1 //===- SROA.cpp - Scalar Replacement Of Aggregates ------------------------===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 /// This transformation implements the well known scalar replacement of
10 /// aggregates transformation. It tries to identify promotable elements of an
11 /// aggregate alloca, and promote them to registers. It will also try to
12 /// convert uses of an element (or set of elements) of an alloca into a vector
13 /// or bitfield-style integer scalar if appropriate.
15 /// It works to do this with minimal slicing of the alloca so that regions
16 /// which are merely transferred in and out of external memory remain unchanged
17 /// and are not decomposed to scalar code.
19 /// Because this also performs alloca promotion, it can be thought of as also
20 /// serving the purpose of SSA formation. The algorithm iterates on the
21 /// function until all opportunities for promotion have been realized.
23 //===----------------------------------------------------------------------===//
25 #include "llvm/Transforms/Scalar/SROA.h"
26 #include "llvm/ADT/APInt.h"
27 #include "llvm/ADT/ArrayRef.h"
28 #include "llvm/ADT/DenseMap.h"
29 #include "llvm/ADT/PointerIntPair.h"
30 #include "llvm/ADT/STLExtras.h"
31 #include "llvm/ADT/SetVector.h"
32 #include "llvm/ADT/SmallBitVector.h"
33 #include "llvm/ADT/SmallPtrSet.h"
34 #include "llvm/ADT/SmallVector.h"
35 #include "llvm/ADT/Statistic.h"
36 #include "llvm/ADT/StringRef.h"
37 #include "llvm/ADT/Twine.h"
38 #include "llvm/ADT/iterator.h"
39 #include "llvm/ADT/iterator_range.h"
40 #include "llvm/Analysis/AssumptionCache.h"
41 #include "llvm/Analysis/GlobalsModRef.h"
42 #include "llvm/Analysis/Loads.h"
43 #include "llvm/Analysis/PtrUseVisitor.h"
44 #include "llvm/Config/llvm-config.h"
45 #include "llvm/IR/BasicBlock.h"
46 #include "llvm/IR/Constant.h"
47 #include "llvm/IR/ConstantFolder.h"
48 #include "llvm/IR/Constants.h"
49 #include "llvm/IR/DIBuilder.h"
50 #include "llvm/IR/DataLayout.h"
51 #include "llvm/IR/DebugInfoMetadata.h"
52 #include "llvm/IR/DerivedTypes.h"
53 #include "llvm/IR/Dominators.h"
54 #include "llvm/IR/Function.h"
55 #include "llvm/IR/GetElementPtrTypeIterator.h"
56 #include "llvm/IR/GlobalAlias.h"
57 #include "llvm/IR/IRBuilder.h"
58 #include "llvm/IR/InstVisitor.h"
59 #include "llvm/IR/InstrTypes.h"
60 #include "llvm/IR/Instruction.h"
61 #include "llvm/IR/Instructions.h"
62 #include "llvm/IR/IntrinsicInst.h"
63 #include "llvm/IR/Intrinsics.h"
64 #include "llvm/IR/LLVMContext.h"
65 #include "llvm/IR/Metadata.h"
66 #include "llvm/IR/Module.h"
67 #include "llvm/IR/Operator.h"
68 #include "llvm/IR/PassManager.h"
69 #include "llvm/IR/Type.h"
70 #include "llvm/IR/Use.h"
71 #include "llvm/IR/User.h"
72 #include "llvm/IR/Value.h"
73 #include "llvm/InitializePasses.h"
74 #include "llvm/Pass.h"
75 #include "llvm/Support/Casting.h"
76 #include "llvm/Support/CommandLine.h"
77 #include "llvm/Support/Compiler.h"
78 #include "llvm/Support/Debug.h"
79 #include "llvm/Support/ErrorHandling.h"
80 #include "llvm/Support/MathExtras.h"
81 #include "llvm/Support/raw_ostream.h"
82 #include "llvm/Transforms/Scalar.h"
83 #include "llvm/Transforms/Utils/Local.h"
84 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
98 using namespace llvm::sroa;
100 #define DEBUG_TYPE "sroa"
102 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
103 STATISTIC(NumAllocaPartitions, "Number of alloca partitions formed");
104 STATISTIC(MaxPartitionsPerAlloca, "Maximum number of partitions per alloca");
105 STATISTIC(NumAllocaPartitionUses, "Number of alloca partition uses rewritten");
106 STATISTIC(MaxUsesPerAllocaPartition, "Maximum number of uses of a partition");
107 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
108 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
109 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
110 STATISTIC(NumDeleted, "Number of instructions deleted");
111 STATISTIC(NumVectorized, "Number of vectorized aggregates");
113 /// Hidden option to experiment with completely strict handling of inbounds
115 static cl::opt<bool> SROAStrictInbounds("sroa-strict-inbounds", cl::init(false),
120 /// A custom IRBuilder inserter which prefixes all names, but only in
122 class IRBuilderPrefixedInserter final : public IRBuilderDefaultInserter {
125 const Twine getNameWithPrefix(const Twine &Name) const {
126 return Name.isTriviallyEmpty() ? Name : Prefix + Name;
130 void SetNamePrefix(const Twine &P) { Prefix = P.str(); }
132 void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
133 BasicBlock::iterator InsertPt) const override {
134 IRBuilderDefaultInserter::InsertHelper(I, getNameWithPrefix(Name), BB,
139 /// Provide a type for IRBuilder that drops names in release builds.
140 using IRBuilderTy = IRBuilder<ConstantFolder, IRBuilderPrefixedInserter>;
142 /// A used slice of an alloca.
144 /// This structure represents a slice of an alloca used by some instruction. It
145 /// stores both the begin and end offsets of this use, a pointer to the use
146 /// itself, and a flag indicating whether we can classify the use as splittable
147 /// or not when forming partitions of the alloca.
149 /// The beginning offset of the range.
150 uint64_t BeginOffset = 0;
152 /// The ending offset, not included in the range.
153 uint64_t EndOffset = 0;
155 /// Storage for both the use of this slice and whether it can be
157 PointerIntPair<Use *, 1, bool> UseAndIsSplittable;
162 Slice(uint64_t BeginOffset, uint64_t EndOffset, Use *U, bool IsSplittable)
163 : BeginOffset(BeginOffset), EndOffset(EndOffset),
164 UseAndIsSplittable(U, IsSplittable) {}
166 uint64_t beginOffset() const { return BeginOffset; }
167 uint64_t endOffset() const { return EndOffset; }
169 bool isSplittable() const { return UseAndIsSplittable.getInt(); }
170 void makeUnsplittable() { UseAndIsSplittable.setInt(false); }
172 Use *getUse() const { return UseAndIsSplittable.getPointer(); }
174 bool isDead() const { return getUse() == nullptr; }
175 void kill() { UseAndIsSplittable.setPointer(nullptr); }
177 /// Support for ordering ranges.
179 /// This provides an ordering over ranges such that start offsets are
180 /// always increasing, and within equal start offsets, the end offsets are
181 /// decreasing. Thus the spanning range comes first in a cluster with the
182 /// same start position.
183 bool operator<(const Slice &RHS) const {
184 if (beginOffset() < RHS.beginOffset())
186 if (beginOffset() > RHS.beginOffset())
188 if (isSplittable() != RHS.isSplittable())
189 return !isSplittable();
190 if (endOffset() > RHS.endOffset())
195 /// Support comparison with a single offset to allow binary searches.
196 friend LLVM_ATTRIBUTE_UNUSED bool operator<(const Slice &LHS,
197 uint64_t RHSOffset) {
198 return LHS.beginOffset() < RHSOffset;
200 friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
202 return LHSOffset < RHS.beginOffset();
205 bool operator==(const Slice &RHS) const {
206 return isSplittable() == RHS.isSplittable() &&
207 beginOffset() == RHS.beginOffset() && endOffset() == RHS.endOffset();
209 bool operator!=(const Slice &RHS) const { return !operator==(RHS); }
212 } // end anonymous namespace
214 /// Representation of the alloca slices.
216 /// This class represents the slices of an alloca which are formed by its
217 /// various uses. If a pointer escapes, we can't fully build a representation
218 /// for the slices used and we reflect that in this structure. The uses are
219 /// stored, sorted by increasing beginning offset and with unsplittable slices
220 /// starting at a particular offset before splittable slices.
221 class llvm::sroa::AllocaSlices {
223 /// Construct the slices of a particular alloca.
224 AllocaSlices(const DataLayout &DL, AllocaInst &AI);
226 /// Test whether a pointer to the allocation escapes our analysis.
228 /// If this is true, the slices are never fully built and should be
230 bool isEscaped() const { return PointerEscapingInstr; }
232 /// Support for iterating over the slices.
234 using iterator = SmallVectorImpl<Slice>::iterator;
235 using range = iterator_range<iterator>;
237 iterator begin() { return Slices.begin(); }
238 iterator end() { return Slices.end(); }
240 using const_iterator = SmallVectorImpl<Slice>::const_iterator;
241 using const_range = iterator_range<const_iterator>;
243 const_iterator begin() const { return Slices.begin(); }
244 const_iterator end() const { return Slices.end(); }
247 /// Erase a range of slices.
248 void erase(iterator Start, iterator Stop) { Slices.erase(Start, Stop); }
250 /// Insert new slices for this alloca.
252 /// This moves the slices into the alloca's slices collection, and re-sorts
253 /// everything so that the usual ordering properties of the alloca's slices
255 void insert(ArrayRef<Slice> NewSlices) {
256 int OldSize = Slices.size();
257 Slices.append(NewSlices.begin(), NewSlices.end());
258 auto SliceI = Slices.begin() + OldSize;
259 llvm::sort(SliceI, Slices.end());
260 std::inplace_merge(Slices.begin(), SliceI, Slices.end());
263 // Forward declare the iterator and range accessor for walking the
265 class partition_iterator;
266 iterator_range<partition_iterator> partitions();
268 /// Access the dead users for this alloca.
269 ArrayRef<Instruction *> getDeadUsers() const { return DeadUsers; }
271 /// Access the dead operands referring to this alloca.
273 /// These are operands which have cannot actually be used to refer to the
274 /// alloca as they are outside its range and the user doesn't correct for
275 /// that. These mostly consist of PHI node inputs and the like which we just
276 /// need to replace with undef.
277 ArrayRef<Use *> getDeadOperands() const { return DeadOperands; }
279 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
280 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
281 void printSlice(raw_ostream &OS, const_iterator I,
282 StringRef Indent = " ") const;
283 void printUse(raw_ostream &OS, const_iterator I,
284 StringRef Indent = " ") const;
285 void print(raw_ostream &OS) const;
286 void dump(const_iterator I) const;
291 template <typename DerivedT, typename RetT = void> class BuilderBase;
294 friend class AllocaSlices::SliceBuilder;
296 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
297 /// Handle to alloca instruction to simplify method interfaces.
301 /// The instruction responsible for this alloca not having a known set
304 /// When an instruction (potentially) escapes the pointer to the alloca, we
305 /// store a pointer to that here and abort trying to form slices of the
306 /// alloca. This will be null if the alloca slices are analyzed successfully.
307 Instruction *PointerEscapingInstr;
309 /// The slices of the alloca.
311 /// We store a vector of the slices formed by uses of the alloca here. This
312 /// vector is sorted by increasing begin offset, and then the unsplittable
313 /// slices before the splittable ones. See the Slice inner class for more
315 SmallVector<Slice, 8> Slices;
317 /// Instructions which will become dead if we rewrite the alloca.
319 /// Note that these are not separated by slice. This is because we expect an
320 /// alloca to be completely rewritten or not rewritten at all. If rewritten,
321 /// all these instructions can simply be removed and replaced with undef as
322 /// they come from outside of the allocated space.
323 SmallVector<Instruction *, 8> DeadUsers;
325 /// Operands which will become dead if we rewrite the alloca.
327 /// These are operands that in their particular use can be replaced with
328 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
329 /// to PHI nodes and the like. They aren't entirely dead (there might be
330 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
331 /// want to swap this particular input for undef to simplify the use lists of
333 SmallVector<Use *, 8> DeadOperands;
336 /// A partition of the slices.
338 /// An ephemeral representation for a range of slices which can be viewed as
339 /// a partition of the alloca. This range represents a span of the alloca's
340 /// memory which cannot be split, and provides access to all of the slices
341 /// overlapping some part of the partition.
343 /// Objects of this type are produced by traversing the alloca's slices, but
344 /// are only ephemeral and not persistent.
345 class llvm::sroa::Partition {
347 friend class AllocaSlices;
348 friend class AllocaSlices::partition_iterator;
350 using iterator = AllocaSlices::iterator;
352 /// The beginning and ending offsets of the alloca for this
354 uint64_t BeginOffset = 0, EndOffset = 0;
356 /// The start and end iterators of this partition.
359 /// A collection of split slice tails overlapping the partition.
360 SmallVector<Slice *, 4> SplitTails;
362 /// Raw constructor builds an empty partition starting and ending at
363 /// the given iterator.
364 Partition(iterator SI) : SI(SI), SJ(SI) {}
367 /// The start offset of this partition.
369 /// All of the contained slices start at or after this offset.
370 uint64_t beginOffset() const { return BeginOffset; }
372 /// The end offset of this partition.
374 /// All of the contained slices end at or before this offset.
375 uint64_t endOffset() const { return EndOffset; }
377 /// The size of the partition.
379 /// Note that this can never be zero.
380 uint64_t size() const {
381 assert(BeginOffset < EndOffset && "Partitions must span some bytes!");
382 return EndOffset - BeginOffset;
385 /// Test whether this partition contains no slices, and merely spans
386 /// a region occupied by split slices.
387 bool empty() const { return SI == SJ; }
389 /// \name Iterate slices that start within the partition.
390 /// These may be splittable or unsplittable. They have a begin offset >= the
391 /// partition begin offset.
393 // FIXME: We should probably define a "concat_iterator" helper and use that
394 // to stitch together pointee_iterators over the split tails and the
395 // contiguous iterators of the partition. That would give a much nicer
396 // interface here. We could then additionally expose filtered iterators for
397 // split, unsplit, and unsplittable splices based on the usage patterns.
398 iterator begin() const { return SI; }
399 iterator end() const { return SJ; }
402 /// Get the sequence of split slice tails.
404 /// These tails are of slices which start before this partition but are
405 /// split and overlap into the partition. We accumulate these while forming
407 ArrayRef<Slice *> splitSliceTails() const { return SplitTails; }
410 /// An iterator over partitions of the alloca's slices.
412 /// This iterator implements the core algorithm for partitioning the alloca's
413 /// slices. It is a forward iterator as we don't support backtracking for
414 /// efficiency reasons, and re-use a single storage area to maintain the
415 /// current set of split slices.
417 /// It is templated on the slice iterator type to use so that it can operate
418 /// with either const or non-const slice iterators.
419 class AllocaSlices::partition_iterator
420 : public iterator_facade_base<partition_iterator, std::forward_iterator_tag,
422 friend class AllocaSlices;
424 /// Most of the state for walking the partitions is held in a class
425 /// with a nice interface for examining them.
428 /// We need to keep the end of the slices to know when to stop.
429 AllocaSlices::iterator SE;
431 /// We also need to keep track of the maximum split end offset seen.
432 /// FIXME: Do we really?
433 uint64_t MaxSplitSliceEndOffset = 0;
435 /// Sets the partition to be empty at given iterator, and sets the
437 partition_iterator(AllocaSlices::iterator SI, AllocaSlices::iterator SE)
439 // If not already at the end, advance our state to form the initial
445 /// Advance the iterator to the next partition.
447 /// Requires that the iterator not be at the end of the slices.
449 assert((P.SI != SE || !P.SplitTails.empty()) &&
450 "Cannot advance past the end of the slices!");
452 // Clear out any split uses which have ended.
453 if (!P.SplitTails.empty()) {
454 if (P.EndOffset >= MaxSplitSliceEndOffset) {
455 // If we've finished all splits, this is easy.
456 P.SplitTails.clear();
457 MaxSplitSliceEndOffset = 0;
459 // Remove the uses which have ended in the prior partition. This
460 // cannot change the max split slice end because we just checked that
461 // the prior partition ended prior to that max.
462 P.SplitTails.erase(llvm::remove_if(P.SplitTails,
464 return S->endOffset() <=
468 assert(llvm::any_of(P.SplitTails,
470 return S->endOffset() == MaxSplitSliceEndOffset;
472 "Could not find the current max split slice offset!");
473 assert(llvm::all_of(P.SplitTails,
475 return S->endOffset() <= MaxSplitSliceEndOffset;
477 "Max split slice end offset is not actually the max!");
481 // If P.SI is already at the end, then we've cleared the split tail and
482 // now have an end iterator.
484 assert(P.SplitTails.empty() && "Failed to clear the split slices!");
488 // If we had a non-empty partition previously, set up the state for
489 // subsequent partitions.
491 // Accumulate all the splittable slices which started in the old
492 // partition into the split list.
494 if (S.isSplittable() && S.endOffset() > P.EndOffset) {
495 P.SplitTails.push_back(&S);
496 MaxSplitSliceEndOffset =
497 std::max(S.endOffset(), MaxSplitSliceEndOffset);
500 // Start from the end of the previous partition.
503 // If P.SI is now at the end, we at most have a tail of split slices.
505 P.BeginOffset = P.EndOffset;
506 P.EndOffset = MaxSplitSliceEndOffset;
510 // If the we have split slices and the next slice is after a gap and is
511 // not splittable immediately form an empty partition for the split
512 // slices up until the next slice begins.
513 if (!P.SplitTails.empty() && P.SI->beginOffset() != P.EndOffset &&
514 !P.SI->isSplittable()) {
515 P.BeginOffset = P.EndOffset;
516 P.EndOffset = P.SI->beginOffset();
521 // OK, we need to consume new slices. Set the end offset based on the
522 // current slice, and step SJ past it. The beginning offset of the
523 // partition is the beginning offset of the next slice unless we have
524 // pre-existing split slices that are continuing, in which case we begin
525 // at the prior end offset.
526 P.BeginOffset = P.SplitTails.empty() ? P.SI->beginOffset() : P.EndOffset;
527 P.EndOffset = P.SI->endOffset();
530 // There are two strategies to form a partition based on whether the
531 // partition starts with an unsplittable slice or a splittable slice.
532 if (!P.SI->isSplittable()) {
533 // When we're forming an unsplittable region, it must always start at
534 // the first slice and will extend through its end.
535 assert(P.BeginOffset == P.SI->beginOffset());
537 // Form a partition including all of the overlapping slices with this
538 // unsplittable slice.
539 while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
540 if (!P.SJ->isSplittable())
541 P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
545 // We have a partition across a set of overlapping unsplittable
550 // If we're starting with a splittable slice, then we need to form
551 // a synthetic partition spanning it and any other overlapping splittable
553 assert(P.SI->isSplittable() && "Forming a splittable partition!");
555 // Collect all of the overlapping splittable slices.
556 while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset &&
557 P.SJ->isSplittable()) {
558 P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
562 // Back upiP.EndOffset if we ended the span early when encountering an
563 // unsplittable slice. This synthesizes the early end offset of
564 // a partition spanning only splittable slices.
565 if (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
566 assert(!P.SJ->isSplittable());
567 P.EndOffset = P.SJ->beginOffset();
572 bool operator==(const partition_iterator &RHS) const {
573 assert(SE == RHS.SE &&
574 "End iterators don't match between compared partition iterators!");
576 // The observed positions of partitions is marked by the P.SI iterator and
577 // the emptiness of the split slices. The latter is only relevant when
578 // P.SI == SE, as the end iterator will additionally have an empty split
579 // slices list, but the prior may have the same P.SI and a tail of split
581 if (P.SI == RHS.P.SI && P.SplitTails.empty() == RHS.P.SplitTails.empty()) {
582 assert(P.SJ == RHS.P.SJ &&
583 "Same set of slices formed two different sized partitions!");
584 assert(P.SplitTails.size() == RHS.P.SplitTails.size() &&
585 "Same slice position with differently sized non-empty split "
592 partition_iterator &operator++() {
597 Partition &operator*() { return P; }
600 /// A forward range over the partitions of the alloca's slices.
602 /// This accesses an iterator range over the partitions of the alloca's
603 /// slices. It computes these partitions on the fly based on the overlapping
604 /// offsets of the slices and the ability to split them. It will visit "empty"
605 /// partitions to cover regions of the alloca only accessed via split
607 iterator_range<AllocaSlices::partition_iterator> AllocaSlices::partitions() {
608 return make_range(partition_iterator(begin(), end()),
609 partition_iterator(end(), end()));
612 static Value *foldSelectInst(SelectInst &SI) {
613 // If the condition being selected on is a constant or the same value is
614 // being selected between, fold the select. Yes this does (rarely) happen
616 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
617 return SI.getOperand(1 + CI->isZero());
618 if (SI.getOperand(1) == SI.getOperand(2))
619 return SI.getOperand(1);
624 /// A helper that folds a PHI node or a select.
625 static Value *foldPHINodeOrSelectInst(Instruction &I) {
626 if (PHINode *PN = dyn_cast<PHINode>(&I)) {
627 // If PN merges together the same value, return that value.
628 return PN->hasConstantValue();
630 return foldSelectInst(cast<SelectInst>(I));
633 /// Builder for the alloca slices.
635 /// This class builds a set of alloca slices by recursively visiting the uses
636 /// of an alloca and making a slice for each load and store at each offset.
637 class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
638 friend class PtrUseVisitor<SliceBuilder>;
639 friend class InstVisitor<SliceBuilder>;
641 using Base = PtrUseVisitor<SliceBuilder>;
643 const uint64_t AllocSize;
646 SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
647 SmallDenseMap<Instruction *, uint64_t> PHIOrSelectSizes;
649 /// Set to de-duplicate dead instructions found in the use walk.
650 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
653 SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &AS)
654 : PtrUseVisitor<SliceBuilder>(DL),
655 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType()).getFixedSize()),
659 void markAsDead(Instruction &I) {
660 if (VisitedDeadInsts.insert(&I).second)
661 AS.DeadUsers.push_back(&I);
664 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
665 bool IsSplittable = false) {
666 // Completely skip uses which have a zero size or start either before or
667 // past the end of the allocation.
668 if (Size == 0 || Offset.uge(AllocSize)) {
669 LLVM_DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @"
671 << " which has zero size or starts outside of the "
672 << AllocSize << " byte alloca:\n"
673 << " alloca: " << AS.AI << "\n"
674 << " use: " << I << "\n");
675 return markAsDead(I);
678 uint64_t BeginOffset = Offset.getZExtValue();
679 uint64_t EndOffset = BeginOffset + Size;
681 // Clamp the end offset to the end of the allocation. Note that this is
682 // formulated to handle even the case where "BeginOffset + Size" overflows.
683 // This may appear superficially to be something we could ignore entirely,
684 // but that is not so! There may be widened loads or PHI-node uses where
685 // some instructions are dead but not others. We can't completely ignore
686 // them, and so have to record at least the information here.
687 assert(AllocSize >= BeginOffset); // Established above.
688 if (Size > AllocSize - BeginOffset) {
689 LLVM_DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @"
690 << Offset << " to remain within the " << AllocSize
692 << " alloca: " << AS.AI << "\n"
693 << " use: " << I << "\n");
694 EndOffset = AllocSize;
697 AS.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
700 void visitBitCastInst(BitCastInst &BC) {
702 return markAsDead(BC);
704 return Base::visitBitCastInst(BC);
707 void visitAddrSpaceCastInst(AddrSpaceCastInst &ASC) {
709 return markAsDead(ASC);
711 return Base::visitAddrSpaceCastInst(ASC);
714 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
715 if (GEPI.use_empty())
716 return markAsDead(GEPI);
718 if (SROAStrictInbounds && GEPI.isInBounds()) {
719 // FIXME: This is a manually un-factored variant of the basic code inside
720 // of GEPs with checking of the inbounds invariant specified in the
721 // langref in a very strict sense. If we ever want to enable
722 // SROAStrictInbounds, this code should be factored cleanly into
723 // PtrUseVisitor, but it is easier to experiment with SROAStrictInbounds
724 // by writing out the code here where we have the underlying allocation
725 // size readily available.
726 APInt GEPOffset = Offset;
727 const DataLayout &DL = GEPI.getModule()->getDataLayout();
728 for (gep_type_iterator GTI = gep_type_begin(GEPI),
729 GTE = gep_type_end(GEPI);
731 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
735 // Handle a struct index, which adds its field offset to the pointer.
736 if (StructType *STy = GTI.getStructTypeOrNull()) {
737 unsigned ElementIdx = OpC->getZExtValue();
738 const StructLayout *SL = DL.getStructLayout(STy);
740 APInt(Offset.getBitWidth(), SL->getElementOffset(ElementIdx));
742 // For array or vector indices, scale the index by the size of the
744 APInt Index = OpC->getValue().sextOrTrunc(Offset.getBitWidth());
747 APInt(Offset.getBitWidth(),
748 DL.getTypeAllocSize(GTI.getIndexedType()).getFixedSize());
751 // If this index has computed an intermediate pointer which is not
752 // inbounds, then the result of the GEP is a poison value and we can
753 // delete it and all uses.
754 if (GEPOffset.ugt(AllocSize))
755 return markAsDead(GEPI);
759 return Base::visitGetElementPtrInst(GEPI);
762 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
763 uint64_t Size, bool IsVolatile) {
764 // We allow splitting of non-volatile loads and stores where the type is an
765 // integer type. These may be used to implement 'memcpy' or other "transfer
766 // of bits" patterns.
767 bool IsSplittable = Ty->isIntegerTy() && !IsVolatile;
769 insertUse(I, Offset, Size, IsSplittable);
772 void visitLoadInst(LoadInst &LI) {
773 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
774 "All simple FCA loads should have been pre-split");
777 return PI.setAborted(&LI);
779 if (LI.isVolatile() &&
780 LI.getPointerAddressSpace() != DL.getAllocaAddrSpace())
781 return PI.setAborted(&LI);
783 uint64_t Size = DL.getTypeStoreSize(LI.getType()).getFixedSize();
784 return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
787 void visitStoreInst(StoreInst &SI) {
788 Value *ValOp = SI.getValueOperand();
790 return PI.setEscapedAndAborted(&SI);
792 return PI.setAborted(&SI);
794 if (SI.isVolatile() &&
795 SI.getPointerAddressSpace() != DL.getAllocaAddrSpace())
796 return PI.setAborted(&SI);
798 uint64_t Size = DL.getTypeStoreSize(ValOp->getType()).getFixedSize();
800 // If this memory access can be shown to *statically* extend outside the
801 // bounds of the allocation, it's behavior is undefined, so simply
802 // ignore it. Note that this is more strict than the generic clamping
803 // behavior of insertUse. We also try to handle cases which might run the
805 // FIXME: We should instead consider the pointer to have escaped if this
806 // function is being instrumented for addressing bugs or race conditions.
807 if (Size > AllocSize || Offset.ugt(AllocSize - Size)) {
808 LLVM_DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @"
809 << Offset << " which extends past the end of the "
810 << AllocSize << " byte alloca:\n"
811 << " alloca: " << AS.AI << "\n"
812 << " use: " << SI << "\n");
813 return markAsDead(SI);
816 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
817 "All simple FCA stores should have been pre-split");
818 handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
821 void visitMemSetInst(MemSetInst &II) {
822 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
823 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
824 if ((Length && Length->getValue() == 0) ||
825 (IsOffsetKnown && Offset.uge(AllocSize)))
826 // Zero-length mem transfer intrinsics can be ignored entirely.
827 return markAsDead(II);
830 return PI.setAborted(&II);
832 // Don't replace this with a store with a different address space. TODO:
833 // Use a store with the casted new alloca?
834 if (II.isVolatile() && II.getDestAddressSpace() != DL.getAllocaAddrSpace())
835 return PI.setAborted(&II);
837 insertUse(II, Offset, Length ? Length->getLimitedValue()
838 : AllocSize - Offset.getLimitedValue(),
842 void visitMemTransferInst(MemTransferInst &II) {
843 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
844 if (Length && Length->getValue() == 0)
845 // Zero-length mem transfer intrinsics can be ignored entirely.
846 return markAsDead(II);
848 // Because we can visit these intrinsics twice, also check to see if the
849 // first time marked this instruction as dead. If so, skip it.
850 if (VisitedDeadInsts.count(&II))
854 return PI.setAborted(&II);
856 // Don't replace this with a load/store with a different address space.
857 // TODO: Use a store with the casted new alloca?
858 if (II.isVolatile() &&
859 (II.getDestAddressSpace() != DL.getAllocaAddrSpace() ||
860 II.getSourceAddressSpace() != DL.getAllocaAddrSpace()))
861 return PI.setAborted(&II);
863 // This side of the transfer is completely out-of-bounds, and so we can
864 // nuke the entire transfer. However, we also need to nuke the other side
865 // if already added to our partitions.
866 // FIXME: Yet another place we really should bypass this when
867 // instrumenting for ASan.
868 if (Offset.uge(AllocSize)) {
869 SmallDenseMap<Instruction *, unsigned>::iterator MTPI =
870 MemTransferSliceMap.find(&II);
871 if (MTPI != MemTransferSliceMap.end())
872 AS.Slices[MTPI->second].kill();
873 return markAsDead(II);
876 uint64_t RawOffset = Offset.getLimitedValue();
877 uint64_t Size = Length ? Length->getLimitedValue() : AllocSize - RawOffset;
879 // Check for the special case where the same exact value is used for both
881 if (*U == II.getRawDest() && *U == II.getRawSource()) {
882 // For non-volatile transfers this is a no-op.
883 if (!II.isVolatile())
884 return markAsDead(II);
886 return insertUse(II, Offset, Size, /*IsSplittable=*/false);
889 // If we have seen both source and destination for a mem transfer, then
890 // they both point to the same alloca.
892 SmallDenseMap<Instruction *, unsigned>::iterator MTPI;
893 std::tie(MTPI, Inserted) =
894 MemTransferSliceMap.insert(std::make_pair(&II, AS.Slices.size()));
895 unsigned PrevIdx = MTPI->second;
897 Slice &PrevP = AS.Slices[PrevIdx];
899 // Check if the begin offsets match and this is a non-volatile transfer.
900 // In that case, we can completely elide the transfer.
901 if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
903 return markAsDead(II);
906 // Otherwise we have an offset transfer within the same alloca. We can't
908 PrevP.makeUnsplittable();
911 // Insert the use now that we've fixed up the splittable nature.
912 insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
914 // Check that we ended up with a valid index in the map.
915 assert(AS.Slices[PrevIdx].getUse()->getUser() == &II &&
916 "Map index doesn't point back to a slice with this user.");
919 // Disable SRoA for any intrinsics except for lifetime invariants.
920 // FIXME: What about debug intrinsics? This matches old behavior, but
921 // doesn't make sense.
922 void visitIntrinsicInst(IntrinsicInst &II) {
924 return PI.setAborted(&II);
926 if (II.isLifetimeStartOrEnd()) {
927 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
928 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
929 Length->getLimitedValue());
930 insertUse(II, Offset, Size, true);
934 Base::visitIntrinsicInst(II);
937 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
938 // We consider any PHI or select that results in a direct load or store of
939 // the same offset to be a viable use for slicing purposes. These uses
940 // are considered unsplittable and the size is the maximum loaded or stored
942 SmallPtrSet<Instruction *, 4> Visited;
943 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
944 Visited.insert(Root);
945 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
946 const DataLayout &DL = Root->getModule()->getDataLayout();
947 // If there are no loads or stores, the access is dead. We mark that as
948 // a size zero access.
951 Instruction *I, *UsedI;
952 std::tie(UsedI, I) = Uses.pop_back_val();
954 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
955 Size = std::max(Size,
956 DL.getTypeStoreSize(LI->getType()).getFixedSize());
959 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
960 Value *Op = SI->getOperand(0);
963 Size = std::max(Size,
964 DL.getTypeStoreSize(Op->getType()).getFixedSize());
968 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
969 if (!GEP->hasAllZeroIndices())
971 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
972 !isa<SelectInst>(I) && !isa<AddrSpaceCastInst>(I)) {
976 for (User *U : I->users())
977 if (Visited.insert(cast<Instruction>(U)).second)
978 Uses.push_back(std::make_pair(I, cast<Instruction>(U)));
979 } while (!Uses.empty());
984 void visitPHINodeOrSelectInst(Instruction &I) {
985 assert(isa<PHINode>(I) || isa<SelectInst>(I));
987 return markAsDead(I);
989 // TODO: We could use SimplifyInstruction here to fold PHINodes and
990 // SelectInsts. However, doing so requires to change the current
991 // dead-operand-tracking mechanism. For instance, suppose neither loading
992 // from %U nor %other traps. Then "load (select undef, %U, %other)" does not
993 // trap either. However, if we simply replace %U with undef using the
994 // current dead-operand-tracking mechanism, "load (select undef, undef,
995 // %other)" may trap because the select may return the first operand
997 if (Value *Result = foldPHINodeOrSelectInst(I)) {
999 // If the result of the constant fold will be the pointer, recurse
1000 // through the PHI/select as if we had RAUW'ed it.
1003 // Otherwise the operand to the PHI/select is dead, and we can replace
1005 AS.DeadOperands.push_back(U);
1011 return PI.setAborted(&I);
1013 // See if we already have computed info on this node.
1014 uint64_t &Size = PHIOrSelectSizes[&I];
1016 // This is a new PHI/Select, check for an unsafe use of it.
1017 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&I, Size))
1018 return PI.setAborted(UnsafeI);
1021 // For PHI and select operands outside the alloca, we can't nuke the entire
1022 // phi or select -- the other side might still be relevant, so we special
1023 // case them here and use a separate structure to track the operands
1024 // themselves which should be replaced with undef.
1025 // FIXME: This should instead be escaped in the event we're instrumenting
1026 // for address sanitization.
1027 if (Offset.uge(AllocSize)) {
1028 AS.DeadOperands.push_back(U);
1032 insertUse(I, Offset, Size);
1035 void visitPHINode(PHINode &PN) { visitPHINodeOrSelectInst(PN); }
1037 void visitSelectInst(SelectInst &SI) { visitPHINodeOrSelectInst(SI); }
1039 /// Disable SROA entirely if there are unhandled users of the alloca.
1040 void visitInstruction(Instruction &I) { PI.setAborted(&I); }
1043 AllocaSlices::AllocaSlices(const DataLayout &DL, AllocaInst &AI)
1045 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1048 PointerEscapingInstr(nullptr) {
1049 SliceBuilder PB(DL, AI, *this);
1050 SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
1051 if (PtrI.isEscaped() || PtrI.isAborted()) {
1052 // FIXME: We should sink the escape vs. abort info into the caller nicely,
1053 // possibly by just storing the PtrInfo in the AllocaSlices.
1054 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
1055 : PtrI.getAbortingInst();
1056 assert(PointerEscapingInstr && "Did not track a bad instruction");
1061 llvm::remove_if(Slices, [](const Slice &S) { return S.isDead(); }),
1064 // Sort the uses. This arranges for the offsets to be in ascending order,
1065 // and the sizes to be in descending order.
1066 std::stable_sort(Slices.begin(), Slices.end());
1069 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1071 void AllocaSlices::print(raw_ostream &OS, const_iterator I,
1072 StringRef Indent) const {
1073 printSlice(OS, I, Indent);
1075 printUse(OS, I, Indent);
1078 void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
1079 StringRef Indent) const {
1080 OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
1081 << " slice #" << (I - begin())
1082 << (I->isSplittable() ? " (splittable)" : "");
1085 void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
1086 StringRef Indent) const {
1087 OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
1090 void AllocaSlices::print(raw_ostream &OS) const {
1091 if (PointerEscapingInstr) {
1092 OS << "Can't analyze slices for alloca: " << AI << "\n"
1093 << " A pointer to this alloca escaped by:\n"
1094 << " " << *PointerEscapingInstr << "\n";
1098 OS << "Slices of alloca: " << AI << "\n";
1099 for (const_iterator I = begin(), E = end(); I != E; ++I)
1103 LLVM_DUMP_METHOD void AllocaSlices::dump(const_iterator I) const {
1106 LLVM_DUMP_METHOD void AllocaSlices::dump() const { print(dbgs()); }
1108 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1110 /// Walk the range of a partitioning looking for a common type to cover this
1111 /// sequence of slices.
1112 static Type *findCommonType(AllocaSlices::const_iterator B,
1113 AllocaSlices::const_iterator E,
1114 uint64_t EndOffset) {
1116 bool TyIsCommon = true;
1117 IntegerType *ITy = nullptr;
1119 // Note that we need to look at *every* alloca slice's Use to ensure we
1120 // always get consistent results regardless of the order of slices.
1121 for (AllocaSlices::const_iterator I = B; I != E; ++I) {
1122 Use *U = I->getUse();
1123 if (isa<IntrinsicInst>(*U->getUser()))
1125 if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
1128 Type *UserTy = nullptr;
1129 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1130 UserTy = LI->getType();
1131 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1132 UserTy = SI->getValueOperand()->getType();
1135 if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
1136 // If the type is larger than the partition, skip it. We only encounter
1137 // this for split integer operations where we want to use the type of the
1138 // entity causing the split. Also skip if the type is not a byte width
1140 if (UserITy->getBitWidth() % 8 != 0 ||
1141 UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
1144 // Track the largest bitwidth integer type used in this way in case there
1145 // is no common type.
1146 if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
1150 // To avoid depending on the order of slices, Ty and TyIsCommon must not
1151 // depend on types skipped above.
1152 if (!UserTy || (Ty && Ty != UserTy))
1153 TyIsCommon = false; // Give up on anything but an iN type.
1158 return TyIsCommon ? Ty : ITy;
1161 /// PHI instructions that use an alloca and are subsequently loaded can be
1162 /// rewritten to load both input pointers in the pred blocks and then PHI the
1163 /// results, allowing the load of the alloca to be promoted.
1165 /// %P2 = phi [i32* %Alloca, i32* %Other]
1166 /// %V = load i32* %P2
1168 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1170 /// %V2 = load i32* %Other
1172 /// %V = phi [i32 %V1, i32 %V2]
1174 /// We can do this to a select if its only uses are loads and if the operands
1175 /// to the select can be loaded unconditionally.
1177 /// FIXME: This should be hoisted into a generic utility, likely in
1178 /// Transforms/Util/Local.h
1179 static bool isSafePHIToSpeculate(PHINode &PN) {
1180 const DataLayout &DL = PN.getModule()->getDataLayout();
1182 // For now, we can only do this promotion if the load is in the same block
1183 // as the PHI, and if there are no stores between the phi and load.
1184 // TODO: Allow recursive phi users.
1185 // TODO: Allow stores.
1186 BasicBlock *BB = PN.getParent();
1188 uint64_t APWidth = DL.getIndexTypeSizeInBits(PN.getType());
1189 APInt MaxSize(APWidth, 0);
1190 bool HaveLoad = false;
1191 for (User *U : PN.users()) {
1192 LoadInst *LI = dyn_cast<LoadInst>(U);
1193 if (!LI || !LI->isSimple())
1196 // For now we only allow loads in the same block as the PHI. This is
1197 // a common case that happens when instcombine merges two loads through
1199 if (LI->getParent() != BB)
1202 // Ensure that there are no instructions between the PHI and the load that
1204 for (BasicBlock::iterator BBI(PN); &*BBI != LI; ++BBI)
1205 if (BBI->mayWriteToMemory())
1208 uint64_t Size = DL.getTypeStoreSize(LI->getType()).getFixedSize();
1209 MaxAlign = std::max(MaxAlign, LI->getAlign());
1210 MaxSize = MaxSize.ult(Size) ? APInt(APWidth, Size) : MaxSize;
1217 // We can only transform this if it is safe to push the loads into the
1218 // predecessor blocks. The only thing to watch out for is that we can't put
1219 // a possibly trapping load in the predecessor if it is a critical edge.
1220 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1221 Instruction *TI = PN.getIncomingBlock(Idx)->getTerminator();
1222 Value *InVal = PN.getIncomingValue(Idx);
1224 // If the value is produced by the terminator of the predecessor (an
1225 // invoke) or it has side-effects, there is no valid place to put a load
1226 // in the predecessor.
1227 if (TI == InVal || TI->mayHaveSideEffects())
1230 // If the predecessor has a single successor, then the edge isn't
1232 if (TI->getNumSuccessors() == 1)
1235 // If this pointer is always safe to load, or if we can prove that there
1236 // is already a load in the block, then we can move the load to the pred
1238 if (isSafeToLoadUnconditionally(InVal, MaxAlign, MaxSize, DL, TI))
1247 static void speculatePHINodeLoads(PHINode &PN) {
1248 LLVM_DEBUG(dbgs() << " original: " << PN << "\n");
1250 LoadInst *SomeLoad = cast<LoadInst>(PN.user_back());
1251 Type *LoadTy = SomeLoad->getType();
1252 IRBuilderTy PHIBuilder(&PN);
1253 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1254 PN.getName() + ".sroa.speculated");
1256 // Get the AA tags and alignment to use from one of the loads. It does not
1257 // matter which one we get and if any differ.
1259 SomeLoad->getAAMetadata(AATags);
1260 Align Alignment = SomeLoad->getAlign();
1262 // Rewrite all loads of the PN to use the new PHI.
1263 while (!PN.use_empty()) {
1264 LoadInst *LI = cast<LoadInst>(PN.user_back());
1265 LI->replaceAllUsesWith(NewPN);
1266 LI->eraseFromParent();
1269 // Inject loads into all of the pred blocks.
1270 DenseMap<BasicBlock*, Value*> InjectedLoads;
1271 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1272 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1273 Value *InVal = PN.getIncomingValue(Idx);
1275 // A PHI node is allowed to have multiple (duplicated) entries for the same
1276 // basic block, as long as the value is the same. So if we already injected
1277 // a load in the predecessor, then we should reuse the same load for all
1278 // duplicated entries.
1279 if (Value* V = InjectedLoads.lookup(Pred)) {
1280 NewPN->addIncoming(V, Pred);
1284 Instruction *TI = Pred->getTerminator();
1285 IRBuilderTy PredBuilder(TI);
1287 LoadInst *Load = PredBuilder.CreateAlignedLoad(
1288 LoadTy, InVal, Alignment,
1289 (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1290 ++NumLoadsSpeculated;
1292 Load->setAAMetadata(AATags);
1293 NewPN->addIncoming(Load, Pred);
1294 InjectedLoads[Pred] = Load;
1297 LLVM_DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1298 PN.eraseFromParent();
1301 /// Select instructions that use an alloca and are subsequently loaded can be
1302 /// rewritten to load both input pointers and then select between the result,
1303 /// allowing the load of the alloca to be promoted.
1305 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1306 /// %V = load i32* %P2
1308 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1309 /// %V2 = load i32* %Other
1310 /// %V = select i1 %cond, i32 %V1, i32 %V2
1312 /// We can do this to a select if its only uses are loads and if the operand
1313 /// to the select can be loaded unconditionally.
1314 static bool isSafeSelectToSpeculate(SelectInst &SI) {
1315 Value *TValue = SI.getTrueValue();
1316 Value *FValue = SI.getFalseValue();
1317 const DataLayout &DL = SI.getModule()->getDataLayout();
1319 for (User *U : SI.users()) {
1320 LoadInst *LI = dyn_cast<LoadInst>(U);
1321 if (!LI || !LI->isSimple())
1324 // Both operands to the select need to be dereferenceable, either
1325 // absolutely (e.g. allocas) or at this point because we can see other
1327 if (!isSafeToLoadUnconditionally(TValue, LI->getType(),
1328 LI->getAlign(), DL, LI))
1330 if (!isSafeToLoadUnconditionally(FValue, LI->getType(),
1331 LI->getAlign(), DL, LI))
1338 static void speculateSelectInstLoads(SelectInst &SI) {
1339 LLVM_DEBUG(dbgs() << " original: " << SI << "\n");
1341 IRBuilderTy IRB(&SI);
1342 Value *TV = SI.getTrueValue();
1343 Value *FV = SI.getFalseValue();
1344 // Replace the loads of the select with a select of two loads.
1345 while (!SI.use_empty()) {
1346 LoadInst *LI = cast<LoadInst>(SI.user_back());
1347 assert(LI->isSimple() && "We only speculate simple loads");
1349 IRB.SetInsertPoint(LI);
1350 LoadInst *TL = IRB.CreateLoad(LI->getType(), TV,
1351 LI->getName() + ".sroa.speculate.load.true");
1352 LoadInst *FL = IRB.CreateLoad(LI->getType(), FV,
1353 LI->getName() + ".sroa.speculate.load.false");
1354 NumLoadsSpeculated += 2;
1356 // Transfer alignment and AA info if present.
1357 TL->setAlignment(LI->getAlign());
1358 FL->setAlignment(LI->getAlign());
1361 LI->getAAMetadata(Tags);
1363 TL->setAAMetadata(Tags);
1364 FL->setAAMetadata(Tags);
1367 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1368 LI->getName() + ".sroa.speculated");
1370 LLVM_DEBUG(dbgs() << " speculated to: " << *V << "\n");
1371 LI->replaceAllUsesWith(V);
1372 LI->eraseFromParent();
1374 SI.eraseFromParent();
1377 /// Build a GEP out of a base pointer and indices.
1379 /// This will return the BasePtr if that is valid, or build a new GEP
1380 /// instruction using the IRBuilder if GEP-ing is needed.
1381 static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1382 SmallVectorImpl<Value *> &Indices, Twine NamePrefix) {
1383 if (Indices.empty())
1386 // A single zero index is a no-op, so check for this and avoid building a GEP
1388 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1391 return IRB.CreateInBoundsGEP(BasePtr->getType()->getPointerElementType(),
1392 BasePtr, Indices, NamePrefix + "sroa_idx");
1395 /// Get a natural GEP off of the BasePtr walking through Ty toward
1396 /// TargetTy without changing the offset of the pointer.
1398 /// This routine assumes we've already established a properly offset GEP with
1399 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1400 /// zero-indices down through type layers until we find one the same as
1401 /// TargetTy. If we can't find one with the same type, we at least try to use
1402 /// one with the same size. If none of that works, we just produce the GEP as
1403 /// indicated by Indices to have the correct offset.
1404 static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1405 Value *BasePtr, Type *Ty, Type *TargetTy,
1406 SmallVectorImpl<Value *> &Indices,
1409 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1411 // Offset size to use for the indices.
1412 unsigned OffsetSize = DL.getIndexTypeSizeInBits(BasePtr->getType());
1414 // See if we can descend into a struct and locate a field with the correct
1416 unsigned NumLayers = 0;
1417 Type *ElementTy = Ty;
1419 if (ElementTy->isPointerTy())
1422 if (ArrayType *ArrayTy = dyn_cast<ArrayType>(ElementTy)) {
1423 ElementTy = ArrayTy->getElementType();
1424 Indices.push_back(IRB.getIntN(OffsetSize, 0));
1425 } else if (VectorType *VectorTy = dyn_cast<VectorType>(ElementTy)) {
1426 ElementTy = VectorTy->getElementType();
1427 Indices.push_back(IRB.getInt32(0));
1428 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1429 if (STy->element_begin() == STy->element_end())
1430 break; // Nothing left to descend into.
1431 ElementTy = *STy->element_begin();
1432 Indices.push_back(IRB.getInt32(0));
1437 } while (ElementTy != TargetTy);
1438 if (ElementTy != TargetTy)
1439 Indices.erase(Indices.end() - NumLayers, Indices.end());
1441 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1444 /// Recursively compute indices for a natural GEP.
1446 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1447 /// element types adding appropriate indices for the GEP.
1448 static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1449 Value *Ptr, Type *Ty, APInt &Offset,
1451 SmallVectorImpl<Value *> &Indices,
1454 return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices,
1457 // We can't recurse through pointer types.
1458 if (Ty->isPointerTy())
1461 // We try to analyze GEPs over vectors here, but note that these GEPs are
1462 // extremely poorly defined currently. The long-term goal is to remove GEPing
1463 // over a vector from the IR completely.
1464 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1465 unsigned ElementSizeInBits =
1466 DL.getTypeSizeInBits(VecTy->getScalarType()).getFixedSize();
1467 if (ElementSizeInBits % 8 != 0) {
1468 // GEPs over non-multiple of 8 size vector elements are invalid.
1471 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1472 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1473 if (NumSkippedElements.ugt(cast<FixedVectorType>(VecTy)->getNumElements()))
1475 Offset -= NumSkippedElements * ElementSize;
1476 Indices.push_back(IRB.getInt(NumSkippedElements));
1477 return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1478 Offset, TargetTy, Indices, NamePrefix);
1481 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1482 Type *ElementTy = ArrTy->getElementType();
1483 APInt ElementSize(Offset.getBitWidth(),
1484 DL.getTypeAllocSize(ElementTy).getFixedSize());
1485 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1486 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1489 Offset -= NumSkippedElements * ElementSize;
1490 Indices.push_back(IRB.getInt(NumSkippedElements));
1491 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1492 Indices, NamePrefix);
1495 StructType *STy = dyn_cast<StructType>(Ty);
1499 const StructLayout *SL = DL.getStructLayout(STy);
1500 uint64_t StructOffset = Offset.getZExtValue();
1501 if (StructOffset >= SL->getSizeInBytes())
1503 unsigned Index = SL->getElementContainingOffset(StructOffset);
1504 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1505 Type *ElementTy = STy->getElementType(Index);
1506 if (Offset.uge(DL.getTypeAllocSize(ElementTy).getFixedSize()))
1507 return nullptr; // The offset points into alignment padding.
1509 Indices.push_back(IRB.getInt32(Index));
1510 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1511 Indices, NamePrefix);
1514 /// Get a natural GEP from a base pointer to a particular offset and
1515 /// resulting in a particular type.
1517 /// The goal is to produce a "natural" looking GEP that works with the existing
1518 /// composite types to arrive at the appropriate offset and element type for
1519 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1520 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1521 /// Indices, and setting Ty to the result subtype.
1523 /// If no natural GEP can be constructed, this function returns null.
1524 static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1525 Value *Ptr, APInt Offset, Type *TargetTy,
1526 SmallVectorImpl<Value *> &Indices,
1528 PointerType *Ty = cast<PointerType>(Ptr->getType());
1530 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1532 if (Ty == IRB.getInt8PtrTy(Ty->getAddressSpace()) && TargetTy->isIntegerTy(8))
1535 Type *ElementTy = Ty->getElementType();
1536 if (!ElementTy->isSized())
1537 return nullptr; // We can't GEP through an unsized element.
1538 APInt ElementSize(Offset.getBitWidth(),
1539 DL.getTypeAllocSize(ElementTy).getFixedSize());
1540 if (ElementSize == 0)
1541 return nullptr; // Zero-length arrays can't help us build a natural GEP.
1542 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1544 Offset -= NumSkippedElements * ElementSize;
1545 Indices.push_back(IRB.getInt(NumSkippedElements));
1546 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1547 Indices, NamePrefix);
1550 /// Compute an adjusted pointer from Ptr by Offset bytes where the
1551 /// resulting pointer has PointerTy.
1553 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1554 /// and produces the pointer type desired. Where it cannot, it will try to use
1555 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1556 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1557 /// bitcast to the type.
1559 /// The strategy for finding the more natural GEPs is to peel off layers of the
1560 /// pointer, walking back through bit casts and GEPs, searching for a base
1561 /// pointer from which we can compute a natural GEP with the desired
1562 /// properties. The algorithm tries to fold as many constant indices into
1563 /// a single GEP as possible, thus making each GEP more independent of the
1564 /// surrounding code.
1565 static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr,
1566 APInt Offset, Type *PointerTy, Twine NamePrefix) {
1567 // Even though we don't look through PHI nodes, we could be called on an
1568 // instruction in an unreachable block, which may be on a cycle.
1569 SmallPtrSet<Value *, 4> Visited;
1570 Visited.insert(Ptr);
1571 SmallVector<Value *, 4> Indices;
1573 // We may end up computing an offset pointer that has the wrong type. If we
1574 // never are able to compute one directly that has the correct type, we'll
1575 // fall back to it, so keep it and the base it was computed from around here.
1576 Value *OffsetPtr = nullptr;
1577 Value *OffsetBasePtr;
1579 // Remember any i8 pointer we come across to re-use if we need to do a raw
1581 Value *Int8Ptr = nullptr;
1582 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1584 PointerType *TargetPtrTy = cast<PointerType>(PointerTy);
1585 Type *TargetTy = TargetPtrTy->getElementType();
1587 // As `addrspacecast` is , `Ptr` (the storage pointer) may have different
1588 // address space from the expected `PointerTy` (the pointer to be used).
1589 // Adjust the pointer type based the original storage pointer.
1590 auto AS = cast<PointerType>(Ptr->getType())->getAddressSpace();
1591 PointerTy = TargetTy->getPointerTo(AS);
1594 // First fold any existing GEPs into the offset.
1595 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1596 APInt GEPOffset(Offset.getBitWidth(), 0);
1597 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1599 Offset += GEPOffset;
1600 Ptr = GEP->getPointerOperand();
1601 if (!Visited.insert(Ptr).second)
1605 // See if we can perform a natural GEP here.
1607 if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1608 Indices, NamePrefix)) {
1609 // If we have a new natural pointer at the offset, clear out any old
1610 // offset pointer we computed. Unless it is the base pointer or
1611 // a non-instruction, we built a GEP we don't need. Zap it.
1612 if (OffsetPtr && OffsetPtr != OffsetBasePtr)
1613 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr)) {
1614 assert(I->use_empty() && "Built a GEP with uses some how!");
1615 I->eraseFromParent();
1618 OffsetBasePtr = Ptr;
1619 // If we also found a pointer of the right type, we're done.
1620 if (P->getType() == PointerTy)
1624 // Stash this pointer if we've found an i8*.
1625 if (Ptr->getType()->isIntegerTy(8)) {
1627 Int8PtrOffset = Offset;
1630 // Peel off a layer of the pointer and update the offset appropriately.
1631 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1632 Ptr = cast<Operator>(Ptr)->getOperand(0);
1633 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1634 if (GA->isInterposable())
1636 Ptr = GA->getAliasee();
1640 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1641 } while (Visited.insert(Ptr).second);
1645 Int8Ptr = IRB.CreateBitCast(
1646 Ptr, IRB.getInt8PtrTy(PointerTy->getPointerAddressSpace()),
1647 NamePrefix + "sroa_raw_cast");
1648 Int8PtrOffset = Offset;
1651 OffsetPtr = Int8PtrOffset == 0
1653 : IRB.CreateInBoundsGEP(IRB.getInt8Ty(), Int8Ptr,
1654 IRB.getInt(Int8PtrOffset),
1655 NamePrefix + "sroa_raw_idx");
1659 // On the off chance we were targeting i8*, guard the bitcast here.
1660 if (cast<PointerType>(Ptr->getType()) != TargetPtrTy) {
1661 Ptr = IRB.CreatePointerBitCastOrAddrSpaceCast(Ptr,
1663 NamePrefix + "sroa_cast");
1669 /// Compute the adjusted alignment for a load or store from an offset.
1670 static Align getAdjustedAlignment(Instruction *I, uint64_t Offset) {
1671 return commonAlignment(getLoadStoreAlignment(I), Offset);
1674 /// Test whether we can convert a value from the old to the new type.
1676 /// This predicate should be used to guard calls to convertValue in order to
1677 /// ensure that we only try to convert viable values. The strategy is that we
1678 /// will peel off single element struct and array wrappings to get to an
1679 /// underlying value, and convert that value.
1680 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1684 // For integer types, we can't handle any bit-width differences. This would
1685 // break both vector conversions with extension and introduce endianness
1686 // issues when in conjunction with loads and stores.
1687 if (isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) {
1688 assert(cast<IntegerType>(OldTy)->getBitWidth() !=
1689 cast<IntegerType>(NewTy)->getBitWidth() &&
1690 "We can't have the same bitwidth for different int types");
1694 if (DL.getTypeSizeInBits(NewTy).getFixedSize() !=
1695 DL.getTypeSizeInBits(OldTy).getFixedSize())
1697 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1700 // We can convert pointers to integers and vice-versa. Same for vectors
1701 // of pointers and integers.
1702 OldTy = OldTy->getScalarType();
1703 NewTy = NewTy->getScalarType();
1704 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1705 if (NewTy->isPointerTy() && OldTy->isPointerTy()) {
1706 unsigned OldAS = OldTy->getPointerAddressSpace();
1707 unsigned NewAS = NewTy->getPointerAddressSpace();
1708 // Convert pointers if they are pointers from the same address space or
1709 // different integral (not non-integral) address spaces with the same
1711 return OldAS == NewAS ||
1712 (!DL.isNonIntegralAddressSpace(OldAS) &&
1713 !DL.isNonIntegralAddressSpace(NewAS) &&
1714 DL.getPointerSize(OldAS) == DL.getPointerSize(NewAS));
1717 // We can convert integers to integral pointers, but not to non-integral
1719 if (OldTy->isIntegerTy())
1720 return !DL.isNonIntegralPointerType(NewTy);
1722 // We can convert integral pointers to integers, but non-integral pointers
1723 // need to remain pointers.
1724 if (!DL.isNonIntegralPointerType(OldTy))
1725 return NewTy->isIntegerTy();
1733 /// Generic routine to convert an SSA value to a value of a different
1736 /// This will try various different casting techniques, such as bitcasts,
1737 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1738 /// two types for viability with this routine.
1739 static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1741 Type *OldTy = V->getType();
1742 assert(canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type");
1747 assert(!(isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) &&
1748 "Integer types must be the exact same to convert.");
1750 // See if we need inttoptr for this type pair. May require additional bitcast.
1751 if (OldTy->isIntOrIntVectorTy() && NewTy->isPtrOrPtrVectorTy()) {
1752 // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1753 // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1754 // Expand <4 x i32> to <2 x i8*> --> <4 x i32> to <2 x i64> to <2 x i8*>
1755 // Directly handle i64 to i8*
1756 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1760 // See if we need ptrtoint for this type pair. May require additional bitcast.
1761 if (OldTy->isPtrOrPtrVectorTy() && NewTy->isIntOrIntVectorTy()) {
1762 // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1763 // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1764 // Expand <2 x i8*> to <4 x i32> --> <2 x i8*> to <2 x i64> to <4 x i32>
1765 // Expand i8* to i64 --> i8* to i64 to i64
1766 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1770 if (OldTy->isPtrOrPtrVectorTy() && NewTy->isPtrOrPtrVectorTy()) {
1771 unsigned OldAS = OldTy->getPointerAddressSpace();
1772 unsigned NewAS = NewTy->getPointerAddressSpace();
1773 // To convert pointers with different address spaces (they are already
1774 // checked convertible, i.e. they have the same pointer size), so far we
1775 // cannot use `bitcast` (which has restrict on the same address space) or
1776 // `addrspacecast` (which is not always no-op casting). Instead, use a pair
1777 // of no-op `ptrtoint`/`inttoptr` casts through an integer with the same bit
1779 if (OldAS != NewAS) {
1780 assert(DL.getPointerSize(OldAS) == DL.getPointerSize(NewAS));
1781 return IRB.CreateIntToPtr(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1786 return IRB.CreateBitCast(V, NewTy);
1789 /// Test whether the given slice use can be promoted to a vector.
1791 /// This function is called to test each entry in a partition which is slated
1792 /// for a single slice.
1793 static bool isVectorPromotionViableForSlice(Partition &P, const Slice &S,
1795 uint64_t ElementSize,
1796 const DataLayout &DL) {
1797 // First validate the slice offsets.
1798 uint64_t BeginOffset =
1799 std::max(S.beginOffset(), P.beginOffset()) - P.beginOffset();
1800 uint64_t BeginIndex = BeginOffset / ElementSize;
1801 if (BeginIndex * ElementSize != BeginOffset ||
1802 BeginIndex >= cast<FixedVectorType>(Ty)->getNumElements())
1804 uint64_t EndOffset =
1805 std::min(S.endOffset(), P.endOffset()) - P.beginOffset();
1806 uint64_t EndIndex = EndOffset / ElementSize;
1807 if (EndIndex * ElementSize != EndOffset ||
1808 EndIndex > cast<FixedVectorType>(Ty)->getNumElements())
1811 assert(EndIndex > BeginIndex && "Empty vector!");
1812 uint64_t NumElements = EndIndex - BeginIndex;
1813 Type *SliceTy = (NumElements == 1)
1814 ? Ty->getElementType()
1815 : FixedVectorType::get(Ty->getElementType(), NumElements);
1818 Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1820 Use *U = S.getUse();
1822 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1823 if (MI->isVolatile())
1825 if (!S.isSplittable())
1826 return false; // Skip any unsplittable intrinsics.
1827 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1828 if (!II->isLifetimeStartOrEnd())
1830 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1831 // Disable vector promotion when there are loads or stores of an FCA.
1833 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1834 if (LI->isVolatile())
1836 Type *LTy = LI->getType();
1837 if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1838 assert(LTy->isIntegerTy());
1841 if (!canConvertValue(DL, SliceTy, LTy))
1843 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1844 if (SI->isVolatile())
1846 Type *STy = SI->getValueOperand()->getType();
1847 if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1848 assert(STy->isIntegerTy());
1851 if (!canConvertValue(DL, STy, SliceTy))
1860 /// Test whether the given alloca partitioning and range of slices can be
1861 /// promoted to a vector.
1863 /// This is a quick test to check whether we can rewrite a particular alloca
1864 /// partition (and its newly formed alloca) into a vector alloca with only
1865 /// whole-vector loads and stores such that it could be promoted to a vector
1866 /// SSA value. We only can ensure this for a limited set of operations, and we
1867 /// don't want to do the rewrites unless we are confident that the result will
1868 /// be promotable, so we have an early test here.
1869 static VectorType *isVectorPromotionViable(Partition &P, const DataLayout &DL) {
1870 // Collect the candidate types for vector-based promotion. Also track whether
1871 // we have different element types.
1872 SmallVector<VectorType *, 4> CandidateTys;
1873 Type *CommonEltTy = nullptr;
1874 bool HaveCommonEltTy = true;
1875 auto CheckCandidateType = [&](Type *Ty) {
1876 if (auto *VTy = dyn_cast<VectorType>(Ty)) {
1877 // Return if bitcast to vectors is different for total size in bits.
1878 if (!CandidateTys.empty()) {
1879 VectorType *V = CandidateTys[0];
1880 if (DL.getTypeSizeInBits(VTy).getFixedSize() !=
1881 DL.getTypeSizeInBits(V).getFixedSize()) {
1882 CandidateTys.clear();
1886 CandidateTys.push_back(VTy);
1888 CommonEltTy = VTy->getElementType();
1889 else if (CommonEltTy != VTy->getElementType())
1890 HaveCommonEltTy = false;
1893 // Consider any loads or stores that are the exact size of the slice.
1894 for (const Slice &S : P)
1895 if (S.beginOffset() == P.beginOffset() &&
1896 S.endOffset() == P.endOffset()) {
1897 if (auto *LI = dyn_cast<LoadInst>(S.getUse()->getUser()))
1898 CheckCandidateType(LI->getType());
1899 else if (auto *SI = dyn_cast<StoreInst>(S.getUse()->getUser()))
1900 CheckCandidateType(SI->getValueOperand()->getType());
1903 // If we didn't find a vector type, nothing to do here.
1904 if (CandidateTys.empty())
1907 // Remove non-integer vector types if we had multiple common element types.
1908 // FIXME: It'd be nice to replace them with integer vector types, but we can't
1909 // do that until all the backends are known to produce good code for all
1910 // integer vector types.
1911 if (!HaveCommonEltTy) {
1913 llvm::remove_if(CandidateTys,
1914 [](VectorType *VTy) {
1915 return !VTy->getElementType()->isIntegerTy();
1917 CandidateTys.end());
1919 // If there were no integer vector types, give up.
1920 if (CandidateTys.empty())
1923 // Rank the remaining candidate vector types. This is easy because we know
1924 // they're all integer vectors. We sort by ascending number of elements.
1925 auto RankVectorTypes = [&DL](VectorType *RHSTy, VectorType *LHSTy) {
1927 assert(DL.getTypeSizeInBits(RHSTy).getFixedSize() ==
1928 DL.getTypeSizeInBits(LHSTy).getFixedSize() &&
1929 "Cannot have vector types of different sizes!");
1930 assert(RHSTy->getElementType()->isIntegerTy() &&
1931 "All non-integer types eliminated!");
1932 assert(LHSTy->getElementType()->isIntegerTy() &&
1933 "All non-integer types eliminated!");
1934 return cast<FixedVectorType>(RHSTy)->getNumElements() <
1935 cast<FixedVectorType>(LHSTy)->getNumElements();
1937 llvm::sort(CandidateTys, RankVectorTypes);
1939 std::unique(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes),
1940 CandidateTys.end());
1942 // The only way to have the same element type in every vector type is to
1943 // have the same vector type. Check that and remove all but one.
1945 for (VectorType *VTy : CandidateTys) {
1946 assert(VTy->getElementType() == CommonEltTy &&
1947 "Unaccounted for element type!");
1948 assert(VTy == CandidateTys[0] &&
1949 "Different vector types with the same element type!");
1952 CandidateTys.resize(1);
1955 // Try each vector type, and return the one which works.
1956 auto CheckVectorTypeForPromotion = [&](VectorType *VTy) {
1957 uint64_t ElementSize =
1958 DL.getTypeSizeInBits(VTy->getElementType()).getFixedSize();
1960 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1961 // that aren't byte sized.
1962 if (ElementSize % 8)
1964 assert((DL.getTypeSizeInBits(VTy).getFixedSize() % 8) == 0 &&
1965 "vector size not a multiple of element size?");
1968 for (const Slice &S : P)
1969 if (!isVectorPromotionViableForSlice(P, S, VTy, ElementSize, DL))
1972 for (const Slice *S : P.splitSliceTails())
1973 if (!isVectorPromotionViableForSlice(P, *S, VTy, ElementSize, DL))
1978 for (VectorType *VTy : CandidateTys)
1979 if (CheckVectorTypeForPromotion(VTy))
1985 /// Test whether a slice of an alloca is valid for integer widening.
1987 /// This implements the necessary checking for the \c isIntegerWideningViable
1988 /// test below on a single slice of the alloca.
1989 static bool isIntegerWideningViableForSlice(const Slice &S,
1990 uint64_t AllocBeginOffset,
1992 const DataLayout &DL,
1993 bool &WholeAllocaOp) {
1994 uint64_t Size = DL.getTypeStoreSize(AllocaTy).getFixedSize();
1996 uint64_t RelBegin = S.beginOffset() - AllocBeginOffset;
1997 uint64_t RelEnd = S.endOffset() - AllocBeginOffset;
1999 // We can't reasonably handle cases where the load or store extends past
2000 // the end of the alloca's type and into its padding.
2004 Use *U = S.getUse();
2006 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
2007 if (LI->isVolatile())
2009 // We can't handle loads that extend past the allocated memory.
2010 if (DL.getTypeStoreSize(LI->getType()).getFixedSize() > Size)
2012 // So far, AllocaSliceRewriter does not support widening split slice tails
2013 // in rewriteIntegerLoad.
2014 if (S.beginOffset() < AllocBeginOffset)
2016 // Note that we don't count vector loads or stores as whole-alloca
2017 // operations which enable integer widening because we would prefer to use
2018 // vector widening instead.
2019 if (!isa<VectorType>(LI->getType()) && RelBegin == 0 && RelEnd == Size)
2020 WholeAllocaOp = true;
2021 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
2022 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy).getFixedSize())
2024 } else if (RelBegin != 0 || RelEnd != Size ||
2025 !canConvertValue(DL, AllocaTy, LI->getType())) {
2026 // Non-integer loads need to be convertible from the alloca type so that
2027 // they are promotable.
2030 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
2031 Type *ValueTy = SI->getValueOperand()->getType();
2032 if (SI->isVolatile())
2034 // We can't handle stores that extend past the allocated memory.
2035 if (DL.getTypeStoreSize(ValueTy).getFixedSize() > Size)
2037 // So far, AllocaSliceRewriter does not support widening split slice tails
2038 // in rewriteIntegerStore.
2039 if (S.beginOffset() < AllocBeginOffset)
2041 // Note that we don't count vector loads or stores as whole-alloca
2042 // operations which enable integer widening because we would prefer to use
2043 // vector widening instead.
2044 if (!isa<VectorType>(ValueTy) && RelBegin == 0 && RelEnd == Size)
2045 WholeAllocaOp = true;
2046 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
2047 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy).getFixedSize())
2049 } else if (RelBegin != 0 || RelEnd != Size ||
2050 !canConvertValue(DL, ValueTy, AllocaTy)) {
2051 // Non-integer stores need to be convertible to the alloca type so that
2052 // they are promotable.
2055 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
2056 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
2058 if (!S.isSplittable())
2059 return false; // Skip any unsplittable intrinsics.
2060 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
2061 if (!II->isLifetimeStartOrEnd())
2070 /// Test whether the given alloca partition's integer operations can be
2071 /// widened to promotable ones.
2073 /// This is a quick test to check whether we can rewrite the integer loads and
2074 /// stores to a particular alloca into wider loads and stores and be able to
2075 /// promote the resulting alloca.
2076 static bool isIntegerWideningViable(Partition &P, Type *AllocaTy,
2077 const DataLayout &DL) {
2078 uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy).getFixedSize();
2079 // Don't create integer types larger than the maximum bitwidth.
2080 if (SizeInBits > IntegerType::MAX_INT_BITS)
2083 // Don't try to handle allocas with bit-padding.
2084 if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy).getFixedSize())
2087 // We need to ensure that an integer type with the appropriate bitwidth can
2088 // be converted to the alloca type, whatever that is. We don't want to force
2089 // the alloca itself to have an integer type if there is a more suitable one.
2090 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
2091 if (!canConvertValue(DL, AllocaTy, IntTy) ||
2092 !canConvertValue(DL, IntTy, AllocaTy))
2095 // While examining uses, we ensure that the alloca has a covering load or
2096 // store. We don't want to widen the integer operations only to fail to
2097 // promote due to some other unsplittable entry (which we may make splittable
2098 // later). However, if there are only splittable uses, go ahead and assume
2099 // that we cover the alloca.
2100 // FIXME: We shouldn't consider split slices that happen to start in the
2101 // partition here...
2102 bool WholeAllocaOp =
2103 P.begin() != P.end() ? false : DL.isLegalInteger(SizeInBits);
2105 for (const Slice &S : P)
2106 if (!isIntegerWideningViableForSlice(S, P.beginOffset(), AllocaTy, DL,
2110 for (const Slice *S : P.splitSliceTails())
2111 if (!isIntegerWideningViableForSlice(*S, P.beginOffset(), AllocaTy, DL,
2115 return WholeAllocaOp;
2118 static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
2119 IntegerType *Ty, uint64_t Offset,
2120 const Twine &Name) {
2121 LLVM_DEBUG(dbgs() << " start: " << *V << "\n");
2122 IntegerType *IntTy = cast<IntegerType>(V->getType());
2123 assert(DL.getTypeStoreSize(Ty).getFixedSize() + Offset <=
2124 DL.getTypeStoreSize(IntTy).getFixedSize() &&
2125 "Element extends past full value");
2126 uint64_t ShAmt = 8 * Offset;
2127 if (DL.isBigEndian())
2128 ShAmt = 8 * (DL.getTypeStoreSize(IntTy).getFixedSize() -
2129 DL.getTypeStoreSize(Ty).getFixedSize() - Offset);
2131 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
2132 LLVM_DEBUG(dbgs() << " shifted: " << *V << "\n");
2134 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2135 "Cannot extract to a larger integer!");
2137 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
2138 LLVM_DEBUG(dbgs() << " trunced: " << *V << "\n");
2143 static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
2144 Value *V, uint64_t Offset, const Twine &Name) {
2145 IntegerType *IntTy = cast<IntegerType>(Old->getType());
2146 IntegerType *Ty = cast<IntegerType>(V->getType());
2147 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2148 "Cannot insert a larger integer!");
2149 LLVM_DEBUG(dbgs() << " start: " << *V << "\n");
2151 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
2152 LLVM_DEBUG(dbgs() << " extended: " << *V << "\n");
2154 assert(DL.getTypeStoreSize(Ty).getFixedSize() + Offset <=
2155 DL.getTypeStoreSize(IntTy).getFixedSize() &&
2156 "Element store outside of alloca store");
2157 uint64_t ShAmt = 8 * Offset;
2158 if (DL.isBigEndian())
2159 ShAmt = 8 * (DL.getTypeStoreSize(IntTy).getFixedSize() -
2160 DL.getTypeStoreSize(Ty).getFixedSize() - Offset);
2162 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
2163 LLVM_DEBUG(dbgs() << " shifted: " << *V << "\n");
2166 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
2167 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
2168 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
2169 LLVM_DEBUG(dbgs() << " masked: " << *Old << "\n");
2170 V = IRB.CreateOr(Old, V, Name + ".insert");
2171 LLVM_DEBUG(dbgs() << " inserted: " << *V << "\n");
2176 static Value *extractVector(IRBuilderTy &IRB, Value *V, unsigned BeginIndex,
2177 unsigned EndIndex, const Twine &Name) {
2178 auto *VecTy = cast<FixedVectorType>(V->getType());
2179 unsigned NumElements = EndIndex - BeginIndex;
2180 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2182 if (NumElements == VecTy->getNumElements())
2185 if (NumElements == 1) {
2186 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
2188 LLVM_DEBUG(dbgs() << " extract: " << *V << "\n");
2192 SmallVector<int, 8> Mask;
2193 Mask.reserve(NumElements);
2194 for (unsigned i = BeginIndex; i != EndIndex; ++i)
2196 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()), Mask,
2198 LLVM_DEBUG(dbgs() << " shuffle: " << *V << "\n");
2202 static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
2203 unsigned BeginIndex, const Twine &Name) {
2204 VectorType *VecTy = cast<VectorType>(Old->getType());
2205 assert(VecTy && "Can only insert a vector into a vector");
2207 VectorType *Ty = dyn_cast<VectorType>(V->getType());
2209 // Single element to insert.
2210 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
2212 LLVM_DEBUG(dbgs() << " insert: " << *V << "\n");
2216 assert(cast<FixedVectorType>(Ty)->getNumElements() <=
2217 cast<FixedVectorType>(VecTy)->getNumElements() &&
2218 "Too many elements!");
2219 if (cast<FixedVectorType>(Ty)->getNumElements() ==
2220 cast<FixedVectorType>(VecTy)->getNumElements()) {
2221 assert(V->getType() == VecTy && "Vector type mismatch");
2224 unsigned EndIndex = BeginIndex + cast<FixedVectorType>(Ty)->getNumElements();
2226 // When inserting a smaller vector into the larger to store, we first
2227 // use a shuffle vector to widen it with undef elements, and then
2228 // a second shuffle vector to select between the loaded vector and the
2230 SmallVector<Constant *, 8> Mask;
2231 Mask.reserve(cast<FixedVectorType>(VecTy)->getNumElements());
2232 for (unsigned i = 0; i != cast<FixedVectorType>(VecTy)->getNumElements(); ++i)
2233 if (i >= BeginIndex && i < EndIndex)
2234 Mask.push_back(IRB.getInt32(i - BeginIndex));
2236 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
2237 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2238 ConstantVector::get(Mask), Name + ".expand");
2239 LLVM_DEBUG(dbgs() << " shuffle: " << *V << "\n");
2242 for (unsigned i = 0; i != cast<FixedVectorType>(VecTy)->getNumElements(); ++i)
2243 Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
2245 V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
2247 LLVM_DEBUG(dbgs() << " blend: " << *V << "\n");
2251 /// Visitor to rewrite instructions using p particular slice of an alloca
2252 /// to use a new alloca.
2254 /// Also implements the rewriting to vector-based accesses when the partition
2255 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2257 class llvm::sroa::AllocaSliceRewriter
2258 : public InstVisitor<AllocaSliceRewriter, bool> {
2259 // Befriend the base class so it can delegate to private visit methods.
2260 friend class InstVisitor<AllocaSliceRewriter, bool>;
2262 using Base = InstVisitor<AllocaSliceRewriter, bool>;
2264 const DataLayout &DL;
2267 AllocaInst &OldAI, &NewAI;
2268 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2271 // This is a convenience and flag variable that will be null unless the new
2272 // alloca's integer operations should be widened to this integer type due to
2273 // passing isIntegerWideningViable above. If it is non-null, the desired
2274 // integer type will be stored here for easy access during rewriting.
2277 // If we are rewriting an alloca partition which can be written as pure
2278 // vector operations, we stash extra information here. When VecTy is
2279 // non-null, we have some strict guarantees about the rewritten alloca:
2280 // - The new alloca is exactly the size of the vector type here.
2281 // - The accesses all either map to the entire vector or to a single
2283 // - The set of accessing instructions is only one of those handled above
2284 // in isVectorPromotionViable. Generally these are the same access kinds
2285 // which are promotable via mem2reg.
2288 uint64_t ElementSize;
2290 // The original offset of the slice currently being rewritten relative to
2291 // the original alloca.
2292 uint64_t BeginOffset = 0;
2293 uint64_t EndOffset = 0;
2295 // The new offsets of the slice currently being rewritten relative to the
2297 uint64_t NewBeginOffset = 0, NewEndOffset = 0;
2299 uint64_t SliceSize = 0;
2300 bool IsSplittable = false;
2301 bool IsSplit = false;
2302 Use *OldUse = nullptr;
2303 Instruction *OldPtr = nullptr;
2305 // Track post-rewrite users which are PHI nodes and Selects.
2306 SmallSetVector<PHINode *, 8> &PHIUsers;
2307 SmallSetVector<SelectInst *, 8> &SelectUsers;
2309 // Utility IR builder, whose name prefix is setup for each visited use, and
2310 // the insertion point is set to point to the user.
2314 AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &AS, SROA &Pass,
2315 AllocaInst &OldAI, AllocaInst &NewAI,
2316 uint64_t NewAllocaBeginOffset,
2317 uint64_t NewAllocaEndOffset, bool IsIntegerPromotable,
2318 VectorType *PromotableVecTy,
2319 SmallSetVector<PHINode *, 8> &PHIUsers,
2320 SmallSetVector<SelectInst *, 8> &SelectUsers)
2321 : DL(DL), AS(AS), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
2322 NewAllocaBeginOffset(NewAllocaBeginOffset),
2323 NewAllocaEndOffset(NewAllocaEndOffset),
2324 NewAllocaTy(NewAI.getAllocatedType()),
2327 ? Type::getIntNTy(NewAI.getContext(),
2328 DL.getTypeSizeInBits(NewAI.getAllocatedType())
2331 VecTy(PromotableVecTy),
2332 ElementTy(VecTy ? VecTy->getElementType() : nullptr),
2333 ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy).getFixedSize() / 8
2335 PHIUsers(PHIUsers), SelectUsers(SelectUsers),
2336 IRB(NewAI.getContext(), ConstantFolder()) {
2338 assert((DL.getTypeSizeInBits(ElementTy).getFixedSize() % 8) == 0 &&
2339 "Only multiple-of-8 sized vector elements are viable");
2342 assert((!IntTy && !VecTy) || (IntTy && !VecTy) || (!IntTy && VecTy));
2345 bool visit(AllocaSlices::const_iterator I) {
2346 bool CanSROA = true;
2347 BeginOffset = I->beginOffset();
2348 EndOffset = I->endOffset();
2349 IsSplittable = I->isSplittable();
2351 BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
2352 LLVM_DEBUG(dbgs() << " rewriting " << (IsSplit ? "split " : ""));
2353 LLVM_DEBUG(AS.printSlice(dbgs(), I, ""));
2354 LLVM_DEBUG(dbgs() << "\n");
2356 // Compute the intersecting offset range.
2357 assert(BeginOffset < NewAllocaEndOffset);
2358 assert(EndOffset > NewAllocaBeginOffset);
2359 NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2360 NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2362 SliceSize = NewEndOffset - NewBeginOffset;
2364 OldUse = I->getUse();
2365 OldPtr = cast<Instruction>(OldUse->get());
2367 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2368 IRB.SetInsertPoint(OldUserI);
2369 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2370 IRB.getInserter().SetNamePrefix(
2371 Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
2373 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2380 // Make sure the other visit overloads are visible.
2383 // Every instruction which can end up as a user must have a rewrite rule.
2384 bool visitInstruction(Instruction &I) {
2385 LLVM_DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2386 llvm_unreachable("No rewrite rule for this instruction!");
2389 Value *getNewAllocaSlicePtr(IRBuilderTy &IRB, Type *PointerTy) {
2390 // Note that the offset computation can use BeginOffset or NewBeginOffset
2391 // interchangeably for unsplit slices.
2392 assert(IsSplit || BeginOffset == NewBeginOffset);
2393 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2396 StringRef OldName = OldPtr->getName();
2397 // Skip through the last '.sroa.' component of the name.
2398 size_t LastSROAPrefix = OldName.rfind(".sroa.");
2399 if (LastSROAPrefix != StringRef::npos) {
2400 OldName = OldName.substr(LastSROAPrefix + strlen(".sroa."));
2401 // Look for an SROA slice index.
2402 size_t IndexEnd = OldName.find_first_not_of("0123456789");
2403 if (IndexEnd != StringRef::npos && OldName[IndexEnd] == '.') {
2404 // Strip the index and look for the offset.
2405 OldName = OldName.substr(IndexEnd + 1);
2406 size_t OffsetEnd = OldName.find_first_not_of("0123456789");
2407 if (OffsetEnd != StringRef::npos && OldName[OffsetEnd] == '.')
2408 // Strip the offset.
2409 OldName = OldName.substr(OffsetEnd + 1);
2412 // Strip any SROA suffixes as well.
2413 OldName = OldName.substr(0, OldName.find(".sroa_"));
2416 return getAdjustedPtr(IRB, DL, &NewAI,
2417 APInt(DL.getIndexTypeSizeInBits(PointerTy), Offset),
2420 Twine(OldName) + "."
2427 /// Compute suitable alignment to access this slice of the *new*
2430 /// You can optionally pass a type to this routine and if that type's ABI
2431 /// alignment is itself suitable, this will return zero.
2432 Align getSliceAlign() {
2433 return commonAlignment(NewAI.getAlign(),
2434 NewBeginOffset - NewAllocaBeginOffset);
2437 unsigned getIndex(uint64_t Offset) {
2438 assert(VecTy && "Can only call getIndex when rewriting a vector");
2439 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2440 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2441 uint32_t Index = RelOffset / ElementSize;
2442 assert(Index * ElementSize == RelOffset);
2446 void deleteIfTriviallyDead(Value *V) {
2447 Instruction *I = cast<Instruction>(V);
2448 if (isInstructionTriviallyDead(I))
2449 Pass.DeadInsts.insert(I);
2452 Value *rewriteVectorizedLoadInst() {
2453 unsigned BeginIndex = getIndex(NewBeginOffset);
2454 unsigned EndIndex = getIndex(NewEndOffset);
2455 assert(EndIndex > BeginIndex && "Empty vector!");
2457 Value *V = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2458 NewAI.getAlign(), "load");
2459 return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2462 Value *rewriteIntegerLoad(LoadInst &LI) {
2463 assert(IntTy && "We cannot insert an integer to the alloca");
2464 assert(!LI.isVolatile());
2465 Value *V = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2466 NewAI.getAlign(), "load");
2467 V = convertValue(DL, IRB, V, IntTy);
2468 assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2469 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2470 if (Offset > 0 || NewEndOffset < NewAllocaEndOffset) {
2471 IntegerType *ExtractTy = Type::getIntNTy(LI.getContext(), SliceSize * 8);
2472 V = extractInteger(DL, IRB, V, ExtractTy, Offset, "extract");
2474 // It is possible that the extracted type is not the load type. This
2475 // happens if there is a load past the end of the alloca, and as
2476 // a consequence the slice is narrower but still a candidate for integer
2477 // lowering. To handle this case, we just zero extend the extracted
2479 assert(cast<IntegerType>(LI.getType())->getBitWidth() >= SliceSize * 8 &&
2480 "Can only handle an extract for an overly wide load");
2481 if (cast<IntegerType>(LI.getType())->getBitWidth() > SliceSize * 8)
2482 V = IRB.CreateZExt(V, LI.getType());
2486 bool visitLoadInst(LoadInst &LI) {
2487 LLVM_DEBUG(dbgs() << " original: " << LI << "\n");
2488 Value *OldOp = LI.getOperand(0);
2489 assert(OldOp == OldPtr);
2492 LI.getAAMetadata(AATags);
2494 unsigned AS = LI.getPointerAddressSpace();
2496 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), SliceSize * 8)
2498 const bool IsLoadPastEnd =
2499 DL.getTypeStoreSize(TargetTy).getFixedSize() > SliceSize;
2500 bool IsPtrAdjusted = false;
2503 V = rewriteVectorizedLoadInst();
2504 } else if (IntTy && LI.getType()->isIntegerTy()) {
2505 V = rewriteIntegerLoad(LI);
2506 } else if (NewBeginOffset == NewAllocaBeginOffset &&
2507 NewEndOffset == NewAllocaEndOffset &&
2508 (canConvertValue(DL, NewAllocaTy, TargetTy) ||
2509 (IsLoadPastEnd && NewAllocaTy->isIntegerTy() &&
2510 TargetTy->isIntegerTy()))) {
2511 LoadInst *NewLI = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2512 NewAI.getAlign(), LI.isVolatile(),
2515 NewLI->setAAMetadata(AATags);
2516 if (LI.isVolatile())
2517 NewLI->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
2518 if (NewLI->isAtomic())
2519 NewLI->setAlignment(LI.getAlign());
2521 // Any !nonnull metadata or !range metadata on the old load is also valid
2522 // on the new load. This is even true in some cases even when the loads
2523 // are different types, for example by mapping !nonnull metadata to
2524 // !range metadata by modeling the null pointer constant converted to the
2526 // FIXME: Add support for range metadata here. Currently the utilities
2527 // for this don't propagate range metadata in trivial cases from one
2528 // integer load to another, don't handle non-addrspace-0 null pointers
2529 // correctly, and don't have any support for mapping ranges as the
2530 // integer type becomes winder or narrower.
2531 if (MDNode *N = LI.getMetadata(LLVMContext::MD_nonnull))
2532 copyNonnullMetadata(LI, N, *NewLI);
2534 // Try to preserve nonnull metadata
2537 // If this is an integer load past the end of the slice (which means the
2538 // bytes outside the slice are undef or this load is dead) just forcibly
2539 // fix the integer size with correct handling of endianness.
2540 if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2541 if (auto *TITy = dyn_cast<IntegerType>(TargetTy))
2542 if (AITy->getBitWidth() < TITy->getBitWidth()) {
2543 V = IRB.CreateZExt(V, TITy, "load.ext");
2544 if (DL.isBigEndian())
2545 V = IRB.CreateShl(V, TITy->getBitWidth() - AITy->getBitWidth(),
2549 Type *LTy = TargetTy->getPointerTo(AS);
2551 IRB.CreateAlignedLoad(TargetTy, getNewAllocaSlicePtr(IRB, LTy),
2552 getSliceAlign(), LI.isVolatile(), LI.getName());
2554 NewLI->setAAMetadata(AATags);
2555 if (LI.isVolatile())
2556 NewLI->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
2559 IsPtrAdjusted = true;
2561 V = convertValue(DL, IRB, V, TargetTy);
2564 assert(!LI.isVolatile());
2565 assert(LI.getType()->isIntegerTy() &&
2566 "Only integer type loads and stores are split");
2567 assert(SliceSize < DL.getTypeStoreSize(LI.getType()).getFixedSize() &&
2568 "Split load isn't smaller than original load");
2569 assert(DL.typeSizeEqualsStoreSize(LI.getType()) &&
2570 "Non-byte-multiple bit width");
2571 // Move the insertion point just past the load so that we can refer to it.
2572 IRB.SetInsertPoint(&*std::next(BasicBlock::iterator(&LI)));
2573 // Create a placeholder value with the same type as LI to use as the
2574 // basis for the new value. This allows us to replace the uses of LI with
2575 // the computed value, and then replace the placeholder with LI, leaving
2576 // LI only used for this computation.
2577 Value *Placeholder = new LoadInst(
2578 LI.getType(), UndefValue::get(LI.getType()->getPointerTo(AS)), "",
2580 V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset - BeginOffset,
2582 LI.replaceAllUsesWith(V);
2583 Placeholder->replaceAllUsesWith(&LI);
2584 Placeholder->deleteValue();
2586 LI.replaceAllUsesWith(V);
2589 Pass.DeadInsts.insert(&LI);
2590 deleteIfTriviallyDead(OldOp);
2591 LLVM_DEBUG(dbgs() << " to: " << *V << "\n");
2592 return !LI.isVolatile() && !IsPtrAdjusted;
2595 bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp,
2597 if (V->getType() != VecTy) {
2598 unsigned BeginIndex = getIndex(NewBeginOffset);
2599 unsigned EndIndex = getIndex(NewEndOffset);
2600 assert(EndIndex > BeginIndex && "Empty vector!");
2601 unsigned NumElements = EndIndex - BeginIndex;
2602 assert(NumElements <= cast<FixedVectorType>(VecTy)->getNumElements() &&
2603 "Too many elements!");
2604 Type *SliceTy = (NumElements == 1)
2606 : FixedVectorType::get(ElementTy, NumElements);
2607 if (V->getType() != SliceTy)
2608 V = convertValue(DL, IRB, V, SliceTy);
2610 // Mix in the existing elements.
2611 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2612 NewAI.getAlign(), "load");
2613 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2615 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlign());
2617 Store->setAAMetadata(AATags);
2618 Pass.DeadInsts.insert(&SI);
2620 LLVM_DEBUG(dbgs() << " to: " << *Store << "\n");
2624 bool rewriteIntegerStore(Value *V, StoreInst &SI, AAMDNodes AATags) {
2625 assert(IntTy && "We cannot extract an integer from the alloca");
2626 assert(!SI.isVolatile());
2627 if (DL.getTypeSizeInBits(V->getType()).getFixedSize() !=
2628 IntTy->getBitWidth()) {
2629 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2630 NewAI.getAlign(), "oldload");
2631 Old = convertValue(DL, IRB, Old, IntTy);
2632 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2633 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2634 V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset, "insert");
2636 V = convertValue(DL, IRB, V, NewAllocaTy);
2637 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlign());
2638 Store->copyMetadata(SI, {LLVMContext::MD_mem_parallel_loop_access,
2639 LLVMContext::MD_access_group});
2641 Store->setAAMetadata(AATags);
2642 Pass.DeadInsts.insert(&SI);
2643 LLVM_DEBUG(dbgs() << " to: " << *Store << "\n");
2647 bool visitStoreInst(StoreInst &SI) {
2648 LLVM_DEBUG(dbgs() << " original: " << SI << "\n");
2649 Value *OldOp = SI.getOperand(1);
2650 assert(OldOp == OldPtr);
2653 SI.getAAMetadata(AATags);
2655 Value *V = SI.getValueOperand();
2657 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2658 // alloca that should be re-examined after promoting this alloca.
2659 if (V->getType()->isPointerTy())
2660 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2661 Pass.PostPromotionWorklist.insert(AI);
2663 if (SliceSize < DL.getTypeStoreSize(V->getType()).getFixedSize()) {
2664 assert(!SI.isVolatile());
2665 assert(V->getType()->isIntegerTy() &&
2666 "Only integer type loads and stores are split");
2667 assert(DL.typeSizeEqualsStoreSize(V->getType()) &&
2668 "Non-byte-multiple bit width");
2669 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), SliceSize * 8);
2670 V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset - BeginOffset,
2675 return rewriteVectorizedStoreInst(V, SI, OldOp, AATags);
2676 if (IntTy && V->getType()->isIntegerTy())
2677 return rewriteIntegerStore(V, SI, AATags);
2679 const bool IsStorePastEnd =
2680 DL.getTypeStoreSize(V->getType()).getFixedSize() > SliceSize;
2682 if (NewBeginOffset == NewAllocaBeginOffset &&
2683 NewEndOffset == NewAllocaEndOffset &&
2684 (canConvertValue(DL, V->getType(), NewAllocaTy) ||
2685 (IsStorePastEnd && NewAllocaTy->isIntegerTy() &&
2686 V->getType()->isIntegerTy()))) {
2687 // If this is an integer store past the end of slice (and thus the bytes
2688 // past that point are irrelevant or this is unreachable), truncate the
2689 // value prior to storing.
2690 if (auto *VITy = dyn_cast<IntegerType>(V->getType()))
2691 if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2692 if (VITy->getBitWidth() > AITy->getBitWidth()) {
2693 if (DL.isBigEndian())
2694 V = IRB.CreateLShr(V, VITy->getBitWidth() - AITy->getBitWidth(),
2696 V = IRB.CreateTrunc(V, AITy, "load.trunc");
2699 V = convertValue(DL, IRB, V, NewAllocaTy);
2701 IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlign(), SI.isVolatile());
2703 unsigned AS = SI.getPointerAddressSpace();
2704 Value *NewPtr = getNewAllocaSlicePtr(IRB, V->getType()->getPointerTo(AS));
2706 IRB.CreateAlignedStore(V, NewPtr, getSliceAlign(), SI.isVolatile());
2708 NewSI->copyMetadata(SI, {LLVMContext::MD_mem_parallel_loop_access,
2709 LLVMContext::MD_access_group});
2711 NewSI->setAAMetadata(AATags);
2712 if (SI.isVolatile())
2713 NewSI->setAtomic(SI.getOrdering(), SI.getSyncScopeID());
2714 if (NewSI->isAtomic())
2715 NewSI->setAlignment(SI.getAlign());
2716 Pass.DeadInsts.insert(&SI);
2717 deleteIfTriviallyDead(OldOp);
2719 LLVM_DEBUG(dbgs() << " to: " << *NewSI << "\n");
2720 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2723 /// Compute an integer value from splatting an i8 across the given
2724 /// number of bytes.
2726 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2727 /// call this routine.
2728 /// FIXME: Heed the advice above.
2730 /// \param V The i8 value to splat.
2731 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2732 Value *getIntegerSplat(Value *V, unsigned Size) {
2733 assert(Size > 0 && "Expected a positive number of bytes.");
2734 IntegerType *VTy = cast<IntegerType>(V->getType());
2735 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2739 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size * 8);
2741 IRB.CreateZExt(V, SplatIntTy, "zext"),
2742 ConstantExpr::getUDiv(
2743 Constant::getAllOnesValue(SplatIntTy),
2744 ConstantExpr::getZExt(Constant::getAllOnesValue(V->getType()),
2750 /// Compute a vector splat for a given element value.
2751 Value *getVectorSplat(Value *V, unsigned NumElements) {
2752 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2753 LLVM_DEBUG(dbgs() << " splat: " << *V << "\n");
2757 bool visitMemSetInst(MemSetInst &II) {
2758 LLVM_DEBUG(dbgs() << " original: " << II << "\n");
2759 assert(II.getRawDest() == OldPtr);
2762 II.getAAMetadata(AATags);
2764 // If the memset has a variable size, it cannot be split, just adjust the
2765 // pointer to the new alloca.
2766 if (!isa<Constant>(II.getLength())) {
2768 assert(NewBeginOffset == BeginOffset);
2769 II.setDest(getNewAllocaSlicePtr(IRB, OldPtr->getType()));
2770 II.setDestAlignment(getSliceAlign());
2772 deleteIfTriviallyDead(OldPtr);
2776 // Record this instruction for deletion.
2777 Pass.DeadInsts.insert(&II);
2779 Type *AllocaTy = NewAI.getAllocatedType();
2780 Type *ScalarTy = AllocaTy->getScalarType();
2782 const bool CanContinue = [&]() {
2785 if (BeginOffset > NewAllocaBeginOffset ||
2786 EndOffset < NewAllocaEndOffset)
2788 auto *C = cast<ConstantInt>(II.getLength());
2789 if (C->getBitWidth() > 64)
2791 const auto Len = C->getZExtValue();
2792 auto *Int8Ty = IntegerType::getInt8Ty(NewAI.getContext());
2793 auto *SrcTy = FixedVectorType::get(Int8Ty, Len);
2794 return canConvertValue(DL, SrcTy, AllocaTy) &&
2795 DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy).getFixedSize());
2798 // If this doesn't map cleanly onto the alloca type, and that type isn't
2799 // a single value type, just emit a memset.
2801 Type *SizeTy = II.getLength()->getType();
2802 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2803 CallInst *New = IRB.CreateMemSet(
2804 getNewAllocaSlicePtr(IRB, OldPtr->getType()), II.getValue(), Size,
2805 MaybeAlign(getSliceAlign()), II.isVolatile());
2807 New->setAAMetadata(AATags);
2808 LLVM_DEBUG(dbgs() << " to: " << *New << "\n");
2812 // If we can represent this as a simple value, we have to build the actual
2813 // value to store, which requires expanding the byte present in memset to
2814 // a sensible representation for the alloca type. This is essentially
2815 // splatting the byte to a sufficiently wide integer, splatting it across
2816 // any desired vector width, and bitcasting to the final type.
2820 // If this is a memset of a vectorized alloca, insert it.
2821 assert(ElementTy == ScalarTy);
2823 unsigned BeginIndex = getIndex(NewBeginOffset);
2824 unsigned EndIndex = getIndex(NewEndOffset);
2825 assert(EndIndex > BeginIndex && "Empty vector!");
2826 unsigned NumElements = EndIndex - BeginIndex;
2827 assert(NumElements <= cast<FixedVectorType>(VecTy)->getNumElements() &&
2828 "Too many elements!");
2830 Value *Splat = getIntegerSplat(
2831 II.getValue(), DL.getTypeSizeInBits(ElementTy).getFixedSize() / 8);
2832 Splat = convertValue(DL, IRB, Splat, ElementTy);
2833 if (NumElements > 1)
2834 Splat = getVectorSplat(Splat, NumElements);
2836 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2837 NewAI.getAlign(), "oldload");
2838 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2840 // If this is a memset on an alloca where we can widen stores, insert the
2842 assert(!II.isVolatile());
2844 uint64_t Size = NewEndOffset - NewBeginOffset;
2845 V = getIntegerSplat(II.getValue(), Size);
2847 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2848 EndOffset != NewAllocaBeginOffset)) {
2849 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
2850 NewAI.getAlign(), "oldload");
2851 Old = convertValue(DL, IRB, Old, IntTy);
2852 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2853 V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2855 assert(V->getType() == IntTy &&
2856 "Wrong type for an alloca wide integer!");
2858 V = convertValue(DL, IRB, V, AllocaTy);
2860 // Established these invariants above.
2861 assert(NewBeginOffset == NewAllocaBeginOffset);
2862 assert(NewEndOffset == NewAllocaEndOffset);
2864 V = getIntegerSplat(II.getValue(),
2865 DL.getTypeSizeInBits(ScalarTy).getFixedSize() / 8);
2866 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2868 V, cast<FixedVectorType>(AllocaVecTy)->getNumElements());
2870 V = convertValue(DL, IRB, V, AllocaTy);
2874 IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlign(), II.isVolatile());
2876 New->setAAMetadata(AATags);
2877 LLVM_DEBUG(dbgs() << " to: " << *New << "\n");
2878 return !II.isVolatile();
2881 bool visitMemTransferInst(MemTransferInst &II) {
2882 // Rewriting of memory transfer instructions can be a bit tricky. We break
2883 // them into two categories: split intrinsics and unsplit intrinsics.
2885 LLVM_DEBUG(dbgs() << " original: " << II << "\n");
2888 II.getAAMetadata(AATags);
2890 bool IsDest = &II.getRawDestUse() == OldUse;
2891 assert((IsDest && II.getRawDest() == OldPtr) ||
2892 (!IsDest && II.getRawSource() == OldPtr));
2894 MaybeAlign SliceAlign = getSliceAlign();
2896 // For unsplit intrinsics, we simply modify the source and destination
2897 // pointers in place. This isn't just an optimization, it is a matter of
2898 // correctness. With unsplit intrinsics we may be dealing with transfers
2899 // within a single alloca before SROA ran, or with transfers that have
2900 // a variable length. We may also be dealing with memmove instead of
2901 // memcpy, and so simply updating the pointers is the necessary for us to
2902 // update both source and dest of a single call.
2903 if (!IsSplittable) {
2904 Value *AdjustedPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2906 II.setDest(AdjustedPtr);
2907 II.setDestAlignment(SliceAlign);
2910 II.setSource(AdjustedPtr);
2911 II.setSourceAlignment(SliceAlign);
2914 LLVM_DEBUG(dbgs() << " to: " << II << "\n");
2915 deleteIfTriviallyDead(OldPtr);
2918 // For split transfer intrinsics we have an incredibly useful assurance:
2919 // the source and destination do not reside within the same alloca, and at
2920 // least one of them does not escape. This means that we can replace
2921 // memmove with memcpy, and we don't need to worry about all manner of
2922 // downsides to splitting and transforming the operations.
2924 // If this doesn't map cleanly onto the alloca type, and that type isn't
2925 // a single value type, just emit a memcpy.
2928 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2930 DL.getTypeStoreSize(NewAI.getAllocatedType()).getFixedSize() ||
2931 !NewAI.getAllocatedType()->isSingleValueType());
2933 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2934 // size hasn't been shrunk based on analysis of the viable range, this is
2936 if (EmitMemCpy && &OldAI == &NewAI) {
2937 // Ensure the start lines up.
2938 assert(NewBeginOffset == BeginOffset);
2940 // Rewrite the size as needed.
2941 if (NewEndOffset != EndOffset)
2942 II.setLength(ConstantInt::get(II.getLength()->getType(),
2943 NewEndOffset - NewBeginOffset));
2946 // Record this instruction for deletion.
2947 Pass.DeadInsts.insert(&II);
2949 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2950 // alloca that should be re-examined after rewriting this instruction.
2951 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2952 if (AllocaInst *AI =
2953 dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
2954 assert(AI != &OldAI && AI != &NewAI &&
2955 "Splittable transfers cannot reach the same alloca on both ends.");
2956 Pass.Worklist.insert(AI);
2959 Type *OtherPtrTy = OtherPtr->getType();
2960 unsigned OtherAS = OtherPtrTy->getPointerAddressSpace();
2962 // Compute the relative offset for the other pointer within the transfer.
2963 unsigned OffsetWidth = DL.getIndexSizeInBits(OtherAS);
2964 APInt OtherOffset(OffsetWidth, NewBeginOffset - BeginOffset);
2966 (IsDest ? II.getSourceAlign() : II.getDestAlign()).valueOrOne();
2968 commonAlignment(OtherAlign, OtherOffset.zextOrTrunc(64).getZExtValue());
2971 // Compute the other pointer, folding as much as possible to produce
2972 // a single, simple GEP in most cases.
2973 OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2974 OtherPtr->getName() + ".");
2976 Value *OurPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2977 Type *SizeTy = II.getLength()->getType();
2978 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2980 Value *DestPtr, *SrcPtr;
2981 MaybeAlign DestAlign, SrcAlign;
2982 // Note: IsDest is true iff we're copying into the new alloca slice
2985 DestAlign = SliceAlign;
2987 SrcAlign = OtherAlign;
2990 DestAlign = OtherAlign;
2992 SrcAlign = SliceAlign;
2994 CallInst *New = IRB.CreateMemCpy(DestPtr, DestAlign, SrcPtr, SrcAlign,
2995 Size, II.isVolatile());
2997 New->setAAMetadata(AATags);
2998 LLVM_DEBUG(dbgs() << " to: " << *New << "\n");
3002 bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
3003 NewEndOffset == NewAllocaEndOffset;
3004 uint64_t Size = NewEndOffset - NewBeginOffset;
3005 unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
3006 unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
3007 unsigned NumElements = EndIndex - BeginIndex;
3008 IntegerType *SubIntTy =
3009 IntTy ? Type::getIntNTy(IntTy->getContext(), Size * 8) : nullptr;
3011 // Reset the other pointer type to match the register type we're going to
3012 // use, but using the address space of the original other pointer.
3014 if (VecTy && !IsWholeAlloca) {
3015 if (NumElements == 1)
3016 OtherTy = VecTy->getElementType();
3018 OtherTy = FixedVectorType::get(VecTy->getElementType(), NumElements);
3019 } else if (IntTy && !IsWholeAlloca) {
3022 OtherTy = NewAllocaTy;
3024 OtherPtrTy = OtherTy->getPointerTo(OtherAS);
3026 Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
3027 OtherPtr->getName() + ".");
3028 MaybeAlign SrcAlign = OtherAlign;
3029 Value *DstPtr = &NewAI;
3030 MaybeAlign DstAlign = SliceAlign;
3032 std::swap(SrcPtr, DstPtr);
3033 std::swap(SrcAlign, DstAlign);
3037 if (VecTy && !IsWholeAlloca && !IsDest) {
3038 Src = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
3039 NewAI.getAlign(), "load");
3040 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
3041 } else if (IntTy && !IsWholeAlloca && !IsDest) {
3042 Src = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
3043 NewAI.getAlign(), "load");
3044 Src = convertValue(DL, IRB, Src, IntTy);
3045 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
3046 Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
3048 LoadInst *Load = IRB.CreateAlignedLoad(OtherTy, SrcPtr, SrcAlign,
3049 II.isVolatile(), "copyload");
3051 Load->setAAMetadata(AATags);
3055 if (VecTy && !IsWholeAlloca && IsDest) {
3056 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
3057 NewAI.getAlign(), "oldload");
3058 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
3059 } else if (IntTy && !IsWholeAlloca && IsDest) {
3060 Value *Old = IRB.CreateAlignedLoad(NewAI.getAllocatedType(), &NewAI,
3061 NewAI.getAlign(), "oldload");
3062 Old = convertValue(DL, IRB, Old, IntTy);
3063 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
3064 Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
3065 Src = convertValue(DL, IRB, Src, NewAllocaTy);
3068 StoreInst *Store = cast<StoreInst>(
3069 IRB.CreateAlignedStore(Src, DstPtr, DstAlign, II.isVolatile()));
3071 Store->setAAMetadata(AATags);
3072 LLVM_DEBUG(dbgs() << " to: " << *Store << "\n");
3073 return !II.isVolatile();
3076 bool visitIntrinsicInst(IntrinsicInst &II) {
3077 assert(II.isLifetimeStartOrEnd());
3078 LLVM_DEBUG(dbgs() << " original: " << II << "\n");
3079 assert(II.getArgOperand(1) == OldPtr);
3081 // Record this instruction for deletion.
3082 Pass.DeadInsts.insert(&II);
3084 // Lifetime intrinsics are only promotable if they cover the whole alloca.
3085 // Therefore, we drop lifetime intrinsics which don't cover the whole
3087 // (In theory, intrinsics which partially cover an alloca could be
3088 // promoted, but PromoteMemToReg doesn't handle that case.)
3089 // FIXME: Check whether the alloca is promotable before dropping the
3090 // lifetime intrinsics?
3091 if (NewBeginOffset != NewAllocaBeginOffset ||
3092 NewEndOffset != NewAllocaEndOffset)
3096 ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
3097 NewEndOffset - NewBeginOffset);
3098 // Lifetime intrinsics always expect an i8* so directly get such a pointer
3099 // for the new alloca slice.
3100 Type *PointerTy = IRB.getInt8PtrTy(OldPtr->getType()->getPointerAddressSpace());
3101 Value *Ptr = getNewAllocaSlicePtr(IRB, PointerTy);
3103 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
3104 New = IRB.CreateLifetimeStart(Ptr, Size);
3106 New = IRB.CreateLifetimeEnd(Ptr, Size);
3109 LLVM_DEBUG(dbgs() << " to: " << *New << "\n");
3114 void fixLoadStoreAlign(Instruction &Root) {
3115 // This algorithm implements the same visitor loop as
3116 // hasUnsafePHIOrSelectUse, and fixes the alignment of each load
3118 SmallPtrSet<Instruction *, 4> Visited;
3119 SmallVector<Instruction *, 4> Uses;
3120 Visited.insert(&Root);
3121 Uses.push_back(&Root);
3123 Instruction *I = Uses.pop_back_val();
3125 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
3126 LI->setAlignment(std::min(LI->getAlign(), getSliceAlign()));
3129 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
3130 SI->setAlignment(std::min(SI->getAlign(), getSliceAlign()));
3134 assert(isa<BitCastInst>(I) || isa<AddrSpaceCastInst>(I) ||
3135 isa<PHINode>(I) || isa<SelectInst>(I) ||
3136 isa<GetElementPtrInst>(I));
3137 for (User *U : I->users())
3138 if (Visited.insert(cast<Instruction>(U)).second)
3139 Uses.push_back(cast<Instruction>(U));
3140 } while (!Uses.empty());
3143 bool visitPHINode(PHINode &PN) {
3144 LLVM_DEBUG(dbgs() << " original: " << PN << "\n");
3145 assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
3146 assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
3148 // We would like to compute a new pointer in only one place, but have it be
3149 // as local as possible to the PHI. To do that, we re-use the location of
3150 // the old pointer, which necessarily must be in the right position to
3151 // dominate the PHI.
3152 IRBuilderBase::InsertPointGuard Guard(IRB);
3153 if (isa<PHINode>(OldPtr))
3154 IRB.SetInsertPoint(&*OldPtr->getParent()->getFirstInsertionPt());
3156 IRB.SetInsertPoint(OldPtr);
3157 IRB.SetCurrentDebugLocation(OldPtr->getDebugLoc());
3159 Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
3160 // Replace the operands which were using the old pointer.
3161 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
3163 LLVM_DEBUG(dbgs() << " to: " << PN << "\n");
3164 deleteIfTriviallyDead(OldPtr);
3166 // Fix the alignment of any loads or stores using this PHI node.
3167 fixLoadStoreAlign(PN);
3169 // PHIs can't be promoted on their own, but often can be speculated. We
3170 // check the speculation outside of the rewriter so that we see the
3171 // fully-rewritten alloca.
3172 PHIUsers.insert(&PN);
3176 bool visitSelectInst(SelectInst &SI) {
3177 LLVM_DEBUG(dbgs() << " original: " << SI << "\n");
3178 assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
3179 "Pointer isn't an operand!");
3180 assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
3181 assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
3183 Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
3184 // Replace the operands which were using the old pointer.
3185 if (SI.getOperand(1) == OldPtr)
3186 SI.setOperand(1, NewPtr);
3187 if (SI.getOperand(2) == OldPtr)
3188 SI.setOperand(2, NewPtr);
3190 LLVM_DEBUG(dbgs() << " to: " << SI << "\n");
3191 deleteIfTriviallyDead(OldPtr);
3193 // Fix the alignment of any loads or stores using this select.
3194 fixLoadStoreAlign(SI);
3196 // Selects can't be promoted on their own, but often can be speculated. We
3197 // check the speculation outside of the rewriter so that we see the
3198 // fully-rewritten alloca.
3199 SelectUsers.insert(&SI);
3206 /// Visitor to rewrite aggregate loads and stores as scalar.
3208 /// This pass aggressively rewrites all aggregate loads and stores on
3209 /// a particular pointer (or any pointer derived from it which we can identify)
3210 /// with scalar loads and stores.
3211 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
3212 // Befriend the base class so it can delegate to private visit methods.
3213 friend class InstVisitor<AggLoadStoreRewriter, bool>;
3215 /// Queue of pointer uses to analyze and potentially rewrite.
3216 SmallVector<Use *, 8> Queue;
3218 /// Set to prevent us from cycling with phi nodes and loops.
3219 SmallPtrSet<User *, 8> Visited;
3221 /// The current pointer use being rewritten. This is used to dig up the used
3222 /// value (as opposed to the user).
3225 /// Used to calculate offsets, and hence alignment, of subobjects.
3226 const DataLayout &DL;
3229 AggLoadStoreRewriter(const DataLayout &DL) : DL(DL) {}
3231 /// Rewrite loads and stores through a pointer and all pointers derived from
3233 bool rewrite(Instruction &I) {
3234 LLVM_DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
3236 bool Changed = false;
3237 while (!Queue.empty()) {
3238 U = Queue.pop_back_val();
3239 Changed |= visit(cast<Instruction>(U->getUser()));
3245 /// Enqueue all the users of the given instruction for further processing.
3246 /// This uses a set to de-duplicate users.
3247 void enqueueUsers(Instruction &I) {
3248 for (Use &U : I.uses())
3249 if (Visited.insert(U.getUser()).second)
3250 Queue.push_back(&U);
3253 // Conservative default is to not rewrite anything.
3254 bool visitInstruction(Instruction &I) { return false; }
3256 /// Generic recursive split emission class.
3257 template <typename Derived> class OpSplitter {
3259 /// The builder used to form new instructions.
3262 /// The indices which to be used with insert- or extractvalue to select the
3263 /// appropriate value within the aggregate.
3264 SmallVector<unsigned, 4> Indices;
3266 /// The indices to a GEP instruction which will move Ptr to the correct slot
3267 /// within the aggregate.
3268 SmallVector<Value *, 4> GEPIndices;
3270 /// The base pointer of the original op, used as a base for GEPing the
3271 /// split operations.
3274 /// The base pointee type being GEPed into.
3277 /// Known alignment of the base pointer.
3280 /// To calculate offset of each component so we can correctly deduce
3282 const DataLayout &DL;
3284 /// Initialize the splitter with an insertion point, Ptr and start with a
3285 /// single zero GEP index.
3286 OpSplitter(Instruction *InsertionPoint, Value *Ptr, Type *BaseTy,
3287 Align BaseAlign, const DataLayout &DL)
3288 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr),
3289 BaseTy(BaseTy), BaseAlign(BaseAlign), DL(DL) {}
3292 /// Generic recursive split emission routine.
3294 /// This method recursively splits an aggregate op (load or store) into
3295 /// scalar or vector ops. It splits recursively until it hits a single value
3296 /// and emits that single value operation via the template argument.
3298 /// The logic of this routine relies on GEPs and insertvalue and
3299 /// extractvalue all operating with the same fundamental index list, merely
3300 /// formatted differently (GEPs need actual values).
3302 /// \param Ty The type being split recursively into smaller ops.
3303 /// \param Agg The aggregate value being built up or stored, depending on
3304 /// whether this is splitting a load or a store respectively.
3305 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
3306 if (Ty->isSingleValueType()) {
3307 unsigned Offset = DL.getIndexedOffsetInType(BaseTy, GEPIndices);
3308 return static_cast<Derived *>(this)->emitFunc(
3309 Ty, Agg, commonAlignment(BaseAlign, Offset), Name);
3312 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
3313 unsigned OldSize = Indices.size();
3315 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
3317 assert(Indices.size() == OldSize && "Did not return to the old size");
3318 Indices.push_back(Idx);
3319 GEPIndices.push_back(IRB.getInt32(Idx));
3320 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
3321 GEPIndices.pop_back();
3327 if (StructType *STy = dyn_cast<StructType>(Ty)) {
3328 unsigned OldSize = Indices.size();
3330 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
3332 assert(Indices.size() == OldSize && "Did not return to the old size");
3333 Indices.push_back(Idx);
3334 GEPIndices.push_back(IRB.getInt32(Idx));
3335 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
3336 GEPIndices.pop_back();
3342 llvm_unreachable("Only arrays and structs are aggregate loadable types");
3346 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
3349 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr, Type *BaseTy,
3350 AAMDNodes AATags, Align BaseAlign, const DataLayout &DL)
3351 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr, BaseTy, BaseAlign,
3355 /// Emit a leaf load of a single value. This is called at the leaves of the
3356 /// recursive emission to actually load values.
3357 void emitFunc(Type *Ty, Value *&Agg, Align Alignment, const Twine &Name) {
3358 assert(Ty->isSingleValueType());
3359 // Load the single value and insert it using the indices.
3361 IRB.CreateInBoundsGEP(BaseTy, Ptr, GEPIndices, Name + ".gep");
3363 IRB.CreateAlignedLoad(Ty, GEP, Alignment, Name + ".load");
3365 Load->setAAMetadata(AATags);
3366 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
3367 LLVM_DEBUG(dbgs() << " to: " << *Load << "\n");
3371 bool visitLoadInst(LoadInst &LI) {
3372 assert(LI.getPointerOperand() == *U);
3373 if (!LI.isSimple() || LI.getType()->isSingleValueType())
3376 // We have an aggregate being loaded, split it apart.
3377 LLVM_DEBUG(dbgs() << " original: " << LI << "\n");
3379 LI.getAAMetadata(AATags);
3380 LoadOpSplitter Splitter(&LI, *U, LI.getType(), AATags,
3381 getAdjustedAlignment(&LI, 0), DL);
3382 Value *V = UndefValue::get(LI.getType());
3383 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
3385 LI.replaceAllUsesWith(V);
3386 LI.eraseFromParent();
3390 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
3391 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr, Type *BaseTy,
3392 AAMDNodes AATags, Align BaseAlign, const DataLayout &DL)
3393 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr, BaseTy, BaseAlign,
3397 /// Emit a leaf store of a single value. This is called at the leaves of the
3398 /// recursive emission to actually produce stores.
3399 void emitFunc(Type *Ty, Value *&Agg, Align Alignment, const Twine &Name) {
3400 assert(Ty->isSingleValueType());
3401 // Extract the single value and store it using the indices.
3403 // The gep and extractvalue values are factored out of the CreateStore
3404 // call to make the output independent of the argument evaluation order.
3405 Value *ExtractValue =
3406 IRB.CreateExtractValue(Agg, Indices, Name + ".extract");
3407 Value *InBoundsGEP =
3408 IRB.CreateInBoundsGEP(BaseTy, Ptr, GEPIndices, Name + ".gep");
3410 IRB.CreateAlignedStore(ExtractValue, InBoundsGEP, Alignment);
3412 Store->setAAMetadata(AATags);
3413 LLVM_DEBUG(dbgs() << " to: " << *Store << "\n");
3417 bool visitStoreInst(StoreInst &SI) {
3418 if (!SI.isSimple() || SI.getPointerOperand() != *U)
3420 Value *V = SI.getValueOperand();
3421 if (V->getType()->isSingleValueType())
3424 // We have an aggregate being stored, split it apart.
3425 LLVM_DEBUG(dbgs() << " original: " << SI << "\n");
3427 SI.getAAMetadata(AATags);
3428 StoreOpSplitter Splitter(&SI, *U, V->getType(), AATags,
3429 getAdjustedAlignment(&SI, 0), DL);
3430 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
3432 SI.eraseFromParent();
3436 bool visitBitCastInst(BitCastInst &BC) {
3441 bool visitAddrSpaceCastInst(AddrSpaceCastInst &ASC) {
3446 // Fold gep (select cond, ptr1, ptr2) => select cond, gep(ptr1), gep(ptr2)
3447 bool foldGEPSelect(GetElementPtrInst &GEPI) {
3448 if (!GEPI.hasAllConstantIndices())
3451 SelectInst *Sel = cast<SelectInst>(GEPI.getPointerOperand());
3453 LLVM_DEBUG(dbgs() << " Rewriting gep(select) -> select(gep):"
3454 << "\n original: " << *Sel
3457 IRBuilderTy Builder(&GEPI);
3458 SmallVector<Value *, 4> Index(GEPI.idx_begin(), GEPI.idx_end());
3459 bool IsInBounds = GEPI.isInBounds();
3461 Value *True = Sel->getTrueValue();
3464 ? Builder.CreateInBoundsGEP(True, Index,
3465 True->getName() + ".sroa.gep")
3466 : Builder.CreateGEP(True, Index, True->getName() + ".sroa.gep");
3468 Value *False = Sel->getFalseValue();
3472 ? Builder.CreateInBoundsGEP(False, Index,
3473 False->getName() + ".sroa.gep")
3474 : Builder.CreateGEP(False, Index, False->getName() + ".sroa.gep");
3476 Value *NSel = Builder.CreateSelect(Sel->getCondition(), NTrue, NFalse,
3477 Sel->getName() + ".sroa.sel");
3478 Visited.erase(&GEPI);
3479 GEPI.replaceAllUsesWith(NSel);
3480 GEPI.eraseFromParent();
3481 Instruction *NSelI = cast<Instruction>(NSel);
3482 Visited.insert(NSelI);
3483 enqueueUsers(*NSelI);
3485 LLVM_DEBUG(dbgs() << "\n to: " << *NTrue
3487 << "\n " << *NSel << '\n');
3492 // Fold gep (phi ptr1, ptr2) => phi gep(ptr1), gep(ptr2)
3493 bool foldGEPPhi(GetElementPtrInst &GEPI) {
3494 if (!GEPI.hasAllConstantIndices())
3497 PHINode *PHI = cast<PHINode>(GEPI.getPointerOperand());
3498 if (GEPI.getParent() != PHI->getParent() ||
3499 llvm::any_of(PHI->incoming_values(), [](Value *In)
3500 { Instruction *I = dyn_cast<Instruction>(In);
3501 return !I || isa<GetElementPtrInst>(I) || isa<PHINode>(I) ||
3502 succ_empty(I->getParent()) ||
3503 !I->getParent()->isLegalToHoistInto();
3507 LLVM_DEBUG(dbgs() << " Rewriting gep(phi) -> phi(gep):"
3508 << "\n original: " << *PHI
3512 SmallVector<Value *, 4> Index(GEPI.idx_begin(), GEPI.idx_end());
3513 bool IsInBounds = GEPI.isInBounds();
3514 IRBuilderTy PHIBuilder(GEPI.getParent()->getFirstNonPHI());
3515 PHINode *NewPN = PHIBuilder.CreatePHI(GEPI.getType(),
3516 PHI->getNumIncomingValues(),
3517 PHI->getName() + ".sroa.phi");
3518 for (unsigned I = 0, E = PHI->getNumIncomingValues(); I != E; ++I) {
3519 Instruction *In = cast<Instruction>(PHI->getIncomingValue(I));
3521 IRBuilderTy B(In->getParent(), std::next(In->getIterator()));
3522 Value *NewVal = IsInBounds
3523 ? B.CreateInBoundsGEP(In, Index, In->getName() + ".sroa.gep")
3524 : B.CreateGEP(In, Index, In->getName() + ".sroa.gep");
3525 NewPN->addIncoming(NewVal, PHI->getIncomingBlock(I));
3528 Visited.erase(&GEPI);
3529 GEPI.replaceAllUsesWith(NewPN);
3530 GEPI.eraseFromParent();
3531 Visited.insert(NewPN);
3532 enqueueUsers(*NewPN);
3534 LLVM_DEBUG(for (Value *In : NewPN->incoming_values())
3535 dbgs() << "\n " << *In;
3536 dbgs() << "\n " << *NewPN << '\n');
3541 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
3542 if (isa<SelectInst>(GEPI.getPointerOperand()) &&
3543 foldGEPSelect(GEPI))
3546 if (isa<PHINode>(GEPI.getPointerOperand()) &&
3554 bool visitPHINode(PHINode &PN) {
3559 bool visitSelectInst(SelectInst &SI) {
3565 } // end anonymous namespace
3567 /// Strip aggregate type wrapping.
3569 /// This removes no-op aggregate types wrapping an underlying type. It will
3570 /// strip as many layers of types as it can without changing either the type
3571 /// size or the allocated size.
3572 static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
3573 if (Ty->isSingleValueType())
3576 uint64_t AllocSize = DL.getTypeAllocSize(Ty).getFixedSize();
3577 uint64_t TypeSize = DL.getTypeSizeInBits(Ty).getFixedSize();
3580 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
3581 InnerTy = ArrTy->getElementType();
3582 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
3583 const StructLayout *SL = DL.getStructLayout(STy);
3584 unsigned Index = SL->getElementContainingOffset(0);
3585 InnerTy = STy->getElementType(Index);
3590 if (AllocSize > DL.getTypeAllocSize(InnerTy).getFixedSize() ||
3591 TypeSize > DL.getTypeSizeInBits(InnerTy).getFixedSize())
3594 return stripAggregateTypeWrapping(DL, InnerTy);
3597 /// Try to find a partition of the aggregate type passed in for a given
3598 /// offset and size.
3600 /// This recurses through the aggregate type and tries to compute a subtype
3601 /// based on the offset and size. When the offset and size span a sub-section
3602 /// of an array, it will even compute a new array type for that sub-section,
3603 /// and the same for structs.
3605 /// Note that this routine is very strict and tries to find a partition of the
3606 /// type which produces the *exact* right offset and size. It is not forgiving
3607 /// when the size or offset cause either end of type-based partition to be off.
3608 /// Also, this is a best-effort routine. It is reasonable to give up and not
3609 /// return a type if necessary.
3610 static Type *getTypePartition(const DataLayout &DL, Type *Ty, uint64_t Offset,
3612 if (Offset == 0 && DL.getTypeAllocSize(Ty).getFixedSize() == Size)
3613 return stripAggregateTypeWrapping(DL, Ty);
3614 if (Offset > DL.getTypeAllocSize(Ty).getFixedSize() ||
3615 (DL.getTypeAllocSize(Ty).getFixedSize() - Offset) < Size)
3618 if (isa<ArrayType>(Ty) || isa<VectorType>(Ty)) {
3620 uint64_t TyNumElements;
3621 if (auto *AT = dyn_cast<ArrayType>(Ty)) {
3622 ElementTy = AT->getElementType();
3623 TyNumElements = AT->getNumElements();
3625 // FIXME: This isn't right for vectors with non-byte-sized or
3626 // non-power-of-two sized elements.
3627 auto *VT = cast<FixedVectorType>(Ty);
3628 ElementTy = VT->getElementType();
3629 TyNumElements = VT->getNumElements();
3631 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy).getFixedSize();
3632 uint64_t NumSkippedElements = Offset / ElementSize;
3633 if (NumSkippedElements >= TyNumElements)
3635 Offset -= NumSkippedElements * ElementSize;
3637 // First check if we need to recurse.
3638 if (Offset > 0 || Size < ElementSize) {
3639 // Bail if the partition ends in a different array element.
3640 if ((Offset + Size) > ElementSize)
3642 // Recurse through the element type trying to peel off offset bytes.
3643 return getTypePartition(DL, ElementTy, Offset, Size);
3645 assert(Offset == 0);
3647 if (Size == ElementSize)
3648 return stripAggregateTypeWrapping(DL, ElementTy);
3649 assert(Size > ElementSize);
3650 uint64_t NumElements = Size / ElementSize;
3651 if (NumElements * ElementSize != Size)
3653 return ArrayType::get(ElementTy, NumElements);
3656 StructType *STy = dyn_cast<StructType>(Ty);
3660 const StructLayout *SL = DL.getStructLayout(STy);
3661 if (Offset >= SL->getSizeInBytes())
3663 uint64_t EndOffset = Offset + Size;
3664 if (EndOffset > SL->getSizeInBytes())
3667 unsigned Index = SL->getElementContainingOffset(Offset);
3668 Offset -= SL->getElementOffset(Index);
3670 Type *ElementTy = STy->getElementType(Index);
3671 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy).getFixedSize();
3672 if (Offset >= ElementSize)
3673 return nullptr; // The offset points into alignment padding.
3675 // See if any partition must be contained by the element.
3676 if (Offset > 0 || Size < ElementSize) {
3677 if ((Offset + Size) > ElementSize)
3679 return getTypePartition(DL, ElementTy, Offset, Size);
3681 assert(Offset == 0);
3683 if (Size == ElementSize)
3684 return stripAggregateTypeWrapping(DL, ElementTy);
3686 StructType::element_iterator EI = STy->element_begin() + Index,
3687 EE = STy->element_end();
3688 if (EndOffset < SL->getSizeInBytes()) {
3689 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3690 if (Index == EndIndex)
3691 return nullptr; // Within a single element and its padding.
3693 // Don't try to form "natural" types if the elements don't line up with the
3695 // FIXME: We could potentially recurse down through the last element in the
3696 // sub-struct to find a natural end point.
3697 if (SL->getElementOffset(EndIndex) != EndOffset)
3700 assert(Index < EndIndex);
3701 EE = STy->element_begin() + EndIndex;
3704 // Try to build up a sub-structure.
3706 StructType::get(STy->getContext(), makeArrayRef(EI, EE), STy->isPacked());
3707 const StructLayout *SubSL = DL.getStructLayout(SubTy);
3708 if (Size != SubSL->getSizeInBytes())
3709 return nullptr; // The sub-struct doesn't have quite the size needed.
3714 /// Pre-split loads and stores to simplify rewriting.
3716 /// We want to break up the splittable load+store pairs as much as
3717 /// possible. This is important to do as a preprocessing step, as once we
3718 /// start rewriting the accesses to partitions of the alloca we lose the
3719 /// necessary information to correctly split apart paired loads and stores
3720 /// which both point into this alloca. The case to consider is something like
3723 /// %a = alloca [12 x i8]
3724 /// %gep1 = getelementptr [12 x i8]* %a, i32 0, i32 0
3725 /// %gep2 = getelementptr [12 x i8]* %a, i32 0, i32 4
3726 /// %gep3 = getelementptr [12 x i8]* %a, i32 0, i32 8
3727 /// %iptr1 = bitcast i8* %gep1 to i64*
3728 /// %iptr2 = bitcast i8* %gep2 to i64*
3729 /// %fptr1 = bitcast i8* %gep1 to float*
3730 /// %fptr2 = bitcast i8* %gep2 to float*
3731 /// %fptr3 = bitcast i8* %gep3 to float*
3732 /// store float 0.0, float* %fptr1
3733 /// store float 1.0, float* %fptr2
3734 /// %v = load i64* %iptr1
3735 /// store i64 %v, i64* %iptr2
3736 /// %f1 = load float* %fptr2
3737 /// %f2 = load float* %fptr3
3739 /// Here we want to form 3 partitions of the alloca, each 4 bytes large, and
3740 /// promote everything so we recover the 2 SSA values that should have been
3741 /// there all along.
3743 /// \returns true if any changes are made.
3744 bool SROA::presplitLoadsAndStores(AllocaInst &AI, AllocaSlices &AS) {
3745 LLVM_DEBUG(dbgs() << "Pre-splitting loads and stores\n");
3747 // Track the loads and stores which are candidates for pre-splitting here, in
3748 // the order they first appear during the partition scan. These give stable
3749 // iteration order and a basis for tracking which loads and stores we
3751 SmallVector<LoadInst *, 4> Loads;
3752 SmallVector<StoreInst *, 4> Stores;
3754 // We need to accumulate the splits required of each load or store where we
3755 // can find them via a direct lookup. This is important to cross-check loads
3756 // and stores against each other. We also track the slice so that we can kill
3757 // all the slices that end up split.
3758 struct SplitOffsets {
3760 std::vector<uint64_t> Splits;
3762 SmallDenseMap<Instruction *, SplitOffsets, 8> SplitOffsetsMap;
3764 // Track loads out of this alloca which cannot, for any reason, be pre-split.
3765 // This is important as we also cannot pre-split stores of those loads!
3766 // FIXME: This is all pretty gross. It means that we can be more aggressive
3767 // in pre-splitting when the load feeding the store happens to come from
3768 // a separate alloca. Put another way, the effectiveness of SROA would be
3769 // decreased by a frontend which just concatenated all of its local allocas
3770 // into one big flat alloca. But defeating such patterns is exactly the job
3771 // SROA is tasked with! Sadly, to not have this discrepancy we would have
3772 // change store pre-splitting to actually force pre-splitting of the load
3773 // that feeds it *and all stores*. That makes pre-splitting much harder, but
3774 // maybe it would make it more principled?
3775 SmallPtrSet<LoadInst *, 8> UnsplittableLoads;
3777 LLVM_DEBUG(dbgs() << " Searching for candidate loads and stores\n");
3778 for (auto &P : AS.partitions()) {
3779 for (Slice &S : P) {
3780 Instruction *I = cast<Instruction>(S.getUse()->getUser());
3781 if (!S.isSplittable() || S.endOffset() <= P.endOffset()) {
3782 // If this is a load we have to track that it can't participate in any
3783 // pre-splitting. If this is a store of a load we have to track that
3784 // that load also can't participate in any pre-splitting.
3785 if (auto *LI = dyn_cast<LoadInst>(I))
3786 UnsplittableLoads.insert(LI);
3787 else if (auto *SI = dyn_cast<StoreInst>(I))
3788 if (auto *LI = dyn_cast<LoadInst>(SI->getValueOperand()))
3789 UnsplittableLoads.insert(LI);
3792 assert(P.endOffset() > S.beginOffset() &&
3793 "Empty or backwards partition!");
3795 // Determine if this is a pre-splittable slice.
3796 if (auto *LI = dyn_cast<LoadInst>(I)) {
3797 assert(!LI->isVolatile() && "Cannot split volatile loads!");
3799 // The load must be used exclusively to store into other pointers for
3800 // us to be able to arbitrarily pre-split it. The stores must also be
3801 // simple to avoid changing semantics.
3802 auto IsLoadSimplyStored = [](LoadInst *LI) {
3803 for (User *LU : LI->users()) {
3804 auto *SI = dyn_cast<StoreInst>(LU);
3805 if (!SI || !SI->isSimple())
3810 if (!IsLoadSimplyStored(LI)) {
3811 UnsplittableLoads.insert(LI);
3815 Loads.push_back(LI);
3816 } else if (auto *SI = dyn_cast<StoreInst>(I)) {
3817 if (S.getUse() != &SI->getOperandUse(SI->getPointerOperandIndex()))
3818 // Skip stores *of* pointers. FIXME: This shouldn't even be possible!
3820 auto *StoredLoad = dyn_cast<LoadInst>(SI->getValueOperand());
3821 if (!StoredLoad || !StoredLoad->isSimple())
3823 assert(!SI->isVolatile() && "Cannot split volatile stores!");
3825 Stores.push_back(SI);
3827 // Other uses cannot be pre-split.
3831 // Record the initial split.
3832 LLVM_DEBUG(dbgs() << " Candidate: " << *I << "\n");
3833 auto &Offsets = SplitOffsetsMap[I];
3834 assert(Offsets.Splits.empty() &&
3835 "Should not have splits the first time we see an instruction!");
3837 Offsets.Splits.push_back(P.endOffset() - S.beginOffset());
3840 // Now scan the already split slices, and add a split for any of them which
3841 // we're going to pre-split.
3842 for (Slice *S : P.splitSliceTails()) {
3843 auto SplitOffsetsMapI =
3844 SplitOffsetsMap.find(cast<Instruction>(S->getUse()->getUser()));
3845 if (SplitOffsetsMapI == SplitOffsetsMap.end())
3847 auto &Offsets = SplitOffsetsMapI->second;
3849 assert(Offsets.S == S && "Found a mismatched slice!");
3850 assert(!Offsets.Splits.empty() &&
3851 "Cannot have an empty set of splits on the second partition!");
3852 assert(Offsets.Splits.back() ==
3853 P.beginOffset() - Offsets.S->beginOffset() &&
3854 "Previous split does not end where this one begins!");
3856 // Record each split. The last partition's end isn't needed as the size
3857 // of the slice dictates that.
3858 if (S->endOffset() > P.endOffset())
3859 Offsets.Splits.push_back(P.endOffset() - Offsets.S->beginOffset());
3863 // We may have split loads where some of their stores are split stores. For
3864 // such loads and stores, we can only pre-split them if their splits exactly
3865 // match relative to their starting offset. We have to verify this prior to
3868 llvm::remove_if(Stores,
3869 [&UnsplittableLoads, &SplitOffsetsMap](StoreInst *SI) {
3870 // Lookup the load we are storing in our map of split
3872 auto *LI = cast<LoadInst>(SI->getValueOperand());
3873 // If it was completely unsplittable, then we're done,
3874 // and this store can't be pre-split.
3875 if (UnsplittableLoads.count(LI))
3878 auto LoadOffsetsI = SplitOffsetsMap.find(LI);
3879 if (LoadOffsetsI == SplitOffsetsMap.end())
3880 return false; // Unrelated loads are definitely safe.
3881 auto &LoadOffsets = LoadOffsetsI->second;
3883 // Now lookup the store's offsets.
3884 auto &StoreOffsets = SplitOffsetsMap[SI];
3886 // If the relative offsets of each split in the load and
3887 // store match exactly, then we can split them and we
3888 // don't need to remove them here.
3889 if (LoadOffsets.Splits == StoreOffsets.Splits)
3894 << " Mismatched splits for load and store:\n"
3895 << " " << *LI << "\n"
3896 << " " << *SI << "\n");
3898 // We've found a store and load that we need to split
3899 // with mismatched relative splits. Just give up on them
3900 // and remove both instructions from our list of
3902 UnsplittableLoads.insert(LI);
3906 // Now we have to go *back* through all the stores, because a later store may
3907 // have caused an earlier store's load to become unsplittable and if it is
3908 // unsplittable for the later store, then we can't rely on it being split in
3909 // the earlier store either.
3910 Stores.erase(llvm::remove_if(Stores,
3911 [&UnsplittableLoads](StoreInst *SI) {
3913 cast<LoadInst>(SI->getValueOperand());
3914 return UnsplittableLoads.count(LI);
3917 // Once we've established all the loads that can't be split for some reason,
3918 // filter any that made it into our list out.
3919 Loads.erase(llvm::remove_if(Loads,
3920 [&UnsplittableLoads](LoadInst *LI) {
3921 return UnsplittableLoads.count(LI);
3925 // If no loads or stores are left, there is no pre-splitting to be done for
3927 if (Loads.empty() && Stores.empty())
3930 // From here on, we can't fail and will be building new accesses, so rig up
3932 IRBuilderTy IRB(&AI);
3934 // Collect the new slices which we will merge into the alloca slices.
3935 SmallVector<Slice, 4> NewSlices;
3937 // Track any allocas we end up splitting loads and stores for so we iterate
3939 SmallPtrSet<AllocaInst *, 4> ResplitPromotableAllocas;
3941 // At this point, we have collected all of the loads and stores we can
3942 // pre-split, and the specific splits needed for them. We actually do the
3943 // splitting in a specific order in order to handle when one of the loads in
3944 // the value operand to one of the stores.
3946 // First, we rewrite all of the split loads, and just accumulate each split
3947 // load in a parallel structure. We also build the slices for them and append
3948 // them to the alloca slices.
3949 SmallDenseMap<LoadInst *, std::vector<LoadInst *>, 1> SplitLoadsMap;
3950 std::vector<LoadInst *> SplitLoads;
3951 const DataLayout &DL = AI.getModule()->getDataLayout();
3952 for (LoadInst *LI : Loads) {
3955 IntegerType *Ty = cast<IntegerType>(LI->getType());
3956 uint64_t LoadSize = Ty->getBitWidth() / 8;
3957 assert(LoadSize > 0 && "Cannot have a zero-sized integer load!");
3959 auto &Offsets = SplitOffsetsMap[LI];
3960 assert(LoadSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&
3961 "Slice size should always match load size exactly!");
3962 uint64_t BaseOffset = Offsets.S->beginOffset();
3963 assert(BaseOffset + LoadSize > BaseOffset &&
3964 "Cannot represent alloca access size using 64-bit integers!");
3966 Instruction *BasePtr = cast<Instruction>(LI->getPointerOperand());
3967 IRB.SetInsertPoint(LI);
3969 LLVM_DEBUG(dbgs() << " Splitting load: " << *LI << "\n");
3971 uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
3972 int Idx = 0, Size = Offsets.Splits.size();
3974 auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
3975 auto AS = LI->getPointerAddressSpace();
3976 auto *PartPtrTy = PartTy->getPointerTo(AS);
3977 LoadInst *PLoad = IRB.CreateAlignedLoad(
3979 getAdjustedPtr(IRB, DL, BasePtr,
3980 APInt(DL.getIndexSizeInBits(AS), PartOffset),
3981 PartPtrTy, BasePtr->getName() + "."),
3982 getAdjustedAlignment(LI, PartOffset),
3983 /*IsVolatile*/ false, LI->getName());
3984 PLoad->copyMetadata(*LI, {LLVMContext::MD_mem_parallel_loop_access,
3985 LLVMContext::MD_access_group});
3987 // Append this load onto the list of split loads so we can find it later
3988 // to rewrite the stores.
3989 SplitLoads.push_back(PLoad);
3991 // Now build a new slice for the alloca.
3992 NewSlices.push_back(
3993 Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
3994 &PLoad->getOperandUse(PLoad->getPointerOperandIndex()),
3995 /*IsSplittable*/ false));
3996 LLVM_DEBUG(dbgs() << " new slice [" << NewSlices.back().beginOffset()
3997 << ", " << NewSlices.back().endOffset()
3998 << "): " << *PLoad << "\n");
4000 // See if we've handled all the splits.
4004 // Setup the next partition.
4005 PartOffset = Offsets.Splits[Idx];
4007 PartSize = (Idx < Size ? Offsets.Splits[Idx] : LoadSize) - PartOffset;
4010 // Now that we have the split loads, do the slow walk over all uses of the
4011 // load and rewrite them as split stores, or save the split loads to use
4012 // below if the store is going to be split there anyways.
4013 bool DeferredStores = false;
4014 for (User *LU : LI->users()) {
4015 StoreInst *SI = cast<StoreInst>(LU);
4016 if (!Stores.empty() && SplitOffsetsMap.count(SI)) {
4017 DeferredStores = true;
4018 LLVM_DEBUG(dbgs() << " Deferred splitting of store: " << *SI
4023 Value *StoreBasePtr = SI->getPointerOperand();
4024 IRB.SetInsertPoint(SI);
4026 LLVM_DEBUG(dbgs() << " Splitting store of load: " << *SI << "\n");
4028 for (int Idx = 0, Size = SplitLoads.size(); Idx < Size; ++Idx) {
4029 LoadInst *PLoad = SplitLoads[Idx];
4030 uint64_t PartOffset = Idx == 0 ? 0 : Offsets.Splits[Idx - 1];
4032 PLoad->getType()->getPointerTo(SI->getPointerAddressSpace());
4034 auto AS = SI->getPointerAddressSpace();
4035 StoreInst *PStore = IRB.CreateAlignedStore(
4037 getAdjustedPtr(IRB, DL, StoreBasePtr,
4038 APInt(DL.getIndexSizeInBits(AS), PartOffset),
4039 PartPtrTy, StoreBasePtr->getName() + "."),
4040 getAdjustedAlignment(SI, PartOffset),
4041 /*IsVolatile*/ false);
4042 PStore->copyMetadata(*LI, {LLVMContext::MD_mem_parallel_loop_access,
4043 LLVMContext::MD_access_group});
4044 LLVM_DEBUG(dbgs() << " +" << PartOffset << ":" << *PStore << "\n");
4047 // We want to immediately iterate on any allocas impacted by splitting
4048 // this store, and we have to track any promotable alloca (indicated by
4049 // a direct store) as needing to be resplit because it is no longer
4051 if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(StoreBasePtr)) {
4052 ResplitPromotableAllocas.insert(OtherAI);
4053 Worklist.insert(OtherAI);
4054 } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
4055 StoreBasePtr->stripInBoundsOffsets())) {
4056 Worklist.insert(OtherAI);
4059 // Mark the original store as dead.
4060 DeadInsts.insert(SI);
4063 // Save the split loads if there are deferred stores among the users.
4065 SplitLoadsMap.insert(std::make_pair(LI, std::move(SplitLoads)));
4067 // Mark the original load as dead and kill the original slice.
4068 DeadInsts.insert(LI);
4072 // Second, we rewrite all of the split stores. At this point, we know that
4073 // all loads from this alloca have been split already. For stores of such
4074 // loads, we can simply look up the pre-existing split loads. For stores of
4075 // other loads, we split those loads first and then write split stores of
4077 for (StoreInst *SI : Stores) {
4078 auto *LI = cast<LoadInst>(SI->getValueOperand());
4079 IntegerType *Ty = cast<IntegerType>(LI->getType());
4080 uint64_t StoreSize = Ty->getBitWidth() / 8;
4081 assert(StoreSize > 0 && "Cannot have a zero-sized integer store!");
4083 auto &Offsets = SplitOffsetsMap[SI];
4084 assert(StoreSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&
4085 "Slice size should always match load size exactly!");
4086 uint64_t BaseOffset = Offsets.S->beginOffset();
4087 assert(BaseOffset + StoreSize > BaseOffset &&
4088 "Cannot represent alloca access size using 64-bit integers!");
4090 Value *LoadBasePtr = LI->getPointerOperand();
4091 Instruction *StoreBasePtr = cast<Instruction>(SI->getPointerOperand());
4093 LLVM_DEBUG(dbgs() << " Splitting store: " << *SI << "\n");
4095 // Check whether we have an already split load.
4096 auto SplitLoadsMapI = SplitLoadsMap.find(LI);
4097 std::vector<LoadInst *> *SplitLoads = nullptr;
4098 if (SplitLoadsMapI != SplitLoadsMap.end()) {
4099 SplitLoads = &SplitLoadsMapI->second;
4100 assert(SplitLoads->size() == Offsets.Splits.size() + 1 &&
4101 "Too few split loads for the number of splits in the store!");
4103 LLVM_DEBUG(dbgs() << " of load: " << *LI << "\n");
4106 uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
4107 int Idx = 0, Size = Offsets.Splits.size();
4109 auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
4110 auto *LoadPartPtrTy = PartTy->getPointerTo(LI->getPointerAddressSpace());
4111 auto *StorePartPtrTy = PartTy->getPointerTo(SI->getPointerAddressSpace());
4113 // Either lookup a split load or create one.
4116 PLoad = (*SplitLoads)[Idx];
4118 IRB.SetInsertPoint(LI);
4119 auto AS = LI->getPointerAddressSpace();
4120 PLoad = IRB.CreateAlignedLoad(
4122 getAdjustedPtr(IRB, DL, LoadBasePtr,
4123 APInt(DL.getIndexSizeInBits(AS), PartOffset),
4124 LoadPartPtrTy, LoadBasePtr->getName() + "."),
4125 getAdjustedAlignment(LI, PartOffset),
4126 /*IsVolatile*/ false, LI->getName());
4129 // And store this partition.
4130 IRB.SetInsertPoint(SI);
4131 auto AS = SI->getPointerAddressSpace();
4132 StoreInst *PStore = IRB.CreateAlignedStore(
4134 getAdjustedPtr(IRB, DL, StoreBasePtr,
4135 APInt(DL.getIndexSizeInBits(AS), PartOffset),
4136 StorePartPtrTy, StoreBasePtr->getName() + "."),
4137 getAdjustedAlignment(SI, PartOffset),
4138 /*IsVolatile*/ false);
4140 // Now build a new slice for the alloca.
4141 NewSlices.push_back(
4142 Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
4143 &PStore->getOperandUse(PStore->getPointerOperandIndex()),
4144 /*IsSplittable*/ false));
4145 LLVM_DEBUG(dbgs() << " new slice [" << NewSlices.back().beginOffset()
4146 << ", " << NewSlices.back().endOffset()
4147 << "): " << *PStore << "\n");
4149 LLVM_DEBUG(dbgs() << " of split load: " << *PLoad << "\n");
4152 // See if we've finished all the splits.
4156 // Setup the next partition.
4157 PartOffset = Offsets.Splits[Idx];
4159 PartSize = (Idx < Size ? Offsets.Splits[Idx] : StoreSize) - PartOffset;
4162 // We want to immediately iterate on any allocas impacted by splitting
4163 // this load, which is only relevant if it isn't a load of this alloca and
4164 // thus we didn't already split the loads above. We also have to keep track
4165 // of any promotable allocas we split loads on as they can no longer be
4168 if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(LoadBasePtr)) {
4169 assert(OtherAI != &AI && "We can't re-split our own alloca!");
4170 ResplitPromotableAllocas.insert(OtherAI);
4171 Worklist.insert(OtherAI);
4172 } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
4173 LoadBasePtr->stripInBoundsOffsets())) {
4174 assert(OtherAI != &AI && "We can't re-split our own alloca!");
4175 Worklist.insert(OtherAI);
4179 // Mark the original store as dead now that we've split it up and kill its
4180 // slice. Note that we leave the original load in place unless this store
4181 // was its only use. It may in turn be split up if it is an alloca load
4182 // for some other alloca, but it may be a normal load. This may introduce
4183 // redundant loads, but where those can be merged the rest of the optimizer
4184 // should handle the merging, and this uncovers SSA splits which is more
4185 // important. In practice, the original loads will almost always be fully
4186 // split and removed eventually, and the splits will be merged by any
4187 // trivial CSE, including instcombine.
4188 if (LI->hasOneUse()) {
4189 assert(*LI->user_begin() == SI && "Single use isn't this store!");
4190 DeadInsts.insert(LI);
4192 DeadInsts.insert(SI);
4196 // Remove the killed slices that have ben pre-split.
4197 AS.erase(llvm::remove_if(AS, [](const Slice &S) { return S.isDead(); }),
4200 // Insert our new slices. This will sort and merge them into the sorted
4202 AS.insert(NewSlices);
4204 LLVM_DEBUG(dbgs() << " Pre-split slices:\n");
4206 for (auto I = AS.begin(), E = AS.end(); I != E; ++I)
4207 LLVM_DEBUG(AS.print(dbgs(), I, " "));
4210 // Finally, don't try to promote any allocas that new require re-splitting.
4211 // They have already been added to the worklist above.
4212 PromotableAllocas.erase(
4215 [&](AllocaInst *AI) { return ResplitPromotableAllocas.count(AI); }),
4216 PromotableAllocas.end());
4221 /// Rewrite an alloca partition's users.
4223 /// This routine drives both of the rewriting goals of the SROA pass. It tries
4224 /// to rewrite uses of an alloca partition to be conducive for SSA value
4225 /// promotion. If the partition needs a new, more refined alloca, this will
4226 /// build that new alloca, preserving as much type information as possible, and
4227 /// rewrite the uses of the old alloca to point at the new one and have the
4228 /// appropriate new offsets. It also evaluates how successful the rewrite was
4229 /// at enabling promotion and if it was successful queues the alloca to be
4231 AllocaInst *SROA::rewritePartition(AllocaInst &AI, AllocaSlices &AS,
4233 // Try to compute a friendly type for this partition of the alloca. This
4234 // won't always succeed, in which case we fall back to a legal integer type
4235 // or an i8 array of an appropriate size.
4236 Type *SliceTy = nullptr;
4237 const DataLayout &DL = AI.getModule()->getDataLayout();
4238 if (Type *CommonUseTy = findCommonType(P.begin(), P.end(), P.endOffset()))
4239 if (DL.getTypeAllocSize(CommonUseTy).getFixedSize() >= P.size())
4240 SliceTy = CommonUseTy;
4242 if (Type *TypePartitionTy = getTypePartition(DL, AI.getAllocatedType(),
4243 P.beginOffset(), P.size()))
4244 SliceTy = TypePartitionTy;
4245 if ((!SliceTy || (SliceTy->isArrayTy() &&
4246 SliceTy->getArrayElementType()->isIntegerTy())) &&
4247 DL.isLegalInteger(P.size() * 8))
4248 SliceTy = Type::getIntNTy(*C, P.size() * 8);
4250 SliceTy = ArrayType::get(Type::getInt8Ty(*C), P.size());
4251 assert(DL.getTypeAllocSize(SliceTy).getFixedSize() >= P.size());
4253 bool IsIntegerPromotable = isIntegerWideningViable(P, SliceTy, DL);
4256 IsIntegerPromotable ? nullptr : isVectorPromotionViable(P, DL);
4260 // Check for the case where we're going to rewrite to a new alloca of the
4261 // exact same type as the original, and with the same access offsets. In that
4262 // case, re-use the existing alloca, but still run through the rewriter to
4263 // perform phi and select speculation.
4264 // P.beginOffset() can be non-zero even with the same type in a case with
4265 // out-of-bounds access (e.g. @PR35657 function in SROA/basictest.ll).
4267 if (SliceTy == AI.getAllocatedType() && P.beginOffset() == 0) {
4269 // FIXME: We should be able to bail at this point with "nothing changed".
4270 // FIXME: We might want to defer PHI speculation until after here.
4271 // FIXME: return nullptr;
4273 // Make sure the alignment is compatible with P.beginOffset().
4274 const Align Alignment = commonAlignment(AI.getAlign(), P.beginOffset());
4275 // If we will get at least this much alignment from the type alone, leave
4276 // the alloca's alignment unconstrained.
4277 const bool IsUnconstrained = Alignment <= DL.getABITypeAlign(SliceTy);
4278 NewAI = new AllocaInst(
4279 SliceTy, AI.getType()->getAddressSpace(), nullptr,
4280 IsUnconstrained ? DL.getPrefTypeAlign(SliceTy) : Alignment,
4281 AI.getName() + ".sroa." + Twine(P.begin() - AS.begin()), &AI);
4282 // Copy the old AI debug location over to the new one.
4283 NewAI->setDebugLoc(AI.getDebugLoc());
4287 LLVM_DEBUG(dbgs() << "Rewriting alloca partition "
4288 << "[" << P.beginOffset() << "," << P.endOffset()
4289 << ") to: " << *NewAI << "\n");
4291 // Track the high watermark on the worklist as it is only relevant for
4292 // promoted allocas. We will reset it to this point if the alloca is not in
4293 // fact scheduled for promotion.
4294 unsigned PPWOldSize = PostPromotionWorklist.size();
4295 unsigned NumUses = 0;
4296 SmallSetVector<PHINode *, 8> PHIUsers;
4297 SmallSetVector<SelectInst *, 8> SelectUsers;
4299 AllocaSliceRewriter Rewriter(DL, AS, *this, AI, *NewAI, P.beginOffset(),
4300 P.endOffset(), IsIntegerPromotable, VecTy,
4301 PHIUsers, SelectUsers);
4302 bool Promotable = true;
4303 for (Slice *S : P.splitSliceTails()) {
4304 Promotable &= Rewriter.visit(S);
4307 for (Slice &S : P) {
4308 Promotable &= Rewriter.visit(&S);
4312 NumAllocaPartitionUses += NumUses;
4313 MaxUsesPerAllocaPartition.updateMax(NumUses);
4315 // Now that we've processed all the slices in the new partition, check if any
4316 // PHIs or Selects would block promotion.
4317 for (PHINode *PHI : PHIUsers)
4318 if (!isSafePHIToSpeculate(*PHI)) {
4321 SelectUsers.clear();
4325 for (SelectInst *Sel : SelectUsers)
4326 if (!isSafeSelectToSpeculate(*Sel)) {
4329 SelectUsers.clear();
4334 if (PHIUsers.empty() && SelectUsers.empty()) {
4335 // Promote the alloca.
4336 PromotableAllocas.push_back(NewAI);
4338 // If we have either PHIs or Selects to speculate, add them to those
4339 // worklists and re-queue the new alloca so that we promote in on the
4341 for (PHINode *PHIUser : PHIUsers)
4342 SpeculatablePHIs.insert(PHIUser);
4343 for (SelectInst *SelectUser : SelectUsers)
4344 SpeculatableSelects.insert(SelectUser);
4345 Worklist.insert(NewAI);
4348 // Drop any post-promotion work items if promotion didn't happen.
4349 while (PostPromotionWorklist.size() > PPWOldSize)
4350 PostPromotionWorklist.pop_back();
4352 // We couldn't promote and we didn't create a new partition, nothing
4357 // If we can't promote the alloca, iterate on it to check for new
4358 // refinements exposed by splitting the current alloca. Don't iterate on an
4359 // alloca which didn't actually change and didn't get promoted.
4360 Worklist.insert(NewAI);
4366 /// Walks the slices of an alloca and form partitions based on them,
4367 /// rewriting each of their uses.
4368 bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &AS) {
4369 if (AS.begin() == AS.end())
4372 unsigned NumPartitions = 0;
4373 bool Changed = false;
4374 const DataLayout &DL = AI.getModule()->getDataLayout();
4376 // First try to pre-split loads and stores.
4377 Changed |= presplitLoadsAndStores(AI, AS);
4379 // Now that we have identified any pre-splitting opportunities,
4380 // mark loads and stores unsplittable except for the following case.
4381 // We leave a slice splittable if all other slices are disjoint or fully
4382 // included in the slice, such as whole-alloca loads and stores.
4383 // If we fail to split these during pre-splitting, we want to force them
4384 // to be rewritten into a partition.
4385 bool IsSorted = true;
4387 uint64_t AllocaSize =
4388 DL.getTypeAllocSize(AI.getAllocatedType()).getFixedSize();
4389 const uint64_t MaxBitVectorSize = 1024;
4390 if (AllocaSize <= MaxBitVectorSize) {
4391 // If a byte boundary is included in any load or store, a slice starting or
4392 // ending at the boundary is not splittable.
4393 SmallBitVector SplittableOffset(AllocaSize + 1, true);
4395 for (unsigned O = S.beginOffset() + 1;
4396 O < S.endOffset() && O < AllocaSize; O++)
4397 SplittableOffset.reset(O);
4399 for (Slice &S : AS) {
4400 if (!S.isSplittable())
4403 if ((S.beginOffset() > AllocaSize || SplittableOffset[S.beginOffset()]) &&
4404 (S.endOffset() > AllocaSize || SplittableOffset[S.endOffset()]))
4407 if (isa<LoadInst>(S.getUse()->getUser()) ||
4408 isa<StoreInst>(S.getUse()->getUser())) {
4409 S.makeUnsplittable();
4415 // We only allow whole-alloca splittable loads and stores
4416 // for a large alloca to avoid creating too large BitVector.
4417 for (Slice &S : AS) {
4418 if (!S.isSplittable())
4421 if (S.beginOffset() == 0 && S.endOffset() >= AllocaSize)
4424 if (isa<LoadInst>(S.getUse()->getUser()) ||
4425 isa<StoreInst>(S.getUse()->getUser())) {
4426 S.makeUnsplittable();
4435 /// Describes the allocas introduced by rewritePartition in order to migrate
4441 Fragment(AllocaInst *AI, uint64_t O, uint64_t S)
4442 : Alloca(AI), Offset(O), Size(S) {}
4444 SmallVector<Fragment, 4> Fragments;
4446 // Rewrite each partition.
4447 for (auto &P : AS.partitions()) {
4448 if (AllocaInst *NewAI = rewritePartition(AI, AS, P)) {
4451 uint64_t SizeOfByte = 8;
4452 uint64_t AllocaSize =
4453 DL.getTypeSizeInBits(NewAI->getAllocatedType()).getFixedSize();
4454 // Don't include any padding.
4455 uint64_t Size = std::min(AllocaSize, P.size() * SizeOfByte);
4456 Fragments.push_back(Fragment(NewAI, P.beginOffset() * SizeOfByte, Size));
4462 NumAllocaPartitions += NumPartitions;
4463 MaxPartitionsPerAlloca.updateMax(NumPartitions);
4465 // Migrate debug information from the old alloca to the new alloca(s)
4466 // and the individual partitions.
4467 TinyPtrVector<DbgVariableIntrinsic *> DbgDeclares = FindDbgAddrUses(&AI);
4468 if (!DbgDeclares.empty()) {
4469 auto *Var = DbgDeclares.front()->getVariable();
4470 auto *Expr = DbgDeclares.front()->getExpression();
4471 auto VarSize = Var->getSizeInBits();
4472 DIBuilder DIB(*AI.getModule(), /*AllowUnresolved*/ false);
4473 uint64_t AllocaSize =
4474 DL.getTypeSizeInBits(AI.getAllocatedType()).getFixedSize();
4475 for (auto Fragment : Fragments) {
4476 // Create a fragment expression describing the new partition or reuse AI's
4477 // expression if there is only one partition.
4478 auto *FragmentExpr = Expr;
4479 if (Fragment.Size < AllocaSize || Expr->isFragment()) {
4480 // If this alloca is already a scalar replacement of a larger aggregate,
4481 // Fragment.Offset describes the offset inside the scalar.
4482 auto ExprFragment = Expr->getFragmentInfo();
4483 uint64_t Offset = ExprFragment ? ExprFragment->OffsetInBits : 0;
4484 uint64_t Start = Offset + Fragment.Offset;
4485 uint64_t Size = Fragment.Size;
4488 ExprFragment->OffsetInBits + ExprFragment->SizeInBits;
4489 if (Start >= AbsEnd)
4490 // No need to describe a SROAed padding.
4492 Size = std::min(Size, AbsEnd - Start);
4494 // The new, smaller fragment is stenciled out from the old fragment.
4495 if (auto OrigFragment = FragmentExpr->getFragmentInfo()) {
4496 assert(Start >= OrigFragment->OffsetInBits &&
4497 "new fragment is outside of original fragment");
4498 Start -= OrigFragment->OffsetInBits;
4501 // The alloca may be larger than the variable.
4503 if (Size > *VarSize)
4505 if (Size == 0 || Start + Size > *VarSize)
4509 // Avoid creating a fragment expression that covers the entire variable.
4510 if (!VarSize || *VarSize != Size) {
4512 DIExpression::createFragmentExpression(Expr, Start, Size))
4519 // Remove any existing intrinsics describing the same alloca.
4520 for (DbgVariableIntrinsic *OldDII : FindDbgAddrUses(Fragment.Alloca))
4521 OldDII->eraseFromParent();
4523 DIB.insertDeclare(Fragment.Alloca, Var, FragmentExpr,
4524 DbgDeclares.front()->getDebugLoc(), &AI);
4530 /// Clobber a use with undef, deleting the used value if it becomes dead.
4531 void SROA::clobberUse(Use &U) {
4533 // Replace the use with an undef value.
4534 U = UndefValue::get(OldV->getType());
4536 // Check for this making an instruction dead. We have to garbage collect
4537 // all the dead instructions to ensure the uses of any alloca end up being
4539 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
4540 if (isInstructionTriviallyDead(OldI)) {
4541 DeadInsts.insert(OldI);
4545 /// Analyze an alloca for SROA.
4547 /// This analyzes the alloca to ensure we can reason about it, builds
4548 /// the slices of the alloca, and then hands it off to be split and
4549 /// rewritten as needed.
4550 bool SROA::runOnAlloca(AllocaInst &AI) {
4551 LLVM_DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
4552 ++NumAllocasAnalyzed;
4554 // Special case dead allocas, as they're trivial.
4555 if (AI.use_empty()) {
4556 AI.eraseFromParent();
4559 const DataLayout &DL = AI.getModule()->getDataLayout();
4561 // Skip alloca forms that this analysis can't handle.
4562 auto *AT = AI.getAllocatedType();
4563 if (AI.isArrayAllocation() || !AT->isSized() || isa<ScalableVectorType>(AT) ||
4564 DL.getTypeAllocSize(AT).getFixedSize() == 0)
4567 bool Changed = false;
4569 // First, split any FCA loads and stores touching this alloca to promote
4570 // better splitting and promotion opportunities.
4571 AggLoadStoreRewriter AggRewriter(DL);
4572 Changed |= AggRewriter.rewrite(AI);
4574 // Build the slices using a recursive instruction-visiting builder.
4575 AllocaSlices AS(DL, AI);
4576 LLVM_DEBUG(AS.print(dbgs()));
4580 // Delete all the dead users of this alloca before splitting and rewriting it.
4581 for (Instruction *DeadUser : AS.getDeadUsers()) {
4582 // Free up everything used by this instruction.
4583 for (Use &DeadOp : DeadUser->operands())
4586 // Now replace the uses of this instruction.
4587 DeadUser->replaceAllUsesWith(UndefValue::get(DeadUser->getType()));
4589 // And mark it for deletion.
4590 DeadInsts.insert(DeadUser);
4593 for (Use *DeadOp : AS.getDeadOperands()) {
4594 clobberUse(*DeadOp);
4598 // No slices to split. Leave the dead alloca for a later pass to clean up.
4599 if (AS.begin() == AS.end())
4602 Changed |= splitAlloca(AI, AS);
4604 LLVM_DEBUG(dbgs() << " Speculating PHIs\n");
4605 while (!SpeculatablePHIs.empty())
4606 speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
4608 LLVM_DEBUG(dbgs() << " Speculating Selects\n");
4609 while (!SpeculatableSelects.empty())
4610 speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
4615 /// Delete the dead instructions accumulated in this run.
4617 /// Recursively deletes the dead instructions we've accumulated. This is done
4618 /// at the very end to maximize locality of the recursive delete and to
4619 /// minimize the problems of invalidated instruction pointers as such pointers
4620 /// are used heavily in the intermediate stages of the algorithm.
4622 /// We also record the alloca instructions deleted here so that they aren't
4623 /// subsequently handed to mem2reg to promote.
4624 bool SROA::deleteDeadInstructions(
4625 SmallPtrSetImpl<AllocaInst *> &DeletedAllocas) {
4626 bool Changed = false;
4627 while (!DeadInsts.empty()) {
4628 Instruction *I = DeadInsts.pop_back_val();
4629 LLVM_DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
4631 // If the instruction is an alloca, find the possible dbg.declare connected
4632 // to it, and remove it too. We must do this before calling RAUW or we will
4633 // not be able to find it.
4634 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
4635 DeletedAllocas.insert(AI);
4636 for (DbgVariableIntrinsic *OldDII : FindDbgAddrUses(AI))
4637 OldDII->eraseFromParent();
4640 I->replaceAllUsesWith(UndefValue::get(I->getType()));
4642 for (Use &Operand : I->operands())
4643 if (Instruction *U = dyn_cast<Instruction>(Operand)) {
4644 // Zero out the operand and see if it becomes trivially dead.
4646 if (isInstructionTriviallyDead(U))
4647 DeadInsts.insert(U);
4651 I->eraseFromParent();
4657 /// Promote the allocas, using the best available technique.
4659 /// This attempts to promote whatever allocas have been identified as viable in
4660 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
4661 /// This function returns whether any promotion occurred.
4662 bool SROA::promoteAllocas(Function &F) {
4663 if (PromotableAllocas.empty())
4666 NumPromoted += PromotableAllocas.size();
4668 LLVM_DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
4669 PromoteMemToReg(PromotableAllocas, *DT, AC);
4670 PromotableAllocas.clear();
4674 PreservedAnalyses SROA::runImpl(Function &F, DominatorTree &RunDT,
4675 AssumptionCache &RunAC) {
4676 LLVM_DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
4677 C = &F.getContext();
4681 BasicBlock &EntryBB = F.getEntryBlock();
4682 for (BasicBlock::iterator I = EntryBB.begin(), E = std::prev(EntryBB.end());
4684 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
4685 if (isa<ScalableVectorType>(AI->getAllocatedType())) {
4686 if (isAllocaPromotable(AI))
4687 PromotableAllocas.push_back(AI);
4689 Worklist.insert(AI);
4694 bool Changed = false;
4695 // A set of deleted alloca instruction pointers which should be removed from
4696 // the list of promotable allocas.
4697 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
4700 while (!Worklist.empty()) {
4701 Changed |= runOnAlloca(*Worklist.pop_back_val());
4702 Changed |= deleteDeadInstructions(DeletedAllocas);
4704 // Remove the deleted allocas from various lists so that we don't try to
4705 // continue processing them.
4706 if (!DeletedAllocas.empty()) {
4707 auto IsInSet = [&](AllocaInst *AI) { return DeletedAllocas.count(AI); };
4708 Worklist.remove_if(IsInSet);
4709 PostPromotionWorklist.remove_if(IsInSet);
4710 PromotableAllocas.erase(llvm::remove_if(PromotableAllocas, IsInSet),
4711 PromotableAllocas.end());
4712 DeletedAllocas.clear();
4716 Changed |= promoteAllocas(F);
4718 Worklist = PostPromotionWorklist;
4719 PostPromotionWorklist.clear();
4720 } while (!Worklist.empty());
4723 return PreservedAnalyses::all();
4725 PreservedAnalyses PA;
4726 PA.preserveSet<CFGAnalyses>();
4727 PA.preserve<GlobalsAA>();
4731 PreservedAnalyses SROA::run(Function &F, FunctionAnalysisManager &AM) {
4732 return runImpl(F, AM.getResult<DominatorTreeAnalysis>(F),
4733 AM.getResult<AssumptionAnalysis>(F));
4736 /// A legacy pass for the legacy pass manager that wraps the \c SROA pass.
4738 /// This is in the llvm namespace purely to allow it to be a friend of the \c
4740 class llvm::sroa::SROALegacyPass : public FunctionPass {
4741 /// The SROA implementation.
4747 SROALegacyPass() : FunctionPass(ID) {
4748 initializeSROALegacyPassPass(*PassRegistry::getPassRegistry());
4751 bool runOnFunction(Function &F) override {
4752 if (skipFunction(F))
4755 auto PA = Impl.runImpl(
4756 F, getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
4757 getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F));
4758 return !PA.areAllPreserved();
4761 void getAnalysisUsage(AnalysisUsage &AU) const override {
4762 AU.addRequired<AssumptionCacheTracker>();
4763 AU.addRequired<DominatorTreeWrapperPass>();
4764 AU.addPreserved<GlobalsAAWrapperPass>();
4765 AU.setPreservesCFG();
4768 StringRef getPassName() const override { return "SROA"; }
4771 char SROALegacyPass::ID = 0;
4773 FunctionPass *llvm::createSROAPass() { return new SROALegacyPass(); }
4775 INITIALIZE_PASS_BEGIN(SROALegacyPass, "sroa",
4776 "Scalar Replacement Of Aggregates", false, false)
4777 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
4778 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
4779 INITIALIZE_PASS_END(SROALegacyPass, "sroa", "Scalar Replacement Of Aggregates",