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 if (!GEP->getPointerOperandType()->isPointerTy())
458 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(),
460 // See if all uses can be converted.
461 if (!CanConvertToScalar(GEP, Offset+GEPOffset))
463 IsNotTrivial = true; // Can't be mem2reg'd.
464 HadNonMemTransferAccess = true;
468 // If this is a constant sized memset of a constant value (e.g. 0) we can
470 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
471 // Store of constant value.
472 if (!isa<ConstantInt>(MSI->getValue()))
475 // Store of constant size.
476 ConstantInt *Len = dyn_cast<ConstantInt>(MSI->getLength());
480 // If the size differs from the alloca, we can only convert the alloca to
481 // an integer bag-of-bits.
482 // FIXME: This should handle all of the cases that are currently accepted
483 // as vector element insertions.
484 if (Len->getZExtValue() != AllocaSize || Offset != 0)
485 ScalarKind = Integer;
487 IsNotTrivial = true; // Can't be mem2reg'd.
488 HadNonMemTransferAccess = true;
492 // If this is a memcpy or memmove into or out of the whole allocation, we
493 // can handle it like a load or store of the scalar type.
494 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
495 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength());
496 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0)
499 IsNotTrivial = true; // Can't be mem2reg'd.
503 // If this is a lifetime intrinsic, we can handle it.
504 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
505 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
506 II->getIntrinsicID() == Intrinsic::lifetime_end) {
511 // Otherwise, we cannot handle this!
518 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
519 /// directly. This happens when we are converting an "integer union" to a
520 /// single integer scalar, or when we are converting a "vector union" to a
521 /// vector with insert/extractelement instructions.
523 /// Offset is an offset from the original alloca, in bits that need to be
524 /// shifted to the right. By the end of this, there should be no uses of Ptr.
525 void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI,
527 while (!Ptr->use_empty()) {
528 Instruction *User = cast<Instruction>(Ptr->use_back());
530 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
531 ConvertUsesToScalar(CI, NewAI, Offset);
532 CI->eraseFromParent();
536 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
537 // Compute the offset that this GEP adds to the pointer.
538 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
539 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(),
541 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8);
542 GEP->eraseFromParent();
546 IRBuilder<> Builder(User);
548 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
549 // The load is a bit extract from NewAI shifted right by Offset bits.
550 Value *LoadedVal = Builder.CreateLoad(NewAI);
552 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder);
553 LI->replaceAllUsesWith(NewLoadVal);
554 LI->eraseFromParent();
558 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
559 assert(SI->getOperand(0) != Ptr && "Consistency error!");
560 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
561 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset,
563 Builder.CreateStore(New, NewAI);
564 SI->eraseFromParent();
566 // If the load we just inserted is now dead, then the inserted store
567 // overwrote the entire thing.
568 if (Old->use_empty())
569 Old->eraseFromParent();
573 // If this is a constant sized memset of a constant value (e.g. 0) we can
574 // transform it into a store of the expanded constant value.
575 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
576 assert(MSI->getRawDest() == Ptr && "Consistency error!");
577 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
579 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue();
581 // Compute the value replicated the right number of times.
582 APInt APVal(NumBytes*8, Val);
584 // Splat the value if non-zero.
586 for (unsigned i = 1; i != NumBytes; ++i)
589 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
590 Value *New = ConvertScalar_InsertValue(
591 ConstantInt::get(User->getContext(), APVal),
592 Old, Offset, Builder);
593 Builder.CreateStore(New, NewAI);
595 // If the load we just inserted is now dead, then the memset overwrote
597 if (Old->use_empty())
598 Old->eraseFromParent();
600 MSI->eraseFromParent();
604 // If this is a memcpy or memmove into or out of the whole allocation, we
605 // can handle it like a load or store of the scalar type.
606 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
607 assert(Offset == 0 && "must be store to start of alloca");
609 // If the source and destination are both to the same alloca, then this is
610 // a noop copy-to-self, just delete it. Otherwise, emit a load and store
612 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &TD, 0));
614 if (GetUnderlyingObject(MTI->getSource(), &TD, 0) != OrigAI) {
615 // Dest must be OrigAI, change this to be a load from the original
616 // pointer (bitcasted), then a store to our new alloca.
617 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?");
618 Value *SrcPtr = MTI->getSource();
619 PointerType* SPTy = cast<PointerType>(SrcPtr->getType());
620 PointerType* AIPTy = cast<PointerType>(NewAI->getType());
621 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
622 AIPTy = PointerType::get(AIPTy->getElementType(),
623 SPTy->getAddressSpace());
625 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy);
627 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval");
628 SrcVal->setAlignment(MTI->getAlignment());
629 Builder.CreateStore(SrcVal, NewAI);
630 } else if (GetUnderlyingObject(MTI->getDest(), &TD, 0) != OrigAI) {
631 // Src must be OrigAI, change this to be a load from NewAI then a store
632 // through the original dest pointer (bitcasted).
633 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?");
634 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval");
636 PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType());
637 PointerType* AIPTy = cast<PointerType>(NewAI->getType());
638 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
639 AIPTy = PointerType::get(AIPTy->getElementType(),
640 DPTy->getAddressSpace());
642 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy);
644 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr);
645 NewStore->setAlignment(MTI->getAlignment());
647 // Noop transfer. Src == Dst
650 MTI->eraseFromParent();
654 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
655 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
656 II->getIntrinsicID() == Intrinsic::lifetime_end) {
657 // There's no need to preserve these, as the resulting alloca will be
658 // converted to a register anyways.
659 II->eraseFromParent();
664 llvm_unreachable("Unsupported operation!");
668 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer
669 /// or vector value FromVal, extracting the bits from the offset specified by
670 /// Offset. This returns the value, which is of type ToType.
672 /// This happens when we are converting an "integer union" to a single
673 /// integer scalar, or when we are converting a "vector union" to a vector with
674 /// insert/extractelement instructions.
676 /// Offset is an offset from the original alloca, in bits that need to be
677 /// shifted to the right.
678 Value *ConvertToScalarInfo::
679 ConvertScalar_ExtractValue(Value *FromVal, Type *ToType,
680 uint64_t Offset, IRBuilder<> &Builder) {
681 // If the load is of the whole new alloca, no conversion is needed.
682 Type *FromType = FromVal->getType();
683 if (FromType == ToType && Offset == 0)
686 // If the result alloca is a vector type, this is either an element
687 // access or a bitcast to another vector type of the same size.
688 if (VectorType *VTy = dyn_cast<VectorType>(FromType)) {
689 unsigned FromTypeSize = TD.getTypeAllocSize(FromType);
690 unsigned ToTypeSize = TD.getTypeAllocSize(ToType);
691 if (FromTypeSize == ToTypeSize)
692 return Builder.CreateBitCast(FromVal, ToType);
694 // Otherwise it must be an element access.
697 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType());
698 Elt = Offset/EltSize;
699 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
701 // Return the element extracted out of it.
702 Value *V = Builder.CreateExtractElement(FromVal, Builder.getInt32(Elt));
703 if (V->getType() != ToType)
704 V = Builder.CreateBitCast(V, ToType);
708 // If ToType is a first class aggregate, extract out each of the pieces and
709 // use insertvalue's to form the FCA.
710 if (StructType *ST = dyn_cast<StructType>(ToType)) {
711 const StructLayout &Layout = *TD.getStructLayout(ST);
712 Value *Res = UndefValue::get(ST);
713 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
714 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i),
715 Offset+Layout.getElementOffsetInBits(i),
717 Res = Builder.CreateInsertValue(Res, Elt, i);
722 if (ArrayType *AT = dyn_cast<ArrayType>(ToType)) {
723 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType());
724 Value *Res = UndefValue::get(AT);
725 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
726 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(),
727 Offset+i*EltSize, Builder);
728 Res = Builder.CreateInsertValue(Res, Elt, i);
733 // Otherwise, this must be a union that was converted to an integer value.
734 IntegerType *NTy = cast<IntegerType>(FromVal->getType());
736 // If this is a big-endian system and the load is narrower than the
737 // full alloca type, we need to do a shift to get the right bits.
739 if (TD.isBigEndian()) {
740 // On big-endian machines, the lowest bit is stored at the bit offset
741 // from the pointer given by getTypeStoreSizeInBits. This matters for
742 // integers with a bitwidth that is not a multiple of 8.
743 ShAmt = TD.getTypeStoreSizeInBits(NTy) -
744 TD.getTypeStoreSizeInBits(ToType) - Offset;
749 // Note: we support negative bitwidths (with shl) which are not defined.
750 // We do this to support (f.e.) loads off the end of a structure where
751 // only some bits are used.
752 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth())
753 FromVal = Builder.CreateLShr(FromVal,
754 ConstantInt::get(FromVal->getType(), ShAmt));
755 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
756 FromVal = Builder.CreateShl(FromVal,
757 ConstantInt::get(FromVal->getType(), -ShAmt));
759 // Finally, unconditionally truncate the integer to the right width.
760 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType);
761 if (LIBitWidth < NTy->getBitWidth())
763 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(),
765 else if (LIBitWidth > NTy->getBitWidth())
767 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(),
770 // If the result is an integer, this is a trunc or bitcast.
771 if (ToType->isIntegerTy()) {
773 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) {
774 // Just do a bitcast, we know the sizes match up.
775 FromVal = Builder.CreateBitCast(FromVal, ToType);
777 // Otherwise must be a pointer.
778 FromVal = Builder.CreateIntToPtr(FromVal, ToType);
780 assert(FromVal->getType() == ToType && "Didn't convert right?");
784 /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer
785 /// or vector value "Old" at the offset specified by Offset.
787 /// This happens when we are converting an "integer union" to a
788 /// single integer scalar, or when we are converting a "vector union" to a
789 /// vector with insert/extractelement instructions.
791 /// Offset is an offset from the original alloca, in bits that need to be
792 /// shifted to the right.
793 Value *ConvertToScalarInfo::
794 ConvertScalar_InsertValue(Value *SV, Value *Old,
795 uint64_t Offset, IRBuilder<> &Builder) {
796 // Convert the stored type to the actual type, shift it left to insert
797 // then 'or' into place.
798 Type *AllocaType = Old->getType();
799 LLVMContext &Context = Old->getContext();
801 if (VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
802 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy);
803 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType());
805 // Changing the whole vector with memset or with an access of a different
807 if (ValSize == VecSize)
808 return Builder.CreateBitCast(SV, AllocaType);
810 // Must be an element insertion.
811 Type *EltTy = VTy->getElementType();
812 if (SV->getType() != EltTy)
813 SV = Builder.CreateBitCast(SV, EltTy);
814 uint64_t EltSize = TD.getTypeAllocSizeInBits(EltTy);
815 unsigned Elt = Offset/EltSize;
816 return Builder.CreateInsertElement(Old, SV, Builder.getInt32(Elt));
819 // If SV is a first-class aggregate value, insert each value recursively.
820 if (StructType *ST = dyn_cast<StructType>(SV->getType())) {
821 const StructLayout &Layout = *TD.getStructLayout(ST);
822 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
823 Value *Elt = Builder.CreateExtractValue(SV, i);
824 Old = ConvertScalar_InsertValue(Elt, Old,
825 Offset+Layout.getElementOffsetInBits(i),
831 if (ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) {
832 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType());
833 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
834 Value *Elt = Builder.CreateExtractValue(SV, i);
835 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder);
840 // If SV is a float, convert it to the appropriate integer type.
841 // If it is a pointer, do the same.
842 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType());
843 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType);
844 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType());
845 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType);
846 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy())
847 SV = Builder.CreateBitCast(SV, IntegerType::get(SV->getContext(),SrcWidth));
848 else if (SV->getType()->isPointerTy())
849 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()));
851 // Zero extend or truncate the value if needed.
852 if (SV->getType() != AllocaType) {
853 if (SV->getType()->getPrimitiveSizeInBits() <
854 AllocaType->getPrimitiveSizeInBits())
855 SV = Builder.CreateZExt(SV, AllocaType);
857 // Truncation may be needed if storing more than the alloca can hold
858 // (undefined behavior).
859 SV = Builder.CreateTrunc(SV, AllocaType);
860 SrcWidth = DestWidth;
861 SrcStoreWidth = DestStoreWidth;
865 // If this is a big-endian system and the store is narrower than the
866 // full alloca type, we need to do a shift to get the right bits.
868 if (TD.isBigEndian()) {
869 // On big-endian machines, the lowest bit is stored at the bit offset
870 // from the pointer given by getTypeStoreSizeInBits. This matters for
871 // integers with a bitwidth that is not a multiple of 8.
872 ShAmt = DestStoreWidth - SrcStoreWidth - Offset;
877 // Note: we support negative bitwidths (with shr) which are not defined.
878 // We do this to support (f.e.) stores off the end of a structure where
879 // only some bits in the structure are set.
880 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
881 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
882 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), ShAmt));
884 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
885 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), -ShAmt));
886 Mask = Mask.lshr(-ShAmt);
889 // Mask out the bits we are about to insert from the old value, and or
891 if (SrcWidth != DestWidth) {
892 assert(DestWidth > SrcWidth);
893 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask");
894 SV = Builder.CreateOr(Old, SV, "ins");
900 //===----------------------------------------------------------------------===//
902 //===----------------------------------------------------------------------===//
905 bool SROA::runOnFunction(Function &F) {
906 TD = getAnalysisIfAvailable<TargetData>();
908 bool Changed = performPromotion(F);
910 // FIXME: ScalarRepl currently depends on TargetData more than it
911 // theoretically needs to. It should be refactored in order to support
912 // target-independent IR. Until this is done, just skip the actual
913 // scalar-replacement portion of this pass.
914 if (!TD) return Changed;
917 bool LocalChange = performScalarRepl(F);
918 if (!LocalChange) break; // No need to repromote if no scalarrepl
920 LocalChange = performPromotion(F);
921 if (!LocalChange) break; // No need to re-scalarrepl if no promotion
928 class AllocaPromoter : public LoadAndStorePromoter {
931 SmallVector<DbgDeclareInst *, 4> DDIs;
932 SmallVector<DbgValueInst *, 4> DVIs;
934 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
936 : LoadAndStorePromoter(Insts, S), AI(0), DIB(DB) {}
938 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) {
939 // Remember which alloca we're promoting (for isInstInList).
941 if (MDNode *DebugNode = MDNode::getIfExists(AI->getContext(), AI))
942 for (Value::use_iterator UI = DebugNode->use_begin(),
943 E = DebugNode->use_end(); UI != E; ++UI)
944 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
946 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI))
949 LoadAndStorePromoter::run(Insts);
950 AI->eraseFromParent();
951 for (SmallVector<DbgDeclareInst *, 4>::iterator I = DDIs.begin(),
952 E = DDIs.end(); I != E; ++I) {
953 DbgDeclareInst *DDI = *I;
954 DDI->eraseFromParent();
956 for (SmallVector<DbgValueInst *, 4>::iterator I = DVIs.begin(),
957 E = DVIs.end(); I != E; ++I) {
958 DbgValueInst *DVI = *I;
959 DVI->eraseFromParent();
963 virtual bool isInstInList(Instruction *I,
964 const SmallVectorImpl<Instruction*> &Insts) const {
965 if (LoadInst *LI = dyn_cast<LoadInst>(I))
966 return LI->getOperand(0) == AI;
967 return cast<StoreInst>(I)->getPointerOperand() == AI;
970 virtual void updateDebugInfo(Instruction *Inst) const {
971 for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(),
972 E = DDIs.end(); I != E; ++I) {
973 DbgDeclareInst *DDI = *I;
974 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
975 ConvertDebugDeclareToDebugValue(DDI, SI, *DIB);
976 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
977 ConvertDebugDeclareToDebugValue(DDI, LI, *DIB);
979 for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(),
980 E = DVIs.end(); I != E; ++I) {
981 DbgValueInst *DVI = *I;
982 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
983 Instruction *DbgVal = NULL;
984 // If an argument is zero extended then use argument directly. The ZExt
985 // may be zapped by an optimization pass in future.
986 Argument *ExtendedArg = NULL;
987 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
988 ExtendedArg = dyn_cast<Argument>(ZExt->getOperand(0));
989 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
990 ExtendedArg = dyn_cast<Argument>(SExt->getOperand(0));
992 DbgVal = DIB->insertDbgValueIntrinsic(ExtendedArg, 0,
993 DIVariable(DVI->getVariable()),
996 DbgVal = DIB->insertDbgValueIntrinsic(SI->getOperand(0), 0,
997 DIVariable(DVI->getVariable()),
999 DbgVal->setDebugLoc(DVI->getDebugLoc());
1000 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
1001 Instruction *DbgVal =
1002 DIB->insertDbgValueIntrinsic(LI->getOperand(0), 0,
1003 DIVariable(DVI->getVariable()), LI);
1004 DbgVal->setDebugLoc(DVI->getDebugLoc());
1009 } // end anon namespace
1011 /// isSafeSelectToSpeculate - Select instructions that use an alloca and are
1012 /// subsequently loaded can be rewritten to load both input pointers and then
1013 /// select between the result, allowing the load of the alloca to be promoted.
1015 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1016 /// %V = load i32* %P2
1018 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1019 /// %V2 = load i32* %Other
1020 /// %V = select i1 %cond, i32 %V1, i32 %V2
1022 /// We can do this to a select if its only uses are loads and if the operand to
1023 /// the select can be loaded unconditionally.
1024 static bool isSafeSelectToSpeculate(SelectInst *SI, const TargetData *TD) {
1025 bool TDerefable = SI->getTrueValue()->isDereferenceablePointer();
1026 bool FDerefable = SI->getFalseValue()->isDereferenceablePointer();
1028 for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end();
1030 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1031 if (LI == 0 || !LI->isSimple()) return false;
1033 // Both operands to the select need to be dereferencable, either absolutely
1034 // (e.g. allocas) or at this point because we can see other accesses to it.
1035 if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI,
1036 LI->getAlignment(), TD))
1038 if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI,
1039 LI->getAlignment(), TD))
1046 /// isSafePHIToSpeculate - PHI instructions that use an alloca and are
1047 /// subsequently loaded can be rewritten to load both input pointers in the pred
1048 /// blocks and then PHI the results, allowing the load of the alloca to be
1051 /// %P2 = phi [i32* %Alloca, i32* %Other]
1052 /// %V = load i32* %P2
1054 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1056 /// %V2 = load i32* %Other
1058 /// %V = phi [i32 %V1, i32 %V2]
1060 /// We can do this to a select if its only uses are loads and if the operand to
1061 /// the select can be loaded unconditionally.
1062 static bool isSafePHIToSpeculate(PHINode *PN, const TargetData *TD) {
1063 // For now, we can only do this promotion if the load is in the same block as
1064 // the PHI, and if there are no stores between the phi and load.
1065 // TODO: Allow recursive phi users.
1066 // TODO: Allow stores.
1067 BasicBlock *BB = PN->getParent();
1068 unsigned MaxAlign = 0;
1069 for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end();
1071 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1072 if (LI == 0 || !LI->isSimple()) return false;
1074 // For now we only allow loads in the same block as the PHI. This is a
1075 // common case that happens when instcombine merges two loads through a PHI.
1076 if (LI->getParent() != BB) return false;
1078 // Ensure that there are no instructions between the PHI and the load that
1080 for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI)
1081 if (BBI->mayWriteToMemory())
1084 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1087 // Okay, we know that we have one or more loads in the same block as the PHI.
1088 // We can transform this if it is safe to push the loads into the predecessor
1089 // blocks. The only thing to watch out for is that we can't put a possibly
1090 // trapping load in the predecessor if it is a critical edge.
1091 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1092 BasicBlock *Pred = PN->getIncomingBlock(i);
1093 Value *InVal = PN->getIncomingValue(i);
1095 // If the terminator of the predecessor has side-effects (an invoke),
1096 // there is no safe place to put a load in the predecessor.
1097 if (Pred->getTerminator()->mayHaveSideEffects())
1100 // If the value is produced by the terminator of the predecessor
1101 // (an invoke), there is no valid place to put a load in the predecessor.
1102 if (Pred->getTerminator() == InVal)
1105 // If the predecessor has a single successor, then the edge isn't critical.
1106 if (Pred->getTerminator()->getNumSuccessors() == 1)
1109 // If this pointer is always safe to load, or if we can prove that there is
1110 // already a load in the block, then we can move the load to the pred block.
1111 if (InVal->isDereferenceablePointer() ||
1112 isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, TD))
1122 /// tryToMakeAllocaBePromotable - This returns true if the alloca only has
1123 /// direct (non-volatile) loads and stores to it. If the alloca is close but
1124 /// not quite there, this will transform the code to allow promotion. As such,
1125 /// it is a non-pure predicate.
1126 static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) {
1127 SetVector<Instruction*, SmallVector<Instruction*, 4>,
1128 SmallPtrSet<Instruction*, 4> > InstsToRewrite;
1130 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
1133 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
1134 if (!LI->isSimple())
1139 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1140 if (SI->getOperand(0) == AI || !SI->isSimple())
1141 return false; // Don't allow a store OF the AI, only INTO the AI.
1145 if (SelectInst *SI = dyn_cast<SelectInst>(U)) {
1146 // If the condition being selected on is a constant, fold the select, yes
1147 // this does (rarely) happen early on.
1148 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) {
1149 Value *Result = SI->getOperand(1+CI->isZero());
1150 SI->replaceAllUsesWith(Result);
1151 SI->eraseFromParent();
1153 // This is very rare and we just scrambled the use list of AI, start
1155 return tryToMakeAllocaBePromotable(AI, TD);
1158 // If it is safe to turn "load (select c, AI, ptr)" into a select of two
1159 // loads, then we can transform this by rewriting the select.
1160 if (!isSafeSelectToSpeculate(SI, TD))
1163 InstsToRewrite.insert(SI);
1167 if (PHINode *PN = dyn_cast<PHINode>(U)) {
1168 if (PN->use_empty()) { // Dead PHIs can be stripped.
1169 InstsToRewrite.insert(PN);
1173 // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads
1174 // in the pred blocks, then we can transform this by rewriting the PHI.
1175 if (!isSafePHIToSpeculate(PN, TD))
1178 InstsToRewrite.insert(PN);
1182 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
1183 if (onlyUsedByLifetimeMarkers(BCI)) {
1184 InstsToRewrite.insert(BCI);
1192 // If there are no instructions to rewrite, then all uses are load/stores and
1194 if (InstsToRewrite.empty())
1197 // If we have instructions that need to be rewritten for this to be promotable
1198 // take care of it now.
1199 for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) {
1200 if (BitCastInst *BCI = dyn_cast<BitCastInst>(InstsToRewrite[i])) {
1201 // This could only be a bitcast used by nothing but lifetime intrinsics.
1202 for (BitCastInst::use_iterator I = BCI->use_begin(), E = BCI->use_end();
1204 Use &U = I.getUse();
1206 cast<Instruction>(U.getUser())->eraseFromParent();
1208 BCI->eraseFromParent();
1212 if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) {
1213 // Selects in InstsToRewrite only have load uses. Rewrite each as two
1214 // loads with a new select.
1215 while (!SI->use_empty()) {
1216 LoadInst *LI = cast<LoadInst>(SI->use_back());
1218 IRBuilder<> Builder(LI);
1219 LoadInst *TrueLoad =
1220 Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t");
1221 LoadInst *FalseLoad =
1222 Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".f");
1224 // Transfer alignment and TBAA info if present.
1225 TrueLoad->setAlignment(LI->getAlignment());
1226 FalseLoad->setAlignment(LI->getAlignment());
1227 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
1228 TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1229 FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1232 Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad);
1234 LI->replaceAllUsesWith(V);
1235 LI->eraseFromParent();
1238 // Now that all the loads are gone, the select is gone too.
1239 SI->eraseFromParent();
1243 // Otherwise, we have a PHI node which allows us to push the loads into the
1245 PHINode *PN = cast<PHINode>(InstsToRewrite[i]);
1246 if (PN->use_empty()) {
1247 PN->eraseFromParent();
1251 Type *LoadTy = cast<PointerType>(PN->getType())->getElementType();
1252 PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(),
1253 PN->getName()+".ld", PN);
1255 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1256 // matter which one we get and if any differ, it doesn't matter.
1257 LoadInst *SomeLoad = cast<LoadInst>(PN->use_back());
1258 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1259 unsigned Align = SomeLoad->getAlignment();
1261 // Rewrite all loads of the PN to use the new PHI.
1262 while (!PN->use_empty()) {
1263 LoadInst *LI = cast<LoadInst>(PN->use_back());
1264 LI->replaceAllUsesWith(NewPN);
1265 LI->eraseFromParent();
1268 // Inject loads into all of the pred blocks. Keep track of which blocks we
1269 // insert them into in case we have multiple edges from the same block.
1270 DenseMap<BasicBlock*, LoadInst*> InsertedLoads;
1272 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1273 BasicBlock *Pred = PN->getIncomingBlock(i);
1274 LoadInst *&Load = InsertedLoads[Pred];
1276 Load = new LoadInst(PN->getIncomingValue(i),
1277 PN->getName() + "." + Pred->getName(),
1278 Pred->getTerminator());
1279 Load->setAlignment(Align);
1280 if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1283 NewPN->addIncoming(Load, Pred);
1286 PN->eraseFromParent();
1293 bool SROA::performPromotion(Function &F) {
1294 std::vector<AllocaInst*> Allocas;
1295 DominatorTree *DT = 0;
1297 DT = &getAnalysis<DominatorTree>();
1299 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
1300 DIBuilder DIB(*F.getParent());
1301 bool Changed = false;
1302 SmallVector<Instruction*, 64> Insts;
1306 // Find allocas that are safe to promote, by looking at all instructions in
1308 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
1309 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
1310 if (tryToMakeAllocaBePromotable(AI, TD))
1311 Allocas.push_back(AI);
1313 if (Allocas.empty()) break;
1316 PromoteMemToReg(Allocas, *DT);
1319 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
1320 AllocaInst *AI = Allocas[i];
1322 // Build list of instructions to promote.
1323 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
1325 Insts.push_back(cast<Instruction>(*UI));
1326 AllocaPromoter(Insts, SSA, &DIB).run(AI, Insts);
1330 NumPromoted += Allocas.size();
1338 /// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for
1339 /// SROA. It must be a struct or array type with a small number of elements.
1340 static bool ShouldAttemptScalarRepl(AllocaInst *AI) {
1341 Type *T = AI->getAllocatedType();
1342 // Do not promote any struct into more than 32 separate vars.
1343 if (StructType *ST = dyn_cast<StructType>(T))
1344 return ST->getNumElements() <= 32;
1345 // Arrays are much less likely to be safe for SROA; only consider
1346 // them if they are very small.
1347 if (ArrayType *AT = dyn_cast<ArrayType>(T))
1348 return AT->getNumElements() <= 8;
1353 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
1354 // which runs on all of the alloca instructions in the function, removing them
1355 // if they are only used by getelementptr instructions.
1357 bool SROA::performScalarRepl(Function &F) {
1358 std::vector<AllocaInst*> WorkList;
1360 // Scan the entry basic block, adding allocas to the worklist.
1361 BasicBlock &BB = F.getEntryBlock();
1362 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
1363 if (AllocaInst *A = dyn_cast<AllocaInst>(I))
1364 WorkList.push_back(A);
1366 // Process the worklist
1367 bool Changed = false;
1368 while (!WorkList.empty()) {
1369 AllocaInst *AI = WorkList.back();
1370 WorkList.pop_back();
1372 // Handle dead allocas trivially. These can be formed by SROA'ing arrays
1373 // with unused elements.
1374 if (AI->use_empty()) {
1375 AI->eraseFromParent();
1380 // If this alloca is impossible for us to promote, reject it early.
1381 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized())
1384 // Check to see if this allocation is only modified by a memcpy/memmove from
1385 // a constant global. If this is the case, we can change all users to use
1386 // the constant global instead. This is commonly produced by the CFE by
1387 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
1388 // is only subsequently read.
1389 SmallVector<Instruction *, 4> ToDelete;
1390 if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(AI, ToDelete)) {
1391 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n');
1392 DEBUG(dbgs() << " memcpy = " << *Copy << '\n');
1393 for (unsigned i = 0, e = ToDelete.size(); i != e; ++i)
1394 ToDelete[i]->eraseFromParent();
1395 Constant *TheSrc = cast<Constant>(Copy->getSource());
1396 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType()));
1397 Copy->eraseFromParent(); // Don't mutate the global.
1398 AI->eraseFromParent();
1404 // Check to see if we can perform the core SROA transformation. We cannot
1405 // transform the allocation instruction if it is an array allocation
1406 // (allocations OF arrays are ok though), and an allocation of a scalar
1407 // value cannot be decomposed at all.
1408 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType());
1410 // Do not promote [0 x %struct].
1411 if (AllocaSize == 0) continue;
1413 // Do not promote any struct whose size is too big.
1414 if (AllocaSize > SRThreshold) continue;
1416 // If the alloca looks like a good candidate for scalar replacement, and if
1417 // all its users can be transformed, then split up the aggregate into its
1418 // separate elements.
1419 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) {
1420 DoScalarReplacement(AI, WorkList);
1425 // If we can turn this aggregate value (potentially with casts) into a
1426 // simple scalar value that can be mem2reg'd into a register value.
1427 // IsNotTrivial tracks whether this is something that mem2reg could have
1428 // promoted itself. If so, we don't want to transform it needlessly. Note
1429 // that we can't just check based on the type: the alloca may be of an i32
1430 // but that has pointer arithmetic to set byte 3 of it or something.
1431 if (AllocaInst *NewAI =
1432 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) {
1433 NewAI->takeName(AI);
1434 AI->eraseFromParent();
1440 // Otherwise, couldn't process this alloca.
1446 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl
1447 /// predicate, do SROA now.
1448 void SROA::DoScalarReplacement(AllocaInst *AI,
1449 std::vector<AllocaInst*> &WorkList) {
1450 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n');
1451 SmallVector<AllocaInst*, 32> ElementAllocas;
1452 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
1453 ElementAllocas.reserve(ST->getNumContainedTypes());
1454 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
1455 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
1457 AI->getName() + "." + Twine(i), AI);
1458 ElementAllocas.push_back(NA);
1459 WorkList.push_back(NA); // Add to worklist for recursive processing
1462 ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
1463 ElementAllocas.reserve(AT->getNumElements());
1464 Type *ElTy = AT->getElementType();
1465 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1466 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
1467 AI->getName() + "." + Twine(i), AI);
1468 ElementAllocas.push_back(NA);
1469 WorkList.push_back(NA); // Add to worklist for recursive processing
1473 // Now that we have created the new alloca instructions, rewrite all the
1474 // uses of the old alloca.
1475 RewriteForScalarRepl(AI, AI, 0, ElementAllocas);
1477 // Now erase any instructions that were made dead while rewriting the alloca.
1478 DeleteDeadInstructions();
1479 AI->eraseFromParent();
1484 /// DeleteDeadInstructions - Erase instructions on the DeadInstrs list,
1485 /// recursively including all their operands that become trivially dead.
1486 void SROA::DeleteDeadInstructions() {
1487 while (!DeadInsts.empty()) {
1488 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val());
1490 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
1491 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
1492 // Zero out the operand and see if it becomes trivially dead.
1493 // (But, don't add allocas to the dead instruction list -- they are
1494 // already on the worklist and will be deleted separately.)
1496 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U))
1497 DeadInsts.push_back(U);
1500 I->eraseFromParent();
1504 /// isSafeForScalarRepl - Check if instruction I is a safe use with regard to
1505 /// performing scalar replacement of alloca AI. The results are flagged in
1506 /// the Info parameter. Offset indicates the position within AI that is
1507 /// referenced by this instruction.
1508 void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset,
1510 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
1511 Instruction *User = cast<Instruction>(*UI);
1513 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1514 isSafeForScalarRepl(BC, Offset, Info);
1515 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1516 uint64_t GEPOffset = Offset;
1517 isSafeGEP(GEPI, GEPOffset, Info);
1519 isSafeForScalarRepl(GEPI, GEPOffset, Info);
1520 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
1521 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
1523 return MarkUnsafe(Info, User);
1524 isSafeMemAccess(Offset, Length->getZExtValue(), 0,
1525 UI.getOperandNo() == 0, Info, MI,
1526 true /*AllowWholeAccess*/);
1527 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1528 if (!LI->isSimple())
1529 return MarkUnsafe(Info, User);
1530 Type *LIType = LI->getType();
1531 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType),
1532 LIType, false, Info, LI, true /*AllowWholeAccess*/);
1533 Info.hasALoadOrStore = true;
1535 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1536 // Store is ok if storing INTO the pointer, not storing the pointer
1537 if (!SI->isSimple() || SI->getOperand(0) == I)
1538 return MarkUnsafe(Info, User);
1540 Type *SIType = SI->getOperand(0)->getType();
1541 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType),
1542 SIType, true, Info, SI, true /*AllowWholeAccess*/);
1543 Info.hasALoadOrStore = true;
1544 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
1545 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1546 II->getIntrinsicID() != Intrinsic::lifetime_end)
1547 return MarkUnsafe(Info, User);
1548 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) {
1549 isSafePHISelectUseForScalarRepl(User, Offset, Info);
1551 return MarkUnsafe(Info, User);
1553 if (Info.isUnsafe) return;
1558 /// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer
1559 /// derived from the alloca, we can often still split the alloca into elements.
1560 /// This is useful if we have a large alloca where one element is phi'd
1561 /// together somewhere: we can SRoA and promote all the other elements even if
1562 /// we end up not being able to promote this one.
1564 /// All we require is that the uses of the PHI do not index into other parts of
1565 /// the alloca. The most important use case for this is single load and stores
1566 /// that are PHI'd together, which can happen due to code sinking.
1567 void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset,
1569 // If we've already checked this PHI, don't do it again.
1570 if (PHINode *PN = dyn_cast<PHINode>(I))
1571 if (!Info.CheckedPHIs.insert(PN))
1574 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
1575 Instruction *User = cast<Instruction>(*UI);
1577 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1578 isSafePHISelectUseForScalarRepl(BC, Offset, Info);
1579 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1580 // Only allow "bitcast" GEPs for simplicity. We could generalize this,
1581 // but would have to prove that we're staying inside of an element being
1583 if (!GEPI->hasAllZeroIndices())
1584 return MarkUnsafe(Info, User);
1585 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info);
1586 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1587 if (!LI->isSimple())
1588 return MarkUnsafe(Info, User);
1589 Type *LIType = LI->getType();
1590 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType),
1591 LIType, false, Info, LI, false /*AllowWholeAccess*/);
1592 Info.hasALoadOrStore = true;
1594 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1595 // Store is ok if storing INTO the pointer, not storing the pointer
1596 if (!SI->isSimple() || SI->getOperand(0) == I)
1597 return MarkUnsafe(Info, User);
1599 Type *SIType = SI->getOperand(0)->getType();
1600 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType),
1601 SIType, true, Info, SI, false /*AllowWholeAccess*/);
1602 Info.hasALoadOrStore = true;
1603 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) {
1604 isSafePHISelectUseForScalarRepl(User, Offset, Info);
1606 return MarkUnsafe(Info, User);
1608 if (Info.isUnsafe) return;
1612 /// isSafeGEP - Check if a GEP instruction can be handled for scalar
1613 /// replacement. It is safe when all the indices are constant, in-bounds
1614 /// references, and when the resulting offset corresponds to an element within
1615 /// the alloca type. The results are flagged in the Info parameter. Upon
1616 /// return, Offset is adjusted as specified by the GEP indices.
1617 void SROA::isSafeGEP(GetElementPtrInst *GEPI,
1618 uint64_t &Offset, AllocaInfo &Info) {
1619 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI);
1623 // Walk through the GEP type indices, checking the types that this indexes
1625 for (; GEPIt != E; ++GEPIt) {
1626 // Ignore struct elements, no extra checking needed for these.
1627 if ((*GEPIt)->isStructTy())
1630 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand());
1632 return MarkUnsafe(Info, GEPI);
1635 // Compute the offset due to this GEP and check if the alloca has a
1636 // component element at that offset.
1637 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
1638 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices);
1639 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0))
1640 MarkUnsafe(Info, GEPI);
1643 /// isHomogeneousAggregate - Check if type T is a struct or array containing
1644 /// elements of the same type (which is always true for arrays). If so,
1645 /// return true with NumElts and EltTy set to the number of elements and the
1646 /// element type, respectively.
1647 static bool isHomogeneousAggregate(Type *T, unsigned &NumElts,
1649 if (ArrayType *AT = dyn_cast<ArrayType>(T)) {
1650 NumElts = AT->getNumElements();
1651 EltTy = (NumElts == 0 ? 0 : AT->getElementType());
1654 if (StructType *ST = dyn_cast<StructType>(T)) {
1655 NumElts = ST->getNumContainedTypes();
1656 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0));
1657 for (unsigned n = 1; n < NumElts; ++n) {
1658 if (ST->getContainedType(n) != EltTy)
1666 /// isCompatibleAggregate - Check if T1 and T2 are either the same type or are
1667 /// "homogeneous" aggregates with the same element type and number of elements.
1668 static bool isCompatibleAggregate(Type *T1, Type *T2) {
1672 unsigned NumElts1, NumElts2;
1673 Type *EltTy1, *EltTy2;
1674 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) &&
1675 isHomogeneousAggregate(T2, NumElts2, EltTy2) &&
1676 NumElts1 == NumElts2 &&
1683 /// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI
1684 /// alloca or has an offset and size that corresponds to a component element
1685 /// within it. The offset checked here may have been formed from a GEP with a
1686 /// pointer bitcasted to a different type.
1688 /// If AllowWholeAccess is true, then this allows uses of the entire alloca as a
1689 /// unit. If false, it only allows accesses known to be in a single element.
1690 void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize,
1691 Type *MemOpType, bool isStore,
1692 AllocaInfo &Info, Instruction *TheAccess,
1693 bool AllowWholeAccess) {
1694 // Check if this is a load/store of the entire alloca.
1695 if (Offset == 0 && AllowWholeAccess &&
1696 MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) {
1697 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer
1698 // loads/stores (which are essentially the same as the MemIntrinsics with
1699 // regard to copying padding between elements). But, if an alloca is
1700 // flagged as both a source and destination of such operations, we'll need
1701 // to check later for padding between elements.
1702 if (!MemOpType || MemOpType->isIntegerTy()) {
1704 Info.isMemCpyDst = true;
1706 Info.isMemCpySrc = true;
1709 // This is also safe for references using a type that is compatible with
1710 // the type of the alloca, so that loads/stores can be rewritten using
1711 // insertvalue/extractvalue.
1712 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) {
1713 Info.hasSubelementAccess = true;
1717 // Check if the offset/size correspond to a component within the alloca type.
1718 Type *T = Info.AI->getAllocatedType();
1719 if (TypeHasComponent(T, Offset, MemSize)) {
1720 Info.hasSubelementAccess = true;
1724 return MarkUnsafe(Info, TheAccess);
1727 /// TypeHasComponent - Return true if T has a component type with the
1728 /// specified offset and size. If Size is zero, do not check the size.
1729 bool SROA::TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size) {
1732 if (StructType *ST = dyn_cast<StructType>(T)) {
1733 const StructLayout *Layout = TD->getStructLayout(ST);
1734 unsigned EltIdx = Layout->getElementContainingOffset(Offset);
1735 EltTy = ST->getContainedType(EltIdx);
1736 EltSize = TD->getTypeAllocSize(EltTy);
1737 Offset -= Layout->getElementOffset(EltIdx);
1738 } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) {
1739 EltTy = AT->getElementType();
1740 EltSize = TD->getTypeAllocSize(EltTy);
1741 if (Offset >= AT->getNumElements() * EltSize)
1747 if (Offset == 0 && (Size == 0 || EltSize == Size))
1749 // Check if the component spans multiple elements.
1750 if (Offset + Size > EltSize)
1752 return TypeHasComponent(EltTy, Offset, Size);
1755 /// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite
1756 /// the instruction I, which references it, to use the separate elements.
1757 /// Offset indicates the position within AI that is referenced by this
1759 void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
1760 SmallVector<AllocaInst*, 32> &NewElts) {
1761 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) {
1762 Use &TheUse = UI.getUse();
1763 Instruction *User = cast<Instruction>(*UI++);
1765 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1766 RewriteBitCast(BC, AI, Offset, NewElts);
1770 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1771 RewriteGEP(GEPI, AI, Offset, NewElts);
1775 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
1776 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
1777 uint64_t MemSize = Length->getZExtValue();
1779 MemSize == TD->getTypeAllocSize(AI->getAllocatedType()))
1780 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts);
1781 // Otherwise the intrinsic can only touch a single element and the
1782 // address operand will be updated, so nothing else needs to be done.
1786 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
1787 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
1788 II->getIntrinsicID() == Intrinsic::lifetime_end) {
1789 RewriteLifetimeIntrinsic(II, AI, Offset, NewElts);
1794 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1795 Type *LIType = LI->getType();
1797 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) {
1799 // %res = load { i32, i32 }* %alloc
1801 // %load.0 = load i32* %alloc.0
1802 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0
1803 // %load.1 = load i32* %alloc.1
1804 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1
1805 // (Also works for arrays instead of structs)
1806 Value *Insert = UndefValue::get(LIType);
1807 IRBuilder<> Builder(LI);
1808 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1809 Value *Load = Builder.CreateLoad(NewElts[i], "load");
1810 Insert = Builder.CreateInsertValue(Insert, Load, i, "insert");
1812 LI->replaceAllUsesWith(Insert);
1813 DeadInsts.push_back(LI);
1814 } else if (LIType->isIntegerTy() &&
1815 TD->getTypeAllocSize(LIType) ==
1816 TD->getTypeAllocSize(AI->getAllocatedType())) {
1817 // If this is a load of the entire alloca to an integer, rewrite it.
1818 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts);
1823 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1824 Value *Val = SI->getOperand(0);
1825 Type *SIType = Val->getType();
1826 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) {
1828 // store { i32, i32 } %val, { i32, i32 }* %alloc
1830 // %val.0 = extractvalue { i32, i32 } %val, 0
1831 // store i32 %val.0, i32* %alloc.0
1832 // %val.1 = extractvalue { i32, i32 } %val, 1
1833 // store i32 %val.1, i32* %alloc.1
1834 // (Also works for arrays instead of structs)
1835 IRBuilder<> Builder(SI);
1836 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1837 Value *Extract = Builder.CreateExtractValue(Val, i, Val->getName());
1838 Builder.CreateStore(Extract, NewElts[i]);
1840 DeadInsts.push_back(SI);
1841 } else if (SIType->isIntegerTy() &&
1842 TD->getTypeAllocSize(SIType) ==
1843 TD->getTypeAllocSize(AI->getAllocatedType())) {
1844 // If this is a store of the entire alloca from an integer, rewrite it.
1845 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts);
1850 if (isa<SelectInst>(User) || isa<PHINode>(User)) {
1851 // If we have a PHI user of the alloca itself (as opposed to a GEP or
1852 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to
1854 if (!isa<AllocaInst>(I)) continue;
1856 assert(Offset == 0 && NewElts[0] &&
1857 "Direct alloca use should have a zero offset");
1859 // If we have a use of the alloca, we know the derived uses will be
1860 // utilizing just the first element of the scalarized result. Insert a
1861 // bitcast of the first alloca before the user as required.
1862 AllocaInst *NewAI = NewElts[0];
1863 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI);
1864 NewAI->moveBefore(BCI);
1871 /// RewriteBitCast - Update a bitcast reference to the alloca being replaced
1872 /// and recursively continue updating all of its uses.
1873 void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
1874 SmallVector<AllocaInst*, 32> &NewElts) {
1875 RewriteForScalarRepl(BC, AI, Offset, NewElts);
1876 if (BC->getOperand(0) != AI)
1879 // The bitcast references the original alloca. Replace its uses with
1880 // references to the alloca containing offset zero (which is normally at
1881 // index zero, but might not be in cases involving structs with elements
1883 Type *T = AI->getAllocatedType();
1884 uint64_t EltOffset = 0;
1886 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy);
1887 Instruction *Val = NewElts[Idx];
1888 if (Val->getType() != BC->getDestTy()) {
1889 Val = new BitCastInst(Val, BC->getDestTy(), "", BC);
1892 BC->replaceAllUsesWith(Val);
1893 DeadInsts.push_back(BC);
1896 /// FindElementAndOffset - Return the index of the element containing Offset
1897 /// within the specified type, which must be either a struct or an array.
1898 /// Sets T to the type of the element and Offset to the offset within that
1899 /// element. IdxTy is set to the type of the index result to be used in a
1900 /// GEP instruction.
1901 uint64_t SROA::FindElementAndOffset(Type *&T, uint64_t &Offset,
1904 if (StructType *ST = dyn_cast<StructType>(T)) {
1905 const StructLayout *Layout = TD->getStructLayout(ST);
1906 Idx = Layout->getElementContainingOffset(Offset);
1907 T = ST->getContainedType(Idx);
1908 Offset -= Layout->getElementOffset(Idx);
1909 IdxTy = Type::getInt32Ty(T->getContext());
1912 ArrayType *AT = cast<ArrayType>(T);
1913 T = AT->getElementType();
1914 uint64_t EltSize = TD->getTypeAllocSize(T);
1915 Idx = Offset / EltSize;
1916 Offset -= Idx * EltSize;
1917 IdxTy = Type::getInt64Ty(T->getContext());
1921 /// RewriteGEP - Check if this GEP instruction moves the pointer across
1922 /// elements of the alloca that are being split apart, and if so, rewrite
1923 /// the GEP to be relative to the new element.
1924 void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
1925 SmallVector<AllocaInst*, 32> &NewElts) {
1926 uint64_t OldOffset = Offset;
1927 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
1928 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices);
1930 RewriteForScalarRepl(GEPI, AI, Offset, NewElts);
1932 Type *T = AI->getAllocatedType();
1934 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy);
1935 if (GEPI->getOperand(0) == AI)
1936 OldIdx = ~0ULL; // Force the GEP to be rewritten.
1938 T = AI->getAllocatedType();
1939 uint64_t EltOffset = Offset;
1940 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy);
1942 // If this GEP does not move the pointer across elements of the alloca
1943 // being split, then it does not needs to be rewritten.
1947 Type *i32Ty = Type::getInt32Ty(AI->getContext());
1948 SmallVector<Value*, 8> NewArgs;
1949 NewArgs.push_back(Constant::getNullValue(i32Ty));
1950 while (EltOffset != 0) {
1951 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy);
1952 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx));
1954 Instruction *Val = NewElts[Idx];
1955 if (NewArgs.size() > 1) {
1956 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs, "", GEPI);
1957 Val->takeName(GEPI);
1959 if (Val->getType() != GEPI->getType())
1960 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI);
1961 GEPI->replaceAllUsesWith(Val);
1962 DeadInsts.push_back(GEPI);
1965 /// RewriteLifetimeIntrinsic - II is a lifetime.start/lifetime.end. Rewrite it
1966 /// to mark the lifetime of the scalarized memory.
1967 void SROA::RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI,
1969 SmallVector<AllocaInst*, 32> &NewElts) {
1970 ConstantInt *OldSize = cast<ConstantInt>(II->getArgOperand(0));
1971 // Put matching lifetime markers on everything from Offset up to
1973 Type *AIType = AI->getAllocatedType();
1974 uint64_t NewOffset = Offset;
1976 uint64_t Idx = FindElementAndOffset(AIType, NewOffset, IdxTy);
1978 IRBuilder<> Builder(II);
1979 uint64_t Size = OldSize->getLimitedValue();
1982 // Splice the first element and index 'NewOffset' bytes in. SROA will
1983 // split the alloca again later.
1984 Value *V = Builder.CreateBitCast(NewElts[Idx], Builder.getInt8PtrTy());
1985 V = Builder.CreateGEP(V, Builder.getInt64(NewOffset));
1987 IdxTy = NewElts[Idx]->getAllocatedType();
1988 uint64_t EltSize = TD->getTypeAllocSize(IdxTy) - NewOffset;
1989 if (EltSize > Size) {
1995 if (II->getIntrinsicID() == Intrinsic::lifetime_start)
1996 Builder.CreateLifetimeStart(V, Builder.getInt64(EltSize));
1998 Builder.CreateLifetimeEnd(V, Builder.getInt64(EltSize));
2002 for (; Idx != NewElts.size() && Size; ++Idx) {
2003 IdxTy = NewElts[Idx]->getAllocatedType();
2004 uint64_t EltSize = TD->getTypeAllocSize(IdxTy);
2005 if (EltSize > Size) {
2011 if (II->getIntrinsicID() == Intrinsic::lifetime_start)
2012 Builder.CreateLifetimeStart(NewElts[Idx],
2013 Builder.getInt64(EltSize));
2015 Builder.CreateLifetimeEnd(NewElts[Idx],
2016 Builder.getInt64(EltSize));
2018 DeadInsts.push_back(II);
2021 /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI.
2022 /// Rewrite it to copy or set the elements of the scalarized memory.
2023 void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
2025 SmallVector<AllocaInst*, 32> &NewElts) {
2026 // If this is a memcpy/memmove, construct the other pointer as the
2027 // appropriate type. The "Other" pointer is the pointer that goes to memory
2028 // that doesn't have anything to do with the alloca that we are promoting. For
2029 // memset, this Value* stays null.
2030 Value *OtherPtr = 0;
2031 unsigned MemAlignment = MI->getAlignment();
2032 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy
2033 if (Inst == MTI->getRawDest())
2034 OtherPtr = MTI->getRawSource();
2036 assert(Inst == MTI->getRawSource());
2037 OtherPtr = MTI->getRawDest();
2041 // If there is an other pointer, we want to convert it to the same pointer
2042 // type as AI has, so we can GEP through it safely.
2044 unsigned AddrSpace =
2045 cast<PointerType>(OtherPtr->getType())->getAddressSpace();
2047 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an
2048 // optimization, but it's also required to detect the corner case where
2049 // both pointer operands are referencing the same memory, and where
2050 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This
2051 // function is only called for mem intrinsics that access the whole
2052 // aggregate, so non-zero GEPs are not an issue here.)
2053 OtherPtr = OtherPtr->stripPointerCasts();
2055 // Copying the alloca to itself is a no-op: just delete it.
2056 if (OtherPtr == AI || OtherPtr == NewElts[0]) {
2057 // This code will run twice for a no-op memcpy -- once for each operand.
2058 // Put only one reference to MI on the DeadInsts list.
2059 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(),
2060 E = DeadInsts.end(); I != E; ++I)
2061 if (*I == MI) return;
2062 DeadInsts.push_back(MI);
2066 // If the pointer is not the right type, insert a bitcast to the right
2069 PointerType::get(AI->getType()->getElementType(), AddrSpace);
2071 if (OtherPtr->getType() != NewTy)
2072 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI);
2075 // Process each element of the aggregate.
2076 bool SROADest = MI->getRawDest() == Inst;
2078 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext()));
2080 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2081 // If this is a memcpy/memmove, emit a GEP of the other element address.
2082 Value *OtherElt = 0;
2083 unsigned OtherEltAlign = MemAlignment;
2086 Value *Idx[2] = { Zero,
2087 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) };
2088 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx,
2089 OtherPtr->getName()+"."+Twine(i),
2092 PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType());
2093 Type *OtherTy = OtherPtrTy->getElementType();
2094 if (StructType *ST = dyn_cast<StructType>(OtherTy)) {
2095 EltOffset = TD->getStructLayout(ST)->getElementOffset(i);
2097 Type *EltTy = cast<SequentialType>(OtherTy)->getElementType();
2098 EltOffset = TD->getTypeAllocSize(EltTy)*i;
2101 // The alignment of the other pointer is the guaranteed alignment of the
2102 // element, which is affected by both the known alignment of the whole
2103 // mem intrinsic and the alignment of the element. If the alignment of
2104 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the
2105 // known alignment is just 4 bytes.
2106 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset);
2109 Value *EltPtr = NewElts[i];
2110 Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType();
2112 // If we got down to a scalar, insert a load or store as appropriate.
2113 if (EltTy->isSingleValueType()) {
2114 if (isa<MemTransferInst>(MI)) {
2116 // From Other to Alloca.
2117 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI);
2118 new StoreInst(Elt, EltPtr, MI);
2120 // From Alloca to Other.
2121 Value *Elt = new LoadInst(EltPtr, "tmp", MI);
2122 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI);
2126 assert(isa<MemSetInst>(MI));
2128 // If the stored element is zero (common case), just store a null
2131 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) {
2133 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0>
2135 // If EltTy is a vector type, get the element type.
2136 Type *ValTy = EltTy->getScalarType();
2138 // Construct an integer with the right value.
2139 unsigned EltSize = TD->getTypeSizeInBits(ValTy);
2140 APInt OneVal(EltSize, CI->getZExtValue());
2141 APInt TotalVal(OneVal);
2143 for (unsigned i = 0; 8*i < EltSize; ++i) {
2144 TotalVal = TotalVal.shl(8);
2148 // Convert the integer value to the appropriate type.
2149 StoreVal = ConstantInt::get(CI->getContext(), TotalVal);
2150 if (ValTy->isPointerTy())
2151 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy);
2152 else if (ValTy->isFloatingPointTy())
2153 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy);
2154 assert(StoreVal->getType() == ValTy && "Type mismatch!");
2156 // If the requested value was a vector constant, create it.
2157 if (EltTy->isVectorTy()) {
2158 unsigned NumElts = cast<VectorType>(EltTy)->getNumElements();
2159 SmallVector<Constant*, 16> Elts(NumElts, StoreVal);
2160 StoreVal = ConstantVector::get(Elts);
2163 new StoreInst(StoreVal, EltPtr, MI);
2166 // Otherwise, if we're storing a byte variable, use a memset call for
2170 unsigned EltSize = TD->getTypeAllocSize(EltTy);
2174 IRBuilder<> Builder(MI);
2176 // Finally, insert the meminst for this element.
2177 if (isa<MemSetInst>(MI)) {
2178 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize,
2181 assert(isa<MemTransferInst>(MI));
2182 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr
2183 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr
2185 if (isa<MemCpyInst>(MI))
2186 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile());
2188 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile());
2191 DeadInsts.push_back(MI);
2194 /// RewriteStoreUserOfWholeAlloca - We found a store of an integer that
2195 /// overwrites the entire allocation. Extract out the pieces of the stored
2196 /// integer and store them individually.
2197 void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
2198 SmallVector<AllocaInst*, 32> &NewElts){
2199 // Extract each element out of the integer according to its structure offset
2200 // and store the element value to the individual alloca.
2201 Value *SrcVal = SI->getOperand(0);
2202 Type *AllocaEltTy = AI->getAllocatedType();
2203 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
2205 IRBuilder<> Builder(SI);
2207 // Handle tail padding by extending the operand
2208 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits)
2209 SrcVal = Builder.CreateZExt(SrcVal,
2210 IntegerType::get(SI->getContext(), AllocaSizeBits));
2212 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI
2215 // There are two forms here: AI could be an array or struct. Both cases
2216 // have different ways to compute the element offset.
2217 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
2218 const StructLayout *Layout = TD->getStructLayout(EltSTy);
2220 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2221 // Get the number of bits to shift SrcVal to get the value.
2222 Type *FieldTy = EltSTy->getElementType(i);
2223 uint64_t Shift = Layout->getElementOffsetInBits(i);
2225 if (TD->isBigEndian())
2226 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy);
2228 Value *EltVal = SrcVal;
2230 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
2231 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt");
2234 // Truncate down to an integer of the right size.
2235 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
2237 // Ignore zero sized fields like {}, they obviously contain no data.
2238 if (FieldSizeBits == 0) continue;
2240 if (FieldSizeBits != AllocaSizeBits)
2241 EltVal = Builder.CreateTrunc(EltVal,
2242 IntegerType::get(SI->getContext(), FieldSizeBits));
2243 Value *DestField = NewElts[i];
2244 if (EltVal->getType() == FieldTy) {
2245 // Storing to an integer field of this size, just do it.
2246 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) {
2247 // Bitcast to the right element type (for fp/vector values).
2248 EltVal = Builder.CreateBitCast(EltVal, FieldTy);
2250 // Otherwise, bitcast the dest pointer (for aggregates).
2251 DestField = Builder.CreateBitCast(DestField,
2252 PointerType::getUnqual(EltVal->getType()));
2254 new StoreInst(EltVal, DestField, SI);
2258 ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
2259 Type *ArrayEltTy = ATy->getElementType();
2260 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
2261 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy);
2265 if (TD->isBigEndian())
2266 Shift = AllocaSizeBits-ElementOffset;
2270 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2271 // Ignore zero sized fields like {}, they obviously contain no data.
2272 if (ElementSizeBits == 0) continue;
2274 Value *EltVal = SrcVal;
2276 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
2277 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt");
2280 // Truncate down to an integer of the right size.
2281 if (ElementSizeBits != AllocaSizeBits)
2282 EltVal = Builder.CreateTrunc(EltVal,
2283 IntegerType::get(SI->getContext(),
2285 Value *DestField = NewElts[i];
2286 if (EltVal->getType() == ArrayEltTy) {
2287 // Storing to an integer field of this size, just do it.
2288 } else if (ArrayEltTy->isFloatingPointTy() ||
2289 ArrayEltTy->isVectorTy()) {
2290 // Bitcast to the right element type (for fp/vector values).
2291 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy);
2293 // Otherwise, bitcast the dest pointer (for aggregates).
2294 DestField = Builder.CreateBitCast(DestField,
2295 PointerType::getUnqual(EltVal->getType()));
2297 new StoreInst(EltVal, DestField, SI);
2299 if (TD->isBigEndian())
2300 Shift -= ElementOffset;
2302 Shift += ElementOffset;
2306 DeadInsts.push_back(SI);
2309 /// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to
2310 /// an integer. Load the individual pieces to form the aggregate value.
2311 void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
2312 SmallVector<AllocaInst*, 32> &NewElts) {
2313 // Extract each element out of the NewElts according to its structure offset
2314 // and form the result value.
2315 Type *AllocaEltTy = AI->getAllocatedType();
2316 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
2318 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI
2321 // There are two forms here: AI could be an array or struct. Both cases
2322 // have different ways to compute the element offset.
2323 const StructLayout *Layout = 0;
2324 uint64_t ArrayEltBitOffset = 0;
2325 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
2326 Layout = TD->getStructLayout(EltSTy);
2328 Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
2329 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
2333 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits));
2335 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2336 // Load the value from the alloca. If the NewElt is an aggregate, cast
2337 // the pointer to an integer of the same size before doing the load.
2338 Value *SrcField = NewElts[i];
2340 cast<PointerType>(SrcField->getType())->getElementType();
2341 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
2343 // Ignore zero sized fields like {}, they obviously contain no data.
2344 if (FieldSizeBits == 0) continue;
2346 IntegerType *FieldIntTy = IntegerType::get(LI->getContext(),
2348 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() &&
2349 !FieldTy->isVectorTy())
2350 SrcField = new BitCastInst(SrcField,
2351 PointerType::getUnqual(FieldIntTy),
2353 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI);
2355 // If SrcField is a fp or vector of the right size but that isn't an
2356 // integer type, bitcast to an integer so we can shift it.
2357 if (SrcField->getType() != FieldIntTy)
2358 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI);
2360 // Zero extend the field to be the same size as the final alloca so that
2361 // we can shift and insert it.
2362 if (SrcField->getType() != ResultVal->getType())
2363 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI);
2365 // Determine the number of bits to shift SrcField.
2367 if (Layout) // Struct case.
2368 Shift = Layout->getElementOffsetInBits(i);
2370 Shift = i*ArrayEltBitOffset;
2372 if (TD->isBigEndian())
2373 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth();
2376 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift);
2377 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI);
2380 // Don't create an 'or x, 0' on the first iteration.
2381 if (!isa<Constant>(ResultVal) ||
2382 !cast<Constant>(ResultVal)->isNullValue())
2383 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI);
2385 ResultVal = SrcField;
2388 // Handle tail padding by truncating the result
2389 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits)
2390 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI);
2392 LI->replaceAllUsesWith(ResultVal);
2393 DeadInsts.push_back(LI);
2396 /// HasPadding - Return true if the specified type has any structure or
2397 /// alignment padding in between the elements that would be split apart
2398 /// by SROA; return false otherwise.
2399 static bool HasPadding(Type *Ty, const TargetData &TD) {
2400 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2401 Ty = ATy->getElementType();
2402 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty);
2405 // SROA currently handles only Arrays and Structs.
2406 StructType *STy = cast<StructType>(Ty);
2407 const StructLayout *SL = TD.getStructLayout(STy);
2408 unsigned PrevFieldBitOffset = 0;
2409 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2410 unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
2412 // Check to see if there is any padding between this element and the
2415 unsigned PrevFieldEnd =
2416 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1));
2417 if (PrevFieldEnd < FieldBitOffset)
2420 PrevFieldBitOffset = FieldBitOffset;
2422 // Check for tail padding.
2423 if (unsigned EltCount = STy->getNumElements()) {
2424 unsigned PrevFieldEnd = PrevFieldBitOffset +
2425 TD.getTypeSizeInBits(STy->getElementType(EltCount-1));
2426 if (PrevFieldEnd < SL->getSizeInBits())
2432 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
2433 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe,
2434 /// or 1 if safe after canonicalization has been performed.
2435 bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) {
2436 // Loop over the use list of the alloca. We can only transform it if all of
2437 // the users are safe to transform.
2438 AllocaInfo Info(AI);
2440 isSafeForScalarRepl(AI, 0, Info);
2441 if (Info.isUnsafe) {
2442 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n');
2446 // Okay, we know all the users are promotable. If the aggregate is a memcpy
2447 // source and destination, we have to be careful. In particular, the memcpy
2448 // could be moving around elements that live in structure padding of the LLVM
2449 // types, but may actually be used. In these cases, we refuse to promote the
2451 if (Info.isMemCpySrc && Info.isMemCpyDst &&
2452 HasPadding(AI->getAllocatedType(), *TD))
2455 // If the alloca never has an access to just *part* of it, but is accessed
2456 // via loads and stores, then we should use ConvertToScalarInfo to promote
2457 // the alloca instead of promoting each piece at a time and inserting fission
2459 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) {
2460 // If the struct/array just has one element, use basic SRoA.
2461 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
2462 if (ST->getNumElements() > 1) return false;
2464 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1)
2474 /// PointsToConstantGlobal - Return true if V (possibly indirectly) points to
2475 /// some part of a constant global variable. This intentionally only accepts
2476 /// constant expressions because we don't can't rewrite arbitrary instructions.
2477 static bool PointsToConstantGlobal(Value *V) {
2478 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
2479 return GV->isConstant();
2480 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
2481 if (CE->getOpcode() == Instruction::BitCast ||
2482 CE->getOpcode() == Instruction::GetElementPtr)
2483 return PointsToConstantGlobal(CE->getOperand(0));
2487 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
2488 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
2489 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
2490 /// track of whether it moves the pointer (with isOffset) but otherwise traverse
2491 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
2492 /// the alloca, and if the source pointer is a pointer to a constant global, we
2493 /// can optimize this.
2495 isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy,
2497 SmallVector<Instruction *, 4> &LifetimeMarkers) {
2498 // We track lifetime intrinsics as we encounter them. If we decide to go
2499 // ahead and replace the value with the global, this lets the caller quickly
2500 // eliminate the markers.
2502 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
2503 User *U = cast<Instruction>(*UI);
2505 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
2506 // Ignore non-volatile loads, they are always ok.
2507 if (!LI->isSimple()) return false;
2511 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
2512 // If uses of the bitcast are ok, we are ok.
2513 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset,
2518 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
2519 // If the GEP has all zero indices, it doesn't offset the pointer. If it
2520 // doesn't, it does.
2521 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy,
2522 isOffset || !GEP->hasAllZeroIndices(),
2528 if (CallSite CS = U) {
2529 // If this is the function being called then we treat it like a load and
2531 if (CS.isCallee(UI))
2534 // If this is a readonly/readnone call site, then we know it is just a
2535 // load (but one that potentially returns the value itself), so we can
2536 // ignore it if we know that the value isn't captured.
2537 unsigned ArgNo = CS.getArgumentNo(UI);
2538 if (CS.onlyReadsMemory() &&
2539 (CS.getInstruction()->use_empty() || CS.doesNotCapture(ArgNo)))
2542 // If this is being passed as a byval argument, the caller is making a
2543 // copy, so it is only a read of the alloca.
2544 if (CS.isByValArgument(ArgNo))
2548 // Lifetime intrinsics can be handled by the caller.
2549 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
2550 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
2551 II->getIntrinsicID() == Intrinsic::lifetime_end) {
2552 assert(II->use_empty() && "Lifetime markers have no result to use!");
2553 LifetimeMarkers.push_back(II);
2558 // If this is isn't our memcpy/memmove, reject it as something we can't
2560 MemTransferInst *MI = dyn_cast<MemTransferInst>(U);
2564 // If the transfer is using the alloca as a source of the transfer, then
2565 // ignore it since it is a load (unless the transfer is volatile).
2566 if (UI.getOperandNo() == 1) {
2567 if (MI->isVolatile()) return false;
2571 // If we already have seen a copy, reject the second one.
2572 if (TheCopy) return false;
2574 // If the pointer has been offset from the start of the alloca, we can't
2575 // safely handle this.
2576 if (isOffset) return false;
2578 // If the memintrinsic isn't using the alloca as the dest, reject it.
2579 if (UI.getOperandNo() != 0) return false;
2581 // If the source of the memcpy/move is not a constant global, reject it.
2582 if (!PointsToConstantGlobal(MI->getSource()))
2585 // Otherwise, the transform is safe. Remember the copy instruction.
2591 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
2592 /// modified by a copy from a constant global. If we can prove this, we can
2593 /// replace any uses of the alloca with uses of the global directly.
2595 SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI,
2596 SmallVector<Instruction*, 4> &ToDelete) {
2597 MemTransferInst *TheCopy = 0;
2598 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false, ToDelete))