1 //===- SROA.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. It tries to identify promotable elements of an
12 /// aggregate alloca, and promote them to registers. It will also try to
13 /// convert uses of an element (or set of elements) of an alloca into a vector
14 /// or bitfield-style integer scalar if appropriate.
16 /// It works to do this with minimal slicing of the alloca so that regions
17 /// which are merely transferred in and out of external memory remain unchanged
18 /// and are not decomposed to scalar code.
20 /// Because this also performs alloca promotion, it can be thought of as also
21 /// serving the purpose of SSA formation. The algorithm iterates on the
22 /// function until all opportunities for promotion have been realized.
24 //===----------------------------------------------------------------------===//
26 #define DEBUG_TYPE "sroa"
27 #include "llvm/Transforms/Scalar.h"
28 #include "llvm/ADT/STLExtras.h"
29 #include "llvm/ADT/SetVector.h"
30 #include "llvm/ADT/SmallVector.h"
31 #include "llvm/ADT/Statistic.h"
32 #include "llvm/Analysis/Dominators.h"
33 #include "llvm/Analysis/Loads.h"
34 #include "llvm/Analysis/PtrUseVisitor.h"
35 #include "llvm/Analysis/ValueTracking.h"
36 #include "llvm/DIBuilder.h"
37 #include "llvm/DebugInfo.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/IR/DerivedTypes.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/IRBuilder.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/IntrinsicInst.h"
45 #include "llvm/IR/LLVMContext.h"
46 #include "llvm/IR/Operator.h"
47 #include "llvm/InstVisitor.h"
48 #include "llvm/Pass.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Compiler.h"
51 #include "llvm/Support/Debug.h"
52 #include "llvm/Support/ErrorHandling.h"
53 #include "llvm/Support/MathExtras.h"
54 #include "llvm/Support/raw_ostream.h"
55 #include "llvm/Transforms/Utils/Local.h"
56 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
57 #include "llvm/Transforms/Utils/SSAUpdater.h"
60 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
61 STATISTIC(NumAllocaPartitions, "Number of alloca partitions formed");
62 STATISTIC(MaxPartitionsPerAlloca, "Maximum number of partitions");
63 STATISTIC(NumAllocaPartitionUses, "Number of alloca partition uses found");
64 STATISTIC(MaxPartitionUsesPerAlloca, "Maximum number of partition uses");
65 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
66 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
67 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
68 STATISTIC(NumDeleted, "Number of instructions deleted");
69 STATISTIC(NumVectorized, "Number of vectorized aggregates");
71 /// Hidden option to force the pass to not use DomTree and mem2reg, instead
72 /// forming SSA values through the SSAUpdater infrastructure.
74 ForceSSAUpdater("force-ssa-updater", cl::init(false), cl::Hidden);
77 /// \brief A custom IRBuilder inserter which prefixes all names if they are
79 template <bool preserveNames = true>
80 class IRBuilderPrefixedInserter :
81 public IRBuilderDefaultInserter<preserveNames> {
85 void SetNamePrefix(const Twine &P) { Prefix = P.str(); }
88 void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
89 BasicBlock::iterator InsertPt) const {
90 IRBuilderDefaultInserter<preserveNames>::InsertHelper(
91 I, Name.isTriviallyEmpty() ? Name : Prefix + Name, BB, InsertPt);
95 // Specialization for not preserving the name is trivial.
97 class IRBuilderPrefixedInserter<false> :
98 public IRBuilderDefaultInserter<false> {
100 void SetNamePrefix(const Twine &P) {}
103 /// \brief Provide a typedef for IRBuilder that drops names in release builds.
105 typedef llvm::IRBuilder<true, ConstantFolder,
106 IRBuilderPrefixedInserter<true> > IRBuilderTy;
108 typedef llvm::IRBuilder<false, ConstantFolder,
109 IRBuilderPrefixedInserter<false> > IRBuilderTy;
114 /// \brief A used slice of an alloca.
116 /// This structure represents a slice of an alloca used by some instruction. It
117 /// stores both the begin and end offsets of this use, a pointer to the use
118 /// itself, and a flag indicating whether we can classify the use as splittable
119 /// or not when forming partitions of the alloca.
121 /// \brief The beginning offset of the range.
122 uint64_t BeginOffset;
124 /// \brief The ending offset, not included in the range.
127 /// \brief Storage for both the use of this slice and whether it can be
129 PointerIntPair<Use *, 1, bool> UseAndIsSplittable;
132 Slice() : BeginOffset(), EndOffset() {}
133 Slice(uint64_t BeginOffset, uint64_t EndOffset, Use *U, bool IsSplittable)
134 : BeginOffset(BeginOffset), EndOffset(EndOffset),
135 UseAndIsSplittable(U, IsSplittable) {}
137 uint64_t beginOffset() const { return BeginOffset; }
138 uint64_t endOffset() const { return EndOffset; }
140 bool isSplittable() const { return UseAndIsSplittable.getInt(); }
141 void makeUnsplittable() { UseAndIsSplittable.setInt(false); }
143 Use *getUse() const { return UseAndIsSplittable.getPointer(); }
145 bool isDead() const { return getUse() == 0; }
146 void kill() { UseAndIsSplittable.setPointer(0); }
148 /// \brief Support for ordering ranges.
150 /// This provides an ordering over ranges such that start offsets are
151 /// always increasing, and within equal start offsets, the end offsets are
152 /// decreasing. Thus the spanning range comes first in a cluster with the
153 /// same start position.
154 bool operator<(const Slice &RHS) const {
155 if (beginOffset() < RHS.beginOffset()) return true;
156 if (beginOffset() > RHS.beginOffset()) return false;
157 if (isSplittable() != RHS.isSplittable()) return !isSplittable();
158 if (endOffset() > RHS.endOffset()) return true;
162 /// \brief Support comparison with a single offset to allow binary searches.
163 friend LLVM_ATTRIBUTE_UNUSED bool operator<(const Slice &LHS,
164 uint64_t RHSOffset) {
165 return LHS.beginOffset() < RHSOffset;
167 friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
169 return LHSOffset < RHS.beginOffset();
172 bool operator==(const Slice &RHS) const {
173 return isSplittable() == RHS.isSplittable() &&
174 beginOffset() == RHS.beginOffset() && endOffset() == RHS.endOffset();
176 bool operator!=(const Slice &RHS) const { return !operator==(RHS); }
178 } // end anonymous namespace
181 template <typename T> struct isPodLike;
182 template <> struct isPodLike<Slice> {
183 static const bool value = true;
188 /// \brief Representation of the alloca slices.
190 /// This class represents the slices of an alloca which are formed by its
191 /// various uses. If a pointer escapes, we can't fully build a representation
192 /// for the slices used and we reflect that in this structure. The uses are
193 /// stored, sorted by increasing beginning offset and with unsplittable slices
194 /// starting at a particular offset before splittable slices.
197 /// \brief Construct the slices of a particular alloca.
198 AllocaSlices(const DataLayout &DL, AllocaInst &AI);
200 /// \brief Test whether a pointer to the allocation escapes our analysis.
202 /// If this is true, the slices are never fully built and should be
204 bool isEscaped() const { return PointerEscapingInstr; }
206 /// \brief Support for iterating over the slices.
208 typedef SmallVectorImpl<Slice>::iterator iterator;
209 iterator begin() { return Slices.begin(); }
210 iterator end() { return Slices.end(); }
212 typedef SmallVectorImpl<Slice>::const_iterator const_iterator;
213 const_iterator begin() const { return Slices.begin(); }
214 const_iterator end() const { return Slices.end(); }
217 /// \brief Allow iterating the dead users for this alloca.
219 /// These are instructions which will never actually use the alloca as they
220 /// are outside the allocated range. They are safe to replace with undef and
223 typedef SmallVectorImpl<Instruction *>::const_iterator dead_user_iterator;
224 dead_user_iterator dead_user_begin() const { return DeadUsers.begin(); }
225 dead_user_iterator dead_user_end() const { return DeadUsers.end(); }
228 /// \brief Allow iterating the dead expressions referring to this alloca.
230 /// These are operands which have cannot actually be used to refer to the
231 /// alloca as they are outside its range and the user doesn't correct for
232 /// that. These mostly consist of PHI node inputs and the like which we just
233 /// need to replace with undef.
235 typedef SmallVectorImpl<Use *>::const_iterator dead_op_iterator;
236 dead_op_iterator dead_op_begin() const { return DeadOperands.begin(); }
237 dead_op_iterator dead_op_end() const { return DeadOperands.end(); }
240 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
241 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
242 void printSlice(raw_ostream &OS, const_iterator I,
243 StringRef Indent = " ") const;
244 void printUse(raw_ostream &OS, const_iterator I,
245 StringRef Indent = " ") const;
246 void print(raw_ostream &OS) const;
247 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump(const_iterator I) const;
248 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump() const;
252 template <typename DerivedT, typename RetT = void> class BuilderBase;
254 friend class AllocaSlices::SliceBuilder;
256 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
257 /// \brief Handle to alloca instruction to simplify method interfaces.
261 /// \brief The instruction responsible for this alloca not having a known set
264 /// When an instruction (potentially) escapes the pointer to the alloca, we
265 /// store a pointer to that here and abort trying to form slices of the
266 /// alloca. This will be null if the alloca slices are analyzed successfully.
267 Instruction *PointerEscapingInstr;
269 /// \brief The slices of the alloca.
271 /// We store a vector of the slices formed by uses of the alloca here. This
272 /// vector is sorted by increasing begin offset, and then the unsplittable
273 /// slices before the splittable ones. See the Slice inner class for more
275 SmallVector<Slice, 8> Slices;
277 /// \brief Instructions which will become dead if we rewrite the alloca.
279 /// Note that these are not separated by slice. This is because we expect an
280 /// alloca to be completely rewritten or not rewritten at all. If rewritten,
281 /// all these instructions can simply be removed and replaced with undef as
282 /// they come from outside of the allocated space.
283 SmallVector<Instruction *, 8> DeadUsers;
285 /// \brief Operands which will become dead if we rewrite the alloca.
287 /// These are operands that in their particular use can be replaced with
288 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
289 /// to PHI nodes and the like. They aren't entirely dead (there might be
290 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
291 /// want to swap this particular input for undef to simplify the use lists of
293 SmallVector<Use *, 8> DeadOperands;
297 static Value *foldSelectInst(SelectInst &SI) {
298 // If the condition being selected on is a constant or the same value is
299 // being selected between, fold the select. Yes this does (rarely) happen
301 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
302 return SI.getOperand(1+CI->isZero());
303 if (SI.getOperand(1) == SI.getOperand(2))
304 return SI.getOperand(1);
309 /// \brief Builder for the alloca slices.
311 /// This class builds a set of alloca slices by recursively visiting the uses
312 /// of an alloca and making a slice for each load and store at each offset.
313 class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
314 friend class PtrUseVisitor<SliceBuilder>;
315 friend class InstVisitor<SliceBuilder>;
316 typedef PtrUseVisitor<SliceBuilder> Base;
318 const uint64_t AllocSize;
321 SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
322 SmallDenseMap<Instruction *, uint64_t> PHIOrSelectSizes;
324 /// \brief Set to de-duplicate dead instructions found in the use walk.
325 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
328 SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &S)
329 : PtrUseVisitor<SliceBuilder>(DL),
330 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())), S(S) {}
333 void markAsDead(Instruction &I) {
334 if (VisitedDeadInsts.insert(&I))
335 S.DeadUsers.push_back(&I);
338 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
339 bool IsSplittable = false) {
340 // Completely skip uses which have a zero size or start either before or
341 // past the end of the allocation.
342 if (Size == 0 || Offset.isNegative() || Offset.uge(AllocSize)) {
343 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
344 << " which has zero size or starts outside of the "
345 << AllocSize << " byte alloca:\n"
346 << " alloca: " << S.AI << "\n"
347 << " use: " << I << "\n");
348 return markAsDead(I);
351 uint64_t BeginOffset = Offset.getZExtValue();
352 uint64_t EndOffset = BeginOffset + Size;
354 // Clamp the end offset to the end of the allocation. Note that this is
355 // formulated to handle even the case where "BeginOffset + Size" overflows.
356 // This may appear superficially to be something we could ignore entirely,
357 // but that is not so! There may be widened loads or PHI-node uses where
358 // some instructions are dead but not others. We can't completely ignore
359 // them, and so have to record at least the information here.
360 assert(AllocSize >= BeginOffset); // Established above.
361 if (Size > AllocSize - BeginOffset) {
362 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
363 << " to remain within the " << AllocSize << " byte alloca:\n"
364 << " alloca: " << S.AI << "\n"
365 << " use: " << I << "\n");
366 EndOffset = AllocSize;
369 S.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
372 void visitBitCastInst(BitCastInst &BC) {
374 return markAsDead(BC);
376 return Base::visitBitCastInst(BC);
379 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
380 if (GEPI.use_empty())
381 return markAsDead(GEPI);
383 return Base::visitGetElementPtrInst(GEPI);
386 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
387 uint64_t Size, bool IsVolatile) {
388 // We allow splitting of loads and stores where the type is an integer type
389 // and cover the entire alloca. This prevents us from splitting over
391 // FIXME: In the great blue eventually, we should eagerly split all integer
392 // loads and stores, and then have a separate step that merges adjacent
393 // alloca partitions into a single partition suitable for integer widening.
394 // Or we should skip the merge step and rely on GVN and other passes to
395 // merge adjacent loads and stores that survive mem2reg.
397 Ty->isIntegerTy() && !IsVolatile && Offset == 0 && Size >= AllocSize;
399 insertUse(I, Offset, Size, IsSplittable);
402 void visitLoadInst(LoadInst &LI) {
403 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
404 "All simple FCA loads should have been pre-split");
407 return PI.setAborted(&LI);
409 uint64_t Size = DL.getTypeStoreSize(LI.getType());
410 return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
413 void visitStoreInst(StoreInst &SI) {
414 Value *ValOp = SI.getValueOperand();
416 return PI.setEscapedAndAborted(&SI);
418 return PI.setAborted(&SI);
420 uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
422 // If this memory access can be shown to *statically* extend outside the
423 // bounds of of the allocation, it's behavior is undefined, so simply
424 // ignore it. Note that this is more strict than the generic clamping
425 // behavior of insertUse. We also try to handle cases which might run the
427 // FIXME: We should instead consider the pointer to have escaped if this
428 // function is being instrumented for addressing bugs or race conditions.
429 if (Offset.isNegative() || Size > AllocSize ||
430 Offset.ugt(AllocSize - Size)) {
431 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @" << Offset
432 << " which extends past the end of the " << AllocSize
434 << " alloca: " << S.AI << "\n"
435 << " use: " << SI << "\n");
436 return markAsDead(SI);
439 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
440 "All simple FCA stores should have been pre-split");
441 handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
445 void visitMemSetInst(MemSetInst &II) {
446 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
447 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
448 if ((Length && Length->getValue() == 0) ||
449 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize)))
450 // Zero-length mem transfer intrinsics can be ignored entirely.
451 return markAsDead(II);
454 return PI.setAborted(&II);
456 insertUse(II, Offset,
457 Length ? Length->getLimitedValue()
458 : AllocSize - Offset.getLimitedValue(),
462 void visitMemTransferInst(MemTransferInst &II) {
463 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
464 if ((Length && Length->getValue() == 0) ||
465 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize)))
466 // Zero-length mem transfer intrinsics can be ignored entirely.
467 return markAsDead(II);
470 return PI.setAborted(&II);
472 uint64_t RawOffset = Offset.getLimitedValue();
473 uint64_t Size = Length ? Length->getLimitedValue()
474 : AllocSize - RawOffset;
476 // Check for the special case where the same exact value is used for both
478 if (*U == II.getRawDest() && *U == II.getRawSource()) {
479 // For non-volatile transfers this is a no-op.
480 if (!II.isVolatile())
481 return markAsDead(II);
483 return insertUse(II, Offset, Size, /*IsSplittable=*/false);;
486 // If we have seen both source and destination for a mem transfer, then
487 // they both point to the same alloca.
489 SmallDenseMap<Instruction *, unsigned>::iterator MTPI;
490 llvm::tie(MTPI, Inserted) =
491 MemTransferSliceMap.insert(std::make_pair(&II, S.Slices.size()));
492 unsigned PrevIdx = MTPI->second;
494 Slice &PrevP = S.Slices[PrevIdx];
496 // Check if the begin offsets match and this is a non-volatile transfer.
497 // In that case, we can completely elide the transfer.
498 if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
500 return markAsDead(II);
503 // Otherwise we have an offset transfer within the same alloca. We can't
505 PrevP.makeUnsplittable();
508 // Insert the use now that we've fixed up the splittable nature.
509 insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
511 // Check that we ended up with a valid index in the map.
512 assert(S.Slices[PrevIdx].getUse()->getUser() == &II &&
513 "Map index doesn't point back to a slice with this user.");
516 // Disable SRoA for any intrinsics except for lifetime invariants.
517 // FIXME: What about debug intrinsics? This matches old behavior, but
518 // doesn't make sense.
519 void visitIntrinsicInst(IntrinsicInst &II) {
521 return PI.setAborted(&II);
523 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
524 II.getIntrinsicID() == Intrinsic::lifetime_end) {
525 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
526 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
527 Length->getLimitedValue());
528 insertUse(II, Offset, Size, true);
532 Base::visitIntrinsicInst(II);
535 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
536 // We consider any PHI or select that results in a direct load or store of
537 // the same offset to be a viable use for slicing purposes. These uses
538 // are considered unsplittable and the size is the maximum loaded or stored
540 SmallPtrSet<Instruction *, 4> Visited;
541 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
542 Visited.insert(Root);
543 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
544 // If there are no loads or stores, the access is dead. We mark that as
545 // a size zero access.
548 Instruction *I, *UsedI;
549 llvm::tie(UsedI, I) = Uses.pop_back_val();
551 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
552 Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
555 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
556 Value *Op = SI->getOperand(0);
559 Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
563 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
564 if (!GEP->hasAllZeroIndices())
566 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
567 !isa<SelectInst>(I)) {
571 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); UI != UE;
573 if (Visited.insert(cast<Instruction>(*UI)))
574 Uses.push_back(std::make_pair(I, cast<Instruction>(*UI)));
575 } while (!Uses.empty());
580 void visitPHINode(PHINode &PN) {
582 return markAsDead(PN);
584 return PI.setAborted(&PN);
586 // See if we already have computed info on this node.
587 uint64_t &PHISize = PHIOrSelectSizes[&PN];
589 // This is a new PHI node, check for an unsafe use of the PHI node.
590 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&PN, PHISize))
591 return PI.setAborted(UnsafeI);
594 // For PHI and select operands outside the alloca, we can't nuke the entire
595 // phi or select -- the other side might still be relevant, so we special
596 // case them here and use a separate structure to track the operands
597 // themselves which should be replaced with undef.
598 // FIXME: This should instead be escaped in the event we're instrumenting
599 // for address sanitization.
600 if ((Offset.isNegative() && (-Offset).uge(PHISize)) ||
601 (!Offset.isNegative() && Offset.uge(AllocSize))) {
602 S.DeadOperands.push_back(U);
606 insertUse(PN, Offset, PHISize);
609 void visitSelectInst(SelectInst &SI) {
611 return markAsDead(SI);
612 if (Value *Result = foldSelectInst(SI)) {
614 // If the result of the constant fold will be the pointer, recurse
615 // through the select as if we had RAUW'ed it.
618 // Otherwise the operand to the select is dead, and we can replace it
620 S.DeadOperands.push_back(U);
625 return PI.setAborted(&SI);
627 // See if we already have computed info on this node.
628 uint64_t &SelectSize = PHIOrSelectSizes[&SI];
630 // This is a new Select, check for an unsafe use of it.
631 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&SI, SelectSize))
632 return PI.setAborted(UnsafeI);
635 // For PHI and select operands outside the alloca, we can't nuke the entire
636 // phi or select -- the other side might still be relevant, so we special
637 // case them here and use a separate structure to track the operands
638 // themselves which should be replaced with undef.
639 // FIXME: This should instead be escaped in the event we're instrumenting
640 // for address sanitization.
641 if ((Offset.isNegative() && Offset.uge(SelectSize)) ||
642 (!Offset.isNegative() && Offset.uge(AllocSize))) {
643 S.DeadOperands.push_back(U);
647 insertUse(SI, Offset, SelectSize);
650 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
651 void visitInstruction(Instruction &I) {
658 bool operator()(const Slice &S) { return S.isDead(); }
662 AllocaSlices::AllocaSlices(const DataLayout &DL, AllocaInst &AI)
664 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
667 PointerEscapingInstr(0) {
668 SliceBuilder PB(DL, AI, *this);
669 SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
670 if (PtrI.isEscaped() || PtrI.isAborted()) {
671 // FIXME: We should sink the escape vs. abort info into the caller nicely,
672 // possibly by just storing the PtrInfo in the AllocaSlices.
673 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
674 : PtrI.getAbortingInst();
675 assert(PointerEscapingInstr && "Did not track a bad instruction");
679 // Sort the uses. This arranges for the offsets to be in ascending order,
680 // and the sizes to be in descending order.
681 std::sort(Slices.begin(), Slices.end());
683 Slices.erase(std::remove_if(Slices.begin(), Slices.end(), IsSliceDead()),
686 // Record how many slices we end up with.
687 NumAllocaPartitions += Slices.size();
688 MaxPartitionsPerAlloca =
689 std::max<unsigned>(Slices.size(), MaxPartitionsPerAlloca);
691 NumAllocaPartitionUses += Slices.size();
692 MaxPartitionUsesPerAlloca =
693 std::max<unsigned>(Slices.size(), MaxPartitionUsesPerAlloca);
696 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
698 void AllocaSlices::print(raw_ostream &OS, const_iterator I,
699 StringRef Indent) const {
700 printSlice(OS, I, Indent);
701 printUse(OS, I, Indent);
704 void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
705 StringRef Indent) const {
706 OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
707 << " slice #" << (I - begin())
708 << (I->isSplittable() ? " (splittable)" : "") << "\n";
711 void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
712 StringRef Indent) const {
713 OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
716 void AllocaSlices::print(raw_ostream &OS) const {
717 if (PointerEscapingInstr) {
718 OS << "Can't analyze slices for alloca: " << AI << "\n"
719 << " A pointer to this alloca escaped by:\n"
720 << " " << *PointerEscapingInstr << "\n";
724 OS << "Slices of alloca: " << AI << "\n";
725 for (const_iterator I = begin(), E = end(); I != E; ++I)
729 void AllocaSlices::dump(const_iterator I) const { print(dbgs(), I); }
730 void AllocaSlices::dump() const { print(dbgs()); }
732 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
735 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
737 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
738 /// the loads and stores of an alloca instruction, as well as updating its
739 /// debug information. This is used when a domtree is unavailable and thus
740 /// mem2reg in its full form can't be used to handle promotion of allocas to
742 class AllocaPromoter : public LoadAndStorePromoter {
746 SmallVector<DbgDeclareInst *, 4> DDIs;
747 SmallVector<DbgValueInst *, 4> DVIs;
750 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
751 AllocaInst &AI, DIBuilder &DIB)
752 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
754 void run(const SmallVectorImpl<Instruction*> &Insts) {
755 // Remember which alloca we're promoting (for isInstInList).
756 if (MDNode *DebugNode = MDNode::getIfExists(AI.getContext(), &AI)) {
757 for (Value::use_iterator UI = DebugNode->use_begin(),
758 UE = DebugNode->use_end();
760 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
762 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI))
766 LoadAndStorePromoter::run(Insts);
767 AI.eraseFromParent();
768 while (!DDIs.empty())
769 DDIs.pop_back_val()->eraseFromParent();
770 while (!DVIs.empty())
771 DVIs.pop_back_val()->eraseFromParent();
774 virtual bool isInstInList(Instruction *I,
775 const SmallVectorImpl<Instruction*> &Insts) const {
776 if (LoadInst *LI = dyn_cast<LoadInst>(I))
777 return LI->getOperand(0) == &AI;
778 return cast<StoreInst>(I)->getPointerOperand() == &AI;
781 virtual void updateDebugInfo(Instruction *Inst) const {
782 for (SmallVectorImpl<DbgDeclareInst *>::const_iterator I = DDIs.begin(),
783 E = DDIs.end(); I != E; ++I) {
784 DbgDeclareInst *DDI = *I;
785 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
786 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
787 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
788 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
790 for (SmallVectorImpl<DbgValueInst *>::const_iterator I = DVIs.begin(),
791 E = DVIs.end(); I != E; ++I) {
792 DbgValueInst *DVI = *I;
794 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
795 // If an argument is zero extended then use argument directly. The ZExt
796 // may be zapped by an optimization pass in future.
797 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
798 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
799 else if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
800 Arg = dyn_cast<Argument>(SExt->getOperand(0));
802 Arg = SI->getValueOperand();
803 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
804 Arg = LI->getPointerOperand();
808 Instruction *DbgVal =
809 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
811 DbgVal->setDebugLoc(DVI->getDebugLoc());
815 } // end anon namespace
819 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
821 /// This pass takes allocations which can be completely analyzed (that is, they
822 /// don't escape) and tries to turn them into scalar SSA values. There are
823 /// a few steps to this process.
825 /// 1) It takes allocations of aggregates and analyzes the ways in which they
826 /// are used to try to split them into smaller allocations, ideally of
827 /// a single scalar data type. It will split up memcpy and memset accesses
828 /// as necessary and try to isolate individual scalar accesses.
829 /// 2) It will transform accesses into forms which are suitable for SSA value
830 /// promotion. This can be replacing a memset with a scalar store of an
831 /// integer value, or it can involve speculating operations on a PHI or
832 /// select to be a PHI or select of the results.
833 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
834 /// onto insert and extract operations on a vector value, and convert them to
835 /// this form. By doing so, it will enable promotion of vector aggregates to
836 /// SSA vector values.
837 class SROA : public FunctionPass {
838 const bool RequiresDomTree;
841 const DataLayout *DL;
844 /// \brief Worklist of alloca instructions to simplify.
846 /// Each alloca in the function is added to this. Each new alloca formed gets
847 /// added to it as well to recursively simplify unless that alloca can be
848 /// directly promoted. Finally, each time we rewrite a use of an alloca other
849 /// the one being actively rewritten, we add it back onto the list if not
850 /// already present to ensure it is re-visited.
851 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > Worklist;
853 /// \brief A collection of instructions to delete.
854 /// We try to batch deletions to simplify code and make things a bit more
856 SetVector<Instruction *, SmallVector<Instruction *, 8> > DeadInsts;
858 /// \brief Post-promotion worklist.
860 /// Sometimes we discover an alloca which has a high probability of becoming
861 /// viable for SROA after a round of promotion takes place. In those cases,
862 /// the alloca is enqueued here for re-processing.
864 /// Note that we have to be very careful to clear allocas out of this list in
865 /// the event they are deleted.
866 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > PostPromotionWorklist;
868 /// \brief A collection of alloca instructions we can directly promote.
869 std::vector<AllocaInst *> PromotableAllocas;
871 /// \brief A worklist of PHIs to speculate prior to promoting allocas.
873 /// All of these PHIs have been checked for the safety of speculation and by
874 /// being speculated will allow promoting allocas currently in the promotable
876 SetVector<PHINode *, SmallVector<PHINode *, 2> > SpeculatablePHIs;
878 /// \brief A worklist of select instructions to speculate prior to promoting
881 /// All of these select instructions have been checked for the safety of
882 /// speculation and by being speculated will allow promoting allocas
883 /// currently in the promotable queue.
884 SetVector<SelectInst *, SmallVector<SelectInst *, 2> > SpeculatableSelects;
887 SROA(bool RequiresDomTree = true)
888 : FunctionPass(ID), RequiresDomTree(RequiresDomTree),
890 initializeSROAPass(*PassRegistry::getPassRegistry());
892 bool runOnFunction(Function &F);
893 void getAnalysisUsage(AnalysisUsage &AU) const;
895 const char *getPassName() const { return "SROA"; }
899 friend class PHIOrSelectSpeculator;
900 friend class AllocaSliceRewriter;
902 bool rewritePartition(AllocaInst &AI, AllocaSlices &S,
903 AllocaSlices::iterator B, AllocaSlices::iterator E,
904 int64_t BeginOffset, int64_t EndOffset,
905 ArrayRef<AllocaSlices::iterator> SplitUses);
906 bool splitAlloca(AllocaInst &AI, AllocaSlices &S);
907 bool runOnAlloca(AllocaInst &AI);
908 void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas);
909 bool promoteAllocas(Function &F);
915 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
916 return new SROA(RequiresDomTree);
919 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates",
921 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
922 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
925 /// Walk the range of a partitioning looking for a common type to cover this
926 /// sequence of slices.
927 static Type *findCommonType(AllocaSlices::const_iterator B,
928 AllocaSlices::const_iterator E,
929 uint64_t EndOffset) {
931 for (AllocaSlices::const_iterator I = B; I != E; ++I) {
932 Use *U = I->getUse();
933 if (isa<IntrinsicInst>(*U->getUser()))
935 if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
939 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser()))
940 UserTy = LI->getType();
941 else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser()))
942 UserTy = SI->getValueOperand()->getType();
944 return 0; // Bail if we have weird uses.
946 if (IntegerType *ITy = dyn_cast<IntegerType>(UserTy)) {
947 // If the type is larger than the partition, skip it. We only encounter
948 // this for split integer operations where we want to use the type of the
949 // entity causing the split.
950 if (ITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
953 // If we have found an integer type use covering the alloca, use that
954 // regardless of the other types, as integers are often used for a
960 if (Ty && Ty != UserTy)
968 /// PHI instructions that use an alloca and are subsequently loaded can be
969 /// rewritten to load both input pointers in the pred blocks and then PHI the
970 /// results, allowing the load of the alloca to be promoted.
972 /// %P2 = phi [i32* %Alloca, i32* %Other]
973 /// %V = load i32* %P2
975 /// %V1 = load i32* %Alloca -> will be mem2reg'd
977 /// %V2 = load i32* %Other
979 /// %V = phi [i32 %V1, i32 %V2]
981 /// We can do this to a select if its only uses are loads and if the operands
982 /// to the select can be loaded unconditionally.
984 /// FIXME: This should be hoisted into a generic utility, likely in
985 /// Transforms/Util/Local.h
986 static bool isSafePHIToSpeculate(PHINode &PN,
987 const DataLayout *DL = 0) {
988 // For now, we can only do this promotion if the load is in the same block
989 // as the PHI, and if there are no stores between the phi and load.
990 // TODO: Allow recursive phi users.
991 // TODO: Allow stores.
992 BasicBlock *BB = PN.getParent();
993 unsigned MaxAlign = 0;
994 bool HaveLoad = false;
995 for (Value::use_iterator UI = PN.use_begin(), UE = PN.use_end(); UI != UE;
997 LoadInst *LI = dyn_cast<LoadInst>(*UI);
998 if (LI == 0 || !LI->isSimple())
1001 // For now we only allow loads in the same block as the PHI. This is
1002 // a common case that happens when instcombine merges two loads through
1004 if (LI->getParent() != BB)
1007 // Ensure that there are no instructions between the PHI and the load that
1009 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1010 if (BBI->mayWriteToMemory())
1013 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1020 // We can only transform this if it is safe to push the loads into the
1021 // predecessor blocks. The only thing to watch out for is that we can't put
1022 // a possibly trapping load in the predecessor if it is a critical edge.
1023 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1024 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1025 Value *InVal = PN.getIncomingValue(Idx);
1027 // If the value is produced by the terminator of the predecessor (an
1028 // invoke) or it has side-effects, there is no valid place to put a load
1029 // in the predecessor.
1030 if (TI == InVal || TI->mayHaveSideEffects())
1033 // If the predecessor has a single successor, then the edge isn't
1035 if (TI->getNumSuccessors() == 1)
1038 // If this pointer is always safe to load, or if we can prove that there
1039 // is already a load in the block, then we can move the load to the pred
1041 if (InVal->isDereferenceablePointer() ||
1042 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, DL))
1051 static void speculatePHINodeLoads(PHINode &PN) {
1052 DEBUG(dbgs() << " original: " << PN << "\n");
1054 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1055 IRBuilderTy PHIBuilder(&PN);
1056 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1057 PN.getName() + ".sroa.speculated");
1059 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1060 // matter which one we get and if any differ.
1061 LoadInst *SomeLoad = cast<LoadInst>(*PN.use_begin());
1062 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1063 unsigned Align = SomeLoad->getAlignment();
1065 // Rewrite all loads of the PN to use the new PHI.
1066 while (!PN.use_empty()) {
1067 LoadInst *LI = cast<LoadInst>(*PN.use_begin());
1068 LI->replaceAllUsesWith(NewPN);
1069 LI->eraseFromParent();
1072 // Inject loads into all of the pred blocks.
1073 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1074 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1075 TerminatorInst *TI = Pred->getTerminator();
1076 Value *InVal = PN.getIncomingValue(Idx);
1077 IRBuilderTy PredBuilder(TI);
1079 LoadInst *Load = PredBuilder.CreateLoad(
1080 InVal, (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1081 ++NumLoadsSpeculated;
1082 Load->setAlignment(Align);
1084 Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1085 NewPN->addIncoming(Load, Pred);
1088 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1089 PN.eraseFromParent();
1092 /// Select instructions that use an alloca and are subsequently loaded can be
1093 /// rewritten to load both input pointers and then select between the result,
1094 /// allowing the load of the alloca to be promoted.
1096 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1097 /// %V = load i32* %P2
1099 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1100 /// %V2 = load i32* %Other
1101 /// %V = select i1 %cond, i32 %V1, i32 %V2
1103 /// We can do this to a select if its only uses are loads and if the operand
1104 /// to the select can be loaded unconditionally.
1105 static bool isSafeSelectToSpeculate(SelectInst &SI, const DataLayout *DL = 0) {
1106 Value *TValue = SI.getTrueValue();
1107 Value *FValue = SI.getFalseValue();
1108 bool TDerefable = TValue->isDereferenceablePointer();
1109 bool FDerefable = FValue->isDereferenceablePointer();
1111 for (Value::use_iterator UI = SI.use_begin(), UE = SI.use_end(); UI != UE;
1113 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1114 if (LI == 0 || !LI->isSimple())
1117 // Both operands to the select need to be dereferencable, either
1118 // absolutely (e.g. allocas) or at this point because we can see other
1121 !isSafeToLoadUnconditionally(TValue, LI, LI->getAlignment(), DL))
1124 !isSafeToLoadUnconditionally(FValue, LI, LI->getAlignment(), DL))
1131 static void speculateSelectInstLoads(SelectInst &SI) {
1132 DEBUG(dbgs() << " original: " << SI << "\n");
1134 IRBuilderTy IRB(&SI);
1135 Value *TV = SI.getTrueValue();
1136 Value *FV = SI.getFalseValue();
1137 // Replace the loads of the select with a select of two loads.
1138 while (!SI.use_empty()) {
1139 LoadInst *LI = cast<LoadInst>(*SI.use_begin());
1140 assert(LI->isSimple() && "We only speculate simple loads");
1142 IRB.SetInsertPoint(LI);
1144 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1146 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1147 NumLoadsSpeculated += 2;
1149 // Transfer alignment and TBAA info if present.
1150 TL->setAlignment(LI->getAlignment());
1151 FL->setAlignment(LI->getAlignment());
1152 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
1153 TL->setMetadata(LLVMContext::MD_tbaa, Tag);
1154 FL->setMetadata(LLVMContext::MD_tbaa, Tag);
1157 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1158 LI->getName() + ".sroa.speculated");
1160 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1161 LI->replaceAllUsesWith(V);
1162 LI->eraseFromParent();
1164 SI.eraseFromParent();
1167 /// \brief Build a GEP out of a base pointer and indices.
1169 /// This will return the BasePtr if that is valid, or build a new GEP
1170 /// instruction using the IRBuilder if GEP-ing is needed.
1171 static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1172 SmallVectorImpl<Value *> &Indices) {
1173 if (Indices.empty())
1176 // A single zero index is a no-op, so check for this and avoid building a GEP
1178 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1181 return IRB.CreateInBoundsGEP(BasePtr, Indices, "idx");
1184 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1185 /// TargetTy without changing the offset of the pointer.
1187 /// This routine assumes we've already established a properly offset GEP with
1188 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1189 /// zero-indices down through type layers until we find one the same as
1190 /// TargetTy. If we can't find one with the same type, we at least try to use
1191 /// one with the same size. If none of that works, we just produce the GEP as
1192 /// indicated by Indices to have the correct offset.
1193 static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1194 Value *BasePtr, Type *Ty, Type *TargetTy,
1195 SmallVectorImpl<Value *> &Indices) {
1197 return buildGEP(IRB, BasePtr, Indices);
1199 // See if we can descend into a struct and locate a field with the correct
1201 unsigned NumLayers = 0;
1202 Type *ElementTy = Ty;
1204 if (ElementTy->isPointerTy())
1206 if (SequentialType *SeqTy = dyn_cast<SequentialType>(ElementTy)) {
1207 ElementTy = SeqTy->getElementType();
1208 // Note that we use the default address space as this index is over an
1209 // array or a vector, not a pointer.
1210 Indices.push_back(IRB.getInt(APInt(DL.getPointerSizeInBits(0), 0)));
1211 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1212 if (STy->element_begin() == STy->element_end())
1213 break; // Nothing left to descend into.
1214 ElementTy = *STy->element_begin();
1215 Indices.push_back(IRB.getInt32(0));
1220 } while (ElementTy != TargetTy);
1221 if (ElementTy != TargetTy)
1222 Indices.erase(Indices.end() - NumLayers, Indices.end());
1224 return buildGEP(IRB, BasePtr, Indices);
1227 /// \brief Recursively compute indices for a natural GEP.
1229 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1230 /// element types adding appropriate indices for the GEP.
1231 static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1232 Value *Ptr, Type *Ty, APInt &Offset,
1234 SmallVectorImpl<Value *> &Indices) {
1236 return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices);
1238 // We can't recurse through pointer types.
1239 if (Ty->isPointerTy())
1242 // We try to analyze GEPs over vectors here, but note that these GEPs are
1243 // extremely poorly defined currently. The long-term goal is to remove GEPing
1244 // over a vector from the IR completely.
1245 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1246 unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
1247 if (ElementSizeInBits % 8)
1248 return 0; // GEPs over non-multiple of 8 size vector elements are invalid.
1249 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1250 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1251 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1253 Offset -= NumSkippedElements * ElementSize;
1254 Indices.push_back(IRB.getInt(NumSkippedElements));
1255 return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1256 Offset, TargetTy, Indices);
1259 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1260 Type *ElementTy = ArrTy->getElementType();
1261 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1262 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1263 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1266 Offset -= NumSkippedElements * ElementSize;
1267 Indices.push_back(IRB.getInt(NumSkippedElements));
1268 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1272 StructType *STy = dyn_cast<StructType>(Ty);
1276 const StructLayout *SL = DL.getStructLayout(STy);
1277 uint64_t StructOffset = Offset.getZExtValue();
1278 if (StructOffset >= SL->getSizeInBytes())
1280 unsigned Index = SL->getElementContainingOffset(StructOffset);
1281 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1282 Type *ElementTy = STy->getElementType(Index);
1283 if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
1284 return 0; // The offset points into alignment padding.
1286 Indices.push_back(IRB.getInt32(Index));
1287 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1291 /// \brief Get a natural GEP from a base pointer to a particular offset and
1292 /// resulting in a particular type.
1294 /// The goal is to produce a "natural" looking GEP that works with the existing
1295 /// composite types to arrive at the appropriate offset and element type for
1296 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1297 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1298 /// Indices, and setting Ty to the result subtype.
1300 /// If no natural GEP can be constructed, this function returns null.
1301 static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1302 Value *Ptr, APInt Offset, Type *TargetTy,
1303 SmallVectorImpl<Value *> &Indices) {
1304 PointerType *Ty = cast<PointerType>(Ptr->getType());
1306 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1308 if (Ty == IRB.getInt8PtrTy() && TargetTy->isIntegerTy(8))
1311 Type *ElementTy = Ty->getElementType();
1312 if (!ElementTy->isSized())
1313 return 0; // We can't GEP through an unsized element.
1314 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1315 if (ElementSize == 0)
1316 return 0; // Zero-length arrays can't help us build a natural GEP.
1317 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1319 Offset -= NumSkippedElements * ElementSize;
1320 Indices.push_back(IRB.getInt(NumSkippedElements));
1321 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1325 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1326 /// resulting pointer has PointerTy.
1328 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1329 /// and produces the pointer type desired. Where it cannot, it will try to use
1330 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1331 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1332 /// bitcast to the type.
1334 /// The strategy for finding the more natural GEPs is to peel off layers of the
1335 /// pointer, walking back through bit casts and GEPs, searching for a base
1336 /// pointer from which we can compute a natural GEP with the desired
1337 /// properties. The algorithm tries to fold as many constant indices into
1338 /// a single GEP as possible, thus making each GEP more independent of the
1339 /// surrounding code.
1340 static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL,
1341 Value *Ptr, APInt Offset, Type *PointerTy) {
1342 // Even though we don't look through PHI nodes, we could be called on an
1343 // instruction in an unreachable block, which may be on a cycle.
1344 SmallPtrSet<Value *, 4> Visited;
1345 Visited.insert(Ptr);
1346 SmallVector<Value *, 4> Indices;
1348 // We may end up computing an offset pointer that has the wrong type. If we
1349 // never are able to compute one directly that has the correct type, we'll
1350 // fall back to it, so keep it around here.
1351 Value *OffsetPtr = 0;
1353 // Remember any i8 pointer we come across to re-use if we need to do a raw
1356 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1358 Type *TargetTy = PointerTy->getPointerElementType();
1361 // First fold any existing GEPs into the offset.
1362 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1363 APInt GEPOffset(Offset.getBitWidth(), 0);
1364 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1366 Offset += GEPOffset;
1367 Ptr = GEP->getPointerOperand();
1368 if (!Visited.insert(Ptr))
1372 // See if we can perform a natural GEP here.
1374 if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1376 if (P->getType() == PointerTy) {
1377 // Zap any offset pointer that we ended up computing in previous rounds.
1378 if (OffsetPtr && OffsetPtr->use_empty())
1379 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1380 I->eraseFromParent();
1388 // Stash this pointer if we've found an i8*.
1389 if (Ptr->getType()->isIntegerTy(8)) {
1391 Int8PtrOffset = Offset;
1394 // Peel off a layer of the pointer and update the offset appropriately.
1395 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1396 Ptr = cast<Operator>(Ptr)->getOperand(0);
1397 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1398 if (GA->mayBeOverridden())
1400 Ptr = GA->getAliasee();
1404 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1405 } while (Visited.insert(Ptr));
1409 Int8Ptr = IRB.CreateBitCast(Ptr, IRB.getInt8PtrTy(),
1411 Int8PtrOffset = Offset;
1414 OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
1415 IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1420 // On the off chance we were targeting i8*, guard the bitcast here.
1421 if (Ptr->getType() != PointerTy)
1422 Ptr = IRB.CreateBitCast(Ptr, PointerTy, "cast");
1427 /// \brief Test whether we can convert a value from the old to the new type.
1429 /// This predicate should be used to guard calls to convertValue in order to
1430 /// ensure that we only try to convert viable values. The strategy is that we
1431 /// will peel off single element struct and array wrappings to get to an
1432 /// underlying value, and convert that value.
1433 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1436 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1437 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1438 if (NewITy->getBitWidth() >= OldITy->getBitWidth())
1440 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1442 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1445 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1446 if (NewTy->isPointerTy() && OldTy->isPointerTy())
1448 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
1456 /// \brief Generic routine to convert an SSA value to a value of a different
1459 /// This will try various different casting techniques, such as bitcasts,
1460 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1461 /// two types for viability with this routine.
1462 static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1464 assert(canConvertValue(DL, V->getType(), Ty) &&
1465 "Value not convertable to type");
1466 if (V->getType() == Ty)
1468 if (IntegerType *OldITy = dyn_cast<IntegerType>(V->getType()))
1469 if (IntegerType *NewITy = dyn_cast<IntegerType>(Ty))
1470 if (NewITy->getBitWidth() > OldITy->getBitWidth())
1471 return IRB.CreateZExt(V, NewITy);
1472 if (V->getType()->isIntegerTy() && Ty->isPointerTy())
1473 return IRB.CreateIntToPtr(V, Ty);
1474 if (V->getType()->isPointerTy() && Ty->isIntegerTy())
1475 return IRB.CreatePtrToInt(V, Ty);
1477 return IRB.CreateBitCast(V, Ty);
1480 /// \brief Test whether the given slice use can be promoted to a vector.
1482 /// This function is called to test each entry in a partioning which is slated
1483 /// for a single slice.
1484 static bool isVectorPromotionViableForSlice(
1485 const DataLayout &DL, AllocaSlices &S, uint64_t SliceBeginOffset,
1486 uint64_t SliceEndOffset, VectorType *Ty, uint64_t ElementSize,
1487 AllocaSlices::const_iterator I) {
1488 // First validate the slice offsets.
1489 uint64_t BeginOffset =
1490 std::max(I->beginOffset(), SliceBeginOffset) - SliceBeginOffset;
1491 uint64_t BeginIndex = BeginOffset / ElementSize;
1492 if (BeginIndex * ElementSize != BeginOffset ||
1493 BeginIndex >= Ty->getNumElements())
1495 uint64_t EndOffset =
1496 std::min(I->endOffset(), SliceEndOffset) - SliceBeginOffset;
1497 uint64_t EndIndex = EndOffset / ElementSize;
1498 if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
1501 assert(EndIndex > BeginIndex && "Empty vector!");
1502 uint64_t NumElements = EndIndex - BeginIndex;
1504 (NumElements == 1) ? Ty->getElementType()
1505 : VectorType::get(Ty->getElementType(), NumElements);
1508 Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1510 Use *U = I->getUse();
1512 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1513 if (MI->isVolatile())
1515 if (!I->isSplittable())
1516 return false; // Skip any unsplittable intrinsics.
1517 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1518 // Disable vector promotion when there are loads or stores of an FCA.
1520 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1521 if (LI->isVolatile())
1523 Type *LTy = LI->getType();
1524 if (SliceBeginOffset > I->beginOffset() ||
1525 SliceEndOffset < I->endOffset()) {
1526 assert(LTy->isIntegerTy());
1529 if (!canConvertValue(DL, SliceTy, LTy))
1531 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1532 if (SI->isVolatile())
1534 Type *STy = SI->getValueOperand()->getType();
1535 if (SliceBeginOffset > I->beginOffset() ||
1536 SliceEndOffset < I->endOffset()) {
1537 assert(STy->isIntegerTy());
1540 if (!canConvertValue(DL, STy, SliceTy))
1549 /// \brief Test whether the given alloca partitioning and range of slices can be
1550 /// promoted to a vector.
1552 /// This is a quick test to check whether we can rewrite a particular alloca
1553 /// partition (and its newly formed alloca) into a vector alloca with only
1554 /// whole-vector loads and stores such that it could be promoted to a vector
1555 /// SSA value. We only can ensure this for a limited set of operations, and we
1556 /// don't want to do the rewrites unless we are confident that the result will
1557 /// be promotable, so we have an early test here.
1559 isVectorPromotionViable(const DataLayout &DL, Type *AllocaTy, AllocaSlices &S,
1560 uint64_t SliceBeginOffset, uint64_t SliceEndOffset,
1561 AllocaSlices::const_iterator I,
1562 AllocaSlices::const_iterator E,
1563 ArrayRef<AllocaSlices::iterator> SplitUses) {
1564 VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
1568 uint64_t ElementSize = DL.getTypeSizeInBits(Ty->getScalarType());
1570 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1571 // that aren't byte sized.
1572 if (ElementSize % 8)
1574 assert((DL.getTypeSizeInBits(Ty) % 8) == 0 &&
1575 "vector size not a multiple of element size?");
1579 if (!isVectorPromotionViableForSlice(DL, S, SliceBeginOffset,
1580 SliceEndOffset, Ty, ElementSize, I))
1583 for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
1584 SUE = SplitUses.end();
1586 if (!isVectorPromotionViableForSlice(DL, S, SliceBeginOffset,
1587 SliceEndOffset, Ty, ElementSize, *SUI))
1593 /// \brief Test whether a slice of an alloca is valid for integer widening.
1595 /// This implements the necessary checking for the \c isIntegerWideningViable
1596 /// test below on a single slice of the alloca.
1597 static bool isIntegerWideningViableForSlice(const DataLayout &DL,
1599 uint64_t AllocBeginOffset,
1600 uint64_t Size, AllocaSlices &S,
1601 AllocaSlices::const_iterator I,
1602 bool &WholeAllocaOp) {
1603 uint64_t RelBegin = I->beginOffset() - AllocBeginOffset;
1604 uint64_t RelEnd = I->endOffset() - AllocBeginOffset;
1606 // We can't reasonably handle cases where the load or store extends past
1607 // the end of the aloca's type and into its padding.
1611 Use *U = I->getUse();
1613 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1614 if (LI->isVolatile())
1616 if (RelBegin == 0 && RelEnd == Size)
1617 WholeAllocaOp = true;
1618 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
1619 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1621 } else if (RelBegin != 0 || RelEnd != Size ||
1622 !canConvertValue(DL, AllocaTy, LI->getType())) {
1623 // Non-integer loads need to be convertible from the alloca type so that
1624 // they are promotable.
1627 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1628 Type *ValueTy = SI->getValueOperand()->getType();
1629 if (SI->isVolatile())
1631 if (RelBegin == 0 && RelEnd == Size)
1632 WholeAllocaOp = true;
1633 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
1634 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1636 } else if (RelBegin != 0 || RelEnd != Size ||
1637 !canConvertValue(DL, ValueTy, AllocaTy)) {
1638 // Non-integer stores need to be convertible to the alloca type so that
1639 // they are promotable.
1642 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1643 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
1645 if (!I->isSplittable())
1646 return false; // Skip any unsplittable intrinsics.
1647 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1648 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1649 II->getIntrinsicID() != Intrinsic::lifetime_end)
1658 /// \brief Test whether the given alloca partition's integer operations can be
1659 /// widened to promotable ones.
1661 /// This is a quick test to check whether we can rewrite the integer loads and
1662 /// stores to a particular alloca into wider loads and stores and be able to
1663 /// promote the resulting alloca.
1665 isIntegerWideningViable(const DataLayout &DL, Type *AllocaTy,
1666 uint64_t AllocBeginOffset, AllocaSlices &S,
1667 AllocaSlices::const_iterator I,
1668 AllocaSlices::const_iterator E,
1669 ArrayRef<AllocaSlices::iterator> SplitUses) {
1670 uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
1671 // Don't create integer types larger than the maximum bitwidth.
1672 if (SizeInBits > IntegerType::MAX_INT_BITS)
1675 // Don't try to handle allocas with bit-padding.
1676 if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
1679 // We need to ensure that an integer type with the appropriate bitwidth can
1680 // be converted to the alloca type, whatever that is. We don't want to force
1681 // the alloca itself to have an integer type if there is a more suitable one.
1682 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
1683 if (!canConvertValue(DL, AllocaTy, IntTy) ||
1684 !canConvertValue(DL, IntTy, AllocaTy))
1687 uint64_t Size = DL.getTypeStoreSize(AllocaTy);
1689 // While examining uses, we ensure that the alloca has a covering load or
1690 // store. We don't want to widen the integer operations only to fail to
1691 // promote due to some other unsplittable entry (which we may make splittable
1692 // later). However, if there are only splittable uses, go ahead and assume
1693 // that we cover the alloca.
1694 bool WholeAllocaOp = (I != E) ? false : DL.isLegalInteger(SizeInBits);
1697 if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
1698 S, I, WholeAllocaOp))
1701 for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
1702 SUE = SplitUses.end();
1704 if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
1705 S, *SUI, WholeAllocaOp))
1708 return WholeAllocaOp;
1711 static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1712 IntegerType *Ty, uint64_t Offset,
1713 const Twine &Name) {
1714 DEBUG(dbgs() << " start: " << *V << "\n");
1715 IntegerType *IntTy = cast<IntegerType>(V->getType());
1716 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
1717 "Element extends past full value");
1718 uint64_t ShAmt = 8*Offset;
1719 if (DL.isBigEndian())
1720 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
1722 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
1723 DEBUG(dbgs() << " shifted: " << *V << "\n");
1725 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
1726 "Cannot extract to a larger integer!");
1728 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
1729 DEBUG(dbgs() << " trunced: " << *V << "\n");
1734 static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
1735 Value *V, uint64_t Offset, const Twine &Name) {
1736 IntegerType *IntTy = cast<IntegerType>(Old->getType());
1737 IntegerType *Ty = cast<IntegerType>(V->getType());
1738 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
1739 "Cannot insert a larger integer!");
1740 DEBUG(dbgs() << " start: " << *V << "\n");
1742 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
1743 DEBUG(dbgs() << " extended: " << *V << "\n");
1745 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
1746 "Element store outside of alloca store");
1747 uint64_t ShAmt = 8*Offset;
1748 if (DL.isBigEndian())
1749 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
1751 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
1752 DEBUG(dbgs() << " shifted: " << *V << "\n");
1755 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
1756 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
1757 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
1758 DEBUG(dbgs() << " masked: " << *Old << "\n");
1759 V = IRB.CreateOr(Old, V, Name + ".insert");
1760 DEBUG(dbgs() << " inserted: " << *V << "\n");
1765 static Value *extractVector(IRBuilderTy &IRB, Value *V,
1766 unsigned BeginIndex, unsigned EndIndex,
1767 const Twine &Name) {
1768 VectorType *VecTy = cast<VectorType>(V->getType());
1769 unsigned NumElements = EndIndex - BeginIndex;
1770 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
1772 if (NumElements == VecTy->getNumElements())
1775 if (NumElements == 1) {
1776 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
1778 DEBUG(dbgs() << " extract: " << *V << "\n");
1782 SmallVector<Constant*, 8> Mask;
1783 Mask.reserve(NumElements);
1784 for (unsigned i = BeginIndex; i != EndIndex; ++i)
1785 Mask.push_back(IRB.getInt32(i));
1786 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
1787 ConstantVector::get(Mask),
1789 DEBUG(dbgs() << " shuffle: " << *V << "\n");
1793 static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
1794 unsigned BeginIndex, const Twine &Name) {
1795 VectorType *VecTy = cast<VectorType>(Old->getType());
1796 assert(VecTy && "Can only insert a vector into a vector");
1798 VectorType *Ty = dyn_cast<VectorType>(V->getType());
1800 // Single element to insert.
1801 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
1803 DEBUG(dbgs() << " insert: " << *V << "\n");
1807 assert(Ty->getNumElements() <= VecTy->getNumElements() &&
1808 "Too many elements!");
1809 if (Ty->getNumElements() == VecTy->getNumElements()) {
1810 assert(V->getType() == VecTy && "Vector type mismatch");
1813 unsigned EndIndex = BeginIndex + Ty->getNumElements();
1815 // When inserting a smaller vector into the larger to store, we first
1816 // use a shuffle vector to widen it with undef elements, and then
1817 // a second shuffle vector to select between the loaded vector and the
1819 SmallVector<Constant*, 8> Mask;
1820 Mask.reserve(VecTy->getNumElements());
1821 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
1822 if (i >= BeginIndex && i < EndIndex)
1823 Mask.push_back(IRB.getInt32(i - BeginIndex));
1825 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
1826 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
1827 ConstantVector::get(Mask),
1829 DEBUG(dbgs() << " shuffle: " << *V << "\n");
1832 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
1833 Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
1835 V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
1837 DEBUG(dbgs() << " blend: " << *V << "\n");
1842 /// \brief Visitor to rewrite instructions using p particular slice of an alloca
1843 /// to use a new alloca.
1845 /// Also implements the rewriting to vector-based accesses when the partition
1846 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
1848 class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
1849 // Befriend the base class so it can delegate to private visit methods.
1850 friend class llvm::InstVisitor<AllocaSliceRewriter, bool>;
1851 typedef llvm::InstVisitor<AllocaSliceRewriter, bool> Base;
1853 const DataLayout &DL;
1856 AllocaInst &OldAI, &NewAI;
1857 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
1860 // If we are rewriting an alloca partition which can be written as pure
1861 // vector operations, we stash extra information here. When VecTy is
1862 // non-null, we have some strict guarantees about the rewritten alloca:
1863 // - The new alloca is exactly the size of the vector type here.
1864 // - The accesses all either map to the entire vector or to a single
1866 // - The set of accessing instructions is only one of those handled above
1867 // in isVectorPromotionViable. Generally these are the same access kinds
1868 // which are promotable via mem2reg.
1871 uint64_t ElementSize;
1873 // This is a convenience and flag variable that will be null unless the new
1874 // alloca's integer operations should be widened to this integer type due to
1875 // passing isIntegerWideningViable above. If it is non-null, the desired
1876 // integer type will be stored here for easy access during rewriting.
1879 // The offset of the slice currently being rewritten.
1880 uint64_t BeginOffset, EndOffset;
1884 Instruction *OldPtr;
1886 // Utility IR builder, whose name prefix is setup for each visited use, and
1887 // the insertion point is set to point to the user.
1891 AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &S, SROA &Pass,
1892 AllocaInst &OldAI, AllocaInst &NewAI,
1893 uint64_t NewBeginOffset, uint64_t NewEndOffset,
1894 bool IsVectorPromotable = false,
1895 bool IsIntegerPromotable = false)
1896 : DL(DL), S(S), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
1897 NewAllocaBeginOffset(NewBeginOffset), NewAllocaEndOffset(NewEndOffset),
1898 NewAllocaTy(NewAI.getAllocatedType()),
1899 VecTy(IsVectorPromotable ? cast<VectorType>(NewAllocaTy) : 0),
1900 ElementTy(VecTy ? VecTy->getElementType() : 0),
1901 ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
1902 IntTy(IsIntegerPromotable
1905 DL.getTypeSizeInBits(NewAI.getAllocatedType()))
1907 BeginOffset(), EndOffset(), IsSplittable(), IsSplit(), OldUse(),
1908 OldPtr(), IRB(NewAI.getContext(), ConstantFolder()) {
1910 assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
1911 "Only multiple-of-8 sized vector elements are viable");
1914 assert((!IsVectorPromotable && !IsIntegerPromotable) ||
1915 IsVectorPromotable != IsIntegerPromotable);
1918 bool visit(AllocaSlices::const_iterator I) {
1919 bool CanSROA = true;
1920 BeginOffset = I->beginOffset();
1921 EndOffset = I->endOffset();
1922 IsSplittable = I->isSplittable();
1924 BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
1926 OldUse = I->getUse();
1927 OldPtr = cast<Instruction>(OldUse->get());
1929 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
1930 IRB.SetInsertPoint(OldUserI);
1931 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
1932 IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
1934 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
1941 // Make sure the other visit overloads are visible.
1944 // Every instruction which can end up as a user must have a rewrite rule.
1945 bool visitInstruction(Instruction &I) {
1946 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
1947 llvm_unreachable("No rewrite rule for this instruction!");
1950 Value *getAdjustedAllocaPtr(IRBuilderTy &IRB, uint64_t Offset,
1952 assert(Offset >= NewAllocaBeginOffset);
1953 return getAdjustedPtr(IRB, DL, &NewAI, APInt(DL.getPointerSizeInBits(),
1954 Offset - NewAllocaBeginOffset),
1958 /// \brief Compute suitable alignment to access an offset into the new alloca.
1959 unsigned getOffsetAlign(uint64_t Offset) {
1960 unsigned NewAIAlign = NewAI.getAlignment();
1962 NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
1963 return MinAlign(NewAIAlign, Offset);
1966 /// \brief Compute suitable alignment to access a type at an offset of the
1969 /// \returns zero if the type's ABI alignment is a suitable alignment,
1970 /// otherwise returns the maximal suitable alignment.
1971 unsigned getOffsetTypeAlign(Type *Ty, uint64_t Offset) {
1972 unsigned Align = getOffsetAlign(Offset);
1973 return Align == DL.getABITypeAlignment(Ty) ? 0 : Align;
1976 unsigned getIndex(uint64_t Offset) {
1977 assert(VecTy && "Can only call getIndex when rewriting a vector");
1978 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
1979 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
1980 uint32_t Index = RelOffset / ElementSize;
1981 assert(Index * ElementSize == RelOffset);
1985 void deleteIfTriviallyDead(Value *V) {
1986 Instruction *I = cast<Instruction>(V);
1987 if (isInstructionTriviallyDead(I))
1988 Pass.DeadInsts.insert(I);
1991 Value *rewriteVectorizedLoadInst(uint64_t NewBeginOffset,
1992 uint64_t NewEndOffset) {
1993 unsigned BeginIndex = getIndex(NewBeginOffset);
1994 unsigned EndIndex = getIndex(NewEndOffset);
1995 assert(EndIndex > BeginIndex && "Empty vector!");
1997 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
1999 return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2002 Value *rewriteIntegerLoad(LoadInst &LI, uint64_t NewBeginOffset,
2003 uint64_t NewEndOffset) {
2004 assert(IntTy && "We cannot insert an integer to the alloca");
2005 assert(!LI.isVolatile());
2006 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2008 V = convertValue(DL, IRB, V, IntTy);
2009 assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2010 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2011 if (Offset > 0 || NewEndOffset < NewAllocaEndOffset)
2012 V = extractInteger(DL, IRB, V, cast<IntegerType>(LI.getType()), Offset,
2017 bool visitLoadInst(LoadInst &LI) {
2018 DEBUG(dbgs() << " original: " << LI << "\n");
2019 Value *OldOp = LI.getOperand(0);
2020 assert(OldOp == OldPtr);
2022 // Compute the intersecting offset range.
2023 assert(BeginOffset < NewAllocaEndOffset);
2024 assert(EndOffset > NewAllocaBeginOffset);
2025 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2026 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2028 uint64_t Size = NewEndOffset - NewBeginOffset;
2030 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), Size * 8)
2032 bool IsPtrAdjusted = false;
2035 V = rewriteVectorizedLoadInst(NewBeginOffset, NewEndOffset);
2036 } else if (IntTy && LI.getType()->isIntegerTy()) {
2037 V = rewriteIntegerLoad(LI, NewBeginOffset, NewEndOffset);
2038 } else if (NewBeginOffset == NewAllocaBeginOffset &&
2039 canConvertValue(DL, NewAllocaTy, LI.getType())) {
2040 V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2041 LI.isVolatile(), "load");
2043 Type *LTy = TargetTy->getPointerTo();
2044 V = IRB.CreateAlignedLoad(
2045 getAdjustedAllocaPtr(IRB, NewBeginOffset, LTy),
2046 getOffsetTypeAlign(TargetTy, NewBeginOffset - NewAllocaBeginOffset),
2047 LI.isVolatile(), "load");
2048 IsPtrAdjusted = true;
2050 V = convertValue(DL, IRB, V, TargetTy);
2053 assert(!LI.isVolatile());
2054 assert(LI.getType()->isIntegerTy() &&
2055 "Only integer type loads and stores are split");
2056 assert(Size < DL.getTypeStoreSize(LI.getType()) &&
2057 "Split load isn't smaller than original load");
2058 assert(LI.getType()->getIntegerBitWidth() ==
2059 DL.getTypeStoreSizeInBits(LI.getType()) &&
2060 "Non-byte-multiple bit width");
2061 // Move the insertion point just past the load so that we can refer to it.
2062 IRB.SetInsertPoint(llvm::next(BasicBlock::iterator(&LI)));
2063 // Create a placeholder value with the same type as LI to use as the
2064 // basis for the new value. This allows us to replace the uses of LI with
2065 // the computed value, and then replace the placeholder with LI, leaving
2066 // LI only used for this computation.
2068 = new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
2069 V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset,
2071 LI.replaceAllUsesWith(V);
2072 Placeholder->replaceAllUsesWith(&LI);
2075 LI.replaceAllUsesWith(V);
2078 Pass.DeadInsts.insert(&LI);
2079 deleteIfTriviallyDead(OldOp);
2080 DEBUG(dbgs() << " to: " << *V << "\n");
2081 return !LI.isVolatile() && !IsPtrAdjusted;
2084 bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp,
2085 uint64_t NewBeginOffset,
2086 uint64_t NewEndOffset) {
2087 if (V->getType() != VecTy) {
2088 unsigned BeginIndex = getIndex(NewBeginOffset);
2089 unsigned EndIndex = getIndex(NewEndOffset);
2090 assert(EndIndex > BeginIndex && "Empty vector!");
2091 unsigned NumElements = EndIndex - BeginIndex;
2092 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2094 (NumElements == 1) ? ElementTy
2095 : VectorType::get(ElementTy, NumElements);
2096 if (V->getType() != SliceTy)
2097 V = convertValue(DL, IRB, V, SliceTy);
2099 // Mix in the existing elements.
2100 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2102 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2104 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2105 Pass.DeadInsts.insert(&SI);
2108 DEBUG(dbgs() << " to: " << *Store << "\n");
2112 bool rewriteIntegerStore(Value *V, StoreInst &SI,
2113 uint64_t NewBeginOffset, uint64_t NewEndOffset) {
2114 assert(IntTy && "We cannot extract an integer from the alloca");
2115 assert(!SI.isVolatile());
2116 if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2117 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2119 Old = convertValue(DL, IRB, Old, IntTy);
2120 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2121 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2122 V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset,
2125 V = convertValue(DL, IRB, V, NewAllocaTy);
2126 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2127 Pass.DeadInsts.insert(&SI);
2129 DEBUG(dbgs() << " to: " << *Store << "\n");
2133 bool visitStoreInst(StoreInst &SI) {
2134 DEBUG(dbgs() << " original: " << SI << "\n");
2135 Value *OldOp = SI.getOperand(1);
2136 assert(OldOp == OldPtr);
2138 Value *V = SI.getValueOperand();
2140 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2141 // alloca that should be re-examined after promoting this alloca.
2142 if (V->getType()->isPointerTy())
2143 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2144 Pass.PostPromotionWorklist.insert(AI);
2146 // Compute the intersecting offset range.
2147 assert(BeginOffset < NewAllocaEndOffset);
2148 assert(EndOffset > NewAllocaBeginOffset);
2149 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2150 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2152 uint64_t Size = NewEndOffset - NewBeginOffset;
2153 if (Size < DL.getTypeStoreSize(V->getType())) {
2154 assert(!SI.isVolatile());
2155 assert(V->getType()->isIntegerTy() &&
2156 "Only integer type loads and stores are split");
2157 assert(V->getType()->getIntegerBitWidth() ==
2158 DL.getTypeStoreSizeInBits(V->getType()) &&
2159 "Non-byte-multiple bit width");
2160 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), Size * 8);
2161 V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset,
2166 return rewriteVectorizedStoreInst(V, SI, OldOp, NewBeginOffset,
2168 if (IntTy && V->getType()->isIntegerTy())
2169 return rewriteIntegerStore(V, SI, NewBeginOffset, NewEndOffset);
2172 if (NewBeginOffset == NewAllocaBeginOffset &&
2173 NewEndOffset == NewAllocaEndOffset &&
2174 canConvertValue(DL, V->getType(), NewAllocaTy)) {
2175 V = convertValue(DL, IRB, V, NewAllocaTy);
2176 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2179 Value *NewPtr = getAdjustedAllocaPtr(IRB, NewBeginOffset,
2180 V->getType()->getPointerTo());
2181 NewSI = IRB.CreateAlignedStore(
2182 V, NewPtr, getOffsetTypeAlign(
2183 V->getType(), NewBeginOffset - NewAllocaBeginOffset),
2187 Pass.DeadInsts.insert(&SI);
2188 deleteIfTriviallyDead(OldOp);
2190 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2191 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2194 /// \brief Compute an integer value from splatting an i8 across the given
2195 /// number of bytes.
2197 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2198 /// call this routine.
2199 /// FIXME: Heed the advice above.
2201 /// \param V The i8 value to splat.
2202 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2203 Value *getIntegerSplat(Value *V, unsigned Size) {
2204 assert(Size > 0 && "Expected a positive number of bytes.");
2205 IntegerType *VTy = cast<IntegerType>(V->getType());
2206 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2210 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size*8);
2211 V = IRB.CreateMul(IRB.CreateZExt(V, SplatIntTy, "zext"),
2212 ConstantExpr::getUDiv(
2213 Constant::getAllOnesValue(SplatIntTy),
2214 ConstantExpr::getZExt(
2215 Constant::getAllOnesValue(V->getType()),
2221 /// \brief Compute a vector splat for a given element value.
2222 Value *getVectorSplat(Value *V, unsigned NumElements) {
2223 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2224 DEBUG(dbgs() << " splat: " << *V << "\n");
2228 bool visitMemSetInst(MemSetInst &II) {
2229 DEBUG(dbgs() << " original: " << II << "\n");
2230 assert(II.getRawDest() == OldPtr);
2232 // If the memset has a variable size, it cannot be split, just adjust the
2233 // pointer to the new alloca.
2234 if (!isa<Constant>(II.getLength())) {
2236 assert(BeginOffset >= NewAllocaBeginOffset);
2238 getAdjustedAllocaPtr(IRB, BeginOffset, II.getRawDest()->getType()));
2239 Type *CstTy = II.getAlignmentCst()->getType();
2240 II.setAlignment(ConstantInt::get(CstTy, getOffsetAlign(BeginOffset)));
2242 deleteIfTriviallyDead(OldPtr);
2246 // Record this instruction for deletion.
2247 Pass.DeadInsts.insert(&II);
2249 Type *AllocaTy = NewAI.getAllocatedType();
2250 Type *ScalarTy = AllocaTy->getScalarType();
2252 // Compute the intersecting offset range.
2253 assert(BeginOffset < NewAllocaEndOffset);
2254 assert(EndOffset > NewAllocaBeginOffset);
2255 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2256 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2257 uint64_t SliceOffset = NewBeginOffset - NewAllocaBeginOffset;
2259 // If this doesn't map cleanly onto the alloca type, and that type isn't
2260 // a single value type, just emit a memset.
2261 if (!VecTy && !IntTy &&
2262 (BeginOffset > NewAllocaBeginOffset ||
2263 EndOffset < NewAllocaEndOffset ||
2264 !AllocaTy->isSingleValueType() ||
2265 !DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy)) ||
2266 DL.getTypeSizeInBits(ScalarTy)%8 != 0)) {
2267 Type *SizeTy = II.getLength()->getType();
2268 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2269 CallInst *New = IRB.CreateMemSet(
2270 getAdjustedAllocaPtr(IRB, NewBeginOffset, II.getRawDest()->getType()),
2271 II.getValue(), Size, getOffsetAlign(SliceOffset), II.isVolatile());
2273 DEBUG(dbgs() << " to: " << *New << "\n");
2277 // If we can represent this as a simple value, we have to build the actual
2278 // value to store, which requires expanding the byte present in memset to
2279 // a sensible representation for the alloca type. This is essentially
2280 // splatting the byte to a sufficiently wide integer, splatting it across
2281 // any desired vector width, and bitcasting to the final type.
2285 // If this is a memset of a vectorized alloca, insert it.
2286 assert(ElementTy == ScalarTy);
2288 unsigned BeginIndex = getIndex(NewBeginOffset);
2289 unsigned EndIndex = getIndex(NewEndOffset);
2290 assert(EndIndex > BeginIndex && "Empty vector!");
2291 unsigned NumElements = EndIndex - BeginIndex;
2292 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2295 getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
2296 Splat = convertValue(DL, IRB, Splat, ElementTy);
2297 if (NumElements > 1)
2298 Splat = getVectorSplat(Splat, NumElements);
2300 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2302 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2304 // If this is a memset on an alloca where we can widen stores, insert the
2306 assert(!II.isVolatile());
2308 uint64_t Size = NewEndOffset - NewBeginOffset;
2309 V = getIntegerSplat(II.getValue(), Size);
2311 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2312 EndOffset != NewAllocaBeginOffset)) {
2313 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2315 Old = convertValue(DL, IRB, Old, IntTy);
2316 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2317 V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2319 assert(V->getType() == IntTy &&
2320 "Wrong type for an alloca wide integer!");
2322 V = convertValue(DL, IRB, V, AllocaTy);
2324 // Established these invariants above.
2325 assert(NewBeginOffset == NewAllocaBeginOffset);
2326 assert(NewEndOffset == NewAllocaEndOffset);
2328 V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
2329 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2330 V = getVectorSplat(V, AllocaVecTy->getNumElements());
2332 V = convertValue(DL, IRB, V, AllocaTy);
2335 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2338 DEBUG(dbgs() << " to: " << *New << "\n");
2339 return !II.isVolatile();
2342 bool visitMemTransferInst(MemTransferInst &II) {
2343 // Rewriting of memory transfer instructions can be a bit tricky. We break
2344 // them into two categories: split intrinsics and unsplit intrinsics.
2346 DEBUG(dbgs() << " original: " << II << "\n");
2348 // Compute the intersecting offset range.
2349 assert(BeginOffset < NewAllocaEndOffset);
2350 assert(EndOffset > NewAllocaBeginOffset);
2351 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2352 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2354 assert(II.getRawSource() == OldPtr || II.getRawDest() == OldPtr);
2355 bool IsDest = II.getRawDest() == OldPtr;
2357 // Compute the relative offset within the transfer.
2358 unsigned IntPtrWidth = DL.getPointerSizeInBits();
2359 APInt RelOffset(IntPtrWidth, NewBeginOffset - BeginOffset);
2361 unsigned Align = II.getAlignment();
2362 uint64_t SliceOffset = NewBeginOffset - NewAllocaBeginOffset;
2365 MinAlign(RelOffset.zextOrTrunc(64).getZExtValue(),
2366 MinAlign(II.getAlignment(), getOffsetAlign(SliceOffset)));
2368 // For unsplit intrinsics, we simply modify the source and destination
2369 // pointers in place. This isn't just an optimization, it is a matter of
2370 // correctness. With unsplit intrinsics we may be dealing with transfers
2371 // within a single alloca before SROA ran, or with transfers that have
2372 // a variable length. We may also be dealing with memmove instead of
2373 // memcpy, and so simply updating the pointers is the necessary for us to
2374 // update both source and dest of a single call.
2375 if (!IsSplittable) {
2376 Value *OldOp = IsDest ? II.getRawDest() : II.getRawSource();
2379 getAdjustedAllocaPtr(IRB, BeginOffset, II.getRawDest()->getType()));
2381 II.setSource(getAdjustedAllocaPtr(IRB, BeginOffset,
2382 II.getRawSource()->getType()));
2384 Type *CstTy = II.getAlignmentCst()->getType();
2385 II.setAlignment(ConstantInt::get(CstTy, Align));
2387 DEBUG(dbgs() << " to: " << II << "\n");
2388 deleteIfTriviallyDead(OldOp);
2391 // For split transfer intrinsics we have an incredibly useful assurance:
2392 // the source and destination do not reside within the same alloca, and at
2393 // least one of them does not escape. This means that we can replace
2394 // memmove with memcpy, and we don't need to worry about all manner of
2395 // downsides to splitting and transforming the operations.
2397 // If this doesn't map cleanly onto the alloca type, and that type isn't
2398 // a single value type, just emit a memcpy.
2400 = !VecTy && !IntTy && (BeginOffset > NewAllocaBeginOffset ||
2401 EndOffset < NewAllocaEndOffset ||
2402 !NewAI.getAllocatedType()->isSingleValueType());
2404 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2405 // size hasn't been shrunk based on analysis of the viable range, this is
2407 if (EmitMemCpy && &OldAI == &NewAI) {
2408 // Ensure the start lines up.
2409 assert(NewBeginOffset == BeginOffset);
2411 // Rewrite the size as needed.
2412 if (NewEndOffset != EndOffset)
2413 II.setLength(ConstantInt::get(II.getLength()->getType(),
2414 NewEndOffset - NewBeginOffset));
2417 // Record this instruction for deletion.
2418 Pass.DeadInsts.insert(&II);
2420 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2421 // alloca that should be re-examined after rewriting this instruction.
2422 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2424 = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets()))
2425 Pass.Worklist.insert(AI);
2428 Type *OtherPtrTy = IsDest ? II.getRawSource()->getType()
2429 : II.getRawDest()->getType();
2431 // Compute the other pointer, folding as much as possible to produce
2432 // a single, simple GEP in most cases.
2433 OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, RelOffset, OtherPtrTy);
2435 Value *OurPtr = getAdjustedAllocaPtr(
2436 IRB, NewBeginOffset,
2437 IsDest ? II.getRawDest()->getType() : II.getRawSource()->getType());
2438 Type *SizeTy = II.getLength()->getType();
2439 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2441 CallInst *New = IRB.CreateMemCpy(IsDest ? OurPtr : OtherPtr,
2442 IsDest ? OtherPtr : OurPtr,
2443 Size, Align, II.isVolatile());
2445 DEBUG(dbgs() << " to: " << *New << "\n");
2449 // Note that we clamp the alignment to 1 here as a 0 alignment for a memcpy
2450 // is equivalent to 1, but that isn't true if we end up rewriting this as
2455 bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
2456 NewEndOffset == NewAllocaEndOffset;
2457 uint64_t Size = NewEndOffset - NewBeginOffset;
2458 unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
2459 unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
2460 unsigned NumElements = EndIndex - BeginIndex;
2461 IntegerType *SubIntTy
2462 = IntTy ? Type::getIntNTy(IntTy->getContext(), Size*8) : 0;
2464 Type *OtherPtrTy = NewAI.getType();
2465 if (VecTy && !IsWholeAlloca) {
2466 if (NumElements == 1)
2467 OtherPtrTy = VecTy->getElementType();
2469 OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2471 OtherPtrTy = OtherPtrTy->getPointerTo();
2472 } else if (IntTy && !IsWholeAlloca) {
2473 OtherPtrTy = SubIntTy->getPointerTo();
2476 Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, RelOffset, OtherPtrTy);
2477 Value *DstPtr = &NewAI;
2479 std::swap(SrcPtr, DstPtr);
2482 if (VecTy && !IsWholeAlloca && !IsDest) {
2483 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2485 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
2486 } else if (IntTy && !IsWholeAlloca && !IsDest) {
2487 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2489 Src = convertValue(DL, IRB, Src, IntTy);
2490 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2491 Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
2493 Src = IRB.CreateAlignedLoad(SrcPtr, Align, II.isVolatile(),
2497 if (VecTy && !IsWholeAlloca && IsDest) {
2498 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2500 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
2501 } else if (IntTy && !IsWholeAlloca && IsDest) {
2502 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2504 Old = convertValue(DL, IRB, Old, IntTy);
2505 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2506 Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
2507 Src = convertValue(DL, IRB, Src, NewAllocaTy);
2510 StoreInst *Store = cast<StoreInst>(
2511 IRB.CreateAlignedStore(Src, DstPtr, Align, II.isVolatile()));
2513 DEBUG(dbgs() << " to: " << *Store << "\n");
2514 return !II.isVolatile();
2517 bool visitIntrinsicInst(IntrinsicInst &II) {
2518 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2519 II.getIntrinsicID() == Intrinsic::lifetime_end);
2520 DEBUG(dbgs() << " original: " << II << "\n");
2521 assert(II.getArgOperand(1) == OldPtr);
2523 // Compute the intersecting offset range.
2524 assert(BeginOffset < NewAllocaEndOffset);
2525 assert(EndOffset > NewAllocaBeginOffset);
2526 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2527 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2529 // Record this instruction for deletion.
2530 Pass.DeadInsts.insert(&II);
2533 = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2534 NewEndOffset - NewBeginOffset);
2536 getAdjustedAllocaPtr(IRB, NewBeginOffset, II.getArgOperand(1)->getType());
2538 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2539 New = IRB.CreateLifetimeStart(Ptr, Size);
2541 New = IRB.CreateLifetimeEnd(Ptr, Size);
2544 DEBUG(dbgs() << " to: " << *New << "\n");
2548 bool visitPHINode(PHINode &PN) {
2549 DEBUG(dbgs() << " original: " << PN << "\n");
2550 assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
2551 assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
2553 // We would like to compute a new pointer in only one place, but have it be
2554 // as local as possible to the PHI. To do that, we re-use the location of
2555 // the old pointer, which necessarily must be in the right position to
2556 // dominate the PHI.
2557 IRBuilderTy PtrBuilder(cast<Instruction>(OldPtr));
2558 PtrBuilder.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) +
2562 getAdjustedAllocaPtr(PtrBuilder, BeginOffset, OldPtr->getType());
2563 // Replace the operands which were using the old pointer.
2564 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
2566 DEBUG(dbgs() << " to: " << PN << "\n");
2567 deleteIfTriviallyDead(OldPtr);
2569 // Check whether we can speculate this PHI node, and if so remember that
2570 // fact and return that this alloca remains viable for promotion to an SSA
2572 if (isSafePHIToSpeculate(PN, &DL)) {
2573 Pass.SpeculatablePHIs.insert(&PN);
2577 return false; // PHIs can't be promoted on their own.
2580 bool visitSelectInst(SelectInst &SI) {
2581 DEBUG(dbgs() << " original: " << SI << "\n");
2582 assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
2583 "Pointer isn't an operand!");
2584 assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
2585 assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
2587 Value *NewPtr = getAdjustedAllocaPtr(IRB, BeginOffset, OldPtr->getType());
2588 // Replace the operands which were using the old pointer.
2589 if (SI.getOperand(1) == OldPtr)
2590 SI.setOperand(1, NewPtr);
2591 if (SI.getOperand(2) == OldPtr)
2592 SI.setOperand(2, NewPtr);
2594 DEBUG(dbgs() << " to: " << SI << "\n");
2595 deleteIfTriviallyDead(OldPtr);
2597 // Check whether we can speculate this select instruction, and if so
2598 // remember that fact and return that this alloca remains viable for
2599 // promotion to an SSA value.
2600 if (isSafeSelectToSpeculate(SI, &DL)) {
2601 Pass.SpeculatableSelects.insert(&SI);
2605 return false; // Selects can't be promoted on their own.
2612 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
2614 /// This pass aggressively rewrites all aggregate loads and stores on
2615 /// a particular pointer (or any pointer derived from it which we can identify)
2616 /// with scalar loads and stores.
2617 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
2618 // Befriend the base class so it can delegate to private visit methods.
2619 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
2621 const DataLayout &DL;
2623 /// Queue of pointer uses to analyze and potentially rewrite.
2624 SmallVector<Use *, 8> Queue;
2626 /// Set to prevent us from cycling with phi nodes and loops.
2627 SmallPtrSet<User *, 8> Visited;
2629 /// The current pointer use being rewritten. This is used to dig up the used
2630 /// value (as opposed to the user).
2634 AggLoadStoreRewriter(const DataLayout &DL) : DL(DL) {}
2636 /// Rewrite loads and stores through a pointer and all pointers derived from
2638 bool rewrite(Instruction &I) {
2639 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
2641 bool Changed = false;
2642 while (!Queue.empty()) {
2643 U = Queue.pop_back_val();
2644 Changed |= visit(cast<Instruction>(U->getUser()));
2650 /// Enqueue all the users of the given instruction for further processing.
2651 /// This uses a set to de-duplicate users.
2652 void enqueueUsers(Instruction &I) {
2653 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE;
2655 if (Visited.insert(*UI))
2656 Queue.push_back(&UI.getUse());
2659 // Conservative default is to not rewrite anything.
2660 bool visitInstruction(Instruction &I) { return false; }
2662 /// \brief Generic recursive split emission class.
2663 template <typename Derived>
2666 /// The builder used to form new instructions.
2668 /// The indices which to be used with insert- or extractvalue to select the
2669 /// appropriate value within the aggregate.
2670 SmallVector<unsigned, 4> Indices;
2671 /// The indices to a GEP instruction which will move Ptr to the correct slot
2672 /// within the aggregate.
2673 SmallVector<Value *, 4> GEPIndices;
2674 /// The base pointer of the original op, used as a base for GEPing the
2675 /// split operations.
2678 /// Initialize the splitter with an insertion point, Ptr and start with a
2679 /// single zero GEP index.
2680 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
2681 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
2684 /// \brief Generic recursive split emission routine.
2686 /// This method recursively splits an aggregate op (load or store) into
2687 /// scalar or vector ops. It splits recursively until it hits a single value
2688 /// and emits that single value operation via the template argument.
2690 /// The logic of this routine relies on GEPs and insertvalue and
2691 /// extractvalue all operating with the same fundamental index list, merely
2692 /// formatted differently (GEPs need actual values).
2694 /// \param Ty The type being split recursively into smaller ops.
2695 /// \param Agg The aggregate value being built up or stored, depending on
2696 /// whether this is splitting a load or a store respectively.
2697 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
2698 if (Ty->isSingleValueType())
2699 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
2701 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2702 unsigned OldSize = Indices.size();
2704 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
2706 assert(Indices.size() == OldSize && "Did not return to the old size");
2707 Indices.push_back(Idx);
2708 GEPIndices.push_back(IRB.getInt32(Idx));
2709 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
2710 GEPIndices.pop_back();
2716 if (StructType *STy = dyn_cast<StructType>(Ty)) {
2717 unsigned OldSize = Indices.size();
2719 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
2721 assert(Indices.size() == OldSize && "Did not return to the old size");
2722 Indices.push_back(Idx);
2723 GEPIndices.push_back(IRB.getInt32(Idx));
2724 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
2725 GEPIndices.pop_back();
2731 llvm_unreachable("Only arrays and structs are aggregate loadable types");
2735 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
2736 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2737 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
2739 /// Emit a leaf load of a single value. This is called at the leaves of the
2740 /// recursive emission to actually load values.
2741 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2742 assert(Ty->isSingleValueType());
2743 // Load the single value and insert it using the indices.
2744 Value *GEP = IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep");
2745 Value *Load = IRB.CreateLoad(GEP, Name + ".load");
2746 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
2747 DEBUG(dbgs() << " to: " << *Load << "\n");
2751 bool visitLoadInst(LoadInst &LI) {
2752 assert(LI.getPointerOperand() == *U);
2753 if (!LI.isSimple() || LI.getType()->isSingleValueType())
2756 // We have an aggregate being loaded, split it apart.
2757 DEBUG(dbgs() << " original: " << LI << "\n");
2758 LoadOpSplitter Splitter(&LI, *U);
2759 Value *V = UndefValue::get(LI.getType());
2760 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
2761 LI.replaceAllUsesWith(V);
2762 LI.eraseFromParent();
2766 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
2767 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2768 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
2770 /// Emit a leaf store of a single value. This is called at the leaves of the
2771 /// recursive emission to actually produce stores.
2772 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2773 assert(Ty->isSingleValueType());
2774 // Extract the single value and store it using the indices.
2775 Value *Store = IRB.CreateStore(
2776 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
2777 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
2779 DEBUG(dbgs() << " to: " << *Store << "\n");
2783 bool visitStoreInst(StoreInst &SI) {
2784 if (!SI.isSimple() || SI.getPointerOperand() != *U)
2786 Value *V = SI.getValueOperand();
2787 if (V->getType()->isSingleValueType())
2790 // We have an aggregate being stored, split it apart.
2791 DEBUG(dbgs() << " original: " << SI << "\n");
2792 StoreOpSplitter Splitter(&SI, *U);
2793 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
2794 SI.eraseFromParent();
2798 bool visitBitCastInst(BitCastInst &BC) {
2803 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
2808 bool visitPHINode(PHINode &PN) {
2813 bool visitSelectInst(SelectInst &SI) {
2820 /// \brief Strip aggregate type wrapping.
2822 /// This removes no-op aggregate types wrapping an underlying type. It will
2823 /// strip as many layers of types as it can without changing either the type
2824 /// size or the allocated size.
2825 static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
2826 if (Ty->isSingleValueType())
2829 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
2830 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
2833 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
2834 InnerTy = ArrTy->getElementType();
2835 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
2836 const StructLayout *SL = DL.getStructLayout(STy);
2837 unsigned Index = SL->getElementContainingOffset(0);
2838 InnerTy = STy->getElementType(Index);
2843 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
2844 TypeSize > DL.getTypeSizeInBits(InnerTy))
2847 return stripAggregateTypeWrapping(DL, InnerTy);
2850 /// \brief Try to find a partition of the aggregate type passed in for a given
2851 /// offset and size.
2853 /// This recurses through the aggregate type and tries to compute a subtype
2854 /// based on the offset and size. When the offset and size span a sub-section
2855 /// of an array, it will even compute a new array type for that sub-section,
2856 /// and the same for structs.
2858 /// Note that this routine is very strict and tries to find a partition of the
2859 /// type which produces the *exact* right offset and size. It is not forgiving
2860 /// when the size or offset cause either end of type-based partition to be off.
2861 /// Also, this is a best-effort routine. It is reasonable to give up and not
2862 /// return a type if necessary.
2863 static Type *getTypePartition(const DataLayout &DL, Type *Ty,
2864 uint64_t Offset, uint64_t Size) {
2865 if (Offset == 0 && DL.getTypeAllocSize(Ty) == Size)
2866 return stripAggregateTypeWrapping(DL, Ty);
2867 if (Offset > DL.getTypeAllocSize(Ty) ||
2868 (DL.getTypeAllocSize(Ty) - Offset) < Size)
2871 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
2872 // We can't partition pointers...
2873 if (SeqTy->isPointerTy())
2876 Type *ElementTy = SeqTy->getElementType();
2877 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
2878 uint64_t NumSkippedElements = Offset / ElementSize;
2879 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy)) {
2880 if (NumSkippedElements >= ArrTy->getNumElements())
2882 } else if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy)) {
2883 if (NumSkippedElements >= VecTy->getNumElements())
2886 Offset -= NumSkippedElements * ElementSize;
2888 // First check if we need to recurse.
2889 if (Offset > 0 || Size < ElementSize) {
2890 // Bail if the partition ends in a different array element.
2891 if ((Offset + Size) > ElementSize)
2893 // Recurse through the element type trying to peel off offset bytes.
2894 return getTypePartition(DL, ElementTy, Offset, Size);
2896 assert(Offset == 0);
2898 if (Size == ElementSize)
2899 return stripAggregateTypeWrapping(DL, ElementTy);
2900 assert(Size > ElementSize);
2901 uint64_t NumElements = Size / ElementSize;
2902 if (NumElements * ElementSize != Size)
2904 return ArrayType::get(ElementTy, NumElements);
2907 StructType *STy = dyn_cast<StructType>(Ty);
2911 const StructLayout *SL = DL.getStructLayout(STy);
2912 if (Offset >= SL->getSizeInBytes())
2914 uint64_t EndOffset = Offset + Size;
2915 if (EndOffset > SL->getSizeInBytes())
2918 unsigned Index = SL->getElementContainingOffset(Offset);
2919 Offset -= SL->getElementOffset(Index);
2921 Type *ElementTy = STy->getElementType(Index);
2922 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
2923 if (Offset >= ElementSize)
2924 return 0; // The offset points into alignment padding.
2926 // See if any partition must be contained by the element.
2927 if (Offset > 0 || Size < ElementSize) {
2928 if ((Offset + Size) > ElementSize)
2930 return getTypePartition(DL, ElementTy, Offset, Size);
2932 assert(Offset == 0);
2934 if (Size == ElementSize)
2935 return stripAggregateTypeWrapping(DL, ElementTy);
2937 StructType::element_iterator EI = STy->element_begin() + Index,
2938 EE = STy->element_end();
2939 if (EndOffset < SL->getSizeInBytes()) {
2940 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
2941 if (Index == EndIndex)
2942 return 0; // Within a single element and its padding.
2944 // Don't try to form "natural" types if the elements don't line up with the
2946 // FIXME: We could potentially recurse down through the last element in the
2947 // sub-struct to find a natural end point.
2948 if (SL->getElementOffset(EndIndex) != EndOffset)
2951 assert(Index < EndIndex);
2952 EE = STy->element_begin() + EndIndex;
2955 // Try to build up a sub-structure.
2956 StructType *SubTy = StructType::get(STy->getContext(), makeArrayRef(EI, EE),
2958 const StructLayout *SubSL = DL.getStructLayout(SubTy);
2959 if (Size != SubSL->getSizeInBytes())
2960 return 0; // The sub-struct doesn't have quite the size needed.
2965 /// \brief Rewrite an alloca partition's users.
2967 /// This routine drives both of the rewriting goals of the SROA pass. It tries
2968 /// to rewrite uses of an alloca partition to be conducive for SSA value
2969 /// promotion. If the partition needs a new, more refined alloca, this will
2970 /// build that new alloca, preserving as much type information as possible, and
2971 /// rewrite the uses of the old alloca to point at the new one and have the
2972 /// appropriate new offsets. It also evaluates how successful the rewrite was
2973 /// at enabling promotion and if it was successful queues the alloca to be
2975 bool SROA::rewritePartition(AllocaInst &AI, AllocaSlices &S,
2976 AllocaSlices::iterator B, AllocaSlices::iterator E,
2977 int64_t BeginOffset, int64_t EndOffset,
2978 ArrayRef<AllocaSlices::iterator> SplitUses) {
2979 assert(BeginOffset < EndOffset);
2980 uint64_t SliceSize = EndOffset - BeginOffset;
2982 // Try to compute a friendly type for this partition of the alloca. This
2983 // won't always succeed, in which case we fall back to a legal integer type
2984 // or an i8 array of an appropriate size.
2986 if (Type *CommonUseTy = findCommonType(B, E, EndOffset))
2987 if (DL->getTypeAllocSize(CommonUseTy) >= SliceSize)
2988 SliceTy = CommonUseTy;
2990 if (Type *TypePartitionTy = getTypePartition(*DL, AI.getAllocatedType(),
2991 BeginOffset, SliceSize))
2992 SliceTy = TypePartitionTy;
2993 if ((!SliceTy || (SliceTy->isArrayTy() &&
2994 SliceTy->getArrayElementType()->isIntegerTy())) &&
2995 DL->isLegalInteger(SliceSize * 8))
2996 SliceTy = Type::getIntNTy(*C, SliceSize * 8);
2998 SliceTy = ArrayType::get(Type::getInt8Ty(*C), SliceSize);
2999 assert(DL->getTypeAllocSize(SliceTy) >= SliceSize);
3001 bool IsVectorPromotable = isVectorPromotionViable(
3002 *DL, SliceTy, S, BeginOffset, EndOffset, B, E, SplitUses);
3004 bool IsIntegerPromotable =
3005 !IsVectorPromotable &&
3006 isIntegerWideningViable(*DL, SliceTy, BeginOffset, S, B, E, SplitUses);
3008 // Check for the case where we're going to rewrite to a new alloca of the
3009 // exact same type as the original, and with the same access offsets. In that
3010 // case, re-use the existing alloca, but still run through the rewriter to
3011 // perform phi and select speculation.
3013 if (SliceTy == AI.getAllocatedType()) {
3014 assert(BeginOffset == 0 &&
3015 "Non-zero begin offset but same alloca type");
3017 // FIXME: We should be able to bail at this point with "nothing changed".
3018 // FIXME: We might want to defer PHI speculation until after here.
3020 unsigned Alignment = AI.getAlignment();
3022 // The minimum alignment which users can rely on when the explicit
3023 // alignment is omitted or zero is that required by the ABI for this
3025 Alignment = DL->getABITypeAlignment(AI.getAllocatedType());
3027 Alignment = MinAlign(Alignment, BeginOffset);
3028 // If we will get at least this much alignment from the type alone, leave
3029 // the alloca's alignment unconstrained.
3030 if (Alignment <= DL->getABITypeAlignment(SliceTy))
3032 NewAI = new AllocaInst(SliceTy, 0, Alignment,
3033 AI.getName() + ".sroa." + Twine(B - S.begin()), &AI);
3037 DEBUG(dbgs() << "Rewriting alloca partition "
3038 << "[" << BeginOffset << "," << EndOffset << ") to: " << *NewAI
3041 // Track the high watermark on several worklists that are only relevant for
3042 // promoted allocas. We will reset it to this point if the alloca is not in
3043 // fact scheduled for promotion.
3044 unsigned PPWOldSize = PostPromotionWorklist.size();
3045 unsigned SPOldSize = SpeculatablePHIs.size();
3046 unsigned SSOldSize = SpeculatableSelects.size();
3048 AllocaSliceRewriter Rewriter(*DL, S, *this, AI, *NewAI, BeginOffset,
3049 EndOffset, IsVectorPromotable,
3050 IsIntegerPromotable);
3051 bool Promotable = true;
3052 for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
3053 SUE = SplitUses.end();
3054 SUI != SUE; ++SUI) {
3055 DEBUG(dbgs() << " rewriting split ");
3056 DEBUG(S.printSlice(dbgs(), *SUI, ""));
3057 Promotable &= Rewriter.visit(*SUI);
3059 for (AllocaSlices::iterator I = B; I != E; ++I) {
3060 DEBUG(dbgs() << " rewriting ");
3061 DEBUG(S.printSlice(dbgs(), I, ""));
3062 Promotable &= Rewriter.visit(I);
3065 if (Promotable && (SpeculatablePHIs.size() > SPOldSize ||
3066 SpeculatableSelects.size() > SSOldSize)) {
3067 // If we have a promotable alloca except for some unspeculated loads below
3068 // PHIs or Selects, iterate once. We will speculate the loads and on the
3069 // next iteration rewrite them into a promotable form.
3070 Worklist.insert(NewAI);
3071 } else if (Promotable) {
3072 DEBUG(dbgs() << " and queuing for promotion\n");
3073 PromotableAllocas.push_back(NewAI);
3074 } else if (NewAI != &AI) {
3075 // If we can't promote the alloca, iterate on it to check for new
3076 // refinements exposed by splitting the current alloca. Don't iterate on an
3077 // alloca which didn't actually change and didn't get promoted.
3078 // FIXME: We should actually track whether the rewriter changed anything.
3079 Worklist.insert(NewAI);
3082 // Drop any post-promotion work items if promotion didn't happen.
3084 while (PostPromotionWorklist.size() > PPWOldSize)
3085 PostPromotionWorklist.pop_back();
3086 while (SpeculatablePHIs.size() > SPOldSize)
3087 SpeculatablePHIs.pop_back();
3088 while (SpeculatableSelects.size() > SSOldSize)
3089 SpeculatableSelects.pop_back();
3096 struct IsSliceEndLessOrEqualTo {
3097 uint64_t UpperBound;
3099 IsSliceEndLessOrEqualTo(uint64_t UpperBound) : UpperBound(UpperBound) {}
3101 bool operator()(const AllocaSlices::iterator &I) {
3102 return I->endOffset() <= UpperBound;
3108 removeFinishedSplitUses(SmallVectorImpl<AllocaSlices::iterator> &SplitUses,
3109 uint64_t &MaxSplitUseEndOffset, uint64_t Offset) {
3110 if (Offset >= MaxSplitUseEndOffset) {
3112 MaxSplitUseEndOffset = 0;
3116 size_t SplitUsesOldSize = SplitUses.size();
3117 SplitUses.erase(std::remove_if(SplitUses.begin(), SplitUses.end(),
3118 IsSliceEndLessOrEqualTo(Offset)),
3120 if (SplitUsesOldSize == SplitUses.size())
3123 // Recompute the max. While this is linear, so is remove_if.
3124 MaxSplitUseEndOffset = 0;
3125 for (SmallVectorImpl<AllocaSlices::iterator>::iterator
3126 SUI = SplitUses.begin(),
3127 SUE = SplitUses.end();
3129 MaxSplitUseEndOffset = std::max((*SUI)->endOffset(), MaxSplitUseEndOffset);
3132 /// \brief Walks the slices of an alloca and form partitions based on them,
3133 /// rewriting each of their uses.
3134 bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &S) {
3135 if (S.begin() == S.end())
3138 bool Changed = false;
3139 SmallVector<AllocaSlices::iterator, 4> SplitUses;
3140 uint64_t MaxSplitUseEndOffset = 0;
3142 uint64_t BeginOffset = S.begin()->beginOffset();
3144 for (AllocaSlices::iterator SI = S.begin(), SJ = llvm::next(SI), SE = S.end();
3145 SI != SE; SI = SJ) {
3146 uint64_t MaxEndOffset = SI->endOffset();
3148 if (!SI->isSplittable()) {
3149 // When we're forming an unsplittable region, it must always start at the
3150 // first slice and will extend through its end.
3151 assert(BeginOffset == SI->beginOffset());
3153 // Form a partition including all of the overlapping slices with this
3154 // unsplittable slice.
3155 while (SJ != SE && SJ->beginOffset() < MaxEndOffset) {
3156 if (!SJ->isSplittable())
3157 MaxEndOffset = std::max(MaxEndOffset, SJ->endOffset());
3161 assert(SI->isSplittable()); // Established above.
3163 // Collect all of the overlapping splittable slices.
3164 while (SJ != SE && SJ->beginOffset() < MaxEndOffset &&
3165 SJ->isSplittable()) {
3166 MaxEndOffset = std::max(MaxEndOffset, SJ->endOffset());
3170 // Back up MaxEndOffset and SJ if we ended the span early when
3171 // encountering an unsplittable slice.
3172 if (SJ != SE && SJ->beginOffset() < MaxEndOffset) {
3173 assert(!SJ->isSplittable());
3174 MaxEndOffset = SJ->beginOffset();
3178 // Check if we have managed to move the end offset forward yet. If so,
3179 // we'll have to rewrite uses and erase old split uses.
3180 if (BeginOffset < MaxEndOffset) {
3181 // Rewrite a sequence of overlapping slices.
3183 rewritePartition(AI, S, SI, SJ, BeginOffset, MaxEndOffset, SplitUses);
3185 removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset, MaxEndOffset);
3188 // Accumulate all the splittable slices from the [SI,SJ) region which
3189 // overlap going forward.
3190 for (AllocaSlices::iterator SK = SI; SK != SJ; ++SK)
3191 if (SK->isSplittable() && SK->endOffset() > MaxEndOffset) {
3192 SplitUses.push_back(SK);
3193 MaxSplitUseEndOffset = std::max(SK->endOffset(), MaxSplitUseEndOffset);
3196 // If we're already at the end and we have no split uses, we're done.
3197 if (SJ == SE && SplitUses.empty())
3200 // If we have no split uses or no gap in offsets, we're ready to move to
3202 if (SplitUses.empty() || (SJ != SE && MaxEndOffset == SJ->beginOffset())) {
3203 BeginOffset = SJ->beginOffset();
3207 // Even if we have split slices, if the next slice is splittable and the
3208 // split slices reach it, we can simply set up the beginning offset of the
3209 // next iteration to bridge between them.
3210 if (SJ != SE && SJ->isSplittable() &&
3211 MaxSplitUseEndOffset > SJ->beginOffset()) {
3212 BeginOffset = MaxEndOffset;
3216 // Otherwise, we have a tail of split slices. Rewrite them with an empty
3218 uint64_t PostSplitEndOffset =
3219 SJ == SE ? MaxSplitUseEndOffset : SJ->beginOffset();
3221 Changed |= rewritePartition(AI, S, SJ, SJ, MaxEndOffset, PostSplitEndOffset,
3224 break; // Skip the rest, we don't need to do any cleanup.
3226 removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset,
3227 PostSplitEndOffset);
3229 // Now just reset the begin offset for the next iteration.
3230 BeginOffset = SJ->beginOffset();
3236 /// \brief Analyze an alloca for SROA.
3238 /// This analyzes the alloca to ensure we can reason about it, builds
3239 /// the slices of the alloca, and then hands it off to be split and
3240 /// rewritten as needed.
3241 bool SROA::runOnAlloca(AllocaInst &AI) {
3242 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3243 ++NumAllocasAnalyzed;
3245 // Special case dead allocas, as they're trivial.
3246 if (AI.use_empty()) {
3247 AI.eraseFromParent();
3251 // Skip alloca forms that this analysis can't handle.
3252 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3253 DL->getTypeAllocSize(AI.getAllocatedType()) == 0)
3256 bool Changed = false;
3258 // First, split any FCA loads and stores touching this alloca to promote
3259 // better splitting and promotion opportunities.
3260 AggLoadStoreRewriter AggRewriter(*DL);
3261 Changed |= AggRewriter.rewrite(AI);
3263 // Build the slices using a recursive instruction-visiting builder.
3264 AllocaSlices S(*DL, AI);
3265 DEBUG(S.print(dbgs()));
3269 // Delete all the dead users of this alloca before splitting and rewriting it.
3270 for (AllocaSlices::dead_user_iterator DI = S.dead_user_begin(),
3271 DE = S.dead_user_end();
3274 (*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType()));
3275 DeadInsts.insert(*DI);
3277 for (AllocaSlices::dead_op_iterator DO = S.dead_op_begin(),
3278 DE = S.dead_op_end();
3281 // Clobber the use with an undef value.
3282 **DO = UndefValue::get(OldV->getType());
3283 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3284 if (isInstructionTriviallyDead(OldI)) {
3286 DeadInsts.insert(OldI);
3290 // No slices to split. Leave the dead alloca for a later pass to clean up.
3291 if (S.begin() == S.end())
3294 Changed |= splitAlloca(AI, S);
3296 DEBUG(dbgs() << " Speculating PHIs\n");
3297 while (!SpeculatablePHIs.empty())
3298 speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
3300 DEBUG(dbgs() << " Speculating Selects\n");
3301 while (!SpeculatableSelects.empty())
3302 speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
3307 /// \brief Delete the dead instructions accumulated in this run.
3309 /// Recursively deletes the dead instructions we've accumulated. This is done
3310 /// at the very end to maximize locality of the recursive delete and to
3311 /// minimize the problems of invalidated instruction pointers as such pointers
3312 /// are used heavily in the intermediate stages of the algorithm.
3314 /// We also record the alloca instructions deleted here so that they aren't
3315 /// subsequently handed to mem2reg to promote.
3316 void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
3317 while (!DeadInsts.empty()) {
3318 Instruction *I = DeadInsts.pop_back_val();
3319 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3321 I->replaceAllUsesWith(UndefValue::get(I->getType()));
3323 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
3324 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
3325 // Zero out the operand and see if it becomes trivially dead.
3327 if (isInstructionTriviallyDead(U))
3328 DeadInsts.insert(U);
3331 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3332 DeletedAllocas.insert(AI);
3335 I->eraseFromParent();
3339 /// \brief Promote the allocas, using the best available technique.
3341 /// This attempts to promote whatever allocas have been identified as viable in
3342 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
3343 /// If there is a domtree available, we attempt to promote using the full power
3344 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3345 /// based on the SSAUpdater utilities. This function returns whether any
3346 /// promotion occurred.
3347 bool SROA::promoteAllocas(Function &F) {
3348 if (PromotableAllocas.empty())
3351 NumPromoted += PromotableAllocas.size();
3353 if (DT && !ForceSSAUpdater) {
3354 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3355 PromoteMemToReg(PromotableAllocas, *DT);
3356 PromotableAllocas.clear();
3360 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3362 DIBuilder DIB(*F.getParent());
3363 SmallVector<Instruction*, 64> Insts;
3365 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3366 AllocaInst *AI = PromotableAllocas[Idx];
3367 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
3369 Instruction *I = cast<Instruction>(*UI++);
3370 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3371 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3372 // leading to them) here. Eventually it should use them to optimize the
3373 // scalar values produced.
3374 if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
3375 assert(onlyUsedByLifetimeMarkers(I) &&
3376 "Found a bitcast used outside of a lifetime marker.");
3377 while (!I->use_empty())
3378 cast<Instruction>(*I->use_begin())->eraseFromParent();
3379 I->eraseFromParent();
3382 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3383 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3384 II->getIntrinsicID() == Intrinsic::lifetime_end);
3385 II->eraseFromParent();
3391 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3395 PromotableAllocas.clear();
3400 /// \brief A predicate to test whether an alloca belongs to a set.
3401 class IsAllocaInSet {
3402 typedef SmallPtrSet<AllocaInst *, 4> SetType;
3406 typedef AllocaInst *argument_type;
3408 IsAllocaInSet(const SetType &Set) : Set(Set) {}
3409 bool operator()(AllocaInst *AI) const { return Set.count(AI); }
3413 bool SROA::runOnFunction(Function &F) {
3414 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3415 C = &F.getContext();
3416 DL = getAnalysisIfAvailable<DataLayout>();
3418 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3421 DT = getAnalysisIfAvailable<DominatorTree>();
3423 BasicBlock &EntryBB = F.getEntryBlock();
3424 for (BasicBlock::iterator I = EntryBB.begin(), E = llvm::prior(EntryBB.end());
3426 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3427 Worklist.insert(AI);
3429 bool Changed = false;
3430 // A set of deleted alloca instruction pointers which should be removed from
3431 // the list of promotable allocas.
3432 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3435 while (!Worklist.empty()) {
3436 Changed |= runOnAlloca(*Worklist.pop_back_val());
3437 deleteDeadInstructions(DeletedAllocas);
3439 // Remove the deleted allocas from various lists so that we don't try to
3440 // continue processing them.
3441 if (!DeletedAllocas.empty()) {
3442 Worklist.remove_if(IsAllocaInSet(DeletedAllocas));
3443 PostPromotionWorklist.remove_if(IsAllocaInSet(DeletedAllocas));
3444 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3445 PromotableAllocas.end(),
3446 IsAllocaInSet(DeletedAllocas)),
3447 PromotableAllocas.end());
3448 DeletedAllocas.clear();
3452 Changed |= promoteAllocas(F);
3454 Worklist = PostPromotionWorklist;
3455 PostPromotionWorklist.clear();
3456 } while (!Worklist.empty());
3461 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3462 if (RequiresDomTree)
3463 AU.addRequired<DominatorTree>();
3464 AU.setPreservesCFG();