1 //===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation implements the well known scalar replacement of
11 // aggregates transformation. This xform breaks up alloca instructions of
12 // aggregate type (structure or array) into individual alloca instructions for
13 // each member (if possible). Then, if possible, it transforms the individual
14 // alloca instructions into nice clean scalar SSA form.
16 // This combines a simple SRoA algorithm with the Mem2Reg algorithm because
17 // often interact, especially for C++ programs. As such, iterating between
18 // SRoA, then Mem2Reg until we run out of things to promote works well.
20 //===----------------------------------------------------------------------===//
22 #define DEBUG_TYPE "scalarrepl"
23 #include "llvm/Transforms/Scalar.h"
24 #include "llvm/Constants.h"
25 #include "llvm/DerivedTypes.h"
26 #include "llvm/Function.h"
27 #include "llvm/GlobalVariable.h"
28 #include "llvm/Instructions.h"
29 #include "llvm/IntrinsicInst.h"
30 #include "llvm/Pass.h"
31 #include "llvm/Analysis/Dominators.h"
32 #include "llvm/Target/TargetData.h"
33 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/GetElementPtrTypeIterator.h"
36 #include "llvm/Support/IRBuilder.h"
37 #include "llvm/Support/MathExtras.h"
38 #include "llvm/Support/Compiler.h"
39 #include "llvm/ADT/SmallVector.h"
40 #include "llvm/ADT/Statistic.h"
41 #include "llvm/ADT/StringExtras.h"
44 STATISTIC(NumReplaced, "Number of allocas broken up");
45 STATISTIC(NumPromoted, "Number of allocas promoted");
46 STATISTIC(NumConverted, "Number of aggregates converted to scalar");
47 STATISTIC(NumGlobals, "Number of allocas copied from constant global");
50 struct VISIBILITY_HIDDEN SROA : public FunctionPass {
51 static char ID; // Pass identification, replacement for typeid
52 explicit SROA(signed T = -1) : FunctionPass(&ID) {
59 bool runOnFunction(Function &F);
61 bool performScalarRepl(Function &F);
62 bool performPromotion(Function &F);
64 // getAnalysisUsage - This pass does not require any passes, but we know it
65 // will not alter the CFG, so say so.
66 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
67 AU.addRequired<DominatorTree>();
68 AU.addRequired<DominanceFrontier>();
69 AU.addRequired<TargetData>();
76 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures
77 /// information about the uses. All these fields are initialized to false
78 /// and set to true when something is learned.
80 /// isUnsafe - This is set to true if the alloca cannot be SROA'd.
83 /// needsCanon - This is set to true if there is some use of the alloca
84 /// that requires canonicalization.
87 /// isMemCpySrc - This is true if this aggregate is memcpy'd from.
90 /// isMemCpyDst - This is true if this aggregate is memcpy'd into.
94 : isUnsafe(false), needsCanon(false),
95 isMemCpySrc(false), isMemCpyDst(false) {}
100 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; }
102 int isSafeAllocaToScalarRepl(AllocationInst *AI);
104 void isSafeUseOfAllocation(Instruction *User, AllocationInst *AI,
106 void isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI,
108 void isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI,
109 unsigned OpNo, AllocaInfo &Info);
110 void isSafeUseOfBitCastedAllocation(BitCastInst *User, AllocationInst *AI,
113 void DoScalarReplacement(AllocationInst *AI,
114 std::vector<AllocationInst*> &WorkList);
115 void CanonicalizeAllocaUsers(AllocationInst *AI);
116 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base);
118 void RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI,
119 SmallVector<AllocaInst*, 32> &NewElts);
121 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst,
123 SmallVector<AllocaInst*, 32> &NewElts);
124 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocationInst *AI,
125 SmallVector<AllocaInst*, 32> &NewElts);
126 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI,
127 SmallVector<AllocaInst*, 32> &NewElts);
129 bool CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy,
130 bool &SawVec, uint64_t Offset, unsigned AllocaSize);
131 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset);
132 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType,
133 uint64_t Offset, IRBuilder<> &Builder);
134 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal,
135 uint64_t Offset, IRBuilder<> &Builder);
136 static Instruction *isOnlyCopiedFromConstantGlobal(AllocationInst *AI);
141 static RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates");
143 // Public interface to the ScalarReplAggregates pass
144 FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) {
145 return new SROA(Threshold);
149 bool SROA::runOnFunction(Function &F) {
150 TD = &getAnalysis<TargetData>();
152 bool Changed = performPromotion(F);
154 bool LocalChange = performScalarRepl(F);
155 if (!LocalChange) break; // No need to repromote if no scalarrepl
157 LocalChange = performPromotion(F);
158 if (!LocalChange) break; // No need to re-scalarrepl if no promotion
165 bool SROA::performPromotion(Function &F) {
166 std::vector<AllocaInst*> Allocas;
167 DominatorTree &DT = getAnalysis<DominatorTree>();
168 DominanceFrontier &DF = getAnalysis<DominanceFrontier>();
170 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
172 bool Changed = false;
177 // Find allocas that are safe to promote, by looking at all instructions in
179 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
180 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
181 if (isAllocaPromotable(AI))
182 Allocas.push_back(AI);
184 if (Allocas.empty()) break;
186 PromoteMemToReg(Allocas, DT, DF);
187 NumPromoted += Allocas.size();
194 /// getNumSAElements - Return the number of elements in the specific struct or
196 static uint64_t getNumSAElements(const Type *T) {
197 if (const StructType *ST = dyn_cast<StructType>(T))
198 return ST->getNumElements();
199 return cast<ArrayType>(T)->getNumElements();
202 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
203 // which runs on all of the malloc/alloca instructions in the function, removing
204 // them if they are only used by getelementptr instructions.
206 bool SROA::performScalarRepl(Function &F) {
207 std::vector<AllocationInst*> WorkList;
209 // Scan the entry basic block, adding any alloca's and mallocs to the worklist
210 BasicBlock &BB = F.getEntryBlock();
211 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
212 if (AllocationInst *A = dyn_cast<AllocationInst>(I))
213 WorkList.push_back(A);
215 // Process the worklist
216 bool Changed = false;
217 while (!WorkList.empty()) {
218 AllocationInst *AI = WorkList.back();
221 // Handle dead allocas trivially. These can be formed by SROA'ing arrays
222 // with unused elements.
223 if (AI->use_empty()) {
224 AI->eraseFromParent();
228 // If this alloca is impossible for us to promote, reject it early.
229 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized())
232 // Check to see if this allocation is only modified by a memcpy/memmove from
233 // a constant global. If this is the case, we can change all users to use
234 // the constant global instead. This is commonly produced by the CFE by
235 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
236 // is only subsequently read.
237 if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) {
238 DOUT << "Found alloca equal to global: " << *AI;
239 DOUT << " memcpy = " << *TheCopy;
240 Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2));
241 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType()));
242 TheCopy->eraseFromParent(); // Don't mutate the global.
243 AI->eraseFromParent();
249 // Check to see if we can perform the core SROA transformation. We cannot
250 // transform the allocation instruction if it is an array allocation
251 // (allocations OF arrays are ok though), and an allocation of a scalar
252 // value cannot be decomposed at all.
253 uint64_t AllocaSize = TD->getTypePaddedSize(AI->getAllocatedType());
255 if ((isa<StructType>(AI->getAllocatedType()) ||
256 isa<ArrayType>(AI->getAllocatedType())) &&
257 // Do not promote any struct whose size is too big.
258 AllocaSize < SRThreshold &&
259 // Do not promote any struct into more than "32" separate vars.
260 getNumSAElements(AI->getAllocatedType()) < SRThreshold/4) {
261 // Check that all of the users of the allocation are capable of being
263 switch (isSafeAllocaToScalarRepl(AI)) {
264 default: assert(0 && "Unexpected value!");
265 case 0: // Not safe to scalar replace.
267 case 1: // Safe, but requires cleanup/canonicalizations first
268 CanonicalizeAllocaUsers(AI);
270 case 3: // Safe to scalar replace.
271 DoScalarReplacement(AI, WorkList);
277 // If we can turn this aggregate value (potentially with casts) into a
278 // simple scalar value that can be mem2reg'd into a register value.
279 // IsNotTrivial tracks whether this is something that mem2reg could have
280 // promoted itself. If so, we don't want to transform it needlessly. Note
281 // that we can't just check based on the type: the alloca may be of an i32
282 // but that has pointer arithmetic to set byte 3 of it or something.
283 bool IsNotTrivial = false;
284 const Type *VectorTy = 0;
285 bool HadAVector = false;
286 if (CanConvertToScalar(AI, IsNotTrivial, VectorTy, HadAVector,
287 0, unsigned(AllocaSize)) && IsNotTrivial) {
289 // If we were able to find a vector type that can handle this with
290 // insert/extract elements, and if there was at least one use that had
291 // a vector type, promote this to a vector. We don't want to promote
292 // random stuff that doesn't use vectors (e.g. <9 x double>) because then
293 // we just get a lot of insert/extracts. If at least one vector is
294 // involved, then we probably really do have a union of vector/array.
295 if (VectorTy && isa<VectorType>(VectorTy) && HadAVector) {
296 DOUT << "CONVERT TO VECTOR: " << *AI << " TYPE = " << *VectorTy <<"\n";
298 // Create and insert the vector alloca.
299 NewAI = new AllocaInst(VectorTy, 0, "", AI->getParent()->begin());
300 ConvertUsesToScalar(AI, NewAI, 0);
302 DOUT << "CONVERT TO SCALAR INTEGER: " << *AI << "\n";
304 // Create and insert the integer alloca.
305 const Type *NewTy = IntegerType::get(AllocaSize*8);
306 NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin());
307 ConvertUsesToScalar(AI, NewAI, 0);
310 AI->eraseFromParent();
316 // Otherwise, couldn't process this alloca.
322 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl
323 /// predicate, do SROA now.
324 void SROA::DoScalarReplacement(AllocationInst *AI,
325 std::vector<AllocationInst*> &WorkList) {
326 DOUT << "Found inst to SROA: " << *AI;
327 SmallVector<AllocaInst*, 32> ElementAllocas;
328 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
329 ElementAllocas.reserve(ST->getNumContainedTypes());
330 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
331 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
333 AI->getName() + "." + utostr(i), AI);
334 ElementAllocas.push_back(NA);
335 WorkList.push_back(NA); // Add to worklist for recursive processing
338 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
339 ElementAllocas.reserve(AT->getNumElements());
340 const Type *ElTy = AT->getElementType();
341 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
342 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
343 AI->getName() + "." + utostr(i), AI);
344 ElementAllocas.push_back(NA);
345 WorkList.push_back(NA); // Add to worklist for recursive processing
349 // Now that we have created the alloca instructions that we want to use,
350 // expand the getelementptr instructions to use them.
352 while (!AI->use_empty()) {
353 Instruction *User = cast<Instruction>(AI->use_back());
354 if (BitCastInst *BCInst = dyn_cast<BitCastInst>(User)) {
355 RewriteBitCastUserOfAlloca(BCInst, AI, ElementAllocas);
356 BCInst->eraseFromParent();
361 // %res = load { i32, i32 }* %alloc
363 // %load.0 = load i32* %alloc.0
364 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0
365 // %load.1 = load i32* %alloc.1
366 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1
367 // (Also works for arrays instead of structs)
368 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
369 Value *Insert = UndefValue::get(LI->getType());
370 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) {
371 Value *Load = new LoadInst(ElementAllocas[i], "load", LI);
372 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI);
374 LI->replaceAllUsesWith(Insert);
375 LI->eraseFromParent();
380 // store { i32, i32 } %val, { i32, i32 }* %alloc
382 // %val.0 = extractvalue { i32, i32 } %val, 0
383 // store i32 %val.0, i32* %alloc.0
384 // %val.1 = extractvalue { i32, i32 } %val, 1
385 // store i32 %val.1, i32* %alloc.1
386 // (Also works for arrays instead of structs)
387 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
388 Value *Val = SI->getOperand(0);
389 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) {
390 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI);
391 new StoreInst(Extract, ElementAllocas[i], SI);
393 SI->eraseFromParent();
397 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User);
398 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst>
400 (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
402 assert(Idx < ElementAllocas.size() && "Index out of range?");
403 AllocaInst *AllocaToUse = ElementAllocas[Idx];
406 if (GEPI->getNumOperands() == 3) {
407 // Do not insert a new getelementptr instruction with zero indices, only
408 // to have it optimized out later.
409 RepValue = AllocaToUse;
411 // We are indexing deeply into the structure, so we still need a
412 // getelement ptr instruction to finish the indexing. This may be
413 // expanded itself once the worklist is rerun.
415 SmallVector<Value*, 8> NewArgs;
416 NewArgs.push_back(Constant::getNullValue(Type::Int32Ty));
417 NewArgs.append(GEPI->op_begin()+3, GEPI->op_end());
418 RepValue = GetElementPtrInst::Create(AllocaToUse, NewArgs.begin(),
419 NewArgs.end(), "", GEPI);
420 RepValue->takeName(GEPI);
423 // If this GEP is to the start of the aggregate, check for memcpys.
424 if (Idx == 0 && GEPI->hasAllZeroIndices())
425 RewriteBitCastUserOfAlloca(GEPI, AI, ElementAllocas);
427 // Move all of the users over to the new GEP.
428 GEPI->replaceAllUsesWith(RepValue);
429 // Delete the old GEP
430 GEPI->eraseFromParent();
433 // Finally, delete the Alloca instruction
434 AI->eraseFromParent();
439 /// isSafeElementUse - Check to see if this use is an allowed use for a
440 /// getelementptr instruction of an array aggregate allocation. isFirstElt
441 /// indicates whether Ptr is known to the start of the aggregate.
443 void SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI,
445 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
447 Instruction *User = cast<Instruction>(*I);
448 switch (User->getOpcode()) {
449 case Instruction::Load: break;
450 case Instruction::Store:
451 // Store is ok if storing INTO the pointer, not storing the pointer
452 if (User->getOperand(0) == Ptr) return MarkUnsafe(Info);
454 case Instruction::GetElementPtr: {
455 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User);
456 bool AreAllZeroIndices = isFirstElt;
457 if (GEP->getNumOperands() > 1) {
458 if (!isa<ConstantInt>(GEP->getOperand(1)) ||
459 !cast<ConstantInt>(GEP->getOperand(1))->isZero())
460 // Using pointer arithmetic to navigate the array.
461 return MarkUnsafe(Info);
463 if (AreAllZeroIndices)
464 AreAllZeroIndices = GEP->hasAllZeroIndices();
466 isSafeElementUse(GEP, AreAllZeroIndices, AI, Info);
467 if (Info.isUnsafe) return;
470 case Instruction::BitCast:
472 isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info);
473 if (Info.isUnsafe) return;
476 DOUT << " Transformation preventing inst: " << *User;
477 return MarkUnsafe(Info);
478 case Instruction::Call:
479 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
481 isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info);
482 if (Info.isUnsafe) return;
486 DOUT << " Transformation preventing inst: " << *User;
487 return MarkUnsafe(Info);
489 DOUT << " Transformation preventing inst: " << *User;
490 return MarkUnsafe(Info);
493 return; // All users look ok :)
496 /// AllUsersAreLoads - Return true if all users of this value are loads.
497 static bool AllUsersAreLoads(Value *Ptr) {
498 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
500 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load)
505 /// isSafeUseOfAllocation - Check to see if this user is an allowed use for an
506 /// aggregate allocation.
508 void SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI,
510 if (BitCastInst *C = dyn_cast<BitCastInst>(User))
511 return isSafeUseOfBitCastedAllocation(C, AI, Info);
513 if (LoadInst *LI = dyn_cast<LoadInst>(User))
514 if (!LI->isVolatile())
515 return;// Loads (returning a first class aggregrate) are always rewritable
517 if (StoreInst *SI = dyn_cast<StoreInst>(User))
518 if (!SI->isVolatile() && SI->getOperand(0) != AI)
519 return;// Store is ok if storing INTO the pointer, not storing the pointer
521 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User);
523 return MarkUnsafe(Info);
525 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI);
527 // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>".
529 I.getOperand() != Constant::getNullValue(I.getOperand()->getType())) {
530 return MarkUnsafe(Info);
534 if (I == E) return MarkUnsafe(Info); // ran out of GEP indices??
536 bool IsAllZeroIndices = true;
538 // If the first index is a non-constant index into an array, see if we can
539 // handle it as a special case.
540 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
541 if (!isa<ConstantInt>(I.getOperand())) {
542 IsAllZeroIndices = 0;
543 uint64_t NumElements = AT->getNumElements();
545 // If this is an array index and the index is not constant, we cannot
546 // promote... that is unless the array has exactly one or two elements in
547 // it, in which case we CAN promote it, but we have to canonicalize this
548 // out if this is the only problem.
549 if ((NumElements == 1 || NumElements == 2) &&
550 AllUsersAreLoads(GEPI)) {
551 Info.needsCanon = true;
552 return; // Canonicalization required!
554 return MarkUnsafe(Info);
558 // Walk through the GEP type indices, checking the types that this indexes
560 for (; I != E; ++I) {
561 // Ignore struct elements, no extra checking needed for these.
562 if (isa<StructType>(*I))
565 ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand());
566 if (!IdxVal) return MarkUnsafe(Info);
568 // Are all indices still zero?
569 IsAllZeroIndices &= IdxVal->isZero();
571 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
572 // This GEP indexes an array. Verify that this is an in-range constant
573 // integer. Specifically, consider A[0][i]. We cannot know that the user
574 // isn't doing invalid things like allowing i to index an out-of-range
575 // subscript that accesses A[1]. Because of this, we have to reject SROA
576 // of any accesses into structs where any of the components are variables.
577 if (IdxVal->getZExtValue() >= AT->getNumElements())
578 return MarkUnsafe(Info);
579 } else if (const VectorType *VT = dyn_cast<VectorType>(*I)) {
580 if (IdxVal->getZExtValue() >= VT->getNumElements())
581 return MarkUnsafe(Info);
585 // If there are any non-simple uses of this getelementptr, make sure to reject
587 return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info);
590 /// isSafeMemIntrinsicOnAllocation - Return true if the specified memory
591 /// intrinsic can be promoted by SROA. At this point, we know that the operand
592 /// of the memintrinsic is a pointer to the beginning of the allocation.
593 void SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI,
594 unsigned OpNo, AllocaInfo &Info) {
595 // If not constant length, give up.
596 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
597 if (!Length) return MarkUnsafe(Info);
599 // If not the whole aggregate, give up.
600 if (Length->getZExtValue() !=
601 TD->getTypePaddedSize(AI->getType()->getElementType()))
602 return MarkUnsafe(Info);
604 // We only know about memcpy/memset/memmove.
605 if (!isa<MemCpyInst>(MI) && !isa<MemSetInst>(MI) && !isa<MemMoveInst>(MI))
606 return MarkUnsafe(Info);
608 // Otherwise, we can transform it. Determine whether this is a memcpy/set
609 // into or out of the aggregate.
611 Info.isMemCpyDst = true;
614 Info.isMemCpySrc = true;
618 /// isSafeUseOfBitCastedAllocation - Return true if all users of this bitcast
620 void SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocationInst *AI,
622 for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end();
624 if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) {
625 isSafeUseOfBitCastedAllocation(BCU, AI, Info);
626 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) {
627 isSafeMemIntrinsicOnAllocation(MI, AI, UI.getOperandNo(), Info);
628 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
629 if (SI->isVolatile())
630 return MarkUnsafe(Info);
632 // If storing the entire alloca in one chunk through a bitcasted pointer
633 // to integer, we can transform it. This happens (for example) when you
634 // cast a {i32,i32}* to i64* and store through it. This is similar to the
635 // memcpy case and occurs in various "byval" cases and emulated memcpys.
636 if (isa<IntegerType>(SI->getOperand(0)->getType()) &&
637 TD->getTypePaddedSize(SI->getOperand(0)->getType()) ==
638 TD->getTypePaddedSize(AI->getType()->getElementType())) {
639 Info.isMemCpyDst = true;
642 return MarkUnsafe(Info);
643 } else if (LoadInst *LI = dyn_cast<LoadInst>(UI)) {
644 if (LI->isVolatile())
645 return MarkUnsafe(Info);
647 // If loading the entire alloca in one chunk through a bitcasted pointer
648 // to integer, we can transform it. This happens (for example) when you
649 // cast a {i32,i32}* to i64* and load through it. This is similar to the
650 // memcpy case and occurs in various "byval" cases and emulated memcpys.
651 if (isa<IntegerType>(LI->getType()) &&
652 TD->getTypePaddedSize(LI->getType()) ==
653 TD->getTypePaddedSize(AI->getType()->getElementType())) {
654 Info.isMemCpySrc = true;
657 return MarkUnsafe(Info);
659 return MarkUnsafe(Info);
661 if (Info.isUnsafe) return;
665 /// RewriteBitCastUserOfAlloca - BCInst (transitively) bitcasts AI, or indexes
666 /// to its first element. Transform users of the cast to use the new values
668 void SROA::RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI,
669 SmallVector<AllocaInst*, 32> &NewElts) {
670 Value::use_iterator UI = BCInst->use_begin(), UE = BCInst->use_end();
672 Instruction *User = cast<Instruction>(*UI++);
673 if (BitCastInst *BCU = dyn_cast<BitCastInst>(User)) {
674 RewriteBitCastUserOfAlloca(BCU, AI, NewElts);
675 if (BCU->use_empty()) BCU->eraseFromParent();
679 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
680 // This must be memcpy/memmove/memset of the entire aggregate.
681 // Split into one per element.
682 RewriteMemIntrinUserOfAlloca(MI, BCInst, AI, NewElts);
686 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
687 // If this is a store of the entire alloca from an integer, rewrite it.
688 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts);
692 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
693 // If this is a load of the entire alloca to an integer, rewrite it.
694 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts);
698 // Otherwise it must be some other user of a gep of the first pointer. Just
699 // leave these alone.
704 /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI.
705 /// Rewrite it to copy or set the elements of the scalarized memory.
706 void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst,
708 SmallVector<AllocaInst*, 32> &NewElts) {
710 // If this is a memcpy/memmove, construct the other pointer as the
713 if (MemCpyInst *MCI = dyn_cast<MemCpyInst>(MI)) {
714 if (BCInst == MCI->getRawDest())
715 OtherPtr = MCI->getRawSource();
717 assert(BCInst == MCI->getRawSource());
718 OtherPtr = MCI->getRawDest();
720 } else if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
721 if (BCInst == MMI->getRawDest())
722 OtherPtr = MMI->getRawSource();
724 assert(BCInst == MMI->getRawSource());
725 OtherPtr = MMI->getRawDest();
729 // If there is an other pointer, we want to convert it to the same pointer
730 // type as AI has, so we can GEP through it safely.
732 // It is likely that OtherPtr is a bitcast, if so, remove it.
733 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr))
734 OtherPtr = BC->getOperand(0);
735 // All zero GEPs are effectively bitcasts.
736 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr))
737 if (GEP->hasAllZeroIndices())
738 OtherPtr = GEP->getOperand(0);
740 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr))
741 if (BCE->getOpcode() == Instruction::BitCast)
742 OtherPtr = BCE->getOperand(0);
744 // If the pointer is not the right type, insert a bitcast to the right
746 if (OtherPtr->getType() != AI->getType())
747 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(),
751 // Process each element of the aggregate.
752 Value *TheFn = MI->getOperand(0);
753 const Type *BytePtrTy = MI->getRawDest()->getType();
754 bool SROADest = MI->getRawDest() == BCInst;
756 Constant *Zero = Constant::getNullValue(Type::Int32Ty);
758 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
759 // If this is a memcpy/memmove, emit a GEP of the other element address.
762 Value *Idx[2] = { Zero, ConstantInt::get(Type::Int32Ty, i) };
763 OtherElt = GetElementPtrInst::Create(OtherPtr, Idx, Idx + 2,
764 OtherPtr->getNameStr()+"."+utostr(i),
768 Value *EltPtr = NewElts[i];
769 const Type *EltTy =cast<PointerType>(EltPtr->getType())->getElementType();
771 // If we got down to a scalar, insert a load or store as appropriate.
772 if (EltTy->isSingleValueType()) {
773 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
774 Value *Elt = new LoadInst(SROADest ? OtherElt : EltPtr, "tmp",
776 new StoreInst(Elt, SROADest ? EltPtr : OtherElt, MI);
779 assert(isa<MemSetInst>(MI));
781 // If the stored element is zero (common case), just store a null
784 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) {
786 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0>
788 // If EltTy is a vector type, get the element type.
789 const Type *ValTy = EltTy;
790 if (const VectorType *VTy = dyn_cast<VectorType>(ValTy))
791 ValTy = VTy->getElementType();
793 // Construct an integer with the right value.
794 unsigned EltSize = TD->getTypeSizeInBits(ValTy);
795 APInt OneVal(EltSize, CI->getZExtValue());
796 APInt TotalVal(OneVal);
798 for (unsigned i = 0; 8*i < EltSize; ++i) {
799 TotalVal = TotalVal.shl(8);
803 // Convert the integer value to the appropriate type.
804 StoreVal = ConstantInt::get(TotalVal);
805 if (isa<PointerType>(ValTy))
806 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy);
807 else if (ValTy->isFloatingPoint())
808 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy);
809 assert(StoreVal->getType() == ValTy && "Type mismatch!");
811 // If the requested value was a vector constant, create it.
812 if (EltTy != ValTy) {
813 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements();
814 SmallVector<Constant*, 16> Elts(NumElts, StoreVal);
815 StoreVal = ConstantVector::get(&Elts[0], NumElts);
818 new StoreInst(StoreVal, EltPtr, MI);
821 // Otherwise, if we're storing a byte variable, use a memset call for
825 // Cast the element pointer to BytePtrTy.
826 if (EltPtr->getType() != BytePtrTy)
827 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI);
829 // Cast the other pointer (if we have one) to BytePtrTy.
830 if (OtherElt && OtherElt->getType() != BytePtrTy)
831 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(),
834 unsigned EltSize = TD->getTypePaddedSize(EltTy);
836 // Finally, insert the meminst for this element.
837 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
839 SROADest ? EltPtr : OtherElt, // Dest ptr
840 SROADest ? OtherElt : EltPtr, // Src ptr
841 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
844 CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
846 assert(isa<MemSetInst>(MI));
848 EltPtr, MI->getOperand(2), // Dest, Value,
849 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
852 CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
855 MI->eraseFromParent();
858 /// RewriteStoreUserOfWholeAlloca - We found an store of an integer that
859 /// overwrites the entire allocation. Extract out the pieces of the stored
860 /// integer and store them individually.
861 void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI,
863 SmallVector<AllocaInst*, 32> &NewElts){
864 // Extract each element out of the integer according to its structure offset
865 // and store the element value to the individual alloca.
866 Value *SrcVal = SI->getOperand(0);
867 const Type *AllocaEltTy = AI->getType()->getElementType();
868 uint64_t AllocaSizeBits = TD->getTypePaddedSizeInBits(AllocaEltTy);
870 // If this isn't a store of an integer to the whole alloca, it may be a store
871 // to the first element. Just ignore the store in this case and normal SROA
873 if (!isa<IntegerType>(SrcVal->getType()) ||
874 TD->getTypePaddedSizeInBits(SrcVal->getType()) != AllocaSizeBits)
877 DOUT << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << *SI;
879 // There are two forms here: AI could be an array or struct. Both cases
880 // have different ways to compute the element offset.
881 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
882 const StructLayout *Layout = TD->getStructLayout(EltSTy);
884 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
885 // Get the number of bits to shift SrcVal to get the value.
886 const Type *FieldTy = EltSTy->getElementType(i);
887 uint64_t Shift = Layout->getElementOffsetInBits(i);
889 if (TD->isBigEndian())
890 Shift = AllocaSizeBits-Shift-TD->getTypePaddedSizeInBits(FieldTy);
892 Value *EltVal = SrcVal;
894 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
895 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
896 "sroa.store.elt", SI);
899 // Truncate down to an integer of the right size.
900 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
902 // Ignore zero sized fields like {}, they obviously contain no data.
903 if (FieldSizeBits == 0) continue;
905 if (FieldSizeBits != AllocaSizeBits)
906 EltVal = new TruncInst(EltVal, IntegerType::get(FieldSizeBits), "", SI);
907 Value *DestField = NewElts[i];
908 if (EltVal->getType() == FieldTy) {
909 // Storing to an integer field of this size, just do it.
910 } else if (FieldTy->isFloatingPoint() || isa<VectorType>(FieldTy)) {
911 // Bitcast to the right element type (for fp/vector values).
912 EltVal = new BitCastInst(EltVal, FieldTy, "", SI);
914 // Otherwise, bitcast the dest pointer (for aggregates).
915 DestField = new BitCastInst(DestField,
916 PointerType::getUnqual(EltVal->getType()),
919 new StoreInst(EltVal, DestField, SI);
923 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
924 const Type *ArrayEltTy = ATy->getElementType();
925 uint64_t ElementOffset = TD->getTypePaddedSizeInBits(ArrayEltTy);
926 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy);
930 if (TD->isBigEndian())
931 Shift = AllocaSizeBits-ElementOffset;
935 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
936 // Ignore zero sized fields like {}, they obviously contain no data.
937 if (ElementSizeBits == 0) continue;
939 Value *EltVal = SrcVal;
941 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
942 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
943 "sroa.store.elt", SI);
946 // Truncate down to an integer of the right size.
947 if (ElementSizeBits != AllocaSizeBits)
948 EltVal = new TruncInst(EltVal, IntegerType::get(ElementSizeBits),"",SI);
949 Value *DestField = NewElts[i];
950 if (EltVal->getType() == ArrayEltTy) {
951 // Storing to an integer field of this size, just do it.
952 } else if (ArrayEltTy->isFloatingPoint() || isa<VectorType>(ArrayEltTy)) {
953 // Bitcast to the right element type (for fp/vector values).
954 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI);
956 // Otherwise, bitcast the dest pointer (for aggregates).
957 DestField = new BitCastInst(DestField,
958 PointerType::getUnqual(EltVal->getType()),
961 new StoreInst(EltVal, DestField, SI);
963 if (TD->isBigEndian())
964 Shift -= ElementOffset;
966 Shift += ElementOffset;
970 SI->eraseFromParent();
973 /// RewriteLoadUserOfWholeAlloca - We found an load of the entire allocation to
974 /// an integer. Load the individual pieces to form the aggregate value.
975 void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI,
976 SmallVector<AllocaInst*, 32> &NewElts) {
977 // Extract each element out of the NewElts according to its structure offset
978 // and form the result value.
979 const Type *AllocaEltTy = AI->getType()->getElementType();
980 uint64_t AllocaSizeBits = TD->getTypePaddedSizeInBits(AllocaEltTy);
982 // If this isn't a load of the whole alloca to an integer, it may be a load
983 // of the first element. Just ignore the load in this case and normal SROA
985 if (!isa<IntegerType>(LI->getType()) ||
986 TD->getTypePaddedSizeInBits(LI->getType()) != AllocaSizeBits)
989 DOUT << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << *LI;
991 // There are two forms here: AI could be an array or struct. Both cases
992 // have different ways to compute the element offset.
993 const StructLayout *Layout = 0;
994 uint64_t ArrayEltBitOffset = 0;
995 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
996 Layout = TD->getStructLayout(EltSTy);
998 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
999 ArrayEltBitOffset = TD->getTypePaddedSizeInBits(ArrayEltTy);
1002 Value *ResultVal = Constant::getNullValue(LI->getType());
1004 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1005 // Load the value from the alloca. If the NewElt is an aggregate, cast
1006 // the pointer to an integer of the same size before doing the load.
1007 Value *SrcField = NewElts[i];
1008 const Type *FieldTy =
1009 cast<PointerType>(SrcField->getType())->getElementType();
1010 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
1012 // Ignore zero sized fields like {}, they obviously contain no data.
1013 if (FieldSizeBits == 0) continue;
1015 const IntegerType *FieldIntTy = IntegerType::get(FieldSizeBits);
1016 if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() &&
1017 !isa<VectorType>(FieldTy))
1018 SrcField = new BitCastInst(SrcField, PointerType::getUnqual(FieldIntTy),
1020 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI);
1022 // If SrcField is a fp or vector of the right size but that isn't an
1023 // integer type, bitcast to an integer so we can shift it.
1024 if (SrcField->getType() != FieldIntTy)
1025 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI);
1027 // Zero extend the field to be the same size as the final alloca so that
1028 // we can shift and insert it.
1029 if (SrcField->getType() != ResultVal->getType())
1030 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI);
1032 // Determine the number of bits to shift SrcField.
1034 if (Layout) // Struct case.
1035 Shift = Layout->getElementOffsetInBits(i);
1037 Shift = i*ArrayEltBitOffset;
1039 if (TD->isBigEndian())
1040 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth();
1043 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift);
1044 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI);
1047 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI);
1050 LI->replaceAllUsesWith(ResultVal);
1051 LI->eraseFromParent();
1055 /// HasPadding - Return true if the specified type has any structure or
1056 /// alignment padding, false otherwise.
1057 static bool HasPadding(const Type *Ty, const TargetData &TD) {
1058 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1059 const StructLayout *SL = TD.getStructLayout(STy);
1060 unsigned PrevFieldBitOffset = 0;
1061 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1062 unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
1064 // Padding in sub-elements?
1065 if (HasPadding(STy->getElementType(i), TD))
1068 // Check to see if there is any padding between this element and the
1071 unsigned PrevFieldEnd =
1072 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1));
1073 if (PrevFieldEnd < FieldBitOffset)
1077 PrevFieldBitOffset = FieldBitOffset;
1080 // Check for tail padding.
1081 if (unsigned EltCount = STy->getNumElements()) {
1082 unsigned PrevFieldEnd = PrevFieldBitOffset +
1083 TD.getTypeSizeInBits(STy->getElementType(EltCount-1));
1084 if (PrevFieldEnd < SL->getSizeInBits())
1088 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1089 return HasPadding(ATy->getElementType(), TD);
1090 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1091 return HasPadding(VTy->getElementType(), TD);
1093 return TD.getTypeSizeInBits(Ty) != TD.getTypePaddedSizeInBits(Ty);
1096 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
1097 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe,
1098 /// or 1 if safe after canonicalization has been performed.
1100 int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) {
1101 // Loop over the use list of the alloca. We can only transform it if all of
1102 // the users are safe to transform.
1105 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end();
1107 isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info);
1108 if (Info.isUnsafe) {
1109 DOUT << "Cannot transform: " << *AI << " due to user: " << **I;
1114 // Okay, we know all the users are promotable. If the aggregate is a memcpy
1115 // source and destination, we have to be careful. In particular, the memcpy
1116 // could be moving around elements that live in structure padding of the LLVM
1117 // types, but may actually be used. In these cases, we refuse to promote the
1119 if (Info.isMemCpySrc && Info.isMemCpyDst &&
1120 HasPadding(AI->getType()->getElementType(), *TD))
1123 // If we require cleanup, return 1, otherwise return 3.
1124 return Info.needsCanon ? 1 : 3;
1127 /// CanonicalizeAllocaUsers - If SROA reported that it can promote the specified
1128 /// allocation, but only if cleaned up, perform the cleanups required.
1129 void SROA::CanonicalizeAllocaUsers(AllocationInst *AI) {
1130 // At this point, we know that the end result will be SROA'd and promoted, so
1131 // we can insert ugly code if required so long as sroa+mem2reg will clean it
1133 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
1135 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI++);
1136 if (!GEPI) continue;
1137 gep_type_iterator I = gep_type_begin(GEPI);
1140 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
1141 uint64_t NumElements = AT->getNumElements();
1143 if (!isa<ConstantInt>(I.getOperand())) {
1144 if (NumElements == 1) {
1145 GEPI->setOperand(2, Constant::getNullValue(Type::Int32Ty));
1147 assert(NumElements == 2 && "Unhandled case!");
1148 // All users of the GEP must be loads. At each use of the GEP, insert
1149 // two loads of the appropriate indexed GEP and select between them.
1150 Value *IsOne = new ICmpInst(ICmpInst::ICMP_NE, I.getOperand(),
1151 Constant::getNullValue(I.getOperand()->getType()),
1153 // Insert the new GEP instructions, which are properly indexed.
1154 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end());
1155 Indices[1] = Constant::getNullValue(Type::Int32Ty);
1156 Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0),
1159 GEPI->getName()+".0", GEPI);
1160 Indices[1] = ConstantInt::get(Type::Int32Ty, 1);
1161 Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0),
1164 GEPI->getName()+".1", GEPI);
1165 // Replace all loads of the variable index GEP with loads from both
1166 // indexes and a select.
1167 while (!GEPI->use_empty()) {
1168 LoadInst *LI = cast<LoadInst>(GEPI->use_back());
1169 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI);
1170 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI);
1171 Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI);
1172 LI->replaceAllUsesWith(R);
1173 LI->eraseFromParent();
1175 GEPI->eraseFromParent();
1182 /// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at
1183 /// the offset specified by Offset (which is specified in bytes).
1185 /// There are two cases we handle here:
1186 /// 1) A union of vector types of the same size and potentially its elements.
1187 /// Here we turn element accesses into insert/extract element operations.
1188 /// This promotes a <4 x float> with a store of float to the third element
1189 /// into a <4 x float> that uses insert element.
1190 /// 2) A fully general blob of memory, which we turn into some (potentially
1191 /// large) integer type with extract and insert operations where the loads
1192 /// and stores would mutate the memory.
1193 static void MergeInType(const Type *In, uint64_t Offset, const Type *&VecTy,
1194 unsigned AllocaSize, const TargetData &TD) {
1195 // If this could be contributing to a vector, analyze it.
1196 if (VecTy != Type::VoidTy) { // either null or a vector type.
1198 // If the In type is a vector that is the same size as the alloca, see if it
1199 // matches the existing VecTy.
1200 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) {
1201 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) {
1202 // If we're storing/loading a vector of the right size, allow it as a
1203 // vector. If this the first vector we see, remember the type so that
1204 // we know the element size.
1209 } else if (In == Type::FloatTy || In == Type::DoubleTy ||
1210 (isa<IntegerType>(In) && In->getPrimitiveSizeInBits() >= 8 &&
1211 isPowerOf2_32(In->getPrimitiveSizeInBits()))) {
1212 // If we're accessing something that could be an element of a vector, see
1213 // if the implied vector agrees with what we already have and if Offset is
1214 // compatible with it.
1215 unsigned EltSize = In->getPrimitiveSizeInBits()/8;
1216 if (Offset % EltSize == 0 &&
1217 AllocaSize % EltSize == 0 &&
1219 cast<VectorType>(VecTy)->getElementType()
1220 ->getPrimitiveSizeInBits()/8 == EltSize)) {
1222 VecTy = VectorType::get(In, AllocaSize/EltSize);
1228 // Otherwise, we have a case that we can't handle with an optimized vector
1229 // form. We can still turn this into a large integer.
1230 VecTy = Type::VoidTy;
1233 /// CanConvertToScalar - V is a pointer. If we can convert the pointee and all
1234 /// its accesses to use a to single vector type, return true, and set VecTy to
1235 /// the new type. If we could convert the alloca into a single promotable
1236 /// integer, return true but set VecTy to VoidTy. Further, if the use is not a
1237 /// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset
1238 /// is the current offset from the base of the alloca being analyzed.
1240 /// If we see at least one access to the value that is as a vector type, set the
1243 bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy,
1244 bool &SawVec, uint64_t Offset,
1245 unsigned AllocaSize) {
1246 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
1247 Instruction *User = cast<Instruction>(*UI);
1249 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1250 // Don't break volatile loads.
1251 if (LI->isVolatile())
1253 MergeInType(LI->getType(), Offset, VecTy, AllocaSize, *TD);
1254 SawVec |= isa<VectorType>(LI->getType());
1258 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1259 // Storing the pointer, not into the value?
1260 if (SI->getOperand(0) == V || SI->isVolatile()) return 0;
1261 MergeInType(SI->getOperand(0)->getType(), Offset, VecTy, AllocaSize, *TD);
1262 SawVec |= isa<VectorType>(SI->getOperand(0)->getType());
1266 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
1267 if (!CanConvertToScalar(BCI, IsNotTrivial, VecTy, SawVec, Offset,
1270 IsNotTrivial = true;
1274 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
1275 // If this is a GEP with a variable indices, we can't handle it.
1276 if (!GEP->hasAllConstantIndices())
1279 // Compute the offset that this GEP adds to the pointer.
1280 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
1281 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(),
1282 &Indices[0], Indices.size());
1283 // See if all uses can be converted.
1284 if (!CanConvertToScalar(GEP, IsNotTrivial, VecTy, SawVec,Offset+GEPOffset,
1287 IsNotTrivial = true;
1291 // If this is a constant sized memset of a constant value (e.g. 0) we can
1293 if (isa<MemSetInst>(User) &&
1294 // Store of constant value.
1295 isa<ConstantInt>(User->getOperand(2)) &&
1296 // Store with constant size.
1297 isa<ConstantInt>(User->getOperand(3))) {
1298 VecTy = Type::VoidTy;
1299 IsNotTrivial = true;
1303 // Otherwise, we cannot handle this!
1311 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
1312 /// directly. This happens when we are converting an "integer union" to a
1313 /// single integer scalar, or when we are converting a "vector union" to a
1314 /// vector with insert/extractelement instructions.
1316 /// Offset is an offset from the original alloca, in bits that need to be
1317 /// shifted to the right. By the end of this, there should be no uses of Ptr.
1318 void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) {
1319 while (!Ptr->use_empty()) {
1320 Instruction *User = cast<Instruction>(Ptr->use_back());
1322 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
1323 ConvertUsesToScalar(CI, NewAI, Offset);
1324 CI->eraseFromParent();
1328 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
1329 // Compute the offset that this GEP adds to the pointer.
1330 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
1331 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(),
1332 &Indices[0], Indices.size());
1333 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8);
1334 GEP->eraseFromParent();
1338 IRBuilder<> Builder(User->getParent(), User);
1340 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1341 // The load is a bit extract from NewAI shifted right by Offset bits.
1342 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp");
1344 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder);
1345 LI->replaceAllUsesWith(NewLoadVal);
1346 LI->eraseFromParent();
1350 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1351 assert(SI->getOperand(0) != Ptr && "Consistency error!");
1352 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").c_str());
1353 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset,
1355 Builder.CreateStore(New, NewAI);
1356 SI->eraseFromParent();
1360 // If this is a constant sized memset of a constant value (e.g. 0) we can
1361 // transform it into a store of the expanded constant value.
1362 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
1363 assert(MSI->getRawDest() == Ptr && "Consistency error!");
1364 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
1365 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue();
1367 // Compute the value replicated the right number of times.
1368 APInt APVal(NumBytes*8, Val);
1370 // Splat the value if non-zero.
1372 for (unsigned i = 1; i != NumBytes; ++i)
1373 APVal |= APVal << 8;
1375 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").c_str());
1376 Value *New = ConvertScalar_InsertValue(ConstantInt::get(APVal), Old,
1378 Builder.CreateStore(New, NewAI);
1379 MSI->eraseFromParent();
1384 assert(0 && "Unsupported operation!");
1389 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer
1390 /// or vector value FromVal, extracting the bits from the offset specified by
1391 /// Offset. This returns the value, which is of type ToType.
1393 /// This happens when we are converting an "integer union" to a single
1394 /// integer scalar, or when we are converting a "vector union" to a vector with
1395 /// insert/extractelement instructions.
1397 /// Offset is an offset from the original alloca, in bits that need to be
1398 /// shifted to the right.
1399 Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType,
1400 uint64_t Offset, IRBuilder<> &Builder) {
1401 // If the load is of the whole new alloca, no conversion is needed.
1402 if (FromVal->getType() == ToType && Offset == 0)
1405 // If the result alloca is a vector type, this is either an element
1406 // access or a bitcast to another vector type of the same size.
1407 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) {
1408 if (isa<VectorType>(ToType))
1409 return Builder.CreateBitCast(FromVal, ToType, "tmp");
1411 // Otherwise it must be an element access.
1414 unsigned EltSize = TD->getTypePaddedSizeInBits(VTy->getElementType());
1415 Elt = Offset/EltSize;
1416 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
1418 // Return the element extracted out of it.
1419 Value *V = Builder.CreateExtractElement(FromVal,
1420 ConstantInt::get(Type::Int32Ty,Elt),
1422 if (V->getType() != ToType)
1423 V = Builder.CreateBitCast(V, ToType, "tmp");
1427 // If ToType is a first class aggregate, extract out each of the pieces and
1428 // use insertvalue's to form the FCA.
1429 if (const StructType *ST = dyn_cast<StructType>(ToType)) {
1430 const StructLayout &Layout = *TD->getStructLayout(ST);
1431 Value *Res = UndefValue::get(ST);
1432 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
1433 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i),
1434 Offset+Layout.getElementOffsetInBits(i),
1436 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
1441 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) {
1442 uint64_t EltSize = TD->getTypePaddedSizeInBits(AT->getElementType());
1443 Value *Res = UndefValue::get(AT);
1444 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1445 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(),
1446 Offset+i*EltSize, Builder);
1447 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
1452 // Otherwise, this must be a union that was converted to an integer value.
1453 const IntegerType *NTy = cast<IntegerType>(FromVal->getType());
1455 // If this is a big-endian system and the load is narrower than the
1456 // full alloca type, we need to do a shift to get the right bits.
1458 if (TD->isBigEndian()) {
1459 // On big-endian machines, the lowest bit is stored at the bit offset
1460 // from the pointer given by getTypeStoreSizeInBits. This matters for
1461 // integers with a bitwidth that is not a multiple of 8.
1462 ShAmt = TD->getTypeStoreSizeInBits(NTy) -
1463 TD->getTypeStoreSizeInBits(ToType) - Offset;
1468 // Note: we support negative bitwidths (with shl) which are not defined.
1469 // We do this to support (f.e.) loads off the end of a structure where
1470 // only some bits are used.
1471 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth())
1472 FromVal = Builder.CreateLShr(FromVal, ConstantInt::get(FromVal->getType(),
1474 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
1475 FromVal = Builder.CreateShl(FromVal, ConstantInt::get(FromVal->getType(),
1478 // Finally, unconditionally truncate the integer to the right width.
1479 unsigned LIBitWidth = TD->getTypeSizeInBits(ToType);
1480 if (LIBitWidth < NTy->getBitWidth())
1481 FromVal = Builder.CreateTrunc(FromVal, IntegerType::get(LIBitWidth), "tmp");
1482 else if (LIBitWidth > NTy->getBitWidth())
1483 FromVal = Builder.CreateZExt(FromVal, IntegerType::get(LIBitWidth), "tmp");
1485 // If the result is an integer, this is a trunc or bitcast.
1486 if (isa<IntegerType>(ToType)) {
1488 } else if (ToType->isFloatingPoint() || isa<VectorType>(ToType)) {
1489 // Just do a bitcast, we know the sizes match up.
1490 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp");
1492 // Otherwise must be a pointer.
1493 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp");
1495 assert(FromVal->getType() == ToType && "Didn't convert right?");
1500 /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer
1501 /// or vector value "Old" at the offset specified by Offset.
1503 /// This happens when we are converting an "integer union" to a
1504 /// single integer scalar, or when we are converting a "vector union" to a
1505 /// vector with insert/extractelement instructions.
1507 /// Offset is an offset from the original alloca, in bits that need to be
1508 /// shifted to the right.
1509 Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old,
1510 uint64_t Offset, IRBuilder<> &Builder) {
1512 // Convert the stored type to the actual type, shift it left to insert
1513 // then 'or' into place.
1514 const Type *AllocaType = Old->getType();
1516 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
1517 // If the result alloca is a vector type, this is either an element
1518 // access or a bitcast to another vector type.
1519 if (isa<VectorType>(SV->getType())) {
1520 SV = Builder.CreateBitCast(SV, AllocaType, "tmp");
1522 // Must be an element insertion.
1523 unsigned Elt = Offset/TD->getTypePaddedSizeInBits(VTy->getElementType());
1525 if (SV->getType() != VTy->getElementType())
1526 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp");
1528 SV = Builder.CreateInsertElement(Old, SV,
1529 ConstantInt::get(Type::Int32Ty, Elt),
1535 // If SV is a first-class aggregate value, insert each value recursively.
1536 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) {
1537 const StructLayout &Layout = *TD->getStructLayout(ST);
1538 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
1539 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
1540 Old = ConvertScalar_InsertValue(Elt, Old,
1541 Offset+Layout.getElementOffsetInBits(i),
1547 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) {
1548 uint64_t EltSize = TD->getTypePaddedSizeInBits(AT->getElementType());
1549 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1550 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
1551 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder);
1556 // If SV is a float, convert it to the appropriate integer type.
1557 // If it is a pointer, do the same.
1558 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType());
1559 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType);
1560 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType());
1561 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType);
1562 if (SV->getType()->isFloatingPoint() || isa<VectorType>(SV->getType()))
1563 SV = Builder.CreateBitCast(SV, IntegerType::get(SrcWidth), "tmp");
1564 else if (isa<PointerType>(SV->getType()))
1565 SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(), "tmp");
1567 // Zero extend or truncate the value if needed.
1568 if (SV->getType() != AllocaType) {
1569 if (SV->getType()->getPrimitiveSizeInBits() <
1570 AllocaType->getPrimitiveSizeInBits())
1571 SV = Builder.CreateZExt(SV, AllocaType, "tmp");
1573 // Truncation may be needed if storing more than the alloca can hold
1574 // (undefined behavior).
1575 SV = Builder.CreateTrunc(SV, AllocaType, "tmp");
1576 SrcWidth = DestWidth;
1577 SrcStoreWidth = DestStoreWidth;
1581 // If this is a big-endian system and the store is narrower than the
1582 // full alloca type, we need to do a shift to get the right bits.
1584 if (TD->isBigEndian()) {
1585 // On big-endian machines, the lowest bit is stored at the bit offset
1586 // from the pointer given by getTypeStoreSizeInBits. This matters for
1587 // integers with a bitwidth that is not a multiple of 8.
1588 ShAmt = DestStoreWidth - SrcStoreWidth - Offset;
1593 // Note: we support negative bitwidths (with shr) which are not defined.
1594 // We do this to support (f.e.) stores off the end of a structure where
1595 // only some bits in the structure are set.
1596 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
1597 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
1598 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), ShAmt), "tmp");
1600 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
1601 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), -ShAmt), "tmp");
1602 Mask = Mask.lshr(-ShAmt);
1605 // Mask out the bits we are about to insert from the old value, and or
1607 if (SrcWidth != DestWidth) {
1608 assert(DestWidth > SrcWidth);
1609 Old = Builder.CreateAnd(Old, ConstantInt::get(~Mask), "mask");
1610 SV = Builder.CreateOr(Old, SV, "ins");
1617 /// PointsToConstantGlobal - Return true if V (possibly indirectly) points to
1618 /// some part of a constant global variable. This intentionally only accepts
1619 /// constant expressions because we don't can't rewrite arbitrary instructions.
1620 static bool PointsToConstantGlobal(Value *V) {
1621 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
1622 return GV->isConstant();
1623 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
1624 if (CE->getOpcode() == Instruction::BitCast ||
1625 CE->getOpcode() == Instruction::GetElementPtr)
1626 return PointsToConstantGlobal(CE->getOperand(0));
1630 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
1631 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
1632 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
1633 /// track of whether it moves the pointer (with isOffset) but otherwise traverse
1634 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
1635 /// the alloca, and if the source pointer is a pointer to a constant global, we
1636 /// can optimize this.
1637 static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy,
1639 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
1640 if (LoadInst *LI = dyn_cast<LoadInst>(*UI))
1641 // Ignore non-volatile loads, they are always ok.
1642 if (!LI->isVolatile())
1645 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) {
1646 // If uses of the bitcast are ok, we are ok.
1647 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset))
1651 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) {
1652 // If the GEP has all zero indices, it doesn't offset the pointer. If it
1653 // doesn't, it does.
1654 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy,
1655 isOffset || !GEP->hasAllZeroIndices()))
1660 // If this is isn't our memcpy/memmove, reject it as something we can't
1662 if (!isa<MemCpyInst>(*UI) && !isa<MemMoveInst>(*UI))
1665 // If we already have seen a copy, reject the second one.
1666 if (TheCopy) return false;
1668 // If the pointer has been offset from the start of the alloca, we can't
1669 // safely handle this.
1670 if (isOffset) return false;
1672 // If the memintrinsic isn't using the alloca as the dest, reject it.
1673 if (UI.getOperandNo() != 1) return false;
1675 MemIntrinsic *MI = cast<MemIntrinsic>(*UI);
1677 // If the source of the memcpy/move is not a constant global, reject it.
1678 if (!PointsToConstantGlobal(MI->getOperand(2)))
1681 // Otherwise, the transform is safe. Remember the copy instruction.
1687 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
1688 /// modified by a copy from a constant global. If we can prove this, we can
1689 /// replace any uses of the alloca with uses of the global directly.
1690 Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocationInst *AI) {
1691 Instruction *TheCopy = 0;
1692 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false))