1 //===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation implements the well known scalar replacement of
11 // aggregates transformation. This xform breaks up alloca instructions of
12 // aggregate type (structure or array) into individual alloca instructions for
13 // each member (if possible). Then, if possible, it transforms the individual
14 // alloca instructions into nice clean scalar SSA form.
16 // This combines a simple SRoA algorithm with the Mem2Reg algorithm because
17 // often interact, especially for C++ programs. As such, iterating between
18 // SRoA, then Mem2Reg until we run out of things to promote works well.
20 //===----------------------------------------------------------------------===//
22 #define DEBUG_TYPE "scalarrepl"
23 #include "llvm/Transforms/Scalar.h"
24 #include "llvm/Constants.h"
25 #include "llvm/DerivedTypes.h"
26 #include "llvm/Function.h"
27 #include "llvm/GlobalVariable.h"
28 #include "llvm/Instructions.h"
29 #include "llvm/IntrinsicInst.h"
30 #include "llvm/LLVMContext.h"
31 #include "llvm/Module.h"
32 #include "llvm/Pass.h"
33 #include "llvm/Analysis/DebugInfo.h"
34 #include "llvm/Analysis/DIBuilder.h"
35 #include "llvm/Analysis/Dominators.h"
36 #include "llvm/Analysis/Loads.h"
37 #include "llvm/Analysis/ValueTracking.h"
38 #include "llvm/Target/TargetData.h"
39 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
40 #include "llvm/Transforms/Utils/Local.h"
41 #include "llvm/Transforms/Utils/SSAUpdater.h"
42 #include "llvm/Support/CallSite.h"
43 #include "llvm/Support/Debug.h"
44 #include "llvm/Support/ErrorHandling.h"
45 #include "llvm/Support/GetElementPtrTypeIterator.h"
46 #include "llvm/Support/IRBuilder.h"
47 #include "llvm/Support/MathExtras.h"
48 #include "llvm/Support/raw_ostream.h"
49 #include "llvm/ADT/SetVector.h"
50 #include "llvm/ADT/SmallVector.h"
51 #include "llvm/ADT/Statistic.h"
54 STATISTIC(NumReplaced, "Number of allocas broken up");
55 STATISTIC(NumPromoted, "Number of allocas promoted");
56 STATISTIC(NumAdjusted, "Number of scalar allocas adjusted to allow promotion");
57 STATISTIC(NumConverted, "Number of aggregates converted to scalar");
58 STATISTIC(NumGlobals, "Number of allocas copied from constant global");
61 struct SROA : public FunctionPass {
62 SROA(int T, bool hasDT, char &ID)
63 : FunctionPass(ID), HasDomTree(hasDT) {
70 bool runOnFunction(Function &F);
72 bool performScalarRepl(Function &F);
73 bool performPromotion(Function &F);
79 /// DeadInsts - Keep track of instructions we have made dead, so that
80 /// we can remove them after we are done working.
81 SmallVector<Value*, 32> DeadInsts;
83 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures
84 /// information about the uses. All these fields are initialized to false
85 /// and set to true when something is learned.
87 /// The alloca to promote.
90 /// CheckedPHIs - This is a set of verified PHI nodes, to prevent infinite
91 /// looping and avoid redundant work.
92 SmallPtrSet<PHINode*, 8> CheckedPHIs;
94 /// isUnsafe - This is set to true if the alloca cannot be SROA'd.
97 /// isMemCpySrc - This is true if this aggregate is memcpy'd from.
100 /// isMemCpyDst - This is true if this aggregate is memcpy'd into.
101 bool isMemCpyDst : 1;
103 /// hasSubelementAccess - This is true if a subelement of the alloca is
104 /// ever accessed, or false if the alloca is only accessed with mem
105 /// intrinsics or load/store that only access the entire alloca at once.
106 bool hasSubelementAccess : 1;
108 /// hasALoadOrStore - This is true if there are any loads or stores to it.
109 /// The alloca may just be accessed with memcpy, for example, which would
111 bool hasALoadOrStore : 1;
113 explicit AllocaInfo(AllocaInst *ai)
114 : AI(ai), isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false),
115 hasSubelementAccess(false), hasALoadOrStore(false) {}
118 unsigned SRThreshold;
120 void MarkUnsafe(AllocaInfo &I, Instruction *User) {
122 DEBUG(dbgs() << " Transformation preventing inst: " << *User << '\n');
125 bool isSafeAllocaToScalarRepl(AllocaInst *AI);
127 void isSafeForScalarRepl(Instruction *I, uint64_t Offset, AllocaInfo &Info);
128 void isSafePHISelectUseForScalarRepl(Instruction *User, uint64_t Offset,
130 void isSafeGEP(GetElementPtrInst *GEPI, uint64_t &Offset, AllocaInfo &Info);
131 void isSafeMemAccess(uint64_t Offset, uint64_t MemSize,
132 Type *MemOpType, bool isStore, AllocaInfo &Info,
133 Instruction *TheAccess, bool AllowWholeAccess);
134 bool TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size);
135 uint64_t FindElementAndOffset(Type *&T, uint64_t &Offset,
138 void DoScalarReplacement(AllocaInst *AI,
139 std::vector<AllocaInst*> &WorkList);
140 void DeleteDeadInstructions();
142 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
143 SmallVector<AllocaInst*, 32> &NewElts);
144 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
145 SmallVector<AllocaInst*, 32> &NewElts);
146 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
147 SmallVector<AllocaInst*, 32> &NewElts);
148 void RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI,
150 SmallVector<AllocaInst*, 32> &NewElts);
151 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
153 SmallVector<AllocaInst*, 32> &NewElts);
154 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
155 SmallVector<AllocaInst*, 32> &NewElts);
156 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
157 SmallVector<AllocaInst*, 32> &NewElts);
159 static MemTransferInst *isOnlyCopiedFromConstantGlobal(
160 AllocaInst *AI, SmallVector<Instruction*, 4> &ToDelete);
163 // SROA_DT - SROA that uses DominatorTree.
164 struct SROA_DT : public SROA {
167 SROA_DT(int T = -1) : SROA(T, true, ID) {
168 initializeSROA_DTPass(*PassRegistry::getPassRegistry());
171 // getAnalysisUsage - This pass does not require any passes, but we know it
172 // will not alter the CFG, so say so.
173 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
174 AU.addRequired<DominatorTree>();
175 AU.setPreservesCFG();
179 // SROA_SSAUp - SROA that uses SSAUpdater.
180 struct SROA_SSAUp : public SROA {
183 SROA_SSAUp(int T = -1) : SROA(T, false, ID) {
184 initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry());
187 // getAnalysisUsage - This pass does not require any passes, but we know it
188 // will not alter the CFG, so say so.
189 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
190 AU.setPreservesCFG();
196 char SROA_DT::ID = 0;
197 char SROA_SSAUp::ID = 0;
199 INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl",
200 "Scalar Replacement of Aggregates (DT)", false, false)
201 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
202 INITIALIZE_PASS_END(SROA_DT, "scalarrepl",
203 "Scalar Replacement of Aggregates (DT)", false, false)
205 INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa",
206 "Scalar Replacement of Aggregates (SSAUp)", false, false)
207 INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa",
208 "Scalar Replacement of Aggregates (SSAUp)", false, false)
210 // Public interface to the ScalarReplAggregates pass
211 FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold,
214 return new SROA_DT(Threshold);
215 return new SROA_SSAUp(Threshold);
219 //===----------------------------------------------------------------------===//
220 // Convert To Scalar Optimization.
221 //===----------------------------------------------------------------------===//
224 /// ConvertToScalarInfo - This class implements the "Convert To Scalar"
225 /// optimization, which scans the uses of an alloca and determines if it can
226 /// rewrite it in terms of a single new alloca that can be mem2reg'd.
227 class ConvertToScalarInfo {
228 /// AllocaSize - The size of the alloca being considered in bytes.
230 const TargetData &TD;
232 /// IsNotTrivial - This is set to true if there is some access to the object
233 /// which means that mem2reg can't promote it.
236 /// ScalarKind - Tracks the kind of alloca being considered for promotion,
237 /// computed based on the uses of the alloca rather than the LLVM type system.
241 // Accesses via GEPs that are consistent with element access of a vector
242 // type. This will not be converted into a vector unless there is a later
243 // access using an actual vector type.
246 // Accesses via vector operations and GEPs that are consistent with the
247 // layout of a vector type.
250 // An integer bag-of-bits with bitwise operations for insertion and
251 // extraction. Any combination of types can be converted into this kind
256 /// VectorTy - This tracks the type that we should promote the vector to if
257 /// it is possible to turn it into a vector. This starts out null, and if it
258 /// isn't possible to turn into a vector type, it gets set to VoidTy.
259 VectorType *VectorTy;
261 /// HadNonMemTransferAccess - True if there is at least one access to the
262 /// alloca that is not a MemTransferInst. We don't want to turn structs into
263 /// large integers unless there is some potential for optimization.
264 bool HadNonMemTransferAccess;
267 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td)
268 : AllocaSize(Size), TD(td), IsNotTrivial(false), ScalarKind(Unknown),
269 VectorTy(0), HadNonMemTransferAccess(false) { }
271 AllocaInst *TryConvert(AllocaInst *AI);
274 bool CanConvertToScalar(Value *V, uint64_t Offset);
275 void MergeInTypeForLoadOrStore(Type *In, uint64_t Offset);
276 bool MergeInVectorType(VectorType *VInTy, uint64_t Offset);
277 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset);
279 Value *ConvertScalar_ExtractValue(Value *NV, Type *ToType,
280 uint64_t Offset, IRBuilder<> &Builder);
281 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal,
282 uint64_t Offset, IRBuilder<> &Builder);
284 } // end anonymous namespace.
287 /// TryConvert - Analyze the specified alloca, and if it is safe to do so,
288 /// rewrite it to be a new alloca which is mem2reg'able. This returns the new
289 /// alloca if possible or null if not.
290 AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) {
291 // If we can't convert this scalar, or if mem2reg can trivially do it, bail
293 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial)
296 // If an alloca has only memset / memcpy uses, it may still have an Unknown
297 // ScalarKind. Treat it as an Integer below.
298 if (ScalarKind == Unknown)
299 ScalarKind = Integer;
301 if (ScalarKind == Vector && VectorTy->getBitWidth() != AllocaSize * 8)
302 ScalarKind = Integer;
304 // If we were able to find a vector type that can handle this with
305 // insert/extract elements, and if there was at least one use that had
306 // a vector type, promote this to a vector. We don't want to promote
307 // random stuff that doesn't use vectors (e.g. <9 x double>) because then
308 // we just get a lot of insert/extracts. If at least one vector is
309 // involved, then we probably really do have a union of vector/array.
311 if (ScalarKind == Vector) {
312 assert(VectorTy && "Missing type for vector scalar.");
313 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = "
314 << *VectorTy << '\n');
315 NewTy = VectorTy; // Use the vector type.
317 unsigned BitWidth = AllocaSize * 8;
318 if ((ScalarKind == ImplicitVector || ScalarKind == Integer) &&
319 !HadNonMemTransferAccess && !TD.fitsInLegalInteger(BitWidth))
322 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n");
323 // Create and insert the integer alloca.
324 NewTy = IntegerType::get(AI->getContext(), BitWidth);
326 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin());
327 ConvertUsesToScalar(AI, NewAI, 0);
331 /// MergeInTypeForLoadOrStore - Add the 'In' type to the accumulated vector type
332 /// (VectorTy) so far at the offset specified by Offset (which is specified in
335 /// There are two cases we handle here:
336 /// 1) A union of vector types of the same size and potentially its elements.
337 /// Here we turn element accesses into insert/extract element operations.
338 /// This promotes a <4 x float> with a store of float to the third element
339 /// into a <4 x float> that uses insert element.
340 /// 2) A fully general blob of memory, which we turn into some (potentially
341 /// large) integer type with extract and insert operations where the loads
342 /// and stores would mutate the memory. We mark this by setting VectorTy
344 void ConvertToScalarInfo::MergeInTypeForLoadOrStore(Type *In,
346 // If we already decided to turn this into a blob of integer memory, there is
347 // nothing to be done.
348 if (ScalarKind == Integer)
351 // If this could be contributing to a vector, analyze it.
353 // If the In type is a vector that is the same size as the alloca, see if it
354 // matches the existing VecTy.
355 if (VectorType *VInTy = dyn_cast<VectorType>(In)) {
356 if (MergeInVectorType(VInTy, Offset))
358 } else if (In->isFloatTy() || In->isDoubleTy() ||
359 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 &&
360 isPowerOf2_32(In->getPrimitiveSizeInBits()))) {
361 // Full width accesses can be ignored, because they can always be turned
363 unsigned EltSize = In->getPrimitiveSizeInBits()/8;
364 if (EltSize == AllocaSize)
367 // If we're accessing something that could be an element of a vector, see
368 // if the implied vector agrees with what we already have and if Offset is
369 // compatible with it.
370 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 &&
371 (!VectorTy || EltSize == VectorTy->getElementType()
372 ->getPrimitiveSizeInBits()/8)) {
374 ScalarKind = ImplicitVector;
375 VectorTy = VectorType::get(In, AllocaSize/EltSize);
381 // Otherwise, we have a case that we can't handle with an optimized vector
382 // form. We can still turn this into a large integer.
383 ScalarKind = Integer;
386 /// MergeInVectorType - Handles the vector case of MergeInTypeForLoadOrStore,
387 /// returning true if the type was successfully merged and false otherwise.
388 bool ConvertToScalarInfo::MergeInVectorType(VectorType *VInTy,
390 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) {
391 // If we're storing/loading a vector of the right size, allow it as a
392 // vector. If this the first vector we see, remember the type so that
393 // we know the element size. If this is a subsequent access, ignore it
394 // even if it is a differing type but the same size. Worst case we can
395 // bitcast the resultant vectors.
405 /// CanConvertToScalar - V is a pointer. If we can convert the pointee and all
406 /// its accesses to a single vector type, return true and set VecTy to
407 /// the new type. If we could convert the alloca into a single promotable
408 /// integer, return true but set VecTy to VoidTy. Further, if the use is not a
409 /// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset
410 /// is the current offset from the base of the alloca being analyzed.
412 /// If we see at least one access to the value that is as a vector type, set the
414 bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) {
415 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
416 Instruction *User = cast<Instruction>(*UI);
418 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
419 // Don't break volatile loads.
422 // Don't touch MMX operations.
423 if (LI->getType()->isX86_MMXTy())
425 HadNonMemTransferAccess = true;
426 MergeInTypeForLoadOrStore(LI->getType(), Offset);
430 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
431 // Storing the pointer, not into the value?
432 if (SI->getOperand(0) == V || !SI->isSimple()) return false;
433 // Don't touch MMX operations.
434 if (SI->getOperand(0)->getType()->isX86_MMXTy())
436 HadNonMemTransferAccess = true;
437 MergeInTypeForLoadOrStore(SI->getOperand(0)->getType(), Offset);
441 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
442 if (!onlyUsedByLifetimeMarkers(BCI))
443 IsNotTrivial = true; // Can't be mem2reg'd.
444 if (!CanConvertToScalar(BCI, Offset))
449 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
450 // If this is a GEP with a variable indices, we can't handle it.
451 if (!GEP->hasAllConstantIndices())
454 // Compute the offset that this GEP adds to the pointer.
455 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
456 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(),
458 // See if all uses can be converted.
459 if (!CanConvertToScalar(GEP, Offset+GEPOffset))
461 IsNotTrivial = true; // Can't be mem2reg'd.
462 HadNonMemTransferAccess = true;
466 // If this is a constant sized memset of a constant value (e.g. 0) we can
468 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
469 // Store of constant value.
470 if (!isa<ConstantInt>(MSI->getValue()))
473 // Store of constant size.
474 ConstantInt *Len = dyn_cast<ConstantInt>(MSI->getLength());
478 // If the size differs from the alloca, we can only convert the alloca to
479 // an integer bag-of-bits.
480 // FIXME: This should handle all of the cases that are currently accepted
481 // as vector element insertions.
482 if (Len->getZExtValue() != AllocaSize || Offset != 0)
483 ScalarKind = Integer;
485 IsNotTrivial = true; // Can't be mem2reg'd.
486 HadNonMemTransferAccess = true;
490 // If this is a memcpy or memmove into or out of the whole allocation, we
491 // can handle it like a load or store of the scalar type.
492 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
493 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength());
494 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0)
497 IsNotTrivial = true; // Can't be mem2reg'd.
501 // If this is a lifetime intrinsic, we can handle it.
502 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
503 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
504 II->getIntrinsicID() == Intrinsic::lifetime_end) {
509 // Otherwise, we cannot handle this!
516 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
517 /// directly. This happens when we are converting an "integer union" to a
518 /// single integer scalar, or when we are converting a "vector union" to a
519 /// vector with insert/extractelement instructions.
521 /// Offset is an offset from the original alloca, in bits that need to be
522 /// shifted to the right. By the end of this, there should be no uses of Ptr.
523 void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI,
525 while (!Ptr->use_empty()) {
526 Instruction *User = cast<Instruction>(Ptr->use_back());
528 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
529 ConvertUsesToScalar(CI, NewAI, Offset);
530 CI->eraseFromParent();
534 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
535 // Compute the offset that this GEP adds to the pointer.
536 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
537 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(),
539 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8);
540 GEP->eraseFromParent();
544 IRBuilder<> Builder(User);
546 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
547 // The load is a bit extract from NewAI shifted right by Offset bits.
548 Value *LoadedVal = Builder.CreateLoad(NewAI);
550 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder);
551 LI->replaceAllUsesWith(NewLoadVal);
552 LI->eraseFromParent();
556 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
557 assert(SI->getOperand(0) != Ptr && "Consistency error!");
558 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
559 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset,
561 Builder.CreateStore(New, NewAI);
562 SI->eraseFromParent();
564 // If the load we just inserted is now dead, then the inserted store
565 // overwrote the entire thing.
566 if (Old->use_empty())
567 Old->eraseFromParent();
571 // If this is a constant sized memset of a constant value (e.g. 0) we can
572 // transform it into a store of the expanded constant value.
573 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
574 assert(MSI->getRawDest() == Ptr && "Consistency error!");
575 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
577 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue();
579 // Compute the value replicated the right number of times.
580 APInt APVal(NumBytes*8, Val);
582 // Splat the value if non-zero.
584 for (unsigned i = 1; i != NumBytes; ++i)
587 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
588 Value *New = ConvertScalar_InsertValue(
589 ConstantInt::get(User->getContext(), APVal),
590 Old, Offset, Builder);
591 Builder.CreateStore(New, NewAI);
593 // If the load we just inserted is now dead, then the memset overwrote
595 if (Old->use_empty())
596 Old->eraseFromParent();
598 MSI->eraseFromParent();
602 // If this is a memcpy or memmove into or out of the whole allocation, we
603 // can handle it like a load or store of the scalar type.
604 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
605 assert(Offset == 0 && "must be store to start of alloca");
607 // If the source and destination are both to the same alloca, then this is
608 // a noop copy-to-self, just delete it. Otherwise, emit a load and store
610 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &TD, 0));
612 if (GetUnderlyingObject(MTI->getSource(), &TD, 0) != OrigAI) {
613 // Dest must be OrigAI, change this to be a load from the original
614 // pointer (bitcasted), then a store to our new alloca.
615 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?");
616 Value *SrcPtr = MTI->getSource();
617 PointerType* SPTy = cast<PointerType>(SrcPtr->getType());
618 PointerType* AIPTy = cast<PointerType>(NewAI->getType());
619 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
620 AIPTy = PointerType::get(AIPTy->getElementType(),
621 SPTy->getAddressSpace());
623 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy);
625 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval");
626 SrcVal->setAlignment(MTI->getAlignment());
627 Builder.CreateStore(SrcVal, NewAI);
628 } else if (GetUnderlyingObject(MTI->getDest(), &TD, 0) != OrigAI) {
629 // Src must be OrigAI, change this to be a load from NewAI then a store
630 // through the original dest pointer (bitcasted).
631 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?");
632 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval");
634 PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType());
635 PointerType* AIPTy = cast<PointerType>(NewAI->getType());
636 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
637 AIPTy = PointerType::get(AIPTy->getElementType(),
638 DPTy->getAddressSpace());
640 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy);
642 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr);
643 NewStore->setAlignment(MTI->getAlignment());
645 // Noop transfer. Src == Dst
648 MTI->eraseFromParent();
652 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
653 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
654 II->getIntrinsicID() == Intrinsic::lifetime_end) {
655 // There's no need to preserve these, as the resulting alloca will be
656 // converted to a register anyways.
657 II->eraseFromParent();
662 llvm_unreachable("Unsupported operation!");
666 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer
667 /// or vector value FromVal, extracting the bits from the offset specified by
668 /// Offset. This returns the value, which is of type ToType.
670 /// This happens when we are converting an "integer union" to a single
671 /// integer scalar, or when we are converting a "vector union" to a vector with
672 /// insert/extractelement instructions.
674 /// Offset is an offset from the original alloca, in bits that need to be
675 /// shifted to the right.
676 Value *ConvertToScalarInfo::
677 ConvertScalar_ExtractValue(Value *FromVal, Type *ToType,
678 uint64_t Offset, IRBuilder<> &Builder) {
679 // If the load is of the whole new alloca, no conversion is needed.
680 Type *FromType = FromVal->getType();
681 if (FromType == ToType && Offset == 0)
684 // If the result alloca is a vector type, this is either an element
685 // access or a bitcast to another vector type of the same size.
686 if (VectorType *VTy = dyn_cast<VectorType>(FromType)) {
687 unsigned FromTypeSize = TD.getTypeAllocSize(FromType);
688 unsigned ToTypeSize = TD.getTypeAllocSize(ToType);
689 if (FromTypeSize == ToTypeSize)
690 return Builder.CreateBitCast(FromVal, ToType);
692 // Otherwise it must be an element access.
695 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType());
696 Elt = Offset/EltSize;
697 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
699 // Return the element extracted out of it.
700 Value *V = Builder.CreateExtractElement(FromVal, Builder.getInt32(Elt));
701 if (V->getType() != ToType)
702 V = Builder.CreateBitCast(V, ToType);
706 // If ToType is a first class aggregate, extract out each of the pieces and
707 // use insertvalue's to form the FCA.
708 if (StructType *ST = dyn_cast<StructType>(ToType)) {
709 const StructLayout &Layout = *TD.getStructLayout(ST);
710 Value *Res = UndefValue::get(ST);
711 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
712 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i),
713 Offset+Layout.getElementOffsetInBits(i),
715 Res = Builder.CreateInsertValue(Res, Elt, i);
720 if (ArrayType *AT = dyn_cast<ArrayType>(ToType)) {
721 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType());
722 Value *Res = UndefValue::get(AT);
723 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
724 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(),
725 Offset+i*EltSize, Builder);
726 Res = Builder.CreateInsertValue(Res, Elt, i);
731 // Otherwise, this must be a union that was converted to an integer value.
732 IntegerType *NTy = cast<IntegerType>(FromVal->getType());
734 // If this is a big-endian system and the load is narrower than the
735 // full alloca type, we need to do a shift to get the right bits.
737 if (TD.isBigEndian()) {
738 // On big-endian machines, the lowest bit is stored at the bit offset
739 // from the pointer given by getTypeStoreSizeInBits. This matters for
740 // integers with a bitwidth that is not a multiple of 8.
741 ShAmt = TD.getTypeStoreSizeInBits(NTy) -
742 TD.getTypeStoreSizeInBits(ToType) - Offset;
747 // Note: we support negative bitwidths (with shl) which are not defined.
748 // We do this to support (f.e.) loads off the end of a structure where
749 // only some bits are used.
750 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth())
751 FromVal = Builder.CreateLShr(FromVal,
752 ConstantInt::get(FromVal->getType(), ShAmt));
753 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
754 FromVal = Builder.CreateShl(FromVal,
755 ConstantInt::get(FromVal->getType(), -ShAmt));
757 // Finally, unconditionally truncate the integer to the right width.
758 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType);
759 if (LIBitWidth < NTy->getBitWidth())
761 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(),
763 else if (LIBitWidth > NTy->getBitWidth())
765 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(),
768 // If the result is an integer, this is a trunc or bitcast.
769 if (ToType->isIntegerTy()) {
771 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) {
772 // Just do a bitcast, we know the sizes match up.
773 FromVal = Builder.CreateBitCast(FromVal, ToType);
775 // Otherwise must be a pointer.
776 FromVal = Builder.CreateIntToPtr(FromVal, ToType);
778 assert(FromVal->getType() == ToType && "Didn't convert right?");
782 /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer
783 /// or vector value "Old" at the offset specified by Offset.
785 /// This happens when we are converting an "integer union" to a
786 /// single integer scalar, or when we are converting a "vector union" to a
787 /// vector with insert/extractelement instructions.
789 /// Offset is an offset from the original alloca, in bits that need to be
790 /// shifted to the right.
791 Value *ConvertToScalarInfo::
792 ConvertScalar_InsertValue(Value *SV, Value *Old,
793 uint64_t Offset, IRBuilder<> &Builder) {
794 // Convert the stored type to the actual type, shift it left to insert
795 // then 'or' into place.
796 Type *AllocaType = Old->getType();
797 LLVMContext &Context = Old->getContext();
799 if (VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
800 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy);
801 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType());
803 // Changing the whole vector with memset or with an access of a different
805 if (ValSize == VecSize)
806 return Builder.CreateBitCast(SV, AllocaType);
808 // Must be an element insertion.
809 Type *EltTy = VTy->getElementType();
810 if (SV->getType() != EltTy)
811 SV = Builder.CreateBitCast(SV, EltTy);
812 uint64_t EltSize = TD.getTypeAllocSizeInBits(EltTy);
813 unsigned Elt = Offset/EltSize;
814 return Builder.CreateInsertElement(Old, SV, Builder.getInt32(Elt));
817 // If SV is a first-class aggregate value, insert each value recursively.
818 if (StructType *ST = dyn_cast<StructType>(SV->getType())) {
819 const StructLayout &Layout = *TD.getStructLayout(ST);
820 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
821 Value *Elt = Builder.CreateExtractValue(SV, i);
822 Old = ConvertScalar_InsertValue(Elt, Old,
823 Offset+Layout.getElementOffsetInBits(i),
829 if (ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) {
830 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType());
831 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
832 Value *Elt = Builder.CreateExtractValue(SV, i);
833 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder);
838 // If SV is a float, convert it to the appropriate integer type.
839 // If it is a pointer, do the same.
840 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType());
841 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType);
842 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType());
843 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType);
844 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy())
845 SV = Builder.CreateBitCast(SV, IntegerType::get(SV->getContext(),SrcWidth));
846 else if (SV->getType()->isPointerTy())
847 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()));
849 // Zero extend or truncate the value if needed.
850 if (SV->getType() != AllocaType) {
851 if (SV->getType()->getPrimitiveSizeInBits() <
852 AllocaType->getPrimitiveSizeInBits())
853 SV = Builder.CreateZExt(SV, AllocaType);
855 // Truncation may be needed if storing more than the alloca can hold
856 // (undefined behavior).
857 SV = Builder.CreateTrunc(SV, AllocaType);
858 SrcWidth = DestWidth;
859 SrcStoreWidth = DestStoreWidth;
863 // If this is a big-endian system and the store is narrower than the
864 // full alloca type, we need to do a shift to get the right bits.
866 if (TD.isBigEndian()) {
867 // On big-endian machines, the lowest bit is stored at the bit offset
868 // from the pointer given by getTypeStoreSizeInBits. This matters for
869 // integers with a bitwidth that is not a multiple of 8.
870 ShAmt = DestStoreWidth - SrcStoreWidth - Offset;
875 // Note: we support negative bitwidths (with shr) which are not defined.
876 // We do this to support (f.e.) stores off the end of a structure where
877 // only some bits in the structure are set.
878 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
879 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
880 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), ShAmt));
882 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
883 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), -ShAmt));
884 Mask = Mask.lshr(-ShAmt);
887 // Mask out the bits we are about to insert from the old value, and or
889 if (SrcWidth != DestWidth) {
890 assert(DestWidth > SrcWidth);
891 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask");
892 SV = Builder.CreateOr(Old, SV, "ins");
898 //===----------------------------------------------------------------------===//
900 //===----------------------------------------------------------------------===//
903 bool SROA::runOnFunction(Function &F) {
904 TD = getAnalysisIfAvailable<TargetData>();
906 bool Changed = performPromotion(F);
908 // FIXME: ScalarRepl currently depends on TargetData more than it
909 // theoretically needs to. It should be refactored in order to support
910 // target-independent IR. Until this is done, just skip the actual
911 // scalar-replacement portion of this pass.
912 if (!TD) return Changed;
915 bool LocalChange = performScalarRepl(F);
916 if (!LocalChange) break; // No need to repromote if no scalarrepl
918 LocalChange = performPromotion(F);
919 if (!LocalChange) break; // No need to re-scalarrepl if no promotion
926 class AllocaPromoter : public LoadAndStorePromoter {
929 SmallVector<DbgDeclareInst *, 4> DDIs;
930 SmallVector<DbgValueInst *, 4> DVIs;
932 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
934 : LoadAndStorePromoter(Insts, S), AI(0), DIB(DB) {}
936 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) {
937 // Remember which alloca we're promoting (for isInstInList).
939 if (MDNode *DebugNode = MDNode::getIfExists(AI->getContext(), AI))
940 for (Value::use_iterator UI = DebugNode->use_begin(),
941 E = DebugNode->use_end(); UI != E; ++UI)
942 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
944 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI))
947 LoadAndStorePromoter::run(Insts);
948 AI->eraseFromParent();
949 for (SmallVector<DbgDeclareInst *, 4>::iterator I = DDIs.begin(),
950 E = DDIs.end(); I != E; ++I) {
951 DbgDeclareInst *DDI = *I;
952 DDI->eraseFromParent();
954 for (SmallVector<DbgValueInst *, 4>::iterator I = DVIs.begin(),
955 E = DVIs.end(); I != E; ++I) {
956 DbgValueInst *DVI = *I;
957 DVI->eraseFromParent();
961 virtual bool isInstInList(Instruction *I,
962 const SmallVectorImpl<Instruction*> &Insts) const {
963 if (LoadInst *LI = dyn_cast<LoadInst>(I))
964 return LI->getOperand(0) == AI;
965 return cast<StoreInst>(I)->getPointerOperand() == AI;
968 virtual void updateDebugInfo(Instruction *Inst) const {
969 for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(),
970 E = DDIs.end(); I != E; ++I) {
971 DbgDeclareInst *DDI = *I;
972 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
973 ConvertDebugDeclareToDebugValue(DDI, SI, *DIB);
974 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
975 ConvertDebugDeclareToDebugValue(DDI, LI, *DIB);
977 for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(),
978 E = DVIs.end(); I != E; ++I) {
979 DbgValueInst *DVI = *I;
980 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
981 Instruction *DbgVal = NULL;
982 // If an argument is zero extended then use argument directly. The ZExt
983 // may be zapped by an optimization pass in future.
984 Argument *ExtendedArg = NULL;
985 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
986 ExtendedArg = dyn_cast<Argument>(ZExt->getOperand(0));
987 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
988 ExtendedArg = dyn_cast<Argument>(SExt->getOperand(0));
990 DbgVal = DIB->insertDbgValueIntrinsic(ExtendedArg, 0,
991 DIVariable(DVI->getVariable()),
994 DbgVal = DIB->insertDbgValueIntrinsic(SI->getOperand(0), 0,
995 DIVariable(DVI->getVariable()),
997 DbgVal->setDebugLoc(DVI->getDebugLoc());
998 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
999 Instruction *DbgVal =
1000 DIB->insertDbgValueIntrinsic(LI->getOperand(0), 0,
1001 DIVariable(DVI->getVariable()), LI);
1002 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;
1351 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
1352 // which runs on all of the alloca instructions in the function, removing them
1353 // if they are only used by getelementptr instructions.
1355 bool SROA::performScalarRepl(Function &F) {
1356 std::vector<AllocaInst*> WorkList;
1358 // Scan the entry basic block, adding allocas to the worklist.
1359 BasicBlock &BB = F.getEntryBlock();
1360 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
1361 if (AllocaInst *A = dyn_cast<AllocaInst>(I))
1362 WorkList.push_back(A);
1364 // Process the worklist
1365 bool Changed = false;
1366 while (!WorkList.empty()) {
1367 AllocaInst *AI = WorkList.back();
1368 WorkList.pop_back();
1370 // Handle dead allocas trivially. These can be formed by SROA'ing arrays
1371 // with unused elements.
1372 if (AI->use_empty()) {
1373 AI->eraseFromParent();
1378 // If this alloca is impossible for us to promote, reject it early.
1379 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized())
1382 // Check to see if this allocation is only modified by a memcpy/memmove from
1383 // a constant global. If this is the case, we can change all users to use
1384 // the constant global instead. This is commonly produced by the CFE by
1385 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
1386 // is only subsequently read.
1387 SmallVector<Instruction *, 4> ToDelete;
1388 if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(AI, ToDelete)) {
1389 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n');
1390 DEBUG(dbgs() << " memcpy = " << *Copy << '\n');
1391 for (unsigned i = 0, e = ToDelete.size(); i != e; ++i)
1392 ToDelete[i]->eraseFromParent();
1393 Constant *TheSrc = cast<Constant>(Copy->getSource());
1394 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType()));
1395 Copy->eraseFromParent(); // Don't mutate the global.
1396 AI->eraseFromParent();
1402 // Check to see if we can perform the core SROA transformation. We cannot
1403 // transform the allocation instruction if it is an array allocation
1404 // (allocations OF arrays are ok though), and an allocation of a scalar
1405 // value cannot be decomposed at all.
1406 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType());
1408 // Do not promote [0 x %struct].
1409 if (AllocaSize == 0) continue;
1411 // Do not promote any struct whose size is too big.
1412 if (AllocaSize > SRThreshold) continue;
1414 // If the alloca looks like a good candidate for scalar replacement, and if
1415 // all its users can be transformed, then split up the aggregate into its
1416 // separate elements.
1417 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) {
1418 DoScalarReplacement(AI, WorkList);
1423 // If we can turn this aggregate value (potentially with casts) into a
1424 // simple scalar value that can be mem2reg'd into a register value.
1425 // IsNotTrivial tracks whether this is something that mem2reg could have
1426 // promoted itself. If so, we don't want to transform it needlessly. Note
1427 // that we can't just check based on the type: the alloca may be of an i32
1428 // but that has pointer arithmetic to set byte 3 of it or something.
1429 if (AllocaInst *NewAI =
1430 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) {
1431 NewAI->takeName(AI);
1432 AI->eraseFromParent();
1438 // Otherwise, couldn't process this alloca.
1444 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl
1445 /// predicate, do SROA now.
1446 void SROA::DoScalarReplacement(AllocaInst *AI,
1447 std::vector<AllocaInst*> &WorkList) {
1448 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n');
1449 SmallVector<AllocaInst*, 32> ElementAllocas;
1450 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
1451 ElementAllocas.reserve(ST->getNumContainedTypes());
1452 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
1453 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
1455 AI->getName() + "." + Twine(i), AI);
1456 ElementAllocas.push_back(NA);
1457 WorkList.push_back(NA); // Add to worklist for recursive processing
1460 ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
1461 ElementAllocas.reserve(AT->getNumElements());
1462 Type *ElTy = AT->getElementType();
1463 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1464 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
1465 AI->getName() + "." + Twine(i), AI);
1466 ElementAllocas.push_back(NA);
1467 WorkList.push_back(NA); // Add to worklist for recursive processing
1471 // Now that we have created the new alloca instructions, rewrite all the
1472 // uses of the old alloca.
1473 RewriteForScalarRepl(AI, AI, 0, ElementAllocas);
1475 // Now erase any instructions that were made dead while rewriting the alloca.
1476 DeleteDeadInstructions();
1477 AI->eraseFromParent();
1482 /// DeleteDeadInstructions - Erase instructions on the DeadInstrs list,
1483 /// recursively including all their operands that become trivially dead.
1484 void SROA::DeleteDeadInstructions() {
1485 while (!DeadInsts.empty()) {
1486 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val());
1488 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
1489 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
1490 // Zero out the operand and see if it becomes trivially dead.
1491 // (But, don't add allocas to the dead instruction list -- they are
1492 // already on the worklist and will be deleted separately.)
1494 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U))
1495 DeadInsts.push_back(U);
1498 I->eraseFromParent();
1502 /// isSafeForScalarRepl - Check if instruction I is a safe use with regard to
1503 /// performing scalar replacement of alloca AI. The results are flagged in
1504 /// the Info parameter. Offset indicates the position within AI that is
1505 /// referenced by this instruction.
1506 void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset,
1508 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
1509 Instruction *User = cast<Instruction>(*UI);
1511 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1512 isSafeForScalarRepl(BC, Offset, Info);
1513 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1514 uint64_t GEPOffset = Offset;
1515 isSafeGEP(GEPI, GEPOffset, Info);
1517 isSafeForScalarRepl(GEPI, GEPOffset, Info);
1518 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
1519 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
1521 return MarkUnsafe(Info, User);
1522 isSafeMemAccess(Offset, Length->getZExtValue(), 0,
1523 UI.getOperandNo() == 0, Info, MI,
1524 true /*AllowWholeAccess*/);
1525 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1526 if (!LI->isSimple())
1527 return MarkUnsafe(Info, User);
1528 Type *LIType = LI->getType();
1529 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType),
1530 LIType, false, Info, LI, true /*AllowWholeAccess*/);
1531 Info.hasALoadOrStore = true;
1533 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1534 // Store is ok if storing INTO the pointer, not storing the pointer
1535 if (!SI->isSimple() || SI->getOperand(0) == I)
1536 return MarkUnsafe(Info, User);
1538 Type *SIType = SI->getOperand(0)->getType();
1539 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType),
1540 SIType, true, Info, SI, true /*AllowWholeAccess*/);
1541 Info.hasALoadOrStore = true;
1542 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
1543 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1544 II->getIntrinsicID() != Intrinsic::lifetime_end)
1545 return MarkUnsafe(Info, User);
1546 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) {
1547 isSafePHISelectUseForScalarRepl(User, Offset, Info);
1549 return MarkUnsafe(Info, User);
1551 if (Info.isUnsafe) return;
1556 /// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer
1557 /// derived from the alloca, we can often still split the alloca into elements.
1558 /// This is useful if we have a large alloca where one element is phi'd
1559 /// together somewhere: we can SRoA and promote all the other elements even if
1560 /// we end up not being able to promote this one.
1562 /// All we require is that the uses of the PHI do not index into other parts of
1563 /// the alloca. The most important use case for this is single load and stores
1564 /// that are PHI'd together, which can happen due to code sinking.
1565 void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset,
1567 // If we've already checked this PHI, don't do it again.
1568 if (PHINode *PN = dyn_cast<PHINode>(I))
1569 if (!Info.CheckedPHIs.insert(PN))
1572 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
1573 Instruction *User = cast<Instruction>(*UI);
1575 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1576 isSafePHISelectUseForScalarRepl(BC, Offset, Info);
1577 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1578 // Only allow "bitcast" GEPs for simplicity. We could generalize this,
1579 // but would have to prove that we're staying inside of an element being
1581 if (!GEPI->hasAllZeroIndices())
1582 return MarkUnsafe(Info, User);
1583 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info);
1584 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1585 if (!LI->isSimple())
1586 return MarkUnsafe(Info, User);
1587 Type *LIType = LI->getType();
1588 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType),
1589 LIType, false, Info, LI, false /*AllowWholeAccess*/);
1590 Info.hasALoadOrStore = true;
1592 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1593 // Store is ok if storing INTO the pointer, not storing the pointer
1594 if (!SI->isSimple() || SI->getOperand(0) == I)
1595 return MarkUnsafe(Info, User);
1597 Type *SIType = SI->getOperand(0)->getType();
1598 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType),
1599 SIType, true, Info, SI, false /*AllowWholeAccess*/);
1600 Info.hasALoadOrStore = true;
1601 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) {
1602 isSafePHISelectUseForScalarRepl(User, Offset, Info);
1604 return MarkUnsafe(Info, User);
1606 if (Info.isUnsafe) return;
1610 /// isSafeGEP - Check if a GEP instruction can be handled for scalar
1611 /// replacement. It is safe when all the indices are constant, in-bounds
1612 /// references, and when the resulting offset corresponds to an element within
1613 /// the alloca type. The results are flagged in the Info parameter. Upon
1614 /// return, Offset is adjusted as specified by the GEP indices.
1615 void SROA::isSafeGEP(GetElementPtrInst *GEPI,
1616 uint64_t &Offset, AllocaInfo &Info) {
1617 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI);
1621 // Walk through the GEP type indices, checking the types that this indexes
1623 for (; GEPIt != E; ++GEPIt) {
1624 // Ignore struct elements, no extra checking needed for these.
1625 if ((*GEPIt)->isStructTy())
1628 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand());
1630 return MarkUnsafe(Info, GEPI);
1633 // Compute the offset due to this GEP and check if the alloca has a
1634 // component element at that offset.
1635 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
1636 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices);
1637 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0))
1638 MarkUnsafe(Info, GEPI);
1641 /// isHomogeneousAggregate - Check if type T is a struct or array containing
1642 /// elements of the same type (which is always true for arrays). If so,
1643 /// return true with NumElts and EltTy set to the number of elements and the
1644 /// element type, respectively.
1645 static bool isHomogeneousAggregate(Type *T, unsigned &NumElts,
1647 if (ArrayType *AT = dyn_cast<ArrayType>(T)) {
1648 NumElts = AT->getNumElements();
1649 EltTy = (NumElts == 0 ? 0 : AT->getElementType());
1652 if (StructType *ST = dyn_cast<StructType>(T)) {
1653 NumElts = ST->getNumContainedTypes();
1654 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0));
1655 for (unsigned n = 1; n < NumElts; ++n) {
1656 if (ST->getContainedType(n) != EltTy)
1664 /// isCompatibleAggregate - Check if T1 and T2 are either the same type or are
1665 /// "homogeneous" aggregates with the same element type and number of elements.
1666 static bool isCompatibleAggregate(Type *T1, Type *T2) {
1670 unsigned NumElts1, NumElts2;
1671 Type *EltTy1, *EltTy2;
1672 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) &&
1673 isHomogeneousAggregate(T2, NumElts2, EltTy2) &&
1674 NumElts1 == NumElts2 &&
1681 /// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI
1682 /// alloca or has an offset and size that corresponds to a component element
1683 /// within it. The offset checked here may have been formed from a GEP with a
1684 /// pointer bitcasted to a different type.
1686 /// If AllowWholeAccess is true, then this allows uses of the entire alloca as a
1687 /// unit. If false, it only allows accesses known to be in a single element.
1688 void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize,
1689 Type *MemOpType, bool isStore,
1690 AllocaInfo &Info, Instruction *TheAccess,
1691 bool AllowWholeAccess) {
1692 // Check if this is a load/store of the entire alloca.
1693 if (Offset == 0 && AllowWholeAccess &&
1694 MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) {
1695 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer
1696 // loads/stores (which are essentially the same as the MemIntrinsics with
1697 // regard to copying padding between elements). But, if an alloca is
1698 // flagged as both a source and destination of such operations, we'll need
1699 // to check later for padding between elements.
1700 if (!MemOpType || MemOpType->isIntegerTy()) {
1702 Info.isMemCpyDst = true;
1704 Info.isMemCpySrc = true;
1707 // This is also safe for references using a type that is compatible with
1708 // the type of the alloca, so that loads/stores can be rewritten using
1709 // insertvalue/extractvalue.
1710 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) {
1711 Info.hasSubelementAccess = true;
1715 // Check if the offset/size correspond to a component within the alloca type.
1716 Type *T = Info.AI->getAllocatedType();
1717 if (TypeHasComponent(T, Offset, MemSize)) {
1718 Info.hasSubelementAccess = true;
1722 return MarkUnsafe(Info, TheAccess);
1725 /// TypeHasComponent - Return true if T has a component type with the
1726 /// specified offset and size. If Size is zero, do not check the size.
1727 bool SROA::TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size) {
1730 if (StructType *ST = dyn_cast<StructType>(T)) {
1731 const StructLayout *Layout = TD->getStructLayout(ST);
1732 unsigned EltIdx = Layout->getElementContainingOffset(Offset);
1733 EltTy = ST->getContainedType(EltIdx);
1734 EltSize = TD->getTypeAllocSize(EltTy);
1735 Offset -= Layout->getElementOffset(EltIdx);
1736 } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) {
1737 EltTy = AT->getElementType();
1738 EltSize = TD->getTypeAllocSize(EltTy);
1739 if (Offset >= AT->getNumElements() * EltSize)
1745 if (Offset == 0 && (Size == 0 || EltSize == Size))
1747 // Check if the component spans multiple elements.
1748 if (Offset + Size > EltSize)
1750 return TypeHasComponent(EltTy, Offset, Size);
1753 /// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite
1754 /// the instruction I, which references it, to use the separate elements.
1755 /// Offset indicates the position within AI that is referenced by this
1757 void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
1758 SmallVector<AllocaInst*, 32> &NewElts) {
1759 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) {
1760 Use &TheUse = UI.getUse();
1761 Instruction *User = cast<Instruction>(*UI++);
1763 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1764 RewriteBitCast(BC, AI, Offset, NewElts);
1768 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1769 RewriteGEP(GEPI, AI, Offset, NewElts);
1773 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
1774 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
1775 uint64_t MemSize = Length->getZExtValue();
1777 MemSize == TD->getTypeAllocSize(AI->getAllocatedType()))
1778 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts);
1779 // Otherwise the intrinsic can only touch a single element and the
1780 // address operand will be updated, so nothing else needs to be done.
1784 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
1785 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
1786 II->getIntrinsicID() == Intrinsic::lifetime_end) {
1787 RewriteLifetimeIntrinsic(II, AI, Offset, NewElts);
1792 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1793 Type *LIType = LI->getType();
1795 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) {
1797 // %res = load { i32, i32 }* %alloc
1799 // %load.0 = load i32* %alloc.0
1800 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0
1801 // %load.1 = load i32* %alloc.1
1802 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1
1803 // (Also works for arrays instead of structs)
1804 Value *Insert = UndefValue::get(LIType);
1805 IRBuilder<> Builder(LI);
1806 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1807 Value *Load = Builder.CreateLoad(NewElts[i], "load");
1808 Insert = Builder.CreateInsertValue(Insert, Load, i, "insert");
1810 LI->replaceAllUsesWith(Insert);
1811 DeadInsts.push_back(LI);
1812 } else if (LIType->isIntegerTy() &&
1813 TD->getTypeAllocSize(LIType) ==
1814 TD->getTypeAllocSize(AI->getAllocatedType())) {
1815 // If this is a load of the entire alloca to an integer, rewrite it.
1816 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts);
1821 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1822 Value *Val = SI->getOperand(0);
1823 Type *SIType = Val->getType();
1824 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) {
1826 // store { i32, i32 } %val, { i32, i32 }* %alloc
1828 // %val.0 = extractvalue { i32, i32 } %val, 0
1829 // store i32 %val.0, i32* %alloc.0
1830 // %val.1 = extractvalue { i32, i32 } %val, 1
1831 // store i32 %val.1, i32* %alloc.1
1832 // (Also works for arrays instead of structs)
1833 IRBuilder<> Builder(SI);
1834 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1835 Value *Extract = Builder.CreateExtractValue(Val, i, Val->getName());
1836 Builder.CreateStore(Extract, NewElts[i]);
1838 DeadInsts.push_back(SI);
1839 } else if (SIType->isIntegerTy() &&
1840 TD->getTypeAllocSize(SIType) ==
1841 TD->getTypeAllocSize(AI->getAllocatedType())) {
1842 // If this is a store of the entire alloca from an integer, rewrite it.
1843 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts);
1848 if (isa<SelectInst>(User) || isa<PHINode>(User)) {
1849 // If we have a PHI user of the alloca itself (as opposed to a GEP or
1850 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to
1852 if (!isa<AllocaInst>(I)) continue;
1854 assert(Offset == 0 && NewElts[0] &&
1855 "Direct alloca use should have a zero offset");
1857 // If we have a use of the alloca, we know the derived uses will be
1858 // utilizing just the first element of the scalarized result. Insert a
1859 // bitcast of the first alloca before the user as required.
1860 AllocaInst *NewAI = NewElts[0];
1861 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI);
1862 NewAI->moveBefore(BCI);
1869 /// RewriteBitCast - Update a bitcast reference to the alloca being replaced
1870 /// and recursively continue updating all of its uses.
1871 void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
1872 SmallVector<AllocaInst*, 32> &NewElts) {
1873 RewriteForScalarRepl(BC, AI, Offset, NewElts);
1874 if (BC->getOperand(0) != AI)
1877 // The bitcast references the original alloca. Replace its uses with
1878 // references to the alloca containing offset zero (which is normally at
1879 // index zero, but might not be in cases involving structs with elements
1881 Type *T = AI->getAllocatedType();
1882 uint64_t EltOffset = 0;
1884 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy);
1885 Instruction *Val = NewElts[Idx];
1886 if (Val->getType() != BC->getDestTy()) {
1887 Val = new BitCastInst(Val, BC->getDestTy(), "", BC);
1890 BC->replaceAllUsesWith(Val);
1891 DeadInsts.push_back(BC);
1894 /// FindElementAndOffset - Return the index of the element containing Offset
1895 /// within the specified type, which must be either a struct or an array.
1896 /// Sets T to the type of the element and Offset to the offset within that
1897 /// element. IdxTy is set to the type of the index result to be used in a
1898 /// GEP instruction.
1899 uint64_t SROA::FindElementAndOffset(Type *&T, uint64_t &Offset,
1902 if (StructType *ST = dyn_cast<StructType>(T)) {
1903 const StructLayout *Layout = TD->getStructLayout(ST);
1904 Idx = Layout->getElementContainingOffset(Offset);
1905 T = ST->getContainedType(Idx);
1906 Offset -= Layout->getElementOffset(Idx);
1907 IdxTy = Type::getInt32Ty(T->getContext());
1910 ArrayType *AT = cast<ArrayType>(T);
1911 T = AT->getElementType();
1912 uint64_t EltSize = TD->getTypeAllocSize(T);
1913 Idx = Offset / EltSize;
1914 Offset -= Idx * EltSize;
1915 IdxTy = Type::getInt64Ty(T->getContext());
1919 /// RewriteGEP - Check if this GEP instruction moves the pointer across
1920 /// elements of the alloca that are being split apart, and if so, rewrite
1921 /// the GEP to be relative to the new element.
1922 void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
1923 SmallVector<AllocaInst*, 32> &NewElts) {
1924 uint64_t OldOffset = Offset;
1925 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
1926 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices);
1928 RewriteForScalarRepl(GEPI, AI, Offset, NewElts);
1930 Type *T = AI->getAllocatedType();
1932 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy);
1933 if (GEPI->getOperand(0) == AI)
1934 OldIdx = ~0ULL; // Force the GEP to be rewritten.
1936 T = AI->getAllocatedType();
1937 uint64_t EltOffset = Offset;
1938 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy);
1940 // If this GEP does not move the pointer across elements of the alloca
1941 // being split, then it does not needs to be rewritten.
1945 Type *i32Ty = Type::getInt32Ty(AI->getContext());
1946 SmallVector<Value*, 8> NewArgs;
1947 NewArgs.push_back(Constant::getNullValue(i32Ty));
1948 while (EltOffset != 0) {
1949 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy);
1950 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx));
1952 Instruction *Val = NewElts[Idx];
1953 if (NewArgs.size() > 1) {
1954 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs, "", GEPI);
1955 Val->takeName(GEPI);
1957 if (Val->getType() != GEPI->getType())
1958 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI);
1959 GEPI->replaceAllUsesWith(Val);
1960 DeadInsts.push_back(GEPI);
1963 /// RewriteLifetimeIntrinsic - II is a lifetime.start/lifetime.end. Rewrite it
1964 /// to mark the lifetime of the scalarized memory.
1965 void SROA::RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI,
1967 SmallVector<AllocaInst*, 32> &NewElts) {
1968 ConstantInt *OldSize = cast<ConstantInt>(II->getArgOperand(0));
1969 // Put matching lifetime markers on everything from Offset up to
1971 Type *AIType = AI->getAllocatedType();
1972 uint64_t NewOffset = Offset;
1974 uint64_t Idx = FindElementAndOffset(AIType, NewOffset, IdxTy);
1976 IRBuilder<> Builder(II);
1977 uint64_t Size = OldSize->getLimitedValue();
1980 // Splice the first element and index 'NewOffset' bytes in. SROA will
1981 // split the alloca again later.
1982 Value *V = Builder.CreateBitCast(NewElts[Idx], Builder.getInt8PtrTy());
1983 V = Builder.CreateGEP(V, Builder.getInt64(NewOffset));
1985 IdxTy = NewElts[Idx]->getAllocatedType();
1986 uint64_t EltSize = TD->getTypeAllocSize(IdxTy) - NewOffset;
1987 if (EltSize > Size) {
1993 if (II->getIntrinsicID() == Intrinsic::lifetime_start)
1994 Builder.CreateLifetimeStart(V, Builder.getInt64(EltSize));
1996 Builder.CreateLifetimeEnd(V, Builder.getInt64(EltSize));
2000 for (; Idx != NewElts.size() && Size; ++Idx) {
2001 IdxTy = NewElts[Idx]->getAllocatedType();
2002 uint64_t EltSize = TD->getTypeAllocSize(IdxTy);
2003 if (EltSize > Size) {
2009 if (II->getIntrinsicID() == Intrinsic::lifetime_start)
2010 Builder.CreateLifetimeStart(NewElts[Idx],
2011 Builder.getInt64(EltSize));
2013 Builder.CreateLifetimeEnd(NewElts[Idx],
2014 Builder.getInt64(EltSize));
2016 DeadInsts.push_back(II);
2019 /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI.
2020 /// Rewrite it to copy or set the elements of the scalarized memory.
2021 void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
2023 SmallVector<AllocaInst*, 32> &NewElts) {
2024 // If this is a memcpy/memmove, construct the other pointer as the
2025 // appropriate type. The "Other" pointer is the pointer that goes to memory
2026 // that doesn't have anything to do with the alloca that we are promoting. For
2027 // memset, this Value* stays null.
2028 Value *OtherPtr = 0;
2029 unsigned MemAlignment = MI->getAlignment();
2030 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy
2031 if (Inst == MTI->getRawDest())
2032 OtherPtr = MTI->getRawSource();
2034 assert(Inst == MTI->getRawSource());
2035 OtherPtr = MTI->getRawDest();
2039 // If there is an other pointer, we want to convert it to the same pointer
2040 // type as AI has, so we can GEP through it safely.
2042 unsigned AddrSpace =
2043 cast<PointerType>(OtherPtr->getType())->getAddressSpace();
2045 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an
2046 // optimization, but it's also required to detect the corner case where
2047 // both pointer operands are referencing the same memory, and where
2048 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This
2049 // function is only called for mem intrinsics that access the whole
2050 // aggregate, so non-zero GEPs are not an issue here.)
2051 OtherPtr = OtherPtr->stripPointerCasts();
2053 // Copying the alloca to itself is a no-op: just delete it.
2054 if (OtherPtr == AI || OtherPtr == NewElts[0]) {
2055 // This code will run twice for a no-op memcpy -- once for each operand.
2056 // Put only one reference to MI on the DeadInsts list.
2057 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(),
2058 E = DeadInsts.end(); I != E; ++I)
2059 if (*I == MI) return;
2060 DeadInsts.push_back(MI);
2064 // If the pointer is not the right type, insert a bitcast to the right
2067 PointerType::get(AI->getType()->getElementType(), AddrSpace);
2069 if (OtherPtr->getType() != NewTy)
2070 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI);
2073 // Process each element of the aggregate.
2074 bool SROADest = MI->getRawDest() == Inst;
2076 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext()));
2078 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2079 // If this is a memcpy/memmove, emit a GEP of the other element address.
2080 Value *OtherElt = 0;
2081 unsigned OtherEltAlign = MemAlignment;
2084 Value *Idx[2] = { Zero,
2085 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) };
2086 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx,
2087 OtherPtr->getName()+"."+Twine(i),
2090 PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType());
2091 Type *OtherTy = OtherPtrTy->getElementType();
2092 if (StructType *ST = dyn_cast<StructType>(OtherTy)) {
2093 EltOffset = TD->getStructLayout(ST)->getElementOffset(i);
2095 Type *EltTy = cast<SequentialType>(OtherTy)->getElementType();
2096 EltOffset = TD->getTypeAllocSize(EltTy)*i;
2099 // The alignment of the other pointer is the guaranteed alignment of the
2100 // element, which is affected by both the known alignment of the whole
2101 // mem intrinsic and the alignment of the element. If the alignment of
2102 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the
2103 // known alignment is just 4 bytes.
2104 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset);
2107 Value *EltPtr = NewElts[i];
2108 Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType();
2110 // If we got down to a scalar, insert a load or store as appropriate.
2111 if (EltTy->isSingleValueType()) {
2112 if (isa<MemTransferInst>(MI)) {
2114 // From Other to Alloca.
2115 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI);
2116 new StoreInst(Elt, EltPtr, MI);
2118 // From Alloca to Other.
2119 Value *Elt = new LoadInst(EltPtr, "tmp", MI);
2120 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI);
2124 assert(isa<MemSetInst>(MI));
2126 // If the stored element is zero (common case), just store a null
2129 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) {
2131 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0>
2133 // If EltTy is a vector type, get the element type.
2134 Type *ValTy = EltTy->getScalarType();
2136 // Construct an integer with the right value.
2137 unsigned EltSize = TD->getTypeSizeInBits(ValTy);
2138 APInt OneVal(EltSize, CI->getZExtValue());
2139 APInt TotalVal(OneVal);
2141 for (unsigned i = 0; 8*i < EltSize; ++i) {
2142 TotalVal = TotalVal.shl(8);
2146 // Convert the integer value to the appropriate type.
2147 StoreVal = ConstantInt::get(CI->getContext(), TotalVal);
2148 if (ValTy->isPointerTy())
2149 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy);
2150 else if (ValTy->isFloatingPointTy())
2151 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy);
2152 assert(StoreVal->getType() == ValTy && "Type mismatch!");
2154 // If the requested value was a vector constant, create it.
2155 if (EltTy->isVectorTy()) {
2156 unsigned NumElts = cast<VectorType>(EltTy)->getNumElements();
2157 SmallVector<Constant*, 16> Elts(NumElts, StoreVal);
2158 StoreVal = ConstantVector::get(Elts);
2161 new StoreInst(StoreVal, EltPtr, MI);
2164 // Otherwise, if we're storing a byte variable, use a memset call for
2168 unsigned EltSize = TD->getTypeAllocSize(EltTy);
2172 IRBuilder<> Builder(MI);
2174 // Finally, insert the meminst for this element.
2175 if (isa<MemSetInst>(MI)) {
2176 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize,
2179 assert(isa<MemTransferInst>(MI));
2180 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr
2181 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr
2183 if (isa<MemCpyInst>(MI))
2184 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile());
2186 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile());
2189 DeadInsts.push_back(MI);
2192 /// RewriteStoreUserOfWholeAlloca - We found a store of an integer that
2193 /// overwrites the entire allocation. Extract out the pieces of the stored
2194 /// integer and store them individually.
2195 void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
2196 SmallVector<AllocaInst*, 32> &NewElts){
2197 // Extract each element out of the integer according to its structure offset
2198 // and store the element value to the individual alloca.
2199 Value *SrcVal = SI->getOperand(0);
2200 Type *AllocaEltTy = AI->getAllocatedType();
2201 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
2203 IRBuilder<> Builder(SI);
2205 // Handle tail padding by extending the operand
2206 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits)
2207 SrcVal = Builder.CreateZExt(SrcVal,
2208 IntegerType::get(SI->getContext(), AllocaSizeBits));
2210 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI
2213 // There are two forms here: AI could be an array or struct. Both cases
2214 // have different ways to compute the element offset.
2215 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
2216 const StructLayout *Layout = TD->getStructLayout(EltSTy);
2218 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2219 // Get the number of bits to shift SrcVal to get the value.
2220 Type *FieldTy = EltSTy->getElementType(i);
2221 uint64_t Shift = Layout->getElementOffsetInBits(i);
2223 if (TD->isBigEndian())
2224 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy);
2226 Value *EltVal = SrcVal;
2228 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
2229 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt");
2232 // Truncate down to an integer of the right size.
2233 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
2235 // Ignore zero sized fields like {}, they obviously contain no data.
2236 if (FieldSizeBits == 0) continue;
2238 if (FieldSizeBits != AllocaSizeBits)
2239 EltVal = Builder.CreateTrunc(EltVal,
2240 IntegerType::get(SI->getContext(), FieldSizeBits));
2241 Value *DestField = NewElts[i];
2242 if (EltVal->getType() == FieldTy) {
2243 // Storing to an integer field of this size, just do it.
2244 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) {
2245 // Bitcast to the right element type (for fp/vector values).
2246 EltVal = Builder.CreateBitCast(EltVal, FieldTy);
2248 // Otherwise, bitcast the dest pointer (for aggregates).
2249 DestField = Builder.CreateBitCast(DestField,
2250 PointerType::getUnqual(EltVal->getType()));
2252 new StoreInst(EltVal, DestField, SI);
2256 ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
2257 Type *ArrayEltTy = ATy->getElementType();
2258 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
2259 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy);
2263 if (TD->isBigEndian())
2264 Shift = AllocaSizeBits-ElementOffset;
2268 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2269 // Ignore zero sized fields like {}, they obviously contain no data.
2270 if (ElementSizeBits == 0) continue;
2272 Value *EltVal = SrcVal;
2274 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
2275 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt");
2278 // Truncate down to an integer of the right size.
2279 if (ElementSizeBits != AllocaSizeBits)
2280 EltVal = Builder.CreateTrunc(EltVal,
2281 IntegerType::get(SI->getContext(),
2283 Value *DestField = NewElts[i];
2284 if (EltVal->getType() == ArrayEltTy) {
2285 // Storing to an integer field of this size, just do it.
2286 } else if (ArrayEltTy->isFloatingPointTy() ||
2287 ArrayEltTy->isVectorTy()) {
2288 // Bitcast to the right element type (for fp/vector values).
2289 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy);
2291 // Otherwise, bitcast the dest pointer (for aggregates).
2292 DestField = Builder.CreateBitCast(DestField,
2293 PointerType::getUnqual(EltVal->getType()));
2295 new StoreInst(EltVal, DestField, SI);
2297 if (TD->isBigEndian())
2298 Shift -= ElementOffset;
2300 Shift += ElementOffset;
2304 DeadInsts.push_back(SI);
2307 /// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to
2308 /// an integer. Load the individual pieces to form the aggregate value.
2309 void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
2310 SmallVector<AllocaInst*, 32> &NewElts) {
2311 // Extract each element out of the NewElts according to its structure offset
2312 // and form the result value.
2313 Type *AllocaEltTy = AI->getAllocatedType();
2314 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
2316 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI
2319 // There are two forms here: AI could be an array or struct. Both cases
2320 // have different ways to compute the element offset.
2321 const StructLayout *Layout = 0;
2322 uint64_t ArrayEltBitOffset = 0;
2323 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
2324 Layout = TD->getStructLayout(EltSTy);
2326 Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
2327 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
2331 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits));
2333 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2334 // Load the value from the alloca. If the NewElt is an aggregate, cast
2335 // the pointer to an integer of the same size before doing the load.
2336 Value *SrcField = NewElts[i];
2338 cast<PointerType>(SrcField->getType())->getElementType();
2339 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
2341 // Ignore zero sized fields like {}, they obviously contain no data.
2342 if (FieldSizeBits == 0) continue;
2344 IntegerType *FieldIntTy = IntegerType::get(LI->getContext(),
2346 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() &&
2347 !FieldTy->isVectorTy())
2348 SrcField = new BitCastInst(SrcField,
2349 PointerType::getUnqual(FieldIntTy),
2351 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI);
2353 // If SrcField is a fp or vector of the right size but that isn't an
2354 // integer type, bitcast to an integer so we can shift it.
2355 if (SrcField->getType() != FieldIntTy)
2356 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI);
2358 // Zero extend the field to be the same size as the final alloca so that
2359 // we can shift and insert it.
2360 if (SrcField->getType() != ResultVal->getType())
2361 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI);
2363 // Determine the number of bits to shift SrcField.
2365 if (Layout) // Struct case.
2366 Shift = Layout->getElementOffsetInBits(i);
2368 Shift = i*ArrayEltBitOffset;
2370 if (TD->isBigEndian())
2371 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth();
2374 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift);
2375 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI);
2378 // Don't create an 'or x, 0' on the first iteration.
2379 if (!isa<Constant>(ResultVal) ||
2380 !cast<Constant>(ResultVal)->isNullValue())
2381 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI);
2383 ResultVal = SrcField;
2386 // Handle tail padding by truncating the result
2387 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits)
2388 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI);
2390 LI->replaceAllUsesWith(ResultVal);
2391 DeadInsts.push_back(LI);
2394 /// HasPadding - Return true if the specified type has any structure or
2395 /// alignment padding in between the elements that would be split apart
2396 /// by SROA; return false otherwise.
2397 static bool HasPadding(Type *Ty, const TargetData &TD) {
2398 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2399 Ty = ATy->getElementType();
2400 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty);
2403 // SROA currently handles only Arrays and Structs.
2404 StructType *STy = cast<StructType>(Ty);
2405 const StructLayout *SL = TD.getStructLayout(STy);
2406 unsigned PrevFieldBitOffset = 0;
2407 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2408 unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
2410 // Check to see if there is any padding between this element and the
2413 unsigned PrevFieldEnd =
2414 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1));
2415 if (PrevFieldEnd < FieldBitOffset)
2418 PrevFieldBitOffset = FieldBitOffset;
2420 // Check for tail padding.
2421 if (unsigned EltCount = STy->getNumElements()) {
2422 unsigned PrevFieldEnd = PrevFieldBitOffset +
2423 TD.getTypeSizeInBits(STy->getElementType(EltCount-1));
2424 if (PrevFieldEnd < SL->getSizeInBits())
2430 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
2431 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe,
2432 /// or 1 if safe after canonicalization has been performed.
2433 bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) {
2434 // Loop over the use list of the alloca. We can only transform it if all of
2435 // the users are safe to transform.
2436 AllocaInfo Info(AI);
2438 isSafeForScalarRepl(AI, 0, Info);
2439 if (Info.isUnsafe) {
2440 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n');
2444 // Okay, we know all the users are promotable. If the aggregate is a memcpy
2445 // source and destination, we have to be careful. In particular, the memcpy
2446 // could be moving around elements that live in structure padding of the LLVM
2447 // types, but may actually be used. In these cases, we refuse to promote the
2449 if (Info.isMemCpySrc && Info.isMemCpyDst &&
2450 HasPadding(AI->getAllocatedType(), *TD))
2453 // If the alloca never has an access to just *part* of it, but is accessed
2454 // via loads and stores, then we should use ConvertToScalarInfo to promote
2455 // the alloca instead of promoting each piece at a time and inserting fission
2457 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) {
2458 // If the struct/array just has one element, use basic SRoA.
2459 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
2460 if (ST->getNumElements() > 1) return false;
2462 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1)
2472 /// PointsToConstantGlobal - Return true if V (possibly indirectly) points to
2473 /// some part of a constant global variable. This intentionally only accepts
2474 /// constant expressions because we don't can't rewrite arbitrary instructions.
2475 static bool PointsToConstantGlobal(Value *V) {
2476 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
2477 return GV->isConstant();
2478 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
2479 if (CE->getOpcode() == Instruction::BitCast ||
2480 CE->getOpcode() == Instruction::GetElementPtr)
2481 return PointsToConstantGlobal(CE->getOperand(0));
2485 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
2486 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
2487 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
2488 /// track of whether it moves the pointer (with isOffset) but otherwise traverse
2489 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
2490 /// the alloca, and if the source pointer is a pointer to a constant global, we
2491 /// can optimize this.
2493 isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy,
2495 SmallVector<Instruction *, 4> &LifetimeMarkers) {
2496 // We track lifetime intrinsics as we encounter them. If we decide to go
2497 // ahead and replace the value with the global, this lets the caller quickly
2498 // eliminate the markers.
2500 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
2501 User *U = cast<Instruction>(*UI);
2503 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
2504 // Ignore non-volatile loads, they are always ok.
2505 if (!LI->isSimple()) return false;
2509 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
2510 // If uses of the bitcast are ok, we are ok.
2511 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset,
2516 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
2517 // If the GEP has all zero indices, it doesn't offset the pointer. If it
2518 // doesn't, it does.
2519 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy,
2520 isOffset || !GEP->hasAllZeroIndices(),
2526 if (CallSite CS = U) {
2527 // If this is the function being called then we treat it like a load and
2529 if (CS.isCallee(UI))
2532 // If this is a readonly/readnone call site, then we know it is just a
2533 // load (but one that potentially returns the value itself), so we can
2534 // ignore it if we know that the value isn't captured.
2535 unsigned ArgNo = CS.getArgumentNo(UI);
2536 if (CS.onlyReadsMemory() &&
2537 (CS.getInstruction()->use_empty() ||
2538 CS.paramHasAttr(ArgNo+1, Attribute::NoCapture)))
2541 // If this is being passed as a byval argument, the caller is making a
2542 // copy, so it is only a read of the alloca.
2543 if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal))
2547 // Lifetime intrinsics can be handled by the caller.
2548 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
2549 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
2550 II->getIntrinsicID() == Intrinsic::lifetime_end) {
2551 assert(II->use_empty() && "Lifetime markers have no result to use!");
2552 LifetimeMarkers.push_back(II);
2557 // If this is isn't our memcpy/memmove, reject it as something we can't
2559 MemTransferInst *MI = dyn_cast<MemTransferInst>(U);
2563 // If the transfer is using the alloca as a source of the transfer, then
2564 // ignore it since it is a load (unless the transfer is volatile).
2565 if (UI.getOperandNo() == 1) {
2566 if (MI->isVolatile()) return false;
2570 // If we already have seen a copy, reject the second one.
2571 if (TheCopy) return false;
2573 // If the pointer has been offset from the start of the alloca, we can't
2574 // safely handle this.
2575 if (isOffset) return false;
2577 // If the memintrinsic isn't using the alloca as the dest, reject it.
2578 if (UI.getOperandNo() != 0) return false;
2580 // If the source of the memcpy/move is not a constant global, reject it.
2581 if (!PointsToConstantGlobal(MI->getSource()))
2584 // Otherwise, the transform is safe. Remember the copy instruction.
2590 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
2591 /// modified by a copy from a constant global. If we can prove this, we can
2592 /// replace any uses of the alloca with uses of the global directly.
2594 SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI,
2595 SmallVector<Instruction*, 4> &ToDelete) {
2596 MemTransferInst *TheCopy = 0;
2597 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false, ToDelete))