.
813 Type *EltTy = VTy->getElementType();
814 if (SV->getType() != EltTy)
815 SV = Builder.CreateBitCast(SV, EltTy);
816 uint64_t EltSize = TD.getTypeAllocSizeInBits(EltTy);
817 unsigned Elt = Offset/EltSize;
818 return Builder.CreateInsertElement(Old, SV, Builder.getInt32(Elt));
821 // If SV is a first-class aggregate value, insert each value recursively.
822 if (StructType *ST = dyn_cast<StructType>(SV->getType())) {
823 const StructLayout &Layout = *TD.getStructLayout(ST);
824 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
825 Value *Elt = Builder.CreateExtractValue(SV, i);
826 Old = ConvertScalar_InsertValue(Elt, Old,
827 Offset+Layout.getElementOffsetInBits(i),
833 if (ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) {
834 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType());
835 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
836 Value *Elt = Builder.CreateExtractValue(SV, i);
837 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder);
842 // If SV is a float, convert it to the appropriate integer type.
843 // If it is a pointer, do the same.
844 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType());
845 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType);
846 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType());
847 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType);
848 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy())
849 SV = Builder.CreateBitCast(SV, IntegerType::get(SV->getContext(),SrcWidth));
850 else if (SV->getType()->isPointerTy())
851 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()));
853 // Zero extend or truncate the value if needed.
854 if (SV->getType() != AllocaType) {
855 if (SV->getType()->getPrimitiveSizeInBits() <
856 AllocaType->getPrimitiveSizeInBits())
857 SV = Builder.CreateZExt(SV, AllocaType);
859 // Truncation may be needed if storing more than the alloca can hold
860 // (undefined behavior).
861 SV = Builder.CreateTrunc(SV, AllocaType);
862 SrcWidth = DestWidth;
863 SrcStoreWidth = DestStoreWidth;
867 // If this is a big-endian system and the store is narrower than the
868 // full alloca type, we need to do a shift to get the right bits.
870 if (TD.isBigEndian()) {
871 // On big-endian machines, the lowest bit is stored at the bit offset
872 // from the pointer given by getTypeStoreSizeInBits. This matters for
873 // integers with a bitwidth that is not a multiple of 8.
874 ShAmt = DestStoreWidth - SrcStoreWidth - Offset;
879 // Note: we support negative bitwidths (with shr) which are not defined.
880 // We do this to support (f.e.) stores off the end of a structure where
881 // only some bits in the structure are set.
882 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
883 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
884 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), ShAmt));
886 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
887 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), -ShAmt));
888 Mask = Mask.lshr(-ShAmt);
891 // Mask out the bits we are about to insert from the old value, and or
893 if (SrcWidth != DestWidth) {
894 assert(DestWidth > SrcWidth);
895 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask");
896 SV = Builder.CreateOr(Old, SV, "ins");
902 //===----------------------------------------------------------------------===//
904 //===----------------------------------------------------------------------===//
907 bool SROA::runOnFunction(Function &F) {
908 TD = getAnalysisIfAvailable<TargetData>();
910 bool Changed = performPromotion(F);
912 // FIXME: ScalarRepl currently depends on TargetData more than it
913 // theoretically needs to. It should be refactored in order to support
914 // target-independent IR. Until this is done, just skip the actual
915 // scalar-replacement portion of this pass.
916 if (!TD) return Changed;
919 bool LocalChange = performScalarRepl(F);
920 if (!LocalChange) break; // No need to repromote if no scalarrepl
922 LocalChange = performPromotion(F);
923 if (!LocalChange) break; // No need to re-scalarrepl if no promotion
930 class AllocaPromoter : public LoadAndStorePromoter {
933 SmallVector<DbgDeclareInst *, 4> DDIs;
934 SmallVector<DbgValueInst *, 4> DVIs;
936 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
938 : LoadAndStorePromoter(Insts, S), AI(0), DIB(DB) {}
940 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) {
941 // Remember which alloca we're promoting (for isInstInList).
943 if (MDNode *DebugNode = MDNode::getIfExists(AI->getContext(), AI)) {
944 for (Value::use_iterator UI = DebugNode->use_begin(),
945 E = DebugNode->use_end(); UI != E; ++UI)
946 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
948 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI))
952 LoadAndStorePromoter::run(Insts);
953 AI->eraseFromParent();
954 for (SmallVector<DbgDeclareInst *, 4>::iterator I = DDIs.begin(),
955 E = DDIs.end(); I != E; ++I) {
956 DbgDeclareInst *DDI = *I;
957 DDI->eraseFromParent();
959 for (SmallVector<DbgValueInst *, 4>::iterator I = DVIs.begin(),
960 E = DVIs.end(); I != E; ++I) {
961 DbgValueInst *DVI = *I;
962 DVI->eraseFromParent();
966 virtual bool isInstInList(Instruction *I,
967 const SmallVectorImpl<Instruction*> &Insts) const {
968 if (LoadInst *LI = dyn_cast<LoadInst>(I))
969 return LI->getOperand(0) == AI;
970 return cast<StoreInst>(I)->getPointerOperand() == AI;
973 virtual void updateDebugInfo(Instruction *Inst) const {
974 for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(),
975 E = DDIs.end(); I != E; ++I) {
976 DbgDeclareInst *DDI = *I;
977 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
978 ConvertDebugDeclareToDebugValue(DDI, SI, *DIB);
979 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
980 ConvertDebugDeclareToDebugValue(DDI, LI, *DIB);
982 for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(),
983 E = DVIs.end(); I != E; ++I) {
984 DbgValueInst *DVI = *I;
986 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
987 // If an argument is zero extended then use argument directly. The ZExt
988 // may be zapped by an optimization pass in future.
989 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
990 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
991 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
992 Arg = dyn_cast<Argument>(SExt->getOperand(0));
994 Arg = SI->getOperand(0);
995 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
996 Arg = LI->getOperand(0);
1000 Instruction *DbgVal =
1001 DIB->insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
1003 DbgVal->setDebugLoc(DVI->getDebugLoc());
1007 } // end anon namespace
1009 /// isSafeSelectToSpeculate - Select instructions that use an alloca and are
1010 /// subsequently loaded can be rewritten to load both input pointers and then
1011 /// select between the result, allowing the load of the alloca to be promoted.
1013 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1014 /// %V = load i32* %P2
1016 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1017 /// %V2 = load i32* %Other
1018 /// %V = select i1 %cond, i32 %V1, i32 %V2
1020 /// We can do this to a select if its only uses are loads and if the operand to
1021 /// the select can be loaded unconditionally.
1022 static bool isSafeSelectToSpeculate(SelectInst *SI, const TargetData *TD) {
1023 bool TDerefable = SI->getTrueValue()->isDereferenceablePointer();
1024 bool FDerefable = SI->getFalseValue()->isDereferenceablePointer();
1026 for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end();
1028 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1029 if (LI == 0 || !LI->isSimple()) return false;
1031 // Both operands to the select need to be dereferencable, either absolutely
1032 // (e.g. allocas) or at this point because we can see other accesses to it.
1033 if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI,
1034 LI->getAlignment(), TD))
1036 if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI,
1037 LI->getAlignment(), TD))
1044 /// isSafePHIToSpeculate - PHI instructions that use an alloca and are
1045 /// subsequently loaded can be rewritten to load both input pointers in the pred
1046 /// blocks and then PHI the results, allowing the load of the alloca to be
1049 /// %P2 = phi [i32* %Alloca, i32* %Other]
1050 /// %V = load i32* %P2
1052 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1054 /// %V2 = load i32* %Other
1056 /// %V = phi [i32 %V1, i32 %V2]
1058 /// We can do this to a select if its only uses are loads and if the operand to
1059 /// the select can be loaded unconditionally.
1060 static bool isSafePHIToSpeculate(PHINode *PN, const TargetData *TD) {
1061 // For now, we can only do this promotion if the load is in the same block as
1062 // the PHI, and if there are no stores between the phi and load.
1063 // TODO: Allow recursive phi users.
1064 // TODO: Allow stores.
1065 BasicBlock *BB = PN->getParent();
1066 unsigned MaxAlign = 0;
1067 for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end();
1069 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1070 if (LI == 0 || !LI->isSimple()) return false;
1072 // For now we only allow loads in the same block as the PHI. This is a
1073 // common case that happens when instcombine merges two loads through a PHI.
1074 if (LI->getParent() != BB) return false;
1076 // Ensure that there are no instructions between the PHI and the load that
1078 for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI)
1079 if (BBI->mayWriteToMemory())
1082 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1085 // Okay, we know that we have one or more loads in the same block as the PHI.
1086 // We can transform this if it is safe to push the loads into the predecessor
1087 // blocks. The only thing to watch out for is that we can't put a possibly
1088 // trapping load in the predecessor if it is a critical edge.
1089 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1090 BasicBlock *Pred = PN->getIncomingBlock(i);
1091 Value *InVal = PN->getIncomingValue(i);
1093 // If the terminator of the predecessor has side-effects (an invoke),
1094 // there is no safe place to put a load in the predecessor.
1095 if (Pred->getTerminator()->mayHaveSideEffects())
1098 // If the value is produced by the terminator of the predecessor
1099 // (an invoke), there is no valid place to put a load in the predecessor.
1100 if (Pred->getTerminator() == InVal)
1103 // If the predecessor has a single successor, then the edge isn't critical.
1104 if (Pred->getTerminator()->getNumSuccessors() == 1)
1107 // If this pointer is always safe to load, or if we can prove that there is
1108 // already a load in the block, then we can move the load to the pred block.
1109 if (InVal->isDereferenceablePointer() ||
1110 isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, TD))
1120 /// tryToMakeAllocaBePromotable - This returns true if the alloca only has
1121 /// direct (non-volatile) loads and stores to it. If the alloca is close but
1122 /// not quite there, this will transform the code to allow promotion. As such,
1123 /// it is a non-pure predicate.
1124 static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) {
1125 SetVector<Instruction*, SmallVector<Instruction*, 4>,
1126 SmallPtrSet<Instruction*, 4> > InstsToRewrite;
1128 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
1131 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
1132 if (!LI->isSimple())
1137 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1138 if (SI->getOperand(0) == AI || !SI->isSimple())
1139 return false; // Don't allow a store OF the AI, only INTO the AI.
1143 if (SelectInst *SI = dyn_cast<SelectInst>(U)) {
1144 // If the condition being selected on is a constant, fold the select, yes
1145 // this does (rarely) happen early on.
1146 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) {
1147 Value *Result = SI->getOperand(1+CI->isZero());
1148 SI->replaceAllUsesWith(Result);
1149 SI->eraseFromParent();
1151 // This is very rare and we just scrambled the use list of AI, start
1153 return tryToMakeAllocaBePromotable(AI, TD);
1156 // If it is safe to turn "load (select c, AI, ptr)" into a select of two
1157 // loads, then we can transform this by rewriting the select.
1158 if (!isSafeSelectToSpeculate(SI, TD))
1161 InstsToRewrite.insert(SI);
1165 if (PHINode *PN = dyn_cast<PHINode>(U)) {
1166 if (PN->use_empty()) { // Dead PHIs can be stripped.
1167 InstsToRewrite.insert(PN);
1171 // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads
1172 // in the pred blocks, then we can transform this by rewriting the PHI.
1173 if (!isSafePHIToSpeculate(PN, TD))
1176 InstsToRewrite.insert(PN);
1180 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
1181 if (onlyUsedByLifetimeMarkers(BCI)) {
1182 InstsToRewrite.insert(BCI);
1190 // If there are no instructions to rewrite, then all uses are load/stores and
1192 if (InstsToRewrite.empty())
1195 // If we have instructions that need to be rewritten for this to be promotable
1196 // take care of it now.
1197 for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) {
1198 if (BitCastInst *BCI = dyn_cast<BitCastInst>(InstsToRewrite[i])) {
1199 // This could only be a bitcast used by nothing but lifetime intrinsics.
1200 for (BitCastInst::use_iterator I = BCI->use_begin(), E = BCI->use_end();
1202 Use &U = I.getUse();
1204 cast<Instruction>(U.getUser())->eraseFromParent();
1206 BCI->eraseFromParent();
1210 if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) {
1211 // Selects in InstsToRewrite only have load uses. Rewrite each as two
1212 // loads with a new select.
1213 while (!SI->use_empty()) {
1214 LoadInst *LI = cast<LoadInst>(SI->use_back());
1216 IRBuilder<> Builder(LI);
1217 LoadInst *TrueLoad =
1218 Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t");
1219 LoadInst *FalseLoad =
1220 Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".f");
1222 // Transfer alignment and TBAA info if present.
1223 TrueLoad->setAlignment(LI->getAlignment());
1224 FalseLoad->setAlignment(LI->getAlignment());
1225 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
1226 TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1227 FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1230 Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad);
1232 LI->replaceAllUsesWith(V);
1233 LI->eraseFromParent();
1236 // Now that all the loads are gone, the select is gone too.
1237 SI->eraseFromParent();
1241 // Otherwise, we have a PHI node which allows us to push the loads into the
1243 PHINode *PN = cast<PHINode>(InstsToRewrite[i]);
1244 if (PN->use_empty()) {
1245 PN->eraseFromParent();
1249 Type *LoadTy = cast<PointerType>(PN->getType())->getElementType();
1250 PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(),
1251 PN->getName()+".ld", PN);
1253 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1254 // matter which one we get and if any differ, it doesn't matter.
1255 LoadInst *SomeLoad = cast<LoadInst>(PN->use_back());
1256 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1257 unsigned Align = SomeLoad->getAlignment();
1259 // Rewrite all loads of the PN to use the new PHI.
1260 while (!PN->use_empty()) {
1261 LoadInst *LI = cast<LoadInst>(PN->use_back());
1262 LI->replaceAllUsesWith(NewPN);
1263 LI->eraseFromParent();
1266 // Inject loads into all of the pred blocks. Keep track of which blocks we
1267 // insert them into in case we have multiple edges from the same block.
1268 DenseMap<BasicBlock*, LoadInst*> InsertedLoads;
1270 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1271 BasicBlock *Pred = PN->getIncomingBlock(i);
1272 LoadInst *&Load = InsertedLoads[Pred];
1274 Load = new LoadInst(PN->getIncomingValue(i),
1275 PN->getName() + "." + Pred->getName(),
1276 Pred->getTerminator());
1277 Load->setAlignment(Align);
1278 if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1281 NewPN->addIncoming(Load, Pred);
1284 PN->eraseFromParent();
1291 bool SROA::performPromotion(Function &F) {
1292 std::vector<AllocaInst*> Allocas;
1293 DominatorTree *DT = 0;
1295 DT = &getAnalysis<DominatorTree>();
1297 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
1298 DIBuilder DIB(*F.getParent());
1299 bool Changed = false;
1300 SmallVector<Instruction*, 64> Insts;
1304 // Find allocas that are safe to promote, by looking at all instructions in
1306 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
1307 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
1308 if (tryToMakeAllocaBePromotable(AI, TD))
1309 Allocas.push_back(AI);
1311 if (Allocas.empty()) break;
1314 PromoteMemToReg(Allocas, *DT);
1317 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
1318 AllocaInst *AI = Allocas[i];
1320 // Build list of instructions to promote.
1321 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
1323 Insts.push_back(cast<Instruction>(*UI));
1324 AllocaPromoter(Insts, SSA, &DIB).run(AI, Insts);
1328 NumPromoted += Allocas.size();
1336 /// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for
1337 /// SROA. It must be a struct or array type with a small number of elements.
1338 static bool ShouldAttemptScalarRepl(AllocaInst *AI) {
1339 Type *T = AI->getAllocatedType();
1340 // Do not promote any struct into more than 32 separate vars.
1341 if (StructType *ST = dyn_cast<StructType>(T))
1342 return ST->getNumElements() <= 32;
1343 // Arrays are much less likely to be safe for SROA; only consider
1344 // them if they are very small.
1345 if (ArrayType *AT = dyn_cast<ArrayType>(T))
1346 return AT->getNumElements() <= 8;
1350 /// getPointeeAlignment - Compute the minimum alignment of the value pointed
1351 /// to by the given pointer.
1352 static unsigned getPointeeAlignment(Value *V, const TargetData &TD) {
1353 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
1354 if (CE->getOpcode() == Instruction::BitCast ||
1355 (CE->getOpcode() == Instruction::GetElementPtr &&
1356 cast<GEPOperator>(CE)->hasAllZeroIndices()))
1357 return getPointeeAlignment(CE->getOperand(0), TD);
1359 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
1360 if (!GV->isDeclaration())
1361 return TD.getPreferredAlignment(GV);
1363 if (PointerType *PT = dyn_cast<PointerType>(V->getType()))
1364 return TD.getABITypeAlignment(PT->getElementType());
1370 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
1371 // which runs on all of the alloca instructions in the function, removing them
1372 // if they are only used by getelementptr instructions.
1374 bool SROA::performScalarRepl(Function &F) {
1375 std::vector<AllocaInst*> WorkList;
1377 // Scan the entry basic block, adding allocas to the worklist.
1378 BasicBlock &BB = F.getEntryBlock();
1379 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
1380 if (AllocaInst *A = dyn_cast<AllocaInst>(I))
1381 WorkList.push_back(A);
1383 // Process the worklist
1384 bool Changed = false;
1385 while (!WorkList.empty()) {
1386 AllocaInst *AI = WorkList.back();
1387 WorkList.pop_back();
1389 // Handle dead allocas trivially. These can be formed by SROA'ing arrays
1390 // with unused elements.
1391 if (AI->use_empty()) {
1392 AI->eraseFromParent();
1397 // If this alloca is impossible for us to promote, reject it early.
1398 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized())
1401 // Check to see if this allocation is only modified by a memcpy/memmove from
1402 // a constant global whose alignment is equal to or exceeds that of the
1403 // allocation. If this is the case, we can change all users to use
1404 // the constant global instead. This is commonly produced by the CFE by
1405 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
1406 // is only subsequently read.
1407 SmallVector<Instruction *, 4> ToDelete;
1408 if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(AI, ToDelete)) {
1409 if (AI->getAlignment() <= getPointeeAlignment(Copy->getSource(), *TD)) {
1410 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n');
1411 DEBUG(dbgs() << " memcpy = " << *Copy << '\n');
1412 for (unsigned i = 0, e = ToDelete.size(); i != e; ++i)
1413 ToDelete[i]->eraseFromParent();
1414 Constant *TheSrc = cast<Constant>(Copy->getSource());
1415 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType()));
1416 Copy->eraseFromParent(); // Don't mutate the global.
1417 AI->eraseFromParent();
1424 // Check to see if we can perform the core SROA transformation. We cannot
1425 // transform the allocation instruction if it is an array allocation
1426 // (allocations OF arrays are ok though), and an allocation of a scalar
1427 // value cannot be decomposed at all.
1428 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType());
1430 // Do not promote [0 x %struct].
1431 if (AllocaSize == 0) continue;
1433 // Do not promote any struct whose size is too big.
1434 if (AllocaSize > SRThreshold) continue;
1436 // If the alloca looks like a good candidate for scalar replacement, and if
1437 // all its users can be transformed, then split up the aggregate into its
1438 // separate elements.
1439 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) {
1440 DoScalarReplacement(AI, WorkList);
1445 // If we can turn this aggregate value (potentially with casts) into a
1446 // simple scalar value that can be mem2reg'd into a register value.
1447 // IsNotTrivial tracks whether this is something that mem2reg could have
1448 // promoted itself. If so, we don't want to transform it needlessly. Note
1449 // that we can't just check based on the type: the alloca may be of an i32
1450 // but that has pointer arithmetic to set byte 3 of it or something.
1451 if (AllocaInst *NewAI =
1452 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) {
1453 NewAI->takeName(AI);
1454 AI->eraseFromParent();
1460 // Otherwise, couldn't process this alloca.
1466 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl
1467 /// predicate, do SROA now.
1468 void SROA::DoScalarReplacement(AllocaInst *AI,
1469 std::vector<AllocaInst*> &WorkList) {
1470 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n');
1471 SmallVector<AllocaInst*, 32> ElementAllocas;
1472 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
1473 ElementAllocas.reserve(ST->getNumContainedTypes());
1474 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
1475 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
1477 AI->getName() + "." + Twine(i), AI);
1478 ElementAllocas.push_back(NA);
1479 WorkList.push_back(NA); // Add to worklist for recursive processing
1482 ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
1483 ElementAllocas.reserve(AT->getNumElements());
1484 Type *ElTy = AT->getElementType();
1485 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1486 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
1487 AI->getName() + "." + Twine(i), AI);
1488 ElementAllocas.push_back(NA);
1489 WorkList.push_back(NA); // Add to worklist for recursive processing
1493 // Now that we have created the new alloca instructions, rewrite all the
1494 // uses of the old alloca.
1495 RewriteForScalarRepl(AI, AI, 0, ElementAllocas);
1497 // Now erase any instructions that were made dead while rewriting the alloca.
1498 DeleteDeadInstructions();
1499 AI->eraseFromParent();
1504 /// DeleteDeadInstructions - Erase instructions on the DeadInstrs list,
1505 /// recursively including all their operands that become trivially dead.
1506 void SROA::DeleteDeadInstructions() {
1507 while (!DeadInsts.empty()) {
1508 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val());
1510 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
1511 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
1512 // Zero out the operand and see if it becomes trivially dead.
1513 // (But, don't add allocas to the dead instruction list -- they are
1514 // already on the worklist and will be deleted separately.)
1516 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U))
1517 DeadInsts.push_back(U);
1520 I->eraseFromParent();
1524 /// isSafeForScalarRepl - Check if instruction I is a safe use with regard to
1525 /// performing scalar replacement of alloca AI. The results are flagged in
1526 /// the Info parameter. Offset indicates the position within AI that is
1527 /// referenced by this instruction.
1528 void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset,
1530 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
1531 Instruction *User = cast<Instruction>(*UI);
1533 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1534 isSafeForScalarRepl(BC, Offset, Info);
1535 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1536 uint64_t GEPOffset = Offset;
1537 isSafeGEP(GEPI, GEPOffset, Info);
1539 isSafeForScalarRepl(GEPI, GEPOffset, Info);
1540 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
1541 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
1543 return MarkUnsafe(Info, User);
1544 if (Length->isNegative())
1545 return MarkUnsafe(Info, User);
1547 isSafeMemAccess(Offset, Length->getZExtValue(), 0,
1548 UI.getOperandNo() == 0, Info, MI,
1549 true /*AllowWholeAccess*/);
1550 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1551 if (!LI->isSimple())
1552 return MarkUnsafe(Info, User);
1553 Type *LIType = LI->getType();
1554 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType),
1555 LIType, false, Info, LI, true /*AllowWholeAccess*/);
1556 Info.hasALoadOrStore = true;
1558 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1559 // Store is ok if storing INTO the pointer, not storing the pointer
1560 if (!SI->isSimple() || SI->getOperand(0) == I)
1561 return MarkUnsafe(Info, User);
1563 Type *SIType = SI->getOperand(0)->getType();
1564 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType),
1565 SIType, true, Info, SI, true /*AllowWholeAccess*/);
1566 Info.hasALoadOrStore = true;
1567 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
1568 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1569 II->getIntrinsicID() != Intrinsic::lifetime_end)
1570 return MarkUnsafe(Info, User);
1571 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) {
1572 isSafePHISelectUseForScalarRepl(User, Offset, Info);
1574 return MarkUnsafe(Info, User);
1576 if (Info.isUnsafe) return;
1581 /// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer
1582 /// derived from the alloca, we can often still split the alloca into elements.
1583 /// This is useful if we have a large alloca where one element is phi'd
1584 /// together somewhere: we can SRoA and promote all the other elements even if
1585 /// we end up not being able to promote this one.
1587 /// All we require is that the uses of the PHI do not index into other parts of
1588 /// the alloca. The most important use case for this is single load and stores
1589 /// that are PHI'd together, which can happen due to code sinking.
1590 void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset,
1592 // If we've already checked this PHI, don't do it again.
1593 if (PHINode *PN = dyn_cast<PHINode>(I))
1594 if (!Info.CheckedPHIs.insert(PN))
1597 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
1598 Instruction *User = cast<Instruction>(*UI);
1600 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1601 isSafePHISelectUseForScalarRepl(BC, Offset, Info);
1602 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1603 // Only allow "bitcast" GEPs for simplicity. We could generalize this,
1604 // but would have to prove that we're staying inside of an element being
1606 if (!GEPI->hasAllZeroIndices())
1607 return MarkUnsafe(Info, User);
1608 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info);
1609 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1610 if (!LI->isSimple())
1611 return MarkUnsafe(Info, User);
1612 Type *LIType = LI->getType();
1613 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType),
1614 LIType, false, Info, LI, false /*AllowWholeAccess*/);
1615 Info.hasALoadOrStore = true;
1617 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1618 // Store is ok if storing INTO the pointer, not storing the pointer
1619 if (!SI->isSimple() || SI->getOperand(0) == I)
1620 return MarkUnsafe(Info, User);
1622 Type *SIType = SI->getOperand(0)->getType();
1623 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType),
1624 SIType, true, Info, SI, false /*AllowWholeAccess*/);
1625 Info.hasALoadOrStore = true;
1626 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) {
1627 isSafePHISelectUseForScalarRepl(User, Offset, Info);
1629 return MarkUnsafe(Info, User);
1631 if (Info.isUnsafe) return;
1635 /// isSafeGEP - Check if a GEP instruction can be handled for scalar
1636 /// replacement. It is safe when all the indices are constant, in-bounds
1637 /// references, and when the resulting offset corresponds to an element within
1638 /// the alloca type. The results are flagged in the Info parameter. Upon
1639 /// return, Offset is adjusted as specified by the GEP indices.
1640 void SROA::isSafeGEP(GetElementPtrInst *GEPI,
1641 uint64_t &Offset, AllocaInfo &Info) {
1642 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI);
1646 // Walk through the GEP type indices, checking the types that this indexes
1648 for (; GEPIt != E; ++GEPIt) {
1649 // Ignore struct elements, no extra checking needed for these.
1650 if ((*GEPIt)->isStructTy())
1653 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand());
1655 return MarkUnsafe(Info, GEPI);
1658 // Compute the offset due to this GEP and check if the alloca has a
1659 // component element at that offset.
1660 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
1661 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices);
1662 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0))
1663 MarkUnsafe(Info, GEPI);
1666 /// isHomogeneousAggregate - Check if type T is a struct or array containing
1667 /// elements of the same type (which is always true for arrays). If so,
1668 /// return true with NumElts and EltTy set to the number of elements and the
1669 /// element type, respectively.
1670 static bool isHomogeneousAggregate(Type *T, unsigned &NumElts,
1672 if (ArrayType *AT = dyn_cast<ArrayType>(T)) {
1673 NumElts = AT->getNumElements();
1674 EltTy = (NumElts == 0 ? 0 : AT->getElementType());
1677 if (StructType *ST = dyn_cast<StructType>(T)) {
1678 NumElts = ST->getNumContainedTypes();
1679 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0));
1680 for (unsigned n = 1; n < NumElts; ++n) {
1681 if (ST->getContainedType(n) != EltTy)
1689 /// isCompatibleAggregate - Check if T1 and T2 are either the same type or are
1690 /// "homogeneous" aggregates with the same element type and number of elements.
1691 static bool isCompatibleAggregate(Type *T1, Type *T2) {
1695 unsigned NumElts1, NumElts2;
1696 Type *EltTy1, *EltTy2;
1697 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) &&
1698 isHomogeneousAggregate(T2, NumElts2, EltTy2) &&
1699 NumElts1 == NumElts2 &&
1706 /// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI
1707 /// alloca or has an offset and size that corresponds to a component element
1708 /// within it. The offset checked here may have been formed from a GEP with a
1709 /// pointer bitcasted to a different type.
1711 /// If AllowWholeAccess is true, then this allows uses of the entire alloca as a
1712 /// unit. If false, it only allows accesses known to be in a single element.
1713 void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize,
1714 Type *MemOpType, bool isStore,
1715 AllocaInfo &Info, Instruction *TheAccess,
1716 bool AllowWholeAccess) {
1717 // Check if this is a load/store of the entire alloca.
1718 if (Offset == 0 && AllowWholeAccess &&
1719 MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) {
1720 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer
1721 // loads/stores (which are essentially the same as the MemIntrinsics with
1722 // regard to copying padding between elements). But, if an alloca is
1723 // flagged as both a source and destination of such operations, we'll need
1724 // to check later for padding between elements.
1725 if (!MemOpType || MemOpType->isIntegerTy()) {
1727 Info.isMemCpyDst = true;
1729 Info.isMemCpySrc = true;
1732 // This is also safe for references using a type that is compatible with
1733 // the type of the alloca, so that loads/stores can be rewritten using
1734 // insertvalue/extractvalue.
1735 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) {
1736 Info.hasSubelementAccess = true;
1740 // Check if the offset/size correspond to a component within the alloca type.
1741 Type *T = Info.AI->getAllocatedType();
1742 if (TypeHasComponent(T, Offset, MemSize)) {
1743 Info.hasSubelementAccess = true;
1747 return MarkUnsafe(Info, TheAccess);
1750 /// TypeHasComponent - Return true if T has a component type with the
1751 /// specified offset and size. If Size is zero, do not check the size.
1752 bool SROA::TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size) {
1755 if (StructType *ST = dyn_cast<StructType>(T)) {
1756 const StructLayout *Layout = TD->getStructLayout(ST);
1757 unsigned EltIdx = Layout->getElementContainingOffset(Offset);
1758 EltTy = ST->getContainedType(EltIdx);
1759 EltSize = TD->getTypeAllocSize(EltTy);
1760 Offset -= Layout->getElementOffset(EltIdx);
1761 } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) {
1762 EltTy = AT->getElementType();
1763 EltSize = TD->getTypeAllocSize(EltTy);
1764 if (Offset >= AT->getNumElements() * EltSize)
1770 if (Offset == 0 && (Size == 0 || EltSize == Size))
1772 // Check if the component spans multiple elements.
1773 if (Offset + Size > EltSize)
1775 return TypeHasComponent(EltTy, Offset, Size);
1778 /// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite
1779 /// the instruction I, which references it, to use the separate elements.
1780 /// Offset indicates the position within AI that is referenced by this
1782 void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
1783 SmallVector<AllocaInst*, 32> &NewElts) {
1784 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) {
1785 Use &TheUse = UI.getUse();
1786 Instruction *User = cast<Instruction>(*UI++);
1788 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1789 RewriteBitCast(BC, AI, Offset, NewElts);
1793 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1794 RewriteGEP(GEPI, AI, Offset, NewElts);
1798 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
1799 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
1800 uint64_t MemSize = Length->getZExtValue();
1802 MemSize == TD->getTypeAllocSize(AI->getAllocatedType()))
1803 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts);
1804 // Otherwise the intrinsic can only touch a single element and the
1805 // address operand will be updated, so nothing else needs to be done.
1809 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
1810 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
1811 II->getIntrinsicID() == Intrinsic::lifetime_end) {
1812 RewriteLifetimeIntrinsic(II, AI, Offset, NewElts);
1817 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1818 Type *LIType = LI->getType();
1820 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) {
1822 // %res = load { i32, i32 }* %alloc
1824 // %load.0 = load i32* %alloc.0
1825 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0
1826 // %load.1 = load i32* %alloc.1
1827 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1
1828 // (Also works for arrays instead of structs)
1829 Value *Insert = UndefValue::get(LIType);
1830 IRBuilder<> Builder(LI);
1831 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1832 Value *Load = Builder.CreateLoad(NewElts[i], "load");
1833 Insert = Builder.CreateInsertValue(Insert, Load, i, "insert");
1835 LI->replaceAllUsesWith(Insert);
1836 DeadInsts.push_back(LI);
1837 } else if (LIType->isIntegerTy() &&
1838 TD->getTypeAllocSize(LIType) ==
1839 TD->getTypeAllocSize(AI->getAllocatedType())) {
1840 // If this is a load of the entire alloca to an integer, rewrite it.
1841 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts);
1846 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1847 Value *Val = SI->getOperand(0);
1848 Type *SIType = Val->getType();
1849 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) {
1851 // store { i32, i32 } %val, { i32, i32 }* %alloc
1853 // %val.0 = extractvalue { i32, i32 } %val, 0
1854 // store i32 %val.0, i32* %alloc.0
1855 // %val.1 = extractvalue { i32, i32 } %val, 1
1856 // store i32 %val.1, i32* %alloc.1
1857 // (Also works for arrays instead of structs)
1858 IRBuilder<> Builder(SI);
1859 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1860 Value *Extract = Builder.CreateExtractValue(Val, i, Val->getName());
1861 Builder.CreateStore(Extract, NewElts[i]);
1863 DeadInsts.push_back(SI);
1864 } else if (SIType->isIntegerTy() &&
1865 TD->getTypeAllocSize(SIType) ==
1866 TD->getTypeAllocSize(AI->getAllocatedType())) {
1867 // If this is a store of the entire alloca from an integer, rewrite it.
1868 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts);
1873 if (isa<SelectInst>(User) || isa<PHINode>(User)) {
1874 // If we have a PHI user of the alloca itself (as opposed to a GEP or
1875 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to
1877 if (!isa<AllocaInst>(I)) continue;
1879 assert(Offset == 0 && NewElts[0] &&
1880 "Direct alloca use should have a zero offset");
1882 // If we have a use of the alloca, we know the derived uses will be
1883 // utilizing just the first element of the scalarized result. Insert a
1884 // bitcast of the first alloca before the user as required.
1885 AllocaInst *NewAI = NewElts[0];
1886 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI);
1887 NewAI->moveBefore(BCI);
1894 /// RewriteBitCast - Update a bitcast reference to the alloca being replaced
1895 /// and recursively continue updating all of its uses.
1896 void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
1897 SmallVector<AllocaInst*, 32> &NewElts) {
1898 RewriteForScalarRepl(BC, AI, Offset, NewElts);
1899 if (BC->getOperand(0) != AI)
1902 // The bitcast references the original alloca. Replace its uses with
1903 // references to the alloca containing offset zero (which is normally at
1904 // index zero, but might not be in cases involving structs with elements
1906 Type *T = AI->getAllocatedType();
1907 uint64_t EltOffset = 0;
1909 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy);
1910 Instruction *Val = NewElts[Idx];
1911 if (Val->getType() != BC->getDestTy()) {
1912 Val = new BitCastInst(Val, BC->getDestTy(), "", BC);
1915 BC->replaceAllUsesWith(Val);
1916 DeadInsts.push_back(BC);
1919 /// FindElementAndOffset - Return the index of the element containing Offset
1920 /// within the specified type, which must be either a struct or an array.
1921 /// Sets T to the type of the element and Offset to the offset within that
1922 /// element. IdxTy is set to the type of the index result to be used in a
1923 /// GEP instruction.
1924 uint64_t SROA::FindElementAndOffset(Type *&T, uint64_t &Offset,
1927 if (StructType *ST = dyn_cast<StructType>(T)) {
1928 const StructLayout *Layout = TD->getStructLayout(ST);
1929 Idx = Layout->getElementContainingOffset(Offset);
1930 T = ST->getContainedType(Idx);
1931 Offset -= Layout->getElementOffset(Idx);
1932 IdxTy = Type::getInt32Ty(T->getContext());
1935 ArrayType *AT = cast<ArrayType>(T);
1936 T = AT->getElementType();
1937 uint64_t EltSize = TD->getTypeAllocSize(T);
1938 Idx = Offset / EltSize;
1939 Offset -= Idx * EltSize;
1940 IdxTy = Type::getInt64Ty(T->getContext());
1944 /// RewriteGEP - Check if this GEP instruction moves the pointer across
1945 /// elements of the alloca that are being split apart, and if so, rewrite
1946 /// the GEP to be relative to the new element.
1947 void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
1948 SmallVector<AllocaInst*, 32> &NewElts) {
1949 uint64_t OldOffset = Offset;
1950 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
1951 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices);
1953 RewriteForScalarRepl(GEPI, AI, Offset, NewElts);
1955 Type *T = AI->getAllocatedType();
1957 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy);
1958 if (GEPI->getOperand(0) == AI)
1959 OldIdx = ~0ULL; // Force the GEP to be rewritten.
1961 T = AI->getAllocatedType();
1962 uint64_t EltOffset = Offset;
1963 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy);
1965 // If this GEP does not move the pointer across elements of the alloca
1966 // being split, then it does not needs to be rewritten.
1970 Type *i32Ty = Type::getInt32Ty(AI->getContext());
1971 SmallVector<Value*, 8> NewArgs;
1972 NewArgs.push_back(Constant::getNullValue(i32Ty));
1973 while (EltOffset != 0) {
1974 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy);
1975 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx));
1977 Instruction *Val = NewElts[Idx];
1978 if (NewArgs.size() > 1) {
1979 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs, "", GEPI);
1980 Val->takeName(GEPI);
1982 if (Val->getType() != GEPI->getType())
1983 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI);
1984 GEPI->replaceAllUsesWith(Val);
1985 DeadInsts.push_back(GEPI);
1988 /// RewriteLifetimeIntrinsic - II is a lifetime.start/lifetime.end. Rewrite it
1989 /// to mark the lifetime of the scalarized memory.
1990 void SROA::RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI,
1992 SmallVector<AllocaInst*, 32> &NewElts) {
1993 ConstantInt *OldSize = cast<ConstantInt>(II->getArgOperand(0));
1994 // Put matching lifetime markers on everything from Offset up to
1996 Type *AIType = AI->getAllocatedType();
1997 uint64_t NewOffset = Offset;
1999 uint64_t Idx = FindElementAndOffset(AIType, NewOffset, IdxTy);
2001 IRBuilder<> Builder(II);
2002 uint64_t Size = OldSize->getLimitedValue();
2005 // Splice the first element and index 'NewOffset' bytes in. SROA will
2006 // split the alloca again later.
2007 Value *V = Builder.CreateBitCast(NewElts[Idx], Builder.getInt8PtrTy());
2008 V = Builder.CreateGEP(V, Builder.getInt64(NewOffset));
2010 IdxTy = NewElts[Idx]->getAllocatedType();
2011 uint64_t EltSize = TD->getTypeAllocSize(IdxTy) - NewOffset;
2012 if (EltSize > Size) {
2018 if (II->getIntrinsicID() == Intrinsic::lifetime_start)
2019 Builder.CreateLifetimeStart(V, Builder.getInt64(EltSize));
2021 Builder.CreateLifetimeEnd(V, Builder.getInt64(EltSize));
2025 for (; Idx != NewElts.size() && Size; ++Idx) {
2026 IdxTy = NewElts[Idx]->getAllocatedType();
2027 uint64_t EltSize = TD->getTypeAllocSize(IdxTy);
2028 if (EltSize > Size) {
2034 if (II->getIntrinsicID() == Intrinsic::lifetime_start)
2035 Builder.CreateLifetimeStart(NewElts[Idx],
2036 Builder.getInt64(EltSize));
2038 Builder.CreateLifetimeEnd(NewElts[Idx],
2039 Builder.getInt64(EltSize));
2041 DeadInsts.push_back(II);
2044 /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI.
2045 /// Rewrite it to copy or set the elements of the scalarized memory.
2046 void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
2048 SmallVector<AllocaInst*, 32> &NewElts) {
2049 // If this is a memcpy/memmove, construct the other pointer as the
2050 // appropriate type. The "Other" pointer is the pointer that goes to memory
2051 // that doesn't have anything to do with the alloca that we are promoting. For
2052 // memset, this Value* stays null.
2053 Value *OtherPtr = 0;
2054 unsigned MemAlignment = MI->getAlignment();
2055 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy
2056 if (Inst == MTI->getRawDest())
2057 OtherPtr = MTI->getRawSource();
2059 assert(Inst == MTI->getRawSource());
2060 OtherPtr = MTI->getRawDest();
2064 // If there is an other pointer, we want to convert it to the same pointer
2065 // type as AI has, so we can GEP through it safely.
2067 unsigned AddrSpace =
2068 cast<PointerType>(OtherPtr->getType())->getAddressSpace();
2070 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an
2071 // optimization, but it's also required to detect the corner case where
2072 // both pointer operands are referencing the same memory, and where
2073 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This
2074 // function is only called for mem intrinsics that access the whole
2075 // aggregate, so non-zero GEPs are not an issue here.)
2076 OtherPtr = OtherPtr->stripPointerCasts();
2078 // Copying the alloca to itself is a no-op: just delete it.
2079 if (OtherPtr == AI || OtherPtr == NewElts[0]) {
2080 // This code will run twice for a no-op memcpy -- once for each operand.
2081 // Put only one reference to MI on the DeadInsts list.
2082 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(),
2083 E = DeadInsts.end(); I != E; ++I)
2084 if (*I == MI) return;
2085 DeadInsts.push_back(MI);
2089 // If the pointer is not the right type, insert a bitcast to the right
2092 PointerType::get(AI->getType()->getElementType(), AddrSpace);
2094 if (OtherPtr->getType() != NewTy)
2095 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI);
2098 // Process each element of the aggregate.
2099 bool SROADest = MI->getRawDest() == Inst;
2101 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext()));
2103 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2104 // If this is a memcpy/memmove, emit a GEP of the other element address.
2105 Value *OtherElt = 0;
2106 unsigned OtherEltAlign = MemAlignment;
2109 Value *Idx[2] = { Zero,
2110 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) };
2111 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx,
2112 OtherPtr->getName()+"."+Twine(i),
2115 PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType());
2116 Type *OtherTy = OtherPtrTy->getElementType();
2117 if (StructType *ST = dyn_cast<StructType>(OtherTy)) {
2118 EltOffset = TD->getStructLayout(ST)->getElementOffset(i);
2120 Type *EltTy = cast<SequentialType>(OtherTy)->getElementType();
2121 EltOffset = TD->getTypeAllocSize(EltTy)*i;
2124 // The alignment of the other pointer is the guaranteed alignment of the
2125 // element, which is affected by both the known alignment of the whole
2126 // mem intrinsic and the alignment of the element. If the alignment of
2127 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the
2128 // known alignment is just 4 bytes.
2129 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset);
2132 Value *EltPtr = NewElts[i];
2133 Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType();
2135 // If we got down to a scalar, insert a load or store as appropriate.
2136 if (EltTy->isSingleValueType()) {
2137 if (isa<MemTransferInst>(MI)) {
2139 // From Other to Alloca.
2140 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI);
2141 new StoreInst(Elt, EltPtr, MI);
2143 // From Alloca to Other.
2144 Value *Elt = new LoadInst(EltPtr, "tmp", MI);
2145 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI);
2149 assert(isa<MemSetInst>(MI));
2151 // If the stored element is zero (common case), just store a null
2154 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) {
2156 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0>
2158 // If EltTy is a vector type, get the element type.
2159 Type *ValTy = EltTy->getScalarType();
2161 // Construct an integer with the right value.
2162 unsigned EltSize = TD->getTypeSizeInBits(ValTy);
2163 APInt OneVal(EltSize, CI->getZExtValue());
2164 APInt TotalVal(OneVal);
2166 for (unsigned i = 0; 8*i < EltSize; ++i) {
2167 TotalVal = TotalVal.shl(8);
2171 // Convert the integer value to the appropriate type.
2172 StoreVal = ConstantInt::get(CI->getContext(), TotalVal);
2173 if (ValTy->isPointerTy())
2174 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy);
2175 else if (ValTy->isFloatingPointTy())
2176 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy);
2177 assert(StoreVal->getType() == ValTy && "Type mismatch!");
2179 // If the requested value was a vector constant, create it.
2180 if (EltTy->isVectorTy()) {
2181 unsigned NumElts = cast<VectorType>(EltTy)->getNumElements();
2182 StoreVal = ConstantVector::getSplat(NumElts, StoreVal);
2185 new StoreInst(StoreVal, EltPtr, MI);
2188 // Otherwise, if we're storing a byte variable, use a memset call for
2192 unsigned EltSize = TD->getTypeAllocSize(EltTy);
2196 IRBuilder<> Builder(MI);
2198 // Finally, insert the meminst for this element.
2199 if (isa<MemSetInst>(MI)) {
2200 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize,
2203 assert(isa<MemTransferInst>(MI));
2204 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr
2205 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr
2207 if (isa<MemCpyInst>(MI))
2208 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile());
2210 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile());
2213 DeadInsts.push_back(MI);
2216 /// RewriteStoreUserOfWholeAlloca - We found a store of an integer that
2217 /// overwrites the entire allocation. Extract out the pieces of the stored
2218 /// integer and store them individually.
2219 void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
2220 SmallVector<AllocaInst*, 32> &NewElts){
2221 // Extract each element out of the integer according to its structure offset
2222 // and store the element value to the individual alloca.
2223 Value *SrcVal = SI->getOperand(0);
2224 Type *AllocaEltTy = AI->getAllocatedType();
2225 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
2227 IRBuilder<> Builder(SI);
2229 // Handle tail padding by extending the operand
2230 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits)
2231 SrcVal = Builder.CreateZExt(SrcVal,
2232 IntegerType::get(SI->getContext(), AllocaSizeBits));
2234 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI
2237 // There are two forms here: AI could be an array or struct. Both cases
2238 // have different ways to compute the element offset.
2239 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
2240 const StructLayout *Layout = TD->getStructLayout(EltSTy);
2242 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2243 // Get the number of bits to shift SrcVal to get the value.
2244 Type *FieldTy = EltSTy->getElementType(i);
2245 uint64_t Shift = Layout->getElementOffsetInBits(i);
2247 if (TD->isBigEndian())
2248 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy);
2250 Value *EltVal = SrcVal;
2252 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
2253 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt");
2256 // Truncate down to an integer of the right size.
2257 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
2259 // Ignore zero sized fields like {}, they obviously contain no data.
2260 if (FieldSizeBits == 0) continue;
2262 if (FieldSizeBits != AllocaSizeBits)
2263 EltVal = Builder.CreateTrunc(EltVal,
2264 IntegerType::get(SI->getContext(), FieldSizeBits));
2265 Value *DestField = NewElts[i];
2266 if (EltVal->getType() == FieldTy) {
2267 // Storing to an integer field of this size, just do it.
2268 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) {
2269 // Bitcast to the right element type (for fp/vector values).
2270 EltVal = Builder.CreateBitCast(EltVal, FieldTy);
2272 // Otherwise, bitcast the dest pointer (for aggregates).
2273 DestField = Builder.CreateBitCast(DestField,
2274 PointerType::getUnqual(EltVal->getType()));
2276 new StoreInst(EltVal, DestField, SI);
2280 ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
2281 Type *ArrayEltTy = ATy->getElementType();
2282 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
2283 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy);
2287 if (TD->isBigEndian())
2288 Shift = AllocaSizeBits-ElementOffset;
2292 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2293 // Ignore zero sized fields like {}, they obviously contain no data.
2294 if (ElementSizeBits == 0) continue;
2296 Value *EltVal = SrcVal;
2298 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
2299 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt");
2302 // Truncate down to an integer of the right size.
2303 if (ElementSizeBits != AllocaSizeBits)
2304 EltVal = Builder.CreateTrunc(EltVal,
2305 IntegerType::get(SI->getContext(),
2307 Value *DestField = NewElts[i];
2308 if (EltVal->getType() == ArrayEltTy) {
2309 // Storing to an integer field of this size, just do it.
2310 } else if (ArrayEltTy->isFloatingPointTy() ||
2311 ArrayEltTy->isVectorTy()) {
2312 // Bitcast to the right element type (for fp/vector values).
2313 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy);
2315 // Otherwise, bitcast the dest pointer (for aggregates).
2316 DestField = Builder.CreateBitCast(DestField,
2317 PointerType::getUnqual(EltVal->getType()));
2319 new StoreInst(EltVal, DestField, SI);
2321 if (TD->isBigEndian())
2322 Shift -= ElementOffset;
2324 Shift += ElementOffset;
2328 DeadInsts.push_back(SI);
2331 /// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to
2332 /// an integer. Load the individual pieces to form the aggregate value.
2333 void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
2334 SmallVector<AllocaInst*, 32> &NewElts) {
2335 // Extract each element out of the NewElts according to its structure offset
2336 // and form the result value.
2337 Type *AllocaEltTy = AI->getAllocatedType();
2338 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
2340 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI
2343 // There are two forms here: AI could be an array or struct. Both cases
2344 // have different ways to compute the element offset.
2345 const StructLayout *Layout = 0;
2346 uint64_t ArrayEltBitOffset = 0;
2347 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
2348 Layout = TD->getStructLayout(EltSTy);
2350 Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
2351 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
2355 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits));
2357 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2358 // Load the value from the alloca. If the NewElt is an aggregate, cast
2359 // the pointer to an integer of the same size before doing the load.
2360 Value *SrcField = NewElts[i];
2362 cast<PointerType>(SrcField->getType())->getElementType();
2363 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
2365 // Ignore zero sized fields like {}, they obviously contain no data.
2366 if (FieldSizeBits == 0) continue;
2368 IntegerType *FieldIntTy = IntegerType::get(LI->getContext(),
2370 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() &&
2371 !FieldTy->isVectorTy())
2372 SrcField = new BitCastInst(SrcField,
2373 PointerType::getUnqual(FieldIntTy),
2375 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI);
2377 // If SrcField is a fp or vector of the right size but that isn't an
2378 // integer type, bitcast to an integer so we can shift it.
2379 if (SrcField->getType() != FieldIntTy)
2380 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI);
2382 // Zero extend the field to be the same size as the final alloca so that
2383 // we can shift and insert it.
2384 if (SrcField->getType() != ResultVal->getType())
2385 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI);
2387 // Determine the number of bits to shift SrcField.
2389 if (Layout) // Struct case.
2390 Shift = Layout->getElementOffsetInBits(i);
2392 Shift = i*ArrayEltBitOffset;
2394 if (TD->isBigEndian())
2395 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth();
2398 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift);
2399 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI);
2402 // Don't create an 'or x, 0' on the first iteration.
2403 if (!isa<Constant>(ResultVal) ||
2404 !cast<Constant>(ResultVal)->isNullValue())
2405 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI);
2407 ResultVal = SrcField;
2410 // Handle tail padding by truncating the result
2411 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits)
2412 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI);
2414 LI->replaceAllUsesWith(ResultVal);
2415 DeadInsts.push_back(LI);
2418 /// HasPadding - Return true if the specified type has any structure or
2419 /// alignment padding in between the elements that would be split apart
2420 /// by SROA; return false otherwise.
2421 static bool HasPadding(Type *Ty, const TargetData &TD) {
2422 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2423 Ty = ATy->getElementType();
2424 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty);
2427 // SROA currently handles only Arrays and Structs.
2428 StructType *STy = cast<StructType>(Ty);
2429 const StructLayout *SL = TD.getStructLayout(STy);
2430 unsigned PrevFieldBitOffset = 0;
2431 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2432 unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
2434 // Check to see if there is any padding between this element and the
2437 unsigned PrevFieldEnd =
2438 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1));
2439 if (PrevFieldEnd < FieldBitOffset)
2442 PrevFieldBitOffset = FieldBitOffset;
2444 // Check for tail padding.
2445 if (unsigned EltCount = STy->getNumElements()) {
2446 unsigned PrevFieldEnd = PrevFieldBitOffset +
2447 TD.getTypeSizeInBits(STy->getElementType(EltCount-1));
2448 if (PrevFieldEnd < SL->getSizeInBits())
2454 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
2455 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe,
2456 /// or 1 if safe after canonicalization has been performed.
2457 bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) {
2458 // Loop over the use list of the alloca. We can only transform it if all of
2459 // the users are safe to transform.
2460 AllocaInfo Info(AI);
2462 isSafeForScalarRepl(AI, 0, Info);
2463 if (Info.isUnsafe) {
2464 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n');
2468 // Okay, we know all the users are promotable. If the aggregate is a memcpy
2469 // source and destination, we have to be careful. In particular, the memcpy
2470 // could be moving around elements that live in structure padding of the LLVM
2471 // types, but may actually be used. In these cases, we refuse to promote the
2473 if (Info.isMemCpySrc && Info.isMemCpyDst &&
2474 HasPadding(AI->getAllocatedType(), *TD))
2477 // If the alloca never has an access to just *part* of it, but is accessed
2478 // via loads and stores, then we should use ConvertToScalarInfo to promote
2479 // the alloca instead of promoting each piece at a time and inserting fission
2481 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) {
2482 // If the struct/array just has one element, use basic SRoA.
2483 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
2484 if (ST->getNumElements() > 1) return false;
2486 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1)
2496 /// PointsToConstantGlobal - Return true if V (possibly indirectly) points to
2497 /// some part of a constant global variable. This intentionally only accepts
2498 /// constant expressions because we don't can't rewrite arbitrary instructions.
2499 static bool PointsToConstantGlobal(Value *V) {
2500 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
2501 return GV->isConstant();
2502 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
2503 if (CE->getOpcode() == Instruction::BitCast ||
2504 CE->getOpcode() == Instruction::GetElementPtr)
2505 return PointsToConstantGlobal(CE->getOperand(0));
2509 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
2510 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
2511 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
2512 /// track of whether it moves the pointer (with isOffset) but otherwise traverse
2513 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
2514 /// the alloca, and if the source pointer is a pointer to a constant global, we
2515 /// can optimize this.
2517 isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy,
2519 SmallVector<Instruction *, 4> &LifetimeMarkers) {
2520 // We track lifetime intrinsics as we encounter them. If we decide to go
2521 // ahead and replace the value with the global, this lets the caller quickly
2522 // eliminate the markers.
2524 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
2525 User *U = cast<Instruction>(*UI);
2527 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
2528 // Ignore non-volatile loads, they are always ok.
2529 if (!LI->isSimple()) return false;
2533 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
2534 // If uses of the bitcast are ok, we are ok.
2535 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset,
2540 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
2541 // If the GEP has all zero indices, it doesn't offset the pointer. If it
2542 // doesn't, it does.
2543 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy,
2544 isOffset || !GEP->hasAllZeroIndices(),
2550 if (CallSite CS = U) {
2551 // If this is the function being called then we treat it like a load and
2553 if (CS.isCallee(UI))
2556 // If this is a readonly/readnone call site, then we know it is just a
2557 // load (but one that potentially returns the value itself), so we can
2558 // ignore it if we know that the value isn't captured.
2559 unsigned ArgNo = CS.getArgumentNo(UI);
2560 if (CS.onlyReadsMemory() &&
2561 (CS.getInstruction()->use_empty() || CS.doesNotCapture(ArgNo)))
2564 // If this is being passed as a byval argument, the caller is making a
2565 // copy, so it is only a read of the alloca.
2566 if (CS.isByValArgument(ArgNo))
2570 // Lifetime intrinsics can be handled by the caller.
2571 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
2572 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
2573 II->getIntrinsicID() == Intrinsic::lifetime_end) {
2574 assert(II->use_empty() && "Lifetime markers have no result to use!");
2575 LifetimeMarkers.push_back(II);
2580 // If this is isn't our memcpy/memmove, reject it as something we can't
2582 MemTransferInst *MI = dyn_cast<MemTransferInst>(U);
2586 // If the transfer is using the alloca as a source of the transfer, then
2587 // ignore it since it is a load (unless the transfer is volatile).
2588 if (UI.getOperandNo() == 1) {
2589 if (MI->isVolatile()) return false;
2593 // If we already have seen a copy, reject the second one.
2594 if (TheCopy) return false;
2596 // If the pointer has been offset from the start of the alloca, we can't
2597 // safely handle this.
2598 if (isOffset) return false;
2600 // If the memintrinsic isn't using the alloca as the dest, reject it.
2601 if (UI.getOperandNo() != 0) return false;
2603 // If the source of the memcpy/move is not a constant global, reject it.
2604 if (!PointsToConstantGlobal(MI->getSource()))
2607 // Otherwise, the transform is safe. Remember the copy instruction.
2613 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
2614 /// modified by a copy from a constant global. If we can prove this, we can
2615 /// replace any uses of the alloca with uses of the global directly.
2617 SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI,
2618 SmallVector<Instruction*, 4> &ToDelete) {
2619 MemTransferInst *TheCopy = 0;
2620 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false, ToDelete))