1 //===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation implements the well known scalar replacement of
11 // aggregates transformation. This xform breaks up alloca instructions of
12 // aggregate type (structure or array) into individual alloca instructions for
13 // each member (if possible). Then, if possible, it transforms the individual
14 // alloca instructions into nice clean scalar SSA form.
16 // This combines a simple SRoA algorithm with the Mem2Reg algorithm because they
17 // often interact, especially for C++ programs. As such, iterating between
18 // SRoA, then Mem2Reg until we run out of things to promote works well.
20 //===----------------------------------------------------------------------===//
22 #include "llvm/Transforms/Scalar.h"
23 #include "llvm/ADT/SetVector.h"
24 #include "llvm/ADT/SmallVector.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/Analysis/AssumptionCache.h"
27 #include "llvm/Analysis/Loads.h"
28 #include "llvm/Analysis/ValueTracking.h"
29 #include "llvm/IR/CallSite.h"
30 #include "llvm/IR/Constants.h"
31 #include "llvm/IR/DIBuilder.h"
32 #include "llvm/IR/DataLayout.h"
33 #include "llvm/IR/DebugInfo.h"
34 #include "llvm/IR/DerivedTypes.h"
35 #include "llvm/IR/Dominators.h"
36 #include "llvm/IR/Function.h"
37 #include "llvm/IR/GetElementPtrTypeIterator.h"
38 #include "llvm/IR/GlobalVariable.h"
39 #include "llvm/IR/IRBuilder.h"
40 #include "llvm/IR/Instructions.h"
41 #include "llvm/IR/IntrinsicInst.h"
42 #include "llvm/IR/LLVMContext.h"
43 #include "llvm/IR/Module.h"
44 #include "llvm/IR/Operator.h"
45 #include "llvm/Pass.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/ErrorHandling.h"
48 #include "llvm/Support/MathExtras.h"
49 #include "llvm/Support/raw_ostream.h"
50 #include "llvm/Transforms/Utils/Local.h"
51 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
52 #include "llvm/Transforms/Utils/SSAUpdater.h"
55 #define DEBUG_TYPE "scalarrepl"
57 STATISTIC(NumReplaced, "Number of allocas broken up");
58 STATISTIC(NumPromoted, "Number of allocas promoted");
59 STATISTIC(NumAdjusted, "Number of scalar allocas adjusted to allow promotion");
60 STATISTIC(NumConverted, "Number of aggregates converted to scalar");
63 struct SROA : public FunctionPass {
64 SROA(int T, bool hasDT, char &ID, int ST, int AT, int SLT)
65 : FunctionPass(ID), HasDomTree(hasDT) {
71 StructMemberThreshold = 32;
73 StructMemberThreshold = ST;
75 ArrayElementThreshold = 8;
77 ArrayElementThreshold = AT;
79 // Do not limit the scalar integer load size if no threshold is given.
80 ScalarLoadThreshold = -1;
82 ScalarLoadThreshold = SLT;
85 bool runOnFunction(Function &F) override;
87 bool performScalarRepl(Function &F);
88 bool performPromotion(Function &F);
93 /// DeadInsts - Keep track of instructions we have made dead, so that
94 /// we can remove them after we are done working.
95 SmallVector<Value*, 32> DeadInsts;
97 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures
98 /// information about the uses. All these fields are initialized to false
99 /// and set to true when something is learned.
101 /// The alloca to promote.
104 /// CheckedPHIs - This is a set of verified PHI nodes, to prevent infinite
105 /// looping and avoid redundant work.
106 SmallPtrSet<PHINode*, 8> CheckedPHIs;
108 /// isUnsafe - This is set to true if the alloca cannot be SROA'd.
111 /// isMemCpySrc - This is true if this aggregate is memcpy'd from.
112 bool isMemCpySrc : 1;
114 /// isMemCpyDst - This is true if this aggregate is memcpy'd into.
115 bool isMemCpyDst : 1;
117 /// hasSubelementAccess - This is true if a subelement of the alloca is
118 /// ever accessed, or false if the alloca is only accessed with mem
119 /// intrinsics or load/store that only access the entire alloca at once.
120 bool hasSubelementAccess : 1;
122 /// hasALoadOrStore - This is true if there are any loads or stores to it.
123 /// The alloca may just be accessed with memcpy, for example, which would
125 bool hasALoadOrStore : 1;
127 explicit AllocaInfo(AllocaInst *ai)
128 : AI(ai), isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false),
129 hasSubelementAccess(false), hasALoadOrStore(false) {}
132 /// SRThreshold - The maximum alloca size to considered for SROA.
133 unsigned SRThreshold;
135 /// StructMemberThreshold - The maximum number of members a struct can
136 /// contain to be considered for SROA.
137 unsigned StructMemberThreshold;
139 /// ArrayElementThreshold - The maximum number of elements an array can
140 /// have to be considered for SROA.
141 unsigned ArrayElementThreshold;
143 /// ScalarLoadThreshold - The maximum size in bits of scalars to load when
144 /// converting to scalar
145 unsigned ScalarLoadThreshold;
147 void MarkUnsafe(AllocaInfo &I, Instruction *User) {
149 DEBUG(dbgs() << " Transformation preventing inst: " << *User << '\n');
152 bool isSafeAllocaToScalarRepl(AllocaInst *AI);
154 void isSafeForScalarRepl(Instruction *I, uint64_t Offset, AllocaInfo &Info);
155 void isSafePHISelectUseForScalarRepl(Instruction *User, uint64_t Offset,
157 void isSafeGEP(GetElementPtrInst *GEPI, uint64_t &Offset, AllocaInfo &Info);
158 void isSafeMemAccess(uint64_t Offset, uint64_t MemSize,
159 Type *MemOpType, bool isStore, AllocaInfo &Info,
160 Instruction *TheAccess, bool AllowWholeAccess);
161 bool TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size,
162 const DataLayout &DL);
163 uint64_t FindElementAndOffset(Type *&T, uint64_t &Offset, Type *&IdxTy,
164 const DataLayout &DL);
166 void DoScalarReplacement(AllocaInst *AI,
167 std::vector<AllocaInst*> &WorkList);
168 void DeleteDeadInstructions();
170 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
171 SmallVectorImpl<AllocaInst *> &NewElts);
172 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
173 SmallVectorImpl<AllocaInst *> &NewElts);
174 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
175 SmallVectorImpl<AllocaInst *> &NewElts);
176 void RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI,
178 SmallVectorImpl<AllocaInst *> &NewElts);
179 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
181 SmallVectorImpl<AllocaInst *> &NewElts);
182 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
183 SmallVectorImpl<AllocaInst *> &NewElts);
184 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
185 SmallVectorImpl<AllocaInst *> &NewElts);
186 bool ShouldAttemptScalarRepl(AllocaInst *AI);
189 // SROA_DT - SROA that uses DominatorTree.
190 struct SROA_DT : public SROA {
193 SROA_DT(int T = -1, int ST = -1, int AT = -1, int SLT = -1) :
194 SROA(T, true, ID, ST, AT, SLT) {
195 initializeSROA_DTPass(*PassRegistry::getPassRegistry());
198 // getAnalysisUsage - This pass does not require any passes, but we know it
199 // will not alter the CFG, so say so.
200 void getAnalysisUsage(AnalysisUsage &AU) const override {
201 AU.addRequired<AssumptionCacheTracker>();
202 AU.addRequired<DominatorTreeWrapperPass>();
203 AU.setPreservesCFG();
207 // SROA_SSAUp - SROA that uses SSAUpdater.
208 struct SROA_SSAUp : public SROA {
211 SROA_SSAUp(int T = -1, int ST = -1, int AT = -1, int SLT = -1) :
212 SROA(T, false, ID, ST, AT, SLT) {
213 initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry());
216 // getAnalysisUsage - This pass does not require any passes, but we know it
217 // will not alter the CFG, so say so.
218 void getAnalysisUsage(AnalysisUsage &AU) const override {
219 AU.addRequired<AssumptionCacheTracker>();
220 AU.setPreservesCFG();
226 char SROA_DT::ID = 0;
227 char SROA_SSAUp::ID = 0;
229 INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl",
230 "Scalar Replacement of Aggregates (DT)", false, false)
231 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
232 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
233 INITIALIZE_PASS_END(SROA_DT, "scalarrepl",
234 "Scalar Replacement of Aggregates (DT)", false, false)
236 INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa",
237 "Scalar Replacement of Aggregates (SSAUp)", false, false)
238 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
239 INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa",
240 "Scalar Replacement of Aggregates (SSAUp)", false, false)
242 // Public interface to the ScalarReplAggregates pass
243 FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold,
245 int StructMemberThreshold,
246 int ArrayElementThreshold,
247 int ScalarLoadThreshold) {
249 return new SROA_DT(Threshold, StructMemberThreshold, ArrayElementThreshold,
250 ScalarLoadThreshold);
251 return new SROA_SSAUp(Threshold, StructMemberThreshold,
252 ArrayElementThreshold, ScalarLoadThreshold);
256 //===----------------------------------------------------------------------===//
257 // Convert To Scalar Optimization.
258 //===----------------------------------------------------------------------===//
261 /// ConvertToScalarInfo - This class implements the "Convert To Scalar"
262 /// optimization, which scans the uses of an alloca and determines if it can
263 /// rewrite it in terms of a single new alloca that can be mem2reg'd.
264 class ConvertToScalarInfo {
265 /// AllocaSize - The size of the alloca being considered in bytes.
267 const DataLayout &DL;
268 unsigned ScalarLoadThreshold;
270 /// IsNotTrivial - This is set to true if there is some access to the object
271 /// which means that mem2reg can't promote it.
274 /// ScalarKind - Tracks the kind of alloca being considered for promotion,
275 /// computed based on the uses of the alloca rather than the LLVM type system.
279 // Accesses via GEPs that are consistent with element access of a vector
280 // type. This will not be converted into a vector unless there is a later
281 // access using an actual vector type.
284 // Accesses via vector operations and GEPs that are consistent with the
285 // layout of a vector type.
288 // An integer bag-of-bits with bitwise operations for insertion and
289 // extraction. Any combination of types can be converted into this kind
294 /// VectorTy - This tracks the type that we should promote the vector to if
295 /// it is possible to turn it into a vector. This starts out null, and if it
296 /// isn't possible to turn into a vector type, it gets set to VoidTy.
297 VectorType *VectorTy;
299 /// HadNonMemTransferAccess - True if there is at least one access to the
300 /// alloca that is not a MemTransferInst. We don't want to turn structs into
301 /// large integers unless there is some potential for optimization.
302 bool HadNonMemTransferAccess;
304 /// HadDynamicAccess - True if some element of this alloca was dynamic.
305 /// We don't yet have support for turning a dynamic access into a large
307 bool HadDynamicAccess;
310 explicit ConvertToScalarInfo(unsigned Size, const DataLayout &DL,
312 : AllocaSize(Size), DL(DL), ScalarLoadThreshold(SLT), IsNotTrivial(false),
313 ScalarKind(Unknown), VectorTy(nullptr), HadNonMemTransferAccess(false),
314 HadDynamicAccess(false) { }
316 AllocaInst *TryConvert(AllocaInst *AI);
319 bool CanConvertToScalar(Value *V, uint64_t Offset, Value* NonConstantIdx);
320 void MergeInTypeForLoadOrStore(Type *In, uint64_t Offset);
321 bool MergeInVectorType(VectorType *VInTy, uint64_t Offset);
322 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset,
323 Value *NonConstantIdx);
325 Value *ConvertScalar_ExtractValue(Value *NV, Type *ToType,
326 uint64_t Offset, Value* NonConstantIdx,
327 IRBuilder<> &Builder);
328 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal,
329 uint64_t Offset, Value* NonConstantIdx,
330 IRBuilder<> &Builder);
332 } // end anonymous namespace.
335 /// TryConvert - Analyze the specified alloca, and if it is safe to do so,
336 /// rewrite it to be a new alloca which is mem2reg'able. This returns the new
337 /// alloca if possible or null if not.
338 AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) {
339 // If we can't convert this scalar, or if mem2reg can trivially do it, bail
341 if (!CanConvertToScalar(AI, 0, nullptr) || !IsNotTrivial)
344 // If an alloca has only memset / memcpy uses, it may still have an Unknown
345 // ScalarKind. Treat it as an Integer below.
346 if (ScalarKind == Unknown)
347 ScalarKind = Integer;
349 if (ScalarKind == Vector && VectorTy->getBitWidth() != AllocaSize * 8)
350 ScalarKind = Integer;
352 // If we were able to find a vector type that can handle this with
353 // insert/extract elements, and if there was at least one use that had
354 // a vector type, promote this to a vector. We don't want to promote
355 // random stuff that doesn't use vectors (e.g. <9 x double>) because then
356 // we just get a lot of insert/extracts. If at least one vector is
357 // involved, then we probably really do have a union of vector/array.
359 if (ScalarKind == Vector) {
360 assert(VectorTy && "Missing type for vector scalar.");
361 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = "
362 << *VectorTy << '\n');
363 NewTy = VectorTy; // Use the vector type.
365 unsigned BitWidth = AllocaSize * 8;
367 // Do not convert to scalar integer if the alloca size exceeds the
368 // scalar load threshold.
369 if (BitWidth > ScalarLoadThreshold)
372 if ((ScalarKind == ImplicitVector || ScalarKind == Integer) &&
373 !HadNonMemTransferAccess && !DL.fitsInLegalInteger(BitWidth))
375 // Dynamic accesses on integers aren't yet supported. They need us to shift
376 // by a dynamic amount which could be difficult to work out as we might not
377 // know whether to use a left or right shift.
378 if (ScalarKind == Integer && HadDynamicAccess)
381 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n");
382 // Create and insert the integer alloca.
383 NewTy = IntegerType::get(AI->getContext(), BitWidth);
385 AllocaInst *NewAI = new AllocaInst(NewTy, nullptr, "",
386 AI->getParent()->begin());
387 ConvertUsesToScalar(AI, NewAI, 0, nullptr);
391 /// MergeInTypeForLoadOrStore - Add the 'In' type to the accumulated vector type
392 /// (VectorTy) so far at the offset specified by Offset (which is specified in
395 /// There are two cases we handle here:
396 /// 1) A union of vector types of the same size and potentially its elements.
397 /// Here we turn element accesses into insert/extract element operations.
398 /// This promotes a <4 x float> with a store of float to the third element
399 /// into a <4 x float> that uses insert element.
400 /// 2) A fully general blob of memory, which we turn into some (potentially
401 /// large) integer type with extract and insert operations where the loads
402 /// and stores would mutate the memory. We mark this by setting VectorTy
404 void ConvertToScalarInfo::MergeInTypeForLoadOrStore(Type *In,
406 // If we already decided to turn this into a blob of integer memory, there is
407 // nothing to be done.
408 if (ScalarKind == Integer)
411 // If this could be contributing to a vector, analyze it.
413 // If the In type is a vector that is the same size as the alloca, see if it
414 // matches the existing VecTy.
415 if (VectorType *VInTy = dyn_cast<VectorType>(In)) {
416 if (MergeInVectorType(VInTy, Offset))
418 } else if (In->isFloatTy() || In->isDoubleTy() ||
419 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 &&
420 isPowerOf2_32(In->getPrimitiveSizeInBits()))) {
421 // Full width accesses can be ignored, because they can always be turned
423 unsigned EltSize = In->getPrimitiveSizeInBits()/8;
424 if (EltSize == AllocaSize)
427 // If we're accessing something that could be an element of a vector, see
428 // if the implied vector agrees with what we already have and if Offset is
429 // compatible with it.
430 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 &&
431 (!VectorTy || EltSize == VectorTy->getElementType()
432 ->getPrimitiveSizeInBits()/8)) {
434 ScalarKind = ImplicitVector;
435 VectorTy = VectorType::get(In, AllocaSize/EltSize);
441 // Otherwise, we have a case that we can't handle with an optimized vector
442 // form. We can still turn this into a large integer.
443 ScalarKind = Integer;
446 /// MergeInVectorType - Handles the vector case of MergeInTypeForLoadOrStore,
447 /// returning true if the type was successfully merged and false otherwise.
448 bool ConvertToScalarInfo::MergeInVectorType(VectorType *VInTy,
450 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) {
451 // If we're storing/loading a vector of the right size, allow it as a
452 // vector. If this the first vector we see, remember the type so that
453 // we know the element size. If this is a subsequent access, ignore it
454 // even if it is a differing type but the same size. Worst case we can
455 // bitcast the resultant vectors.
465 /// CanConvertToScalar - V is a pointer. If we can convert the pointee and all
466 /// its accesses to a single vector type, return true and set VecTy to
467 /// the new type. If we could convert the alloca into a single promotable
468 /// integer, return true but set VecTy to VoidTy. Further, if the use is not a
469 /// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset
470 /// is the current offset from the base of the alloca being analyzed.
472 /// If we see at least one access to the value that is as a vector type, set the
474 bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset,
475 Value* NonConstantIdx) {
476 for (User *U : V->users()) {
477 Instruction *UI = cast<Instruction>(U);
479 if (LoadInst *LI = dyn_cast<LoadInst>(UI)) {
480 // Don't break volatile loads.
483 // Don't touch MMX operations.
484 if (LI->getType()->isX86_MMXTy())
486 HadNonMemTransferAccess = true;
487 MergeInTypeForLoadOrStore(LI->getType(), Offset);
491 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
492 // Storing the pointer, not into the value?
493 if (SI->getOperand(0) == V || !SI->isSimple()) return false;
494 // Don't touch MMX operations.
495 if (SI->getOperand(0)->getType()->isX86_MMXTy())
497 HadNonMemTransferAccess = true;
498 MergeInTypeForLoadOrStore(SI->getOperand(0)->getType(), Offset);
502 if (BitCastInst *BCI = dyn_cast<BitCastInst>(UI)) {
503 if (!onlyUsedByLifetimeMarkers(BCI))
504 IsNotTrivial = true; // Can't be mem2reg'd.
505 if (!CanConvertToScalar(BCI, Offset, NonConstantIdx))
510 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(UI)) {
511 // If this is a GEP with a variable indices, we can't handle it.
512 PointerType* PtrTy = dyn_cast<PointerType>(GEP->getPointerOperandType());
516 // Compute the offset that this GEP adds to the pointer.
517 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
518 Value *GEPNonConstantIdx = nullptr;
519 if (!GEP->hasAllConstantIndices()) {
520 if (!isa<VectorType>(PtrTy->getElementType()))
524 GEPNonConstantIdx = Indices.pop_back_val();
525 if (!GEPNonConstantIdx->getType()->isIntegerTy(32))
527 HadDynamicAccess = true;
529 GEPNonConstantIdx = NonConstantIdx;
530 uint64_t GEPOffset = DL.getIndexedOffset(PtrTy,
532 // See if all uses can be converted.
533 if (!CanConvertToScalar(GEP, Offset+GEPOffset, GEPNonConstantIdx))
535 IsNotTrivial = true; // Can't be mem2reg'd.
536 HadNonMemTransferAccess = true;
540 // If this is a constant sized memset of a constant value (e.g. 0) we can
542 if (MemSetInst *MSI = dyn_cast<MemSetInst>(UI)) {
543 // Store to dynamic index.
546 // Store of constant value.
547 if (!isa<ConstantInt>(MSI->getValue()))
550 // Store of constant size.
551 ConstantInt *Len = dyn_cast<ConstantInt>(MSI->getLength());
555 // If the size differs from the alloca, we can only convert the alloca to
556 // an integer bag-of-bits.
557 // FIXME: This should handle all of the cases that are currently accepted
558 // as vector element insertions.
559 if (Len->getZExtValue() != AllocaSize || Offset != 0)
560 ScalarKind = Integer;
562 IsNotTrivial = true; // Can't be mem2reg'd.
563 HadNonMemTransferAccess = true;
567 // If this is a memcpy or memmove into or out of the whole allocation, we
568 // can handle it like a load or store of the scalar type.
569 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(UI)) {
570 // Store to dynamic index.
573 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength());
574 if (!Len || Len->getZExtValue() != AllocaSize || Offset != 0)
577 IsNotTrivial = true; // Can't be mem2reg'd.
581 // If this is a lifetime intrinsic, we can handle it.
582 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(UI)) {
583 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
584 II->getIntrinsicID() == Intrinsic::lifetime_end) {
589 // Otherwise, we cannot handle this!
596 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
597 /// directly. This happens when we are converting an "integer union" to a
598 /// single integer scalar, or when we are converting a "vector union" to a
599 /// vector with insert/extractelement instructions.
601 /// Offset is an offset from the original alloca, in bits that need to be
602 /// shifted to the right. By the end of this, there should be no uses of Ptr.
603 void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI,
605 Value* NonConstantIdx) {
606 while (!Ptr->use_empty()) {
607 Instruction *User = cast<Instruction>(Ptr->user_back());
609 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
610 ConvertUsesToScalar(CI, NewAI, Offset, NonConstantIdx);
611 CI->eraseFromParent();
615 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
616 // Compute the offset that this GEP adds to the pointer.
617 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
618 Value* GEPNonConstantIdx = nullptr;
619 if (!GEP->hasAllConstantIndices()) {
620 assert(!NonConstantIdx &&
621 "Dynamic GEP reading from dynamic GEP unsupported");
622 GEPNonConstantIdx = Indices.pop_back_val();
624 GEPNonConstantIdx = NonConstantIdx;
625 uint64_t GEPOffset = DL.getIndexedOffset(GEP->getPointerOperandType(),
627 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8, GEPNonConstantIdx);
628 GEP->eraseFromParent();
632 IRBuilder<> Builder(User);
634 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
635 // The load is a bit extract from NewAI shifted right by Offset bits.
636 Value *LoadedVal = Builder.CreateLoad(NewAI);
638 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset,
639 NonConstantIdx, Builder);
640 LI->replaceAllUsesWith(NewLoadVal);
641 LI->eraseFromParent();
645 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
646 assert(SI->getOperand(0) != Ptr && "Consistency error!");
647 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
648 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset,
649 NonConstantIdx, Builder);
650 Builder.CreateStore(New, NewAI);
651 SI->eraseFromParent();
653 // If the load we just inserted is now dead, then the inserted store
654 // overwrote the entire thing.
655 if (Old->use_empty())
656 Old->eraseFromParent();
660 // If this is a constant sized memset of a constant value (e.g. 0) we can
661 // transform it into a store of the expanded constant value.
662 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
663 assert(MSI->getRawDest() == Ptr && "Consistency error!");
664 assert(!NonConstantIdx && "Cannot replace dynamic memset with insert");
665 int64_t SNumBytes = cast<ConstantInt>(MSI->getLength())->getSExtValue();
666 if (SNumBytes > 0 && (SNumBytes >> 32) == 0) {
667 unsigned NumBytes = static_cast<unsigned>(SNumBytes);
668 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue();
670 // Compute the value replicated the right number of times.
671 APInt APVal(NumBytes*8, Val);
673 // Splat the value if non-zero.
675 for (unsigned i = 1; i != NumBytes; ++i)
678 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
679 Value *New = ConvertScalar_InsertValue(
680 ConstantInt::get(User->getContext(), APVal),
681 Old, Offset, nullptr, Builder);
682 Builder.CreateStore(New, NewAI);
684 // If the load we just inserted is now dead, then the memset overwrote
686 if (Old->use_empty())
687 Old->eraseFromParent();
689 MSI->eraseFromParent();
693 // If this is a memcpy or memmove into or out of the whole allocation, we
694 // can handle it like a load or store of the scalar type.
695 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
696 assert(Offset == 0 && "must be store to start of alloca");
697 assert(!NonConstantIdx && "Cannot replace dynamic transfer with insert");
699 // If the source and destination are both to the same alloca, then this is
700 // a noop copy-to-self, just delete it. Otherwise, emit a load and store
702 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, DL, 0));
704 if (GetUnderlyingObject(MTI->getSource(), DL, 0) != OrigAI) {
705 // Dest must be OrigAI, change this to be a load from the original
706 // pointer (bitcasted), then a store to our new alloca.
707 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?");
708 Value *SrcPtr = MTI->getSource();
709 PointerType* SPTy = cast<PointerType>(SrcPtr->getType());
710 PointerType* AIPTy = cast<PointerType>(NewAI->getType());
711 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
712 AIPTy = PointerType::get(AIPTy->getElementType(),
713 SPTy->getAddressSpace());
715 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy);
717 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval");
718 SrcVal->setAlignment(MTI->getAlignment());
719 Builder.CreateStore(SrcVal, NewAI);
720 } else if (GetUnderlyingObject(MTI->getDest(), DL, 0) != OrigAI) {
721 // Src must be OrigAI, change this to be a load from NewAI then a store
722 // through the original dest pointer (bitcasted).
723 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?");
724 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval");
726 PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType());
727 PointerType* AIPTy = cast<PointerType>(NewAI->getType());
728 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
729 AIPTy = PointerType::get(AIPTy->getElementType(),
730 DPTy->getAddressSpace());
732 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy);
734 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr);
735 NewStore->setAlignment(MTI->getAlignment());
737 // Noop transfer. Src == Dst
740 MTI->eraseFromParent();
744 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
745 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
746 II->getIntrinsicID() == Intrinsic::lifetime_end) {
747 // There's no need to preserve these, as the resulting alloca will be
748 // converted to a register anyways.
749 II->eraseFromParent();
754 llvm_unreachable("Unsupported operation!");
758 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer
759 /// or vector value FromVal, extracting the bits from the offset specified by
760 /// Offset. This returns the value, which is of type ToType.
762 /// This happens when we are converting an "integer union" to a single
763 /// integer scalar, or when we are converting a "vector union" to a vector with
764 /// insert/extractelement instructions.
766 /// Offset is an offset from the original alloca, in bits that need to be
767 /// shifted to the right.
768 Value *ConvertToScalarInfo::
769 ConvertScalar_ExtractValue(Value *FromVal, Type *ToType,
770 uint64_t Offset, Value* NonConstantIdx,
771 IRBuilder<> &Builder) {
772 // If the load is of the whole new alloca, no conversion is needed.
773 Type *FromType = FromVal->getType();
774 if (FromType == ToType && Offset == 0)
777 // If the result alloca is a vector type, this is either an element
778 // access or a bitcast to another vector type of the same size.
779 if (VectorType *VTy = dyn_cast<VectorType>(FromType)) {
780 unsigned FromTypeSize = DL.getTypeAllocSize(FromType);
781 unsigned ToTypeSize = DL.getTypeAllocSize(ToType);
782 if (FromTypeSize == ToTypeSize)
783 return Builder.CreateBitCast(FromVal, ToType);
785 // Otherwise it must be an element access.
788 unsigned EltSize = DL.getTypeAllocSizeInBits(VTy->getElementType());
789 Elt = Offset/EltSize;
790 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
792 // Return the element extracted out of it.
794 if (NonConstantIdx) {
796 Idx = Builder.CreateAdd(NonConstantIdx,
797 Builder.getInt32(Elt),
800 Idx = NonConstantIdx;
802 Idx = Builder.getInt32(Elt);
803 Value *V = Builder.CreateExtractElement(FromVal, Idx);
804 if (V->getType() != ToType)
805 V = Builder.CreateBitCast(V, ToType);
809 // If ToType is a first class aggregate, extract out each of the pieces and
810 // use insertvalue's to form the FCA.
811 if (StructType *ST = dyn_cast<StructType>(ToType)) {
812 assert(!NonConstantIdx &&
813 "Dynamic indexing into struct types not supported");
814 const StructLayout &Layout = *DL.getStructLayout(ST);
815 Value *Res = UndefValue::get(ST);
816 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
817 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i),
818 Offset+Layout.getElementOffsetInBits(i),
820 Res = Builder.CreateInsertValue(Res, Elt, i);
825 if (ArrayType *AT = dyn_cast<ArrayType>(ToType)) {
826 assert(!NonConstantIdx &&
827 "Dynamic indexing into array types not supported");
828 uint64_t EltSize = DL.getTypeAllocSizeInBits(AT->getElementType());
829 Value *Res = UndefValue::get(AT);
830 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
831 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(),
832 Offset+i*EltSize, nullptr,
834 Res = Builder.CreateInsertValue(Res, Elt, i);
839 // Otherwise, this must be a union that was converted to an integer value.
840 IntegerType *NTy = cast<IntegerType>(FromVal->getType());
842 // If this is a big-endian system and the load is narrower than the
843 // full alloca type, we need to do a shift to get the right bits.
845 if (DL.isBigEndian()) {
846 // On big-endian machines, the lowest bit is stored at the bit offset
847 // from the pointer given by getTypeStoreSizeInBits. This matters for
848 // integers with a bitwidth that is not a multiple of 8.
849 ShAmt = DL.getTypeStoreSizeInBits(NTy) -
850 DL.getTypeStoreSizeInBits(ToType) - Offset;
855 // Note: we support negative bitwidths (with shl) which are not defined.
856 // We do this to support (f.e.) loads off the end of a structure where
857 // only some bits are used.
858 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth())
859 FromVal = Builder.CreateLShr(FromVal,
860 ConstantInt::get(FromVal->getType(), ShAmt));
861 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
862 FromVal = Builder.CreateShl(FromVal,
863 ConstantInt::get(FromVal->getType(), -ShAmt));
865 // Finally, unconditionally truncate the integer to the right width.
866 unsigned LIBitWidth = DL.getTypeSizeInBits(ToType);
867 if (LIBitWidth < NTy->getBitWidth())
869 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(),
871 else if (LIBitWidth > NTy->getBitWidth())
873 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(),
876 // If the result is an integer, this is a trunc or bitcast.
877 if (ToType->isIntegerTy()) {
879 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) {
880 // Just do a bitcast, we know the sizes match up.
881 FromVal = Builder.CreateBitCast(FromVal, ToType);
883 // Otherwise must be a pointer.
884 FromVal = Builder.CreateIntToPtr(FromVal, ToType);
886 assert(FromVal->getType() == ToType && "Didn't convert right?");
890 /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer
891 /// or vector value "Old" at the offset specified by Offset.
893 /// This happens when we are converting an "integer union" to a
894 /// single integer scalar, or when we are converting a "vector union" to a
895 /// vector with insert/extractelement instructions.
897 /// Offset is an offset from the original alloca, in bits that need to be
898 /// shifted to the right.
900 /// NonConstantIdx is an index value if there was a GEP with a non-constant
901 /// index value. If this is 0 then all GEPs used to find this insert address
903 Value *ConvertToScalarInfo::
904 ConvertScalar_InsertValue(Value *SV, Value *Old,
905 uint64_t Offset, Value* NonConstantIdx,
906 IRBuilder<> &Builder) {
907 // Convert the stored type to the actual type, shift it left to insert
908 // then 'or' into place.
909 Type *AllocaType = Old->getType();
910 LLVMContext &Context = Old->getContext();
912 if (VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
913 uint64_t VecSize = DL.getTypeAllocSizeInBits(VTy);
914 uint64_t ValSize = DL.getTypeAllocSizeInBits(SV->getType());
916 // Changing the whole vector with memset or with an access of a different
918 if (ValSize == VecSize)
919 return Builder.CreateBitCast(SV, AllocaType);
921 // Must be an element insertion.
922 Type *EltTy = VTy->getElementType();
923 if (SV->getType() != EltTy)
924 SV = Builder.CreateBitCast(SV, EltTy);
925 uint64_t EltSize = DL.getTypeAllocSizeInBits(EltTy);
926 unsigned Elt = Offset/EltSize;
928 if (NonConstantIdx) {
930 Idx = Builder.CreateAdd(NonConstantIdx,
931 Builder.getInt32(Elt),
934 Idx = NonConstantIdx;
936 Idx = Builder.getInt32(Elt);
937 return Builder.CreateInsertElement(Old, SV, Idx);
940 // If SV is a first-class aggregate value, insert each value recursively.
941 if (StructType *ST = dyn_cast<StructType>(SV->getType())) {
942 assert(!NonConstantIdx &&
943 "Dynamic indexing into struct types not supported");
944 const StructLayout &Layout = *DL.getStructLayout(ST);
945 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
946 Value *Elt = Builder.CreateExtractValue(SV, i);
947 Old = ConvertScalar_InsertValue(Elt, Old,
948 Offset+Layout.getElementOffsetInBits(i),
954 if (ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) {
955 assert(!NonConstantIdx &&
956 "Dynamic indexing into array types not supported");
957 uint64_t EltSize = DL.getTypeAllocSizeInBits(AT->getElementType());
958 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
959 Value *Elt = Builder.CreateExtractValue(SV, i);
960 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, nullptr,
966 // If SV is a float, convert it to the appropriate integer type.
967 // If it is a pointer, do the same.
968 unsigned SrcWidth = DL.getTypeSizeInBits(SV->getType());
969 unsigned DestWidth = DL.getTypeSizeInBits(AllocaType);
970 unsigned SrcStoreWidth = DL.getTypeStoreSizeInBits(SV->getType());
971 unsigned DestStoreWidth = DL.getTypeStoreSizeInBits(AllocaType);
972 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy())
973 SV = Builder.CreateBitCast(SV, IntegerType::get(SV->getContext(),SrcWidth));
974 else if (SV->getType()->isPointerTy())
975 SV = Builder.CreatePtrToInt(SV, DL.getIntPtrType(SV->getType()));
977 // Zero extend or truncate the value if needed.
978 if (SV->getType() != AllocaType) {
979 if (SV->getType()->getPrimitiveSizeInBits() <
980 AllocaType->getPrimitiveSizeInBits())
981 SV = Builder.CreateZExt(SV, AllocaType);
983 // Truncation may be needed if storing more than the alloca can hold
984 // (undefined behavior).
985 SV = Builder.CreateTrunc(SV, AllocaType);
986 SrcWidth = DestWidth;
987 SrcStoreWidth = DestStoreWidth;
991 // If this is a big-endian system and the store is narrower than the
992 // full alloca type, we need to do a shift to get the right bits.
994 if (DL.isBigEndian()) {
995 // On big-endian machines, the lowest bit is stored at the bit offset
996 // from the pointer given by getTypeStoreSizeInBits. This matters for
997 // integers with a bitwidth that is not a multiple of 8.
998 ShAmt = DestStoreWidth - SrcStoreWidth - Offset;
1003 // Note: we support negative bitwidths (with shr) which are not defined.
1004 // We do this to support (f.e.) stores off the end of a structure where
1005 // only some bits in the structure are set.
1006 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
1007 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
1008 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), ShAmt));
1010 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
1011 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), -ShAmt));
1012 Mask = Mask.lshr(-ShAmt);
1015 // Mask out the bits we are about to insert from the old value, and or
1017 if (SrcWidth != DestWidth) {
1018 assert(DestWidth > SrcWidth);
1019 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask");
1020 SV = Builder.CreateOr(Old, SV, "ins");
1026 //===----------------------------------------------------------------------===//
1028 //===----------------------------------------------------------------------===//
1031 bool SROA::runOnFunction(Function &F) {
1032 if (skipOptnoneFunction(F))
1035 bool Changed = performPromotion(F);
1038 bool LocalChange = performScalarRepl(F);
1039 if (!LocalChange) break; // No need to repromote if no scalarrepl
1041 LocalChange = performPromotion(F);
1042 if (!LocalChange) break; // No need to re-scalarrepl if no promotion
1049 class AllocaPromoter : public LoadAndStorePromoter {
1052 SmallVector<DbgDeclareInst *, 4> DDIs;
1053 SmallVector<DbgValueInst *, 4> DVIs;
1055 AllocaPromoter(ArrayRef<Instruction*> Insts, SSAUpdater &S,
1057 : LoadAndStorePromoter(Insts, S), AI(nullptr), DIB(DB) {}
1059 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) {
1060 // Remember which alloca we're promoting (for isInstInList).
1062 if (auto *L = LocalAsMetadata::getIfExists(AI)) {
1063 if (auto *DINode = MetadataAsValue::getIfExists(AI->getContext(), L)) {
1064 for (User *U : DINode->users())
1065 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U))
1066 DDIs.push_back(DDI);
1067 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(U))
1068 DVIs.push_back(DVI);
1072 LoadAndStorePromoter::run(Insts);
1073 AI->eraseFromParent();
1074 for (SmallVectorImpl<DbgDeclareInst *>::iterator I = DDIs.begin(),
1075 E = DDIs.end(); I != E; ++I) {
1076 DbgDeclareInst *DDI = *I;
1077 DDI->eraseFromParent();
1079 for (SmallVectorImpl<DbgValueInst *>::iterator I = DVIs.begin(),
1080 E = DVIs.end(); I != E; ++I) {
1081 DbgValueInst *DVI = *I;
1082 DVI->eraseFromParent();
1086 bool isInstInList(Instruction *I,
1087 const SmallVectorImpl<Instruction*> &Insts) const override {
1088 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1089 return LI->getOperand(0) == AI;
1090 return cast<StoreInst>(I)->getPointerOperand() == AI;
1093 void updateDebugInfo(Instruction *Inst) const override {
1094 for (SmallVectorImpl<DbgDeclareInst *>::const_iterator I = DDIs.begin(),
1095 E = DDIs.end(); I != E; ++I) {
1096 DbgDeclareInst *DDI = *I;
1097 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
1098 ConvertDebugDeclareToDebugValue(DDI, SI, *DIB);
1099 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
1100 ConvertDebugDeclareToDebugValue(DDI, LI, *DIB);
1102 for (SmallVectorImpl<DbgValueInst *>::const_iterator I = DVIs.begin(),
1103 E = DVIs.end(); I != E; ++I) {
1104 DbgValueInst *DVI = *I;
1105 Value *Arg = nullptr;
1106 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1107 // If an argument is zero extended then use argument directly. The ZExt
1108 // may be zapped by an optimization pass in future.
1109 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
1110 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
1111 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
1112 Arg = dyn_cast<Argument>(SExt->getOperand(0));
1114 Arg = SI->getOperand(0);
1115 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
1116 Arg = LI->getOperand(0);
1120 DIB->insertDbgValueIntrinsic(Arg, 0, DVI->getVariable(),
1121 DVI->getExpression(), DVI->getDebugLoc(),
1126 } // end anon namespace
1128 /// isSafeSelectToSpeculate - Select instructions that use an alloca and are
1129 /// subsequently loaded can be rewritten to load both input pointers and then
1130 /// select between the result, allowing the load of the alloca to be promoted.
1132 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1133 /// %V = load i32* %P2
1135 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1136 /// %V2 = load i32* %Other
1137 /// %V = select i1 %cond, i32 %V1, i32 %V2
1139 /// We can do this to a select if its only uses are loads and if the operand to
1140 /// the select can be loaded unconditionally.
1141 static bool isSafeSelectToSpeculate(SelectInst *SI) {
1142 const DataLayout &DL = SI->getModule()->getDataLayout();
1143 bool TDerefable = isDereferenceablePointer(SI->getTrueValue(), DL);
1144 bool FDerefable = isDereferenceablePointer(SI->getFalseValue(), DL);
1146 for (User *U : SI->users()) {
1147 LoadInst *LI = dyn_cast<LoadInst>(U);
1148 if (!LI || !LI->isSimple()) return false;
1150 // Both operands to the select need to be dereferencable, either absolutely
1151 // (e.g. allocas) or at this point because we can see other accesses to it.
1153 !isSafeToLoadUnconditionally(SI->getTrueValue(), LI,
1154 LI->getAlignment()))
1157 !isSafeToLoadUnconditionally(SI->getFalseValue(), LI,
1158 LI->getAlignment()))
1165 /// isSafePHIToSpeculate - PHI instructions that use an alloca and are
1166 /// subsequently loaded can be rewritten to load both input pointers in the pred
1167 /// blocks and then PHI the results, allowing the load of the alloca to be
1170 /// %P2 = phi [i32* %Alloca, i32* %Other]
1171 /// %V = load i32* %P2
1173 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1175 /// %V2 = load i32* %Other
1177 /// %V = phi [i32 %V1, i32 %V2]
1179 /// We can do this to a select if its only uses are loads and if the operand to
1180 /// the select can be loaded unconditionally.
1181 static bool isSafePHIToSpeculate(PHINode *PN) {
1182 // For now, we can only do this promotion if the load is in the same block as
1183 // the PHI, and if there are no stores between the phi and load.
1184 // TODO: Allow recursive phi users.
1185 // TODO: Allow stores.
1186 BasicBlock *BB = PN->getParent();
1187 unsigned MaxAlign = 0;
1188 for (User *U : PN->users()) {
1189 LoadInst *LI = dyn_cast<LoadInst>(U);
1190 if (!LI || !LI->isSimple()) return false;
1192 // For now we only allow loads in the same block as the PHI. This is a
1193 // common case that happens when instcombine merges two loads through a PHI.
1194 if (LI->getParent() != BB) return false;
1196 // Ensure that there are no instructions between the PHI and the load that
1198 for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI)
1199 if (BBI->mayWriteToMemory())
1202 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1205 const DataLayout &DL = PN->getModule()->getDataLayout();
1207 // Okay, we know that we have one or more loads in the same block as the PHI.
1208 // We can transform this if it is safe to push the loads into the predecessor
1209 // blocks. The only thing to watch out for is that we can't put a possibly
1210 // trapping load in the predecessor if it is a critical edge.
1211 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1212 BasicBlock *Pred = PN->getIncomingBlock(i);
1213 Value *InVal = PN->getIncomingValue(i);
1215 // If the terminator of the predecessor has side-effects (an invoke),
1216 // there is no safe place to put a load in the predecessor.
1217 if (Pred->getTerminator()->mayHaveSideEffects())
1220 // If the value is produced by the terminator of the predecessor
1221 // (an invoke), there is no valid place to put a load in the predecessor.
1222 if (Pred->getTerminator() == InVal)
1225 // If the predecessor has a single successor, then the edge isn't critical.
1226 if (Pred->getTerminator()->getNumSuccessors() == 1)
1229 // If this pointer is always safe to load, or if we can prove that there is
1230 // already a load in the block, then we can move the load to the pred block.
1231 if (isDereferenceablePointer(InVal, DL) ||
1232 isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign))
1242 /// tryToMakeAllocaBePromotable - This returns true if the alloca only has
1243 /// direct (non-volatile) loads and stores to it. If the alloca is close but
1244 /// not quite there, this will transform the code to allow promotion. As such,
1245 /// it is a non-pure predicate.
1246 static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const DataLayout &DL) {
1247 SetVector<Instruction*, SmallVector<Instruction*, 4>,
1248 SmallPtrSet<Instruction*, 4> > InstsToRewrite;
1249 for (User *U : AI->users()) {
1250 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
1251 if (!LI->isSimple())
1256 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1257 if (SI->getOperand(0) == AI || !SI->isSimple())
1258 return false; // Don't allow a store OF the AI, only INTO the AI.
1262 if (SelectInst *SI = dyn_cast<SelectInst>(U)) {
1263 // If the condition being selected on is a constant, fold the select, yes
1264 // this does (rarely) happen early on.
1265 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) {
1266 Value *Result = SI->getOperand(1+CI->isZero());
1267 SI->replaceAllUsesWith(Result);
1268 SI->eraseFromParent();
1270 // This is very rare and we just scrambled the use list of AI, start
1272 return tryToMakeAllocaBePromotable(AI, DL);
1275 // If it is safe to turn "load (select c, AI, ptr)" into a select of two
1276 // loads, then we can transform this by rewriting the select.
1277 if (!isSafeSelectToSpeculate(SI))
1280 InstsToRewrite.insert(SI);
1284 if (PHINode *PN = dyn_cast<PHINode>(U)) {
1285 if (PN->use_empty()) { // Dead PHIs can be stripped.
1286 InstsToRewrite.insert(PN);
1290 // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads
1291 // in the pred blocks, then we can transform this by rewriting the PHI.
1292 if (!isSafePHIToSpeculate(PN))
1295 InstsToRewrite.insert(PN);
1299 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
1300 if (onlyUsedByLifetimeMarkers(BCI)) {
1301 InstsToRewrite.insert(BCI);
1309 // If there are no instructions to rewrite, then all uses are load/stores and
1311 if (InstsToRewrite.empty())
1314 // If we have instructions that need to be rewritten for this to be promotable
1315 // take care of it now.
1316 for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) {
1317 if (BitCastInst *BCI = dyn_cast<BitCastInst>(InstsToRewrite[i])) {
1318 // This could only be a bitcast used by nothing but lifetime intrinsics.
1319 for (BitCastInst::user_iterator I = BCI->user_begin(), E = BCI->user_end();
1321 cast<Instruction>(*I++)->eraseFromParent();
1322 BCI->eraseFromParent();
1326 if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) {
1327 // Selects in InstsToRewrite only have load uses. Rewrite each as two
1328 // loads with a new select.
1329 while (!SI->use_empty()) {
1330 LoadInst *LI = cast<LoadInst>(SI->user_back());
1332 IRBuilder<> Builder(LI);
1333 LoadInst *TrueLoad =
1334 Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t");
1335 LoadInst *FalseLoad =
1336 Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".f");
1338 // Transfer alignment and AA info if present.
1339 TrueLoad->setAlignment(LI->getAlignment());
1340 FalseLoad->setAlignment(LI->getAlignment());
1343 LI->getAAMetadata(Tags);
1345 TrueLoad->setAAMetadata(Tags);
1346 FalseLoad->setAAMetadata(Tags);
1349 Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad);
1351 LI->replaceAllUsesWith(V);
1352 LI->eraseFromParent();
1355 // Now that all the loads are gone, the select is gone too.
1356 SI->eraseFromParent();
1360 // Otherwise, we have a PHI node which allows us to push the loads into the
1362 PHINode *PN = cast<PHINode>(InstsToRewrite[i]);
1363 if (PN->use_empty()) {
1364 PN->eraseFromParent();
1368 Type *LoadTy = cast<PointerType>(PN->getType())->getElementType();
1369 PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(),
1370 PN->getName()+".ld", PN);
1372 // Get the AA tags and alignment to use from one of the loads. It doesn't
1373 // matter which one we get and if any differ, it doesn't matter.
1374 LoadInst *SomeLoad = cast<LoadInst>(PN->user_back());
1377 SomeLoad->getAAMetadata(AATags);
1378 unsigned Align = SomeLoad->getAlignment();
1380 // Rewrite all loads of the PN to use the new PHI.
1381 while (!PN->use_empty()) {
1382 LoadInst *LI = cast<LoadInst>(PN->user_back());
1383 LI->replaceAllUsesWith(NewPN);
1384 LI->eraseFromParent();
1387 // Inject loads into all of the pred blocks. Keep track of which blocks we
1388 // insert them into in case we have multiple edges from the same block.
1389 DenseMap<BasicBlock*, LoadInst*> InsertedLoads;
1391 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1392 BasicBlock *Pred = PN->getIncomingBlock(i);
1393 LoadInst *&Load = InsertedLoads[Pred];
1395 Load = new LoadInst(PN->getIncomingValue(i),
1396 PN->getName() + "." + Pred->getName(),
1397 Pred->getTerminator());
1398 Load->setAlignment(Align);
1399 if (AATags) Load->setAAMetadata(AATags);
1402 NewPN->addIncoming(Load, Pred);
1405 PN->eraseFromParent();
1412 bool SROA::performPromotion(Function &F) {
1413 std::vector<AllocaInst*> Allocas;
1414 const DataLayout &DL = F.getParent()->getDataLayout();
1415 DominatorTree *DT = nullptr;
1417 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1418 AssumptionCache &AC =
1419 getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1421 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
1422 DIBuilder DIB(*F.getParent(), /*AllowUnresolved*/ false);
1423 bool Changed = false;
1424 SmallVector<Instruction*, 64> Insts;
1428 // Find allocas that are safe to promote, by looking at all instructions in
1430 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
1431 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
1432 if (tryToMakeAllocaBePromotable(AI, DL))
1433 Allocas.push_back(AI);
1435 if (Allocas.empty()) break;
1438 PromoteMemToReg(Allocas, *DT, nullptr, &AC);
1441 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
1442 AllocaInst *AI = Allocas[i];
1444 // Build list of instructions to promote.
1445 for (User *U : AI->users())
1446 Insts.push_back(cast<Instruction>(U));
1447 AllocaPromoter(Insts, SSA, &DIB).run(AI, Insts);
1451 NumPromoted += Allocas.size();
1459 /// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for
1460 /// SROA. It must be a struct or array type with a small number of elements.
1461 bool SROA::ShouldAttemptScalarRepl(AllocaInst *AI) {
1462 Type *T = AI->getAllocatedType();
1463 // Do not promote any struct that has too many members.
1464 if (StructType *ST = dyn_cast<StructType>(T))
1465 return ST->getNumElements() <= StructMemberThreshold;
1466 // Do not promote any array that has too many elements.
1467 if (ArrayType *AT = dyn_cast<ArrayType>(T))
1468 return AT->getNumElements() <= ArrayElementThreshold;
1472 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
1473 // which runs on all of the alloca instructions in the entry block, removing
1474 // them if they are only used by getelementptr instructions.
1476 bool SROA::performScalarRepl(Function &F) {
1477 std::vector<AllocaInst*> WorkList;
1478 const DataLayout &DL = F.getParent()->getDataLayout();
1480 // Scan the entry basic block, adding allocas to the worklist.
1481 BasicBlock &BB = F.getEntryBlock();
1482 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
1483 if (AllocaInst *A = dyn_cast<AllocaInst>(I))
1484 WorkList.push_back(A);
1486 // Process the worklist
1487 bool Changed = false;
1488 while (!WorkList.empty()) {
1489 AllocaInst *AI = WorkList.back();
1490 WorkList.pop_back();
1492 // Handle dead allocas trivially. These can be formed by SROA'ing arrays
1493 // with unused elements.
1494 if (AI->use_empty()) {
1495 AI->eraseFromParent();
1500 // If this alloca is impossible for us to promote, reject it early.
1501 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized())
1504 // Check to see if we can perform the core SROA transformation. We cannot
1505 // transform the allocation instruction if it is an array allocation
1506 // (allocations OF arrays are ok though), and an allocation of a scalar
1507 // value cannot be decomposed at all.
1508 uint64_t AllocaSize = DL.getTypeAllocSize(AI->getAllocatedType());
1510 // Do not promote [0 x %struct].
1511 if (AllocaSize == 0) continue;
1513 // Do not promote any struct whose size is too big.
1514 if (AllocaSize > SRThreshold) continue;
1516 // If the alloca looks like a good candidate for scalar replacement, and if
1517 // all its users can be transformed, then split up the aggregate into its
1518 // separate elements.
1519 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) {
1520 DoScalarReplacement(AI, WorkList);
1525 // If we can turn this aggregate value (potentially with casts) into a
1526 // simple scalar value that can be mem2reg'd into a register value.
1527 // IsNotTrivial tracks whether this is something that mem2reg could have
1528 // promoted itself. If so, we don't want to transform it needlessly. Note
1529 // that we can't just check based on the type: the alloca may be of an i32
1530 // but that has pointer arithmetic to set byte 3 of it or something.
1531 if (AllocaInst *NewAI =
1532 ConvertToScalarInfo((unsigned)AllocaSize, DL, ScalarLoadThreshold)
1534 NewAI->takeName(AI);
1535 AI->eraseFromParent();
1541 // Otherwise, couldn't process this alloca.
1547 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl
1548 /// predicate, do SROA now.
1549 void SROA::DoScalarReplacement(AllocaInst *AI,
1550 std::vector<AllocaInst*> &WorkList) {
1551 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n');
1552 SmallVector<AllocaInst*, 32> ElementAllocas;
1553 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
1554 ElementAllocas.reserve(ST->getNumContainedTypes());
1555 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
1556 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), nullptr,
1558 AI->getName() + "." + Twine(i), AI);
1559 ElementAllocas.push_back(NA);
1560 WorkList.push_back(NA); // Add to worklist for recursive processing
1563 ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
1564 ElementAllocas.reserve(AT->getNumElements());
1565 Type *ElTy = AT->getElementType();
1566 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1567 AllocaInst *NA = new AllocaInst(ElTy, nullptr, AI->getAlignment(),
1568 AI->getName() + "." + Twine(i), AI);
1569 ElementAllocas.push_back(NA);
1570 WorkList.push_back(NA); // Add to worklist for recursive processing
1574 // Now that we have created the new alloca instructions, rewrite all the
1575 // uses of the old alloca.
1576 RewriteForScalarRepl(AI, AI, 0, ElementAllocas);
1578 // Now erase any instructions that were made dead while rewriting the alloca.
1579 DeleteDeadInstructions();
1580 AI->eraseFromParent();
1585 /// DeleteDeadInstructions - Erase instructions on the DeadInstrs list,
1586 /// recursively including all their operands that become trivially dead.
1587 void SROA::DeleteDeadInstructions() {
1588 while (!DeadInsts.empty()) {
1589 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val());
1591 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
1592 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
1593 // Zero out the operand and see if it becomes trivially dead.
1594 // (But, don't add allocas to the dead instruction list -- they are
1595 // already on the worklist and will be deleted separately.)
1597 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U))
1598 DeadInsts.push_back(U);
1601 I->eraseFromParent();
1605 /// isSafeForScalarRepl - Check if instruction I is a safe use with regard to
1606 /// performing scalar replacement of alloca AI. The results are flagged in
1607 /// the Info parameter. Offset indicates the position within AI that is
1608 /// referenced by this instruction.
1609 void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset,
1611 const DataLayout &DL = I->getModule()->getDataLayout();
1612 for (Use &U : I->uses()) {
1613 Instruction *User = cast<Instruction>(U.getUser());
1615 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1616 isSafeForScalarRepl(BC, Offset, Info);
1617 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1618 uint64_t GEPOffset = Offset;
1619 isSafeGEP(GEPI, GEPOffset, Info);
1621 isSafeForScalarRepl(GEPI, GEPOffset, Info);
1622 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
1623 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
1624 if (!Length || Length->isNegative())
1625 return MarkUnsafe(Info, User);
1627 isSafeMemAccess(Offset, Length->getZExtValue(), nullptr,
1628 U.getOperandNo() == 0, Info, MI,
1629 true /*AllowWholeAccess*/);
1630 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1631 if (!LI->isSimple())
1632 return MarkUnsafe(Info, User);
1633 Type *LIType = LI->getType();
1634 isSafeMemAccess(Offset, DL.getTypeAllocSize(LIType), LIType, false, Info,
1635 LI, true /*AllowWholeAccess*/);
1636 Info.hasALoadOrStore = true;
1638 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1639 // Store is ok if storing INTO the pointer, not storing the pointer
1640 if (!SI->isSimple() || SI->getOperand(0) == I)
1641 return MarkUnsafe(Info, User);
1643 Type *SIType = SI->getOperand(0)->getType();
1644 isSafeMemAccess(Offset, DL.getTypeAllocSize(SIType), SIType, true, Info,
1645 SI, true /*AllowWholeAccess*/);
1646 Info.hasALoadOrStore = true;
1647 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
1648 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1649 II->getIntrinsicID() != Intrinsic::lifetime_end)
1650 return MarkUnsafe(Info, User);
1651 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) {
1652 isSafePHISelectUseForScalarRepl(User, Offset, Info);
1654 return MarkUnsafe(Info, User);
1656 if (Info.isUnsafe) return;
1661 /// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer
1662 /// derived from the alloca, we can often still split the alloca into elements.
1663 /// This is useful if we have a large alloca where one element is phi'd
1664 /// together somewhere: we can SRoA and promote all the other elements even if
1665 /// we end up not being able to promote this one.
1667 /// All we require is that the uses of the PHI do not index into other parts of
1668 /// the alloca. The most important use case for this is single load and stores
1669 /// that are PHI'd together, which can happen due to code sinking.
1670 void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset,
1672 // If we've already checked this PHI, don't do it again.
1673 if (PHINode *PN = dyn_cast<PHINode>(I))
1674 if (!Info.CheckedPHIs.insert(PN).second)
1677 const DataLayout &DL = I->getModule()->getDataLayout();
1678 for (User *U : I->users()) {
1679 Instruction *UI = cast<Instruction>(U);
1681 if (BitCastInst *BC = dyn_cast<BitCastInst>(UI)) {
1682 isSafePHISelectUseForScalarRepl(BC, Offset, Info);
1683 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(UI)) {
1684 // Only allow "bitcast" GEPs for simplicity. We could generalize this,
1685 // but would have to prove that we're staying inside of an element being
1687 if (!GEPI->hasAllZeroIndices())
1688 return MarkUnsafe(Info, UI);
1689 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info);
1690 } else if (LoadInst *LI = dyn_cast<LoadInst>(UI)) {
1691 if (!LI->isSimple())
1692 return MarkUnsafe(Info, UI);
1693 Type *LIType = LI->getType();
1694 isSafeMemAccess(Offset, DL.getTypeAllocSize(LIType), LIType, false, Info,
1695 LI, false /*AllowWholeAccess*/);
1696 Info.hasALoadOrStore = true;
1698 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1699 // Store is ok if storing INTO the pointer, not storing the pointer
1700 if (!SI->isSimple() || SI->getOperand(0) == I)
1701 return MarkUnsafe(Info, UI);
1703 Type *SIType = SI->getOperand(0)->getType();
1704 isSafeMemAccess(Offset, DL.getTypeAllocSize(SIType), SIType, true, Info,
1705 SI, false /*AllowWholeAccess*/);
1706 Info.hasALoadOrStore = true;
1707 } else if (isa<PHINode>(UI) || isa<SelectInst>(UI)) {
1708 isSafePHISelectUseForScalarRepl(UI, Offset, Info);
1710 return MarkUnsafe(Info, UI);
1712 if (Info.isUnsafe) return;
1716 /// isSafeGEP - Check if a GEP instruction can be handled for scalar
1717 /// replacement. It is safe when all the indices are constant, in-bounds
1718 /// references, and when the resulting offset corresponds to an element within
1719 /// the alloca type. The results are flagged in the Info parameter. Upon
1720 /// return, Offset is adjusted as specified by the GEP indices.
1721 void SROA::isSafeGEP(GetElementPtrInst *GEPI,
1722 uint64_t &Offset, AllocaInfo &Info) {
1723 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI);
1726 bool NonConstant = false;
1727 unsigned NonConstantIdxSize = 0;
1729 // Walk through the GEP type indices, checking the types that this indexes
1731 for (; GEPIt != E; ++GEPIt) {
1732 // Ignore struct elements, no extra checking needed for these.
1733 if ((*GEPIt)->isStructTy())
1736 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand());
1738 return MarkUnsafe(Info, GEPI);
1741 // Compute the offset due to this GEP and check if the alloca has a
1742 // component element at that offset.
1743 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
1744 // If this GEP is non-constant then the last operand must have been a
1745 // dynamic index into a vector. Pop this now as it has no impact on the
1746 // constant part of the offset.
1750 const DataLayout &DL = GEPI->getModule()->getDataLayout();
1751 Offset += DL.getIndexedOffset(GEPI->getPointerOperandType(), Indices);
1752 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, NonConstantIdxSize,
1754 MarkUnsafe(Info, GEPI);
1757 /// isHomogeneousAggregate - Check if type T is a struct or array containing
1758 /// elements of the same type (which is always true for arrays). If so,
1759 /// return true with NumElts and EltTy set to the number of elements and the
1760 /// element type, respectively.
1761 static bool isHomogeneousAggregate(Type *T, unsigned &NumElts,
1763 if (ArrayType *AT = dyn_cast<ArrayType>(T)) {
1764 NumElts = AT->getNumElements();
1765 EltTy = (NumElts == 0 ? nullptr : AT->getElementType());
1768 if (StructType *ST = dyn_cast<StructType>(T)) {
1769 NumElts = ST->getNumContainedTypes();
1770 EltTy = (NumElts == 0 ? nullptr : ST->getContainedType(0));
1771 for (unsigned n = 1; n < NumElts; ++n) {
1772 if (ST->getContainedType(n) != EltTy)
1780 /// isCompatibleAggregate - Check if T1 and T2 are either the same type or are
1781 /// "homogeneous" aggregates with the same element type and number of elements.
1782 static bool isCompatibleAggregate(Type *T1, Type *T2) {
1786 unsigned NumElts1, NumElts2;
1787 Type *EltTy1, *EltTy2;
1788 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) &&
1789 isHomogeneousAggregate(T2, NumElts2, EltTy2) &&
1790 NumElts1 == NumElts2 &&
1797 /// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI
1798 /// alloca or has an offset and size that corresponds to a component element
1799 /// within it. The offset checked here may have been formed from a GEP with a
1800 /// pointer bitcasted to a different type.
1802 /// If AllowWholeAccess is true, then this allows uses of the entire alloca as a
1803 /// unit. If false, it only allows accesses known to be in a single element.
1804 void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize,
1805 Type *MemOpType, bool isStore,
1806 AllocaInfo &Info, Instruction *TheAccess,
1807 bool AllowWholeAccess) {
1808 const DataLayout &DL = TheAccess->getModule()->getDataLayout();
1809 // Check if this is a load/store of the entire alloca.
1810 if (Offset == 0 && AllowWholeAccess &&
1811 MemSize == DL.getTypeAllocSize(Info.AI->getAllocatedType())) {
1812 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer
1813 // loads/stores (which are essentially the same as the MemIntrinsics with
1814 // regard to copying padding between elements). But, if an alloca is
1815 // flagged as both a source and destination of such operations, we'll need
1816 // to check later for padding between elements.
1817 if (!MemOpType || MemOpType->isIntegerTy()) {
1819 Info.isMemCpyDst = true;
1821 Info.isMemCpySrc = true;
1824 // This is also safe for references using a type that is compatible with
1825 // the type of the alloca, so that loads/stores can be rewritten using
1826 // insertvalue/extractvalue.
1827 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) {
1828 Info.hasSubelementAccess = true;
1832 // Check if the offset/size correspond to a component within the alloca type.
1833 Type *T = Info.AI->getAllocatedType();
1834 if (TypeHasComponent(T, Offset, MemSize, DL)) {
1835 Info.hasSubelementAccess = true;
1839 return MarkUnsafe(Info, TheAccess);
1842 /// TypeHasComponent - Return true if T has a component type with the
1843 /// specified offset and size. If Size is zero, do not check the size.
1844 bool SROA::TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size,
1845 const DataLayout &DL) {
1848 if (StructType *ST = dyn_cast<StructType>(T)) {
1849 const StructLayout *Layout = DL.getStructLayout(ST);
1850 unsigned EltIdx = Layout->getElementContainingOffset(Offset);
1851 EltTy = ST->getContainedType(EltIdx);
1852 EltSize = DL.getTypeAllocSize(EltTy);
1853 Offset -= Layout->getElementOffset(EltIdx);
1854 } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) {
1855 EltTy = AT->getElementType();
1856 EltSize = DL.getTypeAllocSize(EltTy);
1857 if (Offset >= AT->getNumElements() * EltSize)
1860 } else if (VectorType *VT = dyn_cast<VectorType>(T)) {
1861 EltTy = VT->getElementType();
1862 EltSize = DL.getTypeAllocSize(EltTy);
1863 if (Offset >= VT->getNumElements() * EltSize)
1869 if (Offset == 0 && (Size == 0 || EltSize == Size))
1871 // Check if the component spans multiple elements.
1872 if (Offset + Size > EltSize)
1874 return TypeHasComponent(EltTy, Offset, Size, DL);
1877 /// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite
1878 /// the instruction I, which references it, to use the separate elements.
1879 /// Offset indicates the position within AI that is referenced by this
1881 void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
1882 SmallVectorImpl<AllocaInst *> &NewElts) {
1883 const DataLayout &DL = I->getModule()->getDataLayout();
1884 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) {
1885 Use &TheUse = *UI++;
1886 Instruction *User = cast<Instruction>(TheUse.getUser());
1888 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1889 RewriteBitCast(BC, AI, Offset, NewElts);
1893 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1894 RewriteGEP(GEPI, AI, Offset, NewElts);
1898 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
1899 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
1900 uint64_t MemSize = Length->getZExtValue();
1901 if (Offset == 0 && MemSize == DL.getTypeAllocSize(AI->getAllocatedType()))
1902 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts);
1903 // Otherwise the intrinsic can only touch a single element and the
1904 // address operand will be updated, so nothing else needs to be done.
1908 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
1909 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
1910 II->getIntrinsicID() == Intrinsic::lifetime_end) {
1911 RewriteLifetimeIntrinsic(II, AI, Offset, NewElts);
1916 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1917 Type *LIType = LI->getType();
1919 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) {
1921 // %res = load { i32, i32 }* %alloc
1923 // %load.0 = load i32* %alloc.0
1924 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0
1925 // %load.1 = load i32* %alloc.1
1926 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1
1927 // (Also works for arrays instead of structs)
1928 Value *Insert = UndefValue::get(LIType);
1929 IRBuilder<> Builder(LI);
1930 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1931 Value *Load = Builder.CreateLoad(NewElts[i], "load");
1932 Insert = Builder.CreateInsertValue(Insert, Load, i, "insert");
1934 LI->replaceAllUsesWith(Insert);
1935 DeadInsts.push_back(LI);
1936 } else if (LIType->isIntegerTy() &&
1937 DL.getTypeAllocSize(LIType) ==
1938 DL.getTypeAllocSize(AI->getAllocatedType())) {
1939 // If this is a load of the entire alloca to an integer, rewrite it.
1940 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts);
1945 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1946 Value *Val = SI->getOperand(0);
1947 Type *SIType = Val->getType();
1948 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) {
1950 // store { i32, i32 } %val, { i32, i32 }* %alloc
1952 // %val.0 = extractvalue { i32, i32 } %val, 0
1953 // store i32 %val.0, i32* %alloc.0
1954 // %val.1 = extractvalue { i32, i32 } %val, 1
1955 // store i32 %val.1, i32* %alloc.1
1956 // (Also works for arrays instead of structs)
1957 IRBuilder<> Builder(SI);
1958 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1959 Value *Extract = Builder.CreateExtractValue(Val, i, Val->getName());
1960 Builder.CreateStore(Extract, NewElts[i]);
1962 DeadInsts.push_back(SI);
1963 } else if (SIType->isIntegerTy() &&
1964 DL.getTypeAllocSize(SIType) ==
1965 DL.getTypeAllocSize(AI->getAllocatedType())) {
1966 // If this is a store of the entire alloca from an integer, rewrite it.
1967 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts);
1972 if (isa<SelectInst>(User) || isa<PHINode>(User)) {
1973 // If we have a PHI user of the alloca itself (as opposed to a GEP or
1974 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to
1976 if (!isa<AllocaInst>(I)) continue;
1978 assert(Offset == 0 && NewElts[0] &&
1979 "Direct alloca use should have a zero offset");
1981 // If we have a use of the alloca, we know the derived uses will be
1982 // utilizing just the first element of the scalarized result. Insert a
1983 // bitcast of the first alloca before the user as required.
1984 AllocaInst *NewAI = NewElts[0];
1985 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI);
1986 NewAI->moveBefore(BCI);
1993 /// RewriteBitCast - Update a bitcast reference to the alloca being replaced
1994 /// and recursively continue updating all of its uses.
1995 void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
1996 SmallVectorImpl<AllocaInst *> &NewElts) {
1997 RewriteForScalarRepl(BC, AI, Offset, NewElts);
1998 if (BC->getOperand(0) != AI)
2001 // The bitcast references the original alloca. Replace its uses with
2002 // references to the alloca containing offset zero (which is normally at
2003 // index zero, but might not be in cases involving structs with elements
2005 Type *T = AI->getAllocatedType();
2006 uint64_t EltOffset = 0;
2008 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy,
2009 BC->getModule()->getDataLayout());
2010 Instruction *Val = NewElts[Idx];
2011 if (Val->getType() != BC->getDestTy()) {
2012 Val = new BitCastInst(Val, BC->getDestTy(), "", BC);
2015 BC->replaceAllUsesWith(Val);
2016 DeadInsts.push_back(BC);
2019 /// FindElementAndOffset - Return the index of the element containing Offset
2020 /// within the specified type, which must be either a struct or an array.
2021 /// Sets T to the type of the element and Offset to the offset within that
2022 /// element. IdxTy is set to the type of the index result to be used in a
2023 /// GEP instruction.
2024 uint64_t SROA::FindElementAndOffset(Type *&T, uint64_t &Offset, Type *&IdxTy,
2025 const DataLayout &DL) {
2028 if (StructType *ST = dyn_cast<StructType>(T)) {
2029 const StructLayout *Layout = DL.getStructLayout(ST);
2030 Idx = Layout->getElementContainingOffset(Offset);
2031 T = ST->getContainedType(Idx);
2032 Offset -= Layout->getElementOffset(Idx);
2033 IdxTy = Type::getInt32Ty(T->getContext());
2035 } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) {
2036 T = AT->getElementType();
2037 uint64_t EltSize = DL.getTypeAllocSize(T);
2038 Idx = Offset / EltSize;
2039 Offset -= Idx * EltSize;
2040 IdxTy = Type::getInt64Ty(T->getContext());
2043 VectorType *VT = cast<VectorType>(T);
2044 T = VT->getElementType();
2045 uint64_t EltSize = DL.getTypeAllocSize(T);
2046 Idx = Offset / EltSize;
2047 Offset -= Idx * EltSize;
2048 IdxTy = Type::getInt64Ty(T->getContext());
2052 /// RewriteGEP - Check if this GEP instruction moves the pointer across
2053 /// elements of the alloca that are being split apart, and if so, rewrite
2054 /// the GEP to be relative to the new element.
2055 void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
2056 SmallVectorImpl<AllocaInst *> &NewElts) {
2057 uint64_t OldOffset = Offset;
2058 const DataLayout &DL = GEPI->getModule()->getDataLayout();
2059 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
2060 // If the GEP was dynamic then it must have been a dynamic vector lookup.
2061 // In this case, it must be the last GEP operand which is dynamic so keep that
2062 // aside until we've found the constant GEP offset then add it back in at the
2064 Value* NonConstantIdx = nullptr;
2065 if (!GEPI->hasAllConstantIndices())
2066 NonConstantIdx = Indices.pop_back_val();
2067 Offset += DL.getIndexedOffset(GEPI->getPointerOperandType(), Indices);
2069 RewriteForScalarRepl(GEPI, AI, Offset, NewElts);
2071 Type *T = AI->getAllocatedType();
2073 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy, DL);
2074 if (GEPI->getOperand(0) == AI)
2075 OldIdx = ~0ULL; // Force the GEP to be rewritten.
2077 T = AI->getAllocatedType();
2078 uint64_t EltOffset = Offset;
2079 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy, DL);
2081 // If this GEP does not move the pointer across elements of the alloca
2082 // being split, then it does not needs to be rewritten.
2086 Type *i32Ty = Type::getInt32Ty(AI->getContext());
2087 SmallVector<Value*, 8> NewArgs;
2088 NewArgs.push_back(Constant::getNullValue(i32Ty));
2089 while (EltOffset != 0) {
2090 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy, DL);
2091 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx));
2093 if (NonConstantIdx) {
2095 // This GEP has a dynamic index. We need to add "i32 0" to index through
2096 // any structs or arrays in the original type until we get to the vector
2098 while (!isa<VectorType>(GepTy)) {
2099 NewArgs.push_back(Constant::getNullValue(i32Ty));
2100 GepTy = cast<CompositeType>(GepTy)->getTypeAtIndex(0U);
2102 NewArgs.push_back(NonConstantIdx);
2104 Instruction *Val = NewElts[Idx];
2105 if (NewArgs.size() > 1) {
2106 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs, "", GEPI);
2107 Val->takeName(GEPI);
2109 if (Val->getType() != GEPI->getType())
2110 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI);
2111 GEPI->replaceAllUsesWith(Val);
2112 DeadInsts.push_back(GEPI);
2115 /// RewriteLifetimeIntrinsic - II is a lifetime.start/lifetime.end. Rewrite it
2116 /// to mark the lifetime of the scalarized memory.
2117 void SROA::RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI,
2119 SmallVectorImpl<AllocaInst *> &NewElts) {
2120 ConstantInt *OldSize = cast<ConstantInt>(II->getArgOperand(0));
2121 // Put matching lifetime markers on everything from Offset up to
2123 Type *AIType = AI->getAllocatedType();
2124 const DataLayout &DL = II->getModule()->getDataLayout();
2125 uint64_t NewOffset = Offset;
2127 uint64_t Idx = FindElementAndOffset(AIType, NewOffset, IdxTy, DL);
2129 IRBuilder<> Builder(II);
2130 uint64_t Size = OldSize->getLimitedValue();
2133 // Splice the first element and index 'NewOffset' bytes in. SROA will
2134 // split the alloca again later.
2135 unsigned AS = AI->getType()->getAddressSpace();
2136 Value *V = Builder.CreateBitCast(NewElts[Idx], Builder.getInt8PtrTy(AS));
2137 V = Builder.CreateGEP(Builder.getInt8Ty(), V, Builder.getInt64(NewOffset));
2139 IdxTy = NewElts[Idx]->getAllocatedType();
2140 uint64_t EltSize = DL.getTypeAllocSize(IdxTy) - NewOffset;
2141 if (EltSize > Size) {
2147 if (II->getIntrinsicID() == Intrinsic::lifetime_start)
2148 Builder.CreateLifetimeStart(V, Builder.getInt64(EltSize));
2150 Builder.CreateLifetimeEnd(V, Builder.getInt64(EltSize));
2154 for (; Idx != NewElts.size() && Size; ++Idx) {
2155 IdxTy = NewElts[Idx]->getAllocatedType();
2156 uint64_t EltSize = DL.getTypeAllocSize(IdxTy);
2157 if (EltSize > Size) {
2163 if (II->getIntrinsicID() == Intrinsic::lifetime_start)
2164 Builder.CreateLifetimeStart(NewElts[Idx],
2165 Builder.getInt64(EltSize));
2167 Builder.CreateLifetimeEnd(NewElts[Idx],
2168 Builder.getInt64(EltSize));
2170 DeadInsts.push_back(II);
2173 /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI.
2174 /// Rewrite it to copy or set the elements of the scalarized memory.
2176 SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
2178 SmallVectorImpl<AllocaInst *> &NewElts) {
2179 // If this is a memcpy/memmove, construct the other pointer as the
2180 // appropriate type. The "Other" pointer is the pointer that goes to memory
2181 // that doesn't have anything to do with the alloca that we are promoting. For
2182 // memset, this Value* stays null.
2183 Value *OtherPtr = nullptr;
2184 unsigned MemAlignment = MI->getAlignment();
2185 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy
2186 if (Inst == MTI->getRawDest())
2187 OtherPtr = MTI->getRawSource();
2189 assert(Inst == MTI->getRawSource());
2190 OtherPtr = MTI->getRawDest();
2194 // If there is an other pointer, we want to convert it to the same pointer
2195 // type as AI has, so we can GEP through it safely.
2197 unsigned AddrSpace =
2198 cast<PointerType>(OtherPtr->getType())->getAddressSpace();
2200 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an
2201 // optimization, but it's also required to detect the corner case where
2202 // both pointer operands are referencing the same memory, and where
2203 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This
2204 // function is only called for mem intrinsics that access the whole
2205 // aggregate, so non-zero GEPs are not an issue here.)
2206 OtherPtr = OtherPtr->stripPointerCasts();
2208 // Copying the alloca to itself is a no-op: just delete it.
2209 if (OtherPtr == AI || OtherPtr == NewElts[0]) {
2210 // This code will run twice for a no-op memcpy -- once for each operand.
2211 // Put only one reference to MI on the DeadInsts list.
2212 for (SmallVectorImpl<Value *>::const_iterator I = DeadInsts.begin(),
2213 E = DeadInsts.end(); I != E; ++I)
2214 if (*I == MI) return;
2215 DeadInsts.push_back(MI);
2219 // If the pointer is not the right type, insert a bitcast to the right
2222 PointerType::get(AI->getType()->getElementType(), AddrSpace);
2224 if (OtherPtr->getType() != NewTy)
2225 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI);
2228 // Process each element of the aggregate.
2229 bool SROADest = MI->getRawDest() == Inst;
2231 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext()));
2232 const DataLayout &DL = MI->getModule()->getDataLayout();
2234 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2235 // If this is a memcpy/memmove, emit a GEP of the other element address.
2236 Value *OtherElt = nullptr;
2237 unsigned OtherEltAlign = MemAlignment;
2240 Value *Idx[2] = { Zero,
2241 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) };
2242 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx,
2243 OtherPtr->getName()+"."+Twine(i),
2246 PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType());
2247 Type *OtherTy = OtherPtrTy->getElementType();
2248 if (StructType *ST = dyn_cast<StructType>(OtherTy)) {
2249 EltOffset = DL.getStructLayout(ST)->getElementOffset(i);
2251 Type *EltTy = cast<SequentialType>(OtherTy)->getElementType();
2252 EltOffset = DL.getTypeAllocSize(EltTy) * i;
2255 // The alignment of the other pointer is the guaranteed alignment of the
2256 // element, which is affected by both the known alignment of the whole
2257 // mem intrinsic and the alignment of the element. If the alignment of
2258 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the
2259 // known alignment is just 4 bytes.
2260 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset);
2263 Value *EltPtr = NewElts[i];
2264 Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType();
2266 // If we got down to a scalar, insert a load or store as appropriate.
2267 if (EltTy->isSingleValueType()) {
2268 if (isa<MemTransferInst>(MI)) {
2270 // From Other to Alloca.
2271 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI);
2272 new StoreInst(Elt, EltPtr, MI);
2274 // From Alloca to Other.
2275 Value *Elt = new LoadInst(EltPtr, "tmp", MI);
2276 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI);
2280 assert(isa<MemSetInst>(MI));
2282 // If the stored element is zero (common case), just store a null
2285 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) {
2287 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0>
2289 // If EltTy is a vector type, get the element type.
2290 Type *ValTy = EltTy->getScalarType();
2292 // Construct an integer with the right value.
2293 unsigned EltSize = DL.getTypeSizeInBits(ValTy);
2294 APInt OneVal(EltSize, CI->getZExtValue());
2295 APInt TotalVal(OneVal);
2297 for (unsigned i = 0; 8*i < EltSize; ++i) {
2298 TotalVal = TotalVal.shl(8);
2302 // Convert the integer value to the appropriate type.
2303 StoreVal = ConstantInt::get(CI->getContext(), TotalVal);
2304 if (ValTy->isPointerTy())
2305 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy);
2306 else if (ValTy->isFloatingPointTy())
2307 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy);
2308 assert(StoreVal->getType() == ValTy && "Type mismatch!");
2310 // If the requested value was a vector constant, create it.
2311 if (EltTy->isVectorTy()) {
2312 unsigned NumElts = cast<VectorType>(EltTy)->getNumElements();
2313 StoreVal = ConstantVector::getSplat(NumElts, StoreVal);
2316 new StoreInst(StoreVal, EltPtr, MI);
2319 // Otherwise, if we're storing a byte variable, use a memset call for
2323 unsigned EltSize = DL.getTypeAllocSize(EltTy);
2327 IRBuilder<> Builder(MI);
2329 // Finally, insert the meminst for this element.
2330 if (isa<MemSetInst>(MI)) {
2331 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize,
2334 assert(isa<MemTransferInst>(MI));
2335 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr
2336 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr
2338 if (isa<MemCpyInst>(MI))
2339 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile());
2341 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile());
2344 DeadInsts.push_back(MI);
2347 /// RewriteStoreUserOfWholeAlloca - We found a store of an integer that
2348 /// overwrites the entire allocation. Extract out the pieces of the stored
2349 /// integer and store them individually.
2351 SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
2352 SmallVectorImpl<AllocaInst *> &NewElts) {
2353 // Extract each element out of the integer according to its structure offset
2354 // and store the element value to the individual alloca.
2355 Value *SrcVal = SI->getOperand(0);
2356 Type *AllocaEltTy = AI->getAllocatedType();
2357 const DataLayout &DL = SI->getModule()->getDataLayout();
2358 uint64_t AllocaSizeBits = DL.getTypeAllocSizeInBits(AllocaEltTy);
2360 IRBuilder<> Builder(SI);
2362 // Handle tail padding by extending the operand
2363 if (DL.getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits)
2364 SrcVal = Builder.CreateZExt(SrcVal,
2365 IntegerType::get(SI->getContext(), AllocaSizeBits));
2367 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI
2370 // There are two forms here: AI could be an array or struct. Both cases
2371 // have different ways to compute the element offset.
2372 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
2373 const StructLayout *Layout = DL.getStructLayout(EltSTy);
2375 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2376 // Get the number of bits to shift SrcVal to get the value.
2377 Type *FieldTy = EltSTy->getElementType(i);
2378 uint64_t Shift = Layout->getElementOffsetInBits(i);
2380 if (DL.isBigEndian())
2381 Shift = AllocaSizeBits - Shift - DL.getTypeAllocSizeInBits(FieldTy);
2383 Value *EltVal = SrcVal;
2385 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
2386 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt");
2389 // Truncate down to an integer of the right size.
2390 uint64_t FieldSizeBits = DL.getTypeSizeInBits(FieldTy);
2392 // Ignore zero sized fields like {}, they obviously contain no data.
2393 if (FieldSizeBits == 0) continue;
2395 if (FieldSizeBits != AllocaSizeBits)
2396 EltVal = Builder.CreateTrunc(EltVal,
2397 IntegerType::get(SI->getContext(), FieldSizeBits));
2398 Value *DestField = NewElts[i];
2399 if (EltVal->getType() == FieldTy) {
2400 // Storing to an integer field of this size, just do it.
2401 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) {
2402 // Bitcast to the right element type (for fp/vector values).
2403 EltVal = Builder.CreateBitCast(EltVal, FieldTy);
2405 // Otherwise, bitcast the dest pointer (for aggregates).
2406 DestField = Builder.CreateBitCast(DestField,
2407 PointerType::getUnqual(EltVal->getType()));
2409 new StoreInst(EltVal, DestField, SI);
2413 ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
2414 Type *ArrayEltTy = ATy->getElementType();
2415 uint64_t ElementOffset = DL.getTypeAllocSizeInBits(ArrayEltTy);
2416 uint64_t ElementSizeBits = DL.getTypeSizeInBits(ArrayEltTy);
2420 if (DL.isBigEndian())
2421 Shift = AllocaSizeBits-ElementOffset;
2425 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2426 // Ignore zero sized fields like {}, they obviously contain no data.
2427 if (ElementSizeBits == 0) continue;
2429 Value *EltVal = SrcVal;
2431 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
2432 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt");
2435 // Truncate down to an integer of the right size.
2436 if (ElementSizeBits != AllocaSizeBits)
2437 EltVal = Builder.CreateTrunc(EltVal,
2438 IntegerType::get(SI->getContext(),
2440 Value *DestField = NewElts[i];
2441 if (EltVal->getType() == ArrayEltTy) {
2442 // Storing to an integer field of this size, just do it.
2443 } else if (ArrayEltTy->isFloatingPointTy() ||
2444 ArrayEltTy->isVectorTy()) {
2445 // Bitcast to the right element type (for fp/vector values).
2446 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy);
2448 // Otherwise, bitcast the dest pointer (for aggregates).
2449 DestField = Builder.CreateBitCast(DestField,
2450 PointerType::getUnqual(EltVal->getType()));
2452 new StoreInst(EltVal, DestField, SI);
2454 if (DL.isBigEndian())
2455 Shift -= ElementOffset;
2457 Shift += ElementOffset;
2461 DeadInsts.push_back(SI);
2464 /// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to
2465 /// an integer. Load the individual pieces to form the aggregate value.
2467 SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
2468 SmallVectorImpl<AllocaInst *> &NewElts) {
2469 // Extract each element out of the NewElts according to its structure offset
2470 // and form the result value.
2471 Type *AllocaEltTy = AI->getAllocatedType();
2472 const DataLayout &DL = LI->getModule()->getDataLayout();
2473 uint64_t AllocaSizeBits = DL.getTypeAllocSizeInBits(AllocaEltTy);
2475 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI
2478 // There are two forms here: AI could be an array or struct. Both cases
2479 // have different ways to compute the element offset.
2480 const StructLayout *Layout = nullptr;
2481 uint64_t ArrayEltBitOffset = 0;
2482 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
2483 Layout = DL.getStructLayout(EltSTy);
2485 Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
2486 ArrayEltBitOffset = DL.getTypeAllocSizeInBits(ArrayEltTy);
2490 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits));
2492 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2493 // Load the value from the alloca. If the NewElt is an aggregate, cast
2494 // the pointer to an integer of the same size before doing the load.
2495 Value *SrcField = NewElts[i];
2497 cast<PointerType>(SrcField->getType())->getElementType();
2498 uint64_t FieldSizeBits = DL.getTypeSizeInBits(FieldTy);
2500 // Ignore zero sized fields like {}, they obviously contain no data.
2501 if (FieldSizeBits == 0) continue;
2503 IntegerType *FieldIntTy = IntegerType::get(LI->getContext(),
2505 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() &&
2506 !FieldTy->isVectorTy())
2507 SrcField = new BitCastInst(SrcField,
2508 PointerType::getUnqual(FieldIntTy),
2510 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI);
2512 // If SrcField is a fp or vector of the right size but that isn't an
2513 // integer type, bitcast to an integer so we can shift it.
2514 if (SrcField->getType() != FieldIntTy)
2515 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI);
2517 // Zero extend the field to be the same size as the final alloca so that
2518 // we can shift and insert it.
2519 if (SrcField->getType() != ResultVal->getType())
2520 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI);
2522 // Determine the number of bits to shift SrcField.
2524 if (Layout) // Struct case.
2525 Shift = Layout->getElementOffsetInBits(i);
2527 Shift = i*ArrayEltBitOffset;
2529 if (DL.isBigEndian())
2530 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth();
2533 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift);
2534 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI);
2537 // Don't create an 'or x, 0' on the first iteration.
2538 if (!isa<Constant>(ResultVal) ||
2539 !cast<Constant>(ResultVal)->isNullValue())
2540 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI);
2542 ResultVal = SrcField;
2545 // Handle tail padding by truncating the result
2546 if (DL.getTypeSizeInBits(LI->getType()) != AllocaSizeBits)
2547 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI);
2549 LI->replaceAllUsesWith(ResultVal);
2550 DeadInsts.push_back(LI);
2553 /// HasPadding - Return true if the specified type has any structure or
2554 /// alignment padding in between the elements that would be split apart
2555 /// by SROA; return false otherwise.
2556 static bool HasPadding(Type *Ty, const DataLayout &DL) {
2557 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2558 Ty = ATy->getElementType();
2559 return DL.getTypeSizeInBits(Ty) != DL.getTypeAllocSizeInBits(Ty);
2562 // SROA currently handles only Arrays and Structs.
2563 StructType *STy = cast<StructType>(Ty);
2564 const StructLayout *SL = DL.getStructLayout(STy);
2565 unsigned PrevFieldBitOffset = 0;
2566 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2567 unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
2569 // Check to see if there is any padding between this element and the
2572 unsigned PrevFieldEnd =
2573 PrevFieldBitOffset+DL.getTypeSizeInBits(STy->getElementType(i-1));
2574 if (PrevFieldEnd < FieldBitOffset)
2577 PrevFieldBitOffset = FieldBitOffset;
2579 // Check for tail padding.
2580 if (unsigned EltCount = STy->getNumElements()) {
2581 unsigned PrevFieldEnd = PrevFieldBitOffset +
2582 DL.getTypeSizeInBits(STy->getElementType(EltCount-1));
2583 if (PrevFieldEnd < SL->getSizeInBits())
2589 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
2590 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe,
2591 /// or 1 if safe after canonicalization has been performed.
2592 bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) {
2593 // Loop over the use list of the alloca. We can only transform it if all of
2594 // the users are safe to transform.
2595 AllocaInfo Info(AI);
2597 isSafeForScalarRepl(AI, 0, Info);
2598 if (Info.isUnsafe) {
2599 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n');
2603 const DataLayout &DL = AI->getModule()->getDataLayout();
2605 // Okay, we know all the users are promotable. If the aggregate is a memcpy
2606 // source and destination, we have to be careful. In particular, the memcpy
2607 // could be moving around elements that live in structure padding of the LLVM
2608 // types, but may actually be used. In these cases, we refuse to promote the
2610 if (Info.isMemCpySrc && Info.isMemCpyDst &&
2611 HasPadding(AI->getAllocatedType(), DL))
2614 // If the alloca never has an access to just *part* of it, but is accessed
2615 // via loads and stores, then we should use ConvertToScalarInfo to promote
2616 // the alloca instead of promoting each piece at a time and inserting fission
2618 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) {
2619 // If the struct/array just has one element, use basic SRoA.
2620 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
2621 if (ST->getNumElements() > 1) return false;
2623 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1)