1 //===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source 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/Instructions.h"
28 #include "llvm/IntrinsicInst.h"
29 #include "llvm/Pass.h"
30 #include "llvm/Analysis/Dominators.h"
31 #include "llvm/Target/TargetData.h"
32 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/GetElementPtrTypeIterator.h"
35 #include "llvm/Support/MathExtras.h"
36 #include "llvm/Support/Compiler.h"
37 #include "llvm/ADT/SmallVector.h"
38 #include "llvm/ADT/Statistic.h"
39 #include "llvm/ADT/StringExtras.h"
42 STATISTIC(NumReplaced, "Number of allocas broken up");
43 STATISTIC(NumPromoted, "Number of allocas promoted");
44 STATISTIC(NumConverted, "Number of aggregates converted to scalar");
47 struct VISIBILITY_HIDDEN SROA : public FunctionPass {
48 bool runOnFunction(Function &F);
50 bool performScalarRepl(Function &F);
51 bool performPromotion(Function &F);
53 // getAnalysisUsage - This pass does not require any passes, but we know it
54 // will not alter the CFG, so say so.
55 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
56 AU.addRequired<DominatorTree>();
57 AU.addRequired<DominanceFrontier>();
58 AU.addRequired<TargetData>();
63 int isSafeElementUse(Value *Ptr);
64 int isSafeUseOfAllocation(Instruction *User, AllocationInst *AI);
65 bool isSafeUseOfBitCastedAllocation(BitCastInst *User, AllocationInst *AI);
66 int isSafeAllocaToScalarRepl(AllocationInst *AI);
67 void CanonicalizeAllocaUsers(AllocationInst *AI);
68 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base);
70 void RewriteBitCastUserOfAlloca(BitCastInst *BCInst, AllocationInst *AI,
71 SmallVector<AllocaInst*, 32> &NewElts);
73 const Type *CanConvertToScalar(Value *V, bool &IsNotTrivial);
74 void ConvertToScalar(AllocationInst *AI, const Type *Ty);
75 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, unsigned Offset);
78 RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates");
81 // Public interface to the ScalarReplAggregates pass
82 FunctionPass *llvm::createScalarReplAggregatesPass() { return new SROA(); }
85 bool SROA::runOnFunction(Function &F) {
86 bool Changed = performPromotion(F);
88 bool LocalChange = performScalarRepl(F);
89 if (!LocalChange) break; // No need to repromote if no scalarrepl
91 LocalChange = performPromotion(F);
92 if (!LocalChange) break; // No need to re-scalarrepl if no promotion
99 bool SROA::performPromotion(Function &F) {
100 std::vector<AllocaInst*> Allocas;
101 const TargetData &TD = getAnalysis<TargetData>();
102 DominatorTree &DT = getAnalysis<DominatorTree>();
103 DominanceFrontier &DF = getAnalysis<DominanceFrontier>();
105 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
107 bool Changed = false;
112 // Find allocas that are safe to promote, by looking at all instructions in
114 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
115 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
116 if (isAllocaPromotable(AI, TD))
117 Allocas.push_back(AI);
119 if (Allocas.empty()) break;
121 PromoteMemToReg(Allocas, DT, DF, TD);
122 NumPromoted += Allocas.size();
129 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
130 // which runs on all of the malloc/alloca instructions in the function, removing
131 // them if they are only used by getelementptr instructions.
133 bool SROA::performScalarRepl(Function &F) {
134 std::vector<AllocationInst*> WorkList;
136 // Scan the entry basic block, adding any alloca's and mallocs to the worklist
137 BasicBlock &BB = F.getEntryBlock();
138 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
139 if (AllocationInst *A = dyn_cast<AllocationInst>(I))
140 WorkList.push_back(A);
142 // Process the worklist
143 bool Changed = false;
144 while (!WorkList.empty()) {
145 AllocationInst *AI = WorkList.back();
148 // Handle dead allocas trivially. These can be formed by SROA'ing arrays
149 // with unused elements.
150 if (AI->use_empty()) {
151 AI->eraseFromParent();
155 // If we can turn this aggregate value (potentially with casts) into a
156 // simple scalar value that can be mem2reg'd into a register value.
157 bool IsNotTrivial = false;
158 if (const Type *ActualType = CanConvertToScalar(AI, IsNotTrivial))
159 if (IsNotTrivial && ActualType != Type::VoidTy) {
160 ConvertToScalar(AI, ActualType);
165 // We cannot transform the allocation instruction if it is an array
166 // allocation (allocations OF arrays are ok though), and an allocation of a
167 // scalar value cannot be decomposed at all.
169 if (AI->isArrayAllocation() ||
170 (!isa<StructType>(AI->getAllocatedType()) &&
171 !isa<ArrayType>(AI->getAllocatedType()))) continue;
173 // Check that all of the users of the allocation are capable of being
175 switch (isSafeAllocaToScalarRepl(AI)) {
176 default: assert(0 && "Unexpected value!");
177 case 0: // Not safe to scalar replace.
179 case 1: // Safe, but requires cleanup/canonicalizations first
180 CanonicalizeAllocaUsers(AI);
181 case 3: // Safe to scalar replace.
185 DOUT << "Found inst to xform: " << *AI;
188 SmallVector<AllocaInst*, 32> ElementAllocas;
189 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
190 ElementAllocas.reserve(ST->getNumContainedTypes());
191 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
192 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
194 AI->getName() + "." + utostr(i), AI);
195 ElementAllocas.push_back(NA);
196 WorkList.push_back(NA); // Add to worklist for recursive processing
199 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
200 ElementAllocas.reserve(AT->getNumElements());
201 const Type *ElTy = AT->getElementType();
202 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
203 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
204 AI->getName() + "." + utostr(i), AI);
205 ElementAllocas.push_back(NA);
206 WorkList.push_back(NA); // Add to worklist for recursive processing
210 // Now that we have created the alloca instructions that we want to use,
211 // expand the getelementptr instructions to use them.
213 while (!AI->use_empty()) {
214 Instruction *User = cast<Instruction>(AI->use_back());
215 if (BitCastInst *BCInst = dyn_cast<BitCastInst>(User)) {
216 RewriteBitCastUserOfAlloca(BCInst, AI, ElementAllocas);
220 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User);
221 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst>
223 (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
225 assert(Idx < ElementAllocas.size() && "Index out of range?");
226 AllocaInst *AllocaToUse = ElementAllocas[Idx];
229 if (GEPI->getNumOperands() == 3) {
230 // Do not insert a new getelementptr instruction with zero indices, only
231 // to have it optimized out later.
232 RepValue = AllocaToUse;
234 // We are indexing deeply into the structure, so we still need a
235 // getelement ptr instruction to finish the indexing. This may be
236 // expanded itself once the worklist is rerun.
238 SmallVector<Value*, 8> NewArgs;
239 NewArgs.push_back(Constant::getNullValue(Type::Int32Ty));
240 NewArgs.append(GEPI->op_begin()+3, GEPI->op_end());
241 RepValue = new GetElementPtrInst(AllocaToUse, &NewArgs[0],
242 NewArgs.size(), "", GEPI);
243 RepValue->takeName(GEPI);
246 // Move all of the users over to the new GEP.
247 GEPI->replaceAllUsesWith(RepValue);
248 // Delete the old GEP
249 GEPI->eraseFromParent();
252 // Finally, delete the Alloca instruction
253 AI->eraseFromParent();
261 /// isSafeElementUse - Check to see if this use is an allowed use for a
262 /// getelementptr instruction of an array aggregate allocation.
264 int SROA::isSafeElementUse(Value *Ptr) {
265 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
267 Instruction *User = cast<Instruction>(*I);
268 switch (User->getOpcode()) {
269 case Instruction::Load: break;
270 case Instruction::Store:
271 // Store is ok if storing INTO the pointer, not storing the pointer
272 if (User->getOperand(0) == Ptr) return 0;
274 case Instruction::GetElementPtr: {
275 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User);
276 if (GEP->getNumOperands() > 1) {
277 if (!isa<Constant>(GEP->getOperand(1)) ||
278 !cast<Constant>(GEP->getOperand(1))->isNullValue())
279 return 0; // Using pointer arithmetic to navigate the array...
281 if (!isSafeElementUse(GEP)) return 0;
285 DOUT << " Transformation preventing inst: " << *User;
289 return 3; // All users look ok :)
292 /// AllUsersAreLoads - Return true if all users of this value are loads.
293 static bool AllUsersAreLoads(Value *Ptr) {
294 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
296 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load)
301 /// isSafeUseOfAllocation - Check to see if this user is an allowed use for an
302 /// aggregate allocation.
304 int SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI) {
305 if (BitCastInst *C = dyn_cast<BitCastInst>(User))
306 return 0 && (isSafeUseOfBitCastedAllocation(C, AI) ? 3 : 0);
307 if (!isa<GetElementPtrInst>(User)) return 0;
309 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User);
310 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI);
312 // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>".
314 I.getOperand() != Constant::getNullValue(I.getOperand()->getType()))
318 if (I == E) return 0; // ran out of GEP indices??
320 // If this is a use of an array allocation, do a bit more checking for sanity.
321 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
322 uint64_t NumElements = AT->getNumElements();
324 if (isa<ConstantInt>(I.getOperand())) {
325 // Check to make sure that index falls within the array. If not,
326 // something funny is going on, so we won't do the optimization.
328 if (cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue() >= NumElements)
331 // We cannot scalar repl this level of the array unless any array
332 // sub-indices are in-range constants. In particular, consider:
333 // A[0][i]. We cannot know that the user isn't doing invalid things like
334 // allowing i to index an out-of-range subscript that accesses A[1].
336 // Scalar replacing *just* the outer index of the array is probably not
337 // going to be a win anyway, so just give up.
338 for (++I; I != E && (isa<ArrayType>(*I) || isa<VectorType>(*I)); ++I) {
339 uint64_t NumElements;
340 if (const ArrayType *SubArrayTy = dyn_cast<ArrayType>(*I))
341 NumElements = SubArrayTy->getNumElements();
343 NumElements = cast<VectorType>(*I)->getNumElements();
345 if (!isa<ConstantInt>(I.getOperand())) return 0;
346 if (cast<ConstantInt>(I.getOperand())->getZExtValue() >= NumElements)
351 // If this is an array index and the index is not constant, we cannot
352 // promote... that is unless the array has exactly one or two elements in
353 // it, in which case we CAN promote it, but we have to canonicalize this
354 // out if this is the only problem.
355 if ((NumElements == 1 || NumElements == 2) &&
356 AllUsersAreLoads(GEPI))
357 return 1; // Canonicalization required!
362 // If there are any non-simple uses of this getelementptr, make sure to reject
364 return isSafeElementUse(GEPI);
367 /// isSafeUseOfBitCastedAllocation - Return true if all users of this bitcast
369 bool SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocationInst *AI) {
370 for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end();
372 if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) {
373 if (!isSafeUseOfBitCastedAllocation(BCU, AI))
375 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) {
376 // If not constant length, give up.
377 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
378 if (!Length) return false;
380 // If not the whole aggregate, give up.
381 const TargetData &TD = getAnalysis<TargetData>();
382 if (Length->getZExtValue() !=
383 TD.getTypeSize(AI->getType()->getElementType()))
386 // We only know about memcpy/memset/memmove.
387 if (!isa<MemCpyInst>(MI) && !isa<MemSetInst>(MI) &&
388 !isa<MemMoveInst>(MI))
390 // Otherwise, we can transform it.
398 /// RewriteBitCastUserOfAlloca - BCInst (transitively) casts AI. Transform
399 /// users of the cast to use the new values instead.
400 void SROA::RewriteBitCastUserOfAlloca(BitCastInst *BCInst, AllocationInst *AI,
401 SmallVector<AllocaInst*, 32> &NewElts) {
402 Constant *Zero = Constant::getNullValue(Type::Int32Ty);
403 const TargetData &TD = getAnalysis<TargetData>();
404 while (!BCInst->use_empty()) {
405 if (BitCastInst *BCU = dyn_cast<BitCastInst>(BCInst->use_back())) {
406 RewriteBitCastUserOfAlloca(BCU, AI, NewElts);
410 // Otherwise, must be memcpy/memmove/memset of the entire aggregate. Split
411 // into one per element.
412 MemIntrinsic *MI = cast<MemIntrinsic>(BCInst->use_back());
414 // If this is a memcpy/memmove, construct the other pointer as the
417 if (MemCpyInst *MCI = dyn_cast<MemCpyInst>(MI)) {
418 if (BCInst == MCI->getRawDest())
419 OtherPtr = MCI->getRawSource();
421 assert(BCInst == MCI->getRawSource());
422 OtherPtr = MCI->getRawDest();
424 } else if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
425 if (BCInst == MMI->getRawDest())
426 OtherPtr = MMI->getRawSource();
428 assert(BCInst == MMI->getRawSource());
429 OtherPtr = MMI->getRawDest();
433 // If there is an other pointer, we want to convert it to the same pointer
434 // type as AI has, so we can GEP through it.
436 // It is likely that OtherPtr is a bitcast, if so, remove it.
437 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr))
438 OtherPtr = BC->getOperand(0);
439 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr))
440 if (BCE->getOpcode() == Instruction::BitCast)
441 OtherPtr = BCE->getOperand(0);
443 // If the pointer is not the right type, insert a bitcast to the right
445 if (OtherPtr->getType() != AI->getType())
446 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(),
450 // Process each element of the aggregate.
451 Value *TheFn = MI->getOperand(0);
452 const Type *BytePtrTy = MI->getRawDest()->getType();
453 bool SROADest = MI->getRawDest() == BCInst;
455 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
456 // If this is a memcpy/memmove, emit a GEP of the other element address.
459 OtherElt = new GetElementPtrInst(OtherPtr, Zero,
460 ConstantInt::get(Type::Int32Ty, i),
461 OtherPtr->getNameStr()+"."+utostr(i),
465 Value *EltPtr = NewElts[i];
466 const Type *EltTy =cast<PointerType>(EltPtr->getType())->getElementType();
468 // If we got down to a scalar, insert a load or store as appropriate.
469 if (EltTy->isFirstClassType()) {
470 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
471 Value *Elt = new LoadInst(SROADest ? OtherElt : EltPtr, "tmp",
473 new StoreInst(Elt, SROADest ? EltPtr : OtherElt, MI);
476 assert(isa<MemSetInst>(MI));
478 // If the stored element is zero (common case), just store a null
481 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) {
483 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0>
485 // If EltTy is a packed type, get the element type.
486 const Type *ValTy = EltTy;
487 if (const VectorType *VTy = dyn_cast<VectorType>(ValTy))
488 ValTy = VTy->getElementType();
490 // Construct an integer with the right value.
491 unsigned EltSize = TD.getTypeSize(ValTy);
492 APInt OneVal(EltSize*8, CI->getZExtValue());
493 APInt TotalVal(OneVal);
495 for (unsigned i = 0; i != EltSize-1; ++i) {
496 TotalVal = TotalVal.shl(8);
500 // Convert the integer value to the appropriate type.
501 StoreVal = ConstantInt::get(TotalVal);
502 if (isa<PointerType>(ValTy))
503 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy);
504 else if (ValTy->isFloatingPoint())
505 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy);
506 assert(StoreVal->getType() == ValTy && "Type mismatch!");
508 // If the requested value was a vector constant, create it.
509 if (EltTy != ValTy) {
510 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements();
511 SmallVector<Constant*, 16> Elts(NumElts, StoreVal);
512 StoreVal = ConstantVector::get(&Elts[0], NumElts);
515 new StoreInst(StoreVal, EltPtr, MI);
518 // Otherwise, if we're storing a byte variable, use a memset call for
523 // Cast the element pointer to BytePtrTy.
524 if (EltPtr->getType() != BytePtrTy)
525 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI);
527 // Cast the other pointer (if we have one) to BytePtrTy.
528 if (OtherElt && OtherElt->getType() != BytePtrTy)
529 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(),
532 unsigned EltSize = TD.getTypeSize(EltTy);
534 // Finally, insert the meminst for this element.
535 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
537 SROADest ? EltPtr : OtherElt, // Dest ptr
538 SROADest ? OtherElt : EltPtr, // Src ptr
539 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
542 new CallInst(TheFn, Ops, 4, "", MI);
544 assert(isa<MemSetInst>(MI));
546 EltPtr, MI->getOperand(2), // Dest, Value,
547 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
550 new CallInst(TheFn, Ops, 4, "", MI);
554 // Finally, MI is now dead, as we've modified its actions to occur on all of
555 // the elements of the aggregate.
556 MI->eraseFromParent();
559 // The cast is dead, remove it.
560 BCInst->eraseFromParent();
564 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
565 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe,
566 /// or 1 if safe after canonicalization has been performed.
568 int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) {
569 // Loop over the use list of the alloca. We can only transform it if all of
570 // the users are safe to transform.
573 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end();
575 isSafe &= isSafeUseOfAllocation(cast<Instruction>(*I), AI);
577 DOUT << "Cannot transform: " << *AI << " due to user: " << **I;
581 // If we require cleanup, isSafe is now 1, otherwise it is 3.
585 /// CanonicalizeAllocaUsers - If SROA reported that it can promote the specified
586 /// allocation, but only if cleaned up, perform the cleanups required.
587 void SROA::CanonicalizeAllocaUsers(AllocationInst *AI) {
588 // At this point, we know that the end result will be SROA'd and promoted, so
589 // we can insert ugly code if required so long as sroa+mem2reg will clean it
591 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
593 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(*UI++);
594 gep_type_iterator I = gep_type_begin(GEPI);
597 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
598 uint64_t NumElements = AT->getNumElements();
600 if (!isa<ConstantInt>(I.getOperand())) {
601 if (NumElements == 1) {
602 GEPI->setOperand(2, Constant::getNullValue(Type::Int32Ty));
604 assert(NumElements == 2 && "Unhandled case!");
605 // All users of the GEP must be loads. At each use of the GEP, insert
606 // two loads of the appropriate indexed GEP and select between them.
607 Value *IsOne = new ICmpInst(ICmpInst::ICMP_NE, I.getOperand(),
608 Constant::getNullValue(I.getOperand()->getType()),
610 // Insert the new GEP instructions, which are properly indexed.
611 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end());
612 Indices[1] = Constant::getNullValue(Type::Int32Ty);
613 Value *ZeroIdx = new GetElementPtrInst(GEPI->getOperand(0),
614 &Indices[0], Indices.size(),
615 GEPI->getName()+".0", GEPI);
616 Indices[1] = ConstantInt::get(Type::Int32Ty, 1);
617 Value *OneIdx = new GetElementPtrInst(GEPI->getOperand(0),
618 &Indices[0], Indices.size(),
619 GEPI->getName()+".1", GEPI);
620 // Replace all loads of the variable index GEP with loads from both
621 // indexes and a select.
622 while (!GEPI->use_empty()) {
623 LoadInst *LI = cast<LoadInst>(GEPI->use_back());
624 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI);
625 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI);
626 Value *R = new SelectInst(IsOne, One, Zero, LI->getName(), LI);
627 LI->replaceAllUsesWith(R);
628 LI->eraseFromParent();
630 GEPI->eraseFromParent();
637 /// MergeInType - Add the 'In' type to the accumulated type so far. If the
638 /// types are incompatible, return true, otherwise update Accum and return
641 /// There are three cases we handle here:
642 /// 1) An effectively-integer union, where the pieces are stored into as
643 /// smaller integers (common with byte swap and other idioms).
644 /// 2) A union of vector types of the same size and potentially its elements.
645 /// Here we turn element accesses into insert/extract element operations.
646 /// 3) A union of scalar types, such as int/float or int/pointer. Here we
647 /// merge together into integers, allowing the xform to work with #1 as
649 static bool MergeInType(const Type *In, const Type *&Accum,
650 const TargetData &TD) {
651 // If this is our first type, just use it.
652 const VectorType *PTy;
653 if (Accum == Type::VoidTy || In == Accum) {
655 } else if (In == Type::VoidTy) {
657 } else if (In->isInteger() && Accum->isInteger()) { // integer union.
658 // Otherwise pick whichever type is larger.
659 if (cast<IntegerType>(In)->getBitWidth() >
660 cast<IntegerType>(Accum)->getBitWidth())
662 } else if (isa<PointerType>(In) && isa<PointerType>(Accum)) {
663 // Pointer unions just stay as one of the pointers.
664 } else if (isa<VectorType>(In) || isa<VectorType>(Accum)) {
665 if ((PTy = dyn_cast<VectorType>(Accum)) &&
666 PTy->getElementType() == In) {
667 // Accum is a vector, and we are accessing an element: ok.
668 } else if ((PTy = dyn_cast<VectorType>(In)) &&
669 PTy->getElementType() == Accum) {
670 // In is a vector, and accum is an element: ok, remember In.
672 } else if ((PTy = dyn_cast<VectorType>(In)) && isa<VectorType>(Accum) &&
673 PTy->getBitWidth() == cast<VectorType>(Accum)->getBitWidth()) {
674 // Two vectors of the same size: keep Accum.
676 // Cannot insert an short into a <4 x int> or handle
677 // <2 x int> -> <4 x int>
681 // Pointer/FP/Integer unions merge together as integers.
682 switch (Accum->getTypeID()) {
683 case Type::PointerTyID: Accum = TD.getIntPtrType(); break;
684 case Type::FloatTyID: Accum = Type::Int32Ty; break;
685 case Type::DoubleTyID: Accum = Type::Int64Ty; break;
687 assert(Accum->isInteger() && "Unknown FP type!");
691 switch (In->getTypeID()) {
692 case Type::PointerTyID: In = TD.getIntPtrType(); break;
693 case Type::FloatTyID: In = Type::Int32Ty; break;
694 case Type::DoubleTyID: In = Type::Int64Ty; break;
696 assert(In->isInteger() && "Unknown FP type!");
699 return MergeInType(In, Accum, TD);
704 /// getUIntAtLeastAsBitAs - Return an unsigned integer type that is at least
705 /// as big as the specified type. If there is no suitable type, this returns
707 const Type *getUIntAtLeastAsBitAs(unsigned NumBits) {
708 if (NumBits > 64) return 0;
709 if (NumBits > 32) return Type::Int64Ty;
710 if (NumBits > 16) return Type::Int32Ty;
711 if (NumBits > 8) return Type::Int16Ty;
715 /// CanConvertToScalar - V is a pointer. If we can convert the pointee to a
716 /// single scalar integer type, return that type. Further, if the use is not
717 /// a completely trivial use that mem2reg could promote, set IsNotTrivial. If
718 /// there are no uses of this pointer, return Type::VoidTy to differentiate from
721 const Type *SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial) {
722 const Type *UsedType = Type::VoidTy; // No uses, no forced type.
723 const TargetData &TD = getAnalysis<TargetData>();
724 const PointerType *PTy = cast<PointerType>(V->getType());
726 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
727 Instruction *User = cast<Instruction>(*UI);
729 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
730 if (MergeInType(LI->getType(), UsedType, TD))
733 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
734 // Storing the pointer, not into the value?
735 if (SI->getOperand(0) == V) return 0;
737 // NOTE: We could handle storing of FP imms into integers here!
739 if (MergeInType(SI->getOperand(0)->getType(), UsedType, TD))
741 } else if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
743 const Type *SubTy = CanConvertToScalar(CI, IsNotTrivial);
744 if (!SubTy || MergeInType(SubTy, UsedType, TD)) return 0;
745 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
746 // Check to see if this is stepping over an element: GEP Ptr, int C
747 if (GEP->getNumOperands() == 2 && isa<ConstantInt>(GEP->getOperand(1))) {
748 unsigned Idx = cast<ConstantInt>(GEP->getOperand(1))->getZExtValue();
749 unsigned ElSize = TD.getTypeSize(PTy->getElementType());
750 unsigned BitOffset = Idx*ElSize*8;
751 if (BitOffset > 64 || !isPowerOf2_32(ElSize)) return 0;
754 const Type *SubElt = CanConvertToScalar(GEP, IsNotTrivial);
755 if (SubElt == 0) return 0;
756 if (SubElt != Type::VoidTy && SubElt->isInteger()) {
758 getUIntAtLeastAsBitAs(TD.getTypeSize(SubElt)*8+BitOffset);
759 if (NewTy == 0 || MergeInType(NewTy, UsedType, TD)) return 0;
762 } else if (GEP->getNumOperands() == 3 &&
763 isa<ConstantInt>(GEP->getOperand(1)) &&
764 isa<ConstantInt>(GEP->getOperand(2)) &&
765 cast<Constant>(GEP->getOperand(1))->isNullValue()) {
766 // We are stepping into an element, e.g. a structure or an array:
767 // GEP Ptr, int 0, uint C
768 const Type *AggTy = PTy->getElementType();
769 unsigned Idx = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
771 if (const ArrayType *ATy = dyn_cast<ArrayType>(AggTy)) {
772 if (Idx >= ATy->getNumElements()) return 0; // Out of range.
773 } else if (const VectorType *VectorTy = dyn_cast<VectorType>(AggTy)) {
774 // Getting an element of the packed vector.
775 if (Idx >= VectorTy->getNumElements()) return 0; // Out of range.
777 // Merge in the vector type.
778 if (MergeInType(VectorTy, UsedType, TD)) return 0;
780 const Type *SubTy = CanConvertToScalar(GEP, IsNotTrivial);
781 if (SubTy == 0) return 0;
783 if (SubTy != Type::VoidTy && MergeInType(SubTy, UsedType, TD))
786 // We'll need to change this to an insert/extract element operation.
788 continue; // Everything looks ok
790 } else if (isa<StructType>(AggTy)) {
791 // Structs are always ok.
795 const Type *NTy = getUIntAtLeastAsBitAs(TD.getTypeSize(AggTy)*8);
796 if (NTy == 0 || MergeInType(NTy, UsedType, TD)) return 0;
797 const Type *SubTy = CanConvertToScalar(GEP, IsNotTrivial);
798 if (SubTy == 0) return 0;
799 if (SubTy != Type::VoidTy && MergeInType(SubTy, UsedType, TD))
801 continue; // Everything looks ok
805 // Cannot handle this!
813 /// ConvertToScalar - The specified alloca passes the CanConvertToScalar
814 /// predicate and is non-trivial. Convert it to something that can be trivially
815 /// promoted into a register by mem2reg.
816 void SROA::ConvertToScalar(AllocationInst *AI, const Type *ActualTy) {
817 DOUT << "CONVERT TO SCALAR: " << *AI << " TYPE = "
818 << *ActualTy << "\n";
821 BasicBlock *EntryBlock = AI->getParent();
822 assert(EntryBlock == &EntryBlock->getParent()->front() &&
823 "Not in the entry block!");
824 EntryBlock->getInstList().remove(AI); // Take the alloca out of the program.
826 // Create and insert the alloca.
827 AllocaInst *NewAI = new AllocaInst(ActualTy, 0, AI->getName(),
828 EntryBlock->begin());
829 ConvertUsesToScalar(AI, NewAI, 0);
834 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
835 /// directly. This happens when we are converting an "integer union" to a
836 /// single integer scalar, or when we are converting a "vector union" to a
837 /// vector with insert/extractelement instructions.
839 /// Offset is an offset from the original alloca, in bits that need to be
840 /// shifted to the right. By the end of this, there should be no uses of Ptr.
841 void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, unsigned Offset) {
842 bool isVectorInsert = isa<VectorType>(NewAI->getType()->getElementType());
843 const TargetData &TD = getAnalysis<TargetData>();
844 while (!Ptr->use_empty()) {
845 Instruction *User = cast<Instruction>(Ptr->use_back());
847 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
848 // The load is a bit extract from NewAI shifted right by Offset bits.
849 Value *NV = new LoadInst(NewAI, LI->getName(), LI);
850 if (NV->getType() != LI->getType()) {
851 if (const VectorType *PTy = dyn_cast<VectorType>(NV->getType())) {
852 // If the result alloca is a vector type, this is either an element
853 // access or a bitcast to another vector type.
854 if (isa<VectorType>(LI->getType())) {
855 NV = new BitCastInst(NV, LI->getType(), LI->getName(), LI);
857 // Must be an element access.
858 unsigned Elt = Offset/(TD.getTypeSize(PTy->getElementType())*8);
859 NV = new ExtractElementInst(
860 NV, ConstantInt::get(Type::Int32Ty, Elt), "tmp", LI);
862 } else if (isa<PointerType>(NV->getType())) {
863 assert(isa<PointerType>(LI->getType()));
864 // Must be ptr->ptr cast. Anything else would result in NV being
866 NV = new BitCastInst(NV, LI->getType(), LI->getName(), LI);
868 assert(NV->getType()->isInteger() && "Unknown promotion!");
869 if (Offset && Offset < TD.getTypeSize(NV->getType())*8) {
870 NV = BinaryOperator::createLShr(NV,
871 ConstantInt::get(NV->getType(), Offset),
875 // If the result is an integer, this is a trunc or bitcast.
876 if (LI->getType()->isInteger()) {
877 NV = CastInst::createTruncOrBitCast(NV, LI->getType(),
879 } else if (LI->getType()->isFloatingPoint()) {
880 // If needed, truncate the integer to the appropriate size.
881 if (NV->getType()->getPrimitiveSizeInBits() >
882 LI->getType()->getPrimitiveSizeInBits()) {
883 switch (LI->getType()->getTypeID()) {
884 default: assert(0 && "Unknown FP type!");
885 case Type::FloatTyID:
886 NV = new TruncInst(NV, Type::Int32Ty, LI->getName(), LI);
888 case Type::DoubleTyID:
889 NV = new TruncInst(NV, Type::Int64Ty, LI->getName(), LI);
894 // Then do a bitcast.
895 NV = new BitCastInst(NV, LI->getType(), LI->getName(), LI);
897 // Otherwise must be a pointer.
898 NV = new IntToPtrInst(NV, LI->getType(), LI->getName(), LI);
902 LI->replaceAllUsesWith(NV);
903 LI->eraseFromParent();
904 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
905 assert(SI->getOperand(0) != Ptr && "Consistency error!");
907 // Convert the stored type to the actual type, shift it left to insert
908 // then 'or' into place.
909 Value *SV = SI->getOperand(0);
910 const Type *AllocaType = NewAI->getType()->getElementType();
911 if (SV->getType() != AllocaType) {
912 Value *Old = new LoadInst(NewAI, NewAI->getName()+".in", SI);
914 if (const VectorType *PTy = dyn_cast<VectorType>(AllocaType)) {
915 // If the result alloca is a vector type, this is either an element
916 // access or a bitcast to another vector type.
917 if (isa<VectorType>(SV->getType())) {
918 SV = new BitCastInst(SV, AllocaType, SV->getName(), SI);
920 // Must be an element insertion.
921 unsigned Elt = Offset/(TD.getTypeSize(PTy->getElementType())*8);
922 SV = new InsertElementInst(Old, SV,
923 ConstantInt::get(Type::Int32Ty, Elt),
927 // If SV is a float, convert it to the appropriate integer type.
928 // If it is a pointer, do the same, and also handle ptr->ptr casts
930 switch (SV->getType()->getTypeID()) {
932 assert(!SV->getType()->isFloatingPoint() && "Unknown FP type!");
934 case Type::FloatTyID:
935 SV = new BitCastInst(SV, Type::Int32Ty, SV->getName(), SI);
937 case Type::DoubleTyID:
938 SV = new BitCastInst(SV, Type::Int64Ty, SV->getName(), SI);
940 case Type::PointerTyID:
941 if (isa<PointerType>(AllocaType))
942 SV = new BitCastInst(SV, AllocaType, SV->getName(), SI);
944 SV = new PtrToIntInst(SV, TD.getIntPtrType(), SV->getName(), SI);
948 unsigned SrcSize = TD.getTypeSize(SV->getType())*8;
950 // Always zero extend the value if needed.
951 if (SV->getType() != AllocaType)
952 SV = CastInst::createZExtOrBitCast(SV, AllocaType,
954 if (Offset && Offset < AllocaType->getPrimitiveSizeInBits())
955 SV = BinaryOperator::createShl(SV,
956 ConstantInt::get(SV->getType(), Offset),
957 SV->getName()+".adj", SI);
958 // Mask out the bits we are about to insert from the old value.
959 unsigned TotalBits = TD.getTypeSize(SV->getType())*8;
960 if (TotalBits != SrcSize) {
961 assert(TotalBits > SrcSize);
962 uint64_t Mask = ~(((1ULL << SrcSize)-1) << Offset);
963 Mask = Mask & cast<IntegerType>(SV->getType())->getBitMask();
964 Old = BinaryOperator::createAnd(Old,
965 ConstantInt::get(Old->getType(), Mask),
966 Old->getName()+".mask", SI);
967 SV = BinaryOperator::createOr(Old, SV, SV->getName()+".ins", SI);
971 new StoreInst(SV, NewAI, SI);
972 SI->eraseFromParent();
974 } else if (CastInst *CI = dyn_cast<CastInst>(User)) {
975 unsigned NewOff = Offset;
976 const TargetData &TD = getAnalysis<TargetData>();
977 if (TD.isBigEndian() && !isVectorInsert) {
978 // Adjust the pointer. For example, storing 16-bits into a 32-bit
979 // alloca with just a cast makes it modify the top 16-bits.
980 const Type *SrcTy = cast<PointerType>(Ptr->getType())->getElementType();
981 const Type *DstTy = cast<PointerType>(CI->getType())->getElementType();
982 int PtrDiffBits = TD.getTypeSize(SrcTy)*8-TD.getTypeSize(DstTy)*8;
983 NewOff += PtrDiffBits;
985 ConvertUsesToScalar(CI, NewAI, NewOff);
986 CI->eraseFromParent();
987 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
988 const PointerType *AggPtrTy =
989 cast<PointerType>(GEP->getOperand(0)->getType());
990 const TargetData &TD = getAnalysis<TargetData>();
991 unsigned AggSizeInBits = TD.getTypeSize(AggPtrTy->getElementType())*8;
993 // Check to see if this is stepping over an element: GEP Ptr, int C
994 unsigned NewOffset = Offset;
995 if (GEP->getNumOperands() == 2) {
996 unsigned Idx = cast<ConstantInt>(GEP->getOperand(1))->getZExtValue();
997 unsigned BitOffset = Idx*AggSizeInBits;
999 if (TD.isLittleEndian() || isVectorInsert)
1000 NewOffset += BitOffset;
1002 NewOffset -= BitOffset;
1004 } else if (GEP->getNumOperands() == 3) {
1005 // We know that operand #2 is zero.
1006 unsigned Idx = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
1007 const Type *AggTy = AggPtrTy->getElementType();
1008 if (const SequentialType *SeqTy = dyn_cast<SequentialType>(AggTy)) {
1009 unsigned ElSizeBits = TD.getTypeSize(SeqTy->getElementType())*8;
1011 if (TD.isLittleEndian() || isVectorInsert)
1012 NewOffset += ElSizeBits*Idx;
1014 NewOffset += AggSizeInBits-ElSizeBits*(Idx+1);
1015 } else if (const StructType *STy = dyn_cast<StructType>(AggTy)) {
1016 unsigned EltBitOffset =
1017 TD.getStructLayout(STy)->getElementOffset(Idx)*8;
1019 if (TD.isLittleEndian() || isVectorInsert)
1020 NewOffset += EltBitOffset;
1022 const PointerType *ElPtrTy = cast<PointerType>(GEP->getType());
1023 unsigned ElSizeBits = TD.getTypeSize(ElPtrTy->getElementType())*8;
1024 NewOffset += AggSizeInBits-(EltBitOffset+ElSizeBits);
1028 assert(0 && "Unsupported operation!");
1032 assert(0 && "Unsupported operation!");
1035 ConvertUsesToScalar(GEP, NewAI, NewOffset);
1036 GEP->eraseFromParent();
1038 assert(0 && "Unsupported operation!");