1 //===- SROA.cpp - Scalar Replacement Of Aggregates ------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 /// This transformation implements the well known scalar replacement of
11 /// aggregates transformation. It tries to identify promotable elements of an
12 /// aggregate alloca, and promote them to registers. It will also try to
13 /// convert uses of an element (or set of elements) of an alloca into a vector
14 /// or bitfield-style integer scalar if appropriate.
16 /// It works to do this with minimal slicing of the alloca so that regions
17 /// which are merely transferred in and out of external memory remain unchanged
18 /// and are not decomposed to scalar code.
20 /// Because this also performs alloca promotion, it can be thought of as also
21 /// serving the purpose of SSA formation. The algorithm iterates on the
22 /// function until all opportunities for promotion have been realized.
24 //===----------------------------------------------------------------------===//
26 #include "llvm/Transforms/Scalar/SROA.h"
27 #include "llvm/ADT/STLExtras.h"
28 #include "llvm/ADT/SmallVector.h"
29 #include "llvm/ADT/Statistic.h"
30 #include "llvm/Analysis/GlobalsModRef.h"
31 #include "llvm/Analysis/Loads.h"
32 #include "llvm/Analysis/PtrUseVisitor.h"
33 #include "llvm/Analysis/ValueTracking.h"
34 #include "llvm/IR/Constants.h"
35 #include "llvm/IR/DIBuilder.h"
36 #include "llvm/IR/DataLayout.h"
37 #include "llvm/IR/DebugInfo.h"
38 #include "llvm/IR/DerivedTypes.h"
39 #include "llvm/IR/IRBuilder.h"
40 #include "llvm/IR/InstVisitor.h"
41 #include "llvm/IR/Instructions.h"
42 #include "llvm/IR/IntrinsicInst.h"
43 #include "llvm/IR/LLVMContext.h"
44 #include "llvm/IR/Operator.h"
45 #include "llvm/Pass.h"
46 #include "llvm/Support/Chrono.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/Compiler.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/ErrorHandling.h"
51 #include "llvm/Support/MathExtras.h"
52 #include "llvm/Support/raw_ostream.h"
53 #include "llvm/Transforms/Scalar.h"
54 #include "llvm/Transforms/Utils/Local.h"
55 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
58 // We only use this for a debug check.
63 using namespace llvm::sroa;
65 #define DEBUG_TYPE "sroa"
67 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
68 STATISTIC(NumAllocaPartitions, "Number of alloca partitions formed");
69 STATISTIC(MaxPartitionsPerAlloca, "Maximum number of partitions per alloca");
70 STATISTIC(NumAllocaPartitionUses, "Number of alloca partition uses rewritten");
71 STATISTIC(MaxUsesPerAllocaPartition, "Maximum number of uses of a partition");
72 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
73 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
74 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
75 STATISTIC(NumDeleted, "Number of instructions deleted");
76 STATISTIC(NumVectorized, "Number of vectorized aggregates");
78 /// Hidden option to enable randomly shuffling the slices to help uncover
79 /// instability in their order.
80 static cl::opt<bool> SROARandomShuffleSlices("sroa-random-shuffle-slices",
81 cl::init(false), cl::Hidden);
83 /// Hidden option to experiment with completely strict handling of inbounds
85 static cl::opt<bool> SROAStrictInbounds("sroa-strict-inbounds", cl::init(false),
89 /// \brief A custom IRBuilder inserter which prefixes all names, but only in
91 class IRBuilderPrefixedInserter : public IRBuilderDefaultInserter {
93 const Twine getNameWithPrefix(const Twine &Name) const {
94 return Name.isTriviallyEmpty() ? Name : Prefix + Name;
98 void SetNamePrefix(const Twine &P) { Prefix = P.str(); }
101 void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
102 BasicBlock::iterator InsertPt) const {
103 IRBuilderDefaultInserter::InsertHelper(I, getNameWithPrefix(Name), BB,
108 /// \brief Provide a typedef for IRBuilder that drops names in release builds.
109 using IRBuilderTy = llvm::IRBuilder<ConstantFolder, IRBuilderPrefixedInserter>;
113 /// \brief A used slice of an alloca.
115 /// This structure represents a slice of an alloca used by some instruction. It
116 /// stores both the begin and end offsets of this use, a pointer to the use
117 /// itself, and a flag indicating whether we can classify the use as splittable
118 /// or not when forming partitions of the alloca.
120 /// \brief The beginning offset of the range.
121 uint64_t BeginOffset;
123 /// \brief The ending offset, not included in the range.
126 /// \brief Storage for both the use of this slice and whether it can be
128 PointerIntPair<Use *, 1, bool> UseAndIsSplittable;
131 Slice() : BeginOffset(), EndOffset() {}
132 Slice(uint64_t BeginOffset, uint64_t EndOffset, Use *U, bool IsSplittable)
133 : BeginOffset(BeginOffset), EndOffset(EndOffset),
134 UseAndIsSplittable(U, IsSplittable) {}
136 uint64_t beginOffset() const { return BeginOffset; }
137 uint64_t endOffset() const { return EndOffset; }
139 bool isSplittable() const { return UseAndIsSplittable.getInt(); }
140 void makeUnsplittable() { UseAndIsSplittable.setInt(false); }
142 Use *getUse() const { return UseAndIsSplittable.getPointer(); }
144 bool isDead() const { return getUse() == nullptr; }
145 void kill() { UseAndIsSplittable.setPointer(nullptr); }
147 /// \brief Support for ordering ranges.
149 /// This provides an ordering over ranges such that start offsets are
150 /// always increasing, and within equal start offsets, the end offsets are
151 /// decreasing. Thus the spanning range comes first in a cluster with the
152 /// same start position.
153 bool operator<(const Slice &RHS) const {
154 if (beginOffset() < RHS.beginOffset())
156 if (beginOffset() > RHS.beginOffset())
158 if (isSplittable() != RHS.isSplittable())
159 return !isSplittable();
160 if (endOffset() > RHS.endOffset())
165 /// \brief Support comparison with a single offset to allow binary searches.
166 friend LLVM_ATTRIBUTE_UNUSED bool operator<(const Slice &LHS,
167 uint64_t RHSOffset) {
168 return LHS.beginOffset() < RHSOffset;
170 friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
172 return LHSOffset < RHS.beginOffset();
175 bool operator==(const Slice &RHS) const {
176 return isSplittable() == RHS.isSplittable() &&
177 beginOffset() == RHS.beginOffset() && endOffset() == RHS.endOffset();
179 bool operator!=(const Slice &RHS) const { return !operator==(RHS); }
181 } // end anonymous namespace
184 template <typename T> struct isPodLike;
185 template <> struct isPodLike<Slice> { static const bool value = true; };
188 /// \brief Representation of the alloca slices.
190 /// This class represents the slices of an alloca which are formed by its
191 /// various uses. If a pointer escapes, we can't fully build a representation
192 /// for the slices used and we reflect that in this structure. The uses are
193 /// stored, sorted by increasing beginning offset and with unsplittable slices
194 /// starting at a particular offset before splittable slices.
195 class llvm::sroa::AllocaSlices {
197 /// \brief Construct the slices of a particular alloca.
198 AllocaSlices(const DataLayout &DL, AllocaInst &AI);
200 /// \brief Test whether a pointer to the allocation escapes our analysis.
202 /// If this is true, the slices are never fully built and should be
204 bool isEscaped() const { return PointerEscapingInstr; }
206 /// \brief Support for iterating over the slices.
208 typedef SmallVectorImpl<Slice>::iterator iterator;
209 typedef iterator_range<iterator> range;
210 iterator begin() { return Slices.begin(); }
211 iterator end() { return Slices.end(); }
213 typedef SmallVectorImpl<Slice>::const_iterator const_iterator;
214 typedef iterator_range<const_iterator> const_range;
215 const_iterator begin() const { return Slices.begin(); }
216 const_iterator end() const { return Slices.end(); }
219 /// \brief Erase a range of slices.
220 void erase(iterator Start, iterator Stop) { Slices.erase(Start, Stop); }
222 /// \brief Insert new slices for this alloca.
224 /// This moves the slices into the alloca's slices collection, and re-sorts
225 /// everything so that the usual ordering properties of the alloca's slices
227 void insert(ArrayRef<Slice> NewSlices) {
228 int OldSize = Slices.size();
229 Slices.append(NewSlices.begin(), NewSlices.end());
230 auto SliceI = Slices.begin() + OldSize;
231 std::sort(SliceI, Slices.end());
232 std::inplace_merge(Slices.begin(), SliceI, Slices.end());
235 // Forward declare the iterator and range accessor for walking the
237 class partition_iterator;
238 iterator_range<partition_iterator> partitions();
240 /// \brief Access the dead users for this alloca.
241 ArrayRef<Instruction *> getDeadUsers() const { return DeadUsers; }
243 /// \brief Access the dead operands referring to this alloca.
245 /// These are operands which have cannot actually be used to refer to the
246 /// alloca as they are outside its range and the user doesn't correct for
247 /// that. These mostly consist of PHI node inputs and the like which we just
248 /// need to replace with undef.
249 ArrayRef<Use *> getDeadOperands() const { return DeadOperands; }
251 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
252 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
253 void printSlice(raw_ostream &OS, const_iterator I,
254 StringRef Indent = " ") const;
255 void printUse(raw_ostream &OS, const_iterator I,
256 StringRef Indent = " ") const;
257 void print(raw_ostream &OS) const;
258 void dump(const_iterator I) const;
263 template <typename DerivedT, typename RetT = void> class BuilderBase;
265 friend class AllocaSlices::SliceBuilder;
267 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
268 /// \brief Handle to alloca instruction to simplify method interfaces.
272 /// \brief The instruction responsible for this alloca not having a known set
275 /// When an instruction (potentially) escapes the pointer to the alloca, we
276 /// store a pointer to that here and abort trying to form slices of the
277 /// alloca. This will be null if the alloca slices are analyzed successfully.
278 Instruction *PointerEscapingInstr;
280 /// \brief The slices of the alloca.
282 /// We store a vector of the slices formed by uses of the alloca here. This
283 /// vector is sorted by increasing begin offset, and then the unsplittable
284 /// slices before the splittable ones. See the Slice inner class for more
286 SmallVector<Slice, 8> Slices;
288 /// \brief Instructions which will become dead if we rewrite the alloca.
290 /// Note that these are not separated by slice. This is because we expect an
291 /// alloca to be completely rewritten or not rewritten at all. If rewritten,
292 /// all these instructions can simply be removed and replaced with undef as
293 /// they come from outside of the allocated space.
294 SmallVector<Instruction *, 8> DeadUsers;
296 /// \brief Operands which will become dead if we rewrite the alloca.
298 /// These are operands that in their particular use can be replaced with
299 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
300 /// to PHI nodes and the like. They aren't entirely dead (there might be
301 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
302 /// want to swap this particular input for undef to simplify the use lists of
304 SmallVector<Use *, 8> DeadOperands;
307 /// \brief A partition of the slices.
309 /// An ephemeral representation for a range of slices which can be viewed as
310 /// a partition of the alloca. This range represents a span of the alloca's
311 /// memory which cannot be split, and provides access to all of the slices
312 /// overlapping some part of the partition.
314 /// Objects of this type are produced by traversing the alloca's slices, but
315 /// are only ephemeral and not persistent.
316 class llvm::sroa::Partition {
318 friend class AllocaSlices;
319 friend class AllocaSlices::partition_iterator;
321 typedef AllocaSlices::iterator iterator;
323 /// \brief The beginning and ending offsets of the alloca for this
325 uint64_t BeginOffset, EndOffset;
327 /// \brief The start end end iterators of this partition.
330 /// \brief A collection of split slice tails overlapping the partition.
331 SmallVector<Slice *, 4> SplitTails;
333 /// \brief Raw constructor builds an empty partition starting and ending at
334 /// the given iterator.
335 Partition(iterator SI) : SI(SI), SJ(SI) {}
338 /// \brief The start offset of this partition.
340 /// All of the contained slices start at or after this offset.
341 uint64_t beginOffset() const { return BeginOffset; }
343 /// \brief The end offset of this partition.
345 /// All of the contained slices end at or before this offset.
346 uint64_t endOffset() const { return EndOffset; }
348 /// \brief The size of the partition.
350 /// Note that this can never be zero.
351 uint64_t size() const {
352 assert(BeginOffset < EndOffset && "Partitions must span some bytes!");
353 return EndOffset - BeginOffset;
356 /// \brief Test whether this partition contains no slices, and merely spans
357 /// a region occupied by split slices.
358 bool empty() const { return SI == SJ; }
360 /// \name Iterate slices that start within the partition.
361 /// These may be splittable or unsplittable. They have a begin offset >= the
362 /// partition begin offset.
364 // FIXME: We should probably define a "concat_iterator" helper and use that
365 // to stitch together pointee_iterators over the split tails and the
366 // contiguous iterators of the partition. That would give a much nicer
367 // interface here. We could then additionally expose filtered iterators for
368 // split, unsplit, and unsplittable splices based on the usage patterns.
369 iterator begin() const { return SI; }
370 iterator end() const { return SJ; }
373 /// \brief Get the sequence of split slice tails.
375 /// These tails are of slices which start before this partition but are
376 /// split and overlap into the partition. We accumulate these while forming
378 ArrayRef<Slice *> splitSliceTails() const { return SplitTails; }
381 /// \brief An iterator over partitions of the alloca's slices.
383 /// This iterator implements the core algorithm for partitioning the alloca's
384 /// slices. It is a forward iterator as we don't support backtracking for
385 /// efficiency reasons, and re-use a single storage area to maintain the
386 /// current set of split slices.
388 /// It is templated on the slice iterator type to use so that it can operate
389 /// with either const or non-const slice iterators.
390 class AllocaSlices::partition_iterator
391 : public iterator_facade_base<partition_iterator, std::forward_iterator_tag,
393 friend class AllocaSlices;
395 /// \brief Most of the state for walking the partitions is held in a class
396 /// with a nice interface for examining them.
399 /// \brief We need to keep the end of the slices to know when to stop.
400 AllocaSlices::iterator SE;
402 /// \brief We also need to keep track of the maximum split end offset seen.
403 /// FIXME: Do we really?
404 uint64_t MaxSplitSliceEndOffset;
406 /// \brief Sets the partition to be empty at given iterator, and sets the
408 partition_iterator(AllocaSlices::iterator SI, AllocaSlices::iterator SE)
409 : P(SI), SE(SE), MaxSplitSliceEndOffset(0) {
410 // If not already at the end, advance our state to form the initial
416 /// \brief Advance the iterator to the next partition.
418 /// Requires that the iterator not be at the end of the slices.
420 assert((P.SI != SE || !P.SplitTails.empty()) &&
421 "Cannot advance past the end of the slices!");
423 // Clear out any split uses which have ended.
424 if (!P.SplitTails.empty()) {
425 if (P.EndOffset >= MaxSplitSliceEndOffset) {
426 // If we've finished all splits, this is easy.
427 P.SplitTails.clear();
428 MaxSplitSliceEndOffset = 0;
430 // Remove the uses which have ended in the prior partition. This
431 // cannot change the max split slice end because we just checked that
432 // the prior partition ended prior to that max.
434 remove_if(P.SplitTails,
435 [&](Slice *S) { return S->endOffset() <= P.EndOffset; }),
437 assert(any_of(P.SplitTails,
439 return S->endOffset() == MaxSplitSliceEndOffset;
441 "Could not find the current max split slice offset!");
442 assert(all_of(P.SplitTails,
444 return S->endOffset() <= MaxSplitSliceEndOffset;
446 "Max split slice end offset is not actually the max!");
450 // If P.SI is already at the end, then we've cleared the split tail and
451 // now have an end iterator.
453 assert(P.SplitTails.empty() && "Failed to clear the split slices!");
457 // If we had a non-empty partition previously, set up the state for
458 // subsequent partitions.
460 // Accumulate all the splittable slices which started in the old
461 // partition into the split list.
463 if (S.isSplittable() && S.endOffset() > P.EndOffset) {
464 P.SplitTails.push_back(&S);
465 MaxSplitSliceEndOffset =
466 std::max(S.endOffset(), MaxSplitSliceEndOffset);
469 // Start from the end of the previous partition.
472 // If P.SI is now at the end, we at most have a tail of split slices.
474 P.BeginOffset = P.EndOffset;
475 P.EndOffset = MaxSplitSliceEndOffset;
479 // If the we have split slices and the next slice is after a gap and is
480 // not splittable immediately form an empty partition for the split
481 // slices up until the next slice begins.
482 if (!P.SplitTails.empty() && P.SI->beginOffset() != P.EndOffset &&
483 !P.SI->isSplittable()) {
484 P.BeginOffset = P.EndOffset;
485 P.EndOffset = P.SI->beginOffset();
490 // OK, we need to consume new slices. Set the end offset based on the
491 // current slice, and step SJ past it. The beginning offset of the
492 // partition is the beginning offset of the next slice unless we have
493 // pre-existing split slices that are continuing, in which case we begin
494 // at the prior end offset.
495 P.BeginOffset = P.SplitTails.empty() ? P.SI->beginOffset() : P.EndOffset;
496 P.EndOffset = P.SI->endOffset();
499 // There are two strategies to form a partition based on whether the
500 // partition starts with an unsplittable slice or a splittable slice.
501 if (!P.SI->isSplittable()) {
502 // When we're forming an unsplittable region, it must always start at
503 // the first slice and will extend through its end.
504 assert(P.BeginOffset == P.SI->beginOffset());
506 // Form a partition including all of the overlapping slices with this
507 // unsplittable slice.
508 while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
509 if (!P.SJ->isSplittable())
510 P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
514 // We have a partition across a set of overlapping unsplittable
519 // If we're starting with a splittable slice, then we need to form
520 // a synthetic partition spanning it and any other overlapping splittable
522 assert(P.SI->isSplittable() && "Forming a splittable partition!");
524 // Collect all of the overlapping splittable slices.
525 while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset &&
526 P.SJ->isSplittable()) {
527 P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
531 // Back upiP.EndOffset if we ended the span early when encountering an
532 // unsplittable slice. This synthesizes the early end offset of
533 // a partition spanning only splittable slices.
534 if (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
535 assert(!P.SJ->isSplittable());
536 P.EndOffset = P.SJ->beginOffset();
541 bool operator==(const partition_iterator &RHS) const {
542 assert(SE == RHS.SE &&
543 "End iterators don't match between compared partition iterators!");
545 // The observed positions of partitions is marked by the P.SI iterator and
546 // the emptiness of the split slices. The latter is only relevant when
547 // P.SI == SE, as the end iterator will additionally have an empty split
548 // slices list, but the prior may have the same P.SI and a tail of split
550 if (P.SI == RHS.P.SI && P.SplitTails.empty() == RHS.P.SplitTails.empty()) {
551 assert(P.SJ == RHS.P.SJ &&
552 "Same set of slices formed two different sized partitions!");
553 assert(P.SplitTails.size() == RHS.P.SplitTails.size() &&
554 "Same slice position with differently sized non-empty split "
561 partition_iterator &operator++() {
566 Partition &operator*() { return P; }
569 /// \brief A forward range over the partitions of the alloca's slices.
571 /// This accesses an iterator range over the partitions of the alloca's
572 /// slices. It computes these partitions on the fly based on the overlapping
573 /// offsets of the slices and the ability to split them. It will visit "empty"
574 /// partitions to cover regions of the alloca only accessed via split
576 iterator_range<AllocaSlices::partition_iterator> AllocaSlices::partitions() {
577 return make_range(partition_iterator(begin(), end()),
578 partition_iterator(end(), end()));
581 static Value *foldSelectInst(SelectInst &SI) {
582 // If the condition being selected on is a constant or the same value is
583 // being selected between, fold the select. Yes this does (rarely) happen
585 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
586 return SI.getOperand(1 + CI->isZero());
587 if (SI.getOperand(1) == SI.getOperand(2))
588 return SI.getOperand(1);
593 /// \brief A helper that folds a PHI node or a select.
594 static Value *foldPHINodeOrSelectInst(Instruction &I) {
595 if (PHINode *PN = dyn_cast<PHINode>(&I)) {
596 // If PN merges together the same value, return that value.
597 return PN->hasConstantValue();
599 return foldSelectInst(cast<SelectInst>(I));
602 /// \brief Builder for the alloca slices.
604 /// This class builds a set of alloca slices by recursively visiting the uses
605 /// of an alloca and making a slice for each load and store at each offset.
606 class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
607 friend class PtrUseVisitor<SliceBuilder>;
608 friend class InstVisitor<SliceBuilder>;
609 typedef PtrUseVisitor<SliceBuilder> Base;
611 const uint64_t AllocSize;
614 SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
615 SmallDenseMap<Instruction *, uint64_t> PHIOrSelectSizes;
617 /// \brief Set to de-duplicate dead instructions found in the use walk.
618 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
621 SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &AS)
622 : PtrUseVisitor<SliceBuilder>(DL),
623 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())), AS(AS) {}
626 void markAsDead(Instruction &I) {
627 if (VisitedDeadInsts.insert(&I).second)
628 AS.DeadUsers.push_back(&I);
631 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
632 bool IsSplittable = false) {
633 // Completely skip uses which have a zero size or start either before or
634 // past the end of the allocation.
635 if (Size == 0 || Offset.uge(AllocSize)) {
636 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
637 << " which has zero size or starts outside of the "
638 << AllocSize << " byte alloca:\n"
639 << " alloca: " << AS.AI << "\n"
640 << " use: " << I << "\n");
641 return markAsDead(I);
644 uint64_t BeginOffset = Offset.getZExtValue();
645 uint64_t EndOffset = BeginOffset + Size;
647 // Clamp the end offset to the end of the allocation. Note that this is
648 // formulated to handle even the case where "BeginOffset + Size" overflows.
649 // This may appear superficially to be something we could ignore entirely,
650 // but that is not so! There may be widened loads or PHI-node uses where
651 // some instructions are dead but not others. We can't completely ignore
652 // them, and so have to record at least the information here.
653 assert(AllocSize >= BeginOffset); // Established above.
654 if (Size > AllocSize - BeginOffset) {
655 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
656 << " to remain within the " << AllocSize << " byte alloca:\n"
657 << " alloca: " << AS.AI << "\n"
658 << " use: " << I << "\n");
659 EndOffset = AllocSize;
662 AS.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
665 void visitBitCastInst(BitCastInst &BC) {
667 return markAsDead(BC);
669 return Base::visitBitCastInst(BC);
672 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
673 if (GEPI.use_empty())
674 return markAsDead(GEPI);
676 if (SROAStrictInbounds && GEPI.isInBounds()) {
677 // FIXME: This is a manually un-factored variant of the basic code inside
678 // of GEPs with checking of the inbounds invariant specified in the
679 // langref in a very strict sense. If we ever want to enable
680 // SROAStrictInbounds, this code should be factored cleanly into
681 // PtrUseVisitor, but it is easier to experiment with SROAStrictInbounds
682 // by writing out the code here where we have the underlying allocation
683 // size readily available.
684 APInt GEPOffset = Offset;
685 const DataLayout &DL = GEPI.getModule()->getDataLayout();
686 for (gep_type_iterator GTI = gep_type_begin(GEPI),
687 GTE = gep_type_end(GEPI);
689 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
693 // Handle a struct index, which adds its field offset to the pointer.
694 if (StructType *STy = GTI.getStructTypeOrNull()) {
695 unsigned ElementIdx = OpC->getZExtValue();
696 const StructLayout *SL = DL.getStructLayout(STy);
698 APInt(Offset.getBitWidth(), SL->getElementOffset(ElementIdx));
700 // For array or vector indices, scale the index by the size of the
702 APInt Index = OpC->getValue().sextOrTrunc(Offset.getBitWidth());
703 GEPOffset += Index * APInt(Offset.getBitWidth(),
704 DL.getTypeAllocSize(GTI.getIndexedType()));
707 // If this index has computed an intermediate pointer which is not
708 // inbounds, then the result of the GEP is a poison value and we can
709 // delete it and all uses.
710 if (GEPOffset.ugt(AllocSize))
711 return markAsDead(GEPI);
715 return Base::visitGetElementPtrInst(GEPI);
718 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
719 uint64_t Size, bool IsVolatile) {
720 // We allow splitting of non-volatile loads and stores where the type is an
721 // integer type. These may be used to implement 'memcpy' or other "transfer
722 // of bits" patterns.
723 bool IsSplittable = Ty->isIntegerTy() && !IsVolatile;
725 insertUse(I, Offset, Size, IsSplittable);
728 void visitLoadInst(LoadInst &LI) {
729 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
730 "All simple FCA loads should have been pre-split");
733 return PI.setAborted(&LI);
735 const DataLayout &DL = LI.getModule()->getDataLayout();
736 uint64_t Size = DL.getTypeStoreSize(LI.getType());
737 return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
740 void visitStoreInst(StoreInst &SI) {
741 Value *ValOp = SI.getValueOperand();
743 return PI.setEscapedAndAborted(&SI);
745 return PI.setAborted(&SI);
747 const DataLayout &DL = SI.getModule()->getDataLayout();
748 uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
750 // If this memory access can be shown to *statically* extend outside the
751 // bounds of of the allocation, it's behavior is undefined, so simply
752 // ignore it. Note that this is more strict than the generic clamping
753 // behavior of insertUse. We also try to handle cases which might run the
755 // FIXME: We should instead consider the pointer to have escaped if this
756 // function is being instrumented for addressing bugs or race conditions.
757 if (Size > AllocSize || Offset.ugt(AllocSize - Size)) {
758 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @" << Offset
759 << " which extends past the end of the " << AllocSize
761 << " alloca: " << AS.AI << "\n"
762 << " use: " << SI << "\n");
763 return markAsDead(SI);
766 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
767 "All simple FCA stores should have been pre-split");
768 handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
771 void visitMemSetInst(MemSetInst &II) {
772 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
773 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
774 if ((Length && Length->getValue() == 0) ||
775 (IsOffsetKnown && Offset.uge(AllocSize)))
776 // Zero-length mem transfer intrinsics can be ignored entirely.
777 return markAsDead(II);
780 return PI.setAborted(&II);
782 insertUse(II, Offset, Length ? Length->getLimitedValue()
783 : AllocSize - Offset.getLimitedValue(),
787 void visitMemTransferInst(MemTransferInst &II) {
788 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
789 if (Length && Length->getValue() == 0)
790 // Zero-length mem transfer intrinsics can be ignored entirely.
791 return markAsDead(II);
793 // Because we can visit these intrinsics twice, also check to see if the
794 // first time marked this instruction as dead. If so, skip it.
795 if (VisitedDeadInsts.count(&II))
799 return PI.setAborted(&II);
801 // This side of the transfer is completely out-of-bounds, and so we can
802 // nuke the entire transfer. However, we also need to nuke the other side
803 // if already added to our partitions.
804 // FIXME: Yet another place we really should bypass this when
805 // instrumenting for ASan.
806 if (Offset.uge(AllocSize)) {
807 SmallDenseMap<Instruction *, unsigned>::iterator MTPI =
808 MemTransferSliceMap.find(&II);
809 if (MTPI != MemTransferSliceMap.end())
810 AS.Slices[MTPI->second].kill();
811 return markAsDead(II);
814 uint64_t RawOffset = Offset.getLimitedValue();
815 uint64_t Size = Length ? Length->getLimitedValue() : AllocSize - RawOffset;
817 // Check for the special case where the same exact value is used for both
819 if (*U == II.getRawDest() && *U == II.getRawSource()) {
820 // For non-volatile transfers this is a no-op.
821 if (!II.isVolatile())
822 return markAsDead(II);
824 return insertUse(II, Offset, Size, /*IsSplittable=*/false);
827 // If we have seen both source and destination for a mem transfer, then
828 // they both point to the same alloca.
830 SmallDenseMap<Instruction *, unsigned>::iterator MTPI;
831 std::tie(MTPI, Inserted) =
832 MemTransferSliceMap.insert(std::make_pair(&II, AS.Slices.size()));
833 unsigned PrevIdx = MTPI->second;
835 Slice &PrevP = AS.Slices[PrevIdx];
837 // Check if the begin offsets match and this is a non-volatile transfer.
838 // In that case, we can completely elide the transfer.
839 if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
841 return markAsDead(II);
844 // Otherwise we have an offset transfer within the same alloca. We can't
846 PrevP.makeUnsplittable();
849 // Insert the use now that we've fixed up the splittable nature.
850 insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
852 // Check that we ended up with a valid index in the map.
853 assert(AS.Slices[PrevIdx].getUse()->getUser() == &II &&
854 "Map index doesn't point back to a slice with this user.");
857 // Disable SRoA for any intrinsics except for lifetime invariants.
858 // FIXME: What about debug intrinsics? This matches old behavior, but
859 // doesn't make sense.
860 void visitIntrinsicInst(IntrinsicInst &II) {
862 return PI.setAborted(&II);
864 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
865 II.getIntrinsicID() == Intrinsic::lifetime_end) {
866 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
867 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
868 Length->getLimitedValue());
869 insertUse(II, Offset, Size, true);
873 Base::visitIntrinsicInst(II);
876 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
877 // We consider any PHI or select that results in a direct load or store of
878 // the same offset to be a viable use for slicing purposes. These uses
879 // are considered unsplittable and the size is the maximum loaded or stored
881 SmallPtrSet<Instruction *, 4> Visited;
882 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
883 Visited.insert(Root);
884 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
885 const DataLayout &DL = Root->getModule()->getDataLayout();
886 // If there are no loads or stores, the access is dead. We mark that as
887 // a size zero access.
890 Instruction *I, *UsedI;
891 std::tie(UsedI, I) = Uses.pop_back_val();
893 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
894 Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
897 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
898 Value *Op = SI->getOperand(0);
901 Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
905 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
906 if (!GEP->hasAllZeroIndices())
908 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
909 !isa<SelectInst>(I)) {
913 for (User *U : I->users())
914 if (Visited.insert(cast<Instruction>(U)).second)
915 Uses.push_back(std::make_pair(I, cast<Instruction>(U)));
916 } while (!Uses.empty());
921 void visitPHINodeOrSelectInst(Instruction &I) {
922 assert(isa<PHINode>(I) || isa<SelectInst>(I));
924 return markAsDead(I);
926 // TODO: We could use SimplifyInstruction here to fold PHINodes and
927 // SelectInsts. However, doing so requires to change the current
928 // dead-operand-tracking mechanism. For instance, suppose neither loading
929 // from %U nor %other traps. Then "load (select undef, %U, %other)" does not
930 // trap either. However, if we simply replace %U with undef using the
931 // current dead-operand-tracking mechanism, "load (select undef, undef,
932 // %other)" may trap because the select may return the first operand
934 if (Value *Result = foldPHINodeOrSelectInst(I)) {
936 // If the result of the constant fold will be the pointer, recurse
937 // through the PHI/select as if we had RAUW'ed it.
940 // Otherwise the operand to the PHI/select is dead, and we can replace
942 AS.DeadOperands.push_back(U);
948 return PI.setAborted(&I);
950 // See if we already have computed info on this node.
951 uint64_t &Size = PHIOrSelectSizes[&I];
953 // This is a new PHI/Select, check for an unsafe use of it.
954 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&I, Size))
955 return PI.setAborted(UnsafeI);
958 // For PHI and select operands outside the alloca, we can't nuke the entire
959 // phi or select -- the other side might still be relevant, so we special
960 // case them here and use a separate structure to track the operands
961 // themselves which should be replaced with undef.
962 // FIXME: This should instead be escaped in the event we're instrumenting
963 // for address sanitization.
964 if (Offset.uge(AllocSize)) {
965 AS.DeadOperands.push_back(U);
969 insertUse(I, Offset, Size);
972 void visitPHINode(PHINode &PN) { visitPHINodeOrSelectInst(PN); }
974 void visitSelectInst(SelectInst &SI) { visitPHINodeOrSelectInst(SI); }
976 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
977 void visitInstruction(Instruction &I) { PI.setAborted(&I); }
980 AllocaSlices::AllocaSlices(const DataLayout &DL, AllocaInst &AI)
982 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
985 PointerEscapingInstr(nullptr) {
986 SliceBuilder PB(DL, AI, *this);
987 SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
988 if (PtrI.isEscaped() || PtrI.isAborted()) {
989 // FIXME: We should sink the escape vs. abort info into the caller nicely,
990 // possibly by just storing the PtrInfo in the AllocaSlices.
991 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
992 : PtrI.getAbortingInst();
993 assert(PointerEscapingInstr && "Did not track a bad instruction");
997 Slices.erase(remove_if(Slices, [](const Slice &S) { return S.isDead(); }),
1001 if (SROARandomShuffleSlices) {
1002 std::mt19937 MT(static_cast<unsigned>(
1003 std::chrono::system_clock::now().time_since_epoch().count()));
1004 std::shuffle(Slices.begin(), Slices.end(), MT);
1008 // Sort the uses. This arranges for the offsets to be in ascending order,
1009 // and the sizes to be in descending order.
1010 std::sort(Slices.begin(), Slices.end());
1013 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1015 void AllocaSlices::print(raw_ostream &OS, const_iterator I,
1016 StringRef Indent) const {
1017 printSlice(OS, I, Indent);
1019 printUse(OS, I, Indent);
1022 void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
1023 StringRef Indent) const {
1024 OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
1025 << " slice #" << (I - begin())
1026 << (I->isSplittable() ? " (splittable)" : "");
1029 void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
1030 StringRef Indent) const {
1031 OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
1034 void AllocaSlices::print(raw_ostream &OS) const {
1035 if (PointerEscapingInstr) {
1036 OS << "Can't analyze slices for alloca: " << AI << "\n"
1037 << " A pointer to this alloca escaped by:\n"
1038 << " " << *PointerEscapingInstr << "\n";
1042 OS << "Slices of alloca: " << AI << "\n";
1043 for (const_iterator I = begin(), E = end(); I != E; ++I)
1047 LLVM_DUMP_METHOD void AllocaSlices::dump(const_iterator I) const {
1050 LLVM_DUMP_METHOD void AllocaSlices::dump() const { print(dbgs()); }
1052 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1054 /// Walk the range of a partitioning looking for a common type to cover this
1055 /// sequence of slices.
1056 static Type *findCommonType(AllocaSlices::const_iterator B,
1057 AllocaSlices::const_iterator E,
1058 uint64_t EndOffset) {
1060 bool TyIsCommon = true;
1061 IntegerType *ITy = nullptr;
1063 // Note that we need to look at *every* alloca slice's Use to ensure we
1064 // always get consistent results regardless of the order of slices.
1065 for (AllocaSlices::const_iterator I = B; I != E; ++I) {
1066 Use *U = I->getUse();
1067 if (isa<IntrinsicInst>(*U->getUser()))
1069 if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
1072 Type *UserTy = nullptr;
1073 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1074 UserTy = LI->getType();
1075 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1076 UserTy = SI->getValueOperand()->getType();
1079 if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
1080 // If the type is larger than the partition, skip it. We only encounter
1081 // this for split integer operations where we want to use the type of the
1082 // entity causing the split. Also skip if the type is not a byte width
1084 if (UserITy->getBitWidth() % 8 != 0 ||
1085 UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
1088 // Track the largest bitwidth integer type used in this way in case there
1089 // is no common type.
1090 if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
1094 // To avoid depending on the order of slices, Ty and TyIsCommon must not
1095 // depend on types skipped above.
1096 if (!UserTy || (Ty && Ty != UserTy))
1097 TyIsCommon = false; // Give up on anything but an iN type.
1102 return TyIsCommon ? Ty : ITy;
1105 /// PHI instructions that use an alloca and are subsequently loaded can be
1106 /// rewritten to load both input pointers in the pred blocks and then PHI the
1107 /// results, allowing the load of the alloca to be promoted.
1109 /// %P2 = phi [i32* %Alloca, i32* %Other]
1110 /// %V = load i32* %P2
1112 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1114 /// %V2 = load i32* %Other
1116 /// %V = phi [i32 %V1, i32 %V2]
1118 /// We can do this to a select if its only uses are loads and if the operands
1119 /// to the select can be loaded unconditionally.
1121 /// FIXME: This should be hoisted into a generic utility, likely in
1122 /// Transforms/Util/Local.h
1123 static bool isSafePHIToSpeculate(PHINode &PN) {
1124 // For now, we can only do this promotion if the load is in the same block
1125 // as the PHI, and if there are no stores between the phi and load.
1126 // TODO: Allow recursive phi users.
1127 // TODO: Allow stores.
1128 BasicBlock *BB = PN.getParent();
1129 unsigned MaxAlign = 0;
1130 bool HaveLoad = false;
1131 for (User *U : PN.users()) {
1132 LoadInst *LI = dyn_cast<LoadInst>(U);
1133 if (!LI || !LI->isSimple())
1136 // For now we only allow loads in the same block as the PHI. This is
1137 // a common case that happens when instcombine merges two loads through
1139 if (LI->getParent() != BB)
1142 // Ensure that there are no instructions between the PHI and the load that
1144 for (BasicBlock::iterator BBI(PN); &*BBI != LI; ++BBI)
1145 if (BBI->mayWriteToMemory())
1148 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1155 const DataLayout &DL = PN.getModule()->getDataLayout();
1157 // We can only transform this if it is safe to push the loads into the
1158 // predecessor blocks. The only thing to watch out for is that we can't put
1159 // a possibly trapping load in the predecessor if it is a critical edge.
1160 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1161 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1162 Value *InVal = PN.getIncomingValue(Idx);
1164 // If the value is produced by the terminator of the predecessor (an
1165 // invoke) or it has side-effects, there is no valid place to put a load
1166 // in the predecessor.
1167 if (TI == InVal || TI->mayHaveSideEffects())
1170 // If the predecessor has a single successor, then the edge isn't
1172 if (TI->getNumSuccessors() == 1)
1175 // If this pointer is always safe to load, or if we can prove that there
1176 // is already a load in the block, then we can move the load to the pred
1178 if (isSafeToLoadUnconditionally(InVal, MaxAlign, DL, TI))
1187 static void speculatePHINodeLoads(PHINode &PN) {
1188 DEBUG(dbgs() << " original: " << PN << "\n");
1190 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1191 IRBuilderTy PHIBuilder(&PN);
1192 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1193 PN.getName() + ".sroa.speculated");
1195 // Get the AA tags and alignment to use from one of the loads. It doesn't
1196 // matter which one we get and if any differ.
1197 LoadInst *SomeLoad = cast<LoadInst>(PN.user_back());
1200 SomeLoad->getAAMetadata(AATags);
1201 unsigned Align = SomeLoad->getAlignment();
1203 // Rewrite all loads of the PN to use the new PHI.
1204 while (!PN.use_empty()) {
1205 LoadInst *LI = cast<LoadInst>(PN.user_back());
1206 LI->replaceAllUsesWith(NewPN);
1207 LI->eraseFromParent();
1210 // Inject loads into all of the pred blocks.
1211 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1212 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1213 TerminatorInst *TI = Pred->getTerminator();
1214 Value *InVal = PN.getIncomingValue(Idx);
1215 IRBuilderTy PredBuilder(TI);
1217 LoadInst *Load = PredBuilder.CreateLoad(
1218 InVal, (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1219 ++NumLoadsSpeculated;
1220 Load->setAlignment(Align);
1222 Load->setAAMetadata(AATags);
1223 NewPN->addIncoming(Load, Pred);
1226 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1227 PN.eraseFromParent();
1230 /// Select instructions that use an alloca and are subsequently loaded can be
1231 /// rewritten to load both input pointers and then select between the result,
1232 /// allowing the load of the alloca to be promoted.
1234 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1235 /// %V = load i32* %P2
1237 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1238 /// %V2 = load i32* %Other
1239 /// %V = select i1 %cond, i32 %V1, i32 %V2
1241 /// We can do this to a select if its only uses are loads and if the operand
1242 /// to the select can be loaded unconditionally.
1243 static bool isSafeSelectToSpeculate(SelectInst &SI) {
1244 Value *TValue = SI.getTrueValue();
1245 Value *FValue = SI.getFalseValue();
1246 const DataLayout &DL = SI.getModule()->getDataLayout();
1248 for (User *U : SI.users()) {
1249 LoadInst *LI = dyn_cast<LoadInst>(U);
1250 if (!LI || !LI->isSimple())
1253 // Both operands to the select need to be dereferencable, either
1254 // absolutely (e.g. allocas) or at this point because we can see other
1256 if (!isSafeToLoadUnconditionally(TValue, LI->getAlignment(), DL, LI))
1258 if (!isSafeToLoadUnconditionally(FValue, LI->getAlignment(), DL, LI))
1265 static void speculateSelectInstLoads(SelectInst &SI) {
1266 DEBUG(dbgs() << " original: " << SI << "\n");
1268 IRBuilderTy IRB(&SI);
1269 Value *TV = SI.getTrueValue();
1270 Value *FV = SI.getFalseValue();
1271 // Replace the loads of the select with a select of two loads.
1272 while (!SI.use_empty()) {
1273 LoadInst *LI = cast<LoadInst>(SI.user_back());
1274 assert(LI->isSimple() && "We only speculate simple loads");
1276 IRB.SetInsertPoint(LI);
1278 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1280 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1281 NumLoadsSpeculated += 2;
1283 // Transfer alignment and AA info if present.
1284 TL->setAlignment(LI->getAlignment());
1285 FL->setAlignment(LI->getAlignment());
1288 LI->getAAMetadata(Tags);
1290 TL->setAAMetadata(Tags);
1291 FL->setAAMetadata(Tags);
1294 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1295 LI->getName() + ".sroa.speculated");
1297 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1298 LI->replaceAllUsesWith(V);
1299 LI->eraseFromParent();
1301 SI.eraseFromParent();
1304 /// \brief Build a GEP out of a base pointer and indices.
1306 /// This will return the BasePtr if that is valid, or build a new GEP
1307 /// instruction using the IRBuilder if GEP-ing is needed.
1308 static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1309 SmallVectorImpl<Value *> &Indices, Twine NamePrefix) {
1310 if (Indices.empty())
1313 // A single zero index is a no-op, so check for this and avoid building a GEP
1315 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1318 return IRB.CreateInBoundsGEP(nullptr, BasePtr, Indices,
1319 NamePrefix + "sroa_idx");
1322 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1323 /// TargetTy without changing the offset of the pointer.
1325 /// This routine assumes we've already established a properly offset GEP with
1326 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1327 /// zero-indices down through type layers until we find one the same as
1328 /// TargetTy. If we can't find one with the same type, we at least try to use
1329 /// one with the same size. If none of that works, we just produce the GEP as
1330 /// indicated by Indices to have the correct offset.
1331 static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1332 Value *BasePtr, Type *Ty, Type *TargetTy,
1333 SmallVectorImpl<Value *> &Indices,
1336 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1338 // Pointer size to use for the indices.
1339 unsigned PtrSize = DL.getPointerTypeSizeInBits(BasePtr->getType());
1341 // See if we can descend into a struct and locate a field with the correct
1343 unsigned NumLayers = 0;
1344 Type *ElementTy = Ty;
1346 if (ElementTy->isPointerTy())
1349 if (ArrayType *ArrayTy = dyn_cast<ArrayType>(ElementTy)) {
1350 ElementTy = ArrayTy->getElementType();
1351 Indices.push_back(IRB.getIntN(PtrSize, 0));
1352 } else if (VectorType *VectorTy = dyn_cast<VectorType>(ElementTy)) {
1353 ElementTy = VectorTy->getElementType();
1354 Indices.push_back(IRB.getInt32(0));
1355 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1356 if (STy->element_begin() == STy->element_end())
1357 break; // Nothing left to descend into.
1358 ElementTy = *STy->element_begin();
1359 Indices.push_back(IRB.getInt32(0));
1364 } while (ElementTy != TargetTy);
1365 if (ElementTy != TargetTy)
1366 Indices.erase(Indices.end() - NumLayers, Indices.end());
1368 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1371 /// \brief Recursively compute indices for a natural GEP.
1373 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1374 /// element types adding appropriate indices for the GEP.
1375 static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1376 Value *Ptr, Type *Ty, APInt &Offset,
1378 SmallVectorImpl<Value *> &Indices,
1381 return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices,
1384 // We can't recurse through pointer types.
1385 if (Ty->isPointerTy())
1388 // We try to analyze GEPs over vectors here, but note that these GEPs are
1389 // extremely poorly defined currently. The long-term goal is to remove GEPing
1390 // over a vector from the IR completely.
1391 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1392 unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
1393 if (ElementSizeInBits % 8 != 0) {
1394 // GEPs over non-multiple of 8 size vector elements are invalid.
1397 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1398 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1399 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1401 Offset -= NumSkippedElements * ElementSize;
1402 Indices.push_back(IRB.getInt(NumSkippedElements));
1403 return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1404 Offset, TargetTy, Indices, NamePrefix);
1407 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1408 Type *ElementTy = ArrTy->getElementType();
1409 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1410 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1411 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1414 Offset -= NumSkippedElements * ElementSize;
1415 Indices.push_back(IRB.getInt(NumSkippedElements));
1416 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1417 Indices, NamePrefix);
1420 StructType *STy = dyn_cast<StructType>(Ty);
1424 const StructLayout *SL = DL.getStructLayout(STy);
1425 uint64_t StructOffset = Offset.getZExtValue();
1426 if (StructOffset >= SL->getSizeInBytes())
1428 unsigned Index = SL->getElementContainingOffset(StructOffset);
1429 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1430 Type *ElementTy = STy->getElementType(Index);
1431 if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
1432 return nullptr; // The offset points into alignment padding.
1434 Indices.push_back(IRB.getInt32(Index));
1435 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1436 Indices, NamePrefix);
1439 /// \brief Get a natural GEP from a base pointer to a particular offset and
1440 /// resulting in a particular type.
1442 /// The goal is to produce a "natural" looking GEP that works with the existing
1443 /// composite types to arrive at the appropriate offset and element type for
1444 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1445 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1446 /// Indices, and setting Ty to the result subtype.
1448 /// If no natural GEP can be constructed, this function returns null.
1449 static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1450 Value *Ptr, APInt Offset, Type *TargetTy,
1451 SmallVectorImpl<Value *> &Indices,
1453 PointerType *Ty = cast<PointerType>(Ptr->getType());
1455 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1457 if (Ty == IRB.getInt8PtrTy(Ty->getAddressSpace()) && TargetTy->isIntegerTy(8))
1460 Type *ElementTy = Ty->getElementType();
1461 if (!ElementTy->isSized())
1462 return nullptr; // We can't GEP through an unsized element.
1463 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1464 if (ElementSize == 0)
1465 return nullptr; // Zero-length arrays can't help us build a natural GEP.
1466 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1468 Offset -= NumSkippedElements * ElementSize;
1469 Indices.push_back(IRB.getInt(NumSkippedElements));
1470 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1471 Indices, NamePrefix);
1474 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1475 /// resulting pointer has PointerTy.
1477 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1478 /// and produces the pointer type desired. Where it cannot, it will try to use
1479 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1480 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1481 /// bitcast to the type.
1483 /// The strategy for finding the more natural GEPs is to peel off layers of the
1484 /// pointer, walking back through bit casts and GEPs, searching for a base
1485 /// pointer from which we can compute a natural GEP with the desired
1486 /// properties. The algorithm tries to fold as many constant indices into
1487 /// a single GEP as possible, thus making each GEP more independent of the
1488 /// surrounding code.
1489 static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr,
1490 APInt Offset, Type *PointerTy, Twine NamePrefix) {
1491 // Even though we don't look through PHI nodes, we could be called on an
1492 // instruction in an unreachable block, which may be on a cycle.
1493 SmallPtrSet<Value *, 4> Visited;
1494 Visited.insert(Ptr);
1495 SmallVector<Value *, 4> Indices;
1497 // We may end up computing an offset pointer that has the wrong type. If we
1498 // never are able to compute one directly that has the correct type, we'll
1499 // fall back to it, so keep it and the base it was computed from around here.
1500 Value *OffsetPtr = nullptr;
1501 Value *OffsetBasePtr;
1503 // Remember any i8 pointer we come across to re-use if we need to do a raw
1505 Value *Int8Ptr = nullptr;
1506 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1508 Type *TargetTy = PointerTy->getPointerElementType();
1511 // First fold any existing GEPs into the offset.
1512 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1513 APInt GEPOffset(Offset.getBitWidth(), 0);
1514 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1516 Offset += GEPOffset;
1517 Ptr = GEP->getPointerOperand();
1518 if (!Visited.insert(Ptr).second)
1522 // See if we can perform a natural GEP here.
1524 if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1525 Indices, NamePrefix)) {
1526 // If we have a new natural pointer at the offset, clear out any old
1527 // offset pointer we computed. Unless it is the base pointer or
1528 // a non-instruction, we built a GEP we don't need. Zap it.
1529 if (OffsetPtr && OffsetPtr != OffsetBasePtr)
1530 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr)) {
1531 assert(I->use_empty() && "Built a GEP with uses some how!");
1532 I->eraseFromParent();
1535 OffsetBasePtr = Ptr;
1536 // If we also found a pointer of the right type, we're done.
1537 if (P->getType() == PointerTy)
1541 // Stash this pointer if we've found an i8*.
1542 if (Ptr->getType()->isIntegerTy(8)) {
1544 Int8PtrOffset = Offset;
1547 // Peel off a layer of the pointer and update the offset appropriately.
1548 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1549 Ptr = cast<Operator>(Ptr)->getOperand(0);
1550 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1551 if (GA->isInterposable())
1553 Ptr = GA->getAliasee();
1557 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1558 } while (Visited.insert(Ptr).second);
1562 Int8Ptr = IRB.CreateBitCast(
1563 Ptr, IRB.getInt8PtrTy(PointerTy->getPointerAddressSpace()),
1564 NamePrefix + "sroa_raw_cast");
1565 Int8PtrOffset = Offset;
1568 OffsetPtr = Int8PtrOffset == 0
1570 : IRB.CreateInBoundsGEP(IRB.getInt8Ty(), Int8Ptr,
1571 IRB.getInt(Int8PtrOffset),
1572 NamePrefix + "sroa_raw_idx");
1576 // On the off chance we were targeting i8*, guard the bitcast here.
1577 if (Ptr->getType() != PointerTy)
1578 Ptr = IRB.CreateBitCast(Ptr, PointerTy, NamePrefix + "sroa_cast");
1583 /// \brief Compute the adjusted alignment for a load or store from an offset.
1584 static unsigned getAdjustedAlignment(Instruction *I, uint64_t Offset,
1585 const DataLayout &DL) {
1588 if (auto *LI = dyn_cast<LoadInst>(I)) {
1589 Alignment = LI->getAlignment();
1591 } else if (auto *SI = dyn_cast<StoreInst>(I)) {
1592 Alignment = SI->getAlignment();
1593 Ty = SI->getValueOperand()->getType();
1595 llvm_unreachable("Only loads and stores are allowed!");
1599 Alignment = DL.getABITypeAlignment(Ty);
1601 return MinAlign(Alignment, Offset);
1604 /// \brief Test whether we can convert a value from the old to the new type.
1606 /// This predicate should be used to guard calls to convertValue in order to
1607 /// ensure that we only try to convert viable values. The strategy is that we
1608 /// will peel off single element struct and array wrappings to get to an
1609 /// underlying value, and convert that value.
1610 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1614 // For integer types, we can't handle any bit-width differences. This would
1615 // break both vector conversions with extension and introduce endianness
1616 // issues when in conjunction with loads and stores.
1617 if (isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) {
1618 assert(cast<IntegerType>(OldTy)->getBitWidth() !=
1619 cast<IntegerType>(NewTy)->getBitWidth() &&
1620 "We can't have the same bitwidth for different int types");
1624 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1626 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1629 // We can convert pointers to integers and vice-versa. Same for vectors
1630 // of pointers and integers.
1631 OldTy = OldTy->getScalarType();
1632 NewTy = NewTy->getScalarType();
1633 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1634 if (NewTy->isPointerTy() && OldTy->isPointerTy()) {
1635 return cast<PointerType>(NewTy)->getPointerAddressSpace() ==
1636 cast<PointerType>(OldTy)->getPointerAddressSpace();
1638 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
1646 /// \brief Generic routine to convert an SSA value to a value of a different
1649 /// This will try various different casting techniques, such as bitcasts,
1650 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1651 /// two types for viability with this routine.
1652 static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1654 Type *OldTy = V->getType();
1655 assert(canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type");
1660 assert(!(isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) &&
1661 "Integer types must be the exact same to convert.");
1663 // See if we need inttoptr for this type pair. A cast involving both scalars
1664 // and vectors requires and additional bitcast.
1665 if (OldTy->getScalarType()->isIntegerTy() &&
1666 NewTy->getScalarType()->isPointerTy()) {
1667 // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1668 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1669 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1672 // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1673 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1674 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1677 return IRB.CreateIntToPtr(V, NewTy);
1680 // See if we need ptrtoint for this type pair. A cast involving both scalars
1681 // and vectors requires and additional bitcast.
1682 if (OldTy->getScalarType()->isPointerTy() &&
1683 NewTy->getScalarType()->isIntegerTy()) {
1684 // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1685 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1686 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1689 // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1690 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1691 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1694 return IRB.CreatePtrToInt(V, NewTy);
1697 return IRB.CreateBitCast(V, NewTy);
1700 /// \brief Test whether the given slice use can be promoted to a vector.
1702 /// This function is called to test each entry in a partition which is slated
1703 /// for a single slice.
1704 static bool isVectorPromotionViableForSlice(Partition &P, const Slice &S,
1706 uint64_t ElementSize,
1707 const DataLayout &DL) {
1708 // First validate the slice offsets.
1709 uint64_t BeginOffset =
1710 std::max(S.beginOffset(), P.beginOffset()) - P.beginOffset();
1711 uint64_t BeginIndex = BeginOffset / ElementSize;
1712 if (BeginIndex * ElementSize != BeginOffset ||
1713 BeginIndex >= Ty->getNumElements())
1715 uint64_t EndOffset =
1716 std::min(S.endOffset(), P.endOffset()) - P.beginOffset();
1717 uint64_t EndIndex = EndOffset / ElementSize;
1718 if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
1721 assert(EndIndex > BeginIndex && "Empty vector!");
1722 uint64_t NumElements = EndIndex - BeginIndex;
1723 Type *SliceTy = (NumElements == 1)
1724 ? Ty->getElementType()
1725 : VectorType::get(Ty->getElementType(), NumElements);
1728 Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1730 Use *U = S.getUse();
1732 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1733 if (MI->isVolatile())
1735 if (!S.isSplittable())
1736 return false; // Skip any unsplittable intrinsics.
1737 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1738 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1739 II->getIntrinsicID() != Intrinsic::lifetime_end)
1741 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1742 // Disable vector promotion when there are loads or stores of an FCA.
1744 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1745 if (LI->isVolatile())
1747 Type *LTy = LI->getType();
1748 if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1749 assert(LTy->isIntegerTy());
1752 if (!canConvertValue(DL, SliceTy, LTy))
1754 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1755 if (SI->isVolatile())
1757 Type *STy = SI->getValueOperand()->getType();
1758 if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1759 assert(STy->isIntegerTy());
1762 if (!canConvertValue(DL, STy, SliceTy))
1771 /// \brief Test whether the given alloca partitioning and range of slices can be
1772 /// promoted to a vector.
1774 /// This is a quick test to check whether we can rewrite a particular alloca
1775 /// partition (and its newly formed alloca) into a vector alloca with only
1776 /// whole-vector loads and stores such that it could be promoted to a vector
1777 /// SSA value. We only can ensure this for a limited set of operations, and we
1778 /// don't want to do the rewrites unless we are confident that the result will
1779 /// be promotable, so we have an early test here.
1780 static VectorType *isVectorPromotionViable(Partition &P, const DataLayout &DL) {
1781 // Collect the candidate types for vector-based promotion. Also track whether
1782 // we have different element types.
1783 SmallVector<VectorType *, 4> CandidateTys;
1784 Type *CommonEltTy = nullptr;
1785 bool HaveCommonEltTy = true;
1786 auto CheckCandidateType = [&](Type *Ty) {
1787 if (auto *VTy = dyn_cast<VectorType>(Ty)) {
1788 CandidateTys.push_back(VTy);
1790 CommonEltTy = VTy->getElementType();
1791 else if (CommonEltTy != VTy->getElementType())
1792 HaveCommonEltTy = false;
1795 // Consider any loads or stores that are the exact size of the slice.
1796 for (const Slice &S : P)
1797 if (S.beginOffset() == P.beginOffset() &&
1798 S.endOffset() == P.endOffset()) {
1799 if (auto *LI = dyn_cast<LoadInst>(S.getUse()->getUser()))
1800 CheckCandidateType(LI->getType());
1801 else if (auto *SI = dyn_cast<StoreInst>(S.getUse()->getUser()))
1802 CheckCandidateType(SI->getValueOperand()->getType());
1805 // If we didn't find a vector type, nothing to do here.
1806 if (CandidateTys.empty())
1809 // Remove non-integer vector types if we had multiple common element types.
1810 // FIXME: It'd be nice to replace them with integer vector types, but we can't
1811 // do that until all the backends are known to produce good code for all
1812 // integer vector types.
1813 if (!HaveCommonEltTy) {
1814 CandidateTys.erase(remove_if(CandidateTys,
1815 [](VectorType *VTy) {
1816 return !VTy->getElementType()->isIntegerTy();
1818 CandidateTys.end());
1820 // If there were no integer vector types, give up.
1821 if (CandidateTys.empty())
1824 // Rank the remaining candidate vector types. This is easy because we know
1825 // they're all integer vectors. We sort by ascending number of elements.
1826 auto RankVectorTypes = [&DL](VectorType *RHSTy, VectorType *LHSTy) {
1827 assert(DL.getTypeSizeInBits(RHSTy) == DL.getTypeSizeInBits(LHSTy) &&
1828 "Cannot have vector types of different sizes!");
1829 assert(RHSTy->getElementType()->isIntegerTy() &&
1830 "All non-integer types eliminated!");
1831 assert(LHSTy->getElementType()->isIntegerTy() &&
1832 "All non-integer types eliminated!");
1833 return RHSTy->getNumElements() < LHSTy->getNumElements();
1835 std::sort(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes);
1837 std::unique(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes),
1838 CandidateTys.end());
1840 // The only way to have the same element type in every vector type is to
1841 // have the same vector type. Check that and remove all but one.
1843 for (VectorType *VTy : CandidateTys) {
1844 assert(VTy->getElementType() == CommonEltTy &&
1845 "Unaccounted for element type!");
1846 assert(VTy == CandidateTys[0] &&
1847 "Different vector types with the same element type!");
1850 CandidateTys.resize(1);
1853 // Try each vector type, and return the one which works.
1854 auto CheckVectorTypeForPromotion = [&](VectorType *VTy) {
1855 uint64_t ElementSize = DL.getTypeSizeInBits(VTy->getElementType());
1857 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1858 // that aren't byte sized.
1859 if (ElementSize % 8)
1861 assert((DL.getTypeSizeInBits(VTy) % 8) == 0 &&
1862 "vector size not a multiple of element size?");
1865 for (const Slice &S : P)
1866 if (!isVectorPromotionViableForSlice(P, S, VTy, ElementSize, DL))
1869 for (const Slice *S : P.splitSliceTails())
1870 if (!isVectorPromotionViableForSlice(P, *S, VTy, ElementSize, DL))
1875 for (VectorType *VTy : CandidateTys)
1876 if (CheckVectorTypeForPromotion(VTy))
1882 /// \brief Test whether a slice of an alloca is valid for integer widening.
1884 /// This implements the necessary checking for the \c isIntegerWideningViable
1885 /// test below on a single slice of the alloca.
1886 static bool isIntegerWideningViableForSlice(const Slice &S,
1887 uint64_t AllocBeginOffset,
1889 const DataLayout &DL,
1890 bool &WholeAllocaOp) {
1891 uint64_t Size = DL.getTypeStoreSize(AllocaTy);
1893 uint64_t RelBegin = S.beginOffset() - AllocBeginOffset;
1894 uint64_t RelEnd = S.endOffset() - AllocBeginOffset;
1896 // We can't reasonably handle cases where the load or store extends past
1897 // the end of the alloca's type and into its padding.
1901 Use *U = S.getUse();
1903 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1904 if (LI->isVolatile())
1906 // We can't handle loads that extend past the allocated memory.
1907 if (DL.getTypeStoreSize(LI->getType()) > Size)
1909 // Note that we don't count vector loads or stores as whole-alloca
1910 // operations which enable integer widening because we would prefer to use
1911 // vector widening instead.
1912 if (!isa<VectorType>(LI->getType()) && RelBegin == 0 && RelEnd == Size)
1913 WholeAllocaOp = true;
1914 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
1915 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1917 } else if (RelBegin != 0 || RelEnd != Size ||
1918 !canConvertValue(DL, AllocaTy, LI->getType())) {
1919 // Non-integer loads need to be convertible from the alloca type so that
1920 // they are promotable.
1923 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1924 Type *ValueTy = SI->getValueOperand()->getType();
1925 if (SI->isVolatile())
1927 // We can't handle stores that extend past the allocated memory.
1928 if (DL.getTypeStoreSize(ValueTy) > Size)
1930 // Note that we don't count vector loads or stores as whole-alloca
1931 // operations which enable integer widening because we would prefer to use
1932 // vector widening instead.
1933 if (!isa<VectorType>(ValueTy) && RelBegin == 0 && RelEnd == Size)
1934 WholeAllocaOp = true;
1935 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
1936 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1938 } else if (RelBegin != 0 || RelEnd != Size ||
1939 !canConvertValue(DL, ValueTy, AllocaTy)) {
1940 // Non-integer stores need to be convertible to the alloca type so that
1941 // they are promotable.
1944 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1945 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
1947 if (!S.isSplittable())
1948 return false; // Skip any unsplittable intrinsics.
1949 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1950 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1951 II->getIntrinsicID() != Intrinsic::lifetime_end)
1960 /// \brief Test whether the given alloca partition's integer operations can be
1961 /// widened to promotable ones.
1963 /// This is a quick test to check whether we can rewrite the integer loads and
1964 /// stores to a particular alloca into wider loads and stores and be able to
1965 /// promote the resulting alloca.
1966 static bool isIntegerWideningViable(Partition &P, Type *AllocaTy,
1967 const DataLayout &DL) {
1968 uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
1969 // Don't create integer types larger than the maximum bitwidth.
1970 if (SizeInBits > IntegerType::MAX_INT_BITS)
1973 // Don't try to handle allocas with bit-padding.
1974 if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
1977 // We need to ensure that an integer type with the appropriate bitwidth can
1978 // be converted to the alloca type, whatever that is. We don't want to force
1979 // the alloca itself to have an integer type if there is a more suitable one.
1980 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
1981 if (!canConvertValue(DL, AllocaTy, IntTy) ||
1982 !canConvertValue(DL, IntTy, AllocaTy))
1985 // While examining uses, we ensure that the alloca has a covering load or
1986 // store. We don't want to widen the integer operations only to fail to
1987 // promote due to some other unsplittable entry (which we may make splittable
1988 // later). However, if there are only splittable uses, go ahead and assume
1989 // that we cover the alloca.
1990 // FIXME: We shouldn't consider split slices that happen to start in the
1991 // partition here...
1992 bool WholeAllocaOp =
1993 P.begin() != P.end() ? false : DL.isLegalInteger(SizeInBits);
1995 for (const Slice &S : P)
1996 if (!isIntegerWideningViableForSlice(S, P.beginOffset(), AllocaTy, DL,
2000 for (const Slice *S : P.splitSliceTails())
2001 if (!isIntegerWideningViableForSlice(*S, P.beginOffset(), AllocaTy, DL,
2005 return WholeAllocaOp;
2008 static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
2009 IntegerType *Ty, uint64_t Offset,
2010 const Twine &Name) {
2011 DEBUG(dbgs() << " start: " << *V << "\n");
2012 IntegerType *IntTy = cast<IntegerType>(V->getType());
2013 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2014 "Element extends past full value");
2015 uint64_t ShAmt = 8 * Offset;
2016 if (DL.isBigEndian())
2017 ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2019 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
2020 DEBUG(dbgs() << " shifted: " << *V << "\n");
2022 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2023 "Cannot extract to a larger integer!");
2025 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
2026 DEBUG(dbgs() << " trunced: " << *V << "\n");
2031 static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
2032 Value *V, uint64_t Offset, const Twine &Name) {
2033 IntegerType *IntTy = cast<IntegerType>(Old->getType());
2034 IntegerType *Ty = cast<IntegerType>(V->getType());
2035 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2036 "Cannot insert a larger integer!");
2037 DEBUG(dbgs() << " start: " << *V << "\n");
2039 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
2040 DEBUG(dbgs() << " extended: " << *V << "\n");
2042 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2043 "Element store outside of alloca store");
2044 uint64_t ShAmt = 8 * Offset;
2045 if (DL.isBigEndian())
2046 ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2048 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
2049 DEBUG(dbgs() << " shifted: " << *V << "\n");
2052 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
2053 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
2054 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
2055 DEBUG(dbgs() << " masked: " << *Old << "\n");
2056 V = IRB.CreateOr(Old, V, Name + ".insert");
2057 DEBUG(dbgs() << " inserted: " << *V << "\n");
2062 static Value *extractVector(IRBuilderTy &IRB, Value *V, unsigned BeginIndex,
2063 unsigned EndIndex, const Twine &Name) {
2064 VectorType *VecTy = cast<VectorType>(V->getType());
2065 unsigned NumElements = EndIndex - BeginIndex;
2066 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2068 if (NumElements == VecTy->getNumElements())
2071 if (NumElements == 1) {
2072 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
2074 DEBUG(dbgs() << " extract: " << *V << "\n");
2078 SmallVector<Constant *, 8> Mask;
2079 Mask.reserve(NumElements);
2080 for (unsigned i = BeginIndex; i != EndIndex; ++i)
2081 Mask.push_back(IRB.getInt32(i));
2082 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2083 ConstantVector::get(Mask), Name + ".extract");
2084 DEBUG(dbgs() << " shuffle: " << *V << "\n");
2088 static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
2089 unsigned BeginIndex, const Twine &Name) {
2090 VectorType *VecTy = cast<VectorType>(Old->getType());
2091 assert(VecTy && "Can only insert a vector into a vector");
2093 VectorType *Ty = dyn_cast<VectorType>(V->getType());
2095 // Single element to insert.
2096 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
2098 DEBUG(dbgs() << " insert: " << *V << "\n");
2102 assert(Ty->getNumElements() <= VecTy->getNumElements() &&
2103 "Too many elements!");
2104 if (Ty->getNumElements() == VecTy->getNumElements()) {
2105 assert(V->getType() == VecTy && "Vector type mismatch");
2108 unsigned EndIndex = BeginIndex + Ty->getNumElements();
2110 // When inserting a smaller vector into the larger to store, we first
2111 // use a shuffle vector to widen it with undef elements, and then
2112 // a second shuffle vector to select between the loaded vector and the
2114 SmallVector<Constant *, 8> Mask;
2115 Mask.reserve(VecTy->getNumElements());
2116 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2117 if (i >= BeginIndex && i < EndIndex)
2118 Mask.push_back(IRB.getInt32(i - BeginIndex));
2120 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
2121 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2122 ConstantVector::get(Mask), Name + ".expand");
2123 DEBUG(dbgs() << " shuffle: " << *V << "\n");
2126 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2127 Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
2129 V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
2131 DEBUG(dbgs() << " blend: " << *V << "\n");
2135 /// \brief Visitor to rewrite instructions using p particular slice of an alloca
2136 /// to use a new alloca.
2138 /// Also implements the rewriting to vector-based accesses when the partition
2139 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2141 class llvm::sroa::AllocaSliceRewriter
2142 : public InstVisitor<AllocaSliceRewriter, bool> {
2143 // Befriend the base class so it can delegate to private visit methods.
2144 friend class llvm::InstVisitor<AllocaSliceRewriter, bool>;
2145 typedef llvm::InstVisitor<AllocaSliceRewriter, bool> Base;
2147 const DataLayout &DL;
2150 AllocaInst &OldAI, &NewAI;
2151 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2154 // This is a convenience and flag variable that will be null unless the new
2155 // alloca's integer operations should be widened to this integer type due to
2156 // passing isIntegerWideningViable above. If it is non-null, the desired
2157 // integer type will be stored here for easy access during rewriting.
2160 // If we are rewriting an alloca partition which can be written as pure
2161 // vector operations, we stash extra information here. When VecTy is
2162 // non-null, we have some strict guarantees about the rewritten alloca:
2163 // - The new alloca is exactly the size of the vector type here.
2164 // - The accesses all either map to the entire vector or to a single
2166 // - The set of accessing instructions is only one of those handled above
2167 // in isVectorPromotionViable. Generally these are the same access kinds
2168 // which are promotable via mem2reg.
2171 uint64_t ElementSize;
2173 // The original offset of the slice currently being rewritten relative to
2174 // the original alloca.
2175 uint64_t BeginOffset, EndOffset;
2176 // The new offsets of the slice currently being rewritten relative to the
2178 uint64_t NewBeginOffset, NewEndOffset;
2184 Instruction *OldPtr;
2186 // Track post-rewrite users which are PHI nodes and Selects.
2187 SmallPtrSetImpl<PHINode *> &PHIUsers;
2188 SmallPtrSetImpl<SelectInst *> &SelectUsers;
2190 // Utility IR builder, whose name prefix is setup for each visited use, and
2191 // the insertion point is set to point to the user.
2195 AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &AS, SROA &Pass,
2196 AllocaInst &OldAI, AllocaInst &NewAI,
2197 uint64_t NewAllocaBeginOffset,
2198 uint64_t NewAllocaEndOffset, bool IsIntegerPromotable,
2199 VectorType *PromotableVecTy,
2200 SmallPtrSetImpl<PHINode *> &PHIUsers,
2201 SmallPtrSetImpl<SelectInst *> &SelectUsers)
2202 : DL(DL), AS(AS), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
2203 NewAllocaBeginOffset(NewAllocaBeginOffset),
2204 NewAllocaEndOffset(NewAllocaEndOffset),
2205 NewAllocaTy(NewAI.getAllocatedType()),
2206 IntTy(IsIntegerPromotable
2209 DL.getTypeSizeInBits(NewAI.getAllocatedType()))
2211 VecTy(PromotableVecTy),
2212 ElementTy(VecTy ? VecTy->getElementType() : nullptr),
2213 ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
2214 BeginOffset(), EndOffset(), IsSplittable(), IsSplit(), OldUse(),
2215 OldPtr(), PHIUsers(PHIUsers), SelectUsers(SelectUsers),
2216 IRB(NewAI.getContext(), ConstantFolder()) {
2218 assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
2219 "Only multiple-of-8 sized vector elements are viable");
2222 assert((!IntTy && !VecTy) || (IntTy && !VecTy) || (!IntTy && VecTy));
2225 bool visit(AllocaSlices::const_iterator I) {
2226 bool CanSROA = true;
2227 BeginOffset = I->beginOffset();
2228 EndOffset = I->endOffset();
2229 IsSplittable = I->isSplittable();
2231 BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
2232 DEBUG(dbgs() << " rewriting " << (IsSplit ? "split " : ""));
2233 DEBUG(AS.printSlice(dbgs(), I, ""));
2234 DEBUG(dbgs() << "\n");
2236 // Compute the intersecting offset range.
2237 assert(BeginOffset < NewAllocaEndOffset);
2238 assert(EndOffset > NewAllocaBeginOffset);
2239 NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2240 NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2242 SliceSize = NewEndOffset - NewBeginOffset;
2244 OldUse = I->getUse();
2245 OldPtr = cast<Instruction>(OldUse->get());
2247 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2248 IRB.SetInsertPoint(OldUserI);
2249 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2250 IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
2252 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2259 // Make sure the other visit overloads are visible.
2262 // Every instruction which can end up as a user must have a rewrite rule.
2263 bool visitInstruction(Instruction &I) {
2264 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2265 llvm_unreachable("No rewrite rule for this instruction!");
2268 Value *getNewAllocaSlicePtr(IRBuilderTy &IRB, Type *PointerTy) {
2269 // Note that the offset computation can use BeginOffset or NewBeginOffset
2270 // interchangeably for unsplit slices.
2271 assert(IsSplit || BeginOffset == NewBeginOffset);
2272 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2275 StringRef OldName = OldPtr->getName();
2276 // Skip through the last '.sroa.' component of the name.
2277 size_t LastSROAPrefix = OldName.rfind(".sroa.");
2278 if (LastSROAPrefix != StringRef::npos) {
2279 OldName = OldName.substr(LastSROAPrefix + strlen(".sroa."));
2280 // Look for an SROA slice index.
2281 size_t IndexEnd = OldName.find_first_not_of("0123456789");
2282 if (IndexEnd != StringRef::npos && OldName[IndexEnd] == '.') {
2283 // Strip the index and look for the offset.
2284 OldName = OldName.substr(IndexEnd + 1);
2285 size_t OffsetEnd = OldName.find_first_not_of("0123456789");
2286 if (OffsetEnd != StringRef::npos && OldName[OffsetEnd] == '.')
2287 // Strip the offset.
2288 OldName = OldName.substr(OffsetEnd + 1);
2291 // Strip any SROA suffixes as well.
2292 OldName = OldName.substr(0, OldName.find(".sroa_"));
2295 return getAdjustedPtr(IRB, DL, &NewAI,
2296 APInt(DL.getPointerSizeInBits(), Offset), PointerTy,
2298 Twine(OldName) + "."
2305 /// \brief Compute suitable alignment to access this slice of the *new*
2308 /// You can optionally pass a type to this routine and if that type's ABI
2309 /// alignment is itself suitable, this will return zero.
2310 unsigned getSliceAlign(Type *Ty = nullptr) {
2311 unsigned NewAIAlign = NewAI.getAlignment();
2313 NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
2315 MinAlign(NewAIAlign, NewBeginOffset - NewAllocaBeginOffset);
2316 return (Ty && Align == DL.getABITypeAlignment(Ty)) ? 0 : Align;
2319 unsigned getIndex(uint64_t Offset) {
2320 assert(VecTy && "Can only call getIndex when rewriting a vector");
2321 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2322 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2323 uint32_t Index = RelOffset / ElementSize;
2324 assert(Index * ElementSize == RelOffset);
2328 void deleteIfTriviallyDead(Value *V) {
2329 Instruction *I = cast<Instruction>(V);
2330 if (isInstructionTriviallyDead(I))
2331 Pass.DeadInsts.insert(I);
2334 Value *rewriteVectorizedLoadInst() {
2335 unsigned BeginIndex = getIndex(NewBeginOffset);
2336 unsigned EndIndex = getIndex(NewEndOffset);
2337 assert(EndIndex > BeginIndex && "Empty vector!");
2339 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2340 return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2343 Value *rewriteIntegerLoad(LoadInst &LI) {
2344 assert(IntTy && "We cannot insert an integer to the alloca");
2345 assert(!LI.isVolatile());
2346 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2347 V = convertValue(DL, IRB, V, IntTy);
2348 assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2349 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2350 if (Offset > 0 || NewEndOffset < NewAllocaEndOffset) {
2351 IntegerType *ExtractTy = Type::getIntNTy(LI.getContext(), SliceSize * 8);
2352 V = extractInteger(DL, IRB, V, ExtractTy, Offset, "extract");
2354 // It is possible that the extracted type is not the load type. This
2355 // happens if there is a load past the end of the alloca, and as
2356 // a consequence the slice is narrower but still a candidate for integer
2357 // lowering. To handle this case, we just zero extend the extracted
2359 assert(cast<IntegerType>(LI.getType())->getBitWidth() >= SliceSize * 8 &&
2360 "Can only handle an extract for an overly wide load");
2361 if (cast<IntegerType>(LI.getType())->getBitWidth() > SliceSize * 8)
2362 V = IRB.CreateZExt(V, LI.getType());
2366 bool visitLoadInst(LoadInst &LI) {
2367 DEBUG(dbgs() << " original: " << LI << "\n");
2368 Value *OldOp = LI.getOperand(0);
2369 assert(OldOp == OldPtr);
2371 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), SliceSize * 8)
2373 const bool IsLoadPastEnd = DL.getTypeStoreSize(TargetTy) > SliceSize;
2374 bool IsPtrAdjusted = false;
2377 V = rewriteVectorizedLoadInst();
2378 } else if (IntTy && LI.getType()->isIntegerTy()) {
2379 V = rewriteIntegerLoad(LI);
2380 } else if (NewBeginOffset == NewAllocaBeginOffset &&
2381 NewEndOffset == NewAllocaEndOffset &&
2382 (canConvertValue(DL, NewAllocaTy, TargetTy) ||
2383 (IsLoadPastEnd && NewAllocaTy->isIntegerTy() &&
2384 TargetTy->isIntegerTy()))) {
2385 LoadInst *NewLI = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2386 LI.isVolatile(), LI.getName());
2387 if (LI.isVolatile())
2388 NewLI->setAtomic(LI.getOrdering(), LI.getSynchScope());
2391 // If this is an integer load past the end of the slice (which means the
2392 // bytes outside the slice are undef or this load is dead) just forcibly
2393 // fix the integer size with correct handling of endianness.
2394 if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2395 if (auto *TITy = dyn_cast<IntegerType>(TargetTy))
2396 if (AITy->getBitWidth() < TITy->getBitWidth()) {
2397 V = IRB.CreateZExt(V, TITy, "load.ext");
2398 if (DL.isBigEndian())
2399 V = IRB.CreateShl(V, TITy->getBitWidth() - AITy->getBitWidth(),
2403 Type *LTy = TargetTy->getPointerTo();
2404 LoadInst *NewLI = IRB.CreateAlignedLoad(getNewAllocaSlicePtr(IRB, LTy),
2405 getSliceAlign(TargetTy),
2406 LI.isVolatile(), LI.getName());
2407 if (LI.isVolatile())
2408 NewLI->setAtomic(LI.getOrdering(), LI.getSynchScope());
2411 IsPtrAdjusted = true;
2413 V = convertValue(DL, IRB, V, TargetTy);
2416 assert(!LI.isVolatile());
2417 assert(LI.getType()->isIntegerTy() &&
2418 "Only integer type loads and stores are split");
2419 assert(SliceSize < DL.getTypeStoreSize(LI.getType()) &&
2420 "Split load isn't smaller than original load");
2421 assert(LI.getType()->getIntegerBitWidth() ==
2422 DL.getTypeStoreSizeInBits(LI.getType()) &&
2423 "Non-byte-multiple bit width");
2424 // Move the insertion point just past the load so that we can refer to it.
2425 IRB.SetInsertPoint(&*std::next(BasicBlock::iterator(&LI)));
2426 // Create a placeholder value with the same type as LI to use as the
2427 // basis for the new value. This allows us to replace the uses of LI with
2428 // the computed value, and then replace the placeholder with LI, leaving
2429 // LI only used for this computation.
2430 Value *Placeholder =
2431 new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
2432 V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset - BeginOffset,
2434 LI.replaceAllUsesWith(V);
2435 Placeholder->replaceAllUsesWith(&LI);
2438 LI.replaceAllUsesWith(V);
2441 Pass.DeadInsts.insert(&LI);
2442 deleteIfTriviallyDead(OldOp);
2443 DEBUG(dbgs() << " to: " << *V << "\n");
2444 return !LI.isVolatile() && !IsPtrAdjusted;
2447 bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp) {
2448 if (V->getType() != VecTy) {
2449 unsigned BeginIndex = getIndex(NewBeginOffset);
2450 unsigned EndIndex = getIndex(NewEndOffset);
2451 assert(EndIndex > BeginIndex && "Empty vector!");
2452 unsigned NumElements = EndIndex - BeginIndex;
2453 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2454 Type *SliceTy = (NumElements == 1)
2456 : VectorType::get(ElementTy, NumElements);
2457 if (V->getType() != SliceTy)
2458 V = convertValue(DL, IRB, V, SliceTy);
2460 // Mix in the existing elements.
2461 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2462 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2464 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2465 Pass.DeadInsts.insert(&SI);
2468 DEBUG(dbgs() << " to: " << *Store << "\n");
2472 bool rewriteIntegerStore(Value *V, StoreInst &SI) {
2473 assert(IntTy && "We cannot extract an integer from the alloca");
2474 assert(!SI.isVolatile());
2475 if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2477 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2478 Old = convertValue(DL, IRB, Old, IntTy);
2479 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2480 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2481 V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset, "insert");
2483 V = convertValue(DL, IRB, V, NewAllocaTy);
2484 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2485 Store->copyMetadata(SI, LLVMContext::MD_mem_parallel_loop_access);
2486 Pass.DeadInsts.insert(&SI);
2487 DEBUG(dbgs() << " to: " << *Store << "\n");
2491 bool visitStoreInst(StoreInst &SI) {
2492 DEBUG(dbgs() << " original: " << SI << "\n");
2493 Value *OldOp = SI.getOperand(1);
2494 assert(OldOp == OldPtr);
2496 Value *V = SI.getValueOperand();
2498 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2499 // alloca that should be re-examined after promoting this alloca.
2500 if (V->getType()->isPointerTy())
2501 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2502 Pass.PostPromotionWorklist.insert(AI);
2504 if (SliceSize < DL.getTypeStoreSize(V->getType())) {
2505 assert(!SI.isVolatile());
2506 assert(V->getType()->isIntegerTy() &&
2507 "Only integer type loads and stores are split");
2508 assert(V->getType()->getIntegerBitWidth() ==
2509 DL.getTypeStoreSizeInBits(V->getType()) &&
2510 "Non-byte-multiple bit width");
2511 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), SliceSize * 8);
2512 V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset - BeginOffset,
2517 return rewriteVectorizedStoreInst(V, SI, OldOp);
2518 if (IntTy && V->getType()->isIntegerTy())
2519 return rewriteIntegerStore(V, SI);
2521 const bool IsStorePastEnd = DL.getTypeStoreSize(V->getType()) > SliceSize;
2523 if (NewBeginOffset == NewAllocaBeginOffset &&
2524 NewEndOffset == NewAllocaEndOffset &&
2525 (canConvertValue(DL, V->getType(), NewAllocaTy) ||
2526 (IsStorePastEnd && NewAllocaTy->isIntegerTy() &&
2527 V->getType()->isIntegerTy()))) {
2528 // If this is an integer store past the end of slice (and thus the bytes
2529 // past that point are irrelevant or this is unreachable), truncate the
2530 // value prior to storing.
2531 if (auto *VITy = dyn_cast<IntegerType>(V->getType()))
2532 if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2533 if (VITy->getBitWidth() > AITy->getBitWidth()) {
2534 if (DL.isBigEndian())
2535 V = IRB.CreateLShr(V, VITy->getBitWidth() - AITy->getBitWidth(),
2537 V = IRB.CreateTrunc(V, AITy, "load.trunc");
2540 V = convertValue(DL, IRB, V, NewAllocaTy);
2541 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2544 Value *NewPtr = getNewAllocaSlicePtr(IRB, V->getType()->getPointerTo());
2545 NewSI = IRB.CreateAlignedStore(V, NewPtr, getSliceAlign(V->getType()),
2548 NewSI->copyMetadata(SI, LLVMContext::MD_mem_parallel_loop_access);
2549 if (SI.isVolatile())
2550 NewSI->setAtomic(SI.getOrdering(), SI.getSynchScope());
2551 Pass.DeadInsts.insert(&SI);
2552 deleteIfTriviallyDead(OldOp);
2554 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2555 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2558 /// \brief Compute an integer value from splatting an i8 across the given
2559 /// number of bytes.
2561 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2562 /// call this routine.
2563 /// FIXME: Heed the advice above.
2565 /// \param V The i8 value to splat.
2566 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2567 Value *getIntegerSplat(Value *V, unsigned Size) {
2568 assert(Size > 0 && "Expected a positive number of bytes.");
2569 IntegerType *VTy = cast<IntegerType>(V->getType());
2570 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2574 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size * 8);
2576 IRB.CreateZExt(V, SplatIntTy, "zext"),
2577 ConstantExpr::getUDiv(
2578 Constant::getAllOnesValue(SplatIntTy),
2579 ConstantExpr::getZExt(Constant::getAllOnesValue(V->getType()),
2585 /// \brief Compute a vector splat for a given element value.
2586 Value *getVectorSplat(Value *V, unsigned NumElements) {
2587 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2588 DEBUG(dbgs() << " splat: " << *V << "\n");
2592 bool visitMemSetInst(MemSetInst &II) {
2593 DEBUG(dbgs() << " original: " << II << "\n");
2594 assert(II.getRawDest() == OldPtr);
2596 // If the memset has a variable size, it cannot be split, just adjust the
2597 // pointer to the new alloca.
2598 if (!isa<Constant>(II.getLength())) {
2600 assert(NewBeginOffset == BeginOffset);
2601 II.setDest(getNewAllocaSlicePtr(IRB, OldPtr->getType()));
2602 Type *CstTy = II.getAlignmentCst()->getType();
2603 II.setAlignment(ConstantInt::get(CstTy, getSliceAlign()));
2605 deleteIfTriviallyDead(OldPtr);
2609 // Record this instruction for deletion.
2610 Pass.DeadInsts.insert(&II);
2612 Type *AllocaTy = NewAI.getAllocatedType();
2613 Type *ScalarTy = AllocaTy->getScalarType();
2615 // If this doesn't map cleanly onto the alloca type, and that type isn't
2616 // a single value type, just emit a memset.
2617 if (!VecTy && !IntTy &&
2618 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2619 SliceSize != DL.getTypeStoreSize(AllocaTy) ||
2620 !AllocaTy->isSingleValueType() ||
2621 !DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy)) ||
2622 DL.getTypeSizeInBits(ScalarTy) % 8 != 0)) {
2623 Type *SizeTy = II.getLength()->getType();
2624 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2625 CallInst *New = IRB.CreateMemSet(
2626 getNewAllocaSlicePtr(IRB, OldPtr->getType()), II.getValue(), Size,
2627 getSliceAlign(), II.isVolatile());
2629 DEBUG(dbgs() << " to: " << *New << "\n");
2633 // If we can represent this as a simple value, we have to build the actual
2634 // value to store, which requires expanding the byte present in memset to
2635 // a sensible representation for the alloca type. This is essentially
2636 // splatting the byte to a sufficiently wide integer, splatting it across
2637 // any desired vector width, and bitcasting to the final type.
2641 // If this is a memset of a vectorized alloca, insert it.
2642 assert(ElementTy == ScalarTy);
2644 unsigned BeginIndex = getIndex(NewBeginOffset);
2645 unsigned EndIndex = getIndex(NewEndOffset);
2646 assert(EndIndex > BeginIndex && "Empty vector!");
2647 unsigned NumElements = EndIndex - BeginIndex;
2648 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2651 getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
2652 Splat = convertValue(DL, IRB, Splat, ElementTy);
2653 if (NumElements > 1)
2654 Splat = getVectorSplat(Splat, NumElements);
2657 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2658 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2660 // If this is a memset on an alloca where we can widen stores, insert the
2662 assert(!II.isVolatile());
2664 uint64_t Size = NewEndOffset - NewBeginOffset;
2665 V = getIntegerSplat(II.getValue(), Size);
2667 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2668 EndOffset != NewAllocaBeginOffset)) {
2670 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2671 Old = convertValue(DL, IRB, Old, IntTy);
2672 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2673 V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2675 assert(V->getType() == IntTy &&
2676 "Wrong type for an alloca wide integer!");
2678 V = convertValue(DL, IRB, V, AllocaTy);
2680 // Established these invariants above.
2681 assert(NewBeginOffset == NewAllocaBeginOffset);
2682 assert(NewEndOffset == NewAllocaEndOffset);
2684 V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
2685 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2686 V = getVectorSplat(V, AllocaVecTy->getNumElements());
2688 V = convertValue(DL, IRB, V, AllocaTy);
2691 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2694 DEBUG(dbgs() << " to: " << *New << "\n");
2695 return !II.isVolatile();
2698 bool visitMemTransferInst(MemTransferInst &II) {
2699 // Rewriting of memory transfer instructions can be a bit tricky. We break
2700 // them into two categories: split intrinsics and unsplit intrinsics.
2702 DEBUG(dbgs() << " original: " << II << "\n");
2704 bool IsDest = &II.getRawDestUse() == OldUse;
2705 assert((IsDest && II.getRawDest() == OldPtr) ||
2706 (!IsDest && II.getRawSource() == OldPtr));
2708 unsigned SliceAlign = getSliceAlign();
2710 // For unsplit intrinsics, we simply modify the source and destination
2711 // pointers in place. This isn't just an optimization, it is a matter of
2712 // correctness. With unsplit intrinsics we may be dealing with transfers
2713 // within a single alloca before SROA ran, or with transfers that have
2714 // a variable length. We may also be dealing with memmove instead of
2715 // memcpy, and so simply updating the pointers is the necessary for us to
2716 // update both source and dest of a single call.
2717 if (!IsSplittable) {
2718 Value *AdjustedPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2720 II.setDest(AdjustedPtr);
2722 II.setSource(AdjustedPtr);
2724 if (II.getAlignment() > SliceAlign) {
2725 Type *CstTy = II.getAlignmentCst()->getType();
2727 ConstantInt::get(CstTy, MinAlign(II.getAlignment(), SliceAlign)));
2730 DEBUG(dbgs() << " to: " << II << "\n");
2731 deleteIfTriviallyDead(OldPtr);
2734 // For split transfer intrinsics we have an incredibly useful assurance:
2735 // the source and destination do not reside within the same alloca, and at
2736 // least one of them does not escape. This means that we can replace
2737 // memmove with memcpy, and we don't need to worry about all manner of
2738 // downsides to splitting and transforming the operations.
2740 // If this doesn't map cleanly onto the alloca type, and that type isn't
2741 // a single value type, just emit a memcpy.
2744 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2745 SliceSize != DL.getTypeStoreSize(NewAI.getAllocatedType()) ||
2746 !NewAI.getAllocatedType()->isSingleValueType());
2748 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2749 // size hasn't been shrunk based on analysis of the viable range, this is
2751 if (EmitMemCpy && &OldAI == &NewAI) {
2752 // Ensure the start lines up.
2753 assert(NewBeginOffset == BeginOffset);
2755 // Rewrite the size as needed.
2756 if (NewEndOffset != EndOffset)
2757 II.setLength(ConstantInt::get(II.getLength()->getType(),
2758 NewEndOffset - NewBeginOffset));
2761 // Record this instruction for deletion.
2762 Pass.DeadInsts.insert(&II);
2764 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2765 // alloca that should be re-examined after rewriting this instruction.
2766 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2767 if (AllocaInst *AI =
2768 dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
2769 assert(AI != &OldAI && AI != &NewAI &&
2770 "Splittable transfers cannot reach the same alloca on both ends.");
2771 Pass.Worklist.insert(AI);
2774 Type *OtherPtrTy = OtherPtr->getType();
2775 unsigned OtherAS = OtherPtrTy->getPointerAddressSpace();
2777 // Compute the relative offset for the other pointer within the transfer.
2778 unsigned IntPtrWidth = DL.getPointerSizeInBits(OtherAS);
2779 APInt OtherOffset(IntPtrWidth, NewBeginOffset - BeginOffset);
2780 unsigned OtherAlign = MinAlign(II.getAlignment() ? II.getAlignment() : 1,
2781 OtherOffset.zextOrTrunc(64).getZExtValue());
2784 // Compute the other pointer, folding as much as possible to produce
2785 // a single, simple GEP in most cases.
2786 OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2787 OtherPtr->getName() + ".");
2789 Value *OurPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2790 Type *SizeTy = II.getLength()->getType();
2791 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2793 CallInst *New = IRB.CreateMemCpy(
2794 IsDest ? OurPtr : OtherPtr, IsDest ? OtherPtr : OurPtr, Size,
2795 MinAlign(SliceAlign, OtherAlign), II.isVolatile());
2797 DEBUG(dbgs() << " to: " << *New << "\n");
2801 bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
2802 NewEndOffset == NewAllocaEndOffset;
2803 uint64_t Size = NewEndOffset - NewBeginOffset;
2804 unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
2805 unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
2806 unsigned NumElements = EndIndex - BeginIndex;
2807 IntegerType *SubIntTy =
2808 IntTy ? Type::getIntNTy(IntTy->getContext(), Size * 8) : nullptr;
2810 // Reset the other pointer type to match the register type we're going to
2811 // use, but using the address space of the original other pointer.
2812 if (VecTy && !IsWholeAlloca) {
2813 if (NumElements == 1)
2814 OtherPtrTy = VecTy->getElementType();
2816 OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2818 OtherPtrTy = OtherPtrTy->getPointerTo(OtherAS);
2819 } else if (IntTy && !IsWholeAlloca) {
2820 OtherPtrTy = SubIntTy->getPointerTo(OtherAS);
2822 OtherPtrTy = NewAllocaTy->getPointerTo(OtherAS);
2825 Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2826 OtherPtr->getName() + ".");
2827 unsigned SrcAlign = OtherAlign;
2828 Value *DstPtr = &NewAI;
2829 unsigned DstAlign = SliceAlign;
2831 std::swap(SrcPtr, DstPtr);
2832 std::swap(SrcAlign, DstAlign);
2836 if (VecTy && !IsWholeAlloca && !IsDest) {
2837 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2838 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
2839 } else if (IntTy && !IsWholeAlloca && !IsDest) {
2840 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2841 Src = convertValue(DL, IRB, Src, IntTy);
2842 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2843 Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
2846 IRB.CreateAlignedLoad(SrcPtr, SrcAlign, II.isVolatile(), "copyload");
2849 if (VecTy && !IsWholeAlloca && IsDest) {
2851 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2852 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
2853 } else if (IntTy && !IsWholeAlloca && IsDest) {
2855 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2856 Old = convertValue(DL, IRB, Old, IntTy);
2857 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2858 Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
2859 Src = convertValue(DL, IRB, Src, NewAllocaTy);
2862 StoreInst *Store = cast<StoreInst>(
2863 IRB.CreateAlignedStore(Src, DstPtr, DstAlign, II.isVolatile()));
2865 DEBUG(dbgs() << " to: " << *Store << "\n");
2866 return !II.isVolatile();
2869 bool visitIntrinsicInst(IntrinsicInst &II) {
2870 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2871 II.getIntrinsicID() == Intrinsic::lifetime_end);
2872 DEBUG(dbgs() << " original: " << II << "\n");
2873 assert(II.getArgOperand(1) == OldPtr);
2875 // Record this instruction for deletion.
2876 Pass.DeadInsts.insert(&II);
2878 // Lifetime intrinsics are only promotable if they cover the whole alloca.
2879 // Therefore, we drop lifetime intrinsics which don't cover the whole
2881 // (In theory, intrinsics which partially cover an alloca could be
2882 // promoted, but PromoteMemToReg doesn't handle that case.)
2883 // FIXME: Check whether the alloca is promotable before dropping the
2884 // lifetime intrinsics?
2885 if (NewBeginOffset != NewAllocaBeginOffset ||
2886 NewEndOffset != NewAllocaEndOffset)
2890 ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2891 NewEndOffset - NewBeginOffset);
2892 Value *Ptr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2894 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2895 New = IRB.CreateLifetimeStart(Ptr, Size);
2897 New = IRB.CreateLifetimeEnd(Ptr, Size);
2900 DEBUG(dbgs() << " to: " << *New << "\n");
2905 bool visitPHINode(PHINode &PN) {
2906 DEBUG(dbgs() << " original: " << PN << "\n");
2907 assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
2908 assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
2910 // We would like to compute a new pointer in only one place, but have it be
2911 // as local as possible to the PHI. To do that, we re-use the location of
2912 // the old pointer, which necessarily must be in the right position to
2913 // dominate the PHI.
2914 IRBuilderTy PtrBuilder(IRB);
2915 if (isa<PHINode>(OldPtr))
2916 PtrBuilder.SetInsertPoint(&*OldPtr->getParent()->getFirstInsertionPt());
2918 PtrBuilder.SetInsertPoint(OldPtr);
2919 PtrBuilder.SetCurrentDebugLocation(OldPtr->getDebugLoc());
2921 Value *NewPtr = getNewAllocaSlicePtr(PtrBuilder, OldPtr->getType());
2922 // Replace the operands which were using the old pointer.
2923 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
2925 DEBUG(dbgs() << " to: " << PN << "\n");
2926 deleteIfTriviallyDead(OldPtr);
2928 // PHIs can't be promoted on their own, but often can be speculated. We
2929 // check the speculation outside of the rewriter so that we see the
2930 // fully-rewritten alloca.
2931 PHIUsers.insert(&PN);
2935 bool visitSelectInst(SelectInst &SI) {
2936 DEBUG(dbgs() << " original: " << SI << "\n");
2937 assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
2938 "Pointer isn't an operand!");
2939 assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
2940 assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
2942 Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2943 // Replace the operands which were using the old pointer.
2944 if (SI.getOperand(1) == OldPtr)
2945 SI.setOperand(1, NewPtr);
2946 if (SI.getOperand(2) == OldPtr)
2947 SI.setOperand(2, NewPtr);
2949 DEBUG(dbgs() << " to: " << SI << "\n");
2950 deleteIfTriviallyDead(OldPtr);
2952 // Selects can't be promoted on their own, but often can be speculated. We
2953 // check the speculation outside of the rewriter so that we see the
2954 // fully-rewritten alloca.
2955 SelectUsers.insert(&SI);
2961 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
2963 /// This pass aggressively rewrites all aggregate loads and stores on
2964 /// a particular pointer (or any pointer derived from it which we can identify)
2965 /// with scalar loads and stores.
2966 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
2967 // Befriend the base class so it can delegate to private visit methods.
2968 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
2970 /// Queue of pointer uses to analyze and potentially rewrite.
2971 SmallVector<Use *, 8> Queue;
2973 /// Set to prevent us from cycling with phi nodes and loops.
2974 SmallPtrSet<User *, 8> Visited;
2976 /// The current pointer use being rewritten. This is used to dig up the used
2977 /// value (as opposed to the user).
2981 /// Rewrite loads and stores through a pointer and all pointers derived from
2983 bool rewrite(Instruction &I) {
2984 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
2986 bool Changed = false;
2987 while (!Queue.empty()) {
2988 U = Queue.pop_back_val();
2989 Changed |= visit(cast<Instruction>(U->getUser()));
2995 /// Enqueue all the users of the given instruction for further processing.
2996 /// This uses a set to de-duplicate users.
2997 void enqueueUsers(Instruction &I) {
2998 for (Use &U : I.uses())
2999 if (Visited.insert(U.getUser()).second)
3000 Queue.push_back(&U);
3003 // Conservative default is to not rewrite anything.
3004 bool visitInstruction(Instruction &I) { return false; }
3006 /// \brief Generic recursive split emission class.
3007 template <typename Derived> class OpSplitter {
3009 /// The builder used to form new instructions.
3011 /// The indices which to be used with insert- or extractvalue to select the
3012 /// appropriate value within the aggregate.
3013 SmallVector<unsigned, 4> Indices;
3014 /// The indices to a GEP instruction which will move Ptr to the correct slot
3015 /// within the aggregate.
3016 SmallVector<Value *, 4> GEPIndices;
3017 /// The base pointer of the original op, used as a base for GEPing the
3018 /// split operations.
3021 /// Initialize the splitter with an insertion point, Ptr and start with a
3022 /// single zero GEP index.
3023 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
3024 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
3027 /// \brief Generic recursive split emission routine.
3029 /// This method recursively splits an aggregate op (load or store) into
3030 /// scalar or vector ops. It splits recursively until it hits a single value
3031 /// and emits that single value operation via the template argument.
3033 /// The logic of this routine relies on GEPs and insertvalue and
3034 /// extractvalue all operating with the same fundamental index list, merely
3035 /// formatted differently (GEPs need actual values).
3037 /// \param Ty The type being split recursively into smaller ops.
3038 /// \param Agg The aggregate value being built up or stored, depending on
3039 /// whether this is splitting a load or a store respectively.
3040 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
3041 if (Ty->isSingleValueType())
3042 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
3044 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
3045 unsigned OldSize = Indices.size();
3047 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
3049 assert(Indices.size() == OldSize && "Did not return to the old size");
3050 Indices.push_back(Idx);
3051 GEPIndices.push_back(IRB.getInt32(Idx));
3052 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
3053 GEPIndices.pop_back();
3059 if (StructType *STy = dyn_cast<StructType>(Ty)) {
3060 unsigned OldSize = Indices.size();
3062 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
3064 assert(Indices.size() == OldSize && "Did not return to the old size");
3065 Indices.push_back(Idx);
3066 GEPIndices.push_back(IRB.getInt32(Idx));
3067 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
3068 GEPIndices.pop_back();
3074 llvm_unreachable("Only arrays and structs are aggregate loadable types");
3078 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
3079 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3080 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
3082 /// Emit a leaf load of a single value. This is called at the leaves of the
3083 /// recursive emission to actually load values.
3084 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3085 assert(Ty->isSingleValueType());
3086 // Load the single value and insert it using the indices.
3088 IRB.CreateInBoundsGEP(nullptr, Ptr, GEPIndices, Name + ".gep");
3089 Value *Load = IRB.CreateLoad(GEP, Name + ".load");
3090 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
3091 DEBUG(dbgs() << " to: " << *Load << "\n");
3095 bool visitLoadInst(LoadInst &LI) {
3096 assert(LI.getPointerOperand() == *U);
3097 if (!LI.isSimple() || LI.getType()->isSingleValueType())
3100 // We have an aggregate being loaded, split it apart.
3101 DEBUG(dbgs() << " original: " << LI << "\n");
3102 LoadOpSplitter Splitter(&LI, *U);
3103 Value *V = UndefValue::get(LI.getType());
3104 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
3105 LI.replaceAllUsesWith(V);
3106 LI.eraseFromParent();
3110 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
3111 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3112 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
3114 /// Emit a leaf store of a single value. This is called at the leaves of the
3115 /// recursive emission to actually produce stores.
3116 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3117 assert(Ty->isSingleValueType());
3118 // Extract the single value and store it using the indices.
3120 // The gep and extractvalue values are factored out of the CreateStore
3121 // call to make the output independent of the argument evaluation order.
3122 Value *ExtractValue =
3123 IRB.CreateExtractValue(Agg, Indices, Name + ".extract");
3124 Value *InBoundsGEP =
3125 IRB.CreateInBoundsGEP(nullptr, Ptr, GEPIndices, Name + ".gep");
3126 Value *Store = IRB.CreateStore(ExtractValue, InBoundsGEP);
3128 DEBUG(dbgs() << " to: " << *Store << "\n");
3132 bool visitStoreInst(StoreInst &SI) {
3133 if (!SI.isSimple() || SI.getPointerOperand() != *U)
3135 Value *V = SI.getValueOperand();
3136 if (V->getType()->isSingleValueType())
3139 // We have an aggregate being stored, split it apart.
3140 DEBUG(dbgs() << " original: " << SI << "\n");
3141 StoreOpSplitter Splitter(&SI, *U);
3142 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
3143 SI.eraseFromParent();
3147 bool visitBitCastInst(BitCastInst &BC) {
3152 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
3157 bool visitPHINode(PHINode &PN) {
3162 bool visitSelectInst(SelectInst &SI) {
3169 /// \brief Strip aggregate type wrapping.
3171 /// This removes no-op aggregate types wrapping an underlying type. It will
3172 /// strip as many layers of types as it can without changing either the type
3173 /// size or the allocated size.
3174 static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
3175 if (Ty->isSingleValueType())
3178 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
3179 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
3182 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
3183 InnerTy = ArrTy->getElementType();
3184 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
3185 const StructLayout *SL = DL.getStructLayout(STy);
3186 unsigned Index = SL->getElementContainingOffset(0);
3187 InnerTy = STy->getElementType(Index);
3192 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
3193 TypeSize > DL.getTypeSizeInBits(InnerTy))
3196 return stripAggregateTypeWrapping(DL, InnerTy);
3199 /// \brief Try to find a partition of the aggregate type passed in for a given
3200 /// offset and size.
3202 /// This recurses through the aggregate type and tries to compute a subtype
3203 /// based on the offset and size. When the offset and size span a sub-section
3204 /// of an array, it will even compute a new array type for that sub-section,
3205 /// and the same for structs.
3207 /// Note that this routine is very strict and tries to find a partition of the
3208 /// type which produces the *exact* right offset and size. It is not forgiving
3209 /// when the size or offset cause either end of type-based partition to be off.
3210 /// Also, this is a best-effort routine. It is reasonable to give up and not
3211 /// return a type if necessary.
3212 static Type *getTypePartition(const DataLayout &DL, Type *Ty, uint64_t Offset,
3214 if (Offset == 0 && DL.getTypeAllocSize(Ty) == Size)
3215 return stripAggregateTypeWrapping(DL, Ty);
3216 if (Offset > DL.getTypeAllocSize(Ty) ||
3217 (DL.getTypeAllocSize(Ty) - Offset) < Size)
3220 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3221 Type *ElementTy = SeqTy->getElementType();
3222 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3223 uint64_t NumSkippedElements = Offset / ElementSize;
3224 if (NumSkippedElements >= SeqTy->getNumElements())
3226 Offset -= NumSkippedElements * ElementSize;
3228 // First check if we need to recurse.
3229 if (Offset > 0 || Size < ElementSize) {
3230 // Bail if the partition ends in a different array element.
3231 if ((Offset + Size) > ElementSize)
3233 // Recurse through the element type trying to peel off offset bytes.
3234 return getTypePartition(DL, ElementTy, Offset, Size);
3236 assert(Offset == 0);
3238 if (Size == ElementSize)
3239 return stripAggregateTypeWrapping(DL, ElementTy);
3240 assert(Size > ElementSize);
3241 uint64_t NumElements = Size / ElementSize;
3242 if (NumElements * ElementSize != Size)
3244 return ArrayType::get(ElementTy, NumElements);
3247 StructType *STy = dyn_cast<StructType>(Ty);
3251 const StructLayout *SL = DL.getStructLayout(STy);
3252 if (Offset >= SL->getSizeInBytes())
3254 uint64_t EndOffset = Offset + Size;
3255 if (EndOffset > SL->getSizeInBytes())
3258 unsigned Index = SL->getElementContainingOffset(Offset);
3259 Offset -= SL->getElementOffset(Index);
3261 Type *ElementTy = STy->getElementType(Index);
3262 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3263 if (Offset >= ElementSize)
3264 return nullptr; // The offset points into alignment padding.
3266 // See if any partition must be contained by the element.
3267 if (Offset > 0 || Size < ElementSize) {
3268 if ((Offset + Size) > ElementSize)
3270 return getTypePartition(DL, ElementTy, Offset, Size);
3272 assert(Offset == 0);
3274 if (Size == ElementSize)
3275 return stripAggregateTypeWrapping(DL, ElementTy);
3277 StructType::element_iterator EI = STy->element_begin() + Index,
3278 EE = STy->element_end();
3279 if (EndOffset < SL->getSizeInBytes()) {
3280 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3281 if (Index == EndIndex)
3282 return nullptr; // Within a single element and its padding.
3284 // Don't try to form "natural" types if the elements don't line up with the
3286 // FIXME: We could potentially recurse down through the last element in the
3287 // sub-struct to find a natural end point.
3288 if (SL->getElementOffset(EndIndex) != EndOffset)
3291 assert(Index < EndIndex);
3292 EE = STy->element_begin() + EndIndex;
3295 // Try to build up a sub-structure.
3297 StructType::get(STy->getContext(), makeArrayRef(EI, EE), STy->isPacked());
3298 const StructLayout *SubSL = DL.getStructLayout(SubTy);
3299 if (Size != SubSL->getSizeInBytes())
3300 return nullptr; // The sub-struct doesn't have quite the size needed.
3305 /// \brief Pre-split loads and stores to simplify rewriting.
3307 /// We want to break up the splittable load+store pairs as much as
3308 /// possible. This is important to do as a preprocessing step, as once we
3309 /// start rewriting the accesses to partitions of the alloca we lose the
3310 /// necessary information to correctly split apart paired loads and stores
3311 /// which both point into this alloca. The case to consider is something like
3314 /// %a = alloca [12 x i8]
3315 /// %gep1 = getelementptr [12 x i8]* %a, i32 0, i32 0
3316 /// %gep2 = getelementptr [12 x i8]* %a, i32 0, i32 4
3317 /// %gep3 = getelementptr [12 x i8]* %a, i32 0, i32 8
3318 /// %iptr1 = bitcast i8* %gep1 to i64*
3319 /// %iptr2 = bitcast i8* %gep2 to i64*
3320 /// %fptr1 = bitcast i8* %gep1 to float*
3321 /// %fptr2 = bitcast i8* %gep2 to float*
3322 /// %fptr3 = bitcast i8* %gep3 to float*
3323 /// store float 0.0, float* %fptr1
3324 /// store float 1.0, float* %fptr2
3325 /// %v = load i64* %iptr1
3326 /// store i64 %v, i64* %iptr2
3327 /// %f1 = load float* %fptr2
3328 /// %f2 = load float* %fptr3
3330 /// Here we want to form 3 partitions of the alloca, each 4 bytes large, and
3331 /// promote everything so we recover the 2 SSA values that should have been
3332 /// there all along.
3334 /// \returns true if any changes are made.
3335 bool SROA::presplitLoadsAndStores(AllocaInst &AI, AllocaSlices &AS) {
3336 DEBUG(dbgs() << "Pre-splitting loads and stores\n");
3338 // Track the loads and stores which are candidates for pre-splitting here, in
3339 // the order they first appear during the partition scan. These give stable
3340 // iteration order and a basis for tracking which loads and stores we
3342 SmallVector<LoadInst *, 4> Loads;
3343 SmallVector<StoreInst *, 4> Stores;
3345 // We need to accumulate the splits required of each load or store where we
3346 // can find them via a direct lookup. This is important to cross-check loads
3347 // and stores against each other. We also track the slice so that we can kill
3348 // all the slices that end up split.
3349 struct SplitOffsets {
3351 std::vector<uint64_t> Splits;
3353 SmallDenseMap<Instruction *, SplitOffsets, 8> SplitOffsetsMap;
3355 // Track loads out of this alloca which cannot, for any reason, be pre-split.
3356 // This is important as we also cannot pre-split stores of those loads!
3357 // FIXME: This is all pretty gross. It means that we can be more aggressive
3358 // in pre-splitting when the load feeding the store happens to come from
3359 // a separate alloca. Put another way, the effectiveness of SROA would be
3360 // decreased by a frontend which just concatenated all of its local allocas
3361 // into one big flat alloca. But defeating such patterns is exactly the job
3362 // SROA is tasked with! Sadly, to not have this discrepancy we would have
3363 // change store pre-splitting to actually force pre-splitting of the load
3364 // that feeds it *and all stores*. That makes pre-splitting much harder, but
3365 // maybe it would make it more principled?
3366 SmallPtrSet<LoadInst *, 8> UnsplittableLoads;
3368 DEBUG(dbgs() << " Searching for candidate loads and stores\n");
3369 for (auto &P : AS.partitions()) {
3370 for (Slice &S : P) {
3371 Instruction *I = cast<Instruction>(S.getUse()->getUser());
3372 if (!S.isSplittable() || S.endOffset() <= P.endOffset()) {
3373 // If this is a load we have to track that it can't participate in any
3374 // pre-splitting. If this is a store of a load we have to track that
3375 // that load also can't participate in any pre-splitting.
3376 if (auto *LI = dyn_cast<LoadInst>(I))
3377 UnsplittableLoads.insert(LI);
3378 else if (auto *SI = dyn_cast<StoreInst>(I))
3379 if (auto *LI = dyn_cast<LoadInst>(SI->getValueOperand()))
3380 UnsplittableLoads.insert(LI);
3383 assert(P.endOffset() > S.beginOffset() &&
3384 "Empty or backwards partition!");
3386 // Determine if this is a pre-splittable slice.
3387 if (auto *LI = dyn_cast<LoadInst>(I)) {
3388 assert(!LI->isVolatile() && "Cannot split volatile loads!");
3390 // The load must be used exclusively to store into other pointers for
3391 // us to be able to arbitrarily pre-split it. The stores must also be
3392 // simple to avoid changing semantics.
3393 auto IsLoadSimplyStored = [](LoadInst *LI) {
3394 for (User *LU : LI->users()) {
3395 auto *SI = dyn_cast<StoreInst>(LU);
3396 if (!SI || !SI->isSimple())
3401 if (!IsLoadSimplyStored(LI)) {
3402 UnsplittableLoads.insert(LI);
3406 Loads.push_back(LI);
3407 } else if (auto *SI = dyn_cast<StoreInst>(I)) {
3408 if (S.getUse() != &SI->getOperandUse(SI->getPointerOperandIndex()))
3409 // Skip stores *of* pointers. FIXME: This shouldn't even be possible!
3411 auto *StoredLoad = dyn_cast<LoadInst>(SI->getValueOperand());
3412 if (!StoredLoad || !StoredLoad->isSimple())
3414 assert(!SI->isVolatile() && "Cannot split volatile stores!");
3416 Stores.push_back(SI);
3418 // Other uses cannot be pre-split.
3422 // Record the initial split.
3423 DEBUG(dbgs() << " Candidate: " << *I << "\n");
3424 auto &Offsets = SplitOffsetsMap[I];
3425 assert(Offsets.Splits.empty() &&
3426 "Should not have splits the first time we see an instruction!");
3428 Offsets.Splits.push_back(P.endOffset() - S.beginOffset());
3431 // Now scan the already split slices, and add a split for any of them which
3432 // we're going to pre-split.
3433 for (Slice *S : P.splitSliceTails()) {
3434 auto SplitOffsetsMapI =
3435 SplitOffsetsMap.find(cast<Instruction>(S->getUse()->getUser()));
3436 if (SplitOffsetsMapI == SplitOffsetsMap.end())
3438 auto &Offsets = SplitOffsetsMapI->second;
3440 assert(Offsets.S == S && "Found a mismatched slice!");
3441 assert(!Offsets.Splits.empty() &&
3442 "Cannot have an empty set of splits on the second partition!");
3443 assert(Offsets.Splits.back() ==
3444 P.beginOffset() - Offsets.S->beginOffset() &&
3445 "Previous split does not end where this one begins!");
3447 // Record each split. The last partition's end isn't needed as the size
3448 // of the slice dictates that.
3449 if (S->endOffset() > P.endOffset())
3450 Offsets.Splits.push_back(P.endOffset() - Offsets.S->beginOffset());
3454 // We may have split loads where some of their stores are split stores. For
3455 // such loads and stores, we can only pre-split them if their splits exactly
3456 // match relative to their starting offset. We have to verify this prior to
3460 [&UnsplittableLoads, &SplitOffsetsMap](StoreInst *SI) {
3461 // Lookup the load we are storing in our map of split
3463 auto *LI = cast<LoadInst>(SI->getValueOperand());
3464 // If it was completely unsplittable, then we're done,
3465 // and this store can't be pre-split.
3466 if (UnsplittableLoads.count(LI))
3469 auto LoadOffsetsI = SplitOffsetsMap.find(LI);
3470 if (LoadOffsetsI == SplitOffsetsMap.end())
3471 return false; // Unrelated loads are definitely safe.
3472 auto &LoadOffsets = LoadOffsetsI->second;
3474 // Now lookup the store's offsets.
3475 auto &StoreOffsets = SplitOffsetsMap[SI];
3477 // If the relative offsets of each split in the load and
3478 // store match exactly, then we can split them and we
3479 // don't need to remove them here.
3480 if (LoadOffsets.Splits == StoreOffsets.Splits)
3483 DEBUG(dbgs() << " Mismatched splits for load and store:\n"
3484 << " " << *LI << "\n"
3485 << " " << *SI << "\n");
3487 // We've found a store and load that we need to split
3488 // with mismatched relative splits. Just give up on them
3489 // and remove both instructions from our list of
3491 UnsplittableLoads.insert(LI);
3495 // Now we have to go *back* through all the stores, because a later store may
3496 // have caused an earlier store's load to become unsplittable and if it is
3497 // unsplittable for the later store, then we can't rely on it being split in
3498 // the earlier store either.
3499 Stores.erase(remove_if(Stores,
3500 [&UnsplittableLoads](StoreInst *SI) {
3501 auto *LI = cast<LoadInst>(SI->getValueOperand());
3502 return UnsplittableLoads.count(LI);
3505 // Once we've established all the loads that can't be split for some reason,
3506 // filter any that made it into our list out.
3507 Loads.erase(remove_if(Loads,
3508 [&UnsplittableLoads](LoadInst *LI) {
3509 return UnsplittableLoads.count(LI);
3513 // If no loads or stores are left, there is no pre-splitting to be done for
3515 if (Loads.empty() && Stores.empty())
3518 // From here on, we can't fail and will be building new accesses, so rig up
3520 IRBuilderTy IRB(&AI);
3522 // Collect the new slices which we will merge into the alloca slices.
3523 SmallVector<Slice, 4> NewSlices;
3525 // Track any allocas we end up splitting loads and stores for so we iterate
3527 SmallPtrSet<AllocaInst *, 4> ResplitPromotableAllocas;
3529 // At this point, we have collected all of the loads and stores we can
3530 // pre-split, and the specific splits needed for them. We actually do the
3531 // splitting in a specific order in order to handle when one of the loads in
3532 // the value operand to one of the stores.
3534 // First, we rewrite all of the split loads, and just accumulate each split
3535 // load in a parallel structure. We also build the slices for them and append
3536 // them to the alloca slices.
3537 SmallDenseMap<LoadInst *, std::vector<LoadInst *>, 1> SplitLoadsMap;
3538 std::vector<LoadInst *> SplitLoads;
3539 const DataLayout &DL = AI.getModule()->getDataLayout();
3540 for (LoadInst *LI : Loads) {
3543 IntegerType *Ty = cast<IntegerType>(LI->getType());
3544 uint64_t LoadSize = Ty->getBitWidth() / 8;
3545 assert(LoadSize > 0 && "Cannot have a zero-sized integer load!");
3547 auto &Offsets = SplitOffsetsMap[LI];
3548 assert(LoadSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&
3549 "Slice size should always match load size exactly!");
3550 uint64_t BaseOffset = Offsets.S->beginOffset();
3551 assert(BaseOffset + LoadSize > BaseOffset &&
3552 "Cannot represent alloca access size using 64-bit integers!");
3554 Instruction *BasePtr = cast<Instruction>(LI->getPointerOperand());
3555 IRB.SetInsertPoint(LI);
3557 DEBUG(dbgs() << " Splitting load: " << *LI << "\n");
3559 uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
3560 int Idx = 0, Size = Offsets.Splits.size();
3562 auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
3563 auto *PartPtrTy = PartTy->getPointerTo(LI->getPointerAddressSpace());
3564 LoadInst *PLoad = IRB.CreateAlignedLoad(
3565 getAdjustedPtr(IRB, DL, BasePtr,
3566 APInt(DL.getPointerSizeInBits(), PartOffset),
3567 PartPtrTy, BasePtr->getName() + "."),
3568 getAdjustedAlignment(LI, PartOffset, DL), /*IsVolatile*/ false,
3570 PLoad->copyMetadata(*LI, LLVMContext::MD_mem_parallel_loop_access);
3572 // Append this load onto the list of split loads so we can find it later
3573 // to rewrite the stores.
3574 SplitLoads.push_back(PLoad);
3576 // Now build a new slice for the alloca.
3577 NewSlices.push_back(
3578 Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
3579 &PLoad->getOperandUse(PLoad->getPointerOperandIndex()),
3580 /*IsSplittable*/ false));
3581 DEBUG(dbgs() << " new slice [" << NewSlices.back().beginOffset()
3582 << ", " << NewSlices.back().endOffset() << "): " << *PLoad
3585 // See if we've handled all the splits.
3589 // Setup the next partition.
3590 PartOffset = Offsets.Splits[Idx];
3592 PartSize = (Idx < Size ? Offsets.Splits[Idx] : LoadSize) - PartOffset;
3595 // Now that we have the split loads, do the slow walk over all uses of the
3596 // load and rewrite them as split stores, or save the split loads to use
3597 // below if the store is going to be split there anyways.
3598 bool DeferredStores = false;
3599 for (User *LU : LI->users()) {
3600 StoreInst *SI = cast<StoreInst>(LU);
3601 if (!Stores.empty() && SplitOffsetsMap.count(SI)) {
3602 DeferredStores = true;
3603 DEBUG(dbgs() << " Deferred splitting of store: " << *SI << "\n");
3607 Value *StoreBasePtr = SI->getPointerOperand();
3608 IRB.SetInsertPoint(SI);
3610 DEBUG(dbgs() << " Splitting store of load: " << *SI << "\n");
3612 for (int Idx = 0, Size = SplitLoads.size(); Idx < Size; ++Idx) {
3613 LoadInst *PLoad = SplitLoads[Idx];
3614 uint64_t PartOffset = Idx == 0 ? 0 : Offsets.Splits[Idx - 1];
3616 PLoad->getType()->getPointerTo(SI->getPointerAddressSpace());
3618 StoreInst *PStore = IRB.CreateAlignedStore(
3619 PLoad, getAdjustedPtr(IRB, DL, StoreBasePtr,
3620 APInt(DL.getPointerSizeInBits(), PartOffset),
3621 PartPtrTy, StoreBasePtr->getName() + "."),
3622 getAdjustedAlignment(SI, PartOffset, DL), /*IsVolatile*/ false);
3623 PStore->copyMetadata(*LI, LLVMContext::MD_mem_parallel_loop_access);
3624 DEBUG(dbgs() << " +" << PartOffset << ":" << *PStore << "\n");
3627 // We want to immediately iterate on any allocas impacted by splitting
3628 // this store, and we have to track any promotable alloca (indicated by
3629 // a direct store) as needing to be resplit because it is no longer
3631 if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(StoreBasePtr)) {
3632 ResplitPromotableAllocas.insert(OtherAI);
3633 Worklist.insert(OtherAI);
3634 } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
3635 StoreBasePtr->stripInBoundsOffsets())) {
3636 Worklist.insert(OtherAI);
3639 // Mark the original store as dead.
3640 DeadInsts.insert(SI);
3643 // Save the split loads if there are deferred stores among the users.
3645 SplitLoadsMap.insert(std::make_pair(LI, std::move(SplitLoads)));
3647 // Mark the original load as dead and kill the original slice.
3648 DeadInsts.insert(LI);
3652 // Second, we rewrite all of the split stores. At this point, we know that
3653 // all loads from this alloca have been split already. For stores of such
3654 // loads, we can simply look up the pre-existing split loads. For stores of
3655 // other loads, we split those loads first and then write split stores of
3657 for (StoreInst *SI : Stores) {
3658 auto *LI = cast<LoadInst>(SI->getValueOperand());
3659 IntegerType *Ty = cast<IntegerType>(LI->getType());
3660 uint64_t StoreSize = Ty->getBitWidth() / 8;
3661 assert(StoreSize > 0 && "Cannot have a zero-sized integer store!");
3663 auto &Offsets = SplitOffsetsMap[SI];
3664 assert(StoreSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&
3665 "Slice size should always match load size exactly!");
3666 uint64_t BaseOffset = Offsets.S->beginOffset();
3667 assert(BaseOffset + StoreSize > BaseOffset &&
3668 "Cannot represent alloca access size using 64-bit integers!");
3670 Value *LoadBasePtr = LI->getPointerOperand();
3671 Instruction *StoreBasePtr = cast<Instruction>(SI->getPointerOperand());
3673 DEBUG(dbgs() << " Splitting store: " << *SI << "\n");
3675 // Check whether we have an already split load.
3676 auto SplitLoadsMapI = SplitLoadsMap.find(LI);
3677 std::vector<LoadInst *> *SplitLoads = nullptr;
3678 if (SplitLoadsMapI != SplitLoadsMap.end()) {
3679 SplitLoads = &SplitLoadsMapI->second;
3680 assert(SplitLoads->size() == Offsets.Splits.size() + 1 &&
3681 "Too few split loads for the number of splits in the store!");
3683 DEBUG(dbgs() << " of load: " << *LI << "\n");
3686 uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
3687 int Idx = 0, Size = Offsets.Splits.size();
3689 auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
3690 auto *PartPtrTy = PartTy->getPointerTo(SI->getPointerAddressSpace());
3692 // Either lookup a split load or create one.
3695 PLoad = (*SplitLoads)[Idx];
3697 IRB.SetInsertPoint(LI);
3698 PLoad = IRB.CreateAlignedLoad(
3699 getAdjustedPtr(IRB, DL, LoadBasePtr,
3700 APInt(DL.getPointerSizeInBits(), PartOffset),
3701 PartPtrTy, LoadBasePtr->getName() + "."),
3702 getAdjustedAlignment(LI, PartOffset, DL), /*IsVolatile*/ false,
3706 // And store this partition.
3707 IRB.SetInsertPoint(SI);
3708 StoreInst *PStore = IRB.CreateAlignedStore(
3709 PLoad, getAdjustedPtr(IRB, DL, StoreBasePtr,
3710 APInt(DL.getPointerSizeInBits(), PartOffset),
3711 PartPtrTy, StoreBasePtr->getName() + "."),
3712 getAdjustedAlignment(SI, PartOffset, DL), /*IsVolatile*/ false);
3714 // Now build a new slice for the alloca.
3715 NewSlices.push_back(
3716 Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
3717 &PStore->getOperandUse(PStore->getPointerOperandIndex()),
3718 /*IsSplittable*/ false));
3719 DEBUG(dbgs() << " new slice [" << NewSlices.back().beginOffset()
3720 << ", " << NewSlices.back().endOffset() << "): " << *PStore
3723 DEBUG(dbgs() << " of split load: " << *PLoad << "\n");
3726 // See if we've finished all the splits.
3730 // Setup the next partition.
3731 PartOffset = Offsets.Splits[Idx];
3733 PartSize = (Idx < Size ? Offsets.Splits[Idx] : StoreSize) - PartOffset;
3736 // We want to immediately iterate on any allocas impacted by splitting
3737 // this load, which is only relevant if it isn't a load of this alloca and
3738 // thus we didn't already split the loads above. We also have to keep track
3739 // of any promotable allocas we split loads on as they can no longer be
3742 if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(LoadBasePtr)) {
3743 assert(OtherAI != &AI && "We can't re-split our own alloca!");
3744 ResplitPromotableAllocas.insert(OtherAI);
3745 Worklist.insert(OtherAI);
3746 } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
3747 LoadBasePtr->stripInBoundsOffsets())) {
3748 assert(OtherAI != &AI && "We can't re-split our own alloca!");
3749 Worklist.insert(OtherAI);
3753 // Mark the original store as dead now that we've split it up and kill its
3754 // slice. Note that we leave the original load in place unless this store
3755 // was its only use. It may in turn be split up if it is an alloca load
3756 // for some other alloca, but it may be a normal load. This may introduce
3757 // redundant loads, but where those can be merged the rest of the optimizer
3758 // should handle the merging, and this uncovers SSA splits which is more
3759 // important. In practice, the original loads will almost always be fully
3760 // split and removed eventually, and the splits will be merged by any
3761 // trivial CSE, including instcombine.
3762 if (LI->hasOneUse()) {
3763 assert(*LI->user_begin() == SI && "Single use isn't this store!");
3764 DeadInsts.insert(LI);
3766 DeadInsts.insert(SI);
3770 // Remove the killed slices that have ben pre-split.
3771 AS.erase(remove_if(AS, [](const Slice &S) { return S.isDead(); }), AS.end());
3773 // Insert our new slices. This will sort and merge them into the sorted
3775 AS.insert(NewSlices);
3777 DEBUG(dbgs() << " Pre-split slices:\n");
3779 for (auto I = AS.begin(), E = AS.end(); I != E; ++I)
3780 DEBUG(AS.print(dbgs(), I, " "));
3783 // Finally, don't try to promote any allocas that new require re-splitting.
3784 // They have already been added to the worklist above.
3785 PromotableAllocas.erase(
3788 [&](AllocaInst *AI) { return ResplitPromotableAllocas.count(AI); }),
3789 PromotableAllocas.end());
3794 /// \brief Rewrite an alloca partition's users.
3796 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3797 /// to rewrite uses of an alloca partition to be conducive for SSA value
3798 /// promotion. If the partition needs a new, more refined alloca, this will
3799 /// build that new alloca, preserving as much type information as possible, and
3800 /// rewrite the uses of the old alloca to point at the new one and have the
3801 /// appropriate new offsets. It also evaluates how successful the rewrite was
3802 /// at enabling promotion and if it was successful queues the alloca to be
3804 AllocaInst *SROA::rewritePartition(AllocaInst &AI, AllocaSlices &AS,
3806 // Try to compute a friendly type for this partition of the alloca. This
3807 // won't always succeed, in which case we fall back to a legal integer type
3808 // or an i8 array of an appropriate size.
3809 Type *SliceTy = nullptr;
3810 const DataLayout &DL = AI.getModule()->getDataLayout();
3811 if (Type *CommonUseTy = findCommonType(P.begin(), P.end(), P.endOffset()))
3812 if (DL.getTypeAllocSize(CommonUseTy) >= P.size())
3813 SliceTy = CommonUseTy;
3815 if (Type *TypePartitionTy = getTypePartition(DL, AI.getAllocatedType(),
3816 P.beginOffset(), P.size()))
3817 SliceTy = TypePartitionTy;
3818 if ((!SliceTy || (SliceTy->isArrayTy() &&
3819 SliceTy->getArrayElementType()->isIntegerTy())) &&
3820 DL.isLegalInteger(P.size() * 8))
3821 SliceTy = Type::getIntNTy(*C, P.size() * 8);
3823 SliceTy = ArrayType::get(Type::getInt8Ty(*C), P.size());
3824 assert(DL.getTypeAllocSize(SliceTy) >= P.size());
3826 bool IsIntegerPromotable = isIntegerWideningViable(P, SliceTy, DL);
3829 IsIntegerPromotable ? nullptr : isVectorPromotionViable(P, DL);
3833 // Check for the case where we're going to rewrite to a new alloca of the
3834 // exact same type as the original, and with the same access offsets. In that
3835 // case, re-use the existing alloca, but still run through the rewriter to
3836 // perform phi and select speculation.
3838 if (SliceTy == AI.getAllocatedType()) {
3839 assert(P.beginOffset() == 0 &&
3840 "Non-zero begin offset but same alloca type");
3842 // FIXME: We should be able to bail at this point with "nothing changed".
3843 // FIXME: We might want to defer PHI speculation until after here.
3844 // FIXME: return nullptr;
3846 unsigned Alignment = AI.getAlignment();
3848 // The minimum alignment which users can rely on when the explicit
3849 // alignment is omitted or zero is that required by the ABI for this
3851 Alignment = DL.getABITypeAlignment(AI.getAllocatedType());
3853 Alignment = MinAlign(Alignment, P.beginOffset());
3854 // If we will get at least this much alignment from the type alone, leave
3855 // the alloca's alignment unconstrained.
3856 if (Alignment <= DL.getABITypeAlignment(SliceTy))
3858 NewAI = new AllocaInst(
3859 SliceTy, nullptr, Alignment,
3860 AI.getName() + ".sroa." + Twine(P.begin() - AS.begin()), &AI);
3864 DEBUG(dbgs() << "Rewriting alloca partition "
3865 << "[" << P.beginOffset() << "," << P.endOffset()
3866 << ") to: " << *NewAI << "\n");
3868 // Track the high watermark on the worklist as it is only relevant for
3869 // promoted allocas. We will reset it to this point if the alloca is not in
3870 // fact scheduled for promotion.
3871 unsigned PPWOldSize = PostPromotionWorklist.size();
3872 unsigned NumUses = 0;
3873 SmallPtrSet<PHINode *, 8> PHIUsers;
3874 SmallPtrSet<SelectInst *, 8> SelectUsers;
3876 AllocaSliceRewriter Rewriter(DL, AS, *this, AI, *NewAI, P.beginOffset(),
3877 P.endOffset(), IsIntegerPromotable, VecTy,
3878 PHIUsers, SelectUsers);
3879 bool Promotable = true;
3880 for (Slice *S : P.splitSliceTails()) {
3881 Promotable &= Rewriter.visit(S);
3884 for (Slice &S : P) {
3885 Promotable &= Rewriter.visit(&S);
3889 NumAllocaPartitionUses += NumUses;
3890 MaxUsesPerAllocaPartition =
3891 std::max<unsigned>(NumUses, MaxUsesPerAllocaPartition);
3893 // Now that we've processed all the slices in the new partition, check if any
3894 // PHIs or Selects would block promotion.
3895 for (SmallPtrSetImpl<PHINode *>::iterator I = PHIUsers.begin(),
3898 if (!isSafePHIToSpeculate(**I)) {
3901 SelectUsers.clear();
3904 for (SmallPtrSetImpl<SelectInst *>::iterator I = SelectUsers.begin(),
3905 E = SelectUsers.end();
3907 if (!isSafeSelectToSpeculate(**I)) {
3910 SelectUsers.clear();
3915 if (PHIUsers.empty() && SelectUsers.empty()) {
3916 // Promote the alloca.
3917 PromotableAllocas.push_back(NewAI);
3919 // If we have either PHIs or Selects to speculate, add them to those
3920 // worklists and re-queue the new alloca so that we promote in on the
3922 for (PHINode *PHIUser : PHIUsers)
3923 SpeculatablePHIs.insert(PHIUser);
3924 for (SelectInst *SelectUser : SelectUsers)
3925 SpeculatableSelects.insert(SelectUser);
3926 Worklist.insert(NewAI);
3929 // Drop any post-promotion work items if promotion didn't happen.
3930 while (PostPromotionWorklist.size() > PPWOldSize)
3931 PostPromotionWorklist.pop_back();
3933 // We couldn't promote and we didn't create a new partition, nothing
3938 // If we can't promote the alloca, iterate on it to check for new
3939 // refinements exposed by splitting the current alloca. Don't iterate on an
3940 // alloca which didn't actually change and didn't get promoted.
3941 Worklist.insert(NewAI);
3947 /// \brief Walks the slices of an alloca and form partitions based on them,
3948 /// rewriting each of their uses.
3949 bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &AS) {
3950 if (AS.begin() == AS.end())
3953 unsigned NumPartitions = 0;
3954 bool Changed = false;
3955 const DataLayout &DL = AI.getModule()->getDataLayout();
3957 // First try to pre-split loads and stores.
3958 Changed |= presplitLoadsAndStores(AI, AS);
3960 // Now that we have identified any pre-splitting opportunities, mark any
3961 // splittable (non-whole-alloca) loads and stores as unsplittable. If we fail
3962 // to split these during pre-splitting, we want to force them to be
3963 // rewritten into a partition.
3964 bool IsSorted = true;
3965 for (Slice &S : AS) {
3966 if (!S.isSplittable())
3968 // FIXME: We currently leave whole-alloca splittable loads and stores. This
3969 // used to be the only splittable loads and stores and we need to be
3970 // confident that the above handling of splittable loads and stores is
3971 // completely sufficient before we forcibly disable the remaining handling.
3972 if (S.beginOffset() == 0 &&
3973 S.endOffset() >= DL.getTypeAllocSize(AI.getAllocatedType()))
3975 if (isa<LoadInst>(S.getUse()->getUser()) ||
3976 isa<StoreInst>(S.getUse()->getUser())) {
3977 S.makeUnsplittable();
3982 std::sort(AS.begin(), AS.end());
3984 /// Describes the allocas introduced by rewritePartition in order to migrate
3990 Fragment(AllocaInst *AI, uint64_t O, uint64_t S)
3991 : Alloca(AI), Offset(O), Size(S) {}
3993 SmallVector<Fragment, 4> Fragments;
3995 // Rewrite each partition.
3996 for (auto &P : AS.partitions()) {
3997 if (AllocaInst *NewAI = rewritePartition(AI, AS, P)) {
4000 uint64_t SizeOfByte = 8;
4001 uint64_t AllocaSize = DL.getTypeSizeInBits(NewAI->getAllocatedType());
4002 // Don't include any padding.
4003 uint64_t Size = std::min(AllocaSize, P.size() * SizeOfByte);
4004 Fragments.push_back(Fragment(NewAI, P.beginOffset() * SizeOfByte, Size));
4010 NumAllocaPartitions += NumPartitions;
4011 MaxPartitionsPerAlloca =
4012 std::max<unsigned>(NumPartitions, MaxPartitionsPerAlloca);
4014 // Migrate debug information from the old alloca to the new alloca(s)
4015 // and the individual partitions.
4016 if (DbgDeclareInst *DbgDecl = FindAllocaDbgDeclare(&AI)) {
4017 auto *Var = DbgDecl->getVariable();
4018 auto *Expr = DbgDecl->getExpression();
4019 DIBuilder DIB(*AI.getModule(), /*AllowUnresolved*/ false);
4020 uint64_t AllocaSize = DL.getTypeSizeInBits(AI.getAllocatedType());
4021 for (auto Fragment : Fragments) {
4022 // Create a fragment expression describing the new partition or reuse AI's
4023 // expression if there is only one partition.
4024 auto *FragmentExpr = Expr;
4025 if (Fragment.Size < AllocaSize || Expr->isFragment()) {
4026 // If this alloca is already a scalar replacement of a larger aggregate,
4027 // Fragment.Offset describes the offset inside the scalar.
4029 Expr->isFragment() ? Expr->getFragmentOffsetInBits() : 0;
4030 uint64_t Start = Offset + Fragment.Offset;
4031 uint64_t Size = Fragment.Size;
4032 if (Expr->isFragment()) {
4034 Expr->getFragmentOffsetInBits() + Expr->getFragmentSizeInBits();
4035 if (Start >= AbsEnd)
4036 // No need to describe a SROAed padding.
4038 Size = std::min(Size, AbsEnd - Start);
4040 FragmentExpr = DIB.createFragmentExpression(Start, Size);
4043 // Remove any existing dbg.declare intrinsic describing the same alloca.
4044 if (DbgDeclareInst *OldDDI = FindAllocaDbgDeclare(Fragment.Alloca))
4045 OldDDI->eraseFromParent();
4047 DIB.insertDeclare(Fragment.Alloca, Var, FragmentExpr,
4048 DbgDecl->getDebugLoc(), &AI);
4054 /// \brief Clobber a use with undef, deleting the used value if it becomes dead.
4055 void SROA::clobberUse(Use &U) {
4057 // Replace the use with an undef value.
4058 U = UndefValue::get(OldV->getType());
4060 // Check for this making an instruction dead. We have to garbage collect
4061 // all the dead instructions to ensure the uses of any alloca end up being
4063 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
4064 if (isInstructionTriviallyDead(OldI)) {
4065 DeadInsts.insert(OldI);
4069 /// \brief Analyze an alloca for SROA.
4071 /// This analyzes the alloca to ensure we can reason about it, builds
4072 /// the slices of the alloca, and then hands it off to be split and
4073 /// rewritten as needed.
4074 bool SROA::runOnAlloca(AllocaInst &AI) {
4075 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
4076 ++NumAllocasAnalyzed;
4078 // Special case dead allocas, as they're trivial.
4079 if (AI.use_empty()) {
4080 AI.eraseFromParent();
4083 const DataLayout &DL = AI.getModule()->getDataLayout();
4085 // Skip alloca forms that this analysis can't handle.
4086 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
4087 DL.getTypeAllocSize(AI.getAllocatedType()) == 0)
4090 bool Changed = false;
4092 // First, split any FCA loads and stores touching this alloca to promote
4093 // better splitting and promotion opportunities.
4094 AggLoadStoreRewriter AggRewriter;
4095 Changed |= AggRewriter.rewrite(AI);
4097 // Build the slices using a recursive instruction-visiting builder.
4098 AllocaSlices AS(DL, AI);
4099 DEBUG(AS.print(dbgs()));
4103 // Delete all the dead users of this alloca before splitting and rewriting it.
4104 for (Instruction *DeadUser : AS.getDeadUsers()) {
4105 // Free up everything used by this instruction.
4106 for (Use &DeadOp : DeadUser->operands())
4109 // Now replace the uses of this instruction.
4110 DeadUser->replaceAllUsesWith(UndefValue::get(DeadUser->getType()));
4112 // And mark it for deletion.
4113 DeadInsts.insert(DeadUser);
4116 for (Use *DeadOp : AS.getDeadOperands()) {
4117 clobberUse(*DeadOp);
4121 // No slices to split. Leave the dead alloca for a later pass to clean up.
4122 if (AS.begin() == AS.end())
4125 Changed |= splitAlloca(AI, AS);
4127 DEBUG(dbgs() << " Speculating PHIs\n");
4128 while (!SpeculatablePHIs.empty())
4129 speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
4131 DEBUG(dbgs() << " Speculating Selects\n");
4132 while (!SpeculatableSelects.empty())
4133 speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
4138 /// \brief Delete the dead instructions accumulated in this run.
4140 /// Recursively deletes the dead instructions we've accumulated. This is done
4141 /// at the very end to maximize locality of the recursive delete and to
4142 /// minimize the problems of invalidated instruction pointers as such pointers
4143 /// are used heavily in the intermediate stages of the algorithm.
4145 /// We also record the alloca instructions deleted here so that they aren't
4146 /// subsequently handed to mem2reg to promote.
4147 void SROA::deleteDeadInstructions(
4148 SmallPtrSetImpl<AllocaInst *> &DeletedAllocas) {
4149 while (!DeadInsts.empty()) {
4150 Instruction *I = DeadInsts.pop_back_val();
4151 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
4153 I->replaceAllUsesWith(UndefValue::get(I->getType()));
4155 for (Use &Operand : I->operands())
4156 if (Instruction *U = dyn_cast<Instruction>(Operand)) {
4157 // Zero out the operand and see if it becomes trivially dead.
4159 if (isInstructionTriviallyDead(U))
4160 DeadInsts.insert(U);
4163 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
4164 DeletedAllocas.insert(AI);
4165 if (DbgDeclareInst *DbgDecl = FindAllocaDbgDeclare(AI))
4166 DbgDecl->eraseFromParent();
4170 I->eraseFromParent();
4174 /// \brief Promote the allocas, using the best available technique.
4176 /// This attempts to promote whatever allocas have been identified as viable in
4177 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
4178 /// This function returns whether any promotion occurred.
4179 bool SROA::promoteAllocas(Function &F) {
4180 if (PromotableAllocas.empty())
4183 NumPromoted += PromotableAllocas.size();
4185 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
4186 PromoteMemToReg(PromotableAllocas, *DT, nullptr);
4187 PromotableAllocas.clear();
4191 PreservedAnalyses SROA::runImpl(Function &F, DominatorTree &RunDT) {
4192 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
4193 C = &F.getContext();
4196 BasicBlock &EntryBB = F.getEntryBlock();
4197 for (BasicBlock::iterator I = EntryBB.begin(), E = std::prev(EntryBB.end());
4199 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
4200 Worklist.insert(AI);
4203 bool Changed = false;
4204 // A set of deleted alloca instruction pointers which should be removed from
4205 // the list of promotable allocas.
4206 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
4209 while (!Worklist.empty()) {
4210 Changed |= runOnAlloca(*Worklist.pop_back_val());
4211 deleteDeadInstructions(DeletedAllocas);
4213 // Remove the deleted allocas from various lists so that we don't try to
4214 // continue processing them.
4215 if (!DeletedAllocas.empty()) {
4216 auto IsInSet = [&](AllocaInst *AI) { return DeletedAllocas.count(AI); };
4217 Worklist.remove_if(IsInSet);
4218 PostPromotionWorklist.remove_if(IsInSet);
4219 PromotableAllocas.erase(remove_if(PromotableAllocas, IsInSet),
4220 PromotableAllocas.end());
4221 DeletedAllocas.clear();
4225 Changed |= promoteAllocas(F);
4227 Worklist = PostPromotionWorklist;
4228 PostPromotionWorklist.clear();
4229 } while (!Worklist.empty());
4232 return PreservedAnalyses::all();
4234 // FIXME: Even when promoting allocas we should preserve some abstract set of
4235 // CFG-specific analyses.
4236 PreservedAnalyses PA;
4237 PA.preserve<GlobalsAA>();
4241 PreservedAnalyses SROA::run(Function &F, FunctionAnalysisManager &AM) {
4242 return runImpl(F, AM.getResult<DominatorTreeAnalysis>(F));
4245 /// A legacy pass for the legacy pass manager that wraps the \c SROA pass.
4247 /// This is in the llvm namespace purely to allow it to be a friend of the \c
4249 class llvm::sroa::SROALegacyPass : public FunctionPass {
4250 /// The SROA implementation.
4254 SROALegacyPass() : FunctionPass(ID) {
4255 initializeSROALegacyPassPass(*PassRegistry::getPassRegistry());
4257 bool runOnFunction(Function &F) override {
4258 if (skipFunction(F))
4261 auto PA = Impl.runImpl(
4262 F, getAnalysis<DominatorTreeWrapperPass>().getDomTree());
4263 return !PA.areAllPreserved();
4265 void getAnalysisUsage(AnalysisUsage &AU) const override {
4266 AU.addRequired<DominatorTreeWrapperPass>();
4267 AU.addPreserved<GlobalsAAWrapperPass>();
4268 AU.setPreservesCFG();
4271 StringRef getPassName() const override { return "SROA"; }
4275 char SROALegacyPass::ID = 0;
4277 FunctionPass *llvm::createSROAPass() { return new SROALegacyPass(); }
4279 INITIALIZE_PASS_BEGIN(SROALegacyPass, "sroa",
4280 "Scalar Replacement Of Aggregates", false, false)
4281 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
4282 INITIALIZE_PASS_END(SROALegacyPass, "sroa", "Scalar Replacement Of Aggregates",