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 per alloca");
63 STATISTIC(NumAllocaPartitionUses, "Number of alloca partition uses rewritten");
64 STATISTIC(MaxUsesPerAllocaPartition, "Maximum number of uses of a partition");
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) {
656 AllocaSlices::AllocaSlices(const DataLayout &DL, AllocaInst &AI)
658 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
661 PointerEscapingInstr(0) {
662 SliceBuilder PB(DL, AI, *this);
663 SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
664 if (PtrI.isEscaped() || PtrI.isAborted()) {
665 // FIXME: We should sink the escape vs. abort info into the caller nicely,
666 // possibly by just storing the PtrInfo in the AllocaSlices.
667 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
668 : PtrI.getAbortingInst();
669 assert(PointerEscapingInstr && "Did not track a bad instruction");
673 Slices.erase(std::remove_if(Slices.begin(), Slices.end(),
674 std::mem_fun_ref(&Slice::isDead)),
677 // Sort the uses. This arranges for the offsets to be in ascending order,
678 // and the sizes to be in descending order.
679 std::sort(Slices.begin(), Slices.end());
682 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
684 void AllocaSlices::print(raw_ostream &OS, const_iterator I,
685 StringRef Indent) const {
686 printSlice(OS, I, Indent);
687 printUse(OS, I, Indent);
690 void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
691 StringRef Indent) const {
692 OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
693 << " slice #" << (I - begin())
694 << (I->isSplittable() ? " (splittable)" : "") << "\n";
697 void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
698 StringRef Indent) const {
699 OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
702 void AllocaSlices::print(raw_ostream &OS) const {
703 if (PointerEscapingInstr) {
704 OS << "Can't analyze slices for alloca: " << AI << "\n"
705 << " A pointer to this alloca escaped by:\n"
706 << " " << *PointerEscapingInstr << "\n";
710 OS << "Slices of alloca: " << AI << "\n";
711 for (const_iterator I = begin(), E = end(); I != E; ++I)
715 void AllocaSlices::dump(const_iterator I) const { print(dbgs(), I); }
716 void AllocaSlices::dump() const { print(dbgs()); }
718 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
721 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
723 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
724 /// the loads and stores of an alloca instruction, as well as updating its
725 /// debug information. This is used when a domtree is unavailable and thus
726 /// mem2reg in its full form can't be used to handle promotion of allocas to
728 class AllocaPromoter : public LoadAndStorePromoter {
732 SmallVector<DbgDeclareInst *, 4> DDIs;
733 SmallVector<DbgValueInst *, 4> DVIs;
736 AllocaPromoter(const SmallVectorImpl<Instruction *> &Insts, SSAUpdater &S,
737 AllocaInst &AI, DIBuilder &DIB)
738 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
740 void run(const SmallVectorImpl<Instruction*> &Insts) {
741 // Retain the debug information attached to the alloca for use when
742 // rewriting loads and stores.
743 if (MDNode *DebugNode = MDNode::getIfExists(AI.getContext(), &AI)) {
744 for (Value::use_iterator UI = DebugNode->use_begin(),
745 UE = DebugNode->use_end();
747 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
749 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI))
753 LoadAndStorePromoter::run(Insts);
755 // While we have the debug information, clear it off of the alloca. The
756 // caller takes care of deleting the alloca.
757 while (!DDIs.empty())
758 DDIs.pop_back_val()->eraseFromParent();
759 while (!DVIs.empty())
760 DVIs.pop_back_val()->eraseFromParent();
763 virtual bool isInstInList(Instruction *I,
764 const SmallVectorImpl<Instruction*> &Insts) const {
766 if (LoadInst *LI = dyn_cast<LoadInst>(I))
767 Ptr = LI->getOperand(0);
769 Ptr = cast<StoreInst>(I)->getPointerOperand();
771 // Only used to detect cycles, which will be rare and quickly found as
772 // we're walking up a chain of defs rather than down through uses.
773 SmallPtrSet<Value *, 4> Visited;
779 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr))
780 Ptr = BCI->getOperand(0);
781 else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Ptr))
782 Ptr = GEPI->getPointerOperand();
786 } while (Visited.insert(Ptr));
791 virtual void updateDebugInfo(Instruction *Inst) const {
792 for (SmallVectorImpl<DbgDeclareInst *>::const_iterator I = DDIs.begin(),
793 E = DDIs.end(); I != E; ++I) {
794 DbgDeclareInst *DDI = *I;
795 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
796 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
797 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
798 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
800 for (SmallVectorImpl<DbgValueInst *>::const_iterator I = DVIs.begin(),
801 E = DVIs.end(); I != E; ++I) {
802 DbgValueInst *DVI = *I;
804 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
805 // If an argument is zero extended then use argument directly. The ZExt
806 // may be zapped by an optimization pass in future.
807 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
808 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
809 else if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
810 Arg = dyn_cast<Argument>(SExt->getOperand(0));
812 Arg = SI->getValueOperand();
813 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
814 Arg = LI->getPointerOperand();
818 Instruction *DbgVal =
819 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
821 DbgVal->setDebugLoc(DVI->getDebugLoc());
825 } // end anon namespace
829 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
831 /// This pass takes allocations which can be completely analyzed (that is, they
832 /// don't escape) and tries to turn them into scalar SSA values. There are
833 /// a few steps to this process.
835 /// 1) It takes allocations of aggregates and analyzes the ways in which they
836 /// are used to try to split them into smaller allocations, ideally of
837 /// a single scalar data type. It will split up memcpy and memset accesses
838 /// as necessary and try to isolate individual scalar accesses.
839 /// 2) It will transform accesses into forms which are suitable for SSA value
840 /// promotion. This can be replacing a memset with a scalar store of an
841 /// integer value, or it can involve speculating operations on a PHI or
842 /// select to be a PHI or select of the results.
843 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
844 /// onto insert and extract operations on a vector value, and convert them to
845 /// this form. By doing so, it will enable promotion of vector aggregates to
846 /// SSA vector values.
847 class SROA : public FunctionPass {
848 const bool RequiresDomTree;
851 const DataLayout *DL;
854 /// \brief Worklist of alloca instructions to simplify.
856 /// Each alloca in the function is added to this. Each new alloca formed gets
857 /// added to it as well to recursively simplify unless that alloca can be
858 /// directly promoted. Finally, each time we rewrite a use of an alloca other
859 /// the one being actively rewritten, we add it back onto the list if not
860 /// already present to ensure it is re-visited.
861 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > Worklist;
863 /// \brief A collection of instructions to delete.
864 /// We try to batch deletions to simplify code and make things a bit more
866 SetVector<Instruction *, SmallVector<Instruction *, 8> > DeadInsts;
868 /// \brief Post-promotion worklist.
870 /// Sometimes we discover an alloca which has a high probability of becoming
871 /// viable for SROA after a round of promotion takes place. In those cases,
872 /// the alloca is enqueued here for re-processing.
874 /// Note that we have to be very careful to clear allocas out of this list in
875 /// the event they are deleted.
876 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > PostPromotionWorklist;
878 /// \brief A collection of alloca instructions we can directly promote.
879 std::vector<AllocaInst *> PromotableAllocas;
881 /// \brief A worklist of PHIs to speculate prior to promoting allocas.
883 /// All of these PHIs have been checked for the safety of speculation and by
884 /// being speculated will allow promoting allocas currently in the promotable
886 SetVector<PHINode *, SmallVector<PHINode *, 2> > SpeculatablePHIs;
888 /// \brief A worklist of select instructions to speculate prior to promoting
891 /// All of these select instructions have been checked for the safety of
892 /// speculation and by being speculated will allow promoting allocas
893 /// currently in the promotable queue.
894 SetVector<SelectInst *, SmallVector<SelectInst *, 2> > SpeculatableSelects;
897 SROA(bool RequiresDomTree = true)
898 : FunctionPass(ID), RequiresDomTree(RequiresDomTree),
900 initializeSROAPass(*PassRegistry::getPassRegistry());
902 bool runOnFunction(Function &F);
903 void getAnalysisUsage(AnalysisUsage &AU) const;
905 const char *getPassName() const { return "SROA"; }
909 friend class PHIOrSelectSpeculator;
910 friend class AllocaSliceRewriter;
912 bool rewritePartition(AllocaInst &AI, AllocaSlices &S,
913 AllocaSlices::iterator B, AllocaSlices::iterator E,
914 int64_t BeginOffset, int64_t EndOffset,
915 ArrayRef<AllocaSlices::iterator> SplitUses);
916 bool splitAlloca(AllocaInst &AI, AllocaSlices &S);
917 bool runOnAlloca(AllocaInst &AI);
918 void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas);
919 bool promoteAllocas(Function &F);
925 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
926 return new SROA(RequiresDomTree);
929 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates",
931 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
932 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
935 /// Walk the range of a partitioning looking for a common type to cover this
936 /// sequence of slices.
937 static Type *findCommonType(AllocaSlices::const_iterator B,
938 AllocaSlices::const_iterator E,
939 uint64_t EndOffset) {
941 bool IgnoreNonIntegralTypes = false;
942 for (AllocaSlices::const_iterator I = B; I != E; ++I) {
943 Use *U = I->getUse();
944 if (isa<IntrinsicInst>(*U->getUser()))
946 if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
950 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
951 UserTy = LI->getType();
952 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
953 UserTy = SI->getValueOperand()->getType();
955 IgnoreNonIntegralTypes = true; // Give up on anything but an iN type.
959 if (IntegerType *ITy = dyn_cast<IntegerType>(UserTy)) {
960 // If the type is larger than the partition, skip it. We only encounter
961 // this for split integer operations where we want to use the type of the
962 // entity causing the split. Also skip if the type is not a byte width
964 if (ITy->getBitWidth() % 8 != 0 ||
965 ITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
968 // If we have found an integer type use covering the alloca, use that
969 // regardless of the other types, as integers are often used for
970 // a "bucket of bits" type.
972 // NB: This *must* be the only return from inside the loop so that the
973 // order of slices doesn't impact the computed type.
975 } else if (IgnoreNonIntegralTypes) {
979 if (Ty && Ty != UserTy)
980 IgnoreNonIntegralTypes = true; // Give up on anything but an iN type.
987 /// PHI instructions that use an alloca and are subsequently loaded can be
988 /// rewritten to load both input pointers in the pred blocks and then PHI the
989 /// results, allowing the load of the alloca to be promoted.
991 /// %P2 = phi [i32* %Alloca, i32* %Other]
992 /// %V = load i32* %P2
994 /// %V1 = load i32* %Alloca -> will be mem2reg'd
996 /// %V2 = load i32* %Other
998 /// %V = phi [i32 %V1, i32 %V2]
1000 /// We can do this to a select if its only uses are loads and if the operands
1001 /// to the select can be loaded unconditionally.
1003 /// FIXME: This should be hoisted into a generic utility, likely in
1004 /// Transforms/Util/Local.h
1005 static bool isSafePHIToSpeculate(PHINode &PN,
1006 const DataLayout *DL = 0) {
1007 // For now, we can only do this promotion if the load is in the same block
1008 // as the PHI, and if there are no stores between the phi and load.
1009 // TODO: Allow recursive phi users.
1010 // TODO: Allow stores.
1011 BasicBlock *BB = PN.getParent();
1012 unsigned MaxAlign = 0;
1013 bool HaveLoad = false;
1014 for (Value::use_iterator UI = PN.use_begin(), UE = PN.use_end(); UI != UE;
1016 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1017 if (LI == 0 || !LI->isSimple())
1020 // For now we only allow loads in the same block as the PHI. This is
1021 // a common case that happens when instcombine merges two loads through
1023 if (LI->getParent() != BB)
1026 // Ensure that there are no instructions between the PHI and the load that
1028 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1029 if (BBI->mayWriteToMemory())
1032 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1039 // We can only transform this if it is safe to push the loads into the
1040 // predecessor blocks. The only thing to watch out for is that we can't put
1041 // a possibly trapping load in the predecessor if it is a critical edge.
1042 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1043 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1044 Value *InVal = PN.getIncomingValue(Idx);
1046 // If the value is produced by the terminator of the predecessor (an
1047 // invoke) or it has side-effects, there is no valid place to put a load
1048 // in the predecessor.
1049 if (TI == InVal || TI->mayHaveSideEffects())
1052 // If the predecessor has a single successor, then the edge isn't
1054 if (TI->getNumSuccessors() == 1)
1057 // If this pointer is always safe to load, or if we can prove that there
1058 // is already a load in the block, then we can move the load to the pred
1060 if (InVal->isDereferenceablePointer() ||
1061 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, DL))
1070 static void speculatePHINodeLoads(PHINode &PN) {
1071 DEBUG(dbgs() << " original: " << PN << "\n");
1073 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1074 IRBuilderTy PHIBuilder(&PN);
1075 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1076 PN.getName() + ".sroa.speculated");
1078 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1079 // matter which one we get and if any differ.
1080 LoadInst *SomeLoad = cast<LoadInst>(*PN.use_begin());
1081 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1082 unsigned Align = SomeLoad->getAlignment();
1084 // Rewrite all loads of the PN to use the new PHI.
1085 while (!PN.use_empty()) {
1086 LoadInst *LI = cast<LoadInst>(*PN.use_begin());
1087 LI->replaceAllUsesWith(NewPN);
1088 LI->eraseFromParent();
1091 // Inject loads into all of the pred blocks.
1092 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1093 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1094 TerminatorInst *TI = Pred->getTerminator();
1095 Value *InVal = PN.getIncomingValue(Idx);
1096 IRBuilderTy PredBuilder(TI);
1098 LoadInst *Load = PredBuilder.CreateLoad(
1099 InVal, (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1100 ++NumLoadsSpeculated;
1101 Load->setAlignment(Align);
1103 Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1104 NewPN->addIncoming(Load, Pred);
1107 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1108 PN.eraseFromParent();
1111 /// Select instructions that use an alloca and are subsequently loaded can be
1112 /// rewritten to load both input pointers and then select between the result,
1113 /// allowing the load of the alloca to be promoted.
1115 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1116 /// %V = load i32* %P2
1118 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1119 /// %V2 = load i32* %Other
1120 /// %V = select i1 %cond, i32 %V1, i32 %V2
1122 /// We can do this to a select if its only uses are loads and if the operand
1123 /// to the select can be loaded unconditionally.
1124 static bool isSafeSelectToSpeculate(SelectInst &SI, const DataLayout *DL = 0) {
1125 Value *TValue = SI.getTrueValue();
1126 Value *FValue = SI.getFalseValue();
1127 bool TDerefable = TValue->isDereferenceablePointer();
1128 bool FDerefable = FValue->isDereferenceablePointer();
1130 for (Value::use_iterator UI = SI.use_begin(), UE = SI.use_end(); UI != UE;
1132 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1133 if (LI == 0 || !LI->isSimple())
1136 // Both operands to the select need to be dereferencable, either
1137 // absolutely (e.g. allocas) or at this point because we can see other
1140 !isSafeToLoadUnconditionally(TValue, LI, LI->getAlignment(), DL))
1143 !isSafeToLoadUnconditionally(FValue, LI, LI->getAlignment(), DL))
1150 static void speculateSelectInstLoads(SelectInst &SI) {
1151 DEBUG(dbgs() << " original: " << SI << "\n");
1153 IRBuilderTy IRB(&SI);
1154 Value *TV = SI.getTrueValue();
1155 Value *FV = SI.getFalseValue();
1156 // Replace the loads of the select with a select of two loads.
1157 while (!SI.use_empty()) {
1158 LoadInst *LI = cast<LoadInst>(*SI.use_begin());
1159 assert(LI->isSimple() && "We only speculate simple loads");
1161 IRB.SetInsertPoint(LI);
1163 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1165 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1166 NumLoadsSpeculated += 2;
1168 // Transfer alignment and TBAA info if present.
1169 TL->setAlignment(LI->getAlignment());
1170 FL->setAlignment(LI->getAlignment());
1171 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
1172 TL->setMetadata(LLVMContext::MD_tbaa, Tag);
1173 FL->setMetadata(LLVMContext::MD_tbaa, Tag);
1176 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1177 LI->getName() + ".sroa.speculated");
1179 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1180 LI->replaceAllUsesWith(V);
1181 LI->eraseFromParent();
1183 SI.eraseFromParent();
1186 /// \brief Build a GEP out of a base pointer and indices.
1188 /// This will return the BasePtr if that is valid, or build a new GEP
1189 /// instruction using the IRBuilder if GEP-ing is needed.
1190 static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1191 SmallVectorImpl<Value *> &Indices) {
1192 if (Indices.empty())
1195 // A single zero index is a no-op, so check for this and avoid building a GEP
1197 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1200 return IRB.CreateInBoundsGEP(BasePtr, Indices, "idx");
1203 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1204 /// TargetTy without changing the offset of the pointer.
1206 /// This routine assumes we've already established a properly offset GEP with
1207 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1208 /// zero-indices down through type layers until we find one the same as
1209 /// TargetTy. If we can't find one with the same type, we at least try to use
1210 /// one with the same size. If none of that works, we just produce the GEP as
1211 /// indicated by Indices to have the correct offset.
1212 static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1213 Value *BasePtr, Type *Ty, Type *TargetTy,
1214 SmallVectorImpl<Value *> &Indices) {
1216 return buildGEP(IRB, BasePtr, Indices);
1218 // See if we can descend into a struct and locate a field with the correct
1220 unsigned NumLayers = 0;
1221 Type *ElementTy = Ty;
1223 if (ElementTy->isPointerTy())
1225 if (SequentialType *SeqTy = dyn_cast<SequentialType>(ElementTy)) {
1226 ElementTy = SeqTy->getElementType();
1227 // Note that we use the default address space as this index is over an
1228 // array or a vector, not a pointer.
1229 Indices.push_back(IRB.getInt(APInt(DL.getPointerSizeInBits(0), 0)));
1230 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1231 if (STy->element_begin() == STy->element_end())
1232 break; // Nothing left to descend into.
1233 ElementTy = *STy->element_begin();
1234 Indices.push_back(IRB.getInt32(0));
1239 } while (ElementTy != TargetTy);
1240 if (ElementTy != TargetTy)
1241 Indices.erase(Indices.end() - NumLayers, Indices.end());
1243 return buildGEP(IRB, BasePtr, Indices);
1246 /// \brief Recursively compute indices for a natural GEP.
1248 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1249 /// element types adding appropriate indices for the GEP.
1250 static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1251 Value *Ptr, Type *Ty, APInt &Offset,
1253 SmallVectorImpl<Value *> &Indices) {
1255 return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices);
1257 // We can't recurse through pointer types.
1258 if (Ty->isPointerTy())
1261 // We try to analyze GEPs over vectors here, but note that these GEPs are
1262 // extremely poorly defined currently. The long-term goal is to remove GEPing
1263 // over a vector from the IR completely.
1264 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1265 unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
1266 if (ElementSizeInBits % 8)
1267 return 0; // GEPs over non-multiple of 8 size vector elements are invalid.
1268 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1269 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1270 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1272 Offset -= NumSkippedElements * ElementSize;
1273 Indices.push_back(IRB.getInt(NumSkippedElements));
1274 return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1275 Offset, TargetTy, Indices);
1278 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1279 Type *ElementTy = ArrTy->getElementType();
1280 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1281 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1282 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1285 Offset -= NumSkippedElements * ElementSize;
1286 Indices.push_back(IRB.getInt(NumSkippedElements));
1287 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1291 StructType *STy = dyn_cast<StructType>(Ty);
1295 const StructLayout *SL = DL.getStructLayout(STy);
1296 uint64_t StructOffset = Offset.getZExtValue();
1297 if (StructOffset >= SL->getSizeInBytes())
1299 unsigned Index = SL->getElementContainingOffset(StructOffset);
1300 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1301 Type *ElementTy = STy->getElementType(Index);
1302 if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
1303 return 0; // The offset points into alignment padding.
1305 Indices.push_back(IRB.getInt32(Index));
1306 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1310 /// \brief Get a natural GEP from a base pointer to a particular offset and
1311 /// resulting in a particular type.
1313 /// The goal is to produce a "natural" looking GEP that works with the existing
1314 /// composite types to arrive at the appropriate offset and element type for
1315 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1316 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1317 /// Indices, and setting Ty to the result subtype.
1319 /// If no natural GEP can be constructed, this function returns null.
1320 static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1321 Value *Ptr, APInt Offset, Type *TargetTy,
1322 SmallVectorImpl<Value *> &Indices) {
1323 PointerType *Ty = cast<PointerType>(Ptr->getType());
1325 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1327 if (Ty == IRB.getInt8PtrTy() && TargetTy->isIntegerTy(8))
1330 Type *ElementTy = Ty->getElementType();
1331 if (!ElementTy->isSized())
1332 return 0; // We can't GEP through an unsized element.
1333 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1334 if (ElementSize == 0)
1335 return 0; // Zero-length arrays can't help us build a natural GEP.
1336 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1338 Offset -= NumSkippedElements * ElementSize;
1339 Indices.push_back(IRB.getInt(NumSkippedElements));
1340 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1344 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1345 /// resulting pointer has PointerTy.
1347 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1348 /// and produces the pointer type desired. Where it cannot, it will try to use
1349 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1350 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1351 /// bitcast to the type.
1353 /// The strategy for finding the more natural GEPs is to peel off layers of the
1354 /// pointer, walking back through bit casts and GEPs, searching for a base
1355 /// pointer from which we can compute a natural GEP with the desired
1356 /// properties. The algorithm tries to fold as many constant indices into
1357 /// a single GEP as possible, thus making each GEP more independent of the
1358 /// surrounding code.
1359 static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL,
1360 Value *Ptr, APInt Offset, Type *PointerTy) {
1361 // Even though we don't look through PHI nodes, we could be called on an
1362 // instruction in an unreachable block, which may be on a cycle.
1363 SmallPtrSet<Value *, 4> Visited;
1364 Visited.insert(Ptr);
1365 SmallVector<Value *, 4> Indices;
1367 // We may end up computing an offset pointer that has the wrong type. If we
1368 // never are able to compute one directly that has the correct type, we'll
1369 // fall back to it, so keep it around here.
1370 Value *OffsetPtr = 0;
1372 // Remember any i8 pointer we come across to re-use if we need to do a raw
1375 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1377 Type *TargetTy = PointerTy->getPointerElementType();
1380 // First fold any existing GEPs into the offset.
1381 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1382 APInt GEPOffset(Offset.getBitWidth(), 0);
1383 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1385 Offset += GEPOffset;
1386 Ptr = GEP->getPointerOperand();
1387 if (!Visited.insert(Ptr))
1391 // See if we can perform a natural GEP here.
1393 if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1395 if (P->getType() == PointerTy) {
1396 // Zap any offset pointer that we ended up computing in previous rounds.
1397 if (OffsetPtr && OffsetPtr->use_empty())
1398 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1399 I->eraseFromParent();
1407 // Stash this pointer if we've found an i8*.
1408 if (Ptr->getType()->isIntegerTy(8)) {
1410 Int8PtrOffset = Offset;
1413 // Peel off a layer of the pointer and update the offset appropriately.
1414 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1415 Ptr = cast<Operator>(Ptr)->getOperand(0);
1416 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1417 if (GA->mayBeOverridden())
1419 Ptr = GA->getAliasee();
1423 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1424 } while (Visited.insert(Ptr));
1428 Int8Ptr = IRB.CreateBitCast(Ptr, IRB.getInt8PtrTy(),
1430 Int8PtrOffset = Offset;
1433 OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
1434 IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1439 // On the off chance we were targeting i8*, guard the bitcast here.
1440 if (Ptr->getType() != PointerTy)
1441 Ptr = IRB.CreateBitCast(Ptr, PointerTy, "cast");
1446 /// \brief Test whether we can convert a value from the old to the new type.
1448 /// This predicate should be used to guard calls to convertValue in order to
1449 /// ensure that we only try to convert viable values. The strategy is that we
1450 /// will peel off single element struct and array wrappings to get to an
1451 /// underlying value, and convert that value.
1452 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1455 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1456 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1457 if (NewITy->getBitWidth() >= OldITy->getBitWidth())
1459 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1461 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1464 // We can convert pointers to integers and vice-versa. Same for vectors
1465 // of pointers and integers.
1466 OldTy = OldTy->getScalarType();
1467 NewTy = NewTy->getScalarType();
1468 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1469 if (NewTy->isPointerTy() && OldTy->isPointerTy())
1471 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
1479 /// \brief Generic routine to convert an SSA value to a value of a different
1482 /// This will try various different casting techniques, such as bitcasts,
1483 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1484 /// two types for viability with this routine.
1485 static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1487 Type *OldTy = V->getType();
1488 assert(canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type");
1493 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1494 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1495 if (NewITy->getBitWidth() > OldITy->getBitWidth())
1496 return IRB.CreateZExt(V, NewITy);
1498 // See if we need inttoptr for this type pair. A cast involving both scalars
1499 // and vectors requires and additional bitcast.
1500 if (OldTy->getScalarType()->isIntegerTy() &&
1501 NewTy->getScalarType()->isPointerTy()) {
1502 // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1503 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1504 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1507 // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1508 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1509 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1512 return IRB.CreateIntToPtr(V, NewTy);
1515 // See if we need ptrtoint for this type pair. A cast involving both scalars
1516 // and vectors requires and additional bitcast.
1517 if (OldTy->getScalarType()->isPointerTy() &&
1518 NewTy->getScalarType()->isIntegerTy()) {
1519 // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1520 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1521 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1524 // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1525 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1526 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1529 return IRB.CreatePtrToInt(V, NewTy);
1532 return IRB.CreateBitCast(V, NewTy);
1535 /// \brief Test whether the given slice use can be promoted to a vector.
1537 /// This function is called to test each entry in a partioning which is slated
1538 /// for a single slice.
1539 static bool isVectorPromotionViableForSlice(
1540 const DataLayout &DL, AllocaSlices &S, uint64_t SliceBeginOffset,
1541 uint64_t SliceEndOffset, VectorType *Ty, uint64_t ElementSize,
1542 AllocaSlices::const_iterator I) {
1543 // First validate the slice offsets.
1544 uint64_t BeginOffset =
1545 std::max(I->beginOffset(), SliceBeginOffset) - SliceBeginOffset;
1546 uint64_t BeginIndex = BeginOffset / ElementSize;
1547 if (BeginIndex * ElementSize != BeginOffset ||
1548 BeginIndex >= Ty->getNumElements())
1550 uint64_t EndOffset =
1551 std::min(I->endOffset(), SliceEndOffset) - SliceBeginOffset;
1552 uint64_t EndIndex = EndOffset / ElementSize;
1553 if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
1556 assert(EndIndex > BeginIndex && "Empty vector!");
1557 uint64_t NumElements = EndIndex - BeginIndex;
1559 (NumElements == 1) ? Ty->getElementType()
1560 : VectorType::get(Ty->getElementType(), NumElements);
1563 Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1565 Use *U = I->getUse();
1567 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1568 if (MI->isVolatile())
1570 if (!I->isSplittable())
1571 return false; // Skip any unsplittable intrinsics.
1572 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1573 // Disable vector promotion when there are loads or stores of an FCA.
1575 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1576 if (LI->isVolatile())
1578 Type *LTy = LI->getType();
1579 if (SliceBeginOffset > I->beginOffset() ||
1580 SliceEndOffset < I->endOffset()) {
1581 assert(LTy->isIntegerTy());
1584 if (!canConvertValue(DL, SliceTy, LTy))
1586 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1587 if (SI->isVolatile())
1589 Type *STy = SI->getValueOperand()->getType();
1590 if (SliceBeginOffset > I->beginOffset() ||
1591 SliceEndOffset < I->endOffset()) {
1592 assert(STy->isIntegerTy());
1595 if (!canConvertValue(DL, STy, SliceTy))
1604 /// \brief Test whether the given alloca partitioning and range of slices can be
1605 /// promoted to a vector.
1607 /// This is a quick test to check whether we can rewrite a particular alloca
1608 /// partition (and its newly formed alloca) into a vector alloca with only
1609 /// whole-vector loads and stores such that it could be promoted to a vector
1610 /// SSA value. We only can ensure this for a limited set of operations, and we
1611 /// don't want to do the rewrites unless we are confident that the result will
1612 /// be promotable, so we have an early test here.
1614 isVectorPromotionViable(const DataLayout &DL, Type *AllocaTy, AllocaSlices &S,
1615 uint64_t SliceBeginOffset, uint64_t SliceEndOffset,
1616 AllocaSlices::const_iterator I,
1617 AllocaSlices::const_iterator E,
1618 ArrayRef<AllocaSlices::iterator> SplitUses) {
1619 VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
1623 uint64_t ElementSize = DL.getTypeSizeInBits(Ty->getScalarType());
1625 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1626 // that aren't byte sized.
1627 if (ElementSize % 8)
1629 assert((DL.getTypeSizeInBits(Ty) % 8) == 0 &&
1630 "vector size not a multiple of element size?");
1634 if (!isVectorPromotionViableForSlice(DL, S, SliceBeginOffset,
1635 SliceEndOffset, Ty, ElementSize, I))
1638 for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
1639 SUE = SplitUses.end();
1641 if (!isVectorPromotionViableForSlice(DL, S, SliceBeginOffset,
1642 SliceEndOffset, Ty, ElementSize, *SUI))
1648 /// \brief Test whether a slice of an alloca is valid for integer widening.
1650 /// This implements the necessary checking for the \c isIntegerWideningViable
1651 /// test below on a single slice of the alloca.
1652 static bool isIntegerWideningViableForSlice(const DataLayout &DL,
1654 uint64_t AllocBeginOffset,
1655 uint64_t Size, AllocaSlices &S,
1656 AllocaSlices::const_iterator I,
1657 bool &WholeAllocaOp) {
1658 uint64_t RelBegin = I->beginOffset() - AllocBeginOffset;
1659 uint64_t RelEnd = I->endOffset() - AllocBeginOffset;
1661 // We can't reasonably handle cases where the load or store extends past
1662 // the end of the aloca's type and into its padding.
1666 Use *U = I->getUse();
1668 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1669 if (LI->isVolatile())
1671 if (RelBegin == 0 && RelEnd == Size)
1672 WholeAllocaOp = true;
1673 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
1674 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1676 } else if (RelBegin != 0 || RelEnd != Size ||
1677 !canConvertValue(DL, AllocaTy, LI->getType())) {
1678 // Non-integer loads need to be convertible from the alloca type so that
1679 // they are promotable.
1682 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1683 Type *ValueTy = SI->getValueOperand()->getType();
1684 if (SI->isVolatile())
1686 if (RelBegin == 0 && RelEnd == Size)
1687 WholeAllocaOp = true;
1688 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
1689 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1691 } else if (RelBegin != 0 || RelEnd != Size ||
1692 !canConvertValue(DL, ValueTy, AllocaTy)) {
1693 // Non-integer stores need to be convertible to the alloca type so that
1694 // they are promotable.
1697 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1698 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
1700 if (!I->isSplittable())
1701 return false; // Skip any unsplittable intrinsics.
1702 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1703 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1704 II->getIntrinsicID() != Intrinsic::lifetime_end)
1713 /// \brief Test whether the given alloca partition's integer operations can be
1714 /// widened to promotable ones.
1716 /// This is a quick test to check whether we can rewrite the integer loads and
1717 /// stores to a particular alloca into wider loads and stores and be able to
1718 /// promote the resulting alloca.
1720 isIntegerWideningViable(const DataLayout &DL, Type *AllocaTy,
1721 uint64_t AllocBeginOffset, AllocaSlices &S,
1722 AllocaSlices::const_iterator I,
1723 AllocaSlices::const_iterator E,
1724 ArrayRef<AllocaSlices::iterator> SplitUses) {
1725 uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
1726 // Don't create integer types larger than the maximum bitwidth.
1727 if (SizeInBits > IntegerType::MAX_INT_BITS)
1730 // Don't try to handle allocas with bit-padding.
1731 if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
1734 // We need to ensure that an integer type with the appropriate bitwidth can
1735 // be converted to the alloca type, whatever that is. We don't want to force
1736 // the alloca itself to have an integer type if there is a more suitable one.
1737 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
1738 if (!canConvertValue(DL, AllocaTy, IntTy) ||
1739 !canConvertValue(DL, IntTy, AllocaTy))
1742 uint64_t Size = DL.getTypeStoreSize(AllocaTy);
1744 // While examining uses, we ensure that the alloca has a covering load or
1745 // store. We don't want to widen the integer operations only to fail to
1746 // promote due to some other unsplittable entry (which we may make splittable
1747 // later). However, if there are only splittable uses, go ahead and assume
1748 // that we cover the alloca.
1749 bool WholeAllocaOp = (I != E) ? false : DL.isLegalInteger(SizeInBits);
1752 if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
1753 S, I, WholeAllocaOp))
1756 for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
1757 SUE = SplitUses.end();
1759 if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
1760 S, *SUI, WholeAllocaOp))
1763 return WholeAllocaOp;
1766 static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1767 IntegerType *Ty, uint64_t Offset,
1768 const Twine &Name) {
1769 DEBUG(dbgs() << " start: " << *V << "\n");
1770 IntegerType *IntTy = cast<IntegerType>(V->getType());
1771 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
1772 "Element extends past full value");
1773 uint64_t ShAmt = 8*Offset;
1774 if (DL.isBigEndian())
1775 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
1777 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
1778 DEBUG(dbgs() << " shifted: " << *V << "\n");
1780 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
1781 "Cannot extract to a larger integer!");
1783 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
1784 DEBUG(dbgs() << " trunced: " << *V << "\n");
1789 static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
1790 Value *V, uint64_t Offset, const Twine &Name) {
1791 IntegerType *IntTy = cast<IntegerType>(Old->getType());
1792 IntegerType *Ty = cast<IntegerType>(V->getType());
1793 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
1794 "Cannot insert a larger integer!");
1795 DEBUG(dbgs() << " start: " << *V << "\n");
1797 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
1798 DEBUG(dbgs() << " extended: " << *V << "\n");
1800 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
1801 "Element store outside of alloca store");
1802 uint64_t ShAmt = 8*Offset;
1803 if (DL.isBigEndian())
1804 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
1806 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
1807 DEBUG(dbgs() << " shifted: " << *V << "\n");
1810 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
1811 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
1812 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
1813 DEBUG(dbgs() << " masked: " << *Old << "\n");
1814 V = IRB.CreateOr(Old, V, Name + ".insert");
1815 DEBUG(dbgs() << " inserted: " << *V << "\n");
1820 static Value *extractVector(IRBuilderTy &IRB, Value *V,
1821 unsigned BeginIndex, unsigned EndIndex,
1822 const Twine &Name) {
1823 VectorType *VecTy = cast<VectorType>(V->getType());
1824 unsigned NumElements = EndIndex - BeginIndex;
1825 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
1827 if (NumElements == VecTy->getNumElements())
1830 if (NumElements == 1) {
1831 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
1833 DEBUG(dbgs() << " extract: " << *V << "\n");
1837 SmallVector<Constant*, 8> Mask;
1838 Mask.reserve(NumElements);
1839 for (unsigned i = BeginIndex; i != EndIndex; ++i)
1840 Mask.push_back(IRB.getInt32(i));
1841 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
1842 ConstantVector::get(Mask),
1844 DEBUG(dbgs() << " shuffle: " << *V << "\n");
1848 static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
1849 unsigned BeginIndex, const Twine &Name) {
1850 VectorType *VecTy = cast<VectorType>(Old->getType());
1851 assert(VecTy && "Can only insert a vector into a vector");
1853 VectorType *Ty = dyn_cast<VectorType>(V->getType());
1855 // Single element to insert.
1856 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
1858 DEBUG(dbgs() << " insert: " << *V << "\n");
1862 assert(Ty->getNumElements() <= VecTy->getNumElements() &&
1863 "Too many elements!");
1864 if (Ty->getNumElements() == VecTy->getNumElements()) {
1865 assert(V->getType() == VecTy && "Vector type mismatch");
1868 unsigned EndIndex = BeginIndex + Ty->getNumElements();
1870 // When inserting a smaller vector into the larger to store, we first
1871 // use a shuffle vector to widen it with undef elements, and then
1872 // a second shuffle vector to select between the loaded vector and the
1874 SmallVector<Constant*, 8> Mask;
1875 Mask.reserve(VecTy->getNumElements());
1876 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
1877 if (i >= BeginIndex && i < EndIndex)
1878 Mask.push_back(IRB.getInt32(i - BeginIndex));
1880 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
1881 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
1882 ConstantVector::get(Mask),
1884 DEBUG(dbgs() << " shuffle: " << *V << "\n");
1887 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
1888 Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
1890 V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
1892 DEBUG(dbgs() << " blend: " << *V << "\n");
1897 /// \brief Visitor to rewrite instructions using p particular slice of an alloca
1898 /// to use a new alloca.
1900 /// Also implements the rewriting to vector-based accesses when the partition
1901 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
1903 class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
1904 // Befriend the base class so it can delegate to private visit methods.
1905 friend class llvm::InstVisitor<AllocaSliceRewriter, bool>;
1906 typedef llvm::InstVisitor<AllocaSliceRewriter, bool> Base;
1908 const DataLayout &DL;
1911 AllocaInst &OldAI, &NewAI;
1912 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
1915 // If we are rewriting an alloca partition which can be written as pure
1916 // vector operations, we stash extra information here. When VecTy is
1917 // non-null, we have some strict guarantees about the rewritten alloca:
1918 // - The new alloca is exactly the size of the vector type here.
1919 // - The accesses all either map to the entire vector or to a single
1921 // - The set of accessing instructions is only one of those handled above
1922 // in isVectorPromotionViable. Generally these are the same access kinds
1923 // which are promotable via mem2reg.
1926 uint64_t ElementSize;
1928 // This is a convenience and flag variable that will be null unless the new
1929 // alloca's integer operations should be widened to this integer type due to
1930 // passing isIntegerWideningViable above. If it is non-null, the desired
1931 // integer type will be stored here for easy access during rewriting.
1934 // The offset of the slice currently being rewritten.
1935 uint64_t BeginOffset, EndOffset;
1939 Instruction *OldPtr;
1941 // Output members carrying state about the result of visiting and rewriting
1942 // the slice of the alloca.
1943 bool IsUsedByRewrittenSpeculatableInstructions;
1945 // Utility IR builder, whose name prefix is setup for each visited use, and
1946 // the insertion point is set to point to the user.
1950 AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &S, SROA &Pass,
1951 AllocaInst &OldAI, AllocaInst &NewAI,
1952 uint64_t NewBeginOffset, uint64_t NewEndOffset,
1953 bool IsVectorPromotable = false,
1954 bool IsIntegerPromotable = false)
1955 : DL(DL), S(S), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
1956 NewAllocaBeginOffset(NewBeginOffset), NewAllocaEndOffset(NewEndOffset),
1957 NewAllocaTy(NewAI.getAllocatedType()),
1958 VecTy(IsVectorPromotable ? cast<VectorType>(NewAllocaTy) : 0),
1959 ElementTy(VecTy ? VecTy->getElementType() : 0),
1960 ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
1961 IntTy(IsIntegerPromotable
1964 DL.getTypeSizeInBits(NewAI.getAllocatedType()))
1966 BeginOffset(), EndOffset(), IsSplittable(), IsSplit(), OldUse(),
1967 OldPtr(), IsUsedByRewrittenSpeculatableInstructions(false),
1968 IRB(NewAI.getContext(), ConstantFolder()) {
1970 assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
1971 "Only multiple-of-8 sized vector elements are viable");
1974 assert((!IsVectorPromotable && !IsIntegerPromotable) ||
1975 IsVectorPromotable != IsIntegerPromotable);
1978 bool visit(AllocaSlices::const_iterator I) {
1979 bool CanSROA = true;
1980 BeginOffset = I->beginOffset();
1981 EndOffset = I->endOffset();
1982 IsSplittable = I->isSplittable();
1984 BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
1986 OldUse = I->getUse();
1987 OldPtr = cast<Instruction>(OldUse->get());
1989 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
1990 IRB.SetInsertPoint(OldUserI);
1991 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
1992 IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
1994 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2000 /// \brief Query whether this slice is used by speculatable instructions after
2003 /// These instructions (PHIs and Selects currently) require the alloca slice
2004 /// to run back through the rewriter. Thus, they are promotable, but not on
2005 /// this iteration. This is distinct from a slice which is unpromotable for
2006 /// some other reason, in which case we don't even want to perform the
2007 /// speculation. This can be querried at any time and reflects whether (at
2008 /// that point) a visit call has rewritten a speculatable instruction on the
2010 bool isUsedByRewrittenSpeculatableInstructions() const {
2011 return IsUsedByRewrittenSpeculatableInstructions;
2015 // Make sure the other visit overloads are visible.
2018 // Every instruction which can end up as a user must have a rewrite rule.
2019 bool visitInstruction(Instruction &I) {
2020 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2021 llvm_unreachable("No rewrite rule for this instruction!");
2024 Value *getAdjustedAllocaPtr(IRBuilderTy &IRB, uint64_t Offset,
2026 assert(Offset >= NewAllocaBeginOffset);
2027 return getAdjustedPtr(IRB, DL, &NewAI, APInt(DL.getPointerSizeInBits(),
2028 Offset - NewAllocaBeginOffset),
2032 /// \brief Compute suitable alignment to access an offset into the new alloca.
2033 unsigned getOffsetAlign(uint64_t Offset) {
2034 unsigned NewAIAlign = NewAI.getAlignment();
2036 NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
2037 return MinAlign(NewAIAlign, Offset);
2040 /// \brief Compute suitable alignment to access a type at an offset of the
2043 /// \returns zero if the type's ABI alignment is a suitable alignment,
2044 /// otherwise returns the maximal suitable alignment.
2045 unsigned getOffsetTypeAlign(Type *Ty, uint64_t Offset) {
2046 unsigned Align = getOffsetAlign(Offset);
2047 return Align == DL.getABITypeAlignment(Ty) ? 0 : Align;
2050 unsigned getIndex(uint64_t Offset) {
2051 assert(VecTy && "Can only call getIndex when rewriting a vector");
2052 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2053 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2054 uint32_t Index = RelOffset / ElementSize;
2055 assert(Index * ElementSize == RelOffset);
2059 void deleteIfTriviallyDead(Value *V) {
2060 Instruction *I = cast<Instruction>(V);
2061 if (isInstructionTriviallyDead(I))
2062 Pass.DeadInsts.insert(I);
2065 Value *rewriteVectorizedLoadInst(uint64_t NewBeginOffset,
2066 uint64_t NewEndOffset) {
2067 unsigned BeginIndex = getIndex(NewBeginOffset);
2068 unsigned EndIndex = getIndex(NewEndOffset);
2069 assert(EndIndex > BeginIndex && "Empty vector!");
2071 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2073 return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2076 Value *rewriteIntegerLoad(LoadInst &LI, uint64_t NewBeginOffset,
2077 uint64_t NewEndOffset) {
2078 assert(IntTy && "We cannot insert an integer to the alloca");
2079 assert(!LI.isVolatile());
2080 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2082 V = convertValue(DL, IRB, V, IntTy);
2083 assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2084 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2085 if (Offset > 0 || NewEndOffset < NewAllocaEndOffset)
2086 V = extractInteger(DL, IRB, V, cast<IntegerType>(LI.getType()), Offset,
2091 bool visitLoadInst(LoadInst &LI) {
2092 DEBUG(dbgs() << " original: " << LI << "\n");
2093 Value *OldOp = LI.getOperand(0);
2094 assert(OldOp == OldPtr);
2096 // Compute the intersecting offset range.
2097 assert(BeginOffset < NewAllocaEndOffset);
2098 assert(EndOffset > NewAllocaBeginOffset);
2099 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2100 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2102 uint64_t Size = NewEndOffset - NewBeginOffset;
2104 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), Size * 8)
2106 bool IsPtrAdjusted = false;
2109 V = rewriteVectorizedLoadInst(NewBeginOffset, NewEndOffset);
2110 } else if (IntTy && LI.getType()->isIntegerTy()) {
2111 V = rewriteIntegerLoad(LI, NewBeginOffset, NewEndOffset);
2112 } else if (NewBeginOffset == NewAllocaBeginOffset &&
2113 canConvertValue(DL, NewAllocaTy, LI.getType())) {
2114 V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2115 LI.isVolatile(), "load");
2117 Type *LTy = TargetTy->getPointerTo();
2118 V = IRB.CreateAlignedLoad(
2119 getAdjustedAllocaPtr(IRB, NewBeginOffset, LTy),
2120 getOffsetTypeAlign(TargetTy, NewBeginOffset - NewAllocaBeginOffset),
2121 LI.isVolatile(), "load");
2122 IsPtrAdjusted = true;
2124 V = convertValue(DL, IRB, V, TargetTy);
2127 assert(!LI.isVolatile());
2128 assert(LI.getType()->isIntegerTy() &&
2129 "Only integer type loads and stores are split");
2130 assert(Size < DL.getTypeStoreSize(LI.getType()) &&
2131 "Split load isn't smaller than original load");
2132 assert(LI.getType()->getIntegerBitWidth() ==
2133 DL.getTypeStoreSizeInBits(LI.getType()) &&
2134 "Non-byte-multiple bit width");
2135 // Move the insertion point just past the load so that we can refer to it.
2136 IRB.SetInsertPoint(llvm::next(BasicBlock::iterator(&LI)));
2137 // Create a placeholder value with the same type as LI to use as the
2138 // basis for the new value. This allows us to replace the uses of LI with
2139 // the computed value, and then replace the placeholder with LI, leaving
2140 // LI only used for this computation.
2142 = new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
2143 V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset,
2145 LI.replaceAllUsesWith(V);
2146 Placeholder->replaceAllUsesWith(&LI);
2149 LI.replaceAllUsesWith(V);
2152 Pass.DeadInsts.insert(&LI);
2153 deleteIfTriviallyDead(OldOp);
2154 DEBUG(dbgs() << " to: " << *V << "\n");
2155 return !LI.isVolatile() && !IsPtrAdjusted;
2158 bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp,
2159 uint64_t NewBeginOffset,
2160 uint64_t NewEndOffset) {
2161 if (V->getType() != VecTy) {
2162 unsigned BeginIndex = getIndex(NewBeginOffset);
2163 unsigned EndIndex = getIndex(NewEndOffset);
2164 assert(EndIndex > BeginIndex && "Empty vector!");
2165 unsigned NumElements = EndIndex - BeginIndex;
2166 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2168 (NumElements == 1) ? ElementTy
2169 : VectorType::get(ElementTy, NumElements);
2170 if (V->getType() != SliceTy)
2171 V = convertValue(DL, IRB, V, SliceTy);
2173 // Mix in the existing elements.
2174 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2176 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2178 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2179 Pass.DeadInsts.insert(&SI);
2182 DEBUG(dbgs() << " to: " << *Store << "\n");
2186 bool rewriteIntegerStore(Value *V, StoreInst &SI,
2187 uint64_t NewBeginOffset, uint64_t NewEndOffset) {
2188 assert(IntTy && "We cannot extract an integer from the alloca");
2189 assert(!SI.isVolatile());
2190 if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2191 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2193 Old = convertValue(DL, IRB, Old, IntTy);
2194 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2195 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2196 V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset,
2199 V = convertValue(DL, IRB, V, NewAllocaTy);
2200 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2201 Pass.DeadInsts.insert(&SI);
2203 DEBUG(dbgs() << " to: " << *Store << "\n");
2207 bool visitStoreInst(StoreInst &SI) {
2208 DEBUG(dbgs() << " original: " << SI << "\n");
2209 Value *OldOp = SI.getOperand(1);
2210 assert(OldOp == OldPtr);
2212 Value *V = SI.getValueOperand();
2214 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2215 // alloca that should be re-examined after promoting this alloca.
2216 if (V->getType()->isPointerTy())
2217 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2218 Pass.PostPromotionWorklist.insert(AI);
2220 // Compute the intersecting offset range.
2221 assert(BeginOffset < NewAllocaEndOffset);
2222 assert(EndOffset > NewAllocaBeginOffset);
2223 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2224 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2226 uint64_t Size = NewEndOffset - NewBeginOffset;
2227 if (Size < DL.getTypeStoreSize(V->getType())) {
2228 assert(!SI.isVolatile());
2229 assert(V->getType()->isIntegerTy() &&
2230 "Only integer type loads and stores are split");
2231 assert(V->getType()->getIntegerBitWidth() ==
2232 DL.getTypeStoreSizeInBits(V->getType()) &&
2233 "Non-byte-multiple bit width");
2234 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), Size * 8);
2235 V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset,
2240 return rewriteVectorizedStoreInst(V, SI, OldOp, NewBeginOffset,
2242 if (IntTy && V->getType()->isIntegerTy())
2243 return rewriteIntegerStore(V, SI, NewBeginOffset, NewEndOffset);
2246 if (NewBeginOffset == NewAllocaBeginOffset &&
2247 NewEndOffset == NewAllocaEndOffset &&
2248 canConvertValue(DL, V->getType(), NewAllocaTy)) {
2249 V = convertValue(DL, IRB, V, NewAllocaTy);
2250 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2253 Value *NewPtr = getAdjustedAllocaPtr(IRB, NewBeginOffset,
2254 V->getType()->getPointerTo());
2255 NewSI = IRB.CreateAlignedStore(
2256 V, NewPtr, getOffsetTypeAlign(
2257 V->getType(), NewBeginOffset - NewAllocaBeginOffset),
2261 Pass.DeadInsts.insert(&SI);
2262 deleteIfTriviallyDead(OldOp);
2264 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2265 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2268 /// \brief Compute an integer value from splatting an i8 across the given
2269 /// number of bytes.
2271 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2272 /// call this routine.
2273 /// FIXME: Heed the advice above.
2275 /// \param V The i8 value to splat.
2276 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2277 Value *getIntegerSplat(Value *V, unsigned Size) {
2278 assert(Size > 0 && "Expected a positive number of bytes.");
2279 IntegerType *VTy = cast<IntegerType>(V->getType());
2280 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2284 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size*8);
2285 V = IRB.CreateMul(IRB.CreateZExt(V, SplatIntTy, "zext"),
2286 ConstantExpr::getUDiv(
2287 Constant::getAllOnesValue(SplatIntTy),
2288 ConstantExpr::getZExt(
2289 Constant::getAllOnesValue(V->getType()),
2295 /// \brief Compute a vector splat for a given element value.
2296 Value *getVectorSplat(Value *V, unsigned NumElements) {
2297 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2298 DEBUG(dbgs() << " splat: " << *V << "\n");
2302 bool visitMemSetInst(MemSetInst &II) {
2303 DEBUG(dbgs() << " original: " << II << "\n");
2304 assert(II.getRawDest() == OldPtr);
2306 // If the memset has a variable size, it cannot be split, just adjust the
2307 // pointer to the new alloca.
2308 if (!isa<Constant>(II.getLength())) {
2310 assert(BeginOffset >= NewAllocaBeginOffset);
2312 getAdjustedAllocaPtr(IRB, BeginOffset, II.getRawDest()->getType()));
2313 Type *CstTy = II.getAlignmentCst()->getType();
2314 II.setAlignment(ConstantInt::get(CstTy, getOffsetAlign(BeginOffset)));
2316 deleteIfTriviallyDead(OldPtr);
2320 // Record this instruction for deletion.
2321 Pass.DeadInsts.insert(&II);
2323 Type *AllocaTy = NewAI.getAllocatedType();
2324 Type *ScalarTy = AllocaTy->getScalarType();
2326 // Compute the intersecting offset range.
2327 assert(BeginOffset < NewAllocaEndOffset);
2328 assert(EndOffset > NewAllocaBeginOffset);
2329 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2330 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2331 uint64_t SliceOffset = NewBeginOffset - NewAllocaBeginOffset;
2333 // If this doesn't map cleanly onto the alloca type, and that type isn't
2334 // a single value type, just emit a memset.
2335 if (!VecTy && !IntTy &&
2336 (BeginOffset > NewAllocaBeginOffset ||
2337 EndOffset < NewAllocaEndOffset ||
2338 !AllocaTy->isSingleValueType() ||
2339 !DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy)) ||
2340 DL.getTypeSizeInBits(ScalarTy)%8 != 0)) {
2341 Type *SizeTy = II.getLength()->getType();
2342 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2343 CallInst *New = IRB.CreateMemSet(
2344 getAdjustedAllocaPtr(IRB, NewBeginOffset, II.getRawDest()->getType()),
2345 II.getValue(), Size, getOffsetAlign(SliceOffset), II.isVolatile());
2347 DEBUG(dbgs() << " to: " << *New << "\n");
2351 // If we can represent this as a simple value, we have to build the actual
2352 // value to store, which requires expanding the byte present in memset to
2353 // a sensible representation for the alloca type. This is essentially
2354 // splatting the byte to a sufficiently wide integer, splatting it across
2355 // any desired vector width, and bitcasting to the final type.
2359 // If this is a memset of a vectorized alloca, insert it.
2360 assert(ElementTy == ScalarTy);
2362 unsigned BeginIndex = getIndex(NewBeginOffset);
2363 unsigned EndIndex = getIndex(NewEndOffset);
2364 assert(EndIndex > BeginIndex && "Empty vector!");
2365 unsigned NumElements = EndIndex - BeginIndex;
2366 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2369 getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
2370 Splat = convertValue(DL, IRB, Splat, ElementTy);
2371 if (NumElements > 1)
2372 Splat = getVectorSplat(Splat, NumElements);
2374 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2376 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2378 // If this is a memset on an alloca where we can widen stores, insert the
2380 assert(!II.isVolatile());
2382 uint64_t Size = NewEndOffset - NewBeginOffset;
2383 V = getIntegerSplat(II.getValue(), Size);
2385 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2386 EndOffset != NewAllocaBeginOffset)) {
2387 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2389 Old = convertValue(DL, IRB, Old, IntTy);
2390 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2391 V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2393 assert(V->getType() == IntTy &&
2394 "Wrong type for an alloca wide integer!");
2396 V = convertValue(DL, IRB, V, AllocaTy);
2398 // Established these invariants above.
2399 assert(NewBeginOffset == NewAllocaBeginOffset);
2400 assert(NewEndOffset == NewAllocaEndOffset);
2402 V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
2403 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2404 V = getVectorSplat(V, AllocaVecTy->getNumElements());
2406 V = convertValue(DL, IRB, V, AllocaTy);
2409 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2412 DEBUG(dbgs() << " to: " << *New << "\n");
2413 return !II.isVolatile();
2416 bool visitMemTransferInst(MemTransferInst &II) {
2417 // Rewriting of memory transfer instructions can be a bit tricky. We break
2418 // them into two categories: split intrinsics and unsplit intrinsics.
2420 DEBUG(dbgs() << " original: " << II << "\n");
2422 // Compute the intersecting offset range.
2423 assert(BeginOffset < NewAllocaEndOffset);
2424 assert(EndOffset > NewAllocaBeginOffset);
2425 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2426 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2428 assert(II.getRawSource() == OldPtr || II.getRawDest() == OldPtr);
2429 bool IsDest = II.getRawDest() == OldPtr;
2431 // Compute the relative offset within the transfer.
2432 unsigned IntPtrWidth = DL.getPointerSizeInBits();
2433 APInt RelOffset(IntPtrWidth, NewBeginOffset - BeginOffset);
2435 unsigned Align = II.getAlignment();
2436 uint64_t SliceOffset = NewBeginOffset - NewAllocaBeginOffset;
2439 MinAlign(RelOffset.zextOrTrunc(64).getZExtValue(),
2440 MinAlign(II.getAlignment(), getOffsetAlign(SliceOffset)));
2442 // For unsplit intrinsics, we simply modify the source and destination
2443 // pointers in place. This isn't just an optimization, it is a matter of
2444 // correctness. With unsplit intrinsics we may be dealing with transfers
2445 // within a single alloca before SROA ran, or with transfers that have
2446 // a variable length. We may also be dealing with memmove instead of
2447 // memcpy, and so simply updating the pointers is the necessary for us to
2448 // update both source and dest of a single call.
2449 if (!IsSplittable) {
2450 Value *OldOp = IsDest ? II.getRawDest() : II.getRawSource();
2453 getAdjustedAllocaPtr(IRB, BeginOffset, II.getRawDest()->getType()));
2455 II.setSource(getAdjustedAllocaPtr(IRB, BeginOffset,
2456 II.getRawSource()->getType()));
2458 Type *CstTy = II.getAlignmentCst()->getType();
2459 II.setAlignment(ConstantInt::get(CstTy, Align));
2461 DEBUG(dbgs() << " to: " << II << "\n");
2462 deleteIfTriviallyDead(OldOp);
2465 // For split transfer intrinsics we have an incredibly useful assurance:
2466 // the source and destination do not reside within the same alloca, and at
2467 // least one of them does not escape. This means that we can replace
2468 // memmove with memcpy, and we don't need to worry about all manner of
2469 // downsides to splitting and transforming the operations.
2471 // If this doesn't map cleanly onto the alloca type, and that type isn't
2472 // a single value type, just emit a memcpy.
2474 = !VecTy && !IntTy && (BeginOffset > NewAllocaBeginOffset ||
2475 EndOffset < NewAllocaEndOffset ||
2476 !NewAI.getAllocatedType()->isSingleValueType());
2478 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2479 // size hasn't been shrunk based on analysis of the viable range, this is
2481 if (EmitMemCpy && &OldAI == &NewAI) {
2482 // Ensure the start lines up.
2483 assert(NewBeginOffset == BeginOffset);
2485 // Rewrite the size as needed.
2486 if (NewEndOffset != EndOffset)
2487 II.setLength(ConstantInt::get(II.getLength()->getType(),
2488 NewEndOffset - NewBeginOffset));
2491 // Record this instruction for deletion.
2492 Pass.DeadInsts.insert(&II);
2494 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2495 // alloca that should be re-examined after rewriting this instruction.
2496 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2498 = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets()))
2499 Pass.Worklist.insert(AI);
2502 Type *OtherPtrTy = IsDest ? II.getRawSource()->getType()
2503 : II.getRawDest()->getType();
2505 // Compute the other pointer, folding as much as possible to produce
2506 // a single, simple GEP in most cases.
2507 OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, RelOffset, OtherPtrTy);
2509 Value *OurPtr = getAdjustedAllocaPtr(
2510 IRB, NewBeginOffset,
2511 IsDest ? II.getRawDest()->getType() : II.getRawSource()->getType());
2512 Type *SizeTy = II.getLength()->getType();
2513 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2515 CallInst *New = IRB.CreateMemCpy(IsDest ? OurPtr : OtherPtr,
2516 IsDest ? OtherPtr : OurPtr,
2517 Size, Align, II.isVolatile());
2519 DEBUG(dbgs() << " to: " << *New << "\n");
2523 // Note that we clamp the alignment to 1 here as a 0 alignment for a memcpy
2524 // is equivalent to 1, but that isn't true if we end up rewriting this as
2529 bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
2530 NewEndOffset == NewAllocaEndOffset;
2531 uint64_t Size = NewEndOffset - NewBeginOffset;
2532 unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
2533 unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
2534 unsigned NumElements = EndIndex - BeginIndex;
2535 IntegerType *SubIntTy
2536 = IntTy ? Type::getIntNTy(IntTy->getContext(), Size*8) : 0;
2538 Type *OtherPtrTy = NewAI.getType();
2539 if (VecTy && !IsWholeAlloca) {
2540 if (NumElements == 1)
2541 OtherPtrTy = VecTy->getElementType();
2543 OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2545 OtherPtrTy = OtherPtrTy->getPointerTo();
2546 } else if (IntTy && !IsWholeAlloca) {
2547 OtherPtrTy = SubIntTy->getPointerTo();
2550 Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, RelOffset, OtherPtrTy);
2551 Value *DstPtr = &NewAI;
2553 std::swap(SrcPtr, DstPtr);
2556 if (VecTy && !IsWholeAlloca && !IsDest) {
2557 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2559 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
2560 } else if (IntTy && !IsWholeAlloca && !IsDest) {
2561 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2563 Src = convertValue(DL, IRB, Src, IntTy);
2564 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2565 Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
2567 Src = IRB.CreateAlignedLoad(SrcPtr, Align, II.isVolatile(),
2571 if (VecTy && !IsWholeAlloca && IsDest) {
2572 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2574 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
2575 } else if (IntTy && !IsWholeAlloca && IsDest) {
2576 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2578 Old = convertValue(DL, IRB, Old, IntTy);
2579 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2580 Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
2581 Src = convertValue(DL, IRB, Src, NewAllocaTy);
2584 StoreInst *Store = cast<StoreInst>(
2585 IRB.CreateAlignedStore(Src, DstPtr, Align, II.isVolatile()));
2587 DEBUG(dbgs() << " to: " << *Store << "\n");
2588 return !II.isVolatile();
2591 bool visitIntrinsicInst(IntrinsicInst &II) {
2592 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2593 II.getIntrinsicID() == Intrinsic::lifetime_end);
2594 DEBUG(dbgs() << " original: " << II << "\n");
2595 assert(II.getArgOperand(1) == OldPtr);
2597 // Compute the intersecting offset range.
2598 assert(BeginOffset < NewAllocaEndOffset);
2599 assert(EndOffset > NewAllocaBeginOffset);
2600 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2601 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2603 // Record this instruction for deletion.
2604 Pass.DeadInsts.insert(&II);
2607 = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2608 NewEndOffset - NewBeginOffset);
2610 getAdjustedAllocaPtr(IRB, NewBeginOffset, II.getArgOperand(1)->getType());
2612 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2613 New = IRB.CreateLifetimeStart(Ptr, Size);
2615 New = IRB.CreateLifetimeEnd(Ptr, Size);
2618 DEBUG(dbgs() << " to: " << *New << "\n");
2622 bool visitPHINode(PHINode &PN) {
2623 DEBUG(dbgs() << " original: " << PN << "\n");
2624 assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
2625 assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
2627 // We would like to compute a new pointer in only one place, but have it be
2628 // as local as possible to the PHI. To do that, we re-use the location of
2629 // the old pointer, which necessarily must be in the right position to
2630 // dominate the PHI.
2631 IRBuilderTy PtrBuilder(OldPtr);
2632 PtrBuilder.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) +
2636 getAdjustedAllocaPtr(PtrBuilder, BeginOffset, OldPtr->getType());
2637 // Replace the operands which were using the old pointer.
2638 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
2640 DEBUG(dbgs() << " to: " << PN << "\n");
2641 deleteIfTriviallyDead(OldPtr);
2643 // Check whether we can speculate this PHI node, and if so remember that
2644 // fact and queue it up for another iteration after the speculation
2646 if (isSafePHIToSpeculate(PN, &DL)) {
2647 Pass.SpeculatablePHIs.insert(&PN);
2648 IsUsedByRewrittenSpeculatableInstructions = true;
2652 return false; // PHIs can't be promoted on their own.
2655 bool visitSelectInst(SelectInst &SI) {
2656 DEBUG(dbgs() << " original: " << SI << "\n");
2657 assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
2658 "Pointer isn't an operand!");
2659 assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
2660 assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
2662 Value *NewPtr = getAdjustedAllocaPtr(IRB, BeginOffset, OldPtr->getType());
2663 // Replace the operands which were using the old pointer.
2664 if (SI.getOperand(1) == OldPtr)
2665 SI.setOperand(1, NewPtr);
2666 if (SI.getOperand(2) == OldPtr)
2667 SI.setOperand(2, NewPtr);
2669 DEBUG(dbgs() << " to: " << SI << "\n");
2670 deleteIfTriviallyDead(OldPtr);
2672 // Check whether we can speculate this select instruction, and if so
2673 // remember that fact and queue it up for another iteration after the
2674 // speculation occurs.
2675 if (isSafeSelectToSpeculate(SI, &DL)) {
2676 Pass.SpeculatableSelects.insert(&SI);
2677 IsUsedByRewrittenSpeculatableInstructions = true;
2681 return false; // Selects can't be promoted on their own.
2688 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
2690 /// This pass aggressively rewrites all aggregate loads and stores on
2691 /// a particular pointer (or any pointer derived from it which we can identify)
2692 /// with scalar loads and stores.
2693 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
2694 // Befriend the base class so it can delegate to private visit methods.
2695 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
2697 const DataLayout &DL;
2699 /// Queue of pointer uses to analyze and potentially rewrite.
2700 SmallVector<Use *, 8> Queue;
2702 /// Set to prevent us from cycling with phi nodes and loops.
2703 SmallPtrSet<User *, 8> Visited;
2705 /// The current pointer use being rewritten. This is used to dig up the used
2706 /// value (as opposed to the user).
2710 AggLoadStoreRewriter(const DataLayout &DL) : DL(DL) {}
2712 /// Rewrite loads and stores through a pointer and all pointers derived from
2714 bool rewrite(Instruction &I) {
2715 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
2717 bool Changed = false;
2718 while (!Queue.empty()) {
2719 U = Queue.pop_back_val();
2720 Changed |= visit(cast<Instruction>(U->getUser()));
2726 /// Enqueue all the users of the given instruction for further processing.
2727 /// This uses a set to de-duplicate users.
2728 void enqueueUsers(Instruction &I) {
2729 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE;
2731 if (Visited.insert(*UI))
2732 Queue.push_back(&UI.getUse());
2735 // Conservative default is to not rewrite anything.
2736 bool visitInstruction(Instruction &I) { return false; }
2738 /// \brief Generic recursive split emission class.
2739 template <typename Derived>
2742 /// The builder used to form new instructions.
2744 /// The indices which to be used with insert- or extractvalue to select the
2745 /// appropriate value within the aggregate.
2746 SmallVector<unsigned, 4> Indices;
2747 /// The indices to a GEP instruction which will move Ptr to the correct slot
2748 /// within the aggregate.
2749 SmallVector<Value *, 4> GEPIndices;
2750 /// The base pointer of the original op, used as a base for GEPing the
2751 /// split operations.
2754 /// Initialize the splitter with an insertion point, Ptr and start with a
2755 /// single zero GEP index.
2756 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
2757 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
2760 /// \brief Generic recursive split emission routine.
2762 /// This method recursively splits an aggregate op (load or store) into
2763 /// scalar or vector ops. It splits recursively until it hits a single value
2764 /// and emits that single value operation via the template argument.
2766 /// The logic of this routine relies on GEPs and insertvalue and
2767 /// extractvalue all operating with the same fundamental index list, merely
2768 /// formatted differently (GEPs need actual values).
2770 /// \param Ty The type being split recursively into smaller ops.
2771 /// \param Agg The aggregate value being built up or stored, depending on
2772 /// whether this is splitting a load or a store respectively.
2773 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
2774 if (Ty->isSingleValueType())
2775 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
2777 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2778 unsigned OldSize = Indices.size();
2780 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
2782 assert(Indices.size() == OldSize && "Did not return to the old size");
2783 Indices.push_back(Idx);
2784 GEPIndices.push_back(IRB.getInt32(Idx));
2785 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
2786 GEPIndices.pop_back();
2792 if (StructType *STy = dyn_cast<StructType>(Ty)) {
2793 unsigned OldSize = Indices.size();
2795 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
2797 assert(Indices.size() == OldSize && "Did not return to the old size");
2798 Indices.push_back(Idx);
2799 GEPIndices.push_back(IRB.getInt32(Idx));
2800 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
2801 GEPIndices.pop_back();
2807 llvm_unreachable("Only arrays and structs are aggregate loadable types");
2811 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
2812 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2813 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
2815 /// Emit a leaf load of a single value. This is called at the leaves of the
2816 /// recursive emission to actually load values.
2817 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2818 assert(Ty->isSingleValueType());
2819 // Load the single value and insert it using the indices.
2820 Value *GEP = IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep");
2821 Value *Load = IRB.CreateLoad(GEP, Name + ".load");
2822 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
2823 DEBUG(dbgs() << " to: " << *Load << "\n");
2827 bool visitLoadInst(LoadInst &LI) {
2828 assert(LI.getPointerOperand() == *U);
2829 if (!LI.isSimple() || LI.getType()->isSingleValueType())
2832 // We have an aggregate being loaded, split it apart.
2833 DEBUG(dbgs() << " original: " << LI << "\n");
2834 LoadOpSplitter Splitter(&LI, *U);
2835 Value *V = UndefValue::get(LI.getType());
2836 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
2837 LI.replaceAllUsesWith(V);
2838 LI.eraseFromParent();
2842 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
2843 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2844 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
2846 /// Emit a leaf store of a single value. This is called at the leaves of the
2847 /// recursive emission to actually produce stores.
2848 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2849 assert(Ty->isSingleValueType());
2850 // Extract the single value and store it using the indices.
2851 Value *Store = IRB.CreateStore(
2852 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
2853 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
2855 DEBUG(dbgs() << " to: " << *Store << "\n");
2859 bool visitStoreInst(StoreInst &SI) {
2860 if (!SI.isSimple() || SI.getPointerOperand() != *U)
2862 Value *V = SI.getValueOperand();
2863 if (V->getType()->isSingleValueType())
2866 // We have an aggregate being stored, split it apart.
2867 DEBUG(dbgs() << " original: " << SI << "\n");
2868 StoreOpSplitter Splitter(&SI, *U);
2869 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
2870 SI.eraseFromParent();
2874 bool visitBitCastInst(BitCastInst &BC) {
2879 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
2884 bool visitPHINode(PHINode &PN) {
2889 bool visitSelectInst(SelectInst &SI) {
2896 /// \brief Strip aggregate type wrapping.
2898 /// This removes no-op aggregate types wrapping an underlying type. It will
2899 /// strip as many layers of types as it can without changing either the type
2900 /// size or the allocated size.
2901 static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
2902 if (Ty->isSingleValueType())
2905 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
2906 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
2909 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
2910 InnerTy = ArrTy->getElementType();
2911 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
2912 const StructLayout *SL = DL.getStructLayout(STy);
2913 unsigned Index = SL->getElementContainingOffset(0);
2914 InnerTy = STy->getElementType(Index);
2919 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
2920 TypeSize > DL.getTypeSizeInBits(InnerTy))
2923 return stripAggregateTypeWrapping(DL, InnerTy);
2926 /// \brief Try to find a partition of the aggregate type passed in for a given
2927 /// offset and size.
2929 /// This recurses through the aggregate type and tries to compute a subtype
2930 /// based on the offset and size. When the offset and size span a sub-section
2931 /// of an array, it will even compute a new array type for that sub-section,
2932 /// and the same for structs.
2934 /// Note that this routine is very strict and tries to find a partition of the
2935 /// type which produces the *exact* right offset and size. It is not forgiving
2936 /// when the size or offset cause either end of type-based partition to be off.
2937 /// Also, this is a best-effort routine. It is reasonable to give up and not
2938 /// return a type if necessary.
2939 static Type *getTypePartition(const DataLayout &DL, Type *Ty,
2940 uint64_t Offset, uint64_t Size) {
2941 if (Offset == 0 && DL.getTypeAllocSize(Ty) == Size)
2942 return stripAggregateTypeWrapping(DL, Ty);
2943 if (Offset > DL.getTypeAllocSize(Ty) ||
2944 (DL.getTypeAllocSize(Ty) - Offset) < Size)
2947 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
2948 // We can't partition pointers...
2949 if (SeqTy->isPointerTy())
2952 Type *ElementTy = SeqTy->getElementType();
2953 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
2954 uint64_t NumSkippedElements = Offset / ElementSize;
2955 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy)) {
2956 if (NumSkippedElements >= ArrTy->getNumElements())
2958 } else if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy)) {
2959 if (NumSkippedElements >= VecTy->getNumElements())
2962 Offset -= NumSkippedElements * ElementSize;
2964 // First check if we need to recurse.
2965 if (Offset > 0 || Size < ElementSize) {
2966 // Bail if the partition ends in a different array element.
2967 if ((Offset + Size) > ElementSize)
2969 // Recurse through the element type trying to peel off offset bytes.
2970 return getTypePartition(DL, ElementTy, Offset, Size);
2972 assert(Offset == 0);
2974 if (Size == ElementSize)
2975 return stripAggregateTypeWrapping(DL, ElementTy);
2976 assert(Size > ElementSize);
2977 uint64_t NumElements = Size / ElementSize;
2978 if (NumElements * ElementSize != Size)
2980 return ArrayType::get(ElementTy, NumElements);
2983 StructType *STy = dyn_cast<StructType>(Ty);
2987 const StructLayout *SL = DL.getStructLayout(STy);
2988 if (Offset >= SL->getSizeInBytes())
2990 uint64_t EndOffset = Offset + Size;
2991 if (EndOffset > SL->getSizeInBytes())
2994 unsigned Index = SL->getElementContainingOffset(Offset);
2995 Offset -= SL->getElementOffset(Index);
2997 Type *ElementTy = STy->getElementType(Index);
2998 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
2999 if (Offset >= ElementSize)
3000 return 0; // The offset points into alignment padding.
3002 // See if any partition must be contained by the element.
3003 if (Offset > 0 || Size < ElementSize) {
3004 if ((Offset + Size) > ElementSize)
3006 return getTypePartition(DL, ElementTy, Offset, Size);
3008 assert(Offset == 0);
3010 if (Size == ElementSize)
3011 return stripAggregateTypeWrapping(DL, ElementTy);
3013 StructType::element_iterator EI = STy->element_begin() + Index,
3014 EE = STy->element_end();
3015 if (EndOffset < SL->getSizeInBytes()) {
3016 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3017 if (Index == EndIndex)
3018 return 0; // Within a single element and its padding.
3020 // Don't try to form "natural" types if the elements don't line up with the
3022 // FIXME: We could potentially recurse down through the last element in the
3023 // sub-struct to find a natural end point.
3024 if (SL->getElementOffset(EndIndex) != EndOffset)
3027 assert(Index < EndIndex);
3028 EE = STy->element_begin() + EndIndex;
3031 // Try to build up a sub-structure.
3032 StructType *SubTy = StructType::get(STy->getContext(), makeArrayRef(EI, EE),
3034 const StructLayout *SubSL = DL.getStructLayout(SubTy);
3035 if (Size != SubSL->getSizeInBytes())
3036 return 0; // The sub-struct doesn't have quite the size needed.
3041 /// \brief Rewrite an alloca partition's users.
3043 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3044 /// to rewrite uses of an alloca partition to be conducive for SSA value
3045 /// promotion. If the partition needs a new, more refined alloca, this will
3046 /// build that new alloca, preserving as much type information as possible, and
3047 /// rewrite the uses of the old alloca to point at the new one and have the
3048 /// appropriate new offsets. It also evaluates how successful the rewrite was
3049 /// at enabling promotion and if it was successful queues the alloca to be
3051 bool SROA::rewritePartition(AllocaInst &AI, AllocaSlices &S,
3052 AllocaSlices::iterator B, AllocaSlices::iterator E,
3053 int64_t BeginOffset, int64_t EndOffset,
3054 ArrayRef<AllocaSlices::iterator> SplitUses) {
3055 assert(BeginOffset < EndOffset);
3056 uint64_t SliceSize = EndOffset - BeginOffset;
3058 // Try to compute a friendly type for this partition of the alloca. This
3059 // won't always succeed, in which case we fall back to a legal integer type
3060 // or an i8 array of an appropriate size.
3062 if (Type *CommonUseTy = findCommonType(B, E, EndOffset))
3063 if (DL->getTypeAllocSize(CommonUseTy) >= SliceSize)
3064 SliceTy = CommonUseTy;
3066 if (Type *TypePartitionTy = getTypePartition(*DL, AI.getAllocatedType(),
3067 BeginOffset, SliceSize))
3068 SliceTy = TypePartitionTy;
3069 if ((!SliceTy || (SliceTy->isArrayTy() &&
3070 SliceTy->getArrayElementType()->isIntegerTy())) &&
3071 DL->isLegalInteger(SliceSize * 8))
3072 SliceTy = Type::getIntNTy(*C, SliceSize * 8);
3074 SliceTy = ArrayType::get(Type::getInt8Ty(*C), SliceSize);
3075 assert(DL->getTypeAllocSize(SliceTy) >= SliceSize);
3077 bool IsVectorPromotable = isVectorPromotionViable(
3078 *DL, SliceTy, S, BeginOffset, EndOffset, B, E, SplitUses);
3080 bool IsIntegerPromotable =
3081 !IsVectorPromotable &&
3082 isIntegerWideningViable(*DL, SliceTy, BeginOffset, S, B, E, SplitUses);
3084 // Check for the case where we're going to rewrite to a new alloca of the
3085 // exact same type as the original, and with the same access offsets. In that
3086 // case, re-use the existing alloca, but still run through the rewriter to
3087 // perform phi and select speculation.
3089 if (SliceTy == AI.getAllocatedType()) {
3090 assert(BeginOffset == 0 &&
3091 "Non-zero begin offset but same alloca type");
3093 // FIXME: We should be able to bail at this point with "nothing changed".
3094 // FIXME: We might want to defer PHI speculation until after here.
3096 unsigned Alignment = AI.getAlignment();
3098 // The minimum alignment which users can rely on when the explicit
3099 // alignment is omitted or zero is that required by the ABI for this
3101 Alignment = DL->getABITypeAlignment(AI.getAllocatedType());
3103 Alignment = MinAlign(Alignment, BeginOffset);
3104 // If we will get at least this much alignment from the type alone, leave
3105 // the alloca's alignment unconstrained.
3106 if (Alignment <= DL->getABITypeAlignment(SliceTy))
3108 NewAI = new AllocaInst(SliceTy, 0, Alignment,
3109 AI.getName() + ".sroa." + Twine(B - S.begin()), &AI);
3113 DEBUG(dbgs() << "Rewriting alloca partition "
3114 << "[" << BeginOffset << "," << EndOffset << ") to: " << *NewAI
3117 // Track the high watermark on several worklists that are only relevant for
3118 // promoted allocas. We will reset it to this point if the alloca is not in
3119 // fact scheduled for promotion.
3120 unsigned PPWOldSize = PostPromotionWorklist.size();
3121 unsigned SPOldSize = SpeculatablePHIs.size();
3122 unsigned SSOldSize = SpeculatableSelects.size();
3123 unsigned NumUses = 0;
3125 AllocaSliceRewriter Rewriter(*DL, S, *this, AI, *NewAI, BeginOffset,
3126 EndOffset, IsVectorPromotable,
3127 IsIntegerPromotable);
3128 bool Promotable = true;
3129 for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
3130 SUE = SplitUses.end();
3131 SUI != SUE; ++SUI) {
3132 DEBUG(dbgs() << " rewriting split ");
3133 DEBUG(S.printSlice(dbgs(), *SUI, ""));
3134 Promotable &= Rewriter.visit(*SUI);
3137 for (AllocaSlices::iterator I = B; I != E; ++I) {
3138 DEBUG(dbgs() << " rewriting ");
3139 DEBUG(S.printSlice(dbgs(), I, ""));
3140 Promotable &= Rewriter.visit(I);
3144 NumAllocaPartitionUses += NumUses;
3145 MaxUsesPerAllocaPartition =
3146 std::max<unsigned>(NumUses, MaxUsesPerAllocaPartition);
3148 if (Promotable && !Rewriter.isUsedByRewrittenSpeculatableInstructions()) {
3149 DEBUG(dbgs() << " and queuing for promotion\n");
3150 PromotableAllocas.push_back(NewAI);
3151 } else if (NewAI != &AI ||
3153 Rewriter.isUsedByRewrittenSpeculatableInstructions())) {
3154 // If we can't promote the alloca, iterate on it to check for new
3155 // refinements exposed by splitting the current alloca. Don't iterate on an
3156 // alloca which didn't actually change and didn't get promoted.
3158 // Alternatively, if we could promote the alloca but have speculatable
3159 // instructions then we will speculate them after finishing our processing
3160 // of the original alloca. Mark the new one for re-visiting in the next
3161 // iteration so the speculated operations can be rewritten.
3163 // FIXME: We should actually track whether the rewriter changed anything.
3164 Worklist.insert(NewAI);
3167 // Drop any post-promotion work items if promotion didn't happen.
3169 while (PostPromotionWorklist.size() > PPWOldSize)
3170 PostPromotionWorklist.pop_back();
3171 while (SpeculatablePHIs.size() > SPOldSize)
3172 SpeculatablePHIs.pop_back();
3173 while (SpeculatableSelects.size() > SSOldSize)
3174 SpeculatableSelects.pop_back();
3181 struct IsSliceEndLessOrEqualTo {
3182 uint64_t UpperBound;
3184 IsSliceEndLessOrEqualTo(uint64_t UpperBound) : UpperBound(UpperBound) {}
3186 bool operator()(const AllocaSlices::iterator &I) {
3187 return I->endOffset() <= UpperBound;
3193 removeFinishedSplitUses(SmallVectorImpl<AllocaSlices::iterator> &SplitUses,
3194 uint64_t &MaxSplitUseEndOffset, uint64_t Offset) {
3195 if (Offset >= MaxSplitUseEndOffset) {
3197 MaxSplitUseEndOffset = 0;
3201 size_t SplitUsesOldSize = SplitUses.size();
3202 SplitUses.erase(std::remove_if(SplitUses.begin(), SplitUses.end(),
3203 IsSliceEndLessOrEqualTo(Offset)),
3205 if (SplitUsesOldSize == SplitUses.size())
3208 // Recompute the max. While this is linear, so is remove_if.
3209 MaxSplitUseEndOffset = 0;
3210 for (SmallVectorImpl<AllocaSlices::iterator>::iterator
3211 SUI = SplitUses.begin(),
3212 SUE = SplitUses.end();
3214 MaxSplitUseEndOffset = std::max((*SUI)->endOffset(), MaxSplitUseEndOffset);
3217 /// \brief Walks the slices of an alloca and form partitions based on them,
3218 /// rewriting each of their uses.
3219 bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &S) {
3220 if (S.begin() == S.end())
3223 unsigned NumPartitions = 0;
3224 bool Changed = false;
3225 SmallVector<AllocaSlices::iterator, 4> SplitUses;
3226 uint64_t MaxSplitUseEndOffset = 0;
3228 uint64_t BeginOffset = S.begin()->beginOffset();
3230 for (AllocaSlices::iterator SI = S.begin(), SJ = llvm::next(SI), SE = S.end();
3231 SI != SE; SI = SJ) {
3232 uint64_t MaxEndOffset = SI->endOffset();
3234 if (!SI->isSplittable()) {
3235 // When we're forming an unsplittable region, it must always start at the
3236 // first slice and will extend through its end.
3237 assert(BeginOffset == SI->beginOffset());
3239 // Form a partition including all of the overlapping slices with this
3240 // unsplittable slice.
3241 while (SJ != SE && SJ->beginOffset() < MaxEndOffset) {
3242 if (!SJ->isSplittable())
3243 MaxEndOffset = std::max(MaxEndOffset, SJ->endOffset());
3247 assert(SI->isSplittable()); // Established above.
3249 // Collect all of the overlapping splittable slices.
3250 while (SJ != SE && SJ->beginOffset() < MaxEndOffset &&
3251 SJ->isSplittable()) {
3252 MaxEndOffset = std::max(MaxEndOffset, SJ->endOffset());
3256 // Back up MaxEndOffset and SJ if we ended the span early when
3257 // encountering an unsplittable slice.
3258 if (SJ != SE && SJ->beginOffset() < MaxEndOffset) {
3259 assert(!SJ->isSplittable());
3260 MaxEndOffset = SJ->beginOffset();
3264 // Check if we have managed to move the end offset forward yet. If so,
3265 // we'll have to rewrite uses and erase old split uses.
3266 if (BeginOffset < MaxEndOffset) {
3267 // Rewrite a sequence of overlapping slices.
3269 rewritePartition(AI, S, SI, SJ, BeginOffset, MaxEndOffset, SplitUses);
3272 removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset, MaxEndOffset);
3275 // Accumulate all the splittable slices from the [SI,SJ) region which
3276 // overlap going forward.
3277 for (AllocaSlices::iterator SK = SI; SK != SJ; ++SK)
3278 if (SK->isSplittable() && SK->endOffset() > MaxEndOffset) {
3279 SplitUses.push_back(SK);
3280 MaxSplitUseEndOffset = std::max(SK->endOffset(), MaxSplitUseEndOffset);
3283 // If we're already at the end and we have no split uses, we're done.
3284 if (SJ == SE && SplitUses.empty())
3287 // If we have no split uses or no gap in offsets, we're ready to move to
3289 if (SplitUses.empty() || (SJ != SE && MaxEndOffset == SJ->beginOffset())) {
3290 BeginOffset = SJ->beginOffset();
3294 // Even if we have split slices, if the next slice is splittable and the
3295 // split slices reach it, we can simply set up the beginning offset of the
3296 // next iteration to bridge between them.
3297 if (SJ != SE && SJ->isSplittable() &&
3298 MaxSplitUseEndOffset > SJ->beginOffset()) {
3299 BeginOffset = MaxEndOffset;
3303 // Otherwise, we have a tail of split slices. Rewrite them with an empty
3305 uint64_t PostSplitEndOffset =
3306 SJ == SE ? MaxSplitUseEndOffset : SJ->beginOffset();
3308 Changed |= rewritePartition(AI, S, SJ, SJ, MaxEndOffset, PostSplitEndOffset,
3313 break; // Skip the rest, we don't need to do any cleanup.
3315 removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset,
3316 PostSplitEndOffset);
3318 // Now just reset the begin offset for the next iteration.
3319 BeginOffset = SJ->beginOffset();
3322 NumAllocaPartitions += NumPartitions;
3323 MaxPartitionsPerAlloca =
3324 std::max<unsigned>(NumPartitions, MaxPartitionsPerAlloca);
3329 /// \brief Analyze an alloca for SROA.
3331 /// This analyzes the alloca to ensure we can reason about it, builds
3332 /// the slices of the alloca, and then hands it off to be split and
3333 /// rewritten as needed.
3334 bool SROA::runOnAlloca(AllocaInst &AI) {
3335 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3336 ++NumAllocasAnalyzed;
3338 // Special case dead allocas, as they're trivial.
3339 if (AI.use_empty()) {
3340 AI.eraseFromParent();
3344 // Skip alloca forms that this analysis can't handle.
3345 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3346 DL->getTypeAllocSize(AI.getAllocatedType()) == 0)
3349 bool Changed = false;
3351 // First, split any FCA loads and stores touching this alloca to promote
3352 // better splitting and promotion opportunities.
3353 AggLoadStoreRewriter AggRewriter(*DL);
3354 Changed |= AggRewriter.rewrite(AI);
3356 // Build the slices using a recursive instruction-visiting builder.
3357 AllocaSlices S(*DL, AI);
3358 DEBUG(S.print(dbgs()));
3362 // Delete all the dead users of this alloca before splitting and rewriting it.
3363 for (AllocaSlices::dead_user_iterator DI = S.dead_user_begin(),
3364 DE = S.dead_user_end();
3367 (*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType()));
3368 DeadInsts.insert(*DI);
3370 for (AllocaSlices::dead_op_iterator DO = S.dead_op_begin(),
3371 DE = S.dead_op_end();
3374 // Clobber the use with an undef value.
3375 **DO = UndefValue::get(OldV->getType());
3376 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3377 if (isInstructionTriviallyDead(OldI)) {
3379 DeadInsts.insert(OldI);
3383 // No slices to split. Leave the dead alloca for a later pass to clean up.
3384 if (S.begin() == S.end())
3387 Changed |= splitAlloca(AI, S);
3389 DEBUG(dbgs() << " Speculating PHIs\n");
3390 while (!SpeculatablePHIs.empty())
3391 speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
3393 DEBUG(dbgs() << " Speculating Selects\n");
3394 while (!SpeculatableSelects.empty())
3395 speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
3400 /// \brief Delete the dead instructions accumulated in this run.
3402 /// Recursively deletes the dead instructions we've accumulated. This is done
3403 /// at the very end to maximize locality of the recursive delete and to
3404 /// minimize the problems of invalidated instruction pointers as such pointers
3405 /// are used heavily in the intermediate stages of the algorithm.
3407 /// We also record the alloca instructions deleted here so that they aren't
3408 /// subsequently handed to mem2reg to promote.
3409 void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
3410 while (!DeadInsts.empty()) {
3411 Instruction *I = DeadInsts.pop_back_val();
3412 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3414 I->replaceAllUsesWith(UndefValue::get(I->getType()));
3416 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
3417 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
3418 // Zero out the operand and see if it becomes trivially dead.
3420 if (isInstructionTriviallyDead(U))
3421 DeadInsts.insert(U);
3424 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3425 DeletedAllocas.insert(AI);
3428 I->eraseFromParent();
3432 static void enqueueUsersInWorklist(Instruction &I,
3433 SmallVectorImpl<Instruction *> &Worklist,
3434 SmallPtrSet<Instruction *, 8> &Visited) {
3435 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE;
3437 if (Visited.insert(cast<Instruction>(*UI)))
3438 Worklist.push_back(cast<Instruction>(*UI));
3441 /// \brief Promote the allocas, using the best available technique.
3443 /// This attempts to promote whatever allocas have been identified as viable in
3444 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
3445 /// If there is a domtree available, we attempt to promote using the full power
3446 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3447 /// based on the SSAUpdater utilities. This function returns whether any
3448 /// promotion occurred.
3449 bool SROA::promoteAllocas(Function &F) {
3450 if (PromotableAllocas.empty())
3453 NumPromoted += PromotableAllocas.size();
3455 if (DT && !ForceSSAUpdater) {
3456 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3457 PromoteMemToReg(PromotableAllocas, *DT);
3458 PromotableAllocas.clear();
3462 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3464 DIBuilder DIB(*F.getParent());
3465 SmallVector<Instruction *, 64> Insts;
3467 // We need a worklist to walk the uses of each alloca.
3468 SmallVector<Instruction *, 8> Worklist;
3469 SmallPtrSet<Instruction *, 8> Visited;
3470 SmallVector<Instruction *, 32> DeadInsts;
3472 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3473 AllocaInst *AI = PromotableAllocas[Idx];
3478 enqueueUsersInWorklist(*AI, Worklist, Visited);
3480 while (!Worklist.empty()) {
3481 Instruction *I = Worklist.pop_back_val();
3483 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3484 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3485 // leading to them) here. Eventually it should use them to optimize the
3486 // scalar values produced.
3487 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3488 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3489 II->getIntrinsicID() == Intrinsic::lifetime_end);
3490 II->eraseFromParent();
3494 // Push the loads and stores we find onto the list. SROA will already
3495 // have validated that all loads and stores are viable candidates for
3497 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
3498 assert(LI->getType() == AI->getAllocatedType());
3499 Insts.push_back(LI);
3502 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
3503 assert(SI->getValueOperand()->getType() == AI->getAllocatedType());
3504 Insts.push_back(SI);
3508 // For everything else, we know that only no-op bitcasts and GEPs will
3509 // make it this far, just recurse through them and recall them for later
3511 DeadInsts.push_back(I);
3512 enqueueUsersInWorklist(*I, Worklist, Visited);
3514 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3515 while (!DeadInsts.empty())
3516 DeadInsts.pop_back_val()->eraseFromParent();
3517 AI->eraseFromParent();
3520 PromotableAllocas.clear();
3525 /// \brief A predicate to test whether an alloca belongs to a set.
3526 class IsAllocaInSet {
3527 typedef SmallPtrSet<AllocaInst *, 4> SetType;
3531 typedef AllocaInst *argument_type;
3533 IsAllocaInSet(const SetType &Set) : Set(Set) {}
3534 bool operator()(AllocaInst *AI) const { return Set.count(AI); }
3538 bool SROA::runOnFunction(Function &F) {
3539 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3540 C = &F.getContext();
3541 DL = getAnalysisIfAvailable<DataLayout>();
3543 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3546 DT = getAnalysisIfAvailable<DominatorTree>();
3548 BasicBlock &EntryBB = F.getEntryBlock();
3549 for (BasicBlock::iterator I = EntryBB.begin(), E = llvm::prior(EntryBB.end());
3551 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3552 Worklist.insert(AI);
3554 bool Changed = false;
3555 // A set of deleted alloca instruction pointers which should be removed from
3556 // the list of promotable allocas.
3557 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3560 while (!Worklist.empty()) {
3561 Changed |= runOnAlloca(*Worklist.pop_back_val());
3562 deleteDeadInstructions(DeletedAllocas);
3564 // Remove the deleted allocas from various lists so that we don't try to
3565 // continue processing them.
3566 if (!DeletedAllocas.empty()) {
3567 Worklist.remove_if(IsAllocaInSet(DeletedAllocas));
3568 PostPromotionWorklist.remove_if(IsAllocaInSet(DeletedAllocas));
3569 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3570 PromotableAllocas.end(),
3571 IsAllocaInSet(DeletedAllocas)),
3572 PromotableAllocas.end());
3573 DeletedAllocas.clear();
3577 Changed |= promoteAllocas(F);
3579 Worklist = PostPromotionWorklist;
3580 PostPromotionWorklist.clear();
3581 } while (!Worklist.empty());
3586 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3587 if (RequiresDomTree)
3588 AU.addRequired<DominatorTree>();
3589 AU.setPreservesCFG();