LLVM  9.0.0svn
SROA.cpp
Go to the documentation of this file.
1 //===- SROA.cpp - Scalar Replacement Of Aggregates ------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 /// \file
9 /// This transformation implements the well known scalar replacement of
10 /// aggregates transformation. It tries to identify promotable elements of an
11 /// aggregate alloca, and promote them to registers. It will also try to
12 /// convert uses of an element (or set of elements) of an alloca into a vector
13 /// or bitfield-style integer scalar if appropriate.
14 ///
15 /// It works to do this with minimal slicing of the alloca so that regions
16 /// which are merely transferred in and out of external memory remain unchanged
17 /// and are not decomposed to scalar code.
18 ///
19 /// Because this also performs alloca promotion, it can be thought of as also
20 /// serving the purpose of SSA formation. The algorithm iterates on the
21 /// function until all opportunities for promotion have been realized.
22 ///
23 //===----------------------------------------------------------------------===//
24 
26 #include "llvm/ADT/APInt.h"
27 #include "llvm/ADT/ArrayRef.h"
28 #include "llvm/ADT/DenseMap.h"
30 #include "llvm/ADT/STLExtras.h"
31 #include "llvm/ADT/SetVector.h"
33 #include "llvm/ADT/SmallPtrSet.h"
34 #include "llvm/ADT/SmallVector.h"
35 #include "llvm/ADT/Statistic.h"
36 #include "llvm/ADT/StringRef.h"
37 #include "llvm/ADT/Twine.h"
38 #include "llvm/ADT/iterator.h"
42 #include "llvm/Analysis/Loads.h"
45 #include "llvm/Config/llvm-config.h"
46 #include "llvm/IR/BasicBlock.h"
47 #include "llvm/IR/Constant.h"
48 #include "llvm/IR/ConstantFolder.h"
49 #include "llvm/IR/Constants.h"
50 #include "llvm/IR/DIBuilder.h"
51 #include "llvm/IR/DataLayout.h"
53 #include "llvm/IR/DerivedTypes.h"
54 #include "llvm/IR/Dominators.h"
55 #include "llvm/IR/Function.h"
57 #include "llvm/IR/GlobalAlias.h"
58 #include "llvm/IR/IRBuilder.h"
59 #include "llvm/IR/InstVisitor.h"
60 #include "llvm/IR/InstrTypes.h"
61 #include "llvm/IR/Instruction.h"
62 #include "llvm/IR/Instructions.h"
63 #include "llvm/IR/IntrinsicInst.h"
64 #include "llvm/IR/Intrinsics.h"
65 #include "llvm/IR/LLVMContext.h"
66 #include "llvm/IR/Metadata.h"
67 #include "llvm/IR/Module.h"
68 #include "llvm/IR/Operator.h"
69 #include "llvm/IR/PassManager.h"
70 #include "llvm/IR/Type.h"
71 #include "llvm/IR/Use.h"
72 #include "llvm/IR/User.h"
73 #include "llvm/IR/Value.h"
74 #include "llvm/Pass.h"
75 #include "llvm/Support/Casting.h"
77 #include "llvm/Support/Compiler.h"
78 #include "llvm/Support/Debug.h"
82 #include "llvm/Transforms/Scalar.h"
84 #include <algorithm>
85 #include <cassert>
86 #include <chrono>
87 #include <cstddef>
88 #include <cstdint>
89 #include <cstring>
90 #include <iterator>
91 #include <string>
92 #include <tuple>
93 #include <utility>
94 #include <vector>
95 
96 #ifndef NDEBUG
97 // We only use this for a debug check.
98 #include <random>
99 #endif
100 
101 using namespace llvm;
102 using namespace llvm::sroa;
103 
104 #define DEBUG_TYPE "sroa"
105 
106 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
107 STATISTIC(NumAllocaPartitions, "Number of alloca partitions formed");
108 STATISTIC(MaxPartitionsPerAlloca, "Maximum number of partitions per alloca");
109 STATISTIC(NumAllocaPartitionUses, "Number of alloca partition uses rewritten");
110 STATISTIC(MaxUsesPerAllocaPartition, "Maximum number of uses of a partition");
111 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
112 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
113 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
114 STATISTIC(NumDeleted, "Number of instructions deleted");
115 STATISTIC(NumVectorized, "Number of vectorized aggregates");
116 
117 /// Hidden option to enable randomly shuffling the slices to help uncover
118 /// instability in their order.
119 static cl::opt<bool> SROARandomShuffleSlices("sroa-random-shuffle-slices",
120  cl::init(false), cl::Hidden);
121 
122 /// Hidden option to experiment with completely strict handling of inbounds
123 /// GEPs.
124 static cl::opt<bool> SROAStrictInbounds("sroa-strict-inbounds", cl::init(false),
125  cl::Hidden);
126 
127 namespace {
128 
129 /// A custom IRBuilder inserter which prefixes all names, but only in
130 /// Assert builds.
131 class IRBuilderPrefixedInserter : public IRBuilderDefaultInserter {
132  std::string Prefix;
133 
134  const Twine getNameWithPrefix(const Twine &Name) const {
135  return Name.isTriviallyEmpty() ? Name : Prefix + Name;
136  }
137 
138 public:
139  void SetNamePrefix(const Twine &P) { Prefix = P.str(); }
140 
141 protected:
142  void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
143  BasicBlock::iterator InsertPt) const {
144  IRBuilderDefaultInserter::InsertHelper(I, getNameWithPrefix(Name), BB,
145  InsertPt);
146  }
147 };
148 
149 /// Provide a type for IRBuilder that drops names in release builds.
151 
152 /// A used slice of an alloca.
153 ///
154 /// This structure represents a slice of an alloca used by some instruction. It
155 /// stores both the begin and end offsets of this use, a pointer to the use
156 /// itself, and a flag indicating whether we can classify the use as splittable
157 /// or not when forming partitions of the alloca.
158 class Slice {
159  /// The beginning offset of the range.
160  uint64_t BeginOffset = 0;
161 
162  /// The ending offset, not included in the range.
163  uint64_t EndOffset = 0;
164 
165  /// Storage for both the use of this slice and whether it can be
166  /// split.
167  PointerIntPair<Use *, 1, bool> UseAndIsSplittable;
168 
169 public:
170  Slice() = default;
171 
172  Slice(uint64_t BeginOffset, uint64_t EndOffset, Use *U, bool IsSplittable)
173  : BeginOffset(BeginOffset), EndOffset(EndOffset),
174  UseAndIsSplittable(U, IsSplittable) {}
175 
176  uint64_t beginOffset() const { return BeginOffset; }
177  uint64_t endOffset() const { return EndOffset; }
178 
179  bool isSplittable() const { return UseAndIsSplittable.getInt(); }
180  void makeUnsplittable() { UseAndIsSplittable.setInt(false); }
181 
182  Use *getUse() const { return UseAndIsSplittable.getPointer(); }
183 
184  bool isDead() const { return getUse() == nullptr; }
185  void kill() { UseAndIsSplittable.setPointer(nullptr); }
186 
187  /// Support for ordering ranges.
188  ///
189  /// This provides an ordering over ranges such that start offsets are
190  /// always increasing, and within equal start offsets, the end offsets are
191  /// decreasing. Thus the spanning range comes first in a cluster with the
192  /// same start position.
193  bool operator<(const Slice &RHS) const {
194  if (beginOffset() < RHS.beginOffset())
195  return true;
196  if (beginOffset() > RHS.beginOffset())
197  return false;
198  if (isSplittable() != RHS.isSplittable())
199  return !isSplittable();
200  if (endOffset() > RHS.endOffset())
201  return true;
202  return false;
203  }
204 
205  /// Support comparison with a single offset to allow binary searches.
206  friend LLVM_ATTRIBUTE_UNUSED bool operator<(const Slice &LHS,
207  uint64_t RHSOffset) {
208  return LHS.beginOffset() < RHSOffset;
209  }
210  friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
211  const Slice &RHS) {
212  return LHSOffset < RHS.beginOffset();
213  }
214 
215  bool operator==(const Slice &RHS) const {
216  return isSplittable() == RHS.isSplittable() &&
217  beginOffset() == RHS.beginOffset() && endOffset() == RHS.endOffset();
218  }
219  bool operator!=(const Slice &RHS) const { return !operator==(RHS); }
220 };
221 
222 } // end anonymous namespace
223 
224 /// Representation of the alloca slices.
225 ///
226 /// This class represents the slices of an alloca which are formed by its
227 /// various uses. If a pointer escapes, we can't fully build a representation
228 /// for the slices used and we reflect that in this structure. The uses are
229 /// stored, sorted by increasing beginning offset and with unsplittable slices
230 /// starting at a particular offset before splittable slices.
232 public:
233  /// Construct the slices of a particular alloca.
234  AllocaSlices(const DataLayout &DL, AllocaInst &AI);
235 
236  /// Test whether a pointer to the allocation escapes our analysis.
237  ///
238  /// If this is true, the slices are never fully built and should be
239  /// ignored.
240  bool isEscaped() const { return PointerEscapingInstr; }
241 
242  /// Support for iterating over the slices.
243  /// @{
246 
247  iterator begin() { return Slices.begin(); }
248  iterator end() { return Slices.end(); }
249 
252 
253  const_iterator begin() const { return Slices.begin(); }
254  const_iterator end() const { return Slices.end(); }
255  /// @}
256 
257  /// Erase a range of slices.
258  void erase(iterator Start, iterator Stop) { Slices.erase(Start, Stop); }
259 
260  /// Insert new slices for this alloca.
261  ///
262  /// This moves the slices into the alloca's slices collection, and re-sorts
263  /// everything so that the usual ordering properties of the alloca's slices
264  /// hold.
265  void insert(ArrayRef<Slice> NewSlices) {
266  int OldSize = Slices.size();
267  Slices.append(NewSlices.begin(), NewSlices.end());
268  auto SliceI = Slices.begin() + OldSize;
269  llvm::sort(SliceI, Slices.end());
270  std::inplace_merge(Slices.begin(), SliceI, Slices.end());
271  }
272 
273  // Forward declare the iterator and range accessor for walking the
274  // partitions.
275  class partition_iterator;
277 
278  /// Access the dead users for this alloca.
279  ArrayRef<Instruction *> getDeadUsers() const { return DeadUsers; }
280 
281  /// Access the dead operands referring to this alloca.
282  ///
283  /// These are operands which have cannot actually be used to refer to the
284  /// alloca as they are outside its range and the user doesn't correct for
285  /// that. These mostly consist of PHI node inputs and the like which we just
286  /// need to replace with undef.
287  ArrayRef<Use *> getDeadOperands() const { return DeadOperands; }
288 
289 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
290  void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
291  void printSlice(raw_ostream &OS, const_iterator I,
292  StringRef Indent = " ") const;
293  void printUse(raw_ostream &OS, const_iterator I,
294  StringRef Indent = " ") const;
295  void print(raw_ostream &OS) const;
296  void dump(const_iterator I) const;
297  void dump() const;
298 #endif
299 
300 private:
301  template <typename DerivedT, typename RetT = void> class BuilderBase;
302  class SliceBuilder;
303 
304  friend class AllocaSlices::SliceBuilder;
305 
306 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
307  /// Handle to alloca instruction to simplify method interfaces.
308  AllocaInst &AI;
309 #endif
310 
311  /// The instruction responsible for this alloca not having a known set
312  /// of slices.
313  ///
314  /// When an instruction (potentially) escapes the pointer to the alloca, we
315  /// store a pointer to that here and abort trying to form slices of the
316  /// alloca. This will be null if the alloca slices are analyzed successfully.
317  Instruction *PointerEscapingInstr;
318 
319  /// The slices of the alloca.
320  ///
321  /// We store a vector of the slices formed by uses of the alloca here. This
322  /// vector is sorted by increasing begin offset, and then the unsplittable
323  /// slices before the splittable ones. See the Slice inner class for more
324  /// details.
325  SmallVector<Slice, 8> Slices;
326 
327  /// Instructions which will become dead if we rewrite the alloca.
328  ///
329  /// Note that these are not separated by slice. This is because we expect an
330  /// alloca to be completely rewritten or not rewritten at all. If rewritten,
331  /// all these instructions can simply be removed and replaced with undef as
332  /// they come from outside of the allocated space.
334 
335  /// Operands which will become dead if we rewrite the alloca.
336  ///
337  /// These are operands that in their particular use can be replaced with
338  /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
339  /// to PHI nodes and the like. They aren't entirely dead (there might be
340  /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
341  /// want to swap this particular input for undef to simplify the use lists of
342  /// the alloca.
343  SmallVector<Use *, 8> DeadOperands;
344 };
345 
346 /// A partition of the slices.
347 ///
348 /// An ephemeral representation for a range of slices which can be viewed as
349 /// a partition of the alloca. This range represents a span of the alloca's
350 /// memory which cannot be split, and provides access to all of the slices
351 /// overlapping some part of the partition.
352 ///
353 /// Objects of this type are produced by traversing the alloca's slices, but
354 /// are only ephemeral and not persistent.
356 private:
357  friend class AllocaSlices;
359 
360  using iterator = AllocaSlices::iterator;
361 
362  /// The beginning and ending offsets of the alloca for this
363  /// partition.
364  uint64_t BeginOffset, EndOffset;
365 
366  /// The start and end iterators of this partition.
367  iterator SI, SJ;
368 
369  /// A collection of split slice tails overlapping the partition.
370  SmallVector<Slice *, 4> SplitTails;
371 
372  /// Raw constructor builds an empty partition starting and ending at
373  /// the given iterator.
374  Partition(iterator SI) : SI(SI), SJ(SI) {}
375 
376 public:
377  /// The start offset of this partition.
378  ///
379  /// All of the contained slices start at or after this offset.
380  uint64_t beginOffset() const { return BeginOffset; }
381 
382  /// The end offset of this partition.
383  ///
384  /// All of the contained slices end at or before this offset.
385  uint64_t endOffset() const { return EndOffset; }
386 
387  /// The size of the partition.
388  ///
389  /// Note that this can never be zero.
390  uint64_t size() const {
391  assert(BeginOffset < EndOffset && "Partitions must span some bytes!");
392  return EndOffset - BeginOffset;
393  }
394 
395  /// Test whether this partition contains no slices, and merely spans
396  /// a region occupied by split slices.
397  bool empty() const { return SI == SJ; }
398 
399  /// \name Iterate slices that start within the partition.
400  /// These may be splittable or unsplittable. They have a begin offset >= the
401  /// partition begin offset.
402  /// @{
403  // FIXME: We should probably define a "concat_iterator" helper and use that
404  // to stitch together pointee_iterators over the split tails and the
405  // contiguous iterators of the partition. That would give a much nicer
406  // interface here. We could then additionally expose filtered iterators for
407  // split, unsplit, and unsplittable splices based on the usage patterns.
408  iterator begin() const { return SI; }
409  iterator end() const { return SJ; }
410  /// @}
411 
412  /// Get the sequence of split slice tails.
413  ///
414  /// These tails are of slices which start before this partition but are
415  /// split and overlap into the partition. We accumulate these while forming
416  /// partitions.
417  ArrayRef<Slice *> splitSliceTails() const { return SplitTails; }
418 };
419 
420 /// An iterator over partitions of the alloca's slices.
421 ///
422 /// This iterator implements the core algorithm for partitioning the alloca's
423 /// slices. It is a forward iterator as we don't support backtracking for
424 /// efficiency reasons, and re-use a single storage area to maintain the
425 /// current set of split slices.
426 ///
427 /// It is templated on the slice iterator type to use so that it can operate
428 /// with either const or non-const slice iterators.
430  : public iterator_facade_base<partition_iterator, std::forward_iterator_tag,
431  Partition> {
432  friend class AllocaSlices;
433 
434  /// Most of the state for walking the partitions is held in a class
435  /// with a nice interface for examining them.
436  Partition P;
437 
438  /// We need to keep the end of the slices to know when to stop.
440 
441  /// We also need to keep track of the maximum split end offset seen.
442  /// FIXME: Do we really?
443  uint64_t MaxSplitSliceEndOffset = 0;
444 
445  /// Sets the partition to be empty at given iterator, and sets the
446  /// end iterator.
448  : P(SI), SE(SE) {
449  // If not already at the end, advance our state to form the initial
450  // partition.
451  if (SI != SE)
452  advance();
453  }
454 
455  /// Advance the iterator to the next partition.
456  ///
457  /// Requires that the iterator not be at the end of the slices.
458  void advance() {
459  assert((P.SI != SE || !P.SplitTails.empty()) &&
460  "Cannot advance past the end of the slices!");
461 
462  // Clear out any split uses which have ended.
463  if (!P.SplitTails.empty()) {
464  if (P.EndOffset >= MaxSplitSliceEndOffset) {
465  // If we've finished all splits, this is easy.
466  P.SplitTails.clear();
467  MaxSplitSliceEndOffset = 0;
468  } else {
469  // Remove the uses which have ended in the prior partition. This
470  // cannot change the max split slice end because we just checked that
471  // the prior partition ended prior to that max.
472  P.SplitTails.erase(llvm::remove_if(P.SplitTails,
473  [&](Slice *S) {
474  return S->endOffset() <=
475  P.EndOffset;
476  }),
477  P.SplitTails.end());
478  assert(llvm::any_of(P.SplitTails,
479  [&](Slice *S) {
480  return S->endOffset() == MaxSplitSliceEndOffset;
481  }) &&
482  "Could not find the current max split slice offset!");
483  assert(llvm::all_of(P.SplitTails,
484  [&](Slice *S) {
485  return S->endOffset() <= MaxSplitSliceEndOffset;
486  }) &&
487  "Max split slice end offset is not actually the max!");
488  }
489  }
490 
491  // If P.SI is already at the end, then we've cleared the split tail and
492  // now have an end iterator.
493  if (P.SI == SE) {
494  assert(P.SplitTails.empty() && "Failed to clear the split slices!");
495  return;
496  }
497 
498  // If we had a non-empty partition previously, set up the state for
499  // subsequent partitions.
500  if (P.SI != P.SJ) {
501  // Accumulate all the splittable slices which started in the old
502  // partition into the split list.
503  for (Slice &S : P)
504  if (S.isSplittable() && S.endOffset() > P.EndOffset) {
505  P.SplitTails.push_back(&S);
506  MaxSplitSliceEndOffset =
507  std::max(S.endOffset(), MaxSplitSliceEndOffset);
508  }
509 
510  // Start from the end of the previous partition.
511  P.SI = P.SJ;
512 
513  // If P.SI is now at the end, we at most have a tail of split slices.
514  if (P.SI == SE) {
515  P.BeginOffset = P.EndOffset;
516  P.EndOffset = MaxSplitSliceEndOffset;
517  return;
518  }
519 
520  // If the we have split slices and the next slice is after a gap and is
521  // not splittable immediately form an empty partition for the split
522  // slices up until the next slice begins.
523  if (!P.SplitTails.empty() && P.SI->beginOffset() != P.EndOffset &&
524  !P.SI->isSplittable()) {
525  P.BeginOffset = P.EndOffset;
526  P.EndOffset = P.SI->beginOffset();
527  return;
528  }
529  }
530 
531  // OK, we need to consume new slices. Set the end offset based on the
532  // current slice, and step SJ past it. The beginning offset of the
533  // partition is the beginning offset of the next slice unless we have
534  // pre-existing split slices that are continuing, in which case we begin
535  // at the prior end offset.
536  P.BeginOffset = P.SplitTails.empty() ? P.SI->beginOffset() : P.EndOffset;
537  P.EndOffset = P.SI->endOffset();
538  ++P.SJ;
539 
540  // There are two strategies to form a partition based on whether the
541  // partition starts with an unsplittable slice or a splittable slice.
542  if (!P.SI->isSplittable()) {
543  // When we're forming an unsplittable region, it must always start at
544  // the first slice and will extend through its end.
545  assert(P.BeginOffset == P.SI->beginOffset());
546 
547  // Form a partition including all of the overlapping slices with this
548  // unsplittable slice.
549  while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
550  if (!P.SJ->isSplittable())
551  P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
552  ++P.SJ;
553  }
554 
555  // We have a partition across a set of overlapping unsplittable
556  // partitions.
557  return;
558  }
559 
560  // If we're starting with a splittable slice, then we need to form
561  // a synthetic partition spanning it and any other overlapping splittable
562  // splices.
563  assert(P.SI->isSplittable() && "Forming a splittable partition!");
564 
565  // Collect all of the overlapping splittable slices.
566  while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset &&
567  P.SJ->isSplittable()) {
568  P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
569  ++P.SJ;
570  }
571 
572  // Back upiP.EndOffset if we ended the span early when encountering an
573  // unsplittable slice. This synthesizes the early end offset of
574  // a partition spanning only splittable slices.
575  if (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
576  assert(!P.SJ->isSplittable());
577  P.EndOffset = P.SJ->beginOffset();
578  }
579  }
580 
581 public:
582  bool operator==(const partition_iterator &RHS) const {
583  assert(SE == RHS.SE &&
584  "End iterators don't match between compared partition iterators!");
585 
586  // The observed positions of partitions is marked by the P.SI iterator and
587  // the emptiness of the split slices. The latter is only relevant when
588  // P.SI == SE, as the end iterator will additionally have an empty split
589  // slices list, but the prior may have the same P.SI and a tail of split
590  // slices.
591  if (P.SI == RHS.P.SI && P.SplitTails.empty() == RHS.P.SplitTails.empty()) {
592  assert(P.SJ == RHS.P.SJ &&
593  "Same set of slices formed two different sized partitions!");
594  assert(P.SplitTails.size() == RHS.P.SplitTails.size() &&
595  "Same slice position with differently sized non-empty split "
596  "slice tails!");
597  return true;
598  }
599  return false;
600  }
601 
603  advance();
604  return *this;
605  }
606 
607  Partition &operator*() { return P; }
608 };
609 
610 /// A forward range over the partitions of the alloca's slices.
611 ///
612 /// This accesses an iterator range over the partitions of the alloca's
613 /// slices. It computes these partitions on the fly based on the overlapping
614 /// offsets of the slices and the ability to split them. It will visit "empty"
615 /// partitions to cover regions of the alloca only accessed via split
616 /// slices.
618  return make_range(partition_iterator(begin(), end()),
619  partition_iterator(end(), end()));
620 }
621 
623  // If the condition being selected on is a constant or the same value is
624  // being selected between, fold the select. Yes this does (rarely) happen
625  // early on.
626  if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
627  return SI.getOperand(1 + CI->isZero());
628  if (SI.getOperand(1) == SI.getOperand(2))
629  return SI.getOperand(1);
630 
631  return nullptr;
632 }
633 
634 /// A helper that folds a PHI node or a select.
636  if (PHINode *PN = dyn_cast<PHINode>(&I)) {
637  // If PN merges together the same value, return that value.
638  return PN->hasConstantValue();
639  }
640  return foldSelectInst(cast<SelectInst>(I));
641 }
642 
643 /// Builder for the alloca slices.
644 ///
645 /// This class builds a set of alloca slices by recursively visiting the uses
646 /// of an alloca and making a slice for each load and store at each offset.
647 class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
649  friend class InstVisitor<SliceBuilder>;
650 
652 
653  const uint64_t AllocSize;
654  AllocaSlices &AS;
655 
656  SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
658 
659  /// Set to de-duplicate dead instructions found in the use walk.
660  SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
661 
662 public:
665  AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())), AS(AS) {}
666 
667 private:
668  void markAsDead(Instruction &I) {
669  if (VisitedDeadInsts.insert(&I).second)
670  AS.DeadUsers.push_back(&I);
671  }
672 
673  void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
674  bool IsSplittable = false) {
675  // Completely skip uses which have a zero size or start either before or
676  // past the end of the allocation.
677  if (Size == 0 || Offset.uge(AllocSize)) {
678  LLVM_DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @"
679  << Offset
680  << " which has zero size or starts outside of the "
681  << AllocSize << " byte alloca:\n"
682  << " alloca: " << AS.AI << "\n"
683  << " use: " << I << "\n");
684  return markAsDead(I);
685  }
686 
687  uint64_t BeginOffset = Offset.getZExtValue();
688  uint64_t EndOffset = BeginOffset + Size;
689 
690  // Clamp the end offset to the end of the allocation. Note that this is
691  // formulated to handle even the case where "BeginOffset + Size" overflows.
692  // This may appear superficially to be something we could ignore entirely,
693  // but that is not so! There may be widened loads or PHI-node uses where
694  // some instructions are dead but not others. We can't completely ignore
695  // them, and so have to record at least the information here.
696  assert(AllocSize >= BeginOffset); // Established above.
697  if (Size > AllocSize - BeginOffset) {
698  LLVM_DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @"
699  << Offset << " to remain within the " << AllocSize
700  << " byte alloca:\n"
701  << " alloca: " << AS.AI << "\n"
702  << " use: " << I << "\n");
703  EndOffset = AllocSize;
704  }
705 
706  AS.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
707  }
708 
709  void visitBitCastInst(BitCastInst &BC) {
710  if (BC.use_empty())
711  return markAsDead(BC);
712 
713  return Base::visitBitCastInst(BC);
714  }
715 
716  void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
717  if (GEPI.use_empty())
718  return markAsDead(GEPI);
719 
720  if (SROAStrictInbounds && GEPI.isInBounds()) {
721  // FIXME: This is a manually un-factored variant of the basic code inside
722  // of GEPs with checking of the inbounds invariant specified in the
723  // langref in a very strict sense. If we ever want to enable
724  // SROAStrictInbounds, this code should be factored cleanly into
725  // PtrUseVisitor, but it is easier to experiment with SROAStrictInbounds
726  // by writing out the code here where we have the underlying allocation
727  // size readily available.
728  APInt GEPOffset = Offset;
729  const DataLayout &DL = GEPI.getModule()->getDataLayout();
730  for (gep_type_iterator GTI = gep_type_begin(GEPI),
731  GTE = gep_type_end(GEPI);
732  GTI != GTE; ++GTI) {
733  ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
734  if (!OpC)
735  break;
736 
737  // Handle a struct index, which adds its field offset to the pointer.
738  if (StructType *STy = GTI.getStructTypeOrNull()) {
739  unsigned ElementIdx = OpC->getZExtValue();
740  const StructLayout *SL = DL.getStructLayout(STy);
741  GEPOffset +=
742  APInt(Offset.getBitWidth(), SL->getElementOffset(ElementIdx));
743  } else {
744  // For array or vector indices, scale the index by the size of the
745  // type.
746  APInt Index = OpC->getValue().sextOrTrunc(Offset.getBitWidth());
747  GEPOffset += Index * APInt(Offset.getBitWidth(),
748  DL.getTypeAllocSize(GTI.getIndexedType()));
749  }
750 
751  // If this index has computed an intermediate pointer which is not
752  // inbounds, then the result of the GEP is a poison value and we can
753  // delete it and all uses.
754  if (GEPOffset.ugt(AllocSize))
755  return markAsDead(GEPI);
756  }
757  }
758 
759  return Base::visitGetElementPtrInst(GEPI);
760  }
761 
762  void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
763  uint64_t Size, bool IsVolatile) {
764  // We allow splitting of non-volatile loads and stores where the type is an
765  // integer type. These may be used to implement 'memcpy' or other "transfer
766  // of bits" patterns.
767  bool IsSplittable = Ty->isIntegerTy() && !IsVolatile;
768 
769  insertUse(I, Offset, Size, IsSplittable);
770  }
771 
772  void visitLoadInst(LoadInst &LI) {
773  assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
774  "All simple FCA loads should have been pre-split");
775 
776  if (!IsOffsetKnown)
777  return PI.setAborted(&LI);
778 
779  const DataLayout &DL = LI.getModule()->getDataLayout();
780  uint64_t Size = DL.getTypeStoreSize(LI.getType());
781  return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
782  }
783 
784  void visitStoreInst(StoreInst &SI) {
785  Value *ValOp = SI.getValueOperand();
786  if (ValOp == *U)
787  return PI.setEscapedAndAborted(&SI);
788  if (!IsOffsetKnown)
789  return PI.setAborted(&SI);
790 
791  const DataLayout &DL = SI.getModule()->getDataLayout();
792  uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
793 
794  // If this memory access can be shown to *statically* extend outside the
795  // bounds of the allocation, it's behavior is undefined, so simply
796  // ignore it. Note that this is more strict than the generic clamping
797  // behavior of insertUse. We also try to handle cases which might run the
798  // risk of overflow.
799  // FIXME: We should instead consider the pointer to have escaped if this
800  // function is being instrumented for addressing bugs or race conditions.
801  if (Size > AllocSize || Offset.ugt(AllocSize - Size)) {
802  LLVM_DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @"
803  << Offset << " which extends past the end of the "
804  << AllocSize << " byte alloca:\n"
805  << " alloca: " << AS.AI << "\n"
806  << " use: " << SI << "\n");
807  return markAsDead(SI);
808  }
809 
810  assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
811  "All simple FCA stores should have been pre-split");
812  handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
813  }
814 
815  void visitMemSetInst(MemSetInst &II) {
816  assert(II.getRawDest() == *U && "Pointer use is not the destination?");
817  ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
818  if ((Length && Length->getValue() == 0) ||
819  (IsOffsetKnown && Offset.uge(AllocSize)))
820  // Zero-length mem transfer intrinsics can be ignored entirely.
821  return markAsDead(II);
822 
823  if (!IsOffsetKnown)
824  return PI.setAborted(&II);
825 
826  insertUse(II, Offset, Length ? Length->getLimitedValue()
827  : AllocSize - Offset.getLimitedValue(),
828  (bool)Length);
829  }
830 
831  void visitMemTransferInst(MemTransferInst &II) {
832  ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
833  if (Length && Length->getValue() == 0)
834  // Zero-length mem transfer intrinsics can be ignored entirely.
835  return markAsDead(II);
836 
837  // Because we can visit these intrinsics twice, also check to see if the
838  // first time marked this instruction as dead. If so, skip it.
839  if (VisitedDeadInsts.count(&II))
840  return;
841 
842  if (!IsOffsetKnown)
843  return PI.setAborted(&II);
844 
845  // This side of the transfer is completely out-of-bounds, and so we can
846  // nuke the entire transfer. However, we also need to nuke the other side
847  // if already added to our partitions.
848  // FIXME: Yet another place we really should bypass this when
849  // instrumenting for ASan.
850  if (Offset.uge(AllocSize)) {
852  MemTransferSliceMap.find(&II);
853  if (MTPI != MemTransferSliceMap.end())
854  AS.Slices[MTPI->second].kill();
855  return markAsDead(II);
856  }
857 
858  uint64_t RawOffset = Offset.getLimitedValue();
859  uint64_t Size = Length ? Length->getLimitedValue() : AllocSize - RawOffset;
860 
861  // Check for the special case where the same exact value is used for both
862  // source and dest.
863  if (*U == II.getRawDest() && *U == II.getRawSource()) {
864  // For non-volatile transfers this is a no-op.
865  if (!II.isVolatile())
866  return markAsDead(II);
867 
868  return insertUse(II, Offset, Size, /*IsSplittable=*/false);
869  }
870 
871  // If we have seen both source and destination for a mem transfer, then
872  // they both point to the same alloca.
873  bool Inserted;
875  std::tie(MTPI, Inserted) =
876  MemTransferSliceMap.insert(std::make_pair(&II, AS.Slices.size()));
877  unsigned PrevIdx = MTPI->second;
878  if (!Inserted) {
879  Slice &PrevP = AS.Slices[PrevIdx];
880 
881  // Check if the begin offsets match and this is a non-volatile transfer.
882  // In that case, we can completely elide the transfer.
883  if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
884  PrevP.kill();
885  return markAsDead(II);
886  }
887 
888  // Otherwise we have an offset transfer within the same alloca. We can't
889  // split those.
890  PrevP.makeUnsplittable();
891  }
892 
893  // Insert the use now that we've fixed up the splittable nature.
894  insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
895 
896  // Check that we ended up with a valid index in the map.
897  assert(AS.Slices[PrevIdx].getUse()->getUser() == &II &&
898  "Map index doesn't point back to a slice with this user.");
899  }
900 
901  // Disable SRoA for any intrinsics except for lifetime invariants.
902  // FIXME: What about debug intrinsics? This matches old behavior, but
903  // doesn't make sense.
904  void visitIntrinsicInst(IntrinsicInst &II) {
905  if (!IsOffsetKnown)
906  return PI.setAborted(&II);
907 
908  if (II.isLifetimeStartOrEnd()) {
909  ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
910  uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
911  Length->getLimitedValue());
912  insertUse(II, Offset, Size, true);
913  return;
914  }
915 
916  Base::visitIntrinsicInst(II);
917  }
918 
919  Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
920  // We consider any PHI or select that results in a direct load or store of
921  // the same offset to be a viable use for slicing purposes. These uses
922  // are considered unsplittable and the size is the maximum loaded or stored
923  // size.
926  Visited.insert(Root);
927  Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
928  const DataLayout &DL = Root->getModule()->getDataLayout();
929  // If there are no loads or stores, the access is dead. We mark that as
930  // a size zero access.
931  Size = 0;
932  do {
933  Instruction *I, *UsedI;
934  std::tie(UsedI, I) = Uses.pop_back_val();
935 
936  if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
937  Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
938  continue;
939  }
940  if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
941  Value *Op = SI->getOperand(0);
942  if (Op == UsedI)
943  return SI;
944  Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
945  continue;
946  }
947 
948  if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
949  if (!GEP->hasAllZeroIndices())
950  return GEP;
951  } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
952  !isa<SelectInst>(I)) {
953  return I;
954  }
955 
956  for (User *U : I->users())
957  if (Visited.insert(cast<Instruction>(U)).second)
958  Uses.push_back(std::make_pair(I, cast<Instruction>(U)));
959  } while (!Uses.empty());
960 
961  return nullptr;
962  }
963 
964  void visitPHINodeOrSelectInst(Instruction &I) {
965  assert(isa<PHINode>(I) || isa<SelectInst>(I));
966  if (I.use_empty())
967  return markAsDead(I);
968 
969  // TODO: We could use SimplifyInstruction here to fold PHINodes and
970  // SelectInsts. However, doing so requires to change the current
971  // dead-operand-tracking mechanism. For instance, suppose neither loading
972  // from %U nor %other traps. Then "load (select undef, %U, %other)" does not
973  // trap either. However, if we simply replace %U with undef using the
974  // current dead-operand-tracking mechanism, "load (select undef, undef,
975  // %other)" may trap because the select may return the first operand
976  // "undef".
977  if (Value *Result = foldPHINodeOrSelectInst(I)) {
978  if (Result == *U)
979  // If the result of the constant fold will be the pointer, recurse
980  // through the PHI/select as if we had RAUW'ed it.
981  enqueueUsers(I);
982  else
983  // Otherwise the operand to the PHI/select is dead, and we can replace
984  // it with undef.
985  AS.DeadOperands.push_back(U);
986 
987  return;
988  }
989 
990  if (!IsOffsetKnown)
991  return PI.setAborted(&I);
992 
993  // See if we already have computed info on this node.
994  uint64_t &Size = PHIOrSelectSizes[&I];
995  if (!Size) {
996  // This is a new PHI/Select, check for an unsafe use of it.
997  if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&I, Size))
998  return PI.setAborted(UnsafeI);
999  }
1000 
1001  // For PHI and select operands outside the alloca, we can't nuke the entire
1002  // phi or select -- the other side might still be relevant, so we special
1003  // case them here and use a separate structure to track the operands
1004  // themselves which should be replaced with undef.
1005  // FIXME: This should instead be escaped in the event we're instrumenting
1006  // for address sanitization.
1007  if (Offset.uge(AllocSize)) {
1008  AS.DeadOperands.push_back(U);
1009  return;
1010  }
1011 
1012  insertUse(I, Offset, Size);
1013  }
1014 
1015  void visitPHINode(PHINode &PN) { visitPHINodeOrSelectInst(PN); }
1016 
1017  void visitSelectInst(SelectInst &SI) { visitPHINodeOrSelectInst(SI); }
1018 
1019  /// Disable SROA entirely if there are unhandled users of the alloca.
1020  void visitInstruction(Instruction &I) { PI.setAborted(&I); }
1021 };
1022 
1024  :
1025 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1026  AI(AI),
1027 #endif
1028  PointerEscapingInstr(nullptr) {
1029  SliceBuilder PB(DL, AI, *this);
1030  SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
1031  if (PtrI.isEscaped() || PtrI.isAborted()) {
1032  // FIXME: We should sink the escape vs. abort info into the caller nicely,
1033  // possibly by just storing the PtrInfo in the AllocaSlices.
1034  PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
1035  : PtrI.getAbortingInst();
1036  assert(PointerEscapingInstr && "Did not track a bad instruction");
1037  return;
1038  }
1039 
1040  Slices.erase(
1041  llvm::remove_if(Slices, [](const Slice &S) { return S.isDead(); }),
1042  Slices.end());
1043 
1044 #ifndef NDEBUG
1046  std::mt19937 MT(static_cast<unsigned>(
1047  std::chrono::system_clock::now().time_since_epoch().count()));
1048  std::shuffle(Slices.begin(), Slices.end(), MT);
1049  }
1050 #endif
1051 
1052  // Sort the uses. This arranges for the offsets to be in ascending order,
1053  // and the sizes to be in descending order.
1054  llvm::sort(Slices);
1055 }
1056 
1057 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1058 
1059 void AllocaSlices::print(raw_ostream &OS, const_iterator I,
1060  StringRef Indent) const {
1061  printSlice(OS, I, Indent);
1062  OS << "\n";
1063  printUse(OS, I, Indent);
1064 }
1065 
1066 void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
1067  StringRef Indent) const {
1068  OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
1069  << " slice #" << (I - begin())
1070  << (I->isSplittable() ? " (splittable)" : "");
1071 }
1072 
1073 void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
1074  StringRef Indent) const {
1075  OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
1076 }
1077 
1078 void AllocaSlices::print(raw_ostream &OS) const {
1079  if (PointerEscapingInstr) {
1080  OS << "Can't analyze slices for alloca: " << AI << "\n"
1081  << " A pointer to this alloca escaped by:\n"
1082  << " " << *PointerEscapingInstr << "\n";
1083  return;
1084  }
1085 
1086  OS << "Slices of alloca: " << AI << "\n";
1087  for (const_iterator I = begin(), E = end(); I != E; ++I)
1088  print(OS, I);
1089 }
1090 
1091 LLVM_DUMP_METHOD void AllocaSlices::dump(const_iterator I) const {
1092  print(dbgs(), I);
1093 }
1094 LLVM_DUMP_METHOD void AllocaSlices::dump() const { print(dbgs()); }
1095 
1096 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1097 
1098 /// Walk the range of a partitioning looking for a common type to cover this
1099 /// sequence of slices.
1102  uint64_t EndOffset) {
1103  Type *Ty = nullptr;
1104  bool TyIsCommon = true;
1105  IntegerType *ITy = nullptr;
1106 
1107  // Note that we need to look at *every* alloca slice's Use to ensure we
1108  // always get consistent results regardless of the order of slices.
1109  for (AllocaSlices::const_iterator I = B; I != E; ++I) {
1110  Use *U = I->getUse();
1111  if (isa<IntrinsicInst>(*U->getUser()))
1112  continue;
1113  if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
1114  continue;
1115 
1116  Type *UserTy = nullptr;
1117  if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1118  UserTy = LI->getType();
1119  } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1120  UserTy = SI->getValueOperand()->getType();
1121  }
1122 
1123  if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
1124  // If the type is larger than the partition, skip it. We only encounter
1125  // this for split integer operations where we want to use the type of the
1126  // entity causing the split. Also skip if the type is not a byte width
1127  // multiple.
1128  if (UserITy->getBitWidth() % 8 != 0 ||
1129  UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
1130  continue;
1131 
1132  // Track the largest bitwidth integer type used in this way in case there
1133  // is no common type.
1134  if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
1135  ITy = UserITy;
1136  }
1137 
1138  // To avoid depending on the order of slices, Ty and TyIsCommon must not
1139  // depend on types skipped above.
1140  if (!UserTy || (Ty && Ty != UserTy))
1141  TyIsCommon = false; // Give up on anything but an iN type.
1142  else
1143  Ty = UserTy;
1144  }
1145 
1146  return TyIsCommon ? Ty : ITy;
1147 }
1148 
1149 /// PHI instructions that use an alloca and are subsequently loaded can be
1150 /// rewritten to load both input pointers in the pred blocks and then PHI the
1151 /// results, allowing the load of the alloca to be promoted.
1152 /// From this:
1153 /// %P2 = phi [i32* %Alloca, i32* %Other]
1154 /// %V = load i32* %P2
1155 /// to:
1156 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1157 /// ...
1158 /// %V2 = load i32* %Other
1159 /// ...
1160 /// %V = phi [i32 %V1, i32 %V2]
1161 ///
1162 /// We can do this to a select if its only uses are loads and if the operands
1163 /// to the select can be loaded unconditionally.
1164 ///
1165 /// FIXME: This should be hoisted into a generic utility, likely in
1166 /// Transforms/Util/Local.h
1167 static bool isSafePHIToSpeculate(PHINode &PN) {
1168  // For now, we can only do this promotion if the load is in the same block
1169  // as the PHI, and if there are no stores between the phi and load.
1170  // TODO: Allow recursive phi users.
1171  // TODO: Allow stores.
1172  BasicBlock *BB = PN.getParent();
1173  unsigned MaxAlign = 0;
1174  bool HaveLoad = false;
1175  for (User *U : PN.users()) {
1176  LoadInst *LI = dyn_cast<LoadInst>(U);
1177  if (!LI || !LI->isSimple())
1178  return false;
1179 
1180  // For now we only allow loads in the same block as the PHI. This is
1181  // a common case that happens when instcombine merges two loads through
1182  // a PHI.
1183  if (LI->getParent() != BB)
1184  return false;
1185 
1186  // Ensure that there are no instructions between the PHI and the load that
1187  // could store.
1188  for (BasicBlock::iterator BBI(PN); &*BBI != LI; ++BBI)
1189  if (BBI->mayWriteToMemory())
1190  return false;
1191 
1192  MaxAlign = std::max(MaxAlign, LI->getAlignment());
1193  HaveLoad = true;
1194  }
1195 
1196  if (!HaveLoad)
1197  return false;
1198 
1199  const DataLayout &DL = PN.getModule()->getDataLayout();
1200 
1201  // We can only transform this if it is safe to push the loads into the
1202  // predecessor blocks. The only thing to watch out for is that we can't put
1203  // a possibly trapping load in the predecessor if it is a critical edge.
1204  for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1205  Instruction *TI = PN.getIncomingBlock(Idx)->getTerminator();
1206  Value *InVal = PN.getIncomingValue(Idx);
1207 
1208  // If the value is produced by the terminator of the predecessor (an
1209  // invoke) or it has side-effects, there is no valid place to put a load
1210  // in the predecessor.
1211  if (TI == InVal || TI->mayHaveSideEffects())
1212  return false;
1213 
1214  // If the predecessor has a single successor, then the edge isn't
1215  // critical.
1216  if (TI->getNumSuccessors() == 1)
1217  continue;
1218 
1219  // If this pointer is always safe to load, or if we can prove that there
1220  // is already a load in the block, then we can move the load to the pred
1221  // block.
1222  if (isSafeToLoadUnconditionally(InVal, MaxAlign, DL, TI))
1223  continue;
1224 
1225  return false;
1226  }
1227 
1228  return true;
1229 }
1230 
1231 static void speculatePHINodeLoads(PHINode &PN) {
1232  LLVM_DEBUG(dbgs() << " original: " << PN << "\n");
1233 
1234  Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1235  IRBuilderTy PHIBuilder(&PN);
1236  PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1237  PN.getName() + ".sroa.speculated");
1238 
1239  // Get the AA tags and alignment to use from one of the loads. It doesn't
1240  // matter which one we get and if any differ.
1241  LoadInst *SomeLoad = cast<LoadInst>(PN.user_back());
1242 
1243  AAMDNodes AATags;
1244  SomeLoad->getAAMetadata(AATags);
1245  unsigned Align = SomeLoad->getAlignment();
1246 
1247  // Rewrite all loads of the PN to use the new PHI.
1248  while (!PN.use_empty()) {
1249  LoadInst *LI = cast<LoadInst>(PN.user_back());
1250  LI->replaceAllUsesWith(NewPN);
1251  LI->eraseFromParent();
1252  }
1253 
1254  // Inject loads into all of the pred blocks.
1255  DenseMap<BasicBlock*, Value*> InjectedLoads;
1256  for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1257  BasicBlock *Pred = PN.getIncomingBlock(Idx);
1258  Value *InVal = PN.getIncomingValue(Idx);
1259 
1260  // A PHI node is allowed to have multiple (duplicated) entries for the same
1261  // basic block, as long as the value is the same. So if we already injected
1262  // a load in the predecessor, then we should reuse the same load for all
1263  // duplicated entries.
1264  if (Value* V = InjectedLoads.lookup(Pred)) {
1265  NewPN->addIncoming(V, Pred);
1266  continue;
1267  }
1268 
1269  Instruction *TI = Pred->getTerminator();
1270  IRBuilderTy PredBuilder(TI);
1271 
1272  LoadInst *Load = PredBuilder.CreateLoad(
1273  InVal, (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1274  ++NumLoadsSpeculated;
1275  Load->setAlignment(Align);
1276  if (AATags)
1277  Load->setAAMetadata(AATags);
1278  NewPN->addIncoming(Load, Pred);
1279  InjectedLoads[Pred] = Load;
1280  }
1281 
1282  LLVM_DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1283  PN.eraseFromParent();
1284 }
1285 
1286 /// Select instructions that use an alloca and are subsequently loaded can be
1287 /// rewritten to load both input pointers and then select between the result,
1288 /// allowing the load of the alloca to be promoted.
1289 /// From this:
1290 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1291 /// %V = load i32* %P2
1292 /// to:
1293 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1294 /// %V2 = load i32* %Other
1295 /// %V = select i1 %cond, i32 %V1, i32 %V2
1296 ///
1297 /// We can do this to a select if its only uses are loads and if the operand
1298 /// to the select can be loaded unconditionally.
1300  Value *TValue = SI.getTrueValue();
1301  Value *FValue = SI.getFalseValue();
1302  const DataLayout &DL = SI.getModule()->getDataLayout();
1303 
1304  for (User *U : SI.users()) {
1305  LoadInst *LI = dyn_cast<LoadInst>(U);
1306  if (!LI || !LI->isSimple())
1307  return false;
1308 
1309  // Both operands to the select need to be dereferenceable, either
1310  // absolutely (e.g. allocas) or at this point because we can see other
1311  // accesses to it.
1312  if (!isSafeToLoadUnconditionally(TValue, LI->getAlignment(), DL, LI))
1313  return false;
1314  if (!isSafeToLoadUnconditionally(FValue, LI->getAlignment(), DL, LI))
1315  return false;
1316  }
1317 
1318  return true;
1319 }
1320 
1322  LLVM_DEBUG(dbgs() << " original: " << SI << "\n");
1323 
1324  IRBuilderTy IRB(&SI);
1325  Value *TV = SI.getTrueValue();
1326  Value *FV = SI.getFalseValue();
1327  // Replace the loads of the select with a select of two loads.
1328  while (!SI.use_empty()) {
1329  LoadInst *LI = cast<LoadInst>(SI.user_back());
1330  assert(LI->isSimple() && "We only speculate simple loads");
1331 
1332  IRB.SetInsertPoint(LI);
1333  LoadInst *TL =
1334  IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1335  LoadInst *FL =
1336  IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1337  NumLoadsSpeculated += 2;
1338 
1339  // Transfer alignment and AA info if present.
1340  TL->setAlignment(LI->getAlignment());
1341  FL->setAlignment(LI->getAlignment());
1342 
1343  AAMDNodes Tags;
1344  LI->getAAMetadata(Tags);
1345  if (Tags) {
1346  TL->setAAMetadata(Tags);
1347  FL->setAAMetadata(Tags);
1348  }
1349 
1350  Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1351  LI->getName() + ".sroa.speculated");
1352 
1353  LLVM_DEBUG(dbgs() << " speculated to: " << *V << "\n");
1354  LI->replaceAllUsesWith(V);
1355  LI->eraseFromParent();
1356  }
1357  SI.eraseFromParent();
1358 }
1359 
1360 /// Build a GEP out of a base pointer and indices.
1361 ///
1362 /// This will return the BasePtr if that is valid, or build a new GEP
1363 /// instruction using the IRBuilder if GEP-ing is needed.
1364 static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1365  SmallVectorImpl<Value *> &Indices, Twine NamePrefix) {
1366  if (Indices.empty())
1367  return BasePtr;
1368 
1369  // A single zero index is a no-op, so check for this and avoid building a GEP
1370  // in that case.
1371  if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1372  return BasePtr;
1373 
1374  return IRB.CreateInBoundsGEP(nullptr, BasePtr, Indices,
1375  NamePrefix + "sroa_idx");
1376 }
1377 
1378 /// Get a natural GEP off of the BasePtr walking through Ty toward
1379 /// TargetTy without changing the offset of the pointer.
1380 ///
1381 /// This routine assumes we've already established a properly offset GEP with
1382 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1383 /// zero-indices down through type layers until we find one the same as
1384 /// TargetTy. If we can't find one with the same type, we at least try to use
1385 /// one with the same size. If none of that works, we just produce the GEP as
1386 /// indicated by Indices to have the correct offset.
1387 static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1388  Value *BasePtr, Type *Ty, Type *TargetTy,
1389  SmallVectorImpl<Value *> &Indices,
1390  Twine NamePrefix) {
1391  if (Ty == TargetTy)
1392  return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1393 
1394  // Offset size to use for the indices.
1395  unsigned OffsetSize = DL.getIndexTypeSizeInBits(BasePtr->getType());
1396 
1397  // See if we can descend into a struct and locate a field with the correct
1398  // type.
1399  unsigned NumLayers = 0;
1400  Type *ElementTy = Ty;
1401  do {
1402  if (ElementTy->isPointerTy())
1403  break;
1404 
1405  if (ArrayType *ArrayTy = dyn_cast<ArrayType>(ElementTy)) {
1406  ElementTy = ArrayTy->getElementType();
1407  Indices.push_back(IRB.getIntN(OffsetSize, 0));
1408  } else if (VectorType *VectorTy = dyn_cast<VectorType>(ElementTy)) {
1409  ElementTy = VectorTy->getElementType();
1410  Indices.push_back(IRB.getInt32(0));
1411  } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1412  if (STy->element_begin() == STy->element_end())
1413  break; // Nothing left to descend into.
1414  ElementTy = *STy->element_begin();
1415  Indices.push_back(IRB.getInt32(0));
1416  } else {
1417  break;
1418  }
1419  ++NumLayers;
1420  } while (ElementTy != TargetTy);
1421  if (ElementTy != TargetTy)
1422  Indices.erase(Indices.end() - NumLayers, Indices.end());
1423 
1424  return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1425 }
1426 
1427 /// Recursively compute indices for a natural GEP.
1428 ///
1429 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1430 /// element types adding appropriate indices for the GEP.
1431 static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1432  Value *Ptr, Type *Ty, APInt &Offset,
1433  Type *TargetTy,
1434  SmallVectorImpl<Value *> &Indices,
1435  Twine NamePrefix) {
1436  if (Offset == 0)
1437  return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices,
1438  NamePrefix);
1439 
1440  // We can't recurse through pointer types.
1441  if (Ty->isPointerTy())
1442  return nullptr;
1443 
1444  // We try to analyze GEPs over vectors here, but note that these GEPs are
1445  // extremely poorly defined currently. The long-term goal is to remove GEPing
1446  // over a vector from the IR completely.
1447  if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1448  unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
1449  if (ElementSizeInBits % 8 != 0) {
1450  // GEPs over non-multiple of 8 size vector elements are invalid.
1451  return nullptr;
1452  }
1453  APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1454  APInt NumSkippedElements = Offset.sdiv(ElementSize);
1455  if (NumSkippedElements.ugt(VecTy->getNumElements()))
1456  return nullptr;
1457  Offset -= NumSkippedElements * ElementSize;
1458  Indices.push_back(IRB.getInt(NumSkippedElements));
1459  return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1460  Offset, TargetTy, Indices, NamePrefix);
1461  }
1462 
1463  if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1464  Type *ElementTy = ArrTy->getElementType();
1465  APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1466  APInt NumSkippedElements = Offset.sdiv(ElementSize);
1467  if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1468  return nullptr;
1469 
1470  Offset -= NumSkippedElements * ElementSize;
1471  Indices.push_back(IRB.getInt(NumSkippedElements));
1472  return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1473  Indices, NamePrefix);
1474  }
1475 
1476  StructType *STy = dyn_cast<StructType>(Ty);
1477  if (!STy)
1478  return nullptr;
1479 
1480  const StructLayout *SL = DL.getStructLayout(STy);
1481  uint64_t StructOffset = Offset.getZExtValue();
1482  if (StructOffset >= SL->getSizeInBytes())
1483  return nullptr;
1484  unsigned Index = SL->getElementContainingOffset(StructOffset);
1485  Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1486  Type *ElementTy = STy->getElementType(Index);
1487  if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
1488  return nullptr; // The offset points into alignment padding.
1489 
1490  Indices.push_back(IRB.getInt32(Index));
1491  return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1492  Indices, NamePrefix);
1493 }
1494 
1495 /// Get a natural GEP from a base pointer to a particular offset and
1496 /// resulting in a particular type.
1497 ///
1498 /// The goal is to produce a "natural" looking GEP that works with the existing
1499 /// composite types to arrive at the appropriate offset and element type for
1500 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1501 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1502 /// Indices, and setting Ty to the result subtype.
1503 ///
1504 /// If no natural GEP can be constructed, this function returns null.
1505 static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1506  Value *Ptr, APInt Offset, Type *TargetTy,
1507  SmallVectorImpl<Value *> &Indices,
1508  Twine NamePrefix) {
1509  PointerType *Ty = cast<PointerType>(Ptr->getType());
1510 
1511  // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1512  // an i8.
1513  if (Ty == IRB.getInt8PtrTy(Ty->getAddressSpace()) && TargetTy->isIntegerTy(8))
1514  return nullptr;
1515 
1516  Type *ElementTy = Ty->getElementType();
1517  if (!ElementTy->isSized())
1518  return nullptr; // We can't GEP through an unsized element.
1519  APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1520  if (ElementSize == 0)
1521  return nullptr; // Zero-length arrays can't help us build a natural GEP.
1522  APInt NumSkippedElements = Offset.sdiv(ElementSize);
1523 
1524  Offset -= NumSkippedElements * ElementSize;
1525  Indices.push_back(IRB.getInt(NumSkippedElements));
1526  return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1527  Indices, NamePrefix);
1528 }
1529 
1530 /// Compute an adjusted pointer from Ptr by Offset bytes where the
1531 /// resulting pointer has PointerTy.
1532 ///
1533 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1534 /// and produces the pointer type desired. Where it cannot, it will try to use
1535 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1536 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1537 /// bitcast to the type.
1538 ///
1539 /// The strategy for finding the more natural GEPs is to peel off layers of the
1540 /// pointer, walking back through bit casts and GEPs, searching for a base
1541 /// pointer from which we can compute a natural GEP with the desired
1542 /// properties. The algorithm tries to fold as many constant indices into
1543 /// a single GEP as possible, thus making each GEP more independent of the
1544 /// surrounding code.
1545 static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr,
1546  APInt Offset, Type *PointerTy, Twine NamePrefix) {
1547  // Even though we don't look through PHI nodes, we could be called on an
1548  // instruction in an unreachable block, which may be on a cycle.
1549  SmallPtrSet<Value *, 4> Visited;
1550  Visited.insert(Ptr);
1551  SmallVector<Value *, 4> Indices;
1552 
1553  // We may end up computing an offset pointer that has the wrong type. If we
1554  // never are able to compute one directly that has the correct type, we'll
1555  // fall back to it, so keep it and the base it was computed from around here.
1556  Value *OffsetPtr = nullptr;
1557  Value *OffsetBasePtr;
1558 
1559  // Remember any i8 pointer we come across to re-use if we need to do a raw
1560  // byte offset.
1561  Value *Int8Ptr = nullptr;
1562  APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1563 
1564  Type *TargetTy = PointerTy->getPointerElementType();
1565 
1566  do {
1567  // First fold any existing GEPs into the offset.
1568  while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1569  APInt GEPOffset(Offset.getBitWidth(), 0);
1570  if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1571  break;
1572  Offset += GEPOffset;
1573  Ptr = GEP->getPointerOperand();
1574  if (!Visited.insert(Ptr).second)
1575  break;
1576  }
1577 
1578  // See if we can perform a natural GEP here.
1579  Indices.clear();
1580  if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1581  Indices, NamePrefix)) {
1582  // If we have a new natural pointer at the offset, clear out any old
1583  // offset pointer we computed. Unless it is the base pointer or
1584  // a non-instruction, we built a GEP we don't need. Zap it.
1585  if (OffsetPtr && OffsetPtr != OffsetBasePtr)
1586  if (Instruction *I = dyn_cast<Instruction>(OffsetPtr)) {
1587  assert(I->use_empty() && "Built a GEP with uses some how!");
1588  I->eraseFromParent();
1589  }
1590  OffsetPtr = P;
1591  OffsetBasePtr = Ptr;
1592  // If we also found a pointer of the right type, we're done.
1593  if (P->getType() == PointerTy)
1594  return P;
1595  }
1596 
1597  // Stash this pointer if we've found an i8*.
1598  if (Ptr->getType()->isIntegerTy(8)) {
1599  Int8Ptr = Ptr;
1600  Int8PtrOffset = Offset;
1601  }
1602 
1603  // Peel off a layer of the pointer and update the offset appropriately.
1604  if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1605  Ptr = cast<Operator>(Ptr)->getOperand(0);
1606  } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1607  if (GA->isInterposable())
1608  break;
1609  Ptr = GA->getAliasee();
1610  } else {
1611  break;
1612  }
1613  assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1614  } while (Visited.insert(Ptr).second);
1615 
1616  if (!OffsetPtr) {
1617  if (!Int8Ptr) {
1618  Int8Ptr = IRB.CreateBitCast(
1619  Ptr, IRB.getInt8PtrTy(PointerTy->getPointerAddressSpace()),
1620  NamePrefix + "sroa_raw_cast");
1621  Int8PtrOffset = Offset;
1622  }
1623 
1624  OffsetPtr = Int8PtrOffset == 0
1625  ? Int8Ptr
1626  : IRB.CreateInBoundsGEP(IRB.getInt8Ty(), Int8Ptr,
1627  IRB.getInt(Int8PtrOffset),
1628  NamePrefix + "sroa_raw_idx");
1629  }
1630  Ptr = OffsetPtr;
1631 
1632  // On the off chance we were targeting i8*, guard the bitcast here.
1633  if (Ptr->getType() != PointerTy)
1634  Ptr = IRB.CreateBitCast(Ptr, PointerTy, NamePrefix + "sroa_cast");
1635 
1636  return Ptr;
1637 }
1638 
1639 /// Compute the adjusted alignment for a load or store from an offset.
1640 static unsigned getAdjustedAlignment(Instruction *I, uint64_t Offset,
1641  const DataLayout &DL) {
1642  unsigned Alignment;
1643  Type *Ty;
1644  if (auto *LI = dyn_cast<LoadInst>(I)) {
1645  Alignment = LI->getAlignment();
1646  Ty = LI->getType();
1647  } else if (auto *SI = dyn_cast<StoreInst>(I)) {
1648  Alignment = SI->getAlignment();
1649  Ty = SI->getValueOperand()->getType();
1650  } else {
1651  llvm_unreachable("Only loads and stores are allowed!");
1652  }
1653 
1654  if (!Alignment)
1655  Alignment = DL.getABITypeAlignment(Ty);
1656 
1657  return MinAlign(Alignment, Offset);
1658 }
1659 
1660 /// Test whether we can convert a value from the old to the new type.
1661 ///
1662 /// This predicate should be used to guard calls to convertValue in order to
1663 /// ensure that we only try to convert viable values. The strategy is that we
1664 /// will peel off single element struct and array wrappings to get to an
1665 /// underlying value, and convert that value.
1666 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1667  if (OldTy == NewTy)
1668  return true;
1669 
1670  // For integer types, we can't handle any bit-width differences. This would
1671  // break both vector conversions with extension and introduce endianness
1672  // issues when in conjunction with loads and stores.
1673  if (isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) {
1674  assert(cast<IntegerType>(OldTy)->getBitWidth() !=
1675  cast<IntegerType>(NewTy)->getBitWidth() &&
1676  "We can't have the same bitwidth for different int types");
1677  return false;
1678  }
1679 
1680  if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1681  return false;
1682  if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1683  return false;
1684 
1685  // We can convert pointers to integers and vice-versa. Same for vectors
1686  // of pointers and integers.
1687  OldTy = OldTy->getScalarType();
1688  NewTy = NewTy->getScalarType();
1689  if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1690  if (NewTy->isPointerTy() && OldTy->isPointerTy()) {
1691  return cast<PointerType>(NewTy)->getPointerAddressSpace() ==
1692  cast<PointerType>(OldTy)->getPointerAddressSpace();
1693  }
1694 
1695  // We can convert integers to integral pointers, but not to non-integral
1696  // pointers.
1697  if (OldTy->isIntegerTy())
1698  return !DL.isNonIntegralPointerType(NewTy);
1699 
1700  // We can convert integral pointers to integers, but non-integral pointers
1701  // need to remain pointers.
1702  if (!DL.isNonIntegralPointerType(OldTy))
1703  return NewTy->isIntegerTy();
1704 
1705  return false;
1706  }
1707 
1708  return true;
1709 }
1710 
1711 /// Generic routine to convert an SSA value to a value of a different
1712 /// type.
1713 ///
1714 /// This will try various different casting techniques, such as bitcasts,
1715 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1716 /// two types for viability with this routine.
1717 static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1718  Type *NewTy) {
1719  Type *OldTy = V->getType();
1720  assert(canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type");
1721 
1722  if (OldTy == NewTy)
1723  return V;
1724 
1725  assert(!(isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) &&
1726  "Integer types must be the exact same to convert.");
1727 
1728  // See if we need inttoptr for this type pair. A cast involving both scalars
1729  // and vectors requires and additional bitcast.
1730  if (OldTy->isIntOrIntVectorTy() && NewTy->isPtrOrPtrVectorTy()) {
1731  // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1732  if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1733  return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1734  NewTy);
1735 
1736  // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1737  if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1738  return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1739  NewTy);
1740 
1741  return IRB.CreateIntToPtr(V, NewTy);
1742  }
1743 
1744  // See if we need ptrtoint for this type pair. A cast involving both scalars
1745  // and vectors requires and additional bitcast.
1746  if (OldTy->isPtrOrPtrVectorTy() && NewTy->isIntOrIntVectorTy()) {
1747  // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1748  if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1749  return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1750  NewTy);
1751 
1752  // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1753  if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1754  return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1755  NewTy);
1756 
1757  return IRB.CreatePtrToInt(V, NewTy);
1758  }
1759 
1760  return IRB.CreateBitCast(V, NewTy);
1761 }
1762 
1763 /// Test whether the given slice use can be promoted to a vector.
1764 ///
1765 /// This function is called to test each entry in a partition which is slated
1766 /// for a single slice.
1767 static bool isVectorPromotionViableForSlice(Partition &P, const Slice &S,
1768  VectorType *Ty,
1769  uint64_t ElementSize,
1770  const DataLayout &DL) {
1771  // First validate the slice offsets.
1772  uint64_t BeginOffset =
1773  std::max(S.beginOffset(), P.beginOffset()) - P.beginOffset();
1774  uint64_t BeginIndex = BeginOffset / ElementSize;
1775  if (BeginIndex * ElementSize != BeginOffset ||
1776  BeginIndex >= Ty->getNumElements())
1777  return false;
1778  uint64_t EndOffset =
1779  std::min(S.endOffset(), P.endOffset()) - P.beginOffset();
1780  uint64_t EndIndex = EndOffset / ElementSize;
1781  if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
1782  return false;
1783 
1784  assert(EndIndex > BeginIndex && "Empty vector!");
1785  uint64_t NumElements = EndIndex - BeginIndex;
1786  Type *SliceTy = (NumElements == 1)
1787  ? Ty->getElementType()
1788  : VectorType::get(Ty->getElementType(), NumElements);
1789 
1790  Type *SplitIntTy =
1791  Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1792 
1793  Use *U = S.getUse();
1794 
1795  if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1796  if (MI->isVolatile())
1797  return false;
1798  if (!S.isSplittable())
1799  return false; // Skip any unsplittable intrinsics.
1800  } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1801  if (!II->isLifetimeStartOrEnd())
1802  return false;
1803  } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1804  // Disable vector promotion when there are loads or stores of an FCA.
1805  return false;
1806  } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1807  if (LI->isVolatile())
1808  return false;
1809  Type *LTy = LI->getType();
1810  if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1811  assert(LTy->isIntegerTy());
1812  LTy = SplitIntTy;
1813  }
1814  if (!canConvertValue(DL, SliceTy, LTy))
1815  return false;
1816  } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1817  if (SI->isVolatile())
1818  return false;
1819  Type *STy = SI->getValueOperand()->getType();
1820  if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1821  assert(STy->isIntegerTy());
1822  STy = SplitIntTy;
1823  }
1824  if (!canConvertValue(DL, STy, SliceTy))
1825  return false;
1826  } else {
1827  return false;
1828  }
1829 
1830  return true;
1831 }
1832 
1833 /// Test whether the given alloca partitioning and range of slices can be
1834 /// promoted to a vector.
1835 ///
1836 /// This is a quick test to check whether we can rewrite a particular alloca
1837 /// partition (and its newly formed alloca) into a vector alloca with only
1838 /// whole-vector loads and stores such that it could be promoted to a vector
1839 /// SSA value. We only can ensure this for a limited set of operations, and we
1840 /// don't want to do the rewrites unless we are confident that the result will
1841 /// be promotable, so we have an early test here.
1843  // Collect the candidate types for vector-based promotion. Also track whether
1844  // we have different element types.
1845  SmallVector<VectorType *, 4> CandidateTys;
1846  Type *CommonEltTy = nullptr;
1847  bool HaveCommonEltTy = true;
1848  auto CheckCandidateType = [&](Type *Ty) {
1849  if (auto *VTy = dyn_cast<VectorType>(Ty)) {
1850  CandidateTys.push_back(VTy);
1851  if (!CommonEltTy)
1852  CommonEltTy = VTy->getElementType();
1853  else if (CommonEltTy != VTy->getElementType())
1854  HaveCommonEltTy = false;
1855  }
1856  };
1857  // Consider any loads or stores that are the exact size of the slice.
1858  for (const Slice &S : P)
1859  if (S.beginOffset() == P.beginOffset() &&
1860  S.endOffset() == P.endOffset()) {
1861  if (auto *LI = dyn_cast<LoadInst>(S.getUse()->getUser()))
1862  CheckCandidateType(LI->getType());
1863  else if (auto *SI = dyn_cast<StoreInst>(S.getUse()->getUser()))
1864  CheckCandidateType(SI->getValueOperand()->getType());
1865  }
1866 
1867  // If we didn't find a vector type, nothing to do here.
1868  if (CandidateTys.empty())
1869  return nullptr;
1870 
1871  // Remove non-integer vector types if we had multiple common element types.
1872  // FIXME: It'd be nice to replace them with integer vector types, but we can't
1873  // do that until all the backends are known to produce good code for all
1874  // integer vector types.
1875  if (!HaveCommonEltTy) {
1876  CandidateTys.erase(
1877  llvm::remove_if(CandidateTys,
1878  [](VectorType *VTy) {
1879  return !VTy->getElementType()->isIntegerTy();
1880  }),
1881  CandidateTys.end());
1882 
1883  // If there were no integer vector types, give up.
1884  if (CandidateTys.empty())
1885  return nullptr;
1886 
1887  // Rank the remaining candidate vector types. This is easy because we know
1888  // they're all integer vectors. We sort by ascending number of elements.
1889  auto RankVectorTypes = [&DL](VectorType *RHSTy, VectorType *LHSTy) {
1890  (void)DL;
1891  assert(DL.getTypeSizeInBits(RHSTy) == DL.getTypeSizeInBits(LHSTy) &&
1892  "Cannot have vector types of different sizes!");
1893  assert(RHSTy->getElementType()->isIntegerTy() &&
1894  "All non-integer types eliminated!");
1895  assert(LHSTy->getElementType()->isIntegerTy() &&
1896  "All non-integer types eliminated!");
1897  return RHSTy->getNumElements() < LHSTy->getNumElements();
1898  };
1899  llvm::sort(CandidateTys, RankVectorTypes);
1900  CandidateTys.erase(
1901  std::unique(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes),
1902  CandidateTys.end());
1903  } else {
1904 // The only way to have the same element type in every vector type is to
1905 // have the same vector type. Check that and remove all but one.
1906 #ifndef NDEBUG
1907  for (VectorType *VTy : CandidateTys) {
1908  assert(VTy->getElementType() == CommonEltTy &&
1909  "Unaccounted for element type!");
1910  assert(VTy == CandidateTys[0] &&
1911  "Different vector types with the same element type!");
1912  }
1913 #endif
1914  CandidateTys.resize(1);
1915  }
1916 
1917  // Try each vector type, and return the one which works.
1918  auto CheckVectorTypeForPromotion = [&](VectorType *VTy) {
1919  uint64_t ElementSize = DL.getTypeSizeInBits(VTy->getElementType());
1920 
1921  // While the definition of LLVM vectors is bitpacked, we don't support sizes
1922  // that aren't byte sized.
1923  if (ElementSize % 8)
1924  return false;
1925  assert((DL.getTypeSizeInBits(VTy) % 8) == 0 &&
1926  "vector size not a multiple of element size?");
1927  ElementSize /= 8;
1928 
1929  for (const Slice &S : P)
1930  if (!isVectorPromotionViableForSlice(P, S, VTy, ElementSize, DL))
1931  return false;
1932 
1933  for (const Slice *S : P.splitSliceTails())
1934  if (!isVectorPromotionViableForSlice(P, *S, VTy, ElementSize, DL))
1935  return false;
1936 
1937  return true;
1938  };
1939  for (VectorType *VTy : CandidateTys)
1940  if (CheckVectorTypeForPromotion(VTy))
1941  return VTy;
1942 
1943  return nullptr;
1944 }
1945 
1946 /// Test whether a slice of an alloca is valid for integer widening.
1947 ///
1948 /// This implements the necessary checking for the \c isIntegerWideningViable
1949 /// test below on a single slice of the alloca.
1950 static bool isIntegerWideningViableForSlice(const Slice &S,
1951  uint64_t AllocBeginOffset,
1952  Type *AllocaTy,
1953  const DataLayout &DL,
1954  bool &WholeAllocaOp) {
1955  uint64_t Size = DL.getTypeStoreSize(AllocaTy);
1956 
1957  uint64_t RelBegin = S.beginOffset() - AllocBeginOffset;
1958  uint64_t RelEnd = S.endOffset() - AllocBeginOffset;
1959 
1960  // We can't reasonably handle cases where the load or store extends past
1961  // the end of the alloca's type and into its padding.
1962  if (RelEnd > Size)
1963  return false;
1964 
1965  Use *U = S.getUse();
1966 
1967  if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1968  if (LI->isVolatile())
1969  return false;
1970  // We can't handle loads that extend past the allocated memory.
1971  if (DL.getTypeStoreSize(LI->getType()) > Size)
1972  return false;
1973  // So far, AllocaSliceRewriter does not support widening split slice tails
1974  // in rewriteIntegerLoad.
1975  if (S.beginOffset() < AllocBeginOffset)
1976  return false;
1977  // Note that we don't count vector loads or stores as whole-alloca
1978  // operations which enable integer widening because we would prefer to use
1979  // vector widening instead.
1980  if (!isa<VectorType>(LI->getType()) && RelBegin == 0 && RelEnd == Size)
1981  WholeAllocaOp = true;
1982  if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
1983  if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1984  return false;
1985  } else if (RelBegin != 0 || RelEnd != Size ||
1986  !canConvertValue(DL, AllocaTy, LI->getType())) {
1987  // Non-integer loads need to be convertible from the alloca type so that
1988  // they are promotable.
1989  return false;
1990  }
1991  } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1992  Type *ValueTy = SI->getValueOperand()->getType();
1993  if (SI->isVolatile())
1994  return false;
1995  // We can't handle stores that extend past the allocated memory.
1996  if (DL.getTypeStoreSize(ValueTy) > Size)
1997  return false;
1998  // So far, AllocaSliceRewriter does not support widening split slice tails
1999  // in rewriteIntegerStore.
2000  if (S.beginOffset() < AllocBeginOffset)
2001  return false;
2002  // Note that we don't count vector loads or stores as whole-alloca
2003  // operations which enable integer widening because we would prefer to use
2004  // vector widening instead.
2005  if (!isa<VectorType>(ValueTy) && RelBegin == 0 && RelEnd == Size)
2006  WholeAllocaOp = true;
2007  if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
2008  if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
2009  return false;
2010  } else if (RelBegin != 0 || RelEnd != Size ||
2011  !canConvertValue(DL, ValueTy, AllocaTy)) {
2012  // Non-integer stores need to be convertible to the alloca type so that
2013  // they are promotable.
2014  return false;
2015  }
2016  } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
2017  if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
2018  return false;
2019  if (!S.isSplittable())
2020  return false; // Skip any unsplittable intrinsics.
2021  } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
2022  if (!II->isLifetimeStartOrEnd())
2023  return false;
2024  } else {
2025  return false;
2026  }
2027 
2028  return true;
2029 }
2030 
2031 /// Test whether the given alloca partition's integer operations can be
2032 /// widened to promotable ones.
2033 ///
2034 /// This is a quick test to check whether we can rewrite the integer loads and
2035 /// stores to a particular alloca into wider loads and stores and be able to
2036 /// promote the resulting alloca.
2037 static bool isIntegerWideningViable(Partition &P, Type *AllocaTy,
2038  const DataLayout &DL) {
2039  uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
2040  // Don't create integer types larger than the maximum bitwidth.
2041  if (SizeInBits > IntegerType::MAX_INT_BITS)
2042  return false;
2043 
2044  // Don't try to handle allocas with bit-padding.
2045  if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
2046  return false;
2047 
2048  // We need to ensure that an integer type with the appropriate bitwidth can
2049  // be converted to the alloca type, whatever that is. We don't want to force
2050  // the alloca itself to have an integer type if there is a more suitable one.
2051  Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
2052  if (!canConvertValue(DL, AllocaTy, IntTy) ||
2053  !canConvertValue(DL, IntTy, AllocaTy))
2054  return false;
2055 
2056  // While examining uses, we ensure that the alloca has a covering load or
2057  // store. We don't want to widen the integer operations only to fail to
2058  // promote due to some other unsplittable entry (which we may make splittable
2059  // later). However, if there are only splittable uses, go ahead and assume
2060  // that we cover the alloca.
2061  // FIXME: We shouldn't consider split slices that happen to start in the
2062  // partition here...
2063  bool WholeAllocaOp =
2064  P.begin() != P.end() ? false : DL.isLegalInteger(SizeInBits);
2065 
2066  for (const Slice &S : P)
2067  if (!isIntegerWideningViableForSlice(S, P.beginOffset(), AllocaTy, DL,
2068  WholeAllocaOp))
2069  return false;
2070 
2071  for (const Slice *S : P.splitSliceTails())
2072  if (!isIntegerWideningViableForSlice(*S, P.beginOffset(), AllocaTy, DL,
2073  WholeAllocaOp))
2074  return false;
2075 
2076  return WholeAllocaOp;
2077 }
2078 
2079 static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
2080  IntegerType *Ty, uint64_t Offset,
2081  const Twine &Name) {
2082  LLVM_DEBUG(dbgs() << " start: " << *V << "\n");
2083  IntegerType *IntTy = cast<IntegerType>(V->getType());
2084  assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2085  "Element extends past full value");
2086  uint64_t ShAmt = 8 * Offset;
2087  if (DL.isBigEndian())
2088  ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2089  if (ShAmt) {
2090  V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
2091  LLVM_DEBUG(dbgs() << " shifted: " << *V << "\n");
2092  }
2093  assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2094  "Cannot extract to a larger integer!");
2095  if (Ty != IntTy) {
2096  V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
2097  LLVM_DEBUG(dbgs() << " trunced: " << *V << "\n");
2098  }
2099  return V;
2100 }
2101 
2102 static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
2103  Value *V, uint64_t Offset, const Twine &Name) {
2104  IntegerType *IntTy = cast<IntegerType>(Old->getType());
2105  IntegerType *Ty = cast<IntegerType>(V->getType());
2106  assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2107  "Cannot insert a larger integer!");
2108  LLVM_DEBUG(dbgs() << " start: " << *V << "\n");
2109  if (Ty != IntTy) {
2110  V = IRB.CreateZExt(V, IntTy, Name + ".ext");
2111  LLVM_DEBUG(dbgs() << " extended: " << *V << "\n");
2112  }
2113  assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2114  "Element store outside of alloca store");
2115  uint64_t ShAmt = 8 * Offset;
2116  if (DL.isBigEndian())
2117  ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2118  if (ShAmt) {
2119  V = IRB.CreateShl(V, ShAmt, Name + ".shift");
2120  LLVM_DEBUG(dbgs() << " shifted: " << *V << "\n");
2121  }
2122 
2123  if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
2124  APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
2125  Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
2126  LLVM_DEBUG(dbgs() << " masked: " << *Old << "\n");
2127  V = IRB.CreateOr(Old, V, Name + ".insert");
2128  LLVM_DEBUG(dbgs() << " inserted: " << *V << "\n");
2129  }
2130  return V;
2131 }
2132 
2133 static Value *extractVector(IRBuilderTy &IRB, Value *V, unsigned BeginIndex,
2134  unsigned EndIndex, const Twine &Name) {
2135  VectorType *VecTy = cast<VectorType>(V->getType());
2136  unsigned NumElements = EndIndex - BeginIndex;
2137  assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2138 
2139  if (NumElements == VecTy->getNumElements())
2140  return V;
2141 
2142  if (NumElements == 1) {
2143  V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
2144  Name + ".extract");
2145  LLVM_DEBUG(dbgs() << " extract: " << *V << "\n");
2146  return V;
2147  }
2148 
2150  Mask.reserve(NumElements);
2151  for (unsigned i = BeginIndex; i != EndIndex; ++i)
2152  Mask.push_back(IRB.getInt32(i));
2153  V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2154  ConstantVector::get(Mask), Name + ".extract");
2155  LLVM_DEBUG(dbgs() << " shuffle: " << *V << "\n");
2156  return V;
2157 }
2158 
2159 static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
2160  unsigned BeginIndex, const Twine &Name) {
2161  VectorType *VecTy = cast<VectorType>(Old->getType());
2162  assert(VecTy && "Can only insert a vector into a vector");
2163 
2164  VectorType *Ty = dyn_cast<VectorType>(V->getType());
2165  if (!Ty) {
2166  // Single element to insert.
2167  V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
2168  Name + ".insert");
2169  LLVM_DEBUG(dbgs() << " insert: " << *V << "\n");
2170  return V;
2171  }
2172 
2173  assert(Ty->getNumElements() <= VecTy->getNumElements() &&
2174  "Too many elements!");
2175  if (Ty->getNumElements() == VecTy->getNumElements()) {
2176  assert(V->getType() == VecTy && "Vector type mismatch");
2177  return V;
2178  }
2179  unsigned EndIndex = BeginIndex + Ty->getNumElements();
2180 
2181  // When inserting a smaller vector into the larger to store, we first
2182  // use a shuffle vector to widen it with undef elements, and then
2183  // a second shuffle vector to select between the loaded vector and the
2184  // incoming vector.
2186  Mask.reserve(VecTy->getNumElements());
2187  for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2188  if (i >= BeginIndex && i < EndIndex)
2189  Mask.push_back(IRB.getInt32(i - BeginIndex));
2190  else
2191  Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
2192  V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2193  ConstantVector::get(Mask), Name + ".expand");
2194  LLVM_DEBUG(dbgs() << " shuffle: " << *V << "\n");
2195 
2196  Mask.clear();
2197  for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2198  Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
2199 
2200  V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
2201 
2202  LLVM_DEBUG(dbgs() << " blend: " << *V << "\n");
2203  return V;
2204 }
2205 
2206 /// Visitor to rewrite instructions using p particular slice of an alloca
2207 /// to use a new alloca.
2208 ///
2209 /// Also implements the rewriting to vector-based accesses when the partition
2210 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2211 /// lives here.
2213  : public InstVisitor<AllocaSliceRewriter, bool> {
2214  // Befriend the base class so it can delegate to private visit methods.
2216 
2218 
2219  const DataLayout &DL;
2220  AllocaSlices &AS;
2221  SROA &Pass;
2222  AllocaInst &OldAI, &NewAI;
2223  const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2224  Type *NewAllocaTy;
2225 
2226  // This is a convenience and flag variable that will be null unless the new
2227  // alloca's integer operations should be widened to this integer type due to
2228  // passing isIntegerWideningViable above. If it is non-null, the desired
2229  // integer type will be stored here for easy access during rewriting.
2230  IntegerType *IntTy;
2231 
2232  // If we are rewriting an alloca partition which can be written as pure
2233  // vector operations, we stash extra information here. When VecTy is
2234  // non-null, we have some strict guarantees about the rewritten alloca:
2235  // - The new alloca is exactly the size of the vector type here.
2236  // - The accesses all either map to the entire vector or to a single
2237  // element.
2238  // - The set of accessing instructions is only one of those handled above
2239  // in isVectorPromotionViable. Generally these are the same access kinds
2240  // which are promotable via mem2reg.
2241  VectorType *VecTy;
2242  Type *ElementTy;
2243  uint64_t ElementSize;
2244 
2245  // The original offset of the slice currently being rewritten relative to
2246  // the original alloca.
2247  uint64_t BeginOffset = 0;
2248  uint64_t EndOffset = 0;
2249 
2250  // The new offsets of the slice currently being rewritten relative to the
2251  // original alloca.
2252  uint64_t NewBeginOffset, NewEndOffset;
2253 
2254  uint64_t SliceSize;
2255  bool IsSplittable = false;
2256  bool IsSplit = false;
2257  Use *OldUse = nullptr;
2258  Instruction *OldPtr = nullptr;
2259 
2260  // Track post-rewrite users which are PHI nodes and Selects.
2261  SmallSetVector<PHINode *, 8> &PHIUsers;
2262  SmallSetVector<SelectInst *, 8> &SelectUsers;
2263 
2264  // Utility IR builder, whose name prefix is setup for each visited use, and
2265  // the insertion point is set to point to the user.
2266  IRBuilderTy IRB;
2267 
2268 public:
2270  AllocaInst &OldAI, AllocaInst &NewAI,
2271  uint64_t NewAllocaBeginOffset,
2272  uint64_t NewAllocaEndOffset, bool IsIntegerPromotable,
2273  VectorType *PromotableVecTy,
2274  SmallSetVector<PHINode *, 8> &PHIUsers,
2275  SmallSetVector<SelectInst *, 8> &SelectUsers)
2276  : DL(DL), AS(AS), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
2277  NewAllocaBeginOffset(NewAllocaBeginOffset),
2278  NewAllocaEndOffset(NewAllocaEndOffset),
2279  NewAllocaTy(NewAI.getAllocatedType()),
2280  IntTy(IsIntegerPromotable
2281  ? Type::getIntNTy(
2282  NewAI.getContext(),
2283  DL.getTypeSizeInBits(NewAI.getAllocatedType()))
2284  : nullptr),
2285  VecTy(PromotableVecTy),
2286  ElementTy(VecTy ? VecTy->getElementType() : nullptr),
2287  ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
2288  PHIUsers(PHIUsers), SelectUsers(SelectUsers),
2289  IRB(NewAI.getContext(), ConstantFolder()) {
2290  if (VecTy) {
2291  assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
2292  "Only multiple-of-8 sized vector elements are viable");
2293  ++NumVectorized;
2294  }
2295  assert((!IntTy && !VecTy) || (IntTy && !VecTy) || (!IntTy && VecTy));
2296  }
2297 
2299  bool CanSROA = true;
2300  BeginOffset = I->beginOffset();
2301  EndOffset = I->endOffset();
2302  IsSplittable = I->isSplittable();
2303  IsSplit =
2304  BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
2305  LLVM_DEBUG(dbgs() << " rewriting " << (IsSplit ? "split " : ""));
2306  LLVM_DEBUG(AS.printSlice(dbgs(), I, ""));
2307  LLVM_DEBUG(dbgs() << "\n");
2308 
2309  // Compute the intersecting offset range.
2310  assert(BeginOffset < NewAllocaEndOffset);
2311  assert(EndOffset > NewAllocaBeginOffset);
2312  NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2313  NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2314 
2315  SliceSize = NewEndOffset - NewBeginOffset;
2316 
2317  OldUse = I->getUse();
2318  OldPtr = cast<Instruction>(OldUse->get());
2319 
2320  Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2321  IRB.SetInsertPoint(OldUserI);
2322  IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2323  IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
2324 
2325  CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2326  if (VecTy || IntTy)
2327  assert(CanSROA);
2328  return CanSROA;
2329  }
2330 
2331 private:
2332  // Make sure the other visit overloads are visible.
2333  using Base::visit;
2334 
2335  // Every instruction which can end up as a user must have a rewrite rule.
2336  bool visitInstruction(Instruction &I) {
2337  LLVM_DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2338  llvm_unreachable("No rewrite rule for this instruction!");
2339  }
2340 
2341  Value *getNewAllocaSlicePtr(IRBuilderTy &IRB, Type *PointerTy) {
2342  // Note that the offset computation can use BeginOffset or NewBeginOffset
2343  // interchangeably for unsplit slices.
2344  assert(IsSplit || BeginOffset == NewBeginOffset);
2345  uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2346 
2347 #ifndef NDEBUG
2348  StringRef OldName = OldPtr->getName();
2349  // Skip through the last '.sroa.' component of the name.
2350  size_t LastSROAPrefix = OldName.rfind(".sroa.");
2351  if (LastSROAPrefix != StringRef::npos) {
2352  OldName = OldName.substr(LastSROAPrefix + strlen(".sroa."));
2353  // Look for an SROA slice index.
2354  size_t IndexEnd = OldName.find_first_not_of("0123456789");
2355  if (IndexEnd != StringRef::npos && OldName[IndexEnd] == '.') {
2356  // Strip the index and look for the offset.
2357  OldName = OldName.substr(IndexEnd + 1);
2358  size_t OffsetEnd = OldName.find_first_not_of("0123456789");
2359  if (OffsetEnd != StringRef::npos && OldName[OffsetEnd] == '.')
2360  // Strip the offset.
2361  OldName = OldName.substr(OffsetEnd + 1);
2362  }
2363  }
2364  // Strip any SROA suffixes as well.
2365  OldName = OldName.substr(0, OldName.find(".sroa_"));
2366 #endif
2367 
2368  return getAdjustedPtr(IRB, DL, &NewAI,
2369  APInt(DL.getIndexTypeSizeInBits(PointerTy), Offset),
2370  PointerTy,
2371 #ifndef NDEBUG
2372  Twine(OldName) + "."
2373 #else
2374  Twine()
2375 #endif
2376  );
2377  }
2378 
2379  /// Compute suitable alignment to access this slice of the *new*
2380  /// alloca.
2381  ///
2382  /// You can optionally pass a type to this routine and if that type's ABI
2383  /// alignment is itself suitable, this will return zero.
2384  unsigned getSliceAlign(Type *Ty = nullptr) {
2385  unsigned NewAIAlign = NewAI.getAlignment();
2386  if (!NewAIAlign)
2387  NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
2388  unsigned Align =
2389  MinAlign(NewAIAlign, NewBeginOffset - NewAllocaBeginOffset);
2390  return (Ty && Align == DL.getABITypeAlignment(Ty)) ? 0 : Align;
2391  }
2392 
2393  unsigned getIndex(uint64_t Offset) {
2394  assert(VecTy && "Can only call getIndex when rewriting a vector");
2395  uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2396  assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2397  uint32_t Index = RelOffset / ElementSize;
2398  assert(Index * ElementSize == RelOffset);
2399  return Index;
2400  }
2401 
2402  void deleteIfTriviallyDead(Value *V) {
2403  Instruction *I = cast<Instruction>(V);
2405  Pass.DeadInsts.insert(I);
2406  }
2407 
2408  Value *rewriteVectorizedLoadInst() {
2409  unsigned BeginIndex = getIndex(NewBeginOffset);
2410  unsigned EndIndex = getIndex(NewEndOffset);
2411  assert(EndIndex > BeginIndex && "Empty vector!");
2412 
2413  Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2414  return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2415  }
2416 
2417  Value *rewriteIntegerLoad(LoadInst &LI) {
2418  assert(IntTy && "We cannot insert an integer to the alloca");
2419  assert(!LI.isVolatile());
2420  Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2421  V = convertValue(DL, IRB, V, IntTy);
2422  assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2423  uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2424  if (Offset > 0 || NewEndOffset < NewAllocaEndOffset) {
2425  IntegerType *ExtractTy = Type::getIntNTy(LI.getContext(), SliceSize * 8);
2426  V = extractInteger(DL, IRB, V, ExtractTy, Offset, "extract");
2427  }
2428  // It is possible that the extracted type is not the load type. This
2429  // happens if there is a load past the end of the alloca, and as
2430  // a consequence the slice is narrower but still a candidate for integer
2431  // lowering. To handle this case, we just zero extend the extracted
2432  // integer.
2433  assert(cast<IntegerType>(LI.getType())->getBitWidth() >= SliceSize * 8 &&
2434  "Can only handle an extract for an overly wide load");
2435  if (cast<IntegerType>(LI.getType())->getBitWidth() > SliceSize * 8)
2436  V = IRB.CreateZExt(V, LI.getType());
2437  return V;
2438  }
2439 
2440  bool visitLoadInst(LoadInst &LI) {
2441  LLVM_DEBUG(dbgs() << " original: " << LI << "\n");
2442  Value *OldOp = LI.getOperand(0);
2443  assert(OldOp == OldPtr);
2444 
2445  AAMDNodes AATags;
2446  LI.getAAMetadata(AATags);
2447 
2448  unsigned AS = LI.getPointerAddressSpace();
2449 
2450  Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), SliceSize * 8)
2451  : LI.getType();
2452  const bool IsLoadPastEnd = DL.getTypeStoreSize(TargetTy) > SliceSize;
2453  bool IsPtrAdjusted = false;
2454  Value *V;
2455  if (VecTy) {
2456  V = rewriteVectorizedLoadInst();
2457  } else if (IntTy && LI.getType()->isIntegerTy()) {
2458  V = rewriteIntegerLoad(LI);
2459  } else if (NewBeginOffset == NewAllocaBeginOffset &&
2460  NewEndOffset == NewAllocaEndOffset &&
2461  (canConvertValue(DL, NewAllocaTy, TargetTy) ||
2462  (IsLoadPastEnd && NewAllocaTy->isIntegerTy() &&
2463  TargetTy->isIntegerTy()))) {
2464  LoadInst *NewLI = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2465  LI.isVolatile(), LI.getName());
2466  if (AATags)
2467  NewLI->setAAMetadata(AATags);
2468  if (LI.isVolatile())
2469  NewLI->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
2470 
2471  // Any !nonnull metadata or !range metadata on the old load is also valid
2472  // on the new load. This is even true in some cases even when the loads
2473  // are different types, for example by mapping !nonnull metadata to
2474  // !range metadata by modeling the null pointer constant converted to the
2475  // integer type.
2476  // FIXME: Add support for range metadata here. Currently the utilities
2477  // for this don't propagate range metadata in trivial cases from one
2478  // integer load to another, don't handle non-addrspace-0 null pointers
2479  // correctly, and don't have any support for mapping ranges as the
2480  // integer type becomes winder or narrower.
2482  copyNonnullMetadata(LI, N, *NewLI);
2483 
2484  // Try to preserve nonnull metadata
2485  V = NewLI;
2486 
2487  // If this is an integer load past the end of the slice (which means the
2488  // bytes outside the slice are undef or this load is dead) just forcibly
2489  // fix the integer size with correct handling of endianness.
2490  if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2491  if (auto *TITy = dyn_cast<IntegerType>(TargetTy))
2492  if (AITy->getBitWidth() < TITy->getBitWidth()) {
2493  V = IRB.CreateZExt(V, TITy, "load.ext");
2494  if (DL.isBigEndian())
2495  V = IRB.CreateShl(V, TITy->getBitWidth() - AITy->getBitWidth(),
2496  "endian_shift");
2497  }
2498  } else {
2499  Type *LTy = TargetTy->getPointerTo(AS);
2500  LoadInst *NewLI = IRB.CreateAlignedLoad(getNewAllocaSlicePtr(IRB, LTy),
2501  getSliceAlign(TargetTy),
2502  LI.isVolatile(), LI.getName());
2503  if (AATags)
2504  NewLI->setAAMetadata(AATags);
2505  if (LI.isVolatile())
2506  NewLI->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
2507 
2508  V = NewLI;
2509  IsPtrAdjusted = true;
2510  }
2511  V = convertValue(DL, IRB, V, TargetTy);
2512 
2513  if (IsSplit) {
2514  assert(!LI.isVolatile());
2515  assert(LI.getType()->isIntegerTy() &&
2516  "Only integer type loads and stores are split");
2517  assert(SliceSize < DL.getTypeStoreSize(LI.getType()) &&
2518  "Split load isn't smaller than original load");
2519  assert(LI.getType()->getIntegerBitWidth() ==
2520  DL.getTypeStoreSizeInBits(LI.getType()) &&
2521  "Non-byte-multiple bit width");
2522  // Move the insertion point just past the load so that we can refer to it.
2523  IRB.SetInsertPoint(&*std::next(BasicBlock::iterator(&LI)));
2524  // Create a placeholder value with the same type as LI to use as the
2525  // basis for the new value. This allows us to replace the uses of LI with
2526  // the computed value, and then replace the placeholder with LI, leaving
2527  // LI only used for this computation.
2528  Value *Placeholder =
2529  new LoadInst(UndefValue::get(LI.getType()->getPointerTo(AS)));
2530  V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset - BeginOffset,
2531  "insert");
2532  LI.replaceAllUsesWith(V);
2533  Placeholder->replaceAllUsesWith(&LI);
2534  Placeholder->deleteValue();
2535  } else {
2536  LI.replaceAllUsesWith(V);
2537  }
2538 
2539  Pass.DeadInsts.insert(&LI);
2540  deleteIfTriviallyDead(OldOp);
2541  LLVM_DEBUG(dbgs() << " to: " << *V << "\n");
2542  return !LI.isVolatile() && !IsPtrAdjusted;
2543  }
2544 
2545  bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp,
2546  AAMDNodes AATags) {
2547  if (V->getType() != VecTy) {
2548  unsigned BeginIndex = getIndex(NewBeginOffset);
2549  unsigned EndIndex = getIndex(NewEndOffset);
2550  assert(EndIndex > BeginIndex && "Empty vector!");
2551  unsigned NumElements = EndIndex - BeginIndex;
2552  assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2553  Type *SliceTy = (NumElements == 1)
2554  ? ElementTy
2555  : VectorType::get(ElementTy, NumElements);
2556  if (V->getType() != SliceTy)
2557  V = convertValue(DL, IRB, V, SliceTy);
2558 
2559  // Mix in the existing elements.
2560  Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2561  V = insertVector(IRB, Old, V, BeginIndex, "vec");
2562  }
2563  StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2564  if (AATags)
2565  Store->setAAMetadata(AATags);
2566  Pass.DeadInsts.insert(&SI);
2567 
2568  LLVM_DEBUG(dbgs() << " to: " << *Store << "\n");
2569  return true;
2570  }
2571 
2572  bool rewriteIntegerStore(Value *V, StoreInst &SI, AAMDNodes AATags) {
2573  assert(IntTy && "We cannot extract an integer from the alloca");
2574  assert(!SI.isVolatile());
2575  if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2576  Value *Old =
2577  IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2578  Old = convertValue(DL, IRB, Old, IntTy);
2579  assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2580  uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2581  V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset, "insert");
2582  }
2583  V = convertValue(DL, IRB, V, NewAllocaTy);
2584  StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2587  if (AATags)
2588  Store->setAAMetadata(AATags);
2589  Pass.DeadInsts.insert(&SI);
2590  LLVM_DEBUG(dbgs() << " to: " << *Store << "\n");
2591  return true;
2592  }
2593 
2594  bool visitStoreInst(StoreInst &SI) {
2595  LLVM_DEBUG(dbgs() << " original: " << SI << "\n");
2596  Value *OldOp = SI.getOperand(1);
2597  assert(OldOp == OldPtr);
2598 
2599  AAMDNodes AATags;
2600  SI.getAAMetadata(AATags);
2601 
2602  Value *V = SI.getValueOperand();
2603 
2604  // Strip all inbounds GEPs and pointer casts to try to dig out any root
2605  // alloca that should be re-examined after promoting this alloca.
2606  if (V->getType()->isPointerTy())
2607  if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2608  Pass.PostPromotionWorklist.insert(AI);
2609 
2610  if (SliceSize < DL.getTypeStoreSize(V->getType())) {
2611  assert(!SI.isVolatile());
2612  assert(V->getType()->isIntegerTy() &&
2613  "Only integer type loads and stores are split");
2614  assert(V->getType()->getIntegerBitWidth() ==
2615  DL.getTypeStoreSizeInBits(V->getType()) &&
2616  "Non-byte-multiple bit width");
2617  IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), SliceSize * 8);
2618  V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset - BeginOffset,
2619  "extract");
2620  }
2621 
2622  if (VecTy)
2623  return rewriteVectorizedStoreInst(V, SI, OldOp, AATags);
2624  if (IntTy && V->getType()->isIntegerTy())
2625  return rewriteIntegerStore(V, SI, AATags);
2626 
2627  const bool IsStorePastEnd = DL.getTypeStoreSize(V->getType()) > SliceSize;
2628  StoreInst *NewSI;
2629  if (NewBeginOffset == NewAllocaBeginOffset &&
2630  NewEndOffset == NewAllocaEndOffset &&
2631  (canConvertValue(DL, V->getType(), NewAllocaTy) ||
2632  (IsStorePastEnd && NewAllocaTy->isIntegerTy() &&
2633  V->getType()->isIntegerTy()))) {
2634  // If this is an integer store past the end of slice (and thus the bytes
2635  // past that point are irrelevant or this is unreachable), truncate the
2636  // value prior to storing.
2637  if (auto *VITy = dyn_cast<IntegerType>(V->getType()))
2638  if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2639  if (VITy->getBitWidth() > AITy->getBitWidth()) {
2640  if (DL.isBigEndian())
2641  V = IRB.CreateLShr(V, VITy->getBitWidth() - AITy->getBitWidth(),
2642  "endian_shift");
2643  V = IRB.CreateTrunc(V, AITy, "load.trunc");
2644  }
2645 
2646  V = convertValue(DL, IRB, V, NewAllocaTy);
2647  NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2648  SI.isVolatile());
2649  } else {
2650  unsigned AS = SI.getPointerAddressSpace();
2651  Value *NewPtr = getNewAllocaSlicePtr(IRB, V->getType()->getPointerTo(AS));
2652  NewSI = IRB.CreateAlignedStore(V, NewPtr, getSliceAlign(V->getType()),
2653  SI.isVolatile());
2654  }
2657  if (AATags)
2658  NewSI->setAAMetadata(AATags);
2659  if (SI.isVolatile())
2660  NewSI->setAtomic(SI.getOrdering(), SI.getSyncScopeID());
2661  Pass.DeadInsts.insert(&SI);
2662  deleteIfTriviallyDead(OldOp);
2663 
2664  LLVM_DEBUG(dbgs() << " to: " << *NewSI << "\n");
2665  return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2666  }
2667 
2668  /// Compute an integer value from splatting an i8 across the given
2669  /// number of bytes.
2670  ///
2671  /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2672  /// call this routine.
2673  /// FIXME: Heed the advice above.
2674  ///
2675  /// \param V The i8 value to splat.
2676  /// \param Size The number of bytes in the output (assuming i8 is one byte)
2677  Value *getIntegerSplat(Value *V, unsigned Size) {
2678  assert(Size > 0 && "Expected a positive number of bytes.");
2679  IntegerType *VTy = cast<IntegerType>(V->getType());
2680  assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2681  if (Size == 1)
2682  return V;
2683 
2684  Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size * 8);
2685  V = IRB.CreateMul(
2686  IRB.CreateZExt(V, SplatIntTy, "zext"),
2688  Constant::getAllOnesValue(SplatIntTy),
2690  SplatIntTy)),
2691  "isplat");
2692  return V;
2693  }
2694 
2695  /// Compute a vector splat for a given element value.
2696  Value *getVectorSplat(Value *V, unsigned NumElements) {
2697  V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2698  LLVM_DEBUG(dbgs() << " splat: " << *V << "\n");
2699  return V;
2700  }
2701 
2702  bool visitMemSetInst(MemSetInst &II) {
2703  LLVM_DEBUG(dbgs() << " original: " << II << "\n");
2704  assert(II.getRawDest() == OldPtr);
2705 
2706  AAMDNodes AATags;
2707  II.getAAMetadata(AATags);
2708 
2709  // If the memset has a variable size, it cannot be split, just adjust the
2710  // pointer to the new alloca.
2711  if (!isa<Constant>(II.getLength())) {
2712  assert(!IsSplit);
2713  assert(NewBeginOffset == BeginOffset);
2714  II.setDest(getNewAllocaSlicePtr(IRB, OldPtr->getType()));
2715  II.setDestAlignment(getSliceAlign());
2716 
2717  deleteIfTriviallyDead(OldPtr);
2718  return false;
2719  }
2720 
2721  // Record this instruction for deletion.
2722  Pass.DeadInsts.insert(&II);
2723 
2724  Type *AllocaTy = NewAI.getAllocatedType();
2725  Type *ScalarTy = AllocaTy->getScalarType();
2726 
2727  // If this doesn't map cleanly onto the alloca type, and that type isn't
2728  // a single value type, just emit a memset.
2729  if (!VecTy && !IntTy &&
2730  (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2731  SliceSize != DL.getTypeStoreSize(AllocaTy) ||
2732  !AllocaTy->isSingleValueType() ||
2733  !DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy)) ||
2734  DL.getTypeSizeInBits(ScalarTy) % 8 != 0)) {
2735  Type *SizeTy = II.getLength()->getType();
2736  Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2737  CallInst *New = IRB.CreateMemSet(
2738  getNewAllocaSlicePtr(IRB, OldPtr->getType()), II.getValue(), Size,
2739  getSliceAlign(), II.isVolatile());
2740  if (AATags)
2741  New->setAAMetadata(AATags);
2742  LLVM_DEBUG(dbgs() << " to: " << *New << "\n");
2743  return false;
2744  }
2745 
2746  // If we can represent this as a simple value, we have to build the actual
2747  // value to store, which requires expanding the byte present in memset to
2748  // a sensible representation for the alloca type. This is essentially
2749  // splatting the byte to a sufficiently wide integer, splatting it across
2750  // any desired vector width, and bitcasting to the final type.
2751  Value *V;
2752 
2753  if (VecTy) {
2754  // If this is a memset of a vectorized alloca, insert it.
2755  assert(ElementTy == ScalarTy);
2756 
2757  unsigned BeginIndex = getIndex(NewBeginOffset);
2758  unsigned EndIndex = getIndex(NewEndOffset);
2759  assert(EndIndex > BeginIndex && "Empty vector!");
2760  unsigned NumElements = EndIndex - BeginIndex;
2761  assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2762 
2763  Value *Splat =
2764  getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
2765  Splat = convertValue(DL, IRB, Splat, ElementTy);
2766  if (NumElements > 1)
2767  Splat = getVectorSplat(Splat, NumElements);
2768 
2769  Value *Old =
2770  IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2771  V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2772  } else if (IntTy) {
2773  // If this is a memset on an alloca where we can widen stores, insert the
2774  // set integer.
2775  assert(!II.isVolatile());
2776 
2777  uint64_t Size = NewEndOffset - NewBeginOffset;
2778  V = getIntegerSplat(II.getValue(), Size);
2779 
2780  if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2781  EndOffset != NewAllocaBeginOffset)) {
2782  Value *Old =
2783  IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2784  Old = convertValue(DL, IRB, Old, IntTy);
2785  uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2786  V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2787  } else {
2788  assert(V->getType() == IntTy &&
2789  "Wrong type for an alloca wide integer!");
2790  }
2791  V = convertValue(DL, IRB, V, AllocaTy);
2792  } else {
2793  // Established these invariants above.
2794  assert(NewBeginOffset == NewAllocaBeginOffset);
2795  assert(NewEndOffset == NewAllocaEndOffset);
2796 
2797  V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
2798  if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2799  V = getVectorSplat(V, AllocaVecTy->getNumElements());
2800 
2801  V = convertValue(DL, IRB, V, AllocaTy);
2802  }
2803 
2804  StoreInst *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2805  II.isVolatile());
2806  if (AATags)
2807  New->setAAMetadata(AATags);
2808  LLVM_DEBUG(dbgs() << " to: " << *New << "\n");
2809  return !II.isVolatile();
2810  }
2811 
2812  bool visitMemTransferInst(MemTransferInst &II) {
2813  // Rewriting of memory transfer instructions can be a bit tricky. We break
2814  // them into two categories: split intrinsics and unsplit intrinsics.
2815 
2816  LLVM_DEBUG(dbgs() << " original: " << II << "\n");
2817 
2818  AAMDNodes AATags;
2819  II.getAAMetadata(AATags);
2820 
2821  bool IsDest = &II.getRawDestUse() == OldUse;
2822  assert((IsDest && II.getRawDest() == OldPtr) ||
2823  (!IsDest && II.getRawSource() == OldPtr));
2824 
2825  unsigned SliceAlign = getSliceAlign();
2826 
2827  // For unsplit intrinsics, we simply modify the source and destination
2828  // pointers in place. This isn't just an optimization, it is a matter of
2829  // correctness. With unsplit intrinsics we may be dealing with transfers
2830  // within a single alloca before SROA ran, or with transfers that have
2831  // a variable length. We may also be dealing with memmove instead of
2832  // memcpy, and so simply updating the pointers is the necessary for us to
2833  // update both source and dest of a single call.
2834  if (!IsSplittable) {
2835  Value *AdjustedPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2836  if (IsDest) {
2837  II.setDest(AdjustedPtr);
2838  II.setDestAlignment(SliceAlign);
2839  }
2840  else {
2841  II.setSource(AdjustedPtr);
2842  II.setSourceAlignment(SliceAlign);
2843  }
2844 
2845  LLVM_DEBUG(dbgs() << " to: " << II << "\n");
2846  deleteIfTriviallyDead(OldPtr);
2847  return false;
2848  }
2849  // For split transfer intrinsics we have an incredibly useful assurance:
2850  // the source and destination do not reside within the same alloca, and at
2851  // least one of them does not escape. This means that we can replace
2852  // memmove with memcpy, and we don't need to worry about all manner of
2853  // downsides to splitting and transforming the operations.
2854 
2855  // If this doesn't map cleanly onto the alloca type, and that type isn't
2856  // a single value type, just emit a memcpy.
2857  bool EmitMemCpy =
2858  !VecTy && !IntTy &&
2859  (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2860  SliceSize != DL.getTypeStoreSize(NewAI.getAllocatedType()) ||
2861  !NewAI.getAllocatedType()->isSingleValueType());
2862 
2863  // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2864  // size hasn't been shrunk based on analysis of the viable range, this is
2865  // a no-op.
2866  if (EmitMemCpy && &OldAI == &NewAI) {
2867  // Ensure the start lines up.
2868  assert(NewBeginOffset == BeginOffset);
2869 
2870  // Rewrite the size as needed.
2871  if (NewEndOffset != EndOffset)
2873  NewEndOffset - NewBeginOffset));
2874  return false;
2875  }
2876  // Record this instruction for deletion.
2877  Pass.DeadInsts.insert(&II);
2878 
2879  // Strip all inbounds GEPs and pointer casts to try to dig out any root
2880  // alloca that should be re-examined after rewriting this instruction.
2881  Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2882  if (AllocaInst *AI =
2883  dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
2884  assert(AI != &OldAI && AI != &NewAI &&
2885  "Splittable transfers cannot reach the same alloca on both ends.");
2886  Pass.Worklist.insert(AI);
2887  }
2888 
2889  Type *OtherPtrTy = OtherPtr->getType();
2890  unsigned OtherAS = OtherPtrTy->getPointerAddressSpace();
2891 
2892  // Compute the relative offset for the other pointer within the transfer.
2893  unsigned OffsetWidth = DL.getIndexSizeInBits(OtherAS);
2894  APInt OtherOffset(OffsetWidth, NewBeginOffset - BeginOffset);
2895  unsigned OtherAlign =
2896  IsDest ? II.getSourceAlignment() : II.getDestAlignment();
2897  OtherAlign = MinAlign(OtherAlign ? OtherAlign : 1,
2898  OtherOffset.zextOrTrunc(64).getZExtValue());
2899 
2900  if (EmitMemCpy) {
2901  // Compute the other pointer, folding as much as possible to produce
2902  // a single, simple GEP in most cases.
2903  OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2904  OtherPtr->getName() + ".");
2905 
2906  Value *OurPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2907  Type *SizeTy = II.getLength()->getType();
2908  Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2909 
2910  Value *DestPtr, *SrcPtr;
2911  unsigned DestAlign, SrcAlign;
2912  // Note: IsDest is true iff we're copying into the new alloca slice
2913  if (IsDest) {
2914  DestPtr = OurPtr;
2915  DestAlign = SliceAlign;
2916  SrcPtr = OtherPtr;
2917  SrcAlign = OtherAlign;
2918  } else {
2919  DestPtr = OtherPtr;
2920  DestAlign = OtherAlign;
2921  SrcPtr = OurPtr;
2922  SrcAlign = SliceAlign;
2923  }
2924  CallInst *New = IRB.CreateMemCpy(DestPtr, DestAlign, SrcPtr, SrcAlign,
2925  Size, II.isVolatile());
2926  if (AATags)
2927  New->setAAMetadata(AATags);
2928  LLVM_DEBUG(dbgs() << " to: " << *New << "\n");
2929  return false;
2930  }
2931 
2932  bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
2933  NewEndOffset == NewAllocaEndOffset;
2934  uint64_t Size = NewEndOffset - NewBeginOffset;
2935  unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
2936  unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
2937  unsigned NumElements = EndIndex - BeginIndex;
2938  IntegerType *SubIntTy =
2939  IntTy ? Type::getIntNTy(IntTy->getContext(), Size * 8) : nullptr;
2940 
2941  // Reset the other pointer type to match the register type we're going to
2942  // use, but using the address space of the original other pointer.
2943  if (VecTy && !IsWholeAlloca) {
2944  if (NumElements == 1)
2945  OtherPtrTy = VecTy->getElementType();
2946  else
2947  OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2948 
2949  OtherPtrTy = OtherPtrTy->getPointerTo(OtherAS);
2950  } else if (IntTy && !IsWholeAlloca) {
2951  OtherPtrTy = SubIntTy->getPointerTo(OtherAS);
2952  } else {
2953  OtherPtrTy = NewAllocaTy->getPointerTo(OtherAS);
2954  }
2955 
2956  Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2957  OtherPtr->getName() + ".");
2958  unsigned SrcAlign = OtherAlign;
2959  Value *DstPtr = &NewAI;
2960  unsigned DstAlign = SliceAlign;
2961  if (!IsDest) {
2962  std::swap(SrcPtr, DstPtr);
2963  std::swap(SrcAlign, DstAlign);
2964  }
2965 
2966  Value *Src;
2967  if (VecTy && !IsWholeAlloca && !IsDest) {
2968  Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2969  Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
2970  } else if (IntTy && !IsWholeAlloca && !IsDest) {
2971  Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2972  Src = convertValue(DL, IRB, Src, IntTy);
2973  uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2974  Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
2975  } else {
2976  LoadInst *Load = IRB.CreateAlignedLoad(SrcPtr, SrcAlign, II.isVolatile(),
2977  "copyload");
2978  if (AATags)
2979  Load->setAAMetadata(AATags);
2980  Src = Load;
2981  }
2982 
2983  if (VecTy && !IsWholeAlloca && IsDest) {
2984  Value *Old =
2985  IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2986  Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
2987  } else if (IntTy && !IsWholeAlloca && IsDest) {
2988  Value *Old =
2989  IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2990  Old = convertValue(DL, IRB, Old, IntTy);
2991  uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2992  Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
2993  Src = convertValue(DL, IRB, Src, NewAllocaTy);
2994  }
2995 
2996  StoreInst *Store = cast<StoreInst>(
2997  IRB.CreateAlignedStore(Src, DstPtr, DstAlign, II.isVolatile()));
2998  if (AATags)
2999  Store->setAAMetadata(AATags);
3000  LLVM_DEBUG(dbgs() << " to: " << *Store << "\n");
3001  return !II.isVolatile();
3002  }
3003 
3004  bool visitIntrinsicInst(IntrinsicInst &II) {
3006  LLVM_DEBUG(dbgs() << " original: " << II << "\n");
3007  assert(II.getArgOperand(1) == OldPtr);
3008 
3009  // Record this instruction for deletion.
3010  Pass.DeadInsts.insert(&II);
3011 
3012  // Lifetime intrinsics are only promotable if they cover the whole alloca.
3013  // Therefore, we drop lifetime intrinsics which don't cover the whole
3014  // alloca.
3015  // (In theory, intrinsics which partially cover an alloca could be
3016  // promoted, but PromoteMemToReg doesn't handle that case.)
3017  // FIXME: Check whether the alloca is promotable before dropping the
3018  // lifetime intrinsics?
3019  if (NewBeginOffset != NewAllocaBeginOffset ||
3020  NewEndOffset != NewAllocaEndOffset)
3021  return true;
3022 
3023  ConstantInt *Size =
3024  ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
3025  NewEndOffset - NewBeginOffset);
3026  // Lifetime intrinsics always expect an i8* so directly get such a pointer
3027  // for the new alloca slice.
3028  Type *PointerTy = IRB.getInt8PtrTy(OldPtr->getType()->getPointerAddressSpace());
3029  Value *Ptr = getNewAllocaSlicePtr(IRB, PointerTy);
3030  Value *New;
3031  if (II.getIntrinsicID() == Intrinsic::lifetime_start)
3032  New = IRB.CreateLifetimeStart(Ptr, Size);
3033  else
3034  New = IRB.CreateLifetimeEnd(Ptr, Size);
3035 
3036  (void)New;
3037  LLVM_DEBUG(dbgs() << " to: " << *New << "\n");
3038 
3039  return true;
3040  }
3041 
3042  void fixLoadStoreAlign(Instruction &Root) {
3043  // This algorithm implements the same visitor loop as
3044  // hasUnsafePHIOrSelectUse, and fixes the alignment of each load
3045  // or store found.
3048  Visited.insert(&Root);
3049  Uses.push_back(&Root);
3050  do {
3051  Instruction *I = Uses.pop_back_val();
3052 
3053  if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
3054  unsigned LoadAlign = LI->getAlignment();
3055  if (!LoadAlign)
3056  LoadAlign = DL.getABITypeAlignment(LI->getType());
3057  LI->setAlignment(std::min(LoadAlign, getSliceAlign()));
3058  continue;
3059  }
3060  if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
3061  unsigned StoreAlign = SI->getAlignment();
3062  if (!StoreAlign) {
3063  Value *Op = SI->getOperand(0);
3064  StoreAlign = DL.getABITypeAlignment(Op->getType());
3065  }
3066  SI->setAlignment(std::min(StoreAlign, getSliceAlign()));
3067  continue;
3068  }
3069 
3070  assert(isa<BitCastInst>(I) || isa<PHINode>(I) ||
3071  isa<SelectInst>(I) || isa<GetElementPtrInst>(I));
3072  for (User *U : I->users())
3073  if (Visited.insert(cast<Instruction>(U)).second)
3074  Uses.push_back(cast<Instruction>(U));
3075  } while (!Uses.empty());
3076  }
3077 
3078  bool visitPHINode(PHINode &PN) {
3079  LLVM_DEBUG(dbgs() << " original: " << PN << "\n");
3080  assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
3081  assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
3082 
3083  // We would like to compute a new pointer in only one place, but have it be
3084  // as local as possible to the PHI. To do that, we re-use the location of
3085  // the old pointer, which necessarily must be in the right position to
3086  // dominate the PHI.
3087  IRBuilderTy PtrBuilder(IRB);
3088  if (isa<PHINode>(OldPtr))
3089  PtrBuilder.SetInsertPoint(&*OldPtr->getParent()->getFirstInsertionPt());
3090  else
3091  PtrBuilder.SetInsertPoint(OldPtr);
3092  PtrBuilder.SetCurrentDebugLocation(OldPtr->getDebugLoc());
3093 
3094  Value *NewPtr = getNewAllocaSlicePtr(PtrBuilder, OldPtr->getType());
3095  // Replace the operands which were using the old pointer.
3096  std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
3097 
3098  LLVM_DEBUG(dbgs() << " to: " << PN << "\n");
3099  deleteIfTriviallyDead(OldPtr);
3100 
3101  // Fix the alignment of any loads or stores using this PHI node.
3102  fixLoadStoreAlign(PN);
3103 
3104  // PHIs can't be promoted on their own, but often can be speculated. We
3105  // check the speculation outside of the rewriter so that we see the
3106  // fully-rewritten alloca.
3107  PHIUsers.insert(&PN);
3108  return true;
3109  }
3110 
3111  bool visitSelectInst(SelectInst &SI) {
3112  LLVM_DEBUG(dbgs() << " original: " << SI << "\n");
3113  assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
3114  "Pointer isn't an operand!");
3115  assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
3116  assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
3117 
3118  Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
3119  // Replace the operands which were using the old pointer.
3120  if (SI.getOperand(1) == OldPtr)
3121  SI.setOperand(1, NewPtr);
3122  if (SI.getOperand(2) == OldPtr)
3123  SI.setOperand(2, NewPtr);
3124 
3125  LLVM_DEBUG(dbgs() << " to: " << SI << "\n");
3126  deleteIfTriviallyDead(OldPtr);
3127 
3128  // Fix the alignment of any loads or stores using this select.
3129  fixLoadStoreAlign(SI);
3130 
3131  // Selects can't be promoted on their own, but often can be speculated. We
3132  // check the speculation outside of the rewriter so that we see the
3133  // fully-rewritten alloca.
3134  SelectUsers.insert(&SI);
3135  return true;
3136  }
3137 };
3138 
3139 namespace {
3140 
3141 /// Visitor to rewrite aggregate loads and stores as scalar.
3142 ///
3143 /// This pass aggressively rewrites all aggregate loads and stores on
3144 /// a particular pointer (or any pointer derived from it which we can identify)
3145 /// with scalar loads and stores.
3146 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
3147  // Befriend the base class so it can delegate to private visit methods.
3148  friend class InstVisitor<AggLoadStoreRewriter, bool>;
3149 
3150  /// Queue of pointer uses to analyze and potentially rewrite.
3151  SmallVector<Use *, 8> Queue;
3152 
3153  /// Set to prevent us from cycling with phi nodes and loops.
3154  SmallPtrSet<User *, 8> Visited;
3155 
3156  /// The current pointer use being rewritten. This is used to dig up the used
3157  /// value (as opposed to the user).
3158  Use *U;
3159 
3160  /// Used to calculate offsets, and hence alignment, of subobjects.
3161  const DataLayout &DL;
3162 
3163 public:
3164  AggLoadStoreRewriter(const DataLayout &DL) : DL(DL) {}
3165 
3166  /// Rewrite loads and stores through a pointer and all pointers derived from
3167  /// it.
3168  bool rewrite(Instruction &I) {
3169  LLVM_DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
3170  enqueueUsers(I);
3171  bool Changed = false;
3172  while (!Queue.empty()) {
3173  U = Queue.pop_back_val();
3174  Changed |= visit(cast<Instruction>(U->getUser()));
3175  }
3176  return Changed;
3177  }
3178 
3179 private:
3180  /// Enqueue all the users of the given instruction for further processing.
3181  /// This uses a set to de-duplicate users.
3182  void enqueueUsers(Instruction &I) {
3183  for (Use &U : I.uses())
3184  if (Visited.insert(U.getUser()).second)
3185  Queue.push_back(&U);
3186  }
3187 
3188  // Conservative default is to not rewrite anything.
3189  bool visitInstruction(Instruction &I) { return false; }
3190 
3191  /// Generic recursive split emission class.
3192  template <typename Derived> class OpSplitter {
3193  protected:
3194  /// The builder used to form new instructions.
3195  IRBuilderTy IRB;
3196 
3197  /// The indices which to be used with insert- or extractvalue to select the
3198  /// appropriate value within the aggregate.
3199  SmallVector<unsigned, 4> Indices;
3200 
3201  /// The indices to a GEP instruction which will move Ptr to the correct slot
3202  /// within the aggregate.
3203  SmallVector<Value *, 4> GEPIndices;
3204 
3205  /// The base pointer of the original op, used as a base for GEPing the
3206  /// split operations.
3207  Value *Ptr;
3208 
3209  /// The base pointee type being GEPed into.
3210  Type *BaseTy;
3211 
3212  /// Known alignment of the base pointer.
3213  unsigned BaseAlign;
3214 
3215  /// To calculate offset of each component so we can correctly deduce
3216  /// alignments.
3217  const DataLayout &DL;
3218 
3219  /// Initialize the splitter with an insertion point, Ptr and start with a
3220  /// single zero GEP index.
3221  OpSplitter(Instruction *InsertionPoint, Value *Ptr, Type *BaseTy,
3222  unsigned BaseAlign, const DataLayout &DL)
3223  : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr),
3224  BaseTy(BaseTy), BaseAlign(BaseAlign), DL(DL) {}
3225 
3226  public:
3227  /// Generic recursive split emission routine.
3228  ///
3229  /// This method recursively splits an aggregate op (load or store) into
3230  /// scalar or vector ops. It splits recursively until it hits a single value
3231  /// and emits that single value operation via the template argument.
3232  ///
3233  /// The logic of this routine relies on GEPs and insertvalue and
3234  /// extractvalue all operating with the same fundamental index list, merely
3235  /// formatted differently (GEPs need actual values).
3236  ///
3237  /// \param Ty The type being split recursively into smaller ops.
3238  /// \param Agg The aggregate value being built up or stored, depending on
3239  /// whether this is splitting a load or a store respectively.
3240  void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
3241  if (Ty->isSingleValueType()) {
3242  unsigned Offset = DL.getIndexedOffsetInType(BaseTy, GEPIndices);
3243  return static_cast<Derived *>(this)->emitFunc(
3244  Ty, Agg, MinAlign(BaseAlign, Offset), Name);
3245  }
3246 
3247  if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
3248  unsigned OldSize = Indices.size();
3249  (void)OldSize;
3250  for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
3251  ++Idx) {
3252  assert(Indices.size() == OldSize && "Did not return to the old size");
3253  Indices.push_back(Idx);
3254  GEPIndices.push_back(IRB.getInt32(Idx));
3255  emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
3256  GEPIndices.pop_back();
3257  Indices.pop_back();
3258  }
3259  return;
3260  }
3261 
3262  if (StructType *STy = dyn_cast<StructType>(Ty)) {
3263  unsigned OldSize = Indices.size();
3264  (void)OldSize;
3265  for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
3266  ++Idx) {
3267  assert(Indices.size() == OldSize && "Did not return to the old size");
3268  Indices.push_back(Idx);
3269  GEPIndices.push_back(IRB.getInt32(Idx));
3270  emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
3271  GEPIndices.pop_back();
3272  Indices.pop_back();
3273  }
3274  return;
3275  }
3276 
3277  llvm_unreachable("Only arrays and structs are aggregate loadable types");
3278  }
3279  };
3280 
3281  struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
3282  AAMDNodes AATags;
3283 
3284  LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr, Type *BaseTy,
3285  AAMDNodes AATags, unsigned BaseAlign, const DataLayout &DL)
3286  : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr, BaseTy, BaseAlign,
3287  DL), AATags(AATags) {}
3288 
3289  /// Emit a leaf load of a single value. This is called at the leaves of the
3290  /// recursive emission to actually load values.
3291  void emitFunc(Type *Ty, Value *&Agg, unsigned Align, const Twine &Name) {
3292  assert(Ty->isSingleValueType());
3293  // Load the single value and insert it using the indices.
3294  Value *GEP =
3295  IRB.CreateInBoundsGEP(nullptr, Ptr, GEPIndices, Name + ".gep");
3296  LoadInst *Load = IRB.CreateAlignedLoad(GEP, Align, Name + ".load");
3297  if (AATags)
3298  Load->setAAMetadata(AATags);
3299  Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
3300  LLVM_DEBUG(dbgs() << " to: " << *Load << "\n");
3301  }
3302  };
3303 
3304  bool visitLoadInst(LoadInst &LI) {
3305  assert(LI.getPointerOperand() == *U);
3306  if (!LI.isSimple() || LI.getType()->isSingleValueType())
3307  return false;
3308 
3309  // We have an aggregate being loaded, split it apart.
3310  LLVM_DEBUG(dbgs() << " original: " << LI << "\n");
3311  AAMDNodes AATags;
3312  LI.getAAMetadata(AATags);
3313  LoadOpSplitter Splitter(&LI, *U, LI.getType(), AATags,
3314  getAdjustedAlignment(&LI, 0, DL), DL);
3315  Value *V = UndefValue::get(LI.getType());
3316  Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
3317  LI.replaceAllUsesWith(V);
3318  LI.eraseFromParent();
3319  return true;
3320  }
3321 
3322  struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
3323  StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr, Type *BaseTy,
3324  AAMDNodes AATags, unsigned BaseAlign, const DataLayout &DL)
3325  : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr, BaseTy, BaseAlign,
3326  DL),
3327  AATags(AATags) {}
3328  AAMDNodes AATags;
3329  /// Emit a leaf store of a single value. This is called at the leaves of the
3330  /// recursive emission to actually produce stores.
3331  void emitFunc(Type *Ty, Value *&Agg, unsigned Align, const Twine &Name) {
3332  assert(Ty->isSingleValueType());
3333  // Extract the single value and store it using the indices.
3334  //
3335  // The gep and extractvalue values are factored out of the CreateStore
3336  // call to make the output independent of the argument evaluation order.
3337  Value *ExtractValue =
3338  IRB.CreateExtractValue(Agg, Indices, Name + ".extract");
3339  Value *InBoundsGEP =
3340  IRB.CreateInBoundsGEP(nullptr, Ptr, GEPIndices, Name + ".gep");
3341  StoreInst *Store =
3342  IRB.CreateAlignedStore(ExtractValue, InBoundsGEP, Align);
3343  if (AATags)
3344  Store->setAAMetadata(AATags);
3345  LLVM_DEBUG(dbgs() << " to: " << *Store << "\n");
3346  }
3347  };
3348 
3349  bool visitStoreInst(StoreInst &SI) {
3350  if (!SI.isSimple() || SI.getPointerOperand() != *U)
3351  return false;
3352  Value *V = SI.getValueOperand();
3353  if (V->getType()->isSingleValueType())
3354  return false;
3355 
3356  // We have an aggregate being stored, split it apart.
3357  LLVM_DEBUG(dbgs() << " original: " << SI << "\n");
3358  AAMDNodes AATags;
3359  SI.getAAMetadata(AATags);
3360  StoreOpSplitter Splitter(&SI, *U, V->getType(), AATags,
3361  getAdjustedAlignment(&SI, 0, DL), DL);
3362  Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
3363  SI.eraseFromParent();
3364  return true;
3365  }
3366 
3367  bool visitBitCastInst(BitCastInst &BC) {
3368  enqueueUsers(BC);
3369  return false;
3370  }
3371 
3372  bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
3373  enqueueUsers(GEPI);
3374  return false;
3375  }
3376 
3377  bool visitPHINode(PHINode &PN) {
3378  enqueueUsers(PN);
3379  return false;
3380  }
3381 
3382  bool visitSelectInst(SelectInst &SI) {
3383  enqueueUsers(SI);
3384  return false;
3385  }
3386 };
3387 
3388 } // end anonymous namespace
3389 
3390 /// Strip aggregate type wrapping.
3391 ///
3392 /// This removes no-op aggregate types wrapping an underlying type. It will
3393 /// strip as many layers of types as it can without changing either the type
3394 /// size or the allocated size.
3396  if (Ty->isSingleValueType())
3397  return Ty;
3398 
3399  uint64_t AllocSize = DL.getTypeAllocSize(Ty);
3400  uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
3401 
3402  Type *InnerTy;
3403  if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
3404  InnerTy = ArrTy->getElementType();
3405  } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
3406  const StructLayout *SL = DL.getStructLayout(STy);
3407  unsigned Index = SL->getElementContainingOffset(0);
3408  InnerTy = STy->getElementType(Index);
3409  } else {
3410  return Ty;
3411  }
3412 
3413  if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
3414  TypeSize > DL.getTypeSizeInBits(InnerTy))
3415  return Ty;
3416 
3417  return stripAggregateTypeWrapping(DL, InnerTy);
3418 }
3419 
3420 /// Try to find a partition of the aggregate type passed in for a given
3421 /// offset and size.
3422 ///
3423 /// This recurses through the aggregate type and tries to compute a subtype
3424 /// based on the offset and size. When the offset and size span a sub-section
3425 /// of an array, it will even compute a new array type for that sub-section,
3426 /// and the same for structs.
3427 ///
3428 /// Note that this routine is very strict and tries to find a partition of the
3429 /// type which produces the *exact* right offset and size. It is not forgiving
3430 /// when the size or offset cause either end of type-based partition to be off.
3431 /// Also, this is a best-effort routine. It is reasonable to give up and not
3432 /// return a type if necessary.
3433 static Type *getTypePartition(const DataLayout &DL, Type *Ty, uint64_t Offset,
3434  uint64_t Size) {
3435  if (Offset == 0 && DL.getTypeAllocSize(Ty) == Size)
3436  return stripAggregateTypeWrapping(DL, Ty);
3437  if (Offset > DL.getTypeAllocSize(Ty) ||
3438  (DL.getTypeAllocSize(Ty) - Offset) < Size)
3439  return nullptr;
3440 
3441  if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3442  Type *ElementTy = SeqTy->getElementType();
3443  uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3444  uint64_t NumSkippedElements = Offset / ElementSize;
3445  if (NumSkippedElements >= SeqTy->getNumElements())
3446  return nullptr;
3447  Offset -= NumSkippedElements * ElementSize;
3448 
3449  // First check if we need to recurse.
3450  if (Offset > 0 || Size < ElementSize) {
3451  // Bail if the partition ends in a different array element.
3452  if ((Offset + Size) > ElementSize)
3453  return nullptr;
3454  // Recurse through the element type trying to peel off offset bytes.
3455  return getTypePartition(DL, ElementTy, Offset, Size);
3456  }
3457  assert(Offset == 0);
3458 
3459  if (Size == ElementSize)
3460  return stripAggregateTypeWrapping(DL, ElementTy);
3461  assert(Size > ElementSize);
3462  uint64_t NumElements = Size / ElementSize;
3463  if (NumElements * ElementSize != Size)
3464  return nullptr;
3465  return ArrayType::get(ElementTy, NumElements);
3466  }
3467 
3468  StructType *STy = dyn_cast<StructType>(Ty);
3469  if (!STy)
3470  return nullptr;
3471 
3472  const StructLayout *SL = DL.getStructLayout(STy);
3473  if (Offset >= SL->getSizeInBytes())
3474  return nullptr;
3475  uint64_t EndOffset = Offset + Size;
3476  if (EndOffset > SL->getSizeInBytes())
3477  return nullptr;
3478 
3479  unsigned Index = SL->getElementContainingOffset(Offset);
3480  Offset -= SL->getElementOffset(Index);
3481 
3482  Type *ElementTy = STy->getElementType(Index);
3483  uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3484  if (Offset >= ElementSize)
3485  return nullptr; // The offset points into alignment padding.
3486 
3487  // See if any partition must be contained by the element.
3488  if (Offset > 0 || Size < ElementSize) {
3489  if ((Offset + Size) > ElementSize)
3490  return nullptr;
3491  return getTypePartition(DL, ElementTy, Offset, Size);
3492  }
3493  assert(Offset == 0);
3494 
3495  if (Size == ElementSize)
3496  return stripAggregateTypeWrapping(DL, ElementTy);
3497 
3499  EE = STy->element_end();
3500  if (EndOffset < SL->getSizeInBytes()) {
3501  unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3502  if (Index == EndIndex)
3503  return nullptr; // Within a single element and its padding.
3504 
3505  // Don't try to form "natural" types if the elements don't line up with the
3506  // expected size.
3507  // FIXME: We could potentially recurse down through the last element in the
3508  // sub-struct to find a natural end point.
3509  if (SL->getElementOffset(EndIndex) != EndOffset)
3510  return nullptr;
3511 
3512  assert(Index < EndIndex);
3513  EE = STy->element_begin() + EndIndex;
3514  }
3515 
3516  // Try to build up a sub-structure.
3517  StructType *SubTy =
3518  StructType::get(STy->getContext(), makeArrayRef(EI, EE), STy->isPacked());
3519  const StructLayout *SubSL = DL.getStructLayout(SubTy);
3520  if (Size != SubSL->getSizeInBytes())
3521  return nullptr; // The sub-struct doesn't have quite the size needed.
3522 
3523  return SubTy;
3524 }
3525 
3526 /// Pre-split loads and stores to simplify rewriting.
3527 ///
3528 /// We want to break up the splittable load+store pairs as much as
3529 /// possible. This is important to do as a preprocessing step, as once we
3530 /// start rewriting the accesses to partitions of the alloca we lose the
3531 /// necessary information to correctly split apart paired loads and stores
3532 /// which both point into this alloca. The case to consider is something like
3533 /// the following:
3534 ///
3535 /// %a = alloca [12 x i8]
3536 /// %gep1 = getelementptr [12 x i8]* %a, i32 0, i32 0
3537 /// %gep2 = getelementptr [12 x i8]* %a, i32 0, i32 4
3538 /// %gep3 = getelementptr [12 x i8]* %a, i32 0, i32 8
3539 /// %iptr1 = bitcast i8* %gep1 to i64*
3540 /// %iptr2 = bitcast i8* %gep2 to i64*
3541 /// %fptr1 = bitcast i8* %gep1 to float*
3542 /// %fptr2 = bitcast i8* %gep2 to float*
3543 /// %fptr3 = bitcast i8* %gep3 to float*
3544 /// store float 0.0, float* %fptr1
3545 /// store float 1.0, float* %fptr2
3546 /// %v = load i64* %iptr1
3547 /// store i64 %v, i64* %iptr2
3548 /// %f1 = load float* %fptr2
3549 /// %f2 = load float* %fptr3
3550 ///
3551 /// Here we want to form 3 partitions of the alloca, each 4 bytes large, and
3552 /// promote everything so we recover the 2 SSA values that should have been
3553 /// there all along.
3554 ///
3555 /// \returns true if any changes are made.
3556 bool SROA::presplitLoadsAndStores(AllocaInst &AI, AllocaSlices &AS) {
3557  LLVM_DEBUG(dbgs() << "Pre-splitting loads and stores\n");
3558 
3559  // Track the loads and stores which are candidates for pre-splitting here, in
3560  // the order they first appear during the partition scan. These give stable
3561  // iteration order and a basis for tracking which loads and stores we
3562  // actually split.
3565 
3566  // We need to accumulate the splits required of each load or store where we
3567  // can find them via a direct lookup. This is important to cross-check loads
3568  // and stores against each other. We also track the slice so that we can kill
3569  // all the slices that end up split.
3570  struct SplitOffsets {
3571  Slice *S;
3572  std::vector<uint64_t> Splits;
3573  };
3575 
3576  // Track loads out of this alloca which cannot, for any reason, be pre-split.
3577  // This is important as we also cannot pre-split stores of those loads!
3578  // FIXME: This is all pretty gross. It means that we can be more aggressive
3579  // in pre-splitting when the load feeding the store happens to come from
3580  // a separate alloca. Put another way, the effectiveness of SROA would be
3581  // decreased by a frontend which just concatenated all of its local allocas
3582  // into one big flat alloca. But defeating such patterns is exactly the job
3583  // SROA is tasked with! Sadly, to not have this discrepancy we would have
3584  // change store pre-splitting to actually force pre-splitting of the load
3585  // that feeds it *and all stores*. That makes pre-splitting much harder, but
3586  // maybe it would make it more principled?
3587  SmallPtrSet<LoadInst *, 8> UnsplittableLoads;
3588 
3589  LLVM_DEBUG(dbgs() << " Searching for candidate loads and stores\n");
3590  for (auto &P : AS.partitions()) {
3591  for (Slice &S : P) {
3592  Instruction *I = cast<Instruction>(S.getUse()->getUser());
3593  if (!S.isSplittable() || S.endOffset() <= P.endOffset()) {
3594  // If this is a load we have to track that it can't participate in any
3595  // pre-splitting. If this is a store of a load we have to track that
3596  // that load also can't participate in any pre-splitting.
3597  if (auto *LI = dyn_cast<LoadInst>(I))
3598  UnsplittableLoads.insert(LI);
3599  else if (auto *SI = dyn_cast<StoreInst>(I))
3600  if (auto *LI = dyn_cast<LoadInst>(SI->getValueOperand()))
3601  UnsplittableLoads.insert(LI);
3602  continue;
3603  }
3604  assert(P.endOffset() > S.beginOffset() &&
3605  "Empty or backwards partition!");
3606 
3607  // Determine if this is a pre-splittable slice.
3608  if (auto *LI = dyn_cast<LoadInst>(I)) {
3609  assert(!LI->isVolatile() && "Cannot split volatile loads!");
3610 
3611  // The load must be used exclusively to store into other pointers for
3612  // us to be able to arbitrarily pre-split it. The stores must also be
3613  // simple to avoid changing semantics.
3614  auto IsLoadSimplyStored = [](LoadInst *LI) {
3615  for (User *LU : LI->users()) {
3616  auto *SI = dyn_cast<StoreInst>(LU);
3617  if (!SI || !SI->isSimple())
3618  return false;
3619  }
3620  return true;
3621  };
3622  if (!IsLoadSimplyStored(LI)) {
3623  UnsplittableLoads.insert(LI);
3624  continue;
3625  }
3626 
3627  Loads.push_back(LI);
3628  } else if (auto *SI = dyn_cast<StoreInst>(I)) {
3629  if (S.getUse() != &SI->getOperandUse(SI->getPointerOperandIndex()))
3630  // Skip stores *of* pointers. FIXME: This shouldn't even be possible!
3631  continue;
3632  auto *StoredLoad = dyn_cast<LoadInst>(SI->getValueOperand());
3633  if (!StoredLoad || !StoredLoad->isSimple())
3634  continue;
3635  assert(!SI->isVolatile() && "Cannot split volatile stores!");
3636 
3637  Stores.push_back(SI);
3638  } else {
3639  // Other uses cannot be pre-split.
3640  continue;
3641  }
3642 
3643  // Record the initial split.
3644  LLVM_DEBUG(dbgs() << " Candidate: " << *I << "\n");
3645  auto &Offsets = SplitOffsetsMap[I];
3646  assert(Offsets.Splits.empty() &&
3647  "Should not have splits the first time we see an instruction!");
3648  Offsets.S = &S;
3649  Offsets.Splits.push_back(P.endOffset() - S.beginOffset());
3650  }
3651 
3652  // Now scan the already split slices, and add a split for any of them which
3653  // we're going to pre-split.
3654  for (Slice *S : P.splitSliceTails()) {
3655  auto SplitOffsetsMapI =
3656  SplitOffsetsMap.find(cast<Instruction>(S->getUse()->getUser()));
3657  if (SplitOffsetsMapI == SplitOffsetsMap.end())
3658  continue;
3659  auto &Offsets = SplitOffsetsMapI->second;
3660 
3661  assert(Offsets.S == S && "Found a mismatched slice!");
3662  assert(!Offsets.Splits.empty() &&
3663  "Cannot have an empty set of splits on the second partition!");
3664  assert(Offsets.Splits.back() ==
3665  P.beginOffset() - Offsets.S->beginOffset() &&
3666  "Previous split does not end where this one begins!");
3667 
3668  // Record each split. The last partition's end isn't needed as the size
3669  // of the slice dictates that.
3670  if (S->endOffset() > P.endOffset())
3671  Offsets.Splits.push_back(P.endOffset() - Offsets.S->beginOffset());
3672  }
3673  }
3674 
3675  // We may have split loads where some of their stores are split stores. For
3676  // such loads and stores, we can only pre-split them if their splits exactly
3677  // match relative to their starting offset. We have to verify this prior to
3678  // any rewriting.
3679  Stores.erase(
3680  llvm::remove_if(Stores,
3681  [&UnsplittableLoads, &SplitOffsetsMap](StoreInst *SI) {
3682  // Lookup the load we are storing in our map of split
3683  // offsets.
3684  auto *LI = cast<LoadInst>(SI->getValueOperand());
3685  // If it was completely unsplittable, then we're done,
3686  // and this store can't be pre-split.
3687  if (UnsplittableLoads.count(LI))
3688  return true;
3689 
3690  auto LoadOffsetsI = SplitOffsetsMap.find(LI);
3691  if (LoadOffsetsI == SplitOffsetsMap.end())
3692  return false; // Unrelated loads are definitely safe.
3693  auto &LoadOffsets = LoadOffsetsI->second;
3694 
3695  // Now lookup the store's offsets.
3696  auto &StoreOffsets = SplitOffsetsMap[SI];
3697 
3698  // If the relative offsets of each split in the load and
3699  // store match exactly, then we can split them and we
3700  // don't need to remove them here.
3701  if (LoadOffsets.Splits == StoreOffsets.Splits)
3702  return false;
3703 
3704  LLVM_DEBUG(
3705  dbgs()
3706  << " Mismatched splits for load and store:\n"
3707  << " " << *LI << "\n"
3708  << " " << *SI << "\n");
3709 
3710  // We've found a store and load that we need to split
3711  // with mismatched relative splits. Just give up on them
3712  // and remove both instructions from our list of
3713  // candidates.
3714  UnsplittableLoads.insert(LI);
3715  return true;
3716  }),
3717  Stores.end());
3718  // Now we have to go *back* through all the stores, because a later store may
3719  // have caused an earlier store's load to become unsplittable and if it is
3720  // unsplittable for the later store, then we can't rely on it being split in
3721  // the earlier store either.
3722  Stores.erase(llvm::remove_if(Stores,
3723  [&UnsplittableLoads](StoreInst *SI) {
3724  auto *LI =
3725  cast<LoadInst>(SI->getValueOperand());
3726  return UnsplittableLoads.count(LI);
3727  }),
3728  Stores.end());
3729  // Once we've established all the loads that can't be split for some reason,
3730  // filter any that made it into our list out.
3731  Loads.erase(llvm::remove_if(Loads,
3732  [&UnsplittableLoads](LoadInst *LI) {
3733  return UnsplittableLoads.count(LI);
3734  }),
3735  Loads.end());
3736 
3737  // If no loads or stores are left, there is no pre-splitting to be done for
3738  // this alloca.
3739  if (Loads.empty() && Stores.empty())
3740  return false;
3741 
3742  // From here on, we can't fail and will be building new accesses, so rig up
3743  // an IR builder.
3744  IRBuilderTy IRB(&AI);
3745 
3746  // Collect the new slices which we will merge into the alloca slices.
3747  SmallVector<Slice, 4> NewSlices;
3748 
3749  // Track any allocas we end up splitting loads and stores for so we iterate
3750  // on them.
3751  SmallPtrSet<AllocaInst *, 4> ResplitPromotableAllocas;
3752 
3753  // At this point, we have collected all of the loads and stores we can
3754  // pre-split, and the specific splits needed for them. We actually do the
3755  // splitting in a specific order in order to handle when one of the loads in
3756  // the value operand to one of the stores.
3757  //
3758  // First, we rewrite all of the split loads, and just accumulate each split
3759  // load in a parallel structure. We also build the slices for them and append
3760  // them to the alloca slices.
3762  std::vector<LoadInst *> SplitLoads;
3763  const DataLayout &DL = AI.getModule()->getDataLayout();
3764  for (LoadInst *LI : Loads) {
3765  SplitLoads.clear();
3766 
3767  IntegerType *Ty = cast<IntegerType>(LI->getType());
3768  uint64_t LoadSize = Ty->getBitWidth() / 8;
3769  assert(LoadSize > 0 && "Cannot have a zero-sized integer load!");
3770 
3771  auto &Offsets = SplitOffsetsMap[LI];
3772  assert(LoadSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&
3773  "Slice size should always match load size exactly!");
3774  uint64_t BaseOffset = Offsets.S->beginOffset();
3775  assert(BaseOffset + LoadSize > BaseOffset &&
3776  "Cannot represent alloca access size using 64-bit integers!");
3777 
3778  Instruction *BasePtr = cast<Instruction>(LI->getPointerOperand());
3779  IRB.SetInsertPoint(LI);
3780 
3781  LLVM_DEBUG(dbgs() << " Splitting load: " << *LI << "\n");
3782 
3783  uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
3784  int Idx = 0, Size = Offsets.Splits.size();
3785  for (;;) {
3786  auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
3787  auto AS = LI->getPointerAddressSpace();
3788  auto *PartPtrTy = PartTy->getPointerTo(AS);
3789  LoadInst *PLoad = IRB.CreateAlignedLoad(
3790  getAdjustedPtr(IRB, DL, BasePtr,
3791  APInt(DL.getIndexSizeInBits(AS), PartOffset),
3792  PartPtrTy, BasePtr->getName() + "."),
3793  getAdjustedAlignment(LI, PartOffset, DL), /*IsVolatile*/ false,
3794  LI->getName());
3797 
3798  // Append this load onto the list of split loads so we can find it later
3799  // to rewrite the stores.
3800  SplitLoads.push_back(PLoad);
3801 
3802  // Now build a new slice for the alloca.
3803  NewSlices.push_back(
3804  Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
3805  &PLoad->getOperandUse(PLoad->getPointerOperandIndex()),
3806  /*IsSplittable*/ false));
3807  LLVM_DEBUG(dbgs() << " new slice [" << NewSlices.back().beginOffset()
3808  << ", " << NewSlices.back().endOffset()
3809  << "): " << *PLoad << "\n");
3810 
3811  // See if we've handled all the splits.
3812  if (Idx >= Size)
3813  break;
3814 
3815  // Setup the next partition.
3816  PartOffset = Offsets.Splits[Idx];
3817  ++Idx;
3818  PartSize = (Idx < Size ? Offsets.Splits[Idx] : LoadSize) - PartOffset;
3819  }
3820 
3821  // Now that we have the split loads, do the slow walk over all uses of the
3822  // load and rewrite them as split stores, or save the split loads to use
3823  // below if the store is going to be split there anyways.
3824  bool DeferredStores = false;
3825  for (User *LU : LI->users()) {
3826  StoreInst *SI = cast<StoreInst>(LU);
3827  if (!Stores.empty() && SplitOffsetsMap.count(SI)) {
3828  DeferredStores = true;
3829  LLVM_DEBUG(dbgs() << " Deferred splitting of store: " << *SI
3830  << "\n");
3831  continue;
3832  }
3833 
3834  Value *StoreBasePtr = SI->getPointerOperand();
3835  IRB.SetInsertPoint(SI);
3836 
3837  LLVM_DEBUG(dbgs() << " Splitting store of load: " << *SI << "\n");
3838 
3839  for (int Idx = 0, Size = SplitLoads.size(); Idx < Size; ++Idx) {
3840  LoadInst *PLoad = SplitLoads[Idx];
3841  uint64_t PartOffset = Idx == 0 ? 0 : Offsets.Splits[Idx - 1];
3842  auto *PartPtrTy =
3843  PLoad->getType()->getPointerTo(SI->getPointerAddressSpace());
3844 
3845  auto AS = SI->getPointerAddressSpace();
3846  StoreInst *PStore = IRB.CreateAlignedStore(
3847  PLoad,
3848  getAdjustedPtr(IRB, DL, StoreBasePtr,
3849  APInt(DL.getIndexSizeInBits(AS), PartOffset),
3850  PartPtrTy, StoreBasePtr->getName() + "."),
3851  getAdjustedAlignment(SI, PartOffset, DL), /*IsVolatile*/ false);
3854  LLVM_DEBUG(dbgs() << " +" << PartOffset << ":" << *PStore << "\n");
3855  }
3856 
3857  // We want to immediately iterate on any allocas impacted by splitting
3858  // this store, and we have to track any promotable alloca (indicated by
3859  // a direct store) as needing to be resplit because it is no longer
3860  // promotable.
3861  if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(StoreBasePtr)) {
3862  ResplitPromotableAllocas.insert(OtherAI);
3863  Worklist.insert(OtherAI);
3864  } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
3865  StoreBasePtr->stripInBoundsOffsets())) {
3866  Worklist.insert(OtherAI);
3867  }
3868 
3869  // Mark the original store as dead.
3870  DeadInsts.insert(SI);
3871  }
3872 
3873  // Save the split loads if there are deferred stores among the users.
3874  if (DeferredStores)
3875  SplitLoadsMap.insert(std::make_pair(LI, std::move(SplitLoads)));
3876 
3877  // Mark the original load as dead and kill the original slice.
3878  DeadInsts.insert(LI);
3879  Offsets.S->kill();
3880  }
3881 
3882  // Second, we rewrite all of the split stores. At this point, we know that
3883  // all loads from this alloca have been split already. For stores of such
3884  // loads, we can simply look up the pre-existing split loads. For stores of
3885  // other loads, we split those loads first and then write split stores of
3886  // them.
3887  for (StoreInst *SI : Stores) {
3888  auto *LI = cast<LoadInst>(SI->getValueOperand());
3889  IntegerType *Ty = cast<IntegerType>(LI->getType());
3890  uint64_t StoreSize = Ty->getBitWidth() / 8;
3891  assert(StoreSize > 0 && "Cannot have a zero-sized integer store!");
3892 
3893  auto &Offsets = SplitOffsetsMap[SI];
3894  assert(StoreSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&
3895  "Slice size should always match load size exactly!");
3896  uint64_t BaseOffset = Offsets.S->beginOffset();
3897  assert(BaseOffset + StoreSize > BaseOffset &&
3898  "Cannot represent alloca access size using 64-bit integers!");
3899 
3900  Value *LoadBasePtr = LI->getPointerOperand();
3901  Instruction *StoreBasePtr = cast<Instruction>(SI->getPointerOperand());
3902 
3903  LLVM_DEBUG(dbgs() << " Splitting store: " << *SI << "\n");
3904 
3905  // Check whether we have an already split load.
3906  auto SplitLoadsMapI = SplitLoadsMap.find(LI);
3907  std::vector<LoadInst *> *SplitLoads = nullptr;
3908  if (SplitLoadsMapI != SplitLoadsMap.end()) {
3909  SplitLoads = &SplitLoadsMapI->second;
3910  assert(SplitLoads->size() == Offsets.Splits.size() + 1 &&
3911  "Too few split loads for the number of splits in the store!");
3912  } else {
3913  LLVM_DEBUG(dbgs() << " of load: " << *LI << "\n");
3914  }
3915 
3916  uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
3917  int Idx = 0, Size = Offsets.Splits.size();
3918  for (;;) {
3919  auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
3920  auto *LoadPartPtrTy = PartTy->getPointerTo(LI->getPointerAddressSpace());
3921  auto *StorePartPtrTy = PartTy->getPointerTo(SI->getPointerAddressSpace());
3922 
3923  // Either lookup a split load or create one.
3924  LoadInst *PLoad;
3925  if (SplitLoads) {
3926  PLoad = (*SplitLoads)[Idx];
3927  } else {
3928  IRB.SetInsertPoint(LI);
3929  auto AS = LI->getPointerAddressSpace();
3930  PLoad = IRB.CreateAlignedLoad(
3931  getAdjustedPtr(IRB, DL, LoadBasePtr,
3932  APInt(DL.getIndexSizeInBits(AS), PartOffset),
3933  LoadPartPtrTy, LoadBasePtr->getName() + "."),
3934  getAdjustedAlignment(LI, PartOffset, DL), /*IsVolatile*/ false,
3935  LI->getName());
3936  }
3937 
3938  // And store this partition.
3939  IRB.SetInsertPoint(SI);
3940  auto AS = SI->getPointerAddressSpace();
3941  StoreInst *PStore = IRB.CreateAlignedStore(
3942  PLoad,
3943  getAdjustedPtr(IRB, DL, StoreBasePtr,
3944  APInt(DL.getIndexSizeInBits(AS), PartOffset),
3945  StorePartPtrTy, StoreBasePtr->getName() + "."),
3946  getAdjustedAlignment(SI, PartOffset, DL), /*IsVolatile*/ false);
3947 
3948  // Now build a new slice for the alloca.
3949  NewSlices.push_back(
3950  Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
3951  &PStore->getOperandUse(PStore->getPointerOperandIndex()),
3952  /*IsSplittable*/ false));
3953  LLVM_DEBUG(dbgs() << " new slice [" << NewSlices.back().beginOffset()
3954  << ", " << NewSlices.back().endOffset()
3955  << "): " << *PStore << "\n");
3956  if (!SplitLoads) {
3957  LLVM_DEBUG(dbgs() << " of split load: " << *PLoad << "\n");
3958  }
3959 
3960  // See if we've finished all the splits.
3961  if (Idx >= Size)
3962  break;
3963 
3964  // Setup the next partition.
3965  PartOffset = Offsets.Splits[Idx];
3966  ++Idx;
3967  PartSize = (Idx < Size ? Offsets.Splits[Idx] : StoreSize) - PartOffset;
3968  }
3969 
3970  // We want to immediately iterate on any allocas impacted by splitting
3971  // this load, which is only relevant if it isn't a load of this alloca and
3972  // thus we didn't already split the loads above. We also have to keep track
3973  // of any promotable allocas we split loads on as they can no longer be
3974  // promoted.
3975  if (!SplitLoads) {
3976  if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(LoadBasePtr)) {
3977  assert(OtherAI != &AI && "We can't re-split our own alloca!");
3978  ResplitPromotableAllocas.insert(OtherAI);
3979  Worklist.insert(OtherAI);
3980  } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
3981  LoadBasePtr->stripInBoundsOffsets())) {
3982  assert(OtherAI != &AI && "We can't re-split our own alloca!");
3983  Worklist.insert(OtherAI);
3984  }
3985  }
3986 
3987  // Mark the original store as dead now that we've split it up and kill its
3988  // slice. Note that we leave the original load in place unless this store
3989  // was its only use. It may in turn be split up if it is an alloca load
3990  // for some other alloca, but it may be a normal load. This may introduce
3991  // redundant loads, but where those can be merged the rest of the optimizer
3992  // should handle the merging, and this uncovers SSA splits which is more
3993  // important. In practice, the original loads will almost always be fully
3994  // split and removed eventually, and the splits will be merged by any
3995  // trivial CSE, including instcombine.
3996  if (LI->hasOneUse()) {
3997  assert(*LI->user_begin() == SI && "Single use isn't this store!");
3998  DeadInsts.insert(LI);
3999  }
4000  DeadInsts.insert(SI);
4001  Offsets.S->kill();
4002  }
4003 
4004  // Remove the killed slices that have ben pre-split.
4005  AS.erase(llvm::remove_if(AS, [](const Slice &S) { return S.isDead(); }),
4006  AS.end());
4007 
4008  // Insert our new slices. This will sort and merge them into the sorted
4009  // sequence.
4010  AS.insert(NewSlices);
4011 
4012  LLVM_DEBUG(dbgs() << " Pre-split slices:\n");
4013 #ifndef NDEBUG
4014  for (auto I = AS.begin(), E = AS.end(); I != E; ++I)
4015  LLVM_DEBUG(AS.print(dbgs(), I, " "));
4016 #endif
4017 
4018  // Finally, don't try to promote any allocas that new require re-splitting.
4019  // They have already been added to the worklist above.
4020  PromotableAllocas.erase(
4022  PromotableAllocas,
4023  [&](AllocaInst *AI) { return ResplitPromotableAllocas.count(AI); }),
4024  PromotableAllocas.end());
4025 
4026  return true;
4027 }
4028 
4029 /// Rewrite an alloca partition's users.
4030 ///
4031 /// This routine drives both of the rewriting goals of the SROA pass. It tries
4032 /// to rewrite uses of an alloca partition to be conducive for SSA value
4033 /// promotion. If the partition needs a new, more refined alloca, this will
4034 /// build that new alloca, preserving as much type information as possible, and
4035 /// rewrite the uses of the old alloca to point at the new one and have the
4036 /// appropriate new offsets. It also evaluates how successful the rewrite was
4037 /// at enabling promotion and if it was successful queues the alloca to be
4038 /// promoted.
4039 AllocaInst *SROA::rewritePartition(AllocaInst &AI, AllocaSlices &AS,
4040  Partition &P) {
4041  // Try to compute a friendly type for this partition of the alloca. This
4042  // won't always succeed, in which case we fall back to a legal integer type
4043  // or an i8 array of an appropriate size.
4044  Type *SliceTy = nullptr;
4045  const DataLayout &DL = AI.getModule()->getDataLayout();
4046  if (Type *CommonUseTy = findCommonType(P.begin(), P.end(), P.endOffset()))
4047  if (DL.getTypeAllocSize(CommonUseTy) >= P.size())
4048  SliceTy = CommonUseTy;
4049  if (!SliceTy)
4050  if (Type *TypePartitionTy = getTypePartition(DL, AI.getAllocatedType(),
4051  P.beginOffset(), P.size()))
4052  SliceTy = TypePartitionTy;
4053  if ((!SliceTy || (SliceTy->isArrayTy() &&
4054  SliceTy->getArrayElementType()->isIntegerTy())) &&
4055  DL.isLegalInteger(P.size() * 8))
4056  SliceTy = Type::getIntNTy(*C, P.size() * 8);
4057  if (!SliceTy)
4058  SliceTy = ArrayType::get(Type::getInt8Ty(*C), P.size());
4059  assert(DL.getTypeAllocSize(SliceTy) >= P.size());
4060 
4061  bool IsIntegerPromotable = isIntegerWideningViable(P, SliceTy, DL);
4062 
4063  VectorType *VecTy =
4064  IsIntegerPromotable ? nullptr : isVectorPromotionViable(P, DL);
4065  if (VecTy)
4066  SliceTy = VecTy;
4067 
4068  // Check for the case where we're going to rewrite to a new alloca of the
4069  // exact same type as the original, and with the same access offsets. In that
4070  // case, re-use the existing alloca, but still run through the rewriter to
4071  // perform phi and select speculation.
4072  // P.beginOffset() can be non-zero even with the same type in a case with
4073  // out-of-bounds access (e.g. @PR35657 function in SROA/basictest.ll).
4074  AllocaInst *NewAI;
4075  if (SliceTy == AI.getAllocatedType() && P.beginOffset() == 0) {
4076  NewAI = &AI;
4077  // FIXME: We should be able to bail at this point with "nothing changed".
4078  // FIXME: We might want to defer PHI speculation until after here.
4079  // FIXME: return nullptr;
4080  } else {
4081  unsigned Alignment = AI.getAlignment();
4082  if (!Alignment) {
4083  // The minimum alignment which users can rely on when the explicit
4084  // alignment is omitted or zero is that required by the ABI for this
4085  // type.
4086  Alignment = DL.getABITypeAlignment(AI.getAllocatedType());
4087  }
4088  Alignment = MinAlign(Alignment, P.beginOffset());
4089  // If we will get at least this much alignment from the type alone, leave
4090  // the alloca's alignment unconstrained.
4091  if (Alignment <= DL.getABITypeAlignment(SliceTy))
4092  Alignment = 0;
4093  NewAI = new AllocaInst(
4094  SliceTy, AI.getType()->getAddressSpace(), nullptr, Alignment,
4095  AI.getName() + ".sroa." + Twine(P.begin() - AS.begin()), &AI);
4096  // Copy the old AI debug location over to the new one.
4097  NewAI->setDebugLoc(AI.getDebugLoc());
4098  ++NumNewAllocas;
4099  }
4100 
4101  LLVM_DEBUG(dbgs() << "Rewriting alloca partition "
4102  << "[" << P.beginOffset() << "," << P.endOffset()
4103  << ") to: " << *NewAI << "\n");
4104 
4105  // Track the high watermark on the worklist as it is only relevant for
4106  // promoted allocas. We will reset it to this point if the alloca is not in
4107  // fact scheduled for promotion.
4108  unsigned PPWOldSize = PostPromotionWorklist.size();
4109  unsigned NumUses = 0;
4111  SmallSetVector<SelectInst *, 8> SelectUsers;
4112 
4113  AllocaSliceRewriter Rewriter(DL, AS, *this, AI, *NewAI, P.beginOffset(),
4114  P.endOffset(), IsIntegerPromotable, VecTy,
4115  PHIUsers, SelectUsers);
4116  bool Promotable = true;
4117  for (Slice *S : P.splitSliceTails()) {
4118  Promotable &= Rewriter.visit(S);
4119  ++NumUses;
4120  }
4121  for (Slice &S : P) {
4122  Promotable &= Rewriter.visit(&S);
4123  ++NumUses;
4124  }
4125 
4126  NumAllocaPartitionUses += NumUses;
4127  MaxUsesPerAllocaPartition.updateMax(NumUses);
4128 
4129  // Now that we've processed all the slices in the new partition, check if any
4130  // PHIs or Selects would block promotion.
4131  for (PHINode *PHI : PHIUsers)
4132  if (!isSafePHIToSpeculate(*PHI)) {
4133  Promotable = false;
4134  PHIUsers.clear();
4135  SelectUsers.clear();
4136  break;
4137  }
4138 
4139  for (SelectInst *Sel : SelectUsers)
4140  if (!isSafeSelectToSpeculate(*Sel)) {
4141  Promotable = false;
4142  PHIUsers.clear();
4143  SelectUsers.clear();
4144  break;
4145  }
4146 
4147  if (Promotable) {
4148  if (PHIUsers.empty() && SelectUsers.empty()) {
4149  // Promote the alloca.
4150  PromotableAllocas.push_back(NewAI);
4151  } else {
4152  // If we have either PHIs or Selects to speculate, add them to those
4153  // worklists and re-queue the new alloca so that we promote in on the
4154  // next iteration.
4155  for (PHINode *PHIUser : PHIUsers)
4156  SpeculatablePHIs.insert(PHIUser);
4157  for (SelectInst *SelectUser : SelectUsers)
4158  SpeculatableSelects.insert(SelectUser);
4159  Worklist.insert(NewAI);
4160  }
4161  } else {
4162  // Drop any post-promotion work items if promotion didn't happen.
4163  while (PostPromotionWorklist.size() > PPWOldSize)
4164  PostPromotionWorklist.pop_back();
4165 
4166  // We couldn't promote and we didn't create a new partition, nothing
4167  // happened.
4168  if (NewAI == &AI)
4169  return nullptr;
4170 
4171  // If we can't promote the alloca, iterate on it to check for new
4172  // refinements exposed by splitting the current alloca. Don't iterate on an
4173  // alloca which didn't actually change and didn't get promoted.
4174  Worklist.insert(NewAI);
4175  }
4176 
4177  return NewAI;
4178 }
4179 
4180 /// Walks the slices of an alloca and form partitions based on them,
4181 /// rewriting each of their uses.
4182 bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &AS) {
4183  if (AS.begin() == AS.end())
4184  return false;
4185 
4186  unsigned NumPartitions = 0;
4187  bool Changed = false;
4188  const DataLayout &DL = AI.getModule()->getDataLayout();
4189 
4190  // First try to pre-split loads and stores.
4191  Changed |= presplitLoadsAndStores(AI, AS);
4192 
4193  // Now that we have identified any pre-splitting opportunities,
4194  // mark loads and stores unsplittable except for the following case.
4195  // We leave a slice splittable if all other slices are disjoint or fully
4196  // included in the slice, such as whole-alloca loads and stores.
4197  // If we fail to split these during pre-splitting, we want to force them
4198  // to be rewritten into a partition.
4199  bool IsSorted = true;
4200 
4201  uint64_t AllocaSize = DL.getTypeAllocSize(AI.getAllocatedType());
4202  const uint64_t MaxBitVectorSize = 1024;
4203  if (AllocaSize <= MaxBitVectorSize) {
4204  // If a byte boundary is included in any load or store, a slice starting or
4205  // ending at the boundary is not splittable.
4206  SmallBitVector SplittableOffset(AllocaSize + 1, true);
4207  for (Slice &S : AS)
4208  for (unsigned O = S.beginOffset() + 1;
4209  O < S.endOffset() && O < AllocaSize; O++)
4210  SplittableOffset.reset(O);
4211 
4212  for (Slice &S : AS) {
4213  if (!S.isSplittable())
4214  continue;
4215 
4216  if ((S.beginOffset() > AllocaSize || SplittableOffset[S.beginOffset()]) &&
4217  (S.endOffset() > AllocaSize || SplittableOffset[S.endOffset()]))
4218  continue;
4219 
4220  if (isa<LoadInst>(S.getUse()->getUser()) ||
4221  isa<StoreInst>(S.getUse()->getUser())) {
4222  S.makeUnsplittable();
4223  IsSorted = false;
4224  }
4225  }
4226  }
4227  else {
4228  // We only allow whole-alloca splittable loads and stores
4229  // for a large alloca to avoid creating too large BitVector.
4230  for (Slice &S : AS) {
4231  if (!S.isSplittable())
4232  continue;
4233 
4234  if (S.beginOffset() == 0 && S.endOffset() >= AllocaSize)
4235  continue;
4236 
4237  if (isa<LoadInst>(S.getUse()->getUser()) ||
4238  isa<StoreInst>(S.getUse()->getUser())) {
4239  S.makeUnsplittable();
4240  IsSorted = false;
4241  }
4242  }
4243  }
4244 
4245  if (!IsSorted)
4246  llvm::sort(AS);
4247 
4248  /// Describes the allocas introduced by rewritePartition in order to migrate
4249  /// the debug info.
4250  struct Fragment {
4251  AllocaInst *Alloca;
4252  uint64_t Offset;
4253  uint64_t Size;
4254  Fragment(AllocaInst *AI, uint64_t O, uint64_t S)
4255  : Alloca(AI), Offset(O), Size(S) {}
4256  };
4257  SmallVector<Fragment, 4> Fragments;
4258 
4259  // Rewrite each partition.
4260  for (auto &P : AS.partitions()) {
4261  if (AllocaInst *NewAI = rewritePartition(AI, AS, P)) {
4262  Changed = true;
4263  if (NewAI != &AI) {
4264  uint64_t SizeOfByte = 8;
4265  uint64_t AllocaSize = DL.getTypeSizeInBits(NewAI->getAllocatedType());
4266  // Don't include any padding.
4267  uint64_t Size = std::min(AllocaSize, P.size() * SizeOfByte);
4268  Fragments.push_back(Fragment(NewAI, P.beginOffset() * SizeOfByte, Size));
4269  }
4270  }
4271  ++NumPartitions;
4272  }
4273 
4274  NumAllocaPartitions += NumPartitions;
4275  MaxPartitionsPerAlloca.updateMax(NumPartitions);
4276 
4277  // Migrate debug information from the old alloca to the new alloca(s)
4278  // and the individual partitions.
4280  if (!DbgDeclares.empty()) {
4281  auto *Var = DbgDeclares.front()->getVariable();
4282  auto *Expr = DbgDeclares.front()->getExpression();
4283  auto VarSize = Var->getSizeInBits();
4284  DIBuilder DIB(*AI.getModule(), /*AllowUnresolved*/ false);
4285  uint64_t AllocaSize = DL.getTypeSizeInBits(AI.getAllocatedType());
4286  for (auto Fragment : Fragments) {
4287  // Create a fragment expression describing the new partition or reuse AI's
4288  // expression if there is only one partition.
4289  auto *FragmentExpr = Expr;
4290  if (Fragment.Size < AllocaSize || Expr->isFragment()) {
4291  // If this alloca is already a scalar replacement of a larger aggregate,
4292  // Fragment.Offset describes the offset inside the scalar.
4293  auto ExprFragment = Expr->getFragmentInfo();
4294  uint64_t Offset = ExprFragment ? ExprFragment->OffsetInBits : 0;
4295  uint64_t Start = Offset + Fragment.Offset;
4296  uint64_t Size = Fragment.Size;
4297  if (ExprFragment) {
4298  uint64_t AbsEnd =
4299  ExprFragment->OffsetInBits + ExprFragment->SizeInBits;
4300  if (Start >= AbsEnd)
4301  // No need to describe a SROAed padding.
4302  continue;
4303  Size = std::min(Size, AbsEnd - Start);
4304  }
4305  // The new, smaller fragment is stenciled out from the old fragment.
4306  if (auto OrigFragment = FragmentExpr->getFragmentInfo()) {
4307  assert(Start >= OrigFragment->OffsetInBits &&
4308  "new fragment is outside of original fragment");
4309  Start -= OrigFragment->OffsetInBits;
4310  }
4311 
4312  // The alloca may be larger than the variable.
4313  if (VarSize) {
4314  if (Size > *VarSize)
4315  Size = *VarSize;
4316  if (Size == 0 || Start + Size > *VarSize)
4317  continue;
4318  }
4319 
4320  // Avoid creating a fragment expression that covers the entire variable.
4321  if (!VarSize || *VarSize != Size) {
4322  if (auto E =
4323  DIExpression::createFragmentExpression(Expr, Start, Size))
4324  FragmentExpr = *E;
4325  else
4326  continue;
4327  }
4328  }
4329 
4330  // Remove any existing intrinsics describing the same alloca.
4331  for (DbgVariableIntrinsic *OldDII : FindDbgAddrUses(Fragment.Alloca))
4332  OldDII->eraseFromParent();
4333 
4334  DIB.insertDeclare(Fragment.Alloca, Var, FragmentExpr,
4335  DbgDeclares.front()->getDebugLoc(), &AI);
4336  }
4337  }
4338  return Changed;
4339 }
4340 
4341 /// Clobber a use with undef, deleting the used value if it becomes dead.
4342 void SROA::clobberUse(Use &U) {
4343  Value *OldV = U;
4344  // Replace the use with an undef value.
4345  U = UndefValue::get(OldV->getType());
4346 
4347  // Check for this making an instruction dead. We have to garbage collect
4348  // all the dead instructions to ensure the uses of any alloca end up being
4349  // minimal.
4350  if (Instruction *OldI = dyn_cast<Instruction>(OldV))
4351  if (isInstructionTriviallyDead(OldI)) {
4352  DeadInsts.insert(OldI);
4353  }
4354 }
4355 
4356 /// Analyze an alloca for SROA.
4357 ///
4358 /// This analyzes the alloca to ensure we can reason about it, builds
4359 /// the slices of the alloca, and then hands it off to be split and
4360 /// rewritten as needed.
4361 bool SROA::runOnAlloca(AllocaInst &AI) {
4362  LLVM_DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
4363  ++NumAllocasAnalyzed;
4364 
4365  // Special case dead allocas, as they're trivial.
4366  if (AI.use_empty()) {
4367  AI.eraseFromParent();
4368  return true;
4369  }
4370  const DataLayout &DL = AI.getModule()->getDataLayout();
4371 
4372  // Skip alloca forms that this analysis can't handle.
4373  if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
4374  DL.getTypeAllocSize(AI.getAllocatedType()) == 0)
4375  return false;
4376 
4377  bool Changed = false;
4378 
4379  // First, split any FCA loads and stores touching this alloca to promote
4380  // better splitting and promotion opportunities.
4381  AggLoadStoreRewriter AggRewriter(DL);
4382  Changed |= AggRewriter.rewrite(AI);
4383 
4384  // Build the slices using a recursive instruction-visiting builder.
4385  AllocaSlices AS(DL, AI);
4386  LLVM_DEBUG(AS.print(dbgs()));
4387  if (AS.isEscaped())
4388  return Changed;
4389 
4390  // Delete all the dead users of this alloca before splitting and rewriting it.
4391  for (Instruction *DeadUser : AS.getDeadUsers()) {
4392  // Free up everything used by this instruction.
4393  for (Use &DeadOp : DeadUser->operands())
4394  clobberUse(DeadOp);
4395 
4396  // Now replace the uses of this instruction.
4397  DeadUser->replaceAllUsesWith(UndefValue::get(DeadUser->getType()));
4398 
4399  // And mark it for deletion.
4400  DeadInsts.insert(DeadUser);
4401  Changed = true;
4402  }
4403  for (Use *DeadOp : AS.getDeadOperands()) {
4404  clobberUse(*DeadOp);
4405  Changed = true;
4406  }
4407 
4408  // No slices to split. Leave the dead alloca for a later pass to clean up.
4409  if (AS.begin() == AS.end())
4410  return Changed;
4411 
4412  Changed |= splitAlloca(AI, AS);
4413 
4414  LLVM_DEBUG(dbgs() << " Speculating PHIs\n");
4415  while (!SpeculatablePHIs.empty())
4416  speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
4417 
4418  LLVM_DEBUG(dbgs() << " Speculating Selects\n");
4419  while (!SpeculatableSelects.empty())
4420  speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
4421 
4422  return Changed;
4423 }
4424 
4425 /// Delete the dead instructions accumulated in this run.
4426 ///
4427 /// Recursively deletes the dead instructions we've accumulated. This is done
4428 /// at the very end to maximize locality of the recursive delete and to
4429 /// minimize the problems of invalidated instruction pointers as such pointers
4430 /// are used heavily in the intermediate stages of the algorithm.
4431 ///
4432 /// We also record the alloca instructions deleted here so that they aren't
4433 /// subsequently handed to mem2reg to promote.
4434 bool SROA::deleteDeadInstructions(
4435  SmallPtrSetImpl<AllocaInst *> &DeletedAllocas) {
4436  bool Changed = false;
4437  while (!DeadInsts.empty()) {
4438  Instruction *I = DeadInsts.pop_back_val();
4439  LLVM_DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
4440 
4441  // If the instruction is an alloca, find the possible dbg.declare connected
4442  // to it, and remove it too. We must do this before calling RAUW or we will
4443  // not be able to find it.
4444  if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
4445  DeletedAllocas.insert(AI);
4446  for (DbgVariableIntrinsic *OldDII : FindDbgAddrUses(AI))
4447  OldDII->eraseFromParent();
4448  }
4449 
4451 
4452  for (Use &Operand : I->operands())
4453  if (Instruction *U = dyn_cast<Instruction>(Operand)) {
4454  // Zero out the operand and see if it becomes trivially dead.
4455  Operand = nullptr;
4457  DeadInsts.insert(U);
4458  }
4459 
4460  ++NumDeleted;
4461  I->eraseFromParent();
4462  Changed = true;
4463  }
4464  return Changed;
4465 }
4466 
4467 /// Promote the allocas, using the best available technique.
4468 ///
4469 /// This attempts to promote whatever allocas have been identified as viable in
4470 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
4471 /// This function returns whether any promotion occurred.
4472 bool SROA::promoteAllocas(Function &F) {
4473  if (PromotableAllocas.empty())
4474  return false;
4475 
4476  NumPromoted += PromotableAllocas.size();
4477 
4478  LLVM_DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
4479  PromoteMemToReg(PromotableAllocas, *DT, AC);
4480  PromotableAllocas.clear();
4481  return true;
4482 }
4483 
4484 PreservedAnalyses SROA::runImpl(Function &F, DominatorTree &RunDT,
4485  AssumptionCache &RunAC) {
4486  LLVM_DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
4487  C = &F.getContext();
4488  DT = &RunDT;
4489  AC = &RunAC;
4490 
4491  BasicBlock &EntryBB = F.getEntryBlock();
4492  for (BasicBlock::iterator I = EntryBB.begin(), E = std::prev(EntryBB.end());
4493  I != E; ++I) {
4494  if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
4495  Worklist.insert(AI);
4496  }
4497 
4498  bool Changed = false;
4499  // A set of deleted alloca instruction pointers which should be removed from
4500  // the list of promotable allocas.
4501  SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
4502 
4503  do {
4504  while (!Worklist.empty()) {
4505  Changed |= runOnAlloca(*Worklist.pop_back_val());
4506  Changed |= deleteDeadInstructions(DeletedAllocas);
4507 
4508  // Remove the deleted allocas from various lists so that we don't try to
4509  // continue processing them.
4510  if (!DeletedAllocas.empty()) {
4511  auto IsInSet = [&](AllocaInst *AI) { return DeletedAllocas.count(AI); };
4512  Worklist.remove_if(IsInSet);
4513  PostPromotionWorklist.remove_if(IsInSet);
4514  PromotableAllocas.erase(llvm::remove_if(PromotableAllocas, IsInSet),
4515  PromotableAllocas.end());
4516  DeletedAllocas.clear();
4517  }
4518  }
4519 
4520  Changed |= promoteAllocas(F);
4521 
4522  Worklist = PostPromotionWorklist;
4523  PostPromotionWorklist.clear();
4524  } while (!Worklist.empty());
4525 
4526  if (!Changed)
4527  return PreservedAnalyses::all();
4528 
4529  PreservedAnalyses PA;
4530  PA.preserveSet<CFGAnalyses>();
4531  PA.preserve<GlobalsAA>();
4532  return PA;
4533 }
4534 
4536  return runImpl(F, AM.getResult<DominatorTreeAnalysis>(F),
4537  AM.getResult<AssumptionAnalysis>(F));
4538 }
4539 
4540 /// A legacy pass for the legacy pass manager that wraps the \c SROA pass.
4541 ///
4542 /// This is in the llvm namespace purely to allow it to be a friend of the \c
4543 /// SROA pass.
4545  /// The SROA implementation.
4546  SROA Impl;
4547 
4548 public:
4549  static char ID;
4550 
4553  }
4554 
4555  bool runOnFunction(Function &F) override {
4556  if (skipFunction(F))
4557  return false;
4558 
4559  auto PA = Impl.runImpl(
4560  F, getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
4561  getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F));
4562  return !PA.areAllPreserved();
4563  }
4564 
4565  void getAnalysisUsage(AnalysisUsage &AU) const override {
4569  AU.setPreservesCFG();
4570  }
4571 
4572  StringRef getPassName() const override { return "SROA"; }
4573 };
4574 
4575 char SROALegacyPass::ID = 0;
4576 
4578 
4580  "Scalar Replacement Of Aggregates", false, false)
4584  false, false)
Legacy wrapper pass to provide the GlobalsAAResult object.
static Value * getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr, APInt Offset, Type *PointerTy, Twine NamePrefix)
Compute an adjusted pointer from Ptr by Offset bytes where the resulting pointer has PointerTy...
Definition: SROA.cpp:1545
static unsigned getBitWidth(Type *Ty, const DataLayout &DL)
Returns the bitwidth of the given scalar or pointer type.
static VectorType * isVectorPromotionViable(Partition &P, const DataLayout &DL)
Test whether the given alloca partitioning and range of slices can be promoted to a vector...
Definition: SROA.cpp:1842
iterator end() const
Definition: SROA.cpp:409
uint64_t CallInst * C
static Value * getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr, Type *Ty, APInt &Offset, Type *TargetTy, SmallVectorImpl< Value *> &Indices, Twine NamePrefix)
Recursively compute indices for a natural GEP.
Definition: SROA.cpp:1431
Value * getValueOperand()
Definition: Instructions.h:409
SymbolTableList< Instruction >::iterator eraseFromParent()
This method unlinks &#39;this&#39; from the containing basic block and deletes it.
Definition: Instruction.cpp:67
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:110
const_iterator end(StringRef path)
Get end iterator over path.
Definition: Path.cpp:258
constexpr char Align[]
Key for Kernel::Arg::Metadata::mAlign.
uint64_t getTypeStoreSizeInBits(Type *Ty) const
Returns the maximum number of bits that may be overwritten by storing the specified type; always a mu...
Definition: DataLayout.h:426
An iterator over partitions of the alloca&#39;s slices.
Definition: SROA.cpp:429
void setInt(IntType IntVal)
bool isSimple() const
Definition: Instructions.h:276
static bool runImpl(Function &F, TargetLibraryInfo &TLI, DominatorTree &DT)
This is the entry point for all transforms.
iterator_range< use_iterator > uses()
Definition: Value.h:354
AnalysisUsage & addPreserved()
Add the specified Pass class to the set of analyses preserved by this pass.
unsigned getIndexSizeInBits(unsigned AS) const
Size in bits of index used for address calculation in getelementptr.
Definition: DataLayout.h:372
void addIncoming(Value *V, BasicBlock *BB)
Add an incoming value to the end of the PHI list.
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
uint64_t getZExtValue() const
Get zero extended value.
Definition: APInt.h:1562
GCNRegPressure max(const GCNRegPressure &P1, const GCNRegPressure &P2)
Base class for instruction visitors.
Definition: InstVisitor.h:80
const_iterator begin(StringRef path, Style style=Style::native)
Get begin iterator over path.
Definition: Path.cpp:249
const Value * stripInBoundsOffsets() const
Strip off pointer casts and inbounds GEPs.
Definition: Value.cpp:588
uint64_t beginOffset() const
The start offset of this partition.
Definition: SROA.cpp:380
This is a &#39;bitvector&#39; (really, a variable-sized bit array), optimized for the case when the array is ...
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
Definition: PassManager.h:769
This class represents lattice values for constants.
Definition: AllocatorList.h:23
PointerTy getPointer() const
#define LLVM_DUMP_METHOD
Mark debug helper function definitions like dump() that should not be stripped from debug builds...
Definition: Compiler.h:464
Type * getElementType(unsigned N) const
Definition: DerivedTypes.h:313
This is the interface for a simple mod/ref and alias analysis over globals.
EltTy front() const
bool isSized(SmallPtrSetImpl< Type *> *Visited=nullptr) const
Return true if it makes sense to take the size of this type.
Definition: Type.h:264
iterator begin() const
Definition: ArrayRef.h:136
constexpr char IsVolatile[]
Key for Kernel::Arg::Metadata::mIsVolatile.
APInt sdiv(const APInt &RHS) const
Signed division function for APInt.
Definition: APInt.cpp:1590
LLVM_NODISCARD size_t rfind(char C, size_t From=npos) const
Search for the last character C in the string.
Definition: StringRef.h:359
const StructLayout * getStructLayout(StructType *Ty) const
Returns a StructLayout object, indicating the alignment of the struct, its size, and the offsets of i...
Definition: DataLayout.cpp:587
This file provides the interface for LLVM&#39;s Scalar Replacement of Aggregates pass.
void push_back(const T &Elt)
Definition: SmallVector.h:217
This provides a very simple, boring adaptor for a begin and end iterator into a range type...
void erase(iterator Start, iterator Stop)
Erase a range of slices.
Definition: SROA.cpp:258
bool isTriviallyEmpty() const
Check if this twine is trivially empty; a false return value does not necessarily mean the twine is e...
Definition: Twine.h:400
This class provides information about the result of a visit.
Definition: PtrUseVisitor.h:62
This class represents a function call, abstracting a target machine&#39;s calling convention.
static Value * extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V, IntegerType *Ty, uint64_t Offset, const Twine &Name)
Definition: SROA.cpp:2079
This file contains the declarations for metadata subclasses.
Representation of the alloca slices.
Definition: SROA.cpp:231
An immutable pass that tracks lazily created AssumptionCache objects.
unsigned getSourceAlignment() const
Instruction * getAbortingInst() const
Get the instruction causing the visit to abort.
Definition: PtrUseVisitor.h:83
SyncScope::ID getSyncScopeID() const
Returns the synchronization scope ID of this store instruction.
Definition: Instructions.h:384
gep_type_iterator gep_type_end(const User *GEP)
const Value * getTrueValue() const
void insert(ArrayRef< Slice > NewSlices)
Insert new slices for this alloca.
Definition: SROA.cpp:265
Value * getValue() const
A cache of @llvm.assume calls within a function.
AtomicOrdering getOrdering() const
Returns the ordering constraint of this load instruction.
Definition: Instructions.h:247
Offsets
Offsets in bytes from the start of the input buffer.
Definition: SIInstrInfo.h:1024
LLVMContext & getContext() const
All values hold a context through their type.
Definition: Value.cpp:704
void deleteValue()
Delete a pointer to a generic Value.
Definition: Value.cpp:97
void setSource(Value *Ptr)
void setDest(Value *Ptr)
Set the specified arguments of the instruction.
This class wraps the llvm.memset intrinsic.
const Use & getOperandUse(unsigned i) const
Definition: User.h:182
unsigned second
Scalar Replacement Of Aggregates
Definition: SROA.cpp:4583
bool all_of(R &&range, UnaryPredicate P)
Provide wrappers to std::all_of which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:1185
STATISTIC(NumFunctions, "Total number of functions")
Metadata node.
Definition: Metadata.h:863
Analysis pass which computes a DominatorTree.
Definition: Dominators.h:230
F(f)
unsigned getPointerAddressSpace() const
Get the address space of this pointer or pointer vector type.
Definition: DerivedTypes.h:502
An instruction for reading from memory.
Definition: Instructions.h:167
APInt zextOrTrunc(unsigned width) const
Zero extend or truncate to width.
Definition: APInt.cpp:875
Hexagon Common GEP
const Instruction * getTerminator() const LLVM_READONLY
Returns the terminator instruction if the block is well formed or null if the block is not well forme...
Definition: BasicBlock.cpp:137
bool isVectorTy() const
True if this is an instance of VectorType.
Definition: Type.h:229
This defines the Use class.
void reserve(size_type N)
Definition: SmallVector.h:374
Value * getLength() const
void setAtomic(AtomicOrdering Ordering, SyncScope::ID SSID=SyncScope::System)
Sets the ordering constraint and the synchronization scope ID of this store instruction.
Definition: Instructions.h:395
bool isEscaped() const
Test whether a pointer to the allocation escapes our analysis.
Definition: SROA.cpp:240
TinyPtrVector - This class is specialized for cases where there are normally 0 or 1 element in a vect...
Definition: TinyPtrVector.h:30
bool isSafeToLoadUnconditionally(Value *V, unsigned Align, const DataLayout &DL, Instruction *ScanFrom=nullptr, const DominatorTree *DT=nullptr)
Return true if we know that executing a load from this value cannot trap.
Definition: Loads.cpp:200
op_iterator op_begin()
Definition: User.h:229
unsigned getElementContainingOffset(uint64_t Offset) const
Given a valid byte offset into the structure, returns the structure index that contains it...
Definition: DataLayout.cpp:83
unsigned getBitWidth() const
Return the number of bits in the APInt.
Definition: APInt.h:1508
static unsigned getAdjustedAlignment(Instruction *I, uint64_t Offset, const DataLayout &DL)
Compute the adjusted alignment for a load or store from an offset.
Definition: SROA.cpp:1640
Builder for the alloca slices.
Definition: SROA.cpp:647
LLVMContext & getContext() const
Return the LLVMContext in which this type was uniqued.
Definition: Type.h:129
A templated base class for SmallPtrSet which provides the typesafe interface that is common across al...
Definition: SmallPtrSet.h:343
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:268
Value * getArgOperand(unsigned i) const
Definition: InstrTypes.h:1134
std::pair< iterator, bool > insert(const std::pair< KeyT, ValueT > &KV)
Definition: DenseMap.h:221
void setAtomic(AtomicOrdering Ordering, SyncScope::ID SSID=SyncScope::System)
Sets the ordering constraint and the synchronization scope ID of this load instruction.
Definition: Instructions.h:270
static bool isVectorPromotionViableForSlice(Partition &P, const Slice &S, VectorType *Ty, uint64_t ElementSize, const DataLayout &DL)
Test whether the given slice use can be promoted to a vector.
Definition: SROA.cpp:1767
PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM)
Run the pass over the function.
Definition: SROA.cpp:4535
void * PointerTy
Definition: GenericValue.h:21
AnalysisUsage & addRequired()
Used to lazily calculate structure layout information for a target machine, based on the DataLayout s...
Definition: DataLayout.h:528
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:50
void setPointer(PointerTy PtrVal)
bool isVolatile() const
Return true if this is a load from a volatile memory location.
Definition: Instructions.h:231
amdgpu Simplify well known AMD library false Value Value const Twine & Name
This class represents the LLVM &#39;select&#39; instruction.
Type * getPointerElementType() const
Definition: Type.h:375
const DataLayout & getDataLayout() const
Get the data layout for the module&#39;s target platform.
Definition: Module.cpp:370
Twine - A lightweight data structure for efficiently representing the concatenation of temporary valu...
Definition: Twine.h:80
unsigned getAlignment() const
Return the alignment of the memory that is being allocated by the instruction.
Definition: Instructions.h:112
ArrayRef< T > makeArrayRef(const T &OneElt)
Construct an ArrayRef from a single element.
Definition: ArrayRef.h:450
PointerType * getType() const
Overload to return most specific pointer type.
Definition: Instructions.h:96
Class to represent struct types.
Definition: DerivedTypes.h:200
A Use represents the edge between a Value definition and its users.
Definition: Use.h:55
static Type * getTypePartition(const DataLayout &DL, Type *Ty, uint64_t Offset, uint64_t Size)
Try to find a partition of the aggregate type passed in for a given offset and size.
Definition: SROA.cpp:3433
PointerType * getPointerTo(unsigned AddrSpace=0) const
Return a pointer to the current type.
Definition: Type.cpp:651
ArrayRef< Slice * > splitSliceTails() const
Get the sequence of split slice tails.
Definition: SROA.cpp:417
bool isAborted() const
Did we abort the visit early?
Definition: PtrUseVisitor.h:75
This class consists of common code factored out of the SmallVector class to reduce code duplication b...
Definition: APFloat.h:41
bool isIntegerTy() const
True if this is an instance of IntegerType.
Definition: Type.h:196
This provides a uniform API for creating instructions and inserting them into a basic block: either a...
Definition: IRBuilder.h:742
static Value * convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V, Type *NewTy)
Generic routine to convert an SSA value to a value of a different type.
Definition: SROA.cpp:1717
A partition of the slices.
Definition: SROA.cpp:355
unsigned getDestAlignment() const
void getAnalysisUsage(AnalysisUsage &AU) const override
getAnalysisUsage - This function should be overriden by passes that need analysis information to do t...
Definition: SROA.cpp:4565
uint64_t getNumElements() const
Definition: DerivedTypes.h:358
static StructType * get(LLVMContext &Context, ArrayRef< Type *> Elements, bool isPacked=false)
This static method is the primary way to create a literal StructType.
Definition: Type.cpp:341
This file implements a class to represent arbitrary precision integral constant values and operations...
bool runOnFunction(Function &F) override
runOnFunction - Virtual method overriden by subclasses to do the per-function processing of the pass...
Definition: SROA.cpp:4555
This file provides a collection of visitors which walk the (instruction) uses of a pointer...
static Constant * getZExt(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:1664
unsigned getPointerAddressSpace() const
Returns the address space of the pointer operand.
Definition: Instructions.h:418
User * getUser() const LLVM_READONLY
Returns the User that contains this Use.
Definition: Use.cpp:40
bool visit(AllocaSlices::const_iterator I)
Definition: SROA.cpp:2298
A base class for visitors over the uses of a pointer value.
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:244
bool insert(const value_type &X)
Insert a new element into the SetVector.
Definition: SetVector.h:141
void copyNonnullMetadata(const LoadInst &OldLI, MDNode *N, LoadInst &NewLI)
Copy a nonnull metadata node to a new load instruction.
Definition: Local.cpp:2478
bool isInBounds() const
Determine whether the GEP has the inbounds flag.
SmallVectorImpl< Slice >::iterator iterator
Support for iterating over the slices.
Definition: SROA.cpp:244
const_iterator end() const
Definition: SROA.cpp:254
IntType getInt() const
void setLength(Value *L)
A legacy pass for the legacy pass manager that wraps the SROA pass.
Definition: SROA.cpp:4544
Class to represent array types.
Definition: DerivedTypes.h:368
This is the common base class for debug info intrinsics for variables.
Definition: IntrinsicInst.h:87
APInt sextOrTrunc(unsigned width) const
Sign extend or truncate to width.
Definition: APInt.cpp:883
This class represents a no-op cast from one type to another.
MDNode * getMetadata(unsigned KindID) const
Get the metadata of given kind attached to this Instruction.
Definition: Instruction.h:220
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory)...
Definition: APInt.h:32
const APInt & getValue() const
Return the constant as an APInt value reference.
Definition: Constants.h:137
ConstantFolder - Create constants with minimum, target independent, folding.
An instruction for storing to memory.
Definition: Instructions.h:320
bool isIntOrIntVectorTy() const
Return true if this is an integer type or a vector of integer types.
Definition: Type.h:202
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
Definition: Value.cpp:428
void printUse(raw_ostream &OS, const_iterator I, StringRef Indent=" ") const
LLVM_NODISCARD LLVM_ATTRIBUTE_ALWAYS_INLINE StringRef substr(size_t Start, size_t N=npos) const
Return a reference to the substring from [Start, Start + N).
Definition: StringRef.h:597
static Constant * getUDiv(Constant *C1, Constant *C2, bool isExact=false)
Definition: Constants.cpp:2270
unsigned getBitWidth() const
Get the number of bits in this IntegerType.
Definition: DerivedTypes.h:65
Type::subtype_iterator element_iterator
Definition: DerivedTypes.h:300
CRTP base class which implements the entire standard iterator facade in terms of a minimal subset of ...
Definition: iterator.h:67
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree...
Definition: Dominators.h:144
static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy)
Test whether we can convert a value from the old to the new type.
Definition: SROA.cpp:1666
unsigned getNumSuccessors() const
Return the number of successors that this instruction has.
Value * getOperand(unsigned i) const
Definition: User.h:169
Class to represent pointers.
Definition: DerivedTypes.h:466
auto count(R &&Range, const E &Element) -> typename std::iterator_traits< decltype(adl_begin(Range))>::difference_type
Wrapper function around std::count to count the number of times an element Element occurs in the give...
Definition: STLExtras.h:1251
Type * getScalarType() const
If this is a vector type, return the element type, otherwise return &#39;this&#39;.
Definition: Type.h:303
iterator find(const_arg_type_t< KeyT > Val)
Definition: DenseMap.h:176
const BasicBlock & getEntryBlock() const
Definition: Function.h:639
constexpr uint64_t MinAlign(uint64_t A, uint64_t B)
A and B are either alignments or offsets.
Definition: MathExtras.h:609
an instruction for type-safe pointer arithmetic to access elements of arrays and structs ...
Definition: Instructions.h:873
IntegerType * getIntPtrType(LLVMContext &C, unsigned AddressSpace=0) const
Returns an integer type with size at least as big as that of a pointer in the given address space...
Definition: DataLayout.cpp:749
void getAAMetadata(AAMDNodes &N, bool Merge=false) const
Fills the AAMDNodes structure with AA metadata from this instruction.
Maximum number of bits that can be specified.
Definition: DerivedTypes.h:51
Scalar Replacement Of false
Definition: SROA.cpp:4583
iterator_range< partition_iterator > partitions()
#define P(N)
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:422
uint64_t getZExtValue() const
Return the constant as a 64-bit unsigned integer value after it has been zero extended as appropriate...
Definition: Constants.h:148
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
void dump(const SparseBitVector< ElementSize > &LHS, raw_ostream &out)
A set of analyses that are preserved following a run of a transformation pass.
Definition: PassManager.h:153
const_iterator getFirstInsertionPt() const
Returns an iterator to the first instruction in this block that is suitable for inserting a non-PHI i...
Definition: BasicBlock.cpp:216
void setDebugLoc(DebugLoc Loc)
Set the debug location information for this instruction.
Definition: Instruction.h:307
bool areAllPreserved() const
Test whether all analyses are preserved (and none are abandoned).
Definition: PassManager.h:321
void setAAMetadata(const AAMDNodes &N)
Sets the metadata on this instruction from the AAMDNodes structure.
Definition: Metadata.cpp:1264
LLVM Basic Block Representation.
Definition: BasicBlock.h:57
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:45
static unsigned getPointerOperandIndex()
Definition: Instructions.h:414
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
This is an important base class in LLVM.
Definition: Constant.h:41
LLVM_NODISCARD bool empty() const
Definition: SmallPtrSet.h:91
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator begin()
Definition: SmallVector.h:128
ArrayRef< Use * > getDeadOperands() const
Access the dead operands referring to this alloca.
Definition: SROA.cpp:287
This file contains the declarations for the subclasses of Constant, which represent the different fla...
bool isPointerTy() const
True if this is an instance of PointerType.
Definition: Type.h:223
static bool isSafeSelectToSpeculate(SelectInst &SI)
Select instructions that use an alloca and are subsequently loaded can be rewritten to load both inpu...
Definition: SROA.cpp:1299
LLVM_NODISCARD size_t find_first_not_of(char C, size_t From=0) const
Find the first character in the string that is not C or npos if not found.
Definition: StringRef.cpp:249
std::pair< iterator, bool > insert(PtrType Ptr)
Inserts Ptr if and only if there is no element in the container equal to Ptr.
Definition: SmallPtrSet.h:370
bool mayHaveSideEffects() const
Return true if the instruction may have side effects.
Definition: Instruction.h:561
static Value * getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr, APInt Offset, Type *TargetTy, SmallVectorImpl< Value *> &Indices, Twine NamePrefix)
Get a natural GEP from a base pointer to a particular offset and resulting in a particular type...
Definition: SROA.cpp:1505
static sys::TimePoint< std::chrono::seconds > now(bool Deterministic)
element_iterator element_end() const
Definition: DerivedTypes.h:303
DIExpression * getExpression() const
Represent the analysis usage information of a pass.
op_iterator op_end()
Definition: User.h:231
bool isLifetimeStartOrEnd() const
Return true if the instruction is a llvm.lifetime.start or llvm.lifetime.end marker.
static cl::opt< bool > SROAStrictInbounds("sroa-strict-inbounds", cl::init(false), cl::Hidden)
Hidden option to experiment with completely strict handling of inbounds GEPs.
bool any_of(R &&range, UnaryPredicate P)
Provide wrappers to std::any_of which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:1192
#define LLVM_ATTRIBUTE_UNUSED
Definition: Compiler.h:159
Analysis pass providing a never-invalidated alias analysis result.
static void replace(Module &M, GlobalVariable *Old, GlobalVariable *New)
void initializeSROALegacyPassPass(PassRegistry &)
sroa
Definition: SROA.cpp:4583
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:284
bool empty() const
op_range operands()
Definition: User.h:237
Value * getPointerOperand()
Definition: Instructions.h:284
unsigned getAddressSpace() const
Return the address space of the Pointer type.
Definition: DerivedTypes.h:494
static void print(raw_ostream &Out, object::Archive::Kind Kind, T Val)
size_type count(ConstPtrType Ptr) const
count - Return 1 if the specified pointer is in the set, 0 otherwise.
Definition: SmallPtrSet.h:381
Class to represent integer types.
Definition: DerivedTypes.h:39
unsigned getIndexTypeSizeInBits(Type *Ty) const
Layout size of the index used in GEP calculation.
Definition: DataLayout.cpp:661
AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &AS, SROA &Pass, AllocaInst &OldAI, AllocaInst &NewAI, uint64_t NewAllocaBeginOffset, uint64_t NewAllocaEndOffset, bool IsIntegerPromotable, VectorType *PromotableVecTy, SmallSetVector< PHINode *, 8 > &PHIUsers, SmallSetVector< SelectInst *, 8 > &SelectUsers)
Definition: SROA.cpp:2269
void setAlignment(unsigned Align)
auto remove_if(R &&Range, UnaryPredicate P) -> decltype(adl_begin(Range))
Provide wrappers to std::remove_if which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:1225
const Value * getCondition() const
static Constant * getAllOnesValue(Type *Ty)
Definition: Constants.cpp:318
LLVMContext & getContext() const
getContext - Return a reference to the LLVMContext associated with this function. ...
Definition: Function.cpp:192
static UndefValue * get(Type *T)
Static factory methods - Return an &#39;undef&#39; object of the specified type.
Definition: Constants.cpp:1414
iterator erase(const_iterator CI)
Definition: SmallVector.h:443
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:159
size_t size() const
Definition: SmallVector.h:52
Value * getIncomingValue(unsigned i) const
Return incoming value number x.
static PointerType * getInt8PtrTy(LLVMContext &C, unsigned AS=0)
Definition: Type.cpp:219
bool isVolatile() const
INITIALIZE_PASS_END(RegBankSelect, DEBUG_TYPE, "Assign register bank of generic virtual registers", false, false) RegBankSelect
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
Type * getAllocatedType() const
Return the type that is being allocated by the instruction.
Definition: Instructions.h:105
static Value * getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL, Value *BasePtr, Type *Ty, Type *TargetTy, SmallVectorImpl< Value *> &Indices, Twine NamePrefix)
Get a natural GEP off of the BasePtr walking through Ty toward TargetTy without changing the offset o...
Definition: SROA.cpp:1387
void sort(IteratorTy Start, IteratorTy End)
Definition: STLExtras.h:1115
void printSlice(raw_ostream &OS, const_iterator I, StringRef Indent=" ") const
static Type * stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty)
Strip aggregate type wrapping.
Definition: SROA.cpp:3395
Intrinsic::ID getIntrinsicID() const
Return the intrinsic ID of this intrinsic.
Definition: IntrinsicInst.h:50
SmallBitVector & reset()
This provides the default implementation of the IRBuilder &#39;InsertHelper&#39; method that is called whenev...
Definition: IRBuilder.h:61
This is the superclass of the array and vector type classes.
Definition: DerivedTypes.h:342
A function analysis which provides an AssumptionCache.
iterator_range< T > make_range(T x, T y)
Convenience function for iterating over sub-ranges.
bool isPtrOrPtrVectorTy() const
Return true if this is a pointer type or a vector of pointer types.
Definition: Type.h:226
A SetVector that performs no allocations if smaller than a certain size.
Definition: SetVector.h:297
print lazy value Lazy Value Info Printer Pass
This is the common base class for memset/memcpy/memmove.
Iterator for intrusive lists based on ilist_node.
uint64_t getLimitedValue(uint64_t Limit=~0ULL) const
getLimitedValue - If the value is smaller than the specified limit, return it, otherwise return the l...
Definition: Constants.h:250
This is the shared class of boolean and integer constants.
Definition: Constants.h:83
TinyPtrVector< DbgVariableIntrinsic * > FindDbgAddrUses(Value *V)
Finds all intrinsics declaring local variables as living in the memory that &#39;V&#39; points to...
Definition: Local.cpp:1472
iterator end()
Definition: BasicBlock.h:270
AllocaSlices(const DataLayout &DL, AllocaInst &AI)
Construct the slices of a particular alloca.
Module.h This file contains the declarations for the Module class.
Instruction * user_back()
Specialize the methods defined in Value, as we know that an instruction can only be used by other ins...
Definition: Instruction.h:63
partition_iterator & operator++()
Definition: SROA.cpp:602
static Value * insertVector(IRBuilderTy &IRB, Value *Old, Value *V, unsigned BeginIndex, const Twine &Name)
Definition: SROA.cpp:2159
iterator end() const
Definition: ArrayRef.h:137
bool isLegalInteger(uint64_t Width) const
Returns true if the specified type is known to be a native integer type supported by the CPU...
Definition: DataLayout.h:242
unsigned getABITypeAlignment(Type *Ty) const
Returns the minimum ABI-required alignment for the specified type.
Definition: DataLayout.cpp:729
A collection of metadata nodes that might be associated with a memory access used by the alias-analys...
Definition: Metadata.h:643
iterator begin() const
Definition: SROA.cpp:408
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:379
uint64_t getSizeInBytes() const
Definition: DataLayout.h:536
SmallVectorImpl< Slice >::const_iterator const_iterator
Definition: SROA.cpp:250
static IntegerType * getIntNTy(LLVMContext &C, unsigned N)
Definition: Type.cpp:179
SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &AS)
Definition: SROA.cpp:663
static Constant * get(Type *Ty, uint64_t V, bool isSigned=false)
If Ty is a vector type, return a Constant with a splat of the given value.
Definition: Constants.cpp:621
void setPreservesCFG()
This function should be called by the pass, iff they do not:
Definition: Pass.cpp:285
Value * getRawSource() const
Return the arguments to the instruction.
unsigned getNumIncomingValues() const
Return the number of incoming edges.
bool uge(const APInt &RHS) const
Unsigned greater or equal comparison.
Definition: APInt.h:1292
void setOperand(unsigned i, Value *Val)
Definition: User.h:174
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:132
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:940
A range adaptor for a pair of iterators.
Class to represent vector types.
Definition: DerivedTypes.h:392
const Module * getModule() const
Return the module owning the function this instruction belongs to or nullptr it the function does not...
Definition: Instruction.cpp:55
Class for arbitrary precision integers.
Definition: APInt.h:69
static bool isIntegerWideningViableForSlice(const Slice &S, uint64_t AllocBeginOffset, Type *AllocaTy, const DataLayout &DL, bool &WholeAllocaOp)
Test whether a slice of an alloca is valid for integer widening.
Definition: SROA.cpp:1950
iterator_range< user_iterator > users()
Definition: Value.h:399
#define NDEBUG
Definition: regutils.h:48
Represents analyses that only rely on functions&#39; control flow.
Definition: PassManager.h:114
const Value * getFalseValue() const
element_iterator element_begin() const
Definition: DerivedTypes.h:302
static Value * insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old, Value *V, uint64_t Offset, const Twine &Name)
Definition: SROA.cpp:2102
bool isNonIntegralPointerType(PointerType *PT) const
Definition: DataLayout.h:348
const_iterator begin() const
Definition: SROA.cpp:253
static Type * findCommonType(AllocaSlices::const_iterator B, AllocaSlices::const_iterator E, uint64_t EndOffset)
Walk the range of a partitioning looking for a common type to cover this sequence of slices...
Definition: SROA.cpp:1100
uint64_t getTypeSizeInBits(Type *Ty) const
Size examples:
Definition: DataLayout.h:567
static Value * foldPHINodeOrSelectInst(Instruction &I)
A helper that folds a PHI node or a select.
Definition: SROA.cpp:635
uint64_t getTypeAllocSize(Type *Ty) const
Returns the offset in bytes between successive objects of the specified type, including alignment pad...
Definition: DataLayout.h:435
Virtual Register Rewriter
Definition: VirtRegMap.cpp:221
bool operator!=(uint64_t V1, const APInt &V2)
Definition: APInt.h:1968
INITIALIZE_PASS_BEGIN(SROALegacyPass, "sroa", "Scalar Replacement Of Aggregates", false, false) INITIALIZE_PASS_END(SROALegacyPass
bool ugt(const APInt &RHS) const
Unsigned greather than comparison.
Definition: APInt.h:1254
static bool isZero(Value *V, const DataLayout &DL, DominatorTree *DT, AssumptionCache *AC)
Definition: Lint.cpp:545
This class wraps the llvm.memcpy/memmove intrinsics.
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator end()
Definition: SmallVector.h:132
bool isPacked() const
Definition: DerivedTypes.h:260
bool isVolatile() const
Return true if this is a store to a volatile memory location.
Definition: Instructions.h:353
const DebugLoc & getDebugLoc() const
Return the debug location for this node as a DebugLoc.
Definition: Instruction.h:310
static const size_t npos
Definition: StringRef.h:50
uint64_t getElementOffset(unsigned Idx) const
Definition: DataLayout.h:550
unsigned getAlignment() const
Return the alignment of the access that is being performed.
Definition: Instructions.h:240
unsigned getIntegerBitWidth() const
Definition: DerivedTypes.h:96
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:55
AtomicOrdering getOrdering() const
Returns the ordering constraint of this store instruction.
Definition: Instructions.h:372
StringRef getPassName() const override
getPassName - Return a nice clean name for a pass.
Definition: SROA.cpp:4572
uint64_t getLimitedValue(uint64_t Limit=UINT64_MAX) const
If this value is smaller than the specified limit, return it, otherwise return the limit value...
Definition: APInt.h:481
Visitor to rewrite instructions using p particular slice of an alloca to use a new alloca...
Definition: SROA.cpp:2212
void preserveSet()
Mark an analysis set as preserved.
Definition: PassManager.h:189
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:213
BasicBlock * getIncomingBlock(unsigned i) const
Return incoming basic block number i.
static void speculateSelectInstLoads(SelectInst &SI)
Definition: SROA.cpp:1321
SyncScope::ID getSyncScopeID() const
Returns the synchronization scope ID of this load instruction.
Definition: Instructions.h:259
#define I(x, y, z)
Definition: MD5.cpp:58
#define N
iterator end()
Definition: DenseMap.h:108
FunctionPass * createSROAPass()
Definition: SROA.cpp:4577
uint64_t endOffset() const
The end offset of this partition.
Definition: SROA.cpp:385
Instruction * getEscapingInst() const
Get the instruction causing the pointer to escape.
Definition: PtrUseVisitor.h:88
static ArrayType * get(Type *ElementType, uint64_t NumElements)
This static method is the primary way to construct an ArrayType.
Definition: Type.cpp:580
LLVM_NODISCARD std::enable_if<!is_simple_type< Y >::value, typename cast_retty< X, const Y >::ret_type >::type dyn_cast(const Y &Val)
Definition: Casting.h:322
uint32_t Size
Definition: Profile.cpp:46
void preserve()
Mark an analysis as preserved.
Definition: PassManager.h:174
DILocalVariable * getVariable() const
static Value * extractVector(IRBuilderTy &IRB, Value *V, unsigned BeginIndex, unsigned EndIndex, const Twine &Name)
Definition: SROA.cpp:2133
size_type count(const_arg_type_t< KeyT > Val) const
Return 1 if the specified key is in the map, 0 otherwise.
Definition: DenseMap.h:171
unsigned getAlignment() const
Return the alignment of the access that is being performed.
Definition: Instructions.h:365
std::string str() const
Return the twine contents as a std::string.
Definition: Twine.cpp:17
ValueT lookup(const_arg_type_t< KeyT > Val) const
lookup - Return the entry for the specified key, or a default constructed value if no such entry exis...
Definition: DenseMap.h:211
unsigned getPointerAddressSpace() const
Returns the address space of the pointer operand.
Definition: Instructions.h:290
static unsigned getPointerOperandIndex()
Definition: Instructions.h:286
bool isArrayAllocation() const
Return true if there is an allocation size parameter to the allocation instruction that is not 1...
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
user_iterator user_begin()
Definition: Value.h:375
Definition: JSON.cpp:591
bool operator<(int64_t V1, const APSInt &V2)
Definition: APSInt.h:325
bool isSingleValueType() const
Return true if the type is a valid type for a register in codegen.
Definition: Type.h:249
bool isInstructionTriviallyDead(Instruction *I, const TargetLibraryInfo *TLI=nullptr)
Return true if the result produced by the instruction is not used, and the instruction has no side ef...
Definition: Local.cpp:348
LLVM Value Representation.
Definition: Value.h:72
void setAlignment(unsigned Align)
static Value * buildGEP(IRBuilderTy &IRB, Value *BasePtr, SmallVectorImpl< Value *> &Indices, Twine NamePrefix)
Build a GEP out of a base pointer and indices.
Definition: SROA.cpp:1364
uint64_t getTypeStoreSize(Type *Ty) const
Returns the maximum number of bytes that may be overwritten by storing the specified type...
Definition: DataLayout.h:418
static bool isSafePHIToSpeculate(PHINode &PN)
PHI instructions that use an alloca and are subsequently loaded can be rewritten to load both input p...
Definition: SROA.cpp:1167
unsigned getOpcode() const
Return the opcode for this Instruction or ConstantExpr.
Definition: Operator.h:40
void copyMetadata(const Instruction &SrcInst, ArrayRef< unsigned > WL=ArrayRef< unsigned >())
Copy metadata from SrcInst to this instruction.
static VectorType * get(Type *ElementType, unsigned NumElements)
This static method is the primary way to construct an VectorType.
Definition: Type.cpp:605
std::underlying_type< E >::type Mask()
Get a bitmask with 1s in all places up to the high-order bit of E&#39;s largest value.
Definition: BitmaskEnum.h:80
void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB, BasicBlock::iterator InsertPt) const
Definition: IRBuilder.h:63
static void speculatePHINodeLoads(PHINode &PN)
Definition: SROA.cpp:1231
This class implements an extremely fast bulk output stream that can only output to a stream...
Definition: raw_ostream.h:45
void PromoteMemToReg(ArrayRef< AllocaInst *> Allocas, DominatorTree &DT, AssumptionCache *AC=nullptr)
Promote the specified list of alloca instructions into scalar registers, inserting PHI nodes as appro...
PtrInfo visitPtr(Instruction &I)
Recursively visit the uses of the given pointer.
An optimization pass providing Scalar Replacement of Aggregates.
Definition: SROA.h:65
Type * getElementType() const
Definition: DerivedTypes.h:359
static cl::opt< bool > SROARandomShuffleSlices("sroa-random-shuffle-slices", cl::init(false), cl::Hidden)
Hidden option to enable randomly shuffling the slices to help uncover instability in their order...
uint64_t size() const
The size of the partition.
Definition: SROA.cpp:390
IRTranslator LLVM IR MI
bool hasOneUse() const
Return true if there is exactly one user of this value.
Definition: Value.h:412
bool empty() const
Test whether this partition contains no slices, and merely spans a region occupied by split slices...
Definition: SROA.cpp:397
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:48
void setSourceAlignment(unsigned Align)
A container for analyses that lazily runs them and caches their results.
void print(raw_ostream &OS, const_iterator I, StringRef Indent=" ") const
Type * getArrayElementType() const
Definition: Type.h:364
Legacy analysis pass which computes a DominatorTree.
Definition: Dominators.h:259
static bool isIntegerWideningViable(Partition &P, Type *AllocaTy, const DataLayout &DL)
Test whether the given alloca partition&#39;s integer operations can be widened to promotable ones...
Definition: SROA.cpp:2037
bool operator==(uint64_t V1, const APInt &V2)
Definition: APInt.h:1966
bool isSimple() const
Definition: Instructions.h:401
This header defines various interfaces for pass management in LLVM.
bool isBigEndian() const
Definition: DataLayout.h:221
LLVM_NODISCARD LLVM_ATTRIBUTE_ALWAYS_INLINE size_t find(char C, size_t From=0) const
Search for the first character C in the string.
Definition: StringRef.h:297
#define LLVM_DEBUG(X)
Definition: