LLVM  8.0.0svn
SROA.cpp
Go to the documentation of this file.
1 //===- SROA.cpp - Scalar Replacement Of Aggregates ------------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 /// \file
10 /// This transformation implements the well known scalar replacement of
11 /// aggregates transformation. It tries to identify promotable elements of an
12 /// aggregate alloca, and promote them to registers. It will also try to
13 /// convert uses of an element (or set of elements) of an alloca into a vector
14 /// or bitfield-style integer scalar if appropriate.
15 ///
16 /// It works to do this with minimal slicing of the alloca so that regions
17 /// which are merely transferred in and out of external memory remain unchanged
18 /// and are not decomposed to scalar code.
19 ///
20 /// Because this also performs alloca promotion, it can be thought of as also
21 /// serving the purpose of SSA formation. The algorithm iterates on the
22 /// function until all opportunities for promotion have been realized.
23 ///
24 //===----------------------------------------------------------------------===//
25 
27 #include "llvm/ADT/APInt.h"
28 #include "llvm/ADT/ArrayRef.h"
29 #include "llvm/ADT/DenseMap.h"
31 #include "llvm/ADT/STLExtras.h"
32 #include "llvm/ADT/SetVector.h"
34 #include "llvm/ADT/SmallPtrSet.h"
35 #include "llvm/ADT/SmallVector.h"
36 #include "llvm/ADT/Statistic.h"
37 #include "llvm/ADT/StringRef.h"
38 #include "llvm/ADT/Twine.h"
39 #include "llvm/ADT/iterator.h"
43 #include "llvm/Analysis/Loads.h"
46 #include "llvm/Config/llvm-config.h"
47 #include "llvm/IR/BasicBlock.h"
48 #include "llvm/IR/Constant.h"
49 #include "llvm/IR/ConstantFolder.h"
50 #include "llvm/IR/Constants.h"
51 #include "llvm/IR/DIBuilder.h"
52 #include "llvm/IR/DataLayout.h"
54 #include "llvm/IR/DerivedTypes.h"
55 #include "llvm/IR/Dominators.h"
56 #include "llvm/IR/Function.h"
58 #include "llvm/IR/GlobalAlias.h"
59 #include "llvm/IR/IRBuilder.h"
60 #include "llvm/IR/InstVisitor.h"
61 #include "llvm/IR/InstrTypes.h"
62 #include "llvm/IR/Instruction.h"
63 #include "llvm/IR/Instructions.h"
64 #include "llvm/IR/IntrinsicInst.h"
65 #include "llvm/IR/Intrinsics.h"
66 #include "llvm/IR/LLVMContext.h"
67 #include "llvm/IR/Metadata.h"
68 #include "llvm/IR/Module.h"
69 #include "llvm/IR/Operator.h"
70 #include "llvm/IR/PassManager.h"
71 #include "llvm/IR/Type.h"
72 #include "llvm/IR/Use.h"
73 #include "llvm/IR/User.h"
74 #include "llvm/IR/Value.h"
75 #include "llvm/Pass.h"
76 #include "llvm/Support/Casting.h"
78 #include "llvm/Support/Compiler.h"
79 #include "llvm/Support/Debug.h"
83 #include "llvm/Transforms/Scalar.h"
85 #include <algorithm>
86 #include <cassert>
87 #include <chrono>
88 #include <cstddef>
89 #include <cstdint>
90 #include <cstring>
91 #include <iterator>
92 #include <string>
93 #include <tuple>
94 #include <utility>
95 #include <vector>
96 
97 #ifndef NDEBUG
98 // We only use this for a debug check.
99 #include <random>
100 #endif
101 
102 using namespace llvm;
103 using namespace llvm::sroa;
104 
105 #define DEBUG_TYPE "sroa"
106 
107 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
108 STATISTIC(NumAllocaPartitions, "Number of alloca partitions formed");
109 STATISTIC(MaxPartitionsPerAlloca, "Maximum number of partitions per alloca");
110 STATISTIC(NumAllocaPartitionUses, "Number of alloca partition uses rewritten");
111 STATISTIC(MaxUsesPerAllocaPartition, "Maximum number of uses of a partition");
112 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
113 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
114 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
115 STATISTIC(NumDeleted, "Number of instructions deleted");
116 STATISTIC(NumVectorized, "Number of vectorized aggregates");
117 
118 /// Hidden option to enable randomly shuffling the slices to help uncover
119 /// instability in their order.
120 static cl::opt<bool> SROARandomShuffleSlices("sroa-random-shuffle-slices",
121  cl::init(false), cl::Hidden);
122 
123 /// Hidden option to experiment with completely strict handling of inbounds
124 /// GEPs.
125 static cl::opt<bool> SROAStrictInbounds("sroa-strict-inbounds", cl::init(false),
126  cl::Hidden);
127 
128 namespace {
129 
130 /// A custom IRBuilder inserter which prefixes all names, but only in
131 /// Assert builds.
132 class IRBuilderPrefixedInserter : public IRBuilderDefaultInserter {
133  std::string Prefix;
134 
135  const Twine getNameWithPrefix(const Twine &Name) const {
136  return Name.isTriviallyEmpty() ? Name : Prefix + Name;
137  }
138 
139 public:
140  void SetNamePrefix(const Twine &P) { Prefix = P.str(); }
141 
142 protected:
143  void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
144  BasicBlock::iterator InsertPt) const {
145  IRBuilderDefaultInserter::InsertHelper(I, getNameWithPrefix(Name), BB,
146  InsertPt);
147  }
148 };
149 
150 /// Provide a type for IRBuilder that drops names in release builds.
152 
153 /// A used slice of an alloca.
154 ///
155 /// This structure represents a slice of an alloca used by some instruction. It
156 /// stores both the begin and end offsets of this use, a pointer to the use
157 /// itself, and a flag indicating whether we can classify the use as splittable
158 /// or not when forming partitions of the alloca.
159 class Slice {
160  /// The beginning offset of the range.
161  uint64_t BeginOffset = 0;
162 
163  /// The ending offset, not included in the range.
164  uint64_t EndOffset = 0;
165 
166  /// Storage for both the use of this slice and whether it can be
167  /// split.
168  PointerIntPair<Use *, 1, bool> UseAndIsSplittable;
169 
170 public:
171  Slice() = default;
172 
173  Slice(uint64_t BeginOffset, uint64_t EndOffset, Use *U, bool IsSplittable)
174  : BeginOffset(BeginOffset), EndOffset(EndOffset),
175  UseAndIsSplittable(U, IsSplittable) {}
176 
177  uint64_t beginOffset() const { return BeginOffset; }
178  uint64_t endOffset() const { return EndOffset; }
179 
180  bool isSplittable() const { return UseAndIsSplittable.getInt(); }
181  void makeUnsplittable() { UseAndIsSplittable.setInt(false); }
182 
183  Use *getUse() const { return UseAndIsSplittable.getPointer(); }
184 
185  bool isDead() const { return getUse() == nullptr; }
186  void kill() { UseAndIsSplittable.setPointer(nullptr); }
187 
188  /// Support for ordering ranges.
189  ///
190  /// This provides an ordering over ranges such that start offsets are
191  /// always increasing, and within equal start offsets, the end offsets are
192  /// decreasing. Thus the spanning range comes first in a cluster with the
193  /// same start position.
194  bool operator<(const Slice &RHS) const {
195  if (beginOffset() < RHS.beginOffset())
196  return true;
197  if (beginOffset() > RHS.beginOffset())
198  return false;
199  if (isSplittable() != RHS.isSplittable())
200  return !isSplittable();
201  if (endOffset() > RHS.endOffset())
202  return true;
203  return false;
204  }
205 
206  /// Support comparison with a single offset to allow binary searches.
207  friend LLVM_ATTRIBUTE_UNUSED bool operator<(const Slice &LHS,
208  uint64_t RHSOffset) {
209  return LHS.beginOffset() < RHSOffset;
210  }
211  friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
212  const Slice &RHS) {
213  return LHSOffset < RHS.beginOffset();
214  }
215 
216  bool operator==(const Slice &RHS) const {
217  return isSplittable() == RHS.isSplittable() &&
218  beginOffset() == RHS.beginOffset() && endOffset() == RHS.endOffset();
219  }
220  bool operator!=(const Slice &RHS) const { return !operator==(RHS); }
221 };
222 
223 } // end anonymous namespace
224 
225 namespace llvm {
226 
227 template <typename T> struct isPodLike;
228 template <> struct isPodLike<Slice> { static const bool value = true; };
229 
230 } // end namespace llvm
231 
232 /// Representation of the alloca slices.
233 ///
234 /// This class represents the slices of an alloca which are formed by its
235 /// various uses. If a pointer escapes, we can't fully build a representation
236 /// for the slices used and we reflect that in this structure. The uses are
237 /// stored, sorted by increasing beginning offset and with unsplittable slices
238 /// starting at a particular offset before splittable slices.
240 public:
241  /// Construct the slices of a particular alloca.
242  AllocaSlices(const DataLayout &DL, AllocaInst &AI);
243 
244  /// Test whether a pointer to the allocation escapes our analysis.
245  ///
246  /// If this is true, the slices are never fully built and should be
247  /// ignored.
248  bool isEscaped() const { return PointerEscapingInstr; }
249 
250  /// Support for iterating over the slices.
251  /// @{
254 
255  iterator begin() { return Slices.begin(); }
256  iterator end() { return Slices.end(); }
257 
260 
261  const_iterator begin() const { return Slices.begin(); }
262  const_iterator end() const { return Slices.end(); }
263  /// @}
264 
265  /// Erase a range of slices.
266  void erase(iterator Start, iterator Stop) { Slices.erase(Start, Stop); }
267 
268  /// Insert new slices for this alloca.
269  ///
270  /// This moves the slices into the alloca's slices collection, and re-sorts
271  /// everything so that the usual ordering properties of the alloca's slices
272  /// hold.
273  void insert(ArrayRef<Slice> NewSlices) {
274  int OldSize = Slices.size();
275  Slices.append(NewSlices.begin(), NewSlices.end());
276  auto SliceI = Slices.begin() + OldSize;
277  llvm::sort(SliceI, Slices.end());
278  std::inplace_merge(Slices.begin(), SliceI, Slices.end());
279  }
280 
281  // Forward declare the iterator and range accessor for walking the
282  // partitions.
283  class partition_iterator;
285 
286  /// Access the dead users for this alloca.
287  ArrayRef<Instruction *> getDeadUsers() const { return DeadUsers; }
288 
289  /// Access the dead operands referring to this alloca.
290  ///
291  /// These are operands which have cannot actually be used to refer to the
292  /// alloca as they are outside its range and the user doesn't correct for
293  /// that. These mostly consist of PHI node inputs and the like which we just
294  /// need to replace with undef.
295  ArrayRef<Use *> getDeadOperands() const { return DeadOperands; }
296 
297 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
298  void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
299  void printSlice(raw_ostream &OS, const_iterator I,
300  StringRef Indent = " ") const;
301  void printUse(raw_ostream &OS, const_iterator I,
302  StringRef Indent = " ") const;
303  void print(raw_ostream &OS) const;
304  void dump(const_iterator I) const;
305  void dump() const;
306 #endif
307 
308 private:
309  template <typename DerivedT, typename RetT = void> class BuilderBase;
310  class SliceBuilder;
311 
312  friend class AllocaSlices::SliceBuilder;
313 
314 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
315  /// Handle to alloca instruction to simplify method interfaces.
316  AllocaInst &AI;
317 #endif
318 
319  /// The instruction responsible for this alloca not having a known set
320  /// of slices.
321  ///
322  /// When an instruction (potentially) escapes the pointer to the alloca, we
323  /// store a pointer to that here and abort trying to form slices of the
324  /// alloca. This will be null if the alloca slices are analyzed successfully.
325  Instruction *PointerEscapingInstr;
326 
327  /// The slices of the alloca.
328  ///
329  /// We store a vector of the slices formed by uses of the alloca here. This
330  /// vector is sorted by increasing begin offset, and then the unsplittable
331  /// slices before the splittable ones. See the Slice inner class for more
332  /// details.
333  SmallVector<Slice, 8> Slices;
334 
335  /// Instructions which will become dead if we rewrite the alloca.
336  ///
337  /// Note that these are not separated by slice. This is because we expect an
338  /// alloca to be completely rewritten or not rewritten at all. If rewritten,
339  /// all these instructions can simply be removed and replaced with undef as
340  /// they come from outside of the allocated space.
342 
343  /// Operands which will become dead if we rewrite the alloca.
344  ///
345  /// These are operands that in their particular use can be replaced with
346  /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
347  /// to PHI nodes and the like. They aren't entirely dead (there might be
348  /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
349  /// want to swap this particular input for undef to simplify the use lists of
350  /// the alloca.
351  SmallVector<Use *, 8> DeadOperands;
352 };
353 
354 /// A partition of the slices.
355 ///
356 /// An ephemeral representation for a range of slices which can be viewed as
357 /// a partition of the alloca. This range represents a span of the alloca's
358 /// memory which cannot be split, and provides access to all of the slices
359 /// overlapping some part of the partition.
360 ///
361 /// Objects of this type are produced by traversing the alloca's slices, but
362 /// are only ephemeral and not persistent.
364 private:
365  friend class AllocaSlices;
367 
368  using iterator = AllocaSlices::iterator;
369 
370  /// The beginning and ending offsets of the alloca for this
371  /// partition.
372  uint64_t BeginOffset, EndOffset;
373 
374  /// The start and end iterators of this partition.
375  iterator SI, SJ;
376 
377  /// A collection of split slice tails overlapping the partition.
378  SmallVector<Slice *, 4> SplitTails;
379 
380  /// Raw constructor builds an empty partition starting and ending at
381  /// the given iterator.
382  Partition(iterator SI) : SI(SI), SJ(SI) {}
383 
384 public:
385  /// The start offset of this partition.
386  ///
387  /// All of the contained slices start at or after this offset.
388  uint64_t beginOffset() const { return BeginOffset; }
389 
390  /// The end offset of this partition.
391  ///
392  /// All of the contained slices end at or before this offset.
393  uint64_t endOffset() const { return EndOffset; }
394 
395  /// The size of the partition.
396  ///
397  /// Note that this can never be zero.
398  uint64_t size() const {
399  assert(BeginOffset < EndOffset && "Partitions must span some bytes!");
400  return EndOffset - BeginOffset;
401  }
402 
403  /// Test whether this partition contains no slices, and merely spans
404  /// a region occupied by split slices.
405  bool empty() const { return SI == SJ; }
406 
407  /// \name Iterate slices that start within the partition.
408  /// These may be splittable or unsplittable. They have a begin offset >= the
409  /// partition begin offset.
410  /// @{
411  // FIXME: We should probably define a "concat_iterator" helper and use that
412  // to stitch together pointee_iterators over the split tails and the
413  // contiguous iterators of the partition. That would give a much nicer
414  // interface here. We could then additionally expose filtered iterators for
415  // split, unsplit, and unsplittable splices based on the usage patterns.
416  iterator begin() const { return SI; }
417  iterator end() const { return SJ; }
418  /// @}
419 
420  /// Get the sequence of split slice tails.
421  ///
422  /// These tails are of slices which start before this partition but are
423  /// split and overlap into the partition. We accumulate these while forming
424  /// partitions.
425  ArrayRef<Slice *> splitSliceTails() const { return SplitTails; }
426 };
427 
428 /// An iterator over partitions of the alloca's slices.
429 ///
430 /// This iterator implements the core algorithm for partitioning the alloca's
431 /// slices. It is a forward iterator as we don't support backtracking for
432 /// efficiency reasons, and re-use a single storage area to maintain the
433 /// current set of split slices.
434 ///
435 /// It is templated on the slice iterator type to use so that it can operate
436 /// with either const or non-const slice iterators.
438  : public iterator_facade_base<partition_iterator, std::forward_iterator_tag,
439  Partition> {
440  friend class AllocaSlices;
441 
442  /// Most of the state for walking the partitions is held in a class
443  /// with a nice interface for examining them.
444  Partition P;
445 
446  /// We need to keep the end of the slices to know when to stop.
448 
449  /// We also need to keep track of the maximum split end offset seen.
450  /// FIXME: Do we really?
451  uint64_t MaxSplitSliceEndOffset = 0;
452 
453  /// Sets the partition to be empty at given iterator, and sets the
454  /// end iterator.
456  : P(SI), SE(SE) {
457  // If not already at the end, advance our state to form the initial
458  // partition.
459  if (SI != SE)
460  advance();
461  }
462 
463  /// Advance the iterator to the next partition.
464  ///
465  /// Requires that the iterator not be at the end of the slices.
466  void advance() {
467  assert((P.SI != SE || !P.SplitTails.empty()) &&
468  "Cannot advance past the end of the slices!");
469 
470  // Clear out any split uses which have ended.
471  if (!P.SplitTails.empty()) {
472  if (P.EndOffset >= MaxSplitSliceEndOffset) {
473  // If we've finished all splits, this is easy.
474  P.SplitTails.clear();
475  MaxSplitSliceEndOffset = 0;
476  } else {
477  // Remove the uses which have ended in the prior partition. This
478  // cannot change the max split slice end because we just checked that
479  // the prior partition ended prior to that max.
480  P.SplitTails.erase(llvm::remove_if(P.SplitTails,
481  [&](Slice *S) {
482  return S->endOffset() <=
483  P.EndOffset;
484  }),
485  P.SplitTails.end());
486  assert(llvm::any_of(P.SplitTails,
487  [&](Slice *S) {
488  return S->endOffset() == MaxSplitSliceEndOffset;
489  }) &&
490  "Could not find the current max split slice offset!");
491  assert(llvm::all_of(P.SplitTails,
492  [&](Slice *S) {
493  return S->endOffset() <= MaxSplitSliceEndOffset;
494  }) &&
495  "Max split slice end offset is not actually the max!");
496  }
497  }
498 
499  // If P.SI is already at the end, then we've cleared the split tail and
500  // now have an end iterator.
501  if (P.SI == SE) {
502  assert(P.SplitTails.empty() && "Failed to clear the split slices!");
503  return;
504  }
505 
506  // If we had a non-empty partition previously, set up the state for
507  // subsequent partitions.
508  if (P.SI != P.SJ) {
509  // Accumulate all the splittable slices which started in the old
510  // partition into the split list.
511  for (Slice &S : P)
512  if (S.isSplittable() && S.endOffset() > P.EndOffset) {
513  P.SplitTails.push_back(&S);
514  MaxSplitSliceEndOffset =
515  std::max(S.endOffset(), MaxSplitSliceEndOffset);
516  }
517 
518  // Start from the end of the previous partition.
519  P.SI = P.SJ;
520 
521  // If P.SI is now at the end, we at most have a tail of split slices.
522  if (P.SI == SE) {
523  P.BeginOffset = P.EndOffset;
524  P.EndOffset = MaxSplitSliceEndOffset;
525  return;
526  }
527 
528  // If the we have split slices and the next slice is after a gap and is
529  // not splittable immediately form an empty partition for the split
530  // slices up until the next slice begins.
531  if (!P.SplitTails.empty() && P.SI->beginOffset() != P.EndOffset &&
532  !P.SI->isSplittable()) {
533  P.BeginOffset = P.EndOffset;
534  P.EndOffset = P.SI->beginOffset();
535  return;
536  }
537  }
538 
539  // OK, we need to consume new slices. Set the end offset based on the
540  // current slice, and step SJ past it. The beginning offset of the
541  // partition is the beginning offset of the next slice unless we have
542  // pre-existing split slices that are continuing, in which case we begin
543  // at the prior end offset.
544  P.BeginOffset = P.SplitTails.empty() ? P.SI->beginOffset() : P.EndOffset;
545  P.EndOffset = P.SI->endOffset();
546  ++P.SJ;
547 
548  // There are two strategies to form a partition based on whether the
549  // partition starts with an unsplittable slice or a splittable slice.
550  if (!P.SI->isSplittable()) {
551  // When we're forming an unsplittable region, it must always start at
552  // the first slice and will extend through its end.
553  assert(P.BeginOffset == P.SI->beginOffset());
554 
555  // Form a partition including all of the overlapping slices with this
556  // unsplittable slice.
557  while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
558  if (!P.SJ->isSplittable())
559  P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
560  ++P.SJ;
561  }
562 
563  // We have a partition across a set of overlapping unsplittable
564  // partitions.
565  return;
566  }
567 
568  // If we're starting with a splittable slice, then we need to form
569  // a synthetic partition spanning it and any other overlapping splittable
570  // splices.
571  assert(P.SI->isSplittable() && "Forming a splittable partition!");
572 
573  // Collect all of the overlapping splittable slices.
574  while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset &&
575  P.SJ->isSplittable()) {
576  P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
577  ++P.SJ;
578  }
579 
580  // Back upiP.EndOffset if we ended the span early when encountering an
581  // unsplittable slice. This synthesizes the early end offset of
582  // a partition spanning only splittable slices.
583  if (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
584  assert(!P.SJ->isSplittable());
585  P.EndOffset = P.SJ->beginOffset();
586  }
587  }
588 
589 public:
590  bool operator==(const partition_iterator &RHS) const {
591  assert(SE == RHS.SE &&
592  "End iterators don't match between compared partition iterators!");
593 
594  // The observed positions of partitions is marked by the P.SI iterator and
595  // the emptiness of the split slices. The latter is only relevant when
596  // P.SI == SE, as the end iterator will additionally have an empty split
597  // slices list, but the prior may have the same P.SI and a tail of split
598  // slices.
599  if (P.SI == RHS.P.SI && P.SplitTails.empty() == RHS.P.SplitTails.empty()) {
600  assert(P.SJ == RHS.P.SJ &&
601  "Same set of slices formed two different sized partitions!");
602  assert(P.SplitTails.size() == RHS.P.SplitTails.size() &&
603  "Same slice position with differently sized non-empty split "
604  "slice tails!");
605  return true;
606  }
607  return false;
608  }
609 
611  advance();
612  return *this;
613  }
614 
615  Partition &operator*() { return P; }
616 };
617 
618 /// A forward range over the partitions of the alloca's slices.
619 ///
620 /// This accesses an iterator range over the partitions of the alloca's
621 /// slices. It computes these partitions on the fly based on the overlapping
622 /// offsets of the slices and the ability to split them. It will visit "empty"
623 /// partitions to cover regions of the alloca only accessed via split
624 /// slices.
626  return make_range(partition_iterator(begin(), end()),
627  partition_iterator(end(), end()));
628 }
629 
631  // If the condition being selected on is a constant or the same value is
632  // being selected between, fold the select. Yes this does (rarely) happen
633  // early on.
634  if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
635  return SI.getOperand(1 + CI->isZero());
636  if (SI.getOperand(1) == SI.getOperand(2))
637  return SI.getOperand(1);
638 
639  return nullptr;
640 }
641 
642 /// A helper that folds a PHI node or a select.
644  if (PHINode *PN = dyn_cast<PHINode>(&I)) {
645  // If PN merges together the same value, return that value.
646  return PN->hasConstantValue();
647  }
648  return foldSelectInst(cast<SelectInst>(I));
649 }
650 
651 /// Builder for the alloca slices.
652 ///
653 /// This class builds a set of alloca slices by recursively visiting the uses
654 /// of an alloca and making a slice for each load and store at each offset.
655 class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
657  friend class InstVisitor<SliceBuilder>;
658 
660 
661  const uint64_t AllocSize;
662  AllocaSlices &AS;
663 
664  SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
666 
667  /// Set to de-duplicate dead instructions found in the use walk.
668  SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
669 
670 public:
673  AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())), AS(AS) {}
674 
675 private:
676  void markAsDead(Instruction &I) {
677  if (VisitedDeadInsts.insert(&I).second)
678  AS.DeadUsers.push_back(&I);
679  }
680 
681  void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
682  bool IsSplittable = false) {
683  // Completely skip uses which have a zero size or start either before or
684  // past the end of the allocation.
685  if (Size == 0 || Offset.uge(AllocSize)) {
686  LLVM_DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @"
687  << Offset
688  << " which has zero size or starts outside of the "
689  << AllocSize << " byte alloca:\n"
690  << " alloca: " << AS.AI << "\n"
691  << " use: " << I << "\n");
692  return markAsDead(I);
693  }
694 
695  uint64_t BeginOffset = Offset.getZExtValue();
696  uint64_t EndOffset = BeginOffset + Size;
697 
698  // Clamp the end offset to the end of the allocation. Note that this is
699  // formulated to handle even the case where "BeginOffset + Size" overflows.
700  // This may appear superficially to be something we could ignore entirely,
701  // but that is not so! There may be widened loads or PHI-node uses where
702  // some instructions are dead but not others. We can't completely ignore
703  // them, and so have to record at least the information here.
704  assert(AllocSize >= BeginOffset); // Established above.
705  if (Size > AllocSize - BeginOffset) {
706  LLVM_DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @"
707  << Offset << " to remain within the " << AllocSize
708  << " byte alloca:\n"
709  << " alloca: " << AS.AI << "\n"
710  << " use: " << I << "\n");
711  EndOffset = AllocSize;
712  }
713 
714  AS.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
715  }
716 
717  void visitBitCastInst(BitCastInst &BC) {
718  if (BC.use_empty())
719  return markAsDead(BC);
720 
721  return Base::visitBitCastInst(BC);
722  }
723 
724  void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
725  if (GEPI.use_empty())
726  return markAsDead(GEPI);
727 
728  if (SROAStrictInbounds && GEPI.isInBounds()) {
729  // FIXME: This is a manually un-factored variant of the basic code inside
730  // of GEPs with checking of the inbounds invariant specified in the
731  // langref in a very strict sense. If we ever want to enable
732  // SROAStrictInbounds, this code should be factored cleanly into
733  // PtrUseVisitor, but it is easier to experiment with SROAStrictInbounds
734  // by writing out the code here where we have the underlying allocation
735  // size readily available.
736  APInt GEPOffset = Offset;
737  const DataLayout &DL = GEPI.getModule()->getDataLayout();
738  for (gep_type_iterator GTI = gep_type_begin(GEPI),
739  GTE = gep_type_end(GEPI);
740  GTI != GTE; ++GTI) {
741  ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
742  if (!OpC)
743  break;
744 
745  // Handle a struct index, which adds its field offset to the pointer.
746  if (StructType *STy = GTI.getStructTypeOrNull()) {
747  unsigned ElementIdx = OpC->getZExtValue();
748  const StructLayout *SL = DL.getStructLayout(STy);
749  GEPOffset +=
750  APInt(Offset.getBitWidth(), SL->getElementOffset(ElementIdx));
751  } else {
752  // For array or vector indices, scale the index by the size of the
753  // type.
754  APInt Index = OpC->getValue().sextOrTrunc(Offset.getBitWidth());
755  GEPOffset += Index * APInt(Offset.getBitWidth(),
756  DL.getTypeAllocSize(GTI.getIndexedType()));
757  }
758 
759  // If this index has computed an intermediate pointer which is not
760  // inbounds, then the result of the GEP is a poison value and we can
761  // delete it and all uses.
762  if (GEPOffset.ugt(AllocSize))
763  return markAsDead(GEPI);
764  }
765  }
766 
767  return Base::visitGetElementPtrInst(GEPI);
768  }
769 
770  void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
771  uint64_t Size, bool IsVolatile) {
772  // We allow splitting of non-volatile loads and stores where the type is an
773  // integer type. These may be used to implement 'memcpy' or other "transfer
774  // of bits" patterns.
775  bool IsSplittable = Ty->isIntegerTy() && !IsVolatile;
776 
777  insertUse(I, Offset, Size, IsSplittable);
778  }
779 
780  void visitLoadInst(LoadInst &LI) {
781  assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
782  "All simple FCA loads should have been pre-split");
783 
784  if (!IsOffsetKnown)
785  return PI.setAborted(&LI);
786 
787  const DataLayout &DL = LI.getModule()->getDataLayout();
788  uint64_t Size = DL.getTypeStoreSize(LI.getType());
789  return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
790  }
791 
792  void visitStoreInst(StoreInst &SI) {
793  Value *ValOp = SI.getValueOperand();
794  if (ValOp == *U)
795  return PI.setEscapedAndAborted(&SI);
796  if (!IsOffsetKnown)
797  return PI.setAborted(&SI);
798 
799  const DataLayout &DL = SI.getModule()->getDataLayout();
800  uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
801 
802  // If this memory access can be shown to *statically* extend outside the
803  // bounds of the allocation, it's behavior is undefined, so simply
804  // ignore it. Note that this is more strict than the generic clamping
805  // behavior of insertUse. We also try to handle cases which might run the
806  // risk of overflow.
807  // FIXME: We should instead consider the pointer to have escaped if this
808  // function is being instrumented for addressing bugs or race conditions.
809  if (Size > AllocSize || Offset.ugt(AllocSize - Size)) {
810  LLVM_DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @"
811  << Offset << " which extends past the end of the "
812  << AllocSize << " byte alloca:\n"
813  << " alloca: " << AS.AI << "\n"
814  << " use: " << SI << "\n");
815  return markAsDead(SI);
816  }
817 
818  assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
819  "All simple FCA stores should have been pre-split");
820  handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
821  }
822 
823  void visitMemSetInst(MemSetInst &II) {
824  assert(II.getRawDest() == *U && "Pointer use is not the destination?");
825  ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
826  if ((Length && Length->getValue() == 0) ||
827  (IsOffsetKnown && Offset.uge(AllocSize)))
828  // Zero-length mem transfer intrinsics can be ignored entirely.
829  return markAsDead(II);
830 
831  if (!IsOffsetKnown)
832  return PI.setAborted(&II);
833 
834  insertUse(II, Offset, Length ? Length->getLimitedValue()
835  : AllocSize - Offset.getLimitedValue(),
836  (bool)Length);
837  }
838 
839  void visitMemTransferInst(MemTransferInst &II) {
840  ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
841  if (Length && Length->getValue() == 0)
842  // Zero-length mem transfer intrinsics can be ignored entirely.
843  return markAsDead(II);
844 
845  // Because we can visit these intrinsics twice, also check to see if the
846  // first time marked this instruction as dead. If so, skip it.
847  if (VisitedDeadInsts.count(&II))
848  return;
849 
850  if (!IsOffsetKnown)
851  return PI.setAborted(&II);
852 
853  // This side of the transfer is completely out-of-bounds, and so we can
854  // nuke the entire transfer. However, we also need to nuke the other side
855  // if already added to our partitions.
856  // FIXME: Yet another place we really should bypass this when
857  // instrumenting for ASan.
858  if (Offset.uge(AllocSize)) {
860  MemTransferSliceMap.find(&II);
861  if (MTPI != MemTransferSliceMap.end())
862  AS.Slices[MTPI->second].kill();
863  return markAsDead(II);
864  }
865 
866  uint64_t RawOffset = Offset.getLimitedValue();
867  uint64_t Size = Length ? Length->getLimitedValue() : AllocSize - RawOffset;
868 
869  // Check for the special case where the same exact value is used for both
870  // source and dest.
871  if (*U == II.getRawDest() && *U == II.getRawSource()) {
872  // For non-volatile transfers this is a no-op.
873  if (!II.isVolatile())
874  return markAsDead(II);
875 
876  return insertUse(II, Offset, Size, /*IsSplittable=*/false);
877  }
878 
879  // If we have seen both source and destination for a mem transfer, then
880  // they both point to the same alloca.
881  bool Inserted;
883  std::tie(MTPI, Inserted) =
884  MemTransferSliceMap.insert(std::make_pair(&II, AS.Slices.size()));
885  unsigned PrevIdx = MTPI->second;
886  if (!Inserted) {
887  Slice &PrevP = AS.Slices[PrevIdx];
888 
889  // Check if the begin offsets match and this is a non-volatile transfer.
890  // In that case, we can completely elide the transfer.
891  if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
892  PrevP.kill();
893  return markAsDead(II);
894  }
895 
896  // Otherwise we have an offset transfer within the same alloca. We can't
897  // split those.
898  PrevP.makeUnsplittable();
899  }
900 
901  // Insert the use now that we've fixed up the splittable nature.
902  insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
903 
904  // Check that we ended up with a valid index in the map.
905  assert(AS.Slices[PrevIdx].getUse()->getUser() == &II &&
906  "Map index doesn't point back to a slice with this user.");
907  }
908 
909  // Disable SRoA for any intrinsics except for lifetime invariants.
910  // FIXME: What about debug intrinsics? This matches old behavior, but
911  // doesn't make sense.
912  void visitIntrinsicInst(IntrinsicInst &II) {
913  if (!IsOffsetKnown)
914  return PI.setAborted(&II);
915 
916  if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
917  II.getIntrinsicID() == Intrinsic::lifetime_end) {
918  ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
919  uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
920  Length->getLimitedValue());
921  insertUse(II, Offset, Size, true);
922  return;
923  }
924 
925  Base::visitIntrinsicInst(II);
926  }
927 
928  Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
929  // We consider any PHI or select that results in a direct load or store of
930  // the same offset to be a viable use for slicing purposes. These uses
931  // are considered unsplittable and the size is the maximum loaded or stored
932  // size.
935  Visited.insert(Root);
936  Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
937  const DataLayout &DL = Root->getModule()->getDataLayout();
938  // If there are no loads or stores, the access is dead. We mark that as
939  // a size zero access.
940  Size = 0;
941  do {
942  Instruction *I, *UsedI;
943  std::tie(UsedI, I) = Uses.pop_back_val();
944 
945  if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
946  Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
947  continue;
948  }
949  if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
950  Value *Op = SI->getOperand(0);
951  if (Op == UsedI)
952  return SI;
953  Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
954  continue;
955  }
956 
957  if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
958  if (!GEP->hasAllZeroIndices())
959  return GEP;
960  } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
961  !isa<SelectInst>(I)) {
962  return I;
963  }
964 
965  for (User *U : I->users())
966  if (Visited.insert(cast<Instruction>(U)).second)
967  Uses.push_back(std::make_pair(I, cast<Instruction>(U)));
968  } while (!Uses.empty());
969 
970  return nullptr;
971  }
972 
973  void visitPHINodeOrSelectInst(Instruction &I) {
974  assert(isa<PHINode>(I) || isa<SelectInst>(I));
975  if (I.use_empty())
976  return markAsDead(I);
977 
978  // TODO: We could use SimplifyInstruction here to fold PHINodes and
979  // SelectInsts. However, doing so requires to change the current
980  // dead-operand-tracking mechanism. For instance, suppose neither loading
981  // from %U nor %other traps. Then "load (select undef, %U, %other)" does not
982  // trap either. However, if we simply replace %U with undef using the
983  // current dead-operand-tracking mechanism, "load (select undef, undef,
984  // %other)" may trap because the select may return the first operand
985  // "undef".
987  if (Result == *U)
988  // If the result of the constant fold will be the pointer, recurse
989  // through the PHI/select as if we had RAUW'ed it.
990  enqueueUsers(I);
991  else
992  // Otherwise the operand to the PHI/select is dead, and we can replace
993  // it with undef.
994  AS.DeadOperands.push_back(U);
995 
996  return;
997  }
998 
999  if (!IsOffsetKnown)
1000  return PI.setAborted(&I);
1001 
1002  // See if we already have computed info on this node.
1003  uint64_t &Size = PHIOrSelectSizes[&I];
1004  if (!Size) {
1005  // This is a new PHI/Select, check for an unsafe use of it.
1006  if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&I, Size))
1007  return PI.setAborted(UnsafeI);
1008  }
1009 
1010  // For PHI and select operands outside the alloca, we can't nuke the entire
1011  // phi or select -- the other side might still be relevant, so we special
1012  // case them here and use a separate structure to track the operands
1013  // themselves which should be replaced with undef.
1014  // FIXME: This should instead be escaped in the event we're instrumenting
1015  // for address sanitization.
1016  if (Offset.uge(AllocSize)) {
1017  AS.DeadOperands.push_back(U);
1018  return;
1019  }
1020 
1021  insertUse(I, Offset, Size);
1022  }
1023 
1024  void visitPHINode(PHINode &PN) { visitPHINodeOrSelectInst(PN); }
1025 
1026  void visitSelectInst(SelectInst &SI) { visitPHINodeOrSelectInst(SI); }
1027 
1028  /// Disable SROA entirely if there are unhandled users of the alloca.
1029  void visitInstruction(Instruction &I) { PI.setAborted(&I); }
1030 };
1031 
1033  :
1034 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1035  AI(AI),
1036 #endif
1037  PointerEscapingInstr(nullptr) {
1038  SliceBuilder PB(DL, AI, *this);
1039  SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
1040  if (PtrI.isEscaped() || PtrI.isAborted()) {
1041  // FIXME: We should sink the escape vs. abort info into the caller nicely,
1042  // possibly by just storing the PtrInfo in the AllocaSlices.
1043  PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
1044  : PtrI.getAbortingInst();
1045  assert(PointerEscapingInstr && "Did not track a bad instruction");
1046  return;
1047  }
1048 
1049  Slices.erase(
1050  llvm::remove_if(Slices, [](const Slice &S) { return S.isDead(); }),
1051  Slices.end());
1052 
1053 #ifndef NDEBUG
1055  std::mt19937 MT(static_cast<unsigned>(
1056  std::chrono::system_clock::now().time_since_epoch().count()));
1057  std::shuffle(Slices.begin(), Slices.end(), MT);
1058  }
1059 #endif
1060 
1061  // Sort the uses. This arranges for the offsets to be in ascending order,
1062  // and the sizes to be in descending order.
1063  llvm::sort(Slices.begin(), Slices.end());
1064 }
1065 
1066 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1067 
1068 void AllocaSlices::print(raw_ostream &OS, const_iterator I,
1069  StringRef Indent) const {
1070  printSlice(OS, I, Indent);
1071  OS << "\n";
1072  printUse(OS, I, Indent);
1073 }
1074 
1075 void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
1076  StringRef Indent) const {
1077  OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
1078  << " slice #" << (I - begin())
1079  << (I->isSplittable() ? " (splittable)" : "");
1080 }
1081 
1082 void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
1083  StringRef Indent) const {
1084  OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
1085 }
1086 
1087 void AllocaSlices::print(raw_ostream &OS) const {
1088  if (PointerEscapingInstr) {
1089  OS << "Can't analyze slices for alloca: " << AI << "\n"
1090  << " A pointer to this alloca escaped by:\n"
1091  << " " << *PointerEscapingInstr << "\n";
1092  return;
1093  }
1094 
1095  OS << "Slices of alloca: " << AI << "\n";
1096  for (const_iterator I = begin(), E = end(); I != E; ++I)
1097  print(OS, I);
1098 }
1099 
1100 LLVM_DUMP_METHOD void AllocaSlices::dump(const_iterator I) const {
1101  print(dbgs(), I);
1102 }
1103 LLVM_DUMP_METHOD void AllocaSlices::dump() const { print(dbgs()); }
1104 
1105 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1106 
1107 /// Walk the range of a partitioning looking for a common type to cover this
1108 /// sequence of slices.
1111  uint64_t EndOffset) {
1112  Type *Ty = nullptr;
1113  bool TyIsCommon = true;
1114  IntegerType *ITy = nullptr;
1115 
1116  // Note that we need to look at *every* alloca slice's Use to ensure we
1117  // always get consistent results regardless of the order of slices.
1118  for (AllocaSlices::const_iterator I = B; I != E; ++I) {
1119  Use *U = I->getUse();
1120  if (isa<IntrinsicInst>(*U->getUser()))
1121  continue;
1122  if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
1123  continue;
1124 
1125  Type *UserTy = nullptr;
1126  if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1127  UserTy = LI->getType();
1128  } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1129  UserTy = SI->getValueOperand()->getType();
1130  }
1131 
1132  if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
1133  // If the type is larger than the partition, skip it. We only encounter
1134  // this for split integer operations where we want to use the type of the
1135  // entity causing the split. Also skip if the type is not a byte width
1136  // multiple.
1137  if (UserITy->getBitWidth() % 8 != 0 ||
1138  UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
1139  continue;
1140 
1141  // Track the largest bitwidth integer type used in this way in case there
1142  // is no common type.
1143  if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
1144  ITy = UserITy;
1145  }
1146 
1147  // To avoid depending on the order of slices, Ty and TyIsCommon must not
1148  // depend on types skipped above.
1149  if (!UserTy || (Ty && Ty != UserTy))
1150  TyIsCommon = false; // Give up on anything but an iN type.
1151  else
1152  Ty = UserTy;
1153  }
1154 
1155  return TyIsCommon ? Ty : ITy;
1156 }
1157 
1158 /// PHI instructions that use an alloca and are subsequently loaded can be
1159 /// rewritten to load both input pointers in the pred blocks and then PHI the
1160 /// results, allowing the load of the alloca to be promoted.
1161 /// From this:
1162 /// %P2 = phi [i32* %Alloca, i32* %Other]
1163 /// %V = load i32* %P2
1164 /// to:
1165 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1166 /// ...
1167 /// %V2 = load i32* %Other
1168 /// ...
1169 /// %V = phi [i32 %V1, i32 %V2]
1170 ///
1171 /// We can do this to a select if its only uses are loads and if the operands
1172 /// to the select can be loaded unconditionally.
1173 ///
1174 /// FIXME: This should be hoisted into a generic utility, likely in
1175 /// Transforms/Util/Local.h
1176 static bool isSafePHIToSpeculate(PHINode &PN) {
1177  // For now, we can only do this promotion if the load is in the same block
1178  // as the PHI, and if there are no stores between the phi and load.
1179  // TODO: Allow recursive phi users.
1180  // TODO: Allow stores.
1181  BasicBlock *BB = PN.getParent();
1182  unsigned MaxAlign = 0;
1183  bool HaveLoad = false;
1184  for (User *U : PN.users()) {
1185  LoadInst *LI = dyn_cast<LoadInst>(U);
1186  if (!LI || !LI->isSimple())
1187  return false;
1188 
1189  // For now we only allow loads in the same block as the PHI. This is
1190  // a common case that happens when instcombine merges two loads through
1191  // a PHI.
1192  if (LI->getParent() != BB)
1193  return false;
1194 
1195  // Ensure that there are no instructions between the PHI and the load that
1196  // could store.
1197  for (BasicBlock::iterator BBI(PN); &*BBI != LI; ++BBI)
1198  if (BBI->mayWriteToMemory())
1199  return false;
1200 
1201  MaxAlign = std::max(MaxAlign, LI->getAlignment());
1202  HaveLoad = true;
1203  }
1204 
1205  if (!HaveLoad)
1206  return false;
1207 
1208  const DataLayout &DL = PN.getModule()->getDataLayout();
1209 
1210  // We can only transform this if it is safe to push the loads into the
1211  // predecessor blocks. The only thing to watch out for is that we can't put
1212  // a possibly trapping load in the predecessor if it is a critical edge.
1213  for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1215  Value *InVal = PN.getIncomingValue(Idx);
1216 
1217  // If the value is produced by the terminator of the predecessor (an
1218  // invoke) or it has side-effects, there is no valid place to put a load
1219  // in the predecessor.
1220  if (TI == InVal || TI->mayHaveSideEffects())
1221  return false;
1222 
1223  // If the predecessor has a single successor, then the edge isn't
1224  // critical.
1225  if (TI->getNumSuccessors() == 1)
1226  continue;
1227 
1228  // If this pointer is always safe to load, or if we can prove that there
1229  // is already a load in the block, then we can move the load to the pred
1230  // block.
1231  if (isSafeToLoadUnconditionally(InVal, MaxAlign, DL, TI))
1232  continue;
1233 
1234  return false;
1235  }
1236 
1237  return true;
1238 }
1239 
1240 static void speculatePHINodeLoads(PHINode &PN) {
1241  LLVM_DEBUG(dbgs() << " original: " << PN << "\n");
1242 
1243  Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1244  IRBuilderTy PHIBuilder(&PN);
1245  PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1246  PN.getName() + ".sroa.speculated");
1247 
1248  // Get the AA tags and alignment to use from one of the loads. It doesn't
1249  // matter which one we get and if any differ.
1250  LoadInst *SomeLoad = cast<LoadInst>(PN.user_back());
1251 
1252  AAMDNodes AATags;
1253  SomeLoad->getAAMetadata(AATags);
1254  unsigned Align = SomeLoad->getAlignment();
1255 
1256  // Rewrite all loads of the PN to use the new PHI.
1257  while (!PN.use_empty()) {
1258  LoadInst *LI = cast<LoadInst>(PN.user_back());
1259  LI->replaceAllUsesWith(NewPN);
1260  LI->eraseFromParent();
1261  }
1262 
1263  // Inject loads into all of the pred blocks.
1264  DenseMap<BasicBlock*, Value*> InjectedLoads;
1265  for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1266  BasicBlock *Pred = PN.getIncomingBlock(Idx);
1267  Value *InVal = PN.getIncomingValue(Idx);
1268 
1269  // A PHI node is allowed to have multiple (duplicated) entries for the same
1270  // basic block, as long as the value is the same. So if we already injected
1271  // a load in the predecessor, then we should reuse the same load for all
1272  // duplicated entries.
1273  if (Value* V = InjectedLoads.lookup(Pred)) {
1274  NewPN->addIncoming(V, Pred);
1275  continue;
1276  }
1277 
1278  TerminatorInst *TI = Pred->getTerminator();
1279  IRBuilderTy PredBuilder(TI);
1280 
1281  LoadInst *Load = PredBuilder.CreateLoad(
1282  InVal, (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1283  ++NumLoadsSpeculated;
1284  Load->setAlignment(Align);
1285  if (AATags)
1286  Load->setAAMetadata(AATags);
1287  NewPN->addIncoming(Load, Pred);
1288  InjectedLoads[Pred] = Load;
1289  }
1290 
1291  LLVM_DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1292  PN.eraseFromParent();
1293 }
1294 
1295 /// Select instructions that use an alloca and are subsequently loaded can be
1296 /// rewritten to load both input pointers and then select between the result,
1297 /// allowing the load of the alloca to be promoted.
1298 /// From this:
1299 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1300 /// %V = load i32* %P2
1301 /// to:
1302 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1303 /// %V2 = load i32* %Other
1304 /// %V = select i1 %cond, i32 %V1, i32 %V2
1305 ///
1306 /// We can do this to a select if its only uses are loads and if the operand
1307 /// to the select can be loaded unconditionally.
1309  Value *TValue = SI.getTrueValue();
1310  Value *FValue = SI.getFalseValue();
1311  const DataLayout &DL = SI.getModule()->getDataLayout();
1312 
1313  for (User *U : SI.users()) {
1314  LoadInst *LI = dyn_cast<LoadInst>(U);
1315  if (!LI || !LI->isSimple())
1316  return false;
1317 
1318  // Both operands to the select need to be dereferenceable, either
1319  // absolutely (e.g. allocas) or at this point because we can see other
1320  // accesses to it.
1321  if (!isSafeToLoadUnconditionally(TValue, LI->getAlignment(), DL, LI))
1322  return false;
1323  if (!isSafeToLoadUnconditionally(FValue, LI->getAlignment(), DL, LI))
1324  return false;
1325  }
1326 
1327  return true;
1328 }
1329 
1331  LLVM_DEBUG(dbgs() << " original: " << SI << "\n");
1332 
1333  IRBuilderTy IRB(&SI);
1334  Value *TV = SI.getTrueValue();
1335  Value *FV = SI.getFalseValue();
1336  // Replace the loads of the select with a select of two loads.
1337  while (!SI.use_empty()) {
1338  LoadInst *LI = cast<LoadInst>(SI.user_back());
1339  assert(LI->isSimple() && "We only speculate simple loads");
1340 
1341  IRB.SetInsertPoint(LI);
1342  LoadInst *TL =
1343  IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1344  LoadInst *FL =
1345  IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1346  NumLoadsSpeculated += 2;
1347 
1348  // Transfer alignment and AA info if present.
1349  TL->setAlignment(LI->getAlignment());
1350  FL->setAlignment(LI->getAlignment());
1351 
1352  AAMDNodes Tags;
1353  LI->getAAMetadata(Tags);
1354  if (Tags) {
1355  TL->setAAMetadata(Tags);
1356  FL->setAAMetadata(Tags);
1357  }
1358 
1359  Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1360  LI->getName() + ".sroa.speculated");
1361 
1362  LLVM_DEBUG(dbgs() << " speculated to: " << *V << "\n");
1363  LI->replaceAllUsesWith(V);
1364  LI->eraseFromParent();
1365  }
1366  SI.eraseFromParent();
1367 }
1368 
1369 /// Build a GEP out of a base pointer and indices.
1370 ///
1371 /// This will return the BasePtr if that is valid, or build a new GEP
1372 /// instruction using the IRBuilder if GEP-ing is needed.
1373 static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1374  SmallVectorImpl<Value *> &Indices, Twine NamePrefix) {
1375  if (Indices.empty())
1376  return BasePtr;
1377 
1378  // A single zero index is a no-op, so check for this and avoid building a GEP
1379  // in that case.
1380  if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1381  return BasePtr;
1382 
1383  return IRB.CreateInBoundsGEP(nullptr, BasePtr, Indices,
1384  NamePrefix + "sroa_idx");
1385 }
1386 
1387 /// Get a natural GEP off of the BasePtr walking through Ty toward
1388 /// TargetTy without changing the offset of the pointer.
1389 ///
1390 /// This routine assumes we've already established a properly offset GEP with
1391 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1392 /// zero-indices down through type layers until we find one the same as
1393 /// TargetTy. If we can't find one with the same type, we at least try to use
1394 /// one with the same size. If none of that works, we just produce the GEP as
1395 /// indicated by Indices to have the correct offset.
1396 static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1397  Value *BasePtr, Type *Ty, Type *TargetTy,
1398  SmallVectorImpl<Value *> &Indices,
1399  Twine NamePrefix) {
1400  if (Ty == TargetTy)
1401  return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1402 
1403  // Pointer size to use for the indices.
1404  unsigned PtrSize = DL.getPointerTypeSizeInBits(BasePtr->getType());
1405 
1406  // See if we can descend into a struct and locate a field with the correct
1407  // type.
1408  unsigned NumLayers = 0;
1409  Type *ElementTy = Ty;
1410  do {
1411  if (ElementTy->isPointerTy())
1412  break;
1413 
1414  if (ArrayType *ArrayTy = dyn_cast<ArrayType>(ElementTy)) {
1415  ElementTy = ArrayTy->getElementType();
1416  Indices.push_back(IRB.getIntN(PtrSize, 0));
1417  } else if (VectorType *VectorTy = dyn_cast<VectorType>(ElementTy)) {
1418  ElementTy = VectorTy->getElementType();
1419  Indices.push_back(IRB.getInt32(0));
1420  } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1421  if (STy->element_begin() == STy->element_end())
1422  break; // Nothing left to descend into.
1423  ElementTy = *STy->element_begin();
1424  Indices.push_back(IRB.getInt32(0));
1425  } else {
1426  break;
1427  }
1428  ++NumLayers;
1429  } while (ElementTy != TargetTy);
1430  if (ElementTy != TargetTy)
1431  Indices.erase(Indices.end() - NumLayers, Indices.end());
1432 
1433  return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1434 }
1435 
1436 /// Recursively compute indices for a natural GEP.
1437 ///
1438 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1439 /// element types adding appropriate indices for the GEP.
1440 static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1441  Value *Ptr, Type *Ty, APInt &Offset,
1442  Type *TargetTy,
1443  SmallVectorImpl<Value *> &Indices,
1444  Twine NamePrefix) {
1445  if (Offset == 0)
1446  return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices,
1447  NamePrefix);
1448 
1449  // We can't recurse through pointer types.
1450  if (Ty->isPointerTy())
1451  return nullptr;
1452 
1453  // We try to analyze GEPs over vectors here, but note that these GEPs are
1454  // extremely poorly defined currently. The long-term goal is to remove GEPing
1455  // over a vector from the IR completely.
1456  if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1457  unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
1458  if (ElementSizeInBits % 8 != 0) {
1459  // GEPs over non-multiple of 8 size vector elements are invalid.
1460  return nullptr;
1461  }
1462  APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1463  APInt NumSkippedElements = Offset.sdiv(ElementSize);
1464  if (NumSkippedElements.ugt(VecTy->getNumElements()))
1465  return nullptr;
1466  Offset -= NumSkippedElements * ElementSize;
1467  Indices.push_back(IRB.getInt(NumSkippedElements));
1468  return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1469  Offset, TargetTy, Indices, NamePrefix);
1470  }
1471 
1472  if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1473  Type *ElementTy = ArrTy->getElementType();
1474  APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1475  APInt NumSkippedElements = Offset.sdiv(ElementSize);
1476  if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1477  return nullptr;
1478 
1479  Offset -= NumSkippedElements * ElementSize;
1480  Indices.push_back(IRB.getInt(NumSkippedElements));
1481  return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1482  Indices, NamePrefix);
1483  }
1484 
1485  StructType *STy = dyn_cast<StructType>(Ty);
1486  if (!STy)
1487  return nullptr;
1488 
1489  const StructLayout *SL = DL.getStructLayout(STy);
1490  uint64_t StructOffset = Offset.getZExtValue();
1491  if (StructOffset >= SL->getSizeInBytes())
1492  return nullptr;
1493  unsigned Index = SL->getElementContainingOffset(StructOffset);
1494  Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1495  Type *ElementTy = STy->getElementType(Index);
1496  if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
1497  return nullptr; // The offset points into alignment padding.
1498 
1499  Indices.push_back(IRB.getInt32(Index));
1500  return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1501  Indices, NamePrefix);
1502 }
1503 
1504 /// Get a natural GEP from a base pointer to a particular offset and
1505 /// resulting in a particular type.
1506 ///
1507 /// The goal is to produce a "natural" looking GEP that works with the existing
1508 /// composite types to arrive at the appropriate offset and element type for
1509 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1510 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1511 /// Indices, and setting Ty to the result subtype.
1512 ///
1513 /// If no natural GEP can be constructed, this function returns null.
1514 static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1515  Value *Ptr, APInt Offset, Type *TargetTy,
1516  SmallVectorImpl<Value *> &Indices,
1517  Twine NamePrefix) {
1518  PointerType *Ty = cast<PointerType>(Ptr->getType());
1519 
1520  // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1521  // an i8.
1522  if (Ty == IRB.getInt8PtrTy(Ty->getAddressSpace()) && TargetTy->isIntegerTy(8))
1523  return nullptr;
1524 
1525  Type *ElementTy = Ty->getElementType();
1526  if (!ElementTy->isSized())
1527  return nullptr; // We can't GEP through an unsized element.
1528  APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1529  if (ElementSize == 0)
1530  return nullptr; // Zero-length arrays can't help us build a natural GEP.
1531  APInt NumSkippedElements = Offset.sdiv(ElementSize);
1532 
1533  Offset -= NumSkippedElements * ElementSize;
1534  Indices.push_back(IRB.getInt(NumSkippedElements));
1535  return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1536  Indices, NamePrefix);
1537 }
1538 
1539 /// Compute an adjusted pointer from Ptr by Offset bytes where the
1540 /// resulting pointer has PointerTy.
1541 ///
1542 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1543 /// and produces the pointer type desired. Where it cannot, it will try to use
1544 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1545 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1546 /// bitcast to the type.
1547 ///
1548 /// The strategy for finding the more natural GEPs is to peel off layers of the
1549 /// pointer, walking back through bit casts and GEPs, searching for a base
1550 /// pointer from which we can compute a natural GEP with the desired
1551 /// properties. The algorithm tries to fold as many constant indices into
1552 /// a single GEP as possible, thus making each GEP more independent of the
1553 /// surrounding code.
1554 static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr,
1555  APInt Offset, Type *PointerTy, Twine NamePrefix) {
1556  // Even though we don't look through PHI nodes, we could be called on an
1557  // instruction in an unreachable block, which may be on a cycle.
1558  SmallPtrSet<Value *, 4> Visited;
1559  Visited.insert(Ptr);
1560  SmallVector<Value *, 4> Indices;
1561 
1562  // We may end up computing an offset pointer that has the wrong type. If we
1563  // never are able to compute one directly that has the correct type, we'll
1564  // fall back to it, so keep it and the base it was computed from around here.
1565  Value *OffsetPtr = nullptr;
1566  Value *OffsetBasePtr;
1567 
1568  // Remember any i8 pointer we come across to re-use if we need to do a raw
1569  // byte offset.
1570  Value *Int8Ptr = nullptr;
1571  APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1572 
1573  Type *TargetTy = PointerTy->getPointerElementType();
1574 
1575  do {
1576  // First fold any existing GEPs into the offset.
1577  while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1578  APInt GEPOffset(Offset.getBitWidth(), 0);
1579  if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1580  break;
1581  Offset += GEPOffset;
1582  Ptr = GEP->getPointerOperand();
1583  if (!Visited.insert(Ptr).second)
1584  break;
1585  }
1586 
1587  // See if we can perform a natural GEP here.
1588  Indices.clear();
1589  if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1590  Indices, NamePrefix)) {
1591  // If we have a new natural pointer at the offset, clear out any old
1592  // offset pointer we computed. Unless it is the base pointer or
1593  // a non-instruction, we built a GEP we don't need. Zap it.
1594  if (OffsetPtr && OffsetPtr != OffsetBasePtr)
1595  if (Instruction *I = dyn_cast<Instruction>(OffsetPtr)) {
1596  assert(I->use_empty() && "Built a GEP with uses some how!");
1597  I->eraseFromParent();
1598  }
1599  OffsetPtr = P;
1600  OffsetBasePtr = Ptr;
1601  // If we also found a pointer of the right type, we're done.
1602  if (P->getType() == PointerTy)
1603  return P;
1604  }
1605 
1606  // Stash this pointer if we've found an i8*.
1607  if (Ptr->getType()->isIntegerTy(8)) {
1608  Int8Ptr = Ptr;
1609  Int8PtrOffset = Offset;
1610  }
1611 
1612  // Peel off a layer of the pointer and update the offset appropriately.
1613  if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1614  Ptr = cast<Operator>(Ptr)->getOperand(0);
1615  } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1616  if (GA->isInterposable())
1617  break;
1618  Ptr = GA->getAliasee();
1619  } else {
1620  break;
1621  }
1622  assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1623  } while (Visited.insert(Ptr).second);
1624 
1625  if (!OffsetPtr) {
1626  if (!Int8Ptr) {
1627  Int8Ptr = IRB.CreateBitCast(
1628  Ptr, IRB.getInt8PtrTy(PointerTy->getPointerAddressSpace()),
1629  NamePrefix + "sroa_raw_cast");
1630  Int8PtrOffset = Offset;
1631  }
1632 
1633  OffsetPtr = Int8PtrOffset == 0
1634  ? Int8Ptr
1635  : IRB.CreateInBoundsGEP(IRB.getInt8Ty(), Int8Ptr,
1636  IRB.getInt(Int8PtrOffset),
1637  NamePrefix + "sroa_raw_idx");
1638  }
1639  Ptr = OffsetPtr;
1640 
1641  // On the off chance we were targeting i8*, guard the bitcast here.
1642  if (Ptr->getType() != PointerTy)
1643  Ptr = IRB.CreateBitCast(Ptr, PointerTy, NamePrefix + "sroa_cast");
1644 
1645  return Ptr;
1646 }
1647 
1648 /// Compute the adjusted alignment for a load or store from an offset.
1649 static unsigned getAdjustedAlignment(Instruction *I, uint64_t Offset,
1650  const DataLayout &DL) {
1651  unsigned Alignment;
1652  Type *Ty;
1653  if (auto *LI = dyn_cast<LoadInst>(I)) {
1654  Alignment = LI->getAlignment();
1655  Ty = LI->getType();
1656  } else if (auto *SI = dyn_cast<StoreInst>(I)) {
1657  Alignment = SI->getAlignment();
1658  Ty = SI->getValueOperand()->getType();
1659  } else {
1660  llvm_unreachable("Only loads and stores are allowed!");
1661  }
1662 
1663  if (!Alignment)
1664  Alignment = DL.getABITypeAlignment(Ty);
1665 
1666  return MinAlign(Alignment, Offset);
1667 }
1668 
1669 /// Test whether we can convert a value from the old to the new type.
1670 ///
1671 /// This predicate should be used to guard calls to convertValue in order to
1672 /// ensure that we only try to convert viable values. The strategy is that we
1673 /// will peel off single element struct and array wrappings to get to an
1674 /// underlying value, and convert that value.
1675 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1676  if (OldTy == NewTy)
1677  return true;
1678 
1679  // For integer types, we can't handle any bit-width differences. This would
1680  // break both vector conversions with extension and introduce endianness
1681  // issues when in conjunction with loads and stores.
1682  if (isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) {
1683  assert(cast<IntegerType>(OldTy)->getBitWidth() !=
1684  cast<IntegerType>(NewTy)->getBitWidth() &&
1685  "We can't have the same bitwidth for different int types");
1686  return false;
1687  }
1688 
1689  if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1690  return false;
1691  if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1692  return false;
1693 
1694  // We can convert pointers to integers and vice-versa. Same for vectors
1695  // of pointers and integers.
1696  OldTy = OldTy->getScalarType();
1697  NewTy = NewTy->getScalarType();
1698  if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1699  if (NewTy->isPointerTy() && OldTy->isPointerTy()) {
1700  return cast<PointerType>(NewTy)->getPointerAddressSpace() ==
1701  cast<PointerType>(OldTy)->getPointerAddressSpace();
1702  }
1703 
1704  // We can convert integers to integral pointers, but not to non-integral
1705  // pointers.
1706  if (OldTy->isIntegerTy())
1707  return !DL.isNonIntegralPointerType(NewTy);
1708 
1709  // We can convert integral pointers to integers, but non-integral pointers
1710  // need to remain pointers.
1711  if (!DL.isNonIntegralPointerType(OldTy))
1712  return NewTy->isIntegerTy();
1713 
1714  return false;
1715  }
1716 
1717  return true;
1718 }
1719 
1720 /// Generic routine to convert an SSA value to a value of a different
1721 /// type.
1722 ///
1723 /// This will try various different casting techniques, such as bitcasts,
1724 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1725 /// two types for viability with this routine.
1726 static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1727  Type *NewTy) {
1728  Type *OldTy = V->getType();
1729  assert(canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type");
1730 
1731  if (OldTy == NewTy)
1732  return V;
1733 
1734  assert(!(isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) &&
1735  "Integer types must be the exact same to convert.");
1736 
1737  // See if we need inttoptr for this type pair. A cast involving both scalars
1738  // and vectors requires and additional bitcast.
1739  if (OldTy->isIntOrIntVectorTy() && NewTy->isPtrOrPtrVectorTy()) {
1740  // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1741  if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1742  return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1743  NewTy);
1744 
1745  // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1746  if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1747  return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1748  NewTy);
1749 
1750  return IRB.CreateIntToPtr(V, NewTy);
1751  }
1752 
1753  // See if we need ptrtoint for this type pair. A cast involving both scalars
1754  // and vectors requires and additional bitcast.
1755  if (OldTy->isPtrOrPtrVectorTy() && NewTy->isIntOrIntVectorTy()) {
1756  // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1757  if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1758  return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1759  NewTy);
1760 
1761  // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1762  if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1763  return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1764  NewTy);
1765 
1766  return IRB.CreatePtrToInt(V, NewTy);
1767  }
1768 
1769  return IRB.CreateBitCast(V, NewTy);
1770 }
1771 
1772 /// Test whether the given slice use can be promoted to a vector.
1773 ///
1774 /// This function is called to test each entry in a partition which is slated
1775 /// for a single slice.
1776 static bool isVectorPromotionViableForSlice(Partition &P, const Slice &S,
1777  VectorType *Ty,
1778  uint64_t ElementSize,
1779  const DataLayout &DL) {
1780  // First validate the slice offsets.
1781  uint64_t BeginOffset =
1782  std::max(S.beginOffset(), P.beginOffset()) - P.beginOffset();
1783  uint64_t BeginIndex = BeginOffset / ElementSize;
1784  if (BeginIndex * ElementSize != BeginOffset ||
1785  BeginIndex >= Ty->getNumElements())
1786  return false;
1787  uint64_t EndOffset =
1788  std::min(S.endOffset(), P.endOffset()) - P.beginOffset();
1789  uint64_t EndIndex = EndOffset / ElementSize;
1790  if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
1791  return false;
1792 
1793  assert(EndIndex > BeginIndex && "Empty vector!");
1794  uint64_t NumElements = EndIndex - BeginIndex;
1795  Type *SliceTy = (NumElements == 1)
1796  ? Ty->getElementType()
1797  : VectorType::get(Ty->getElementType(), NumElements);
1798 
1799  Type *SplitIntTy =
1800  Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1801 
1802  Use *U = S.getUse();
1803 
1804  if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1805  if (MI->isVolatile())
1806  return false;
1807  if (!S.isSplittable())
1808  return false; // Skip any unsplittable intrinsics.
1809  } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1810  if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1811  II->getIntrinsicID() != Intrinsic::lifetime_end)
1812  return false;
1813  } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1814  // Disable vector promotion when there are loads or stores of an FCA.
1815  return false;
1816  } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1817  if (LI->isVolatile())
1818  return false;
1819  Type *LTy = LI->getType();
1820  if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1821  assert(LTy->isIntegerTy());
1822  LTy = SplitIntTy;
1823  }
1824  if (!canConvertValue(DL, SliceTy, LTy))
1825  return false;
1826  } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1827  if (SI->isVolatile())
1828  return false;
1829  Type *STy = SI->getValueOperand()->getType();
1830  if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1831  assert(STy->isIntegerTy());
1832  STy = SplitIntTy;
1833  }
1834  if (!canConvertValue(DL, STy, SliceTy))
1835  return false;
1836  } else {
1837  return false;
1838  }
1839 
1840  return true;
1841 }
1842 
1843 /// Test whether the given alloca partitioning and range of slices can be
1844 /// promoted to a vector.
1845 ///
1846 /// This is a quick test to check whether we can rewrite a particular alloca
1847 /// partition (and its newly formed alloca) into a vector alloca with only
1848 /// whole-vector loads and stores such that it could be promoted to a vector
1849 /// SSA value. We only can ensure this for a limited set of operations, and we
1850 /// don't want to do the rewrites unless we are confident that the result will
1851 /// be promotable, so we have an early test here.
1853  // Collect the candidate types for vector-based promotion. Also track whether
1854  // we have different element types.
1855  SmallVector<VectorType *, 4> CandidateTys;
1856  Type *CommonEltTy = nullptr;
1857  bool HaveCommonEltTy = true;
1858  auto CheckCandidateType = [&](Type *Ty) {
1859  if (auto *VTy = dyn_cast<VectorType>(Ty)) {
1860  CandidateTys.push_back(VTy);
1861  if (!CommonEltTy)
1862  CommonEltTy = VTy->getElementType();
1863  else if (CommonEltTy != VTy->getElementType())
1864  HaveCommonEltTy = false;
1865  }
1866  };
1867  // Consider any loads or stores that are the exact size of the slice.
1868  for (const Slice &S : P)
1869  if (S.beginOffset() == P.beginOffset() &&
1870  S.endOffset() == P.endOffset()) {
1871  if (auto *LI = dyn_cast<LoadInst>(S.getUse()->getUser()))
1872  CheckCandidateType(LI->getType());
1873  else if (auto *SI = dyn_cast<StoreInst>(S.getUse()->getUser()))
1874  CheckCandidateType(SI->getValueOperand()->getType());
1875  }
1876 
1877  // If we didn't find a vector type, nothing to do here.
1878  if (CandidateTys.empty())
1879  return nullptr;
1880 
1881  // Remove non-integer vector types if we had multiple common element types.
1882  // FIXME: It'd be nice to replace them with integer vector types, but we can't
1883  // do that until all the backends are known to produce good code for all
1884  // integer vector types.
1885  if (!HaveCommonEltTy) {
1886  CandidateTys.erase(
1887  llvm::remove_if(CandidateTys,
1888  [](VectorType *VTy) {
1889  return !VTy->getElementType()->isIntegerTy();
1890  }),
1891  CandidateTys.end());
1892 
1893  // If there were no integer vector types, give up.
1894  if (CandidateTys.empty())
1895  return nullptr;
1896 
1897  // Rank the remaining candidate vector types. This is easy because we know
1898  // they're all integer vectors. We sort by ascending number of elements.
1899  auto RankVectorTypes = [&DL](VectorType *RHSTy, VectorType *LHSTy) {
1900  (void)DL;
1901  assert(DL.getTypeSizeInBits(RHSTy) == DL.getTypeSizeInBits(LHSTy) &&
1902  "Cannot have vector types of different sizes!");
1903  assert(RHSTy->getElementType()->isIntegerTy() &&
1904  "All non-integer types eliminated!");
1905  assert(LHSTy->getElementType()->isIntegerTy() &&
1906  "All non-integer types eliminated!");
1907  return RHSTy->getNumElements() < LHSTy->getNumElements();
1908  };
1909  llvm::sort(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes);
1910  CandidateTys.erase(
1911  std::unique(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes),
1912  CandidateTys.end());
1913  } else {
1914 // The only way to have the same element type in every vector type is to
1915 // have the same vector type. Check that and remove all but one.
1916 #ifndef NDEBUG
1917  for (VectorType *VTy : CandidateTys) {
1918  assert(VTy->getElementType() == CommonEltTy &&
1919  "Unaccounted for element type!");
1920  assert(VTy == CandidateTys[0] &&
1921  "Different vector types with the same element type!");
1922  }
1923 #endif
1924  CandidateTys.resize(1);
1925  }
1926 
1927  // Try each vector type, and return the one which works.
1928  auto CheckVectorTypeForPromotion = [&](VectorType *VTy) {
1929  uint64_t ElementSize = DL.getTypeSizeInBits(VTy->getElementType());
1930 
1931  // While the definition of LLVM vectors is bitpacked, we don't support sizes
1932  // that aren't byte sized.
1933  if (ElementSize % 8)
1934  return false;
1935  assert((DL.getTypeSizeInBits(VTy) % 8) == 0 &&
1936  "vector size not a multiple of element size?");
1937  ElementSize /= 8;
1938 
1939  for (const Slice &S : P)
1940  if (!isVectorPromotionViableForSlice(P, S, VTy, ElementSize, DL))
1941  return false;
1942 
1943  for (const Slice *S : P.splitSliceTails())
1944  if (!isVectorPromotionViableForSlice(P, *S, VTy, ElementSize, DL))
1945  return false;
1946 
1947  return true;
1948  };
1949  for (VectorType *VTy : CandidateTys)
1950  if (CheckVectorTypeForPromotion(VTy))
1951  return VTy;
1952 
1953  return nullptr;
1954 }
1955 
1956 /// Test whether a slice of an alloca is valid for integer widening.
1957 ///
1958 /// This implements the necessary checking for the \c isIntegerWideningViable
1959 /// test below on a single slice of the alloca.
1960 static bool isIntegerWideningViableForSlice(const Slice &S,
1961  uint64_t AllocBeginOffset,
1962  Type *AllocaTy,
1963  const DataLayout &DL,
1964  bool &WholeAllocaOp) {
1965  uint64_t Size = DL.getTypeStoreSize(AllocaTy);
1966 
1967  uint64_t RelBegin = S.beginOffset() - AllocBeginOffset;
1968  uint64_t RelEnd = S.endOffset() - AllocBeginOffset;
1969 
1970  // We can't reasonably handle cases where the load or store extends past
1971  // the end of the alloca's type and into its padding.
1972  if (RelEnd > Size)
1973  return false;
1974 
1975  Use *U = S.getUse();
1976 
1977  if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1978  if (LI->isVolatile())
1979  return false;
1980  // We can't handle loads that extend past the allocated memory.
1981  if (DL.getTypeStoreSize(LI->getType()) > Size)
1982  return false;
1983  // So far, AllocaSliceRewriter does not support widening split slice tails
1984  // in rewriteIntegerLoad.
1985  if (S.beginOffset() < AllocBeginOffset)
1986  return false;
1987  // Note that we don't count vector loads or stores as whole-alloca
1988  // operations which enable integer widening because we would prefer to use
1989  // vector widening instead.
1990  if (!isa<VectorType>(LI->getType()) && RelBegin == 0 && RelEnd == Size)
1991  WholeAllocaOp = true;
1992  if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
1993  if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1994  return false;
1995  } else if (RelBegin != 0 || RelEnd != Size ||
1996  !canConvertValue(DL, AllocaTy, LI->getType())) {
1997  // Non-integer loads need to be convertible from the alloca type so that
1998  // they are promotable.
1999  return false;
2000  }
2001  } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
2002  Type *ValueTy = SI->getValueOperand()->getType();
2003  if (SI->isVolatile())
2004  return false;
2005  // We can't handle stores that extend past the allocated memory.
2006  if (DL.getTypeStoreSize(ValueTy) > Size)
2007  return false;
2008  // So far, AllocaSliceRewriter does not support widening split slice tails
2009  // in rewriteIntegerStore.
2010  if (S.beginOffset() < AllocBeginOffset)
2011  return false;
2012  // Note that we don't count vector loads or stores as whole-alloca
2013  // operations which enable integer widening because we would prefer to use
2014  // vector widening instead.
2015  if (!isa<VectorType>(ValueTy) && RelBegin == 0 && RelEnd == Size)
2016  WholeAllocaOp = true;
2017  if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
2018  if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
2019  return false;
2020  } else if (RelBegin != 0 || RelEnd != Size ||
2021  !canConvertValue(DL, ValueTy, AllocaTy)) {
2022  // Non-integer stores need to be convertible to the alloca type so that
2023  // they are promotable.
2024  return false;
2025  }
2026  } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
2027  if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
2028  return false;
2029  if (!S.isSplittable())
2030  return false; // Skip any unsplittable intrinsics.
2031  } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
2032  if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
2033  II->getIntrinsicID() != Intrinsic::lifetime_end)
2034  return false;
2035  } else {
2036  return false;
2037  }
2038 
2039  return true;
2040 }
2041 
2042 /// Test whether the given alloca partition's integer operations can be
2043 /// widened to promotable ones.
2044 ///
2045 /// This is a quick test to check whether we can rewrite the integer loads and
2046 /// stores to a particular alloca into wider loads and stores and be able to
2047 /// promote the resulting alloca.
2048 static bool isIntegerWideningViable(Partition &P, Type *AllocaTy,
2049  const DataLayout &DL) {
2050  uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
2051  // Don't create integer types larger than the maximum bitwidth.
2052  if (SizeInBits > IntegerType::MAX_INT_BITS)
2053  return false;
2054 
2055  // Don't try to handle allocas with bit-padding.
2056  if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
2057  return false;
2058 
2059  // We need to ensure that an integer type with the appropriate bitwidth can
2060  // be converted to the alloca type, whatever that is. We don't want to force
2061  // the alloca itself to have an integer type if there is a more suitable one.
2062  Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
2063  if (!canConvertValue(DL, AllocaTy, IntTy) ||
2064  !canConvertValue(DL, IntTy, AllocaTy))
2065  return false;
2066 
2067  // While examining uses, we ensure that the alloca has a covering load or
2068  // store. We don't want to widen the integer operations only to fail to
2069  // promote due to some other unsplittable entry (which we may make splittable
2070  // later). However, if there are only splittable uses, go ahead and assume
2071  // that we cover the alloca.
2072  // FIXME: We shouldn't consider split slices that happen to start in the
2073  // partition here...
2074  bool WholeAllocaOp =
2075  P.begin() != P.end() ? false : DL.isLegalInteger(SizeInBits);
2076 
2077  for (const Slice &S : P)
2078  if (!isIntegerWideningViableForSlice(S, P.beginOffset(), AllocaTy, DL,
2079  WholeAllocaOp))
2080  return false;
2081 
2082  for (const Slice *S : P.splitSliceTails())
2083  if (!isIntegerWideningViableForSlice(*S, P.beginOffset(), AllocaTy, DL,
2084  WholeAllocaOp))
2085  return false;
2086 
2087  return WholeAllocaOp;
2088 }
2089 
2090 static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
2091  IntegerType *Ty, uint64_t Offset,
2092  const Twine &Name) {
2093  LLVM_DEBUG(dbgs() << " start: " << *V << "\n");
2094  IntegerType *IntTy = cast<IntegerType>(V->getType());
2095  assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2096  "Element extends past full value");
2097  uint64_t ShAmt = 8 * Offset;
2098  if (DL.isBigEndian())
2099  ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2100  if (ShAmt) {
2101  V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
2102  LLVM_DEBUG(dbgs() << " shifted: " << *V << "\n");
2103  }
2104  assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2105  "Cannot extract to a larger integer!");
2106  if (Ty != IntTy) {
2107  V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
2108  LLVM_DEBUG(dbgs() << " trunced: " << *V << "\n");
2109  }
2110  return V;
2111 }
2112 
2113 static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
2114  Value *V, uint64_t Offset, const Twine &Name) {
2115  IntegerType *IntTy = cast<IntegerType>(Old->getType());
2116  IntegerType *Ty = cast<IntegerType>(V->getType());
2117  assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2118  "Cannot insert a larger integer!");
2119  LLVM_DEBUG(dbgs() << " start: " << *V << "\n");
2120  if (Ty != IntTy) {
2121  V = IRB.CreateZExt(V, IntTy, Name + ".ext");
2122  LLVM_DEBUG(dbgs() << " extended: " << *V << "\n");
2123  }
2124  assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2125  "Element store outside of alloca store");
2126  uint64_t ShAmt = 8 * Offset;
2127  if (DL.isBigEndian())
2128  ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2129  if (ShAmt) {
2130  V = IRB.CreateShl(V, ShAmt, Name + ".shift");
2131  LLVM_DEBUG(dbgs() << " shifted: " << *V << "\n");
2132  }
2133 
2134  if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
2135  APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
2136  Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
2137  LLVM_DEBUG(dbgs() << " masked: " << *Old << "\n");
2138  V = IRB.CreateOr(Old, V, Name + ".insert");
2139  LLVM_DEBUG(dbgs() << " inserted: " << *V << "\n");
2140  }
2141  return V;
2142 }
2143 
2144 static Value *extractVector(IRBuilderTy &IRB, Value *V, unsigned BeginIndex,
2145  unsigned EndIndex, const Twine &Name) {
2146  VectorType *VecTy = cast<VectorType>(V->getType());
2147  unsigned NumElements = EndIndex - BeginIndex;
2148  assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2149 
2150  if (NumElements == VecTy->getNumElements())
2151  return V;
2152 
2153  if (NumElements == 1) {
2154  V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
2155  Name + ".extract");
2156  LLVM_DEBUG(dbgs() << " extract: " << *V << "\n");
2157  return V;
2158  }
2159 
2161  Mask.reserve(NumElements);
2162  for (unsigned i = BeginIndex; i != EndIndex; ++i)
2163  Mask.push_back(IRB.getInt32(i));
2164  V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2165  ConstantVector::get(Mask), Name + ".extract");
2166  LLVM_DEBUG(dbgs() << " shuffle: " << *V << "\n");
2167  return V;
2168 }
2169 
2170 static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
2171  unsigned BeginIndex, const Twine &Name) {
2172  VectorType *VecTy = cast<VectorType>(Old->getType());
2173  assert(VecTy && "Can only insert a vector into a vector");
2174 
2175  VectorType *Ty = dyn_cast<VectorType>(V->getType());
2176  if (!Ty) {
2177  // Single element to insert.
2178  V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
2179  Name + ".insert");
2180  LLVM_DEBUG(dbgs() << " insert: " << *V << "\n");
2181  return V;
2182  }
2183 
2184  assert(Ty->getNumElements() <= VecTy->getNumElements() &&
2185  "Too many elements!");
2186  if (Ty->getNumElements() == VecTy->getNumElements()) {
2187  assert(V->getType() == VecTy && "Vector type mismatch");
2188  return V;
2189  }
2190  unsigned EndIndex = BeginIndex + Ty->getNumElements();
2191 
2192  // When inserting a smaller vector into the larger to store, we first
2193  // use a shuffle vector to widen it with undef elements, and then
2194  // a second shuffle vector to select between the loaded vector and the
2195  // incoming vector.
2197  Mask.reserve(VecTy->getNumElements());
2198  for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2199  if (i >= BeginIndex && i < EndIndex)
2200  Mask.push_back(IRB.getInt32(i - BeginIndex));
2201  else
2202  Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
2203  V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2204  ConstantVector::get(Mask), Name + ".expand");
2205  LLVM_DEBUG(dbgs() << " shuffle: " << *V << "\n");
2206 
2207  Mask.clear();
2208  for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2209  Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
2210 
2211  V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
2212 
2213  LLVM_DEBUG(dbgs() << " blend: " << *V << "\n");
2214  return V;
2215 }
2216 
2217 /// Visitor to rewrite instructions using p particular slice of an alloca
2218 /// to use a new alloca.
2219 ///
2220 /// Also implements the rewriting to vector-based accesses when the partition
2221 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2222 /// lives here.
2224  : public InstVisitor<AllocaSliceRewriter, bool> {
2225  // Befriend the base class so it can delegate to private visit methods.
2227 
2229 
2230  const DataLayout &DL;
2231  AllocaSlices &AS;
2232  SROA &Pass;
2233  AllocaInst &OldAI, &NewAI;
2234  const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2235  Type *NewAllocaTy;
2236 
2237  // This is a convenience and flag variable that will be null unless the new
2238  // alloca's integer operations should be widened to this integer type due to
2239  // passing isIntegerWideningViable above. If it is non-null, the desired
2240  // integer type will be stored here for easy access during rewriting.
2241  IntegerType *IntTy;
2242 
2243  // If we are rewriting an alloca partition which can be written as pure
2244  // vector operations, we stash extra information here. When VecTy is
2245  // non-null, we have some strict guarantees about the rewritten alloca:
2246  // - The new alloca is exactly the size of the vector type here.
2247  // - The accesses all either map to the entire vector or to a single
2248  // element.
2249  // - The set of accessing instructions is only one of those handled above
2250  // in isVectorPromotionViable. Generally these are the same access kinds
2251  // which are promotable via mem2reg.
2252  VectorType *VecTy;
2253  Type *ElementTy;
2254  uint64_t ElementSize;
2255 
2256  // The original offset of the slice currently being rewritten relative to
2257  // the original alloca.
2258  uint64_t BeginOffset = 0;
2259  uint64_t EndOffset = 0;
2260 
2261  // The new offsets of the slice currently being rewritten relative to the
2262  // original alloca.
2263  uint64_t NewBeginOffset, NewEndOffset;
2264 
2265  uint64_t SliceSize;
2266  bool IsSplittable = false;
2267  bool IsSplit = false;
2268  Use *OldUse = nullptr;
2269  Instruction *OldPtr = nullptr;
2270 
2271  // Track post-rewrite users which are PHI nodes and Selects.
2272  SmallSetVector<PHINode *, 8> &PHIUsers;
2273  SmallSetVector<SelectInst *, 8> &SelectUsers;
2274 
2275  // Utility IR builder, whose name prefix is setup for each visited use, and
2276  // the insertion point is set to point to the user.
2277  IRBuilderTy IRB;
2278 
2279 public:
2281  AllocaInst &OldAI, AllocaInst &NewAI,
2282  uint64_t NewAllocaBeginOffset,
2283  uint64_t NewAllocaEndOffset, bool IsIntegerPromotable,
2284  VectorType *PromotableVecTy,
2285  SmallSetVector<PHINode *, 8> &PHIUsers,
2286  SmallSetVector<SelectInst *, 8> &SelectUsers)
2287  : DL(DL), AS(AS), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
2288  NewAllocaBeginOffset(NewAllocaBeginOffset),
2289  NewAllocaEndOffset(NewAllocaEndOffset),
2290  NewAllocaTy(NewAI.getAllocatedType()),
2291  IntTy(IsIntegerPromotable
2292  ? Type::getIntNTy(
2293  NewAI.getContext(),
2294  DL.getTypeSizeInBits(NewAI.getAllocatedType()))
2295  : nullptr),
2296  VecTy(PromotableVecTy),
2297  ElementTy(VecTy ? VecTy->getElementType() : nullptr),
2298  ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
2299  PHIUsers(PHIUsers), SelectUsers(SelectUsers),
2300  IRB(NewAI.getContext(), ConstantFolder()) {
2301  if (VecTy) {
2302  assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
2303  "Only multiple-of-8 sized vector elements are viable");
2304  ++NumVectorized;
2305  }
2306  assert((!IntTy && !VecTy) || (IntTy && !VecTy) || (!IntTy && VecTy));
2307  }
2308 
2310  bool CanSROA = true;
2311  BeginOffset = I->beginOffset();
2312  EndOffset = I->endOffset();
2313  IsSplittable = I->isSplittable();
2314  IsSplit =
2315  BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
2316  LLVM_DEBUG(dbgs() << " rewriting " << (IsSplit ? "split " : ""));
2317  LLVM_DEBUG(AS.printSlice(dbgs(), I, ""));
2318  LLVM_DEBUG(dbgs() << "\n");
2319 
2320  // Compute the intersecting offset range.
2321  assert(BeginOffset < NewAllocaEndOffset);
2322  assert(EndOffset > NewAllocaBeginOffset);
2323  NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2324  NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2325 
2326  SliceSize = NewEndOffset - NewBeginOffset;
2327 
2328  OldUse = I->getUse();
2329  OldPtr = cast<Instruction>(OldUse->get());
2330 
2331  Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2332  IRB.SetInsertPoint(OldUserI);
2333  IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2334  IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
2335 
2336  CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2337  if (VecTy || IntTy)
2338  assert(CanSROA);
2339  return CanSROA;
2340  }
2341 
2342 private:
2343  // Make sure the other visit overloads are visible.
2344  using Base::visit;
2345 
2346  // Every instruction which can end up as a user must have a rewrite rule.
2347  bool visitInstruction(Instruction &I) {
2348  LLVM_DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2349  llvm_unreachable("No rewrite rule for this instruction!");
2350  }
2351 
2352  Value *getNewAllocaSlicePtr(IRBuilderTy &IRB, Type *PointerTy) {
2353  // Note that the offset computation can use BeginOffset or NewBeginOffset
2354  // interchangeably for unsplit slices.
2355  assert(IsSplit || BeginOffset == NewBeginOffset);
2356  uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2357 
2358 #ifndef NDEBUG
2359  StringRef OldName = OldPtr->getName();
2360  // Skip through the last '.sroa.' component of the name.
2361  size_t LastSROAPrefix = OldName.rfind(".sroa.");
2362  if (LastSROAPrefix != StringRef::npos) {
2363  OldName = OldName.substr(LastSROAPrefix + strlen(".sroa."));
2364  // Look for an SROA slice index.
2365  size_t IndexEnd = OldName.find_first_not_of("0123456789");
2366  if (IndexEnd != StringRef::npos && OldName[IndexEnd] == '.') {
2367  // Strip the index and look for the offset.
2368  OldName = OldName.substr(IndexEnd + 1);
2369  size_t OffsetEnd = OldName.find_first_not_of("0123456789");
2370  if (OffsetEnd != StringRef::npos && OldName[OffsetEnd] == '.')
2371  // Strip the offset.
2372  OldName = OldName.substr(OffsetEnd + 1);
2373  }
2374  }
2375  // Strip any SROA suffixes as well.
2376  OldName = OldName.substr(0, OldName.find(".sroa_"));
2377 #endif
2378 
2379  return getAdjustedPtr(IRB, DL, &NewAI,
2380  APInt(DL.getPointerTypeSizeInBits(PointerTy), Offset),
2381  PointerTy,
2382 #ifndef NDEBUG
2383  Twine(OldName) + "."
2384 #else
2385  Twine()
2386 #endif
2387  );
2388  }
2389 
2390  /// Compute suitable alignment to access this slice of the *new*
2391  /// alloca.
2392  ///
2393  /// You can optionally pass a type to this routine and if that type's ABI
2394  /// alignment is itself suitable, this will return zero.
2395  unsigned getSliceAlign(Type *Ty = nullptr) {
2396  unsigned NewAIAlign = NewAI.getAlignment();
2397  if (!NewAIAlign)
2398  NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
2399  unsigned Align =
2400  MinAlign(NewAIAlign, NewBeginOffset - NewAllocaBeginOffset);
2401  return (Ty && Align == DL.getABITypeAlignment(Ty)) ? 0 : Align;
2402  }
2403 
2404  unsigned getIndex(uint64_t Offset) {
2405  assert(VecTy && "Can only call getIndex when rewriting a vector");
2406  uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2407  assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2408  uint32_t Index = RelOffset / ElementSize;
2409  assert(Index * ElementSize == RelOffset);
2410  return Index;
2411  }
2412 
2413  void deleteIfTriviallyDead(Value *V) {
2414  Instruction *I = cast<Instruction>(V);
2416  Pass.DeadInsts.insert(I);
2417  }
2418 
2419  Value *rewriteVectorizedLoadInst() {
2420  unsigned BeginIndex = getIndex(NewBeginOffset);
2421  unsigned EndIndex = getIndex(NewEndOffset);
2422  assert(EndIndex > BeginIndex && "Empty vector!");
2423 
2424  Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2425  return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2426  }
2427 
2428  Value *rewriteIntegerLoad(LoadInst &LI) {
2429  assert(IntTy && "We cannot insert an integer to the alloca");
2430  assert(!LI.isVolatile());
2431  Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2432  V = convertValue(DL, IRB, V, IntTy);
2433  assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2434  uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2435  if (Offset > 0 || NewEndOffset < NewAllocaEndOffset) {
2436  IntegerType *ExtractTy = Type::getIntNTy(LI.getContext(), SliceSize * 8);
2437  V = extractInteger(DL, IRB, V, ExtractTy, Offset, "extract");
2438  }
2439  // It is possible that the extracted type is not the load type. This
2440  // happens if there is a load past the end of the alloca, and as
2441  // a consequence the slice is narrower but still a candidate for integer
2442  // lowering. To handle this case, we just zero extend the extracted
2443  // integer.
2444  assert(cast<IntegerType>(LI.getType())->getBitWidth() >= SliceSize * 8 &&
2445  "Can only handle an extract for an overly wide load");
2446  if (cast<IntegerType>(LI.getType())->getBitWidth() > SliceSize * 8)
2447  V = IRB.CreateZExt(V, LI.getType());
2448  return V;
2449  }
2450 
2451  bool visitLoadInst(LoadInst &LI) {
2452  LLVM_DEBUG(dbgs() << " original: " << LI << "\n");
2453  Value *OldOp = LI.getOperand(0);
2454  assert(OldOp == OldPtr);
2455 
2456  AAMDNodes AATags;
2457  LI.getAAMetadata(AATags);
2458 
2459  unsigned AS = LI.getPointerAddressSpace();
2460 
2461  Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), SliceSize * 8)
2462  : LI.getType();
2463  const bool IsLoadPastEnd = DL.getTypeStoreSize(TargetTy) > SliceSize;
2464  bool IsPtrAdjusted = false;
2465  Value *V;
2466  if (VecTy) {
2467  V = rewriteVectorizedLoadInst();
2468  } else if (IntTy && LI.getType()->isIntegerTy()) {
2469  V = rewriteIntegerLoad(LI);
2470  } else if (NewBeginOffset == NewAllocaBeginOffset &&
2471  NewEndOffset == NewAllocaEndOffset &&
2472  (canConvertValue(DL, NewAllocaTy, TargetTy) ||
2473  (IsLoadPastEnd && NewAllocaTy->isIntegerTy() &&
2474  TargetTy->isIntegerTy()))) {
2475  LoadInst *NewLI = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2476  LI.isVolatile(), LI.getName());
2477  if (AATags)
2478  NewLI->setAAMetadata(AATags);
2479  if (LI.isVolatile())
2480  NewLI->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
2481 
2482  // Any !nonnull metadata or !range metadata on the old load is also valid
2483  // on the new load. This is even true in some cases even when the loads
2484  // are different types, for example by mapping !nonnull metadata to
2485  // !range metadata by modeling the null pointer constant converted to the
2486  // integer type.
2487  // FIXME: Add support for range metadata here. Currently the utilities
2488  // for this don't propagate range metadata in trivial cases from one
2489  // integer load to another, don't handle non-addrspace-0 null pointers
2490  // correctly, and don't have any support for mapping ranges as the
2491  // integer type becomes winder or narrower.
2493  copyNonnullMetadata(LI, N, *NewLI);
2494 
2495  // Try to preserve nonnull metadata
2496  V = NewLI;
2497 
2498  // If this is an integer load past the end of the slice (which means the
2499  // bytes outside the slice are undef or this load is dead) just forcibly
2500  // fix the integer size with correct handling of endianness.
2501  if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2502  if (auto *TITy = dyn_cast<IntegerType>(TargetTy))
2503  if (AITy->getBitWidth() < TITy->getBitWidth()) {
2504  V = IRB.CreateZExt(V, TITy, "load.ext");
2505  if (DL.isBigEndian())
2506  V = IRB.CreateShl(V, TITy->getBitWidth() - AITy->getBitWidth(),
2507  "endian_shift");
2508  }
2509  } else {
2510  Type *LTy = TargetTy->getPointerTo(AS);
2511  LoadInst *NewLI = IRB.CreateAlignedLoad(getNewAllocaSlicePtr(IRB, LTy),
2512  getSliceAlign(TargetTy),
2513  LI.isVolatile(), LI.getName());
2514  if (AATags)
2515  NewLI->setAAMetadata(AATags);
2516  if (LI.isVolatile())
2517  NewLI->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
2518 
2519  V = NewLI;
2520  IsPtrAdjusted = true;
2521  }
2522  V = convertValue(DL, IRB, V, TargetTy);
2523 
2524  if (IsSplit) {
2525  assert(!LI.isVolatile());
2526  assert(LI.getType()->isIntegerTy() &&
2527  "Only integer type loads and stores are split");
2528  assert(SliceSize < DL.getTypeStoreSize(LI.getType()) &&
2529  "Split load isn't smaller than original load");
2530  assert(LI.getType()->getIntegerBitWidth() ==
2531  DL.getTypeStoreSizeInBits(LI.getType()) &&
2532  "Non-byte-multiple bit width");
2533  // Move the insertion point just past the load so that we can refer to it.
2534  IRB.SetInsertPoint(&*std::next(BasicBlock::iterator(&LI)));
2535  // Create a placeholder value with the same type as LI to use as the
2536  // basis for the new value. This allows us to replace the uses of LI with
2537  // the computed value, and then replace the placeholder with LI, leaving
2538  // LI only used for this computation.
2539  Value *Placeholder =
2540  new LoadInst(UndefValue::get(LI.getType()->getPointerTo(AS)));
2541  V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset - BeginOffset,
2542  "insert");
2543  LI.replaceAllUsesWith(V);
2544  Placeholder->replaceAllUsesWith(&LI);
2545  Placeholder->deleteValue();
2546  } else {
2547  LI.replaceAllUsesWith(V);
2548  }
2549 
2550  Pass.DeadInsts.insert(&LI);
2551  deleteIfTriviallyDead(OldOp);
2552  LLVM_DEBUG(dbgs() << " to: " << *V << "\n");
2553  return !LI.isVolatile() && !IsPtrAdjusted;
2554  }
2555 
2556  bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp,
2557  AAMDNodes AATags) {
2558  if (V->getType() != VecTy) {
2559  unsigned BeginIndex = getIndex(NewBeginOffset);
2560  unsigned EndIndex = getIndex(NewEndOffset);
2561  assert(EndIndex > BeginIndex && "Empty vector!");
2562  unsigned NumElements = EndIndex - BeginIndex;
2563  assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2564  Type *SliceTy = (NumElements == 1)
2565  ? ElementTy
2566  : VectorType::get(ElementTy, NumElements);
2567  if (V->getType() != SliceTy)
2568  V = convertValue(DL, IRB, V, SliceTy);
2569 
2570  // Mix in the existing elements.
2571  Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2572  V = insertVector(IRB, Old, V, BeginIndex, "vec");
2573  }
2574  StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2575  if (AATags)
2576  Store->setAAMetadata(AATags);
2577  Pass.DeadInsts.insert(&SI);
2578 
2579  LLVM_DEBUG(dbgs() << " to: " << *Store << "\n");
2580  return true;
2581  }
2582 
2583  bool rewriteIntegerStore(Value *V, StoreInst &SI, AAMDNodes AATags) {
2584  assert(IntTy && "We cannot extract an integer from the alloca");
2585  assert(!SI.isVolatile());
2586  if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2587  Value *Old =
2588  IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2589  Old = convertValue(DL, IRB, Old, IntTy);
2590  assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2591  uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2592  V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset, "insert");
2593  }
2594  V = convertValue(DL, IRB, V, NewAllocaTy);
2595  StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2597  if (AATags)
2598  Store->setAAMetadata(AATags);
2599  Pass.DeadInsts.insert(&SI);
2600  LLVM_DEBUG(dbgs() << " to: " << *Store << "\n");
2601  return true;
2602  }
2603 
2604  bool visitStoreInst(StoreInst &SI) {
2605  LLVM_DEBUG(dbgs() << " original: " << SI << "\n");
2606  Value *OldOp = SI.getOperand(1);
2607  assert(OldOp == OldPtr);
2608 
2609  AAMDNodes AATags;
2610  SI.getAAMetadata(AATags);
2611 
2612  Value *V = SI.getValueOperand();
2613 
2614  // Strip all inbounds GEPs and pointer casts to try to dig out any root
2615  // alloca that should be re-examined after promoting this alloca.
2616  if (V->getType()->isPointerTy())
2617  if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2618  Pass.PostPromotionWorklist.insert(AI);
2619 
2620  if (SliceSize < DL.getTypeStoreSize(V->getType())) {
2621  assert(!SI.isVolatile());
2622  assert(V->getType()->isIntegerTy() &&
2623  "Only integer type loads and stores are split");
2624  assert(V->getType()->getIntegerBitWidth() ==
2625  DL.getTypeStoreSizeInBits(V->getType()) &&
2626  "Non-byte-multiple bit width");
2627  IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), SliceSize * 8);
2628  V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset - BeginOffset,
2629  "extract");
2630  }
2631 
2632  if (VecTy)
2633  return rewriteVectorizedStoreInst(V, SI, OldOp, AATags);
2634  if (IntTy && V->getType()->isIntegerTy())
2635  return rewriteIntegerStore(V, SI, AATags);
2636 
2637  const bool IsStorePastEnd = DL.getTypeStoreSize(V->getType()) > SliceSize;
2638  StoreInst *NewSI;
2639  if (NewBeginOffset == NewAllocaBeginOffset &&
2640  NewEndOffset == NewAllocaEndOffset &&
2641  (canConvertValue(DL, V->getType(), NewAllocaTy) ||
2642  (IsStorePastEnd && NewAllocaTy->isIntegerTy() &&
2643  V->getType()->isIntegerTy()))) {
2644  // If this is an integer store past the end of slice (and thus the bytes
2645  // past that point are irrelevant or this is unreachable), truncate the
2646  // value prior to storing.
2647  if (auto *VITy = dyn_cast<IntegerType>(V->getType()))
2648  if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2649  if (VITy->getBitWidth() > AITy->getBitWidth()) {
2650  if (DL.isBigEndian())
2651  V = IRB.CreateLShr(V, VITy->getBitWidth() - AITy->getBitWidth(),
2652  "endian_shift");
2653  V = IRB.CreateTrunc(V, AITy, "load.trunc");
2654  }
2655 
2656  V = convertValue(DL, IRB, V, NewAllocaTy);
2657  NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2658  SI.isVolatile());
2659  } else {
2660  unsigned AS = SI.getPointerAddressSpace();
2661  Value *NewPtr = getNewAllocaSlicePtr(IRB, V->getType()->getPointerTo(AS));
2662  NewSI = IRB.CreateAlignedStore(V, NewPtr, getSliceAlign(V->getType()),
2663  SI.isVolatile());
2664  }
2666  if (AATags)
2667  NewSI->setAAMetadata(AATags);
2668  if (SI.isVolatile())
2669  NewSI->setAtomic(SI.getOrdering(), SI.getSyncScopeID());
2670  Pass.DeadInsts.insert(&SI);
2671  deleteIfTriviallyDead(OldOp);
2672 
2673  LLVM_DEBUG(dbgs() << " to: " << *NewSI << "\n");
2674  return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2675  }
2676 
2677  /// Compute an integer value from splatting an i8 across the given
2678  /// number of bytes.
2679  ///
2680  /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2681  /// call this routine.
2682  /// FIXME: Heed the advice above.
2683  ///
2684  /// \param V The i8 value to splat.
2685  /// \param Size The number of bytes in the output (assuming i8 is one byte)
2686  Value *getIntegerSplat(Value *V, unsigned Size) {
2687  assert(Size > 0 && "Expected a positive number of bytes.");
2688  IntegerType *VTy = cast<IntegerType>(V->getType());
2689  assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2690  if (Size == 1)
2691  return V;
2692 
2693  Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size * 8);
2694  V = IRB.CreateMul(
2695  IRB.CreateZExt(V, SplatIntTy, "zext"),
2697  Constant::getAllOnesValue(SplatIntTy),
2699  SplatIntTy)),
2700  "isplat");
2701  return V;
2702  }
2703 
2704  /// Compute a vector splat for a given element value.
2705  Value *getVectorSplat(Value *V, unsigned NumElements) {
2706  V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2707  LLVM_DEBUG(dbgs() << " splat: " << *V << "\n");
2708  return V;
2709  }
2710 
2711  bool visitMemSetInst(MemSetInst &II) {
2712  LLVM_DEBUG(dbgs() << " original: " << II << "\n");
2713  assert(II.getRawDest() == OldPtr);
2714 
2715  AAMDNodes AATags;
2716  II.getAAMetadata(AATags);
2717 
2718  // If the memset has a variable size, it cannot be split, just adjust the
2719  // pointer to the new alloca.
2720  if (!isa<Constant>(II.getLength())) {
2721  assert(!IsSplit);
2722  assert(NewBeginOffset == BeginOffset);
2723  II.setDest(getNewAllocaSlicePtr(IRB, OldPtr->getType()));
2724  II.setDestAlignment(getSliceAlign());
2725 
2726  deleteIfTriviallyDead(OldPtr);
2727  return false;
2728  }
2729 
2730  // Record this instruction for deletion.
2731  Pass.DeadInsts.insert(&II);
2732 
2733  Type *AllocaTy = NewAI.getAllocatedType();
2734  Type *ScalarTy = AllocaTy->getScalarType();
2735 
2736  // If this doesn't map cleanly onto the alloca type, and that type isn't
2737  // a single value type, just emit a memset.
2738  if (!VecTy && !IntTy &&
2739  (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2740  SliceSize != DL.getTypeStoreSize(AllocaTy) ||
2741  !AllocaTy->isSingleValueType() ||
2742  !DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy)) ||
2743  DL.getTypeSizeInBits(ScalarTy) % 8 != 0)) {
2744  Type *SizeTy = II.getLength()->getType();
2745  Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2746  CallInst *New = IRB.CreateMemSet(
2747  getNewAllocaSlicePtr(IRB, OldPtr->getType()), II.getValue(), Size,
2748  getSliceAlign(), II.isVolatile());
2749  if (AATags)
2750  New->setAAMetadata(AATags);
2751  LLVM_DEBUG(dbgs() << " to: " << *New << "\n");
2752  return false;
2753  }
2754 
2755  // If we can represent this as a simple value, we have to build the actual
2756  // value to store, which requires expanding the byte present in memset to
2757  // a sensible representation for the alloca type. This is essentially
2758  // splatting the byte to a sufficiently wide integer, splatting it across
2759  // any desired vector width, and bitcasting to the final type.
2760  Value *V;
2761 
2762  if (VecTy) {
2763  // If this is a memset of a vectorized alloca, insert it.
2764  assert(ElementTy == ScalarTy);
2765 
2766  unsigned BeginIndex = getIndex(NewBeginOffset);
2767  unsigned EndIndex = getIndex(NewEndOffset);
2768  assert(EndIndex > BeginIndex && "Empty vector!");
2769  unsigned NumElements = EndIndex - BeginIndex;
2770  assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2771 
2772  Value *Splat =
2773  getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
2774  Splat = convertValue(DL, IRB, Splat, ElementTy);
2775  if (NumElements > 1)
2776  Splat = getVectorSplat(Splat, NumElements);
2777 
2778  Value *Old =
2779  IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2780  V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2781  } else if (IntTy) {
2782  // If this is a memset on an alloca where we can widen stores, insert the
2783  // set integer.
2784  assert(!II.isVolatile());
2785 
2786  uint64_t Size = NewEndOffset - NewBeginOffset;
2787  V = getIntegerSplat(II.getValue(), Size);
2788 
2789  if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2790  EndOffset != NewAllocaBeginOffset)) {
2791  Value *Old =
2792  IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2793  Old = convertValue(DL, IRB, Old, IntTy);
2794  uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2795  V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2796  } else {
2797  assert(V->getType() == IntTy &&
2798  "Wrong type for an alloca wide integer!");
2799  }
2800  V = convertValue(DL, IRB, V, AllocaTy);
2801  } else {
2802  // Established these invariants above.
2803  assert(NewBeginOffset == NewAllocaBeginOffset);
2804  assert(NewEndOffset == NewAllocaEndOffset);
2805 
2806  V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
2807  if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2808  V = getVectorSplat(V, AllocaVecTy->getNumElements());
2809 
2810  V = convertValue(DL, IRB, V, AllocaTy);
2811  }
2812 
2813  StoreInst *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2814  II.isVolatile());
2815  if (AATags)
2816  New->setAAMetadata(AATags);
2817  LLVM_DEBUG(dbgs() << " to: " << *New << "\n");
2818  return !II.isVolatile();
2819  }
2820 
2821  bool visitMemTransferInst(MemTransferInst &II) {
2822  // Rewriting of memory transfer instructions can be a bit tricky. We break
2823  // them into two categories: split intrinsics and unsplit intrinsics.
2824 
2825  LLVM_DEBUG(dbgs() << " original: " << II << "\n");
2826 
2827  AAMDNodes AATags;
2828  II.getAAMetadata(AATags);
2829 
2830  bool IsDest = &II.getRawDestUse() == OldUse;
2831  assert((IsDest && II.getRawDest() == OldPtr) ||
2832  (!IsDest && II.getRawSource() == OldPtr));
2833 
2834  unsigned SliceAlign = getSliceAlign();
2835 
2836  // For unsplit intrinsics, we simply modify the source and destination
2837  // pointers in place. This isn't just an optimization, it is a matter of
2838  // correctness. With unsplit intrinsics we may be dealing with transfers
2839  // within a single alloca before SROA ran, or with transfers that have
2840  // a variable length. We may also be dealing with memmove instead of
2841  // memcpy, and so simply updating the pointers is the necessary for us to
2842  // update both source and dest of a single call.
2843  if (!IsSplittable) {
2844  Value *AdjustedPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2845  if (IsDest) {
2846  II.setDest(AdjustedPtr);
2847  II.setDestAlignment(SliceAlign);
2848  }
2849  else {
2850  II.setSource(AdjustedPtr);
2851  II.setSourceAlignment(SliceAlign);
2852  }
2853 
2854  LLVM_DEBUG(dbgs() << " to: " << II << "\n");
2855  deleteIfTriviallyDead(OldPtr);
2856  return false;
2857  }
2858  // For split transfer intrinsics we have an incredibly useful assurance:
2859  // the source and destination do not reside within the same alloca, and at
2860  // least one of them does not escape. This means that we can replace
2861  // memmove with memcpy, and we don't need to worry about all manner of
2862  // downsides to splitting and transforming the operations.
2863 
2864  // If this doesn't map cleanly onto the alloca type, and that type isn't
2865  // a single value type, just emit a memcpy.
2866  bool EmitMemCpy =
2867  !VecTy && !IntTy &&
2868  (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2869  SliceSize != DL.getTypeStoreSize(NewAI.getAllocatedType()) ||
2870  !NewAI.getAllocatedType()->isSingleValueType());
2871 
2872  // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2873  // size hasn't been shrunk based on analysis of the viable range, this is
2874  // a no-op.
2875  if (EmitMemCpy && &OldAI == &NewAI) {
2876  // Ensure the start lines up.
2877  assert(NewBeginOffset == BeginOffset);
2878 
2879  // Rewrite the size as needed.
2880  if (NewEndOffset != EndOffset)
2882  NewEndOffset - NewBeginOffset));
2883  return false;
2884  }
2885  // Record this instruction for deletion.
2886  Pass.DeadInsts.insert(&II);
2887 
2888  // Strip all inbounds GEPs and pointer casts to try to dig out any root
2889  // alloca that should be re-examined after rewriting this instruction.
2890  Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2891  if (AllocaInst *AI =
2892  dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
2893  assert(AI != &OldAI && AI != &NewAI &&
2894  "Splittable transfers cannot reach the same alloca on both ends.");
2895  Pass.Worklist.insert(AI);
2896  }
2897 
2898  Type *OtherPtrTy = OtherPtr->getType();
2899  unsigned OtherAS = OtherPtrTy->getPointerAddressSpace();
2900 
2901  // Compute the relative offset for the other pointer within the transfer.
2902  unsigned IntPtrWidth = DL.getPointerSizeInBits(OtherAS);
2903  APInt OtherOffset(IntPtrWidth, NewBeginOffset - BeginOffset);
2904  unsigned OtherAlign =
2905  IsDest ? II.getSourceAlignment() : II.getDestAlignment();
2906  OtherAlign = MinAlign(OtherAlign ? OtherAlign : 1,
2907  OtherOffset.zextOrTrunc(64).getZExtValue());
2908 
2909  if (EmitMemCpy) {
2910  // Compute the other pointer, folding as much as possible to produce
2911  // a single, simple GEP in most cases.
2912  OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2913  OtherPtr->getName() + ".");
2914 
2915  Value *OurPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2916  Type *SizeTy = II.getLength()->getType();
2917  Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2918 
2919  Value *DestPtr, *SrcPtr;
2920  unsigned DestAlign, SrcAlign;
2921  // Note: IsDest is true iff we're copying into the new alloca slice
2922  if (IsDest) {
2923  DestPtr = OurPtr;
2924  DestAlign = SliceAlign;
2925  SrcPtr = OtherPtr;
2926  SrcAlign = OtherAlign;
2927  } else {
2928  DestPtr = OtherPtr;
2929  DestAlign = OtherAlign;
2930  SrcPtr = OurPtr;
2931  SrcAlign = SliceAlign;
2932  }
2933  CallInst *New = IRB.CreateMemCpy(DestPtr, DestAlign, SrcPtr, SrcAlign,
2934  Size, II.isVolatile());
2935  if (AATags)
2936  New->setAAMetadata(AATags);
2937  LLVM_DEBUG(dbgs() << " to: " << *New << "\n");
2938  return false;
2939  }
2940 
2941  bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
2942  NewEndOffset == NewAllocaEndOffset;
2943  uint64_t Size = NewEndOffset - NewBeginOffset;
2944  unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
2945  unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
2946  unsigned NumElements = EndIndex - BeginIndex;
2947  IntegerType *SubIntTy =
2948  IntTy ? Type::getIntNTy(IntTy->getContext(), Size * 8) : nullptr;
2949 
2950  // Reset the other pointer type to match the register type we're going to
2951  // use, but using the address space of the original other pointer.
2952  if (VecTy && !IsWholeAlloca) {
2953  if (NumElements == 1)
2954  OtherPtrTy = VecTy->getElementType();
2955  else
2956  OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2957 
2958  OtherPtrTy = OtherPtrTy->getPointerTo(OtherAS);
2959  } else if (IntTy && !IsWholeAlloca) {
2960  OtherPtrTy = SubIntTy->getPointerTo(OtherAS);
2961  } else {
2962  OtherPtrTy = NewAllocaTy->getPointerTo(OtherAS);
2963  }
2964 
2965  Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2966  OtherPtr->getName() + ".");
2967  unsigned SrcAlign = OtherAlign;
2968  Value *DstPtr = &NewAI;
2969  unsigned DstAlign = SliceAlign;
2970  if (!IsDest) {
2971  std::swap(SrcPtr, DstPtr);
2972  std::swap(SrcAlign, DstAlign);
2973  }
2974 
2975  Value *Src;
2976  if (VecTy && !IsWholeAlloca && !IsDest) {
2977  Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2978  Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
2979  } else if (IntTy && !IsWholeAlloca && !IsDest) {
2980  Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2981  Src = convertValue(DL, IRB, Src, IntTy);
2982  uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2983  Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
2984  } else {
2985  LoadInst *Load = IRB.CreateAlignedLoad(SrcPtr, SrcAlign, II.isVolatile(),
2986  "copyload");
2987  if (AATags)
2988  Load->setAAMetadata(AATags);
2989  Src = Load;
2990  }
2991 
2992  if (VecTy && !IsWholeAlloca && IsDest) {
2993  Value *Old =
2994  IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2995  Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
2996  } else if (IntTy && !IsWholeAlloca && IsDest) {
2997  Value *Old =
2998  IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2999  Old = convertValue(DL, IRB, Old, IntTy);
3000  uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
3001  Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
3002  Src = convertValue(DL, IRB, Src, NewAllocaTy);
3003  }
3004 
3005  StoreInst *Store = cast<StoreInst>(
3006  IRB.CreateAlignedStore(Src, DstPtr, DstAlign, II.isVolatile()));
3007  if (AATags)
3008  Store->setAAMetadata(AATags);
3009  LLVM_DEBUG(dbgs() << " to: " << *Store << "\n");
3010  return !II.isVolatile();
3011  }
3012 
3013  bool visitIntrinsicInst(IntrinsicInst &II) {
3014  assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
3015  II.getIntrinsicID() == Intrinsic::lifetime_end);
3016  LLVM_DEBUG(dbgs() << " original: " << II << "\n");
3017  assert(II.getArgOperand(1) == OldPtr);
3018 
3019  // Record this instruction for deletion.
3020  Pass.DeadInsts.insert(&II);
3021 
3022  // Lifetime intrinsics are only promotable if they cover the whole alloca.
3023  // Therefore, we drop lifetime intrinsics which don't cover the whole
3024  // alloca.
3025  // (In theory, intrinsics which partially cover an alloca could be
3026  // promoted, but PromoteMemToReg doesn't handle that case.)
3027  // FIXME: Check whether the alloca is promotable before dropping the
3028  // lifetime intrinsics?
3029  if (NewBeginOffset != NewAllocaBeginOffset ||
3030  NewEndOffset != NewAllocaEndOffset)
3031  return true;
3032 
3033  ConstantInt *Size =
3034  ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
3035  NewEndOffset - NewBeginOffset);
3036  Value *Ptr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
3037  Value *New;
3038  if (II.getIntrinsicID() == Intrinsic::lifetime_start)
3039  New = IRB.CreateLifetimeStart(Ptr, Size);
3040  else
3041  New = IRB.CreateLifetimeEnd(Ptr, Size);
3042 
3043  (void)New;
3044  LLVM_DEBUG(dbgs() << " to: " << *New << "\n");
3045 
3046  return true;
3047  }
3048 
3049  bool visitPHINode(PHINode &PN) {
3050  LLVM_DEBUG(dbgs() << " original: " << PN << "\n");
3051  assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
3052  assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
3053 
3054  // We would like to compute a new pointer in only one place, but have it be
3055  // as local as possible to the PHI. To do that, we re-use the location of
3056  // the old pointer, which necessarily must be in the right position to
3057  // dominate the PHI.
3058  IRBuilderTy PtrBuilder(IRB);
3059  if (isa<PHINode>(OldPtr))
3060  PtrBuilder.SetInsertPoint(&*OldPtr->getParent()->getFirstInsertionPt());
3061  else
3062  PtrBuilder.SetInsertPoint(OldPtr);
3063  PtrBuilder.SetCurrentDebugLocation(OldPtr->getDebugLoc());
3064 
3065  Value *NewPtr = getNewAllocaSlicePtr(PtrBuilder, OldPtr->getType());
3066  // Replace the operands which were using the old pointer.
3067  std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
3068 
3069  LLVM_DEBUG(dbgs() << " to: " << PN << "\n");
3070  deleteIfTriviallyDead(OldPtr);
3071 
3072  // PHIs can't be promoted on their own, but often can be speculated. We
3073  // check the speculation outside of the rewriter so that we see the
3074  // fully-rewritten alloca.
3075  PHIUsers.insert(&PN);
3076  return true;
3077  }
3078 
3079  bool visitSelectInst(SelectInst &SI) {
3080  LLVM_DEBUG(dbgs() << " original: " << SI << "\n");
3081  assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
3082  "Pointer isn't an operand!");
3083  assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
3084  assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
3085 
3086  Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
3087  // Replace the operands which were using the old pointer.
3088  if (SI.getOperand(1) == OldPtr)
3089  SI.setOperand(1, NewPtr);
3090  if (SI.getOperand(2) == OldPtr)
3091  SI.setOperand(2, NewPtr);
3092 
3093  LLVM_DEBUG(dbgs() << " to: " << SI << "\n");
3094  deleteIfTriviallyDead(OldPtr);
3095 
3096  // Selects can't be promoted on their own, but often can be speculated. We
3097  // check the speculation outside of the rewriter so that we see the
3098  // fully-rewritten alloca.
3099  SelectUsers.insert(&SI);
3100  return true;
3101  }
3102 };
3103 
3104 namespace {
3105 
3106 /// Visitor to rewrite aggregate loads and stores as scalar.
3107 ///
3108 /// This pass aggressively rewrites all aggregate loads and stores on
3109 /// a particular pointer (or any pointer derived from it which we can identify)
3110 /// with scalar loads and stores.
3111 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
3112  // Befriend the base class so it can delegate to private visit methods.
3113  friend class InstVisitor<AggLoadStoreRewriter, bool>;
3114 
3115  /// Queue of pointer uses to analyze and potentially rewrite.
3116  SmallVector<Use *, 8> Queue;
3117 
3118  /// Set to prevent us from cycling with phi nodes and loops.
3119  SmallPtrSet<User *, 8> Visited;
3120 
3121  /// The current pointer use being rewritten. This is used to dig up the used
3122  /// value (as opposed to the user).
3123  Use *U;
3124 
3125 public:
3126  /// Rewrite loads and stores through a pointer and all pointers derived from
3127  /// it.
3128  bool rewrite(Instruction &I) {
3129  LLVM_DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
3130  enqueueUsers(I);
3131  bool Changed = false;
3132  while (!Queue.empty()) {
3133  U = Queue.pop_back_val();
3134  Changed |= visit(cast<Instruction>(U->getUser()));
3135  }
3136  return Changed;
3137  }
3138 
3139 private:
3140  /// Enqueue all the users of the given instruction for further processing.
3141  /// This uses a set to de-duplicate users.
3142  void enqueueUsers(Instruction &I) {
3143  for (Use &U : I.uses())
3144  if (Visited.insert(U.getUser()).second)
3145  Queue.push_back(&U);
3146  }
3147 
3148  // Conservative default is to not rewrite anything.
3149  bool visitInstruction(Instruction &I) { return false; }
3150 
3151  /// Generic recursive split emission class.
3152  template <typename Derived> class OpSplitter {
3153  protected:
3154  /// The builder used to form new instructions.
3155  IRBuilderTy IRB;
3156 
3157  /// The indices which to be used with insert- or extractvalue to select the
3158  /// appropriate value within the aggregate.
3159  SmallVector<unsigned, 4> Indices;
3160 
3161  /// The indices to a GEP instruction which will move Ptr to the correct slot
3162  /// within the aggregate.
3163  SmallVector<Value *, 4> GEPIndices;
3164 
3165  /// The base pointer of the original op, used as a base for GEPing the
3166  /// split operations.
3167  Value *Ptr;
3168 
3169  /// Initialize the splitter with an insertion point, Ptr and start with a
3170  /// single zero GEP index.
3171  OpSplitter(Instruction *InsertionPoint, Value *Ptr)
3172  : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
3173 
3174  public:
3175  /// Generic recursive split emission routine.
3176  ///
3177  /// This method recursively splits an aggregate op (load or store) into
3178  /// scalar or vector ops. It splits recursively until it hits a single value
3179  /// and emits that single value operation via the template argument.
3180  ///
3181  /// The logic of this routine relies on GEPs and insertvalue and
3182  /// extractvalue all operating with the same fundamental index list, merely
3183  /// formatted differently (GEPs need actual values).
3184  ///
3185  /// \param Ty The type being split recursively into smaller ops.
3186  /// \param Agg The aggregate value being built up or stored, depending on
3187  /// whether this is splitting a load or a store respectively.
3188  void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
3189  if (Ty->isSingleValueType())
3190  return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
3191 
3192  if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
3193  unsigned OldSize = Indices.size();
3194  (void)OldSize;
3195  for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
3196  ++Idx) {
3197  assert(Indices.size() == OldSize && "Did not return to the old size");
3198  Indices.push_back(Idx);
3199  GEPIndices.push_back(IRB.getInt32(Idx));
3200  emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
3201  GEPIndices.pop_back();
3202  Indices.pop_back();
3203  }
3204  return;
3205  }
3206 
3207  if (StructType *STy = dyn_cast<StructType>(Ty)) {
3208  unsigned OldSize = Indices.size();
3209  (void)OldSize;
3210  for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
3211  ++Idx) {
3212  assert(Indices.size() == OldSize && "Did not return to the old size");
3213  Indices.push_back(Idx);
3214  GEPIndices.push_back(IRB.getInt32(Idx));
3215  emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
3216  GEPIndices.pop_back();
3217  Indices.pop_back();
3218  }
3219  return;
3220  }
3221 
3222  llvm_unreachable("Only arrays and structs are aggregate loadable types");
3223  }
3224  };
3225 
3226  struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
3227  AAMDNodes AATags;
3228 
3229  LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr, AAMDNodes AATags)
3230  : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr), AATags(AATags) {}
3231 
3232  /// Emit a leaf load of a single value. This is called at the leaves of the
3233  /// recursive emission to actually load values.
3234  void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3235  assert(Ty->isSingleValueType());
3236  // Load the single value and insert it using the indices.
3237  Value *GEP =
3238  IRB.CreateInBoundsGEP(nullptr, Ptr, GEPIndices, Name + ".gep");
3239  LoadInst *Load = IRB.CreateLoad(GEP, Name + ".load");
3240  if (AATags)
3241  Load->setAAMetadata(AATags);
3242  Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
3243  LLVM_DEBUG(dbgs() << " to: " << *Load << "\n");
3244  }
3245  };
3246 
3247  bool visitLoadInst(LoadInst &LI) {
3248  assert(LI.getPointerOperand() == *U);
3249  if (!LI.isSimple() || LI.getType()->isSingleValueType())
3250  return false;
3251 
3252  // We have an aggregate being loaded, split it apart.
3253  LLVM_DEBUG(dbgs() << " original: " << LI << "\n");
3254  AAMDNodes AATags;
3255  LI.getAAMetadata(AATags);
3256  LoadOpSplitter Splitter(&LI, *U, AATags);
3257  Value *V = UndefValue::get(LI.getType());
3258  Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
3259  LI.replaceAllUsesWith(V);
3260  LI.eraseFromParent();
3261  return true;
3262  }
3263 
3264  struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
3265  StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr, AAMDNodes AATags)
3266  : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr), AATags(AATags) {}
3267  AAMDNodes AATags;
3268 
3269  /// Emit a leaf store of a single value. This is called at the leaves of the
3270  /// recursive emission to actually produce stores.
3271  void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3272  assert(Ty->isSingleValueType());
3273  // Extract the single value and store it using the indices.
3274  //
3275  // The gep and extractvalue values are factored out of the CreateStore
3276  // call to make the output independent of the argument evaluation order.
3277  Value *ExtractValue =
3278  IRB.CreateExtractValue(Agg, Indices, Name + ".extract");
3279  Value *InBoundsGEP =
3280  IRB.CreateInBoundsGEP(nullptr, Ptr, GEPIndices, Name + ".gep");
3281  StoreInst *Store = IRB.CreateStore(ExtractValue, InBoundsGEP);
3282  if (AATags)
3283  Store->setAAMetadata(AATags);
3284  LLVM_DEBUG(dbgs() << " to: " << *Store << "\n");
3285  }
3286  };
3287 
3288  bool visitStoreInst(StoreInst &SI) {
3289  if (!SI.isSimple() || SI.getPointerOperand() != *U)
3290  return false;
3291  Value *V = SI.getValueOperand();
3292  if (V->getType()->isSingleValueType())
3293  return false;
3294 
3295  // We have an aggregate being stored, split it apart.
3296  LLVM_DEBUG(dbgs() << " original: " << SI << "\n");
3297  AAMDNodes AATags;
3298  SI.getAAMetadata(AATags);
3299  StoreOpSplitter Splitter(&SI, *U, AATags);
3300  Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
3301  SI.eraseFromParent();
3302  return true;
3303  }
3304 
3305  bool visitBitCastInst(BitCastInst &BC) {
3306  enqueueUsers(BC);
3307  return false;
3308  }
3309 
3310  bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
3311  enqueueUsers(GEPI);
3312  return false;
3313  }
3314 
3315  bool visitPHINode(PHINode &PN) {
3316  enqueueUsers(PN);
3317  return false;
3318  }
3319 
3320  bool visitSelectInst(SelectInst &SI) {
3321  enqueueUsers(SI);
3322  return false;
3323  }
3324 };
3325 
3326 } // end anonymous namespace
3327 
3328 /// Strip aggregate type wrapping.
3329 ///
3330 /// This removes no-op aggregate types wrapping an underlying type. It will
3331 /// strip as many layers of types as it can without changing either the type
3332 /// size or the allocated size.
3334  if (Ty->isSingleValueType())
3335  return Ty;
3336 
3337  uint64_t AllocSize = DL.getTypeAllocSize(Ty);
3338  uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
3339 
3340  Type *InnerTy;
3341  if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
3342  InnerTy = ArrTy->getElementType();
3343  } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
3344  const StructLayout *SL = DL.getStructLayout(STy);
3345  unsigned Index = SL->getElementContainingOffset(0);
3346  InnerTy = STy->getElementType(Index);
3347  } else {
3348  return Ty;
3349  }
3350 
3351  if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
3352  TypeSize > DL.getTypeSizeInBits(InnerTy))
3353  return Ty;
3354 
3355  return stripAggregateTypeWrapping(DL, InnerTy);
3356 }
3357 
3358 /// Try to find a partition of the aggregate type passed in for a given
3359 /// offset and size.
3360 ///
3361 /// This recurses through the aggregate type and tries to compute a subtype
3362 /// based on the offset and size. When the offset and size span a sub-section
3363 /// of an array, it will even compute a new array type for that sub-section,
3364 /// and the same for structs.
3365 ///
3366 /// Note that this routine is very strict and tries to find a partition of the
3367 /// type which produces the *exact* right offset and size. It is not forgiving
3368 /// when the size or offset cause either end of type-based partition to be off.
3369 /// Also, this is a best-effort routine. It is reasonable to give up and not
3370 /// return a type if necessary.
3371 static Type *getTypePartition(const DataLayout &DL, Type *Ty, uint64_t Offset,
3372  uint64_t Size) {
3373  if (Offset == 0 && DL.getTypeAllocSize(Ty) == Size)
3374  return stripAggregateTypeWrapping(DL, Ty);
3375  if (Offset > DL.getTypeAllocSize(Ty) ||
3376  (DL.getTypeAllocSize(Ty) - Offset) < Size)
3377  return nullptr;
3378 
3379  if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3380  Type *ElementTy = SeqTy->getElementType();
3381  uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3382  uint64_t NumSkippedElements = Offset / ElementSize;
3383  if (NumSkippedElements >= SeqTy->getNumElements())
3384  return nullptr;
3385  Offset -= NumSkippedElements * ElementSize;
3386 
3387  // First check if we need to recurse.
3388  if (Offset > 0 || Size < ElementSize) {
3389  // Bail if the partition ends in a different array element.
3390  if ((Offset + Size) > ElementSize)
3391  return nullptr;
3392  // Recurse through the element type trying to peel off offset bytes.
3393  return getTypePartition(DL, ElementTy, Offset, Size);
3394  }
3395  assert(Offset == 0);
3396 
3397  if (Size == ElementSize)
3398  return stripAggregateTypeWrapping(DL, ElementTy);
3399  assert(Size > ElementSize);
3400  uint64_t NumElements = Size / ElementSize;
3401  if (NumElements * ElementSize != Size)
3402  return nullptr;
3403  return ArrayType::get(ElementTy, NumElements);
3404  }
3405 
3406  StructType *STy = dyn_cast<StructType>(Ty);
3407  if (!STy)
3408  return nullptr;
3409 
3410  const StructLayout *SL = DL.getStructLayout(STy);
3411  if (Offset >= SL->getSizeInBytes())
3412  return nullptr;
3413  uint64_t EndOffset = Offset + Size;
3414  if (EndOffset > SL->getSizeInBytes())
3415  return nullptr;
3416 
3417  unsigned Index = SL->getElementContainingOffset(Offset);
3418  Offset -= SL->getElementOffset(Index);
3419 
3420  Type *ElementTy = STy->getElementType(Index);
3421  uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3422  if (Offset >= ElementSize)
3423  return nullptr; // The offset points into alignment padding.
3424 
3425  // See if any partition must be contained by the element.
3426  if (Offset > 0 || Size < ElementSize) {
3427  if ((Offset + Size) > ElementSize)
3428  return nullptr;
3429  return getTypePartition(DL, ElementTy, Offset, Size);
3430  }
3431  assert(Offset == 0);
3432 
3433  if (Size == ElementSize)
3434  return stripAggregateTypeWrapping(DL, ElementTy);
3435 
3437  EE = STy->element_end();
3438  if (EndOffset < SL->getSizeInBytes()) {
3439  unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3440  if (Index == EndIndex)
3441  return nullptr; // Within a single element and its padding.
3442 
3443  // Don't try to form "natural" types if the elements don't line up with the
3444  // expected size.
3445  // FIXME: We could potentially recurse down through the last element in the
3446  // sub-struct to find a natural end point.
3447  if (SL->getElementOffset(EndIndex) != EndOffset)
3448  return nullptr;
3449 
3450  assert(Index < EndIndex);
3451  EE = STy->element_begin() + EndIndex;
3452  }
3453 
3454  // Try to build up a sub-structure.
3455  StructType *SubTy =
3456  StructType::get(STy->getContext(), makeArrayRef(EI, EE), STy->isPacked());
3457  const StructLayout *SubSL = DL.getStructLayout(SubTy);
3458  if (Size != SubSL->getSizeInBytes())
3459  return nullptr; // The sub-struct doesn't have quite the size needed.
3460 
3461  return SubTy;
3462 }
3463 
3464 /// Pre-split loads and stores to simplify rewriting.
3465 ///
3466 /// We want to break up the splittable load+store pairs as much as
3467 /// possible. This is important to do as a preprocessing step, as once we
3468 /// start rewriting the accesses to partitions of the alloca we lose the
3469 /// necessary information to correctly split apart paired loads and stores
3470 /// which both point into this alloca. The case to consider is something like
3471 /// the following:
3472 ///
3473 /// %a = alloca [12 x i8]
3474 /// %gep1 = getelementptr [12 x i8]* %a, i32 0, i32 0
3475 /// %gep2 = getelementptr [12 x i8]* %a, i32 0, i32 4
3476 /// %gep3 = getelementptr [12 x i8]* %a, i32 0, i32 8
3477 /// %iptr1 = bitcast i8* %gep1 to i64*
3478 /// %iptr2 = bitcast i8* %gep2 to i64*
3479 /// %fptr1 = bitcast i8* %gep1 to float*
3480 /// %fptr2 = bitcast i8* %gep2 to float*
3481 /// %fptr3 = bitcast i8* %gep3 to float*
3482 /// store float 0.0, float* %fptr1
3483 /// store float 1.0, float* %fptr2
3484 /// %v = load i64* %iptr1
3485 /// store i64 %v, i64* %iptr2
3486 /// %f1 = load float* %fptr2
3487 /// %f2 = load float* %fptr3
3488 ///
3489 /// Here we want to form 3 partitions of the alloca, each 4 bytes large, and
3490 /// promote everything so we recover the 2 SSA values that should have been
3491 /// there all along.
3492 ///
3493 /// \returns true if any changes are made.
3494 bool SROA::presplitLoadsAndStores(AllocaInst &AI, AllocaSlices &AS) {
3495  LLVM_DEBUG(dbgs() << "Pre-splitting loads and stores\n");
3496 
3497  // Track the loads and stores which are candidates for pre-splitting here, in
3498  // the order they first appear during the partition scan. These give stable
3499  // iteration order and a basis for tracking which loads and stores we
3500  // actually split.
3503 
3504  // We need to accumulate the splits required of each load or store where we
3505  // can find them via a direct lookup. This is important to cross-check loads
3506  // and stores against each other. We also track the slice so that we can kill
3507  // all the slices that end up split.
3508  struct SplitOffsets {
3509  Slice *S;
3510  std::vector<uint64_t> Splits;
3511  };
3513 
3514  // Track loads out of this alloca which cannot, for any reason, be pre-split.
3515  // This is important as we also cannot pre-split stores of those loads!
3516  // FIXME: This is all pretty gross. It means that we can be more aggressive
3517  // in pre-splitting when the load feeding the store happens to come from
3518  // a separate alloca. Put another way, the effectiveness of SROA would be
3519  // decreased by a frontend which just concatenated all of its local allocas
3520  // into one big flat alloca. But defeating such patterns is exactly the job
3521  // SROA is tasked with! Sadly, to not have this discrepancy we would have
3522  // change store pre-splitting to actually force pre-splitting of the load
3523  // that feeds it *and all stores*. That makes pre-splitting much harder, but
3524  // maybe it would make it more principled?
3525  SmallPtrSet<LoadInst *, 8> UnsplittableLoads;
3526 
3527  LLVM_DEBUG(dbgs() << " Searching for candidate loads and stores\n");
3528  for (auto &P : AS.partitions()) {
3529  for (Slice &S : P) {
3530  Instruction *I = cast<Instruction>(S.getUse()->getUser());
3531  if (!S.isSplittable() || S.endOffset() <= P.endOffset()) {
3532  // If this is a load we have to track that it can't participate in any
3533  // pre-splitting. If this is a store of a load we have to track that
3534  // that load also can't participate in any pre-splitting.
3535  if (auto *LI = dyn_cast<LoadInst>(I))
3536  UnsplittableLoads.insert(LI);
3537  else if (auto *SI = dyn_cast<StoreInst>(I))
3538  if (auto *LI = dyn_cast<LoadInst>(SI->getValueOperand()))
3539  UnsplittableLoads.insert(LI);
3540  continue;
3541  }
3542  assert(P.endOffset() > S.beginOffset() &&
3543  "Empty or backwards partition!");
3544 
3545  // Determine if this is a pre-splittable slice.
3546  if (auto *LI = dyn_cast<LoadInst>(I)) {
3547  assert(!LI->isVolatile() && "Cannot split volatile loads!");
3548 
3549  // The load must be used exclusively to store into other pointers for
3550  // us to be able to arbitrarily pre-split it. The stores must also be
3551  // simple to avoid changing semantics.
3552  auto IsLoadSimplyStored = [](LoadInst *LI) {
3553  for (User *LU : LI->users()) {
3554  auto *SI = dyn_cast<StoreInst>(LU);
3555  if (!SI || !SI->isSimple())
3556  return false;
3557  }
3558  return true;
3559  };
3560  if (!IsLoadSimplyStored(LI)) {
3561  UnsplittableLoads.insert(LI);
3562  continue;
3563  }
3564 
3565  Loads.push_back(LI);
3566  } else if (auto *SI = dyn_cast<StoreInst>(I)) {
3567  if (S.getUse() != &SI->getOperandUse(SI->getPointerOperandIndex()))
3568  // Skip stores *of* pointers. FIXME: This shouldn't even be possible!
3569  continue;
3570  auto *StoredLoad = dyn_cast<LoadInst>(SI->getValueOperand());
3571  if (!StoredLoad || !StoredLoad->isSimple())
3572  continue;
3573  assert(!SI->isVolatile() && "Cannot split volatile stores!");
3574 
3575  Stores.push_back(SI);
3576  } else {
3577  // Other uses cannot be pre-split.
3578  continue;
3579  }
3580 
3581  // Record the initial split.
3582  LLVM_DEBUG(dbgs() << " Candidate: " << *I << "\n");
3583  auto &Offsets = SplitOffsetsMap[I];
3584  assert(Offsets.Splits.empty() &&
3585  "Should not have splits the first time we see an instruction!");
3586  Offsets.S = &S;
3587  Offsets.Splits.push_back(P.endOffset() - S.beginOffset());
3588  }
3589 
3590  // Now scan the already split slices, and add a split for any of them which
3591  // we're going to pre-split.
3592  for (Slice *S : P.splitSliceTails()) {
3593  auto SplitOffsetsMapI =
3594  SplitOffsetsMap.find(cast<Instruction>(S->getUse()->getUser()));
3595  if (SplitOffsetsMapI == SplitOffsetsMap.end())
3596  continue;
3597  auto &Offsets = SplitOffsetsMapI->second;
3598 
3599  assert(Offsets.S == S && "Found a mismatched slice!");
3600  assert(!Offsets.Splits.empty() &&
3601  "Cannot have an empty set of splits on the second partition!");
3602  assert(Offsets.Splits.back() ==
3603  P.beginOffset() - Offsets.S->beginOffset() &&
3604  "Previous split does not end where this one begins!");
3605 
3606  // Record each split. The last partition's end isn't needed as the size
3607  // of the slice dictates that.
3608  if (S->endOffset() > P.endOffset())
3609  Offsets.Splits.push_back(P.endOffset() - Offsets.S->beginOffset());
3610  }
3611  }
3612 
3613  // We may have split loads where some of their stores are split stores. For
3614  // such loads and stores, we can only pre-split them if their splits exactly
3615  // match relative to their starting offset. We have to verify this prior to
3616  // any rewriting.
3617  Stores.erase(
3618  llvm::remove_if(Stores,
3619  [&UnsplittableLoads, &SplitOffsetsMap](StoreInst *SI) {
3620  // Lookup the load we are storing in our map of split
3621  // offsets.
3622  auto *LI = cast<LoadInst>(SI->getValueOperand());
3623  // If it was completely unsplittable, then we're done,
3624  // and this store can't be pre-split.
3625  if (UnsplittableLoads.count(LI))
3626  return true;
3627 
3628  auto LoadOffsetsI = SplitOffsetsMap.find(LI);
3629  if (LoadOffsetsI == SplitOffsetsMap.end())
3630  return false; // Unrelated loads are definitely safe.
3631  auto &LoadOffsets = LoadOffsetsI->second;
3632 
3633  // Now lookup the store's offsets.
3634  auto &StoreOffsets = SplitOffsetsMap[SI];
3635 
3636  // If the relative offsets of each split in the load and
3637  // store match exactly, then we can split them and we
3638  // don't need to remove them here.
3639  if (LoadOffsets.Splits == StoreOffsets.Splits)
3640  return false;
3641 
3642  LLVM_DEBUG(
3643  dbgs()
3644  << " Mismatched splits for load and store:\n"
3645  << " " << *LI << "\n"
3646  << " " << *SI << "\n");
3647 
3648  // We've found a store and load that we need to split
3649  // with mismatched relative splits. Just give up on them
3650  // and remove both instructions from our list of
3651  // candidates.
3652  UnsplittableLoads.insert(LI);
3653  return true;
3654  }),
3655  Stores.end());
3656  // Now we have to go *back* through all the stores, because a later store may
3657  // have caused an earlier store's load to become unsplittable and if it is
3658  // unsplittable for the later store, then we can't rely on it being split in
3659  // the earlier store either.
3660  Stores.erase(llvm::remove_if(Stores,
3661  [&UnsplittableLoads](StoreInst *SI) {
3662  auto *LI =
3663  cast<LoadInst>(SI->getValueOperand());
3664  return UnsplittableLoads.count(LI);
3665  }),
3666  Stores.end());
3667  // Once we've established all the loads that can't be split for some reason,
3668  // filter any that made it into our list out.
3669  Loads.erase(llvm::remove_if(Loads,
3670  [&UnsplittableLoads](LoadInst *LI) {
3671  return UnsplittableLoads.count(LI);
3672  }),
3673  Loads.end());
3674 
3675  // If no loads or stores are left, there is no pre-splitting to be done for
3676  // this alloca.
3677  if (Loads.empty() && Stores.empty())
3678  return false;
3679 
3680  // From here on, we can't fail and will be building new accesses, so rig up
3681  // an IR builder.
3682  IRBuilderTy IRB(&AI);
3683 
3684  // Collect the new slices which we will merge into the alloca slices.
3685  SmallVector<Slice, 4> NewSlices;
3686 
3687  // Track any allocas we end up splitting loads and stores for so we iterate
3688  // on them.
3689  SmallPtrSet<AllocaInst *, 4> ResplitPromotableAllocas;
3690 
3691  // At this point, we have collected all of the loads and stores we can
3692  // pre-split, and the specific splits needed for them. We actually do the
3693  // splitting in a specific order in order to handle when one of the loads in
3694  // the value operand to one of the stores.
3695  //
3696  // First, we rewrite all of the split loads, and just accumulate each split
3697  // load in a parallel structure. We also build the slices for them and append
3698  // them to the alloca slices.
3700  std::vector<LoadInst *> SplitLoads;
3701  const DataLayout &DL = AI.getModule()->getDataLayout();
3702  for (LoadInst *LI : Loads) {
3703  SplitLoads.clear();
3704 
3705  IntegerType *Ty = cast<IntegerType>(LI->getType());
3706  uint64_t LoadSize = Ty->getBitWidth() / 8;
3707  assert(LoadSize > 0 && "Cannot have a zero-sized integer load!");
3708 
3709  auto &Offsets = SplitOffsetsMap[LI];
3710  assert(LoadSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&
3711  "Slice size should always match load size exactly!");
3712  uint64_t BaseOffset = Offsets.S->beginOffset();
3713  assert(BaseOffset + LoadSize > BaseOffset &&
3714  "Cannot represent alloca access size using 64-bit integers!");
3715 
3716  Instruction *BasePtr = cast<Instruction>(LI->getPointerOperand());
3717  IRB.SetInsertPoint(LI);
3718 
3719  LLVM_DEBUG(dbgs() << " Splitting load: " << *LI << "\n");
3720 
3721  uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
3722  int Idx = 0, Size = Offsets.Splits.size();
3723  for (;;) {
3724  auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
3725  auto AS = LI->getPointerAddressSpace();
3726  auto *PartPtrTy = PartTy->getPointerTo(AS);
3727  LoadInst *PLoad = IRB.CreateAlignedLoad(
3728  getAdjustedPtr(IRB, DL, BasePtr,
3729  APInt(DL.getIndexSizeInBits(AS), PartOffset),
3730  PartPtrTy, BasePtr->getName() + "."),
3731  getAdjustedAlignment(LI, PartOffset, DL), /*IsVolatile*/ false,
3732  LI->getName());
3734 
3735  // Append this load onto the list of split loads so we can find it later
3736  // to rewrite the stores.
3737  SplitLoads.push_back(PLoad);
3738 
3739  // Now build a new slice for the alloca.
3740  NewSlices.push_back(
3741  Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
3742  &PLoad->getOperandUse(PLoad->getPointerOperandIndex()),
3743  /*IsSplittable*/ false));
3744  LLVM_DEBUG(dbgs() << " new slice [" << NewSlices.back().beginOffset()
3745  << ", " << NewSlices.back().endOffset()
3746  << "): " << *PLoad << "\n");
3747 
3748  // See if we've handled all the splits.
3749  if (Idx >= Size)
3750  break;
3751 
3752  // Setup the next partition.
3753  PartOffset = Offsets.Splits[Idx];
3754  ++Idx;
3755  PartSize = (Idx < Size ? Offsets.Splits[Idx] : LoadSize) - PartOffset;
3756  }
3757 
3758  // Now that we have the split loads, do the slow walk over all uses of the
3759  // load and rewrite them as split stores, or save the split loads to use
3760  // below if the store is going to be split there anyways.
3761  bool DeferredStores = false;
3762  for (User *LU : LI->users()) {
3763  StoreInst *SI = cast<StoreInst>(LU);
3764  if (!Stores.empty() && SplitOffsetsMap.count(SI)) {
3765  DeferredStores = true;
3766  LLVM_DEBUG(dbgs() << " Deferred splitting of store: " << *SI
3767  << "\n");
3768  continue;
3769  }
3770 
3771  Value *StoreBasePtr = SI->getPointerOperand();
3772  IRB.SetInsertPoint(SI);
3773 
3774  LLVM_DEBUG(dbgs() << " Splitting store of load: " << *SI << "\n");
3775 
3776  for (int Idx = 0, Size = SplitLoads.size(); Idx < Size; ++Idx) {
3777  LoadInst *PLoad = SplitLoads[Idx];
3778  uint64_t PartOffset = Idx == 0 ? 0 : Offsets.Splits[Idx - 1];
3779  auto *PartPtrTy =
3780  PLoad->getType()->getPointerTo(SI->getPointerAddressSpace());
3781 
3782  auto AS = SI->getPointerAddressSpace();
3783  StoreInst *PStore = IRB.CreateAlignedStore(
3784  PLoad,
3785  getAdjustedPtr(IRB, DL, StoreBasePtr,
3786  APInt(DL.getIndexSizeInBits(AS), PartOffset),
3787  PartPtrTy, StoreBasePtr->getName() + "."),
3788  getAdjustedAlignment(SI, PartOffset, DL), /*IsVolatile*/ false);
3790  LLVM_DEBUG(dbgs() << " +" << PartOffset << ":" << *PStore << "\n");
3791  }
3792 
3793  // We want to immediately iterate on any allocas impacted by splitting
3794  // this store, and we have to track any promotable alloca (indicated by
3795  // a direct store) as needing to be resplit because it is no longer
3796  // promotable.
3797  if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(StoreBasePtr)) {
3798  ResplitPromotableAllocas.insert(OtherAI);
3799  Worklist.insert(OtherAI);
3800  } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
3801  StoreBasePtr->stripInBoundsOffsets())) {
3802  Worklist.insert(OtherAI);
3803  }
3804 
3805  // Mark the original store as dead.
3806  DeadInsts.insert(SI);
3807  }
3808 
3809  // Save the split loads if there are deferred stores among the users.
3810  if (DeferredStores)
3811  SplitLoadsMap.insert(std::make_pair(LI, std::move(SplitLoads)));
3812 
3813  // Mark the original load as dead and kill the original slice.
3814  DeadInsts.insert(LI);
3815  Offsets.S->kill();
3816  }
3817 
3818  // Second, we rewrite all of the split stores. At this point, we know that
3819  // all loads from this alloca have been split already. For stores of such
3820  // loads, we can simply look up the pre-existing split loads. For stores of
3821  // other loads, we split those loads first and then write split stores of
3822  // them.
3823  for (StoreInst *SI : Stores) {
3824  auto *LI = cast<LoadInst>(SI->getValueOperand());
3825  IntegerType *Ty = cast<IntegerType>(LI->getType());
3826  uint64_t StoreSize = Ty->getBitWidth() / 8;
3827  assert(StoreSize > 0 && "Cannot have a zero-sized integer store!");
3828 
3829  auto &Offsets = SplitOffsetsMap[SI];
3830  assert(StoreSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&
3831  "Slice size should always match load size exactly!");
3832  uint64_t BaseOffset = Offsets.S->beginOffset();
3833  assert(BaseOffset + StoreSize > BaseOffset &&
3834  "Cannot represent alloca access size using 64-bit integers!");
3835 
3836  Value *LoadBasePtr = LI->getPointerOperand();
3837  Instruction *StoreBasePtr = cast<Instruction>(SI->getPointerOperand());
3838 
3839  LLVM_DEBUG(dbgs() << " Splitting store: " << *SI << "\n");
3840 
3841  // Check whether we have an already split load.
3842  auto SplitLoadsMapI = SplitLoadsMap.find(LI);
3843  std::vector<LoadInst *> *SplitLoads = nullptr;
3844  if (SplitLoadsMapI != SplitLoadsMap.end()) {
3845  SplitLoads = &SplitLoadsMapI->second;
3846  assert(SplitLoads->size() == Offsets.Splits.size() + 1 &&
3847  "Too few split loads for the number of splits in the store!");
3848  } else {
3849  LLVM_DEBUG(dbgs() << " of load: " << *LI << "\n");
3850  }
3851 
3852  uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
3853  int Idx = 0, Size = Offsets.Splits.size();
3854  for (;;) {
3855  auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
3856  auto *LoadPartPtrTy = PartTy->getPointerTo(LI->getPointerAddressSpace());
3857  auto *StorePartPtrTy = PartTy->getPointerTo(SI->getPointerAddressSpace());
3858 
3859  // Either lookup a split load or create one.
3860  LoadInst *PLoad;
3861  if (SplitLoads) {
3862  PLoad = (*SplitLoads)[Idx];
3863  } else {
3864  IRB.SetInsertPoint(LI);
3865  auto AS = LI->getPointerAddressSpace();
3866  PLoad = IRB.CreateAlignedLoad(
3867  getAdjustedPtr(IRB, DL, LoadBasePtr,
3868  APInt(DL.getIndexSizeInBits(AS), PartOffset),
3869  LoadPartPtrTy, LoadBasePtr->getName() + "."),
3870  getAdjustedAlignment(LI, PartOffset, DL), /*IsVolatile*/ false,
3871  LI->getName());
3872  }
3873 
3874  // And store this partition.
3875  IRB.SetInsertPoint(SI);
3876  auto AS = SI->getPointerAddressSpace();
3877  StoreInst *PStore = IRB.CreateAlignedStore(
3878  PLoad,
3879  getAdjustedPtr(IRB, DL, StoreBasePtr,
3880  APInt(DL.getIndexSizeInBits(AS), PartOffset),
3881  StorePartPtrTy, StoreBasePtr->getName() + "."),
3882  getAdjustedAlignment(SI, PartOffset, DL), /*IsVolatile*/ false);
3883 
3884  // Now build a new slice for the alloca.
3885  NewSlices.push_back(
3886  Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
3887  &PStore->getOperandUse(PStore->getPointerOperandIndex()),
3888  /*IsSplittable*/ false));
3889  LLVM_DEBUG(dbgs() << " new slice [" << NewSlices.back().beginOffset()
3890  << ", " << NewSlices.back().endOffset()
3891  << "): " << *PStore << "\n");
3892  if (!SplitLoads) {
3893  LLVM_DEBUG(dbgs() << " of split load: " << *PLoad << "\n");
3894  }
3895 
3896  // See if we've finished all the splits.
3897  if (Idx >= Size)
3898  break;
3899 
3900  // Setup the next partition.
3901  PartOffset = Offsets.Splits[Idx];
3902  ++Idx;
3903  PartSize = (Idx < Size ? Offsets.Splits[Idx] : StoreSize) - PartOffset;
3904  }
3905 
3906  // We want to immediately iterate on any allocas impacted by splitting
3907  // this load, which is only relevant if it isn't a load of this alloca and
3908  // thus we didn't already split the loads above. We also have to keep track
3909  // of any promotable allocas we split loads on as they can no longer be
3910  // promoted.
3911  if (!SplitLoads) {
3912  if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(LoadBasePtr)) {
3913  assert(OtherAI != &AI && "We can't re-split our own alloca!");
3914  ResplitPromotableAllocas.insert(OtherAI);
3915  Worklist.insert(OtherAI);
3916  } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
3917  LoadBasePtr->stripInBoundsOffsets())) {
3918  assert(OtherAI != &AI && "We can't re-split our own alloca!");
3919  Worklist.insert(OtherAI);
3920  }
3921  }
3922 
3923  // Mark the original store as dead now that we've split it up and kill its
3924  // slice. Note that we leave the original load in place unless this store
3925  // was its only use. It may in turn be split up if it is an alloca load
3926  // for some other alloca, but it may be a normal load. This may introduce
3927  // redundant loads, but where those can be merged the rest of the optimizer
3928  // should handle the merging, and this uncovers SSA splits which is more
3929  // important. In practice, the original loads will almost always be fully
3930  // split and removed eventually, and the splits will be merged by any
3931  // trivial CSE, including instcombine.
3932  if (LI->hasOneUse()) {
3933  assert(*LI->user_begin() == SI && "Single use isn't this store!");
3934  DeadInsts.insert(LI);
3935  }
3936  DeadInsts.insert(SI);
3937  Offsets.S->kill();
3938  }
3939 
3940  // Remove the killed slices that have ben pre-split.
3941  AS.erase(llvm::remove_if(AS, [](const Slice &S) { return S.isDead(); }),
3942  AS.end());
3943 
3944  // Insert our new slices. This will sort and merge them into the sorted
3945  // sequence.
3946  AS.insert(NewSlices);
3947 
3948  LLVM_DEBUG(dbgs() << " Pre-split slices:\n");
3949 #ifndef NDEBUG
3950  for (auto I = AS.begin(), E = AS.end(); I != E; ++I)
3951  LLVM_DEBUG(AS.print(dbgs(), I, " "));
3952 #endif
3953 
3954  // Finally, don't try to promote any allocas that new require re-splitting.
3955  // They have already been added to the worklist above.
3956  PromotableAllocas.erase(
3958  PromotableAllocas,
3959  [&](AllocaInst *AI) { return ResplitPromotableAllocas.count(AI); }),
3960  PromotableAllocas.end());
3961 
3962  return true;
3963 }
3964 
3965 /// Rewrite an alloca partition's users.
3966 ///
3967 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3968 /// to rewrite uses of an alloca partition to be conducive for SSA value
3969 /// promotion. If the partition needs a new, more refined alloca, this will
3970 /// build that new alloca, preserving as much type information as possible, and
3971 /// rewrite the uses of the old alloca to point at the new one and have the
3972 /// appropriate new offsets. It also evaluates how successful the rewrite was
3973 /// at enabling promotion and if it was successful queues the alloca to be
3974 /// promoted.
3975 AllocaInst *SROA::rewritePartition(AllocaInst &AI, AllocaSlices &AS,
3976  Partition &P) {
3977  // Try to compute a friendly type for this partition of the alloca. This
3978  // won't always succeed, in which case we fall back to a legal integer type
3979  // or an i8 array of an appropriate size.
3980  Type *SliceTy = nullptr;
3981  const DataLayout &DL = AI.getModule()->getDataLayout();
3982  if (Type *CommonUseTy = findCommonType(P.begin(), P.end(), P.endOffset()))
3983  if (DL.getTypeAllocSize(CommonUseTy) >= P.size())
3984  SliceTy = CommonUseTy;
3985  if (!SliceTy)
3986  if (Type *TypePartitionTy = getTypePartition(DL, AI.getAllocatedType(),
3987  P.beginOffset(), P.size()))
3988  SliceTy = TypePartitionTy;
3989  if ((!SliceTy || (SliceTy->isArrayTy() &&
3990  SliceTy->getArrayElementType()->isIntegerTy())) &&
3991  DL.isLegalInteger(P.size() * 8))
3992  SliceTy = Type::getIntNTy(*C, P.size() * 8);
3993  if (!SliceTy)
3994  SliceTy = ArrayType::get(Type::getInt8Ty(*C), P.size());
3995  assert(DL.getTypeAllocSize(SliceTy) >= P.size());
3996 
3997  bool IsIntegerPromotable = isIntegerWideningViable(P, SliceTy, DL);
3998 
3999  VectorType *VecTy =
4000  IsIntegerPromotable ? nullptr : isVectorPromotionViable(P, DL);
4001  if (VecTy)
4002  SliceTy = VecTy;
4003 
4004  // Check for the case where we're going to rewrite to a new alloca of the
4005  // exact same type as the original, and with the same access offsets. In that
4006  // case, re-use the existing alloca, but still run through the rewriter to
4007  // perform phi and select speculation.
4008  // P.beginOffset() can be non-zero even with the same type in a case with
4009  // out-of-bounds access (e.g. @PR35657 function in SROA/basictest.ll).
4010  AllocaInst *NewAI;
4011  if (SliceTy == AI.getAllocatedType() && P.beginOffset() == 0) {
4012  NewAI = &AI;
4013  // FIXME: We should be able to bail at this point with "nothing changed".
4014  // FIXME: We might want to defer PHI speculation until after here.
4015  // FIXME: return nullptr;
4016  } else {
4017  unsigned Alignment = AI.getAlignment();
4018  if (!Alignment) {
4019  // The minimum alignment which users can rely on when the explicit
4020  // alignment is omitted or zero is that required by the ABI for this
4021  // type.
4022  Alignment = DL.getABITypeAlignment(AI.getAllocatedType());
4023  }
4024  Alignment = MinAlign(Alignment, P.beginOffset());
4025  // If we will get at least this much alignment from the type alone, leave
4026  // the alloca's alignment unconstrained.
4027  if (Alignment <= DL.getABITypeAlignment(SliceTy))
4028  Alignment = 0;
4029  NewAI = new AllocaInst(
4030  SliceTy, AI.getType()->getAddressSpace(), nullptr, Alignment,
4031  AI.getName() + ".sroa." + Twine(P.begin() - AS.begin()), &AI);
4032  // Copy the old AI debug location over to the new one.
4033  NewAI->setDebugLoc(AI.getDebugLoc());
4034  ++NumNewAllocas;
4035  }
4036 
4037  LLVM_DEBUG(dbgs() << "Rewriting alloca partition "
4038  << "[" << P.beginOffset() << "," << P.endOffset()
4039  << ") to: " << *NewAI << "\n");
4040 
4041  // Track the high watermark on the worklist as it is only relevant for
4042  // promoted allocas. We will reset it to this point if the alloca is not in
4043  // fact scheduled for promotion.
4044  unsigned PPWOldSize = PostPromotionWorklist.size();
4045  unsigned NumUses = 0;
4047  SmallSetVector<SelectInst *, 8> SelectUsers;
4048 
4049  AllocaSliceRewriter Rewriter(DL, AS, *this, AI, *NewAI, P.beginOffset(),
4050  P.endOffset(), IsIntegerPromotable, VecTy,
4051  PHIUsers, SelectUsers);
4052  bool Promotable = true;
4053  for (Slice *S : P.splitSliceTails()) {
4054  Promotable &= Rewriter.visit(S);
4055  ++NumUses;
4056  }
4057  for (Slice &S : P) {
4058  Promotable &= Rewriter.visit(&S);
4059  ++NumUses;
4060  }
4061 
4062  NumAllocaPartitionUses += NumUses;
4063  MaxUsesPerAllocaPartition.updateMax(NumUses);
4064 
4065  // Now that we've processed all the slices in the new partition, check if any
4066  // PHIs or Selects would block promotion.
4067  for (PHINode *PHI : PHIUsers)
4068  if (!isSafePHIToSpeculate(*PHI)) {
4069  Promotable = false;
4070  PHIUsers.clear();
4071  SelectUsers.clear();
4072  break;
4073  }
4074 
4075  for (SelectInst *Sel : SelectUsers)
4076  if (!isSafeSelectToSpeculate(*Sel)) {
4077  Promotable = false;
4078  PHIUsers.clear();
4079  SelectUsers.clear();
4080  break;
4081  }
4082 
4083  if (Promotable) {
4084  if (PHIUsers.empty() && SelectUsers.empty()) {
4085  // Promote the alloca.
4086  PromotableAllocas.push_back(NewAI);
4087  } else {
4088  // If we have either PHIs or Selects to speculate, add them to those
4089  // worklists and re-queue the new alloca so that we promote in on the
4090  // next iteration.
4091  for (PHINode *PHIUser : PHIUsers)
4092  SpeculatablePHIs.insert(PHIUser);
4093  for (SelectInst *SelectUser : SelectUsers)
4094  SpeculatableSelects.insert(SelectUser);
4095  Worklist.insert(NewAI);
4096  }
4097  } else {
4098  // Drop any post-promotion work items if promotion didn't happen.
4099  while (PostPromotionWorklist.size() > PPWOldSize)
4100  PostPromotionWorklist.pop_back();
4101 
4102  // We couldn't promote and we didn't create a new partition, nothing
4103  // happened.
4104  if (NewAI == &AI)
4105  return nullptr;
4106 
4107  // If we can't promote the alloca, iterate on it to check for new
4108  // refinements exposed by splitting the current alloca. Don't iterate on an
4109  // alloca which didn't actually change and didn't get promoted.
4110  Worklist.insert(NewAI);
4111  }
4112 
4113  return NewAI;
4114 }
4115 
4116 /// Walks the slices of an alloca and form partitions based on them,
4117 /// rewriting each of their uses.
4118 bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &AS) {
4119  if (AS.begin() == AS.end())
4120  return false;
4121 
4122  unsigned NumPartitions = 0;
4123  bool Changed = false;
4124  const DataLayout &DL = AI.getModule()->getDataLayout();
4125 
4126  // First try to pre-split loads and stores.
4127  Changed |= presplitLoadsAndStores(AI, AS);
4128 
4129  // Now that we have identified any pre-splitting opportunities,
4130  // mark loads and stores unsplittable except for the following case.
4131  // We leave a slice splittable if all other slices are disjoint or fully
4132  // included in the slice, such as whole-alloca loads and stores.
4133  // If we fail to split these during pre-splitting, we want to force them
4134  // to be rewritten into a partition.
4135  bool IsSorted = true;
4136 
4137  uint64_t AllocaSize = DL.getTypeAllocSize(AI.getAllocatedType());
4138  const uint64_t MaxBitVectorSize = 1024;
4139  if (AllocaSize <= MaxBitVectorSize) {
4140  // If a byte boundary is included in any load or store, a slice starting or
4141  // ending at the boundary is not splittable.
4142  SmallBitVector SplittableOffset(AllocaSize + 1, true);
4143  for (Slice &S : AS)
4144  for (unsigned O = S.beginOffset() + 1;
4145  O < S.endOffset() && O < AllocaSize; O++)
4146  SplittableOffset.reset(O);
4147 
4148  for (Slice &S : AS) {
4149  if (!S.isSplittable())
4150  continue;
4151 
4152  if ((S.beginOffset() > AllocaSize || SplittableOffset[S.beginOffset()]) &&
4153  (S.endOffset() > AllocaSize || SplittableOffset[S.endOffset()]))
4154  continue;
4155 
4156  if (isa<LoadInst>(S.getUse()->getUser()) ||
4157  isa<StoreInst>(S.getUse()->getUser())) {
4158  S.makeUnsplittable();
4159  IsSorted = false;
4160  }
4161  }
4162  }
4163  else {
4164  // We only allow whole-alloca splittable loads and stores
4165  // for a large alloca to avoid creating too large BitVector.
4166  for (Slice &S : AS) {
4167  if (!S.isSplittable())
4168  continue;
4169 
4170  if (S.beginOffset() == 0 && S.endOffset() >= AllocaSize)
4171  continue;
4172 
4173  if (isa<LoadInst>(S.getUse()->getUser()) ||
4174  isa<StoreInst>(S.getUse()->getUser())) {
4175  S.makeUnsplittable();
4176  IsSorted = false;
4177  }
4178  }
4179  }
4180 
4181  if (!IsSorted)
4182  llvm::sort(AS.begin(), AS.end());
4183 
4184  /// Describes the allocas introduced by rewritePartition in order to migrate
4185  /// the debug info.
4186  struct Fragment {
4187  AllocaInst *Alloca;
4188  uint64_t Offset;
4189  uint64_t Size;
4190  Fragment(AllocaInst *AI, uint64_t O, uint64_t S)
4191  : Alloca(AI), Offset(O), Size(S) {}
4192  };
4193  SmallVector<Fragment, 4> Fragments;
4194 
4195  // Rewrite each partition.
4196  for (auto &P : AS.partitions()) {
4197  if (AllocaInst *NewAI = rewritePartition(AI, AS, P)) {
4198  Changed = true;
4199  if (NewAI != &AI) {
4200  uint64_t SizeOfByte = 8;
4201  uint64_t AllocaSize = DL.getTypeSizeInBits(NewAI->getAllocatedType());
4202  // Don't include any padding.
4203  uint64_t Size = std::min(AllocaSize, P.size() * SizeOfByte);
4204  Fragments.push_back(Fragment(NewAI, P.beginOffset() * SizeOfByte, Size));
4205  }
4206  }
4207  ++NumPartitions;
4208  }
4209 
4210  NumAllocaPartitions += NumPartitions;
4211  MaxPartitionsPerAlloca.updateMax(NumPartitions);
4212 
4213  // Migrate debug information from the old alloca to the new alloca(s)
4214  // and the individual partitions.
4216  if (!DbgDeclares.empty()) {
4217  auto *Var = DbgDeclares.front()->getVariable();
4218  auto *Expr = DbgDeclares.front()->getExpression();
4219  auto VarSize = Var->getSizeInBits();
4220  DIBuilder DIB(*AI.getModule(), /*AllowUnresolved*/ false);
4221  uint64_t AllocaSize = DL.getTypeSizeInBits(AI.getAllocatedType());
4222  for (auto Fragment : Fragments) {
4223  // Create a fragment expression describing the new partition or reuse AI's
4224  // expression if there is only one partition.
4225  auto *FragmentExpr = Expr;
4226  if (Fragment.Size < AllocaSize || Expr->isFragment()) {
4227  // If this alloca is already a scalar replacement of a larger aggregate,
4228  // Fragment.Offset describes the offset inside the scalar.
4229  auto ExprFragment = Expr->getFragmentInfo();
4230  uint64_t Offset = ExprFragment ? ExprFragment->OffsetInBits : 0;
4231  uint64_t Start = Offset + Fragment.Offset;
4232  uint64_t Size = Fragment.Size;
4233  if (ExprFragment) {
4234  uint64_t AbsEnd =
4235  ExprFragment->OffsetInBits + ExprFragment->SizeInBits;
4236  if (Start >= AbsEnd)
4237  // No need to describe a SROAed padding.
4238  continue;
4239  Size = std::min(Size, AbsEnd - Start);
4240  }
4241  // The new, smaller fragment is stenciled out from the old fragment.
4242  if (auto OrigFragment = FragmentExpr->getFragmentInfo()) {
4243  assert(Start >= OrigFragment->OffsetInBits &&
4244  "new fragment is outside of original fragment");
4245  Start -= OrigFragment->OffsetInBits;
4246  }
4247 
4248  // The alloca may be larger than the variable.
4249  if (VarSize) {
4250  if (Size > *VarSize)
4251  Size = *VarSize;
4252  if (Size == 0 || Start + Size > *VarSize)
4253  continue;
4254  }
4255 
4256  // Avoid creating a fragment expression that covers the entire variable.
4257  if (!VarSize || *VarSize != Size) {
4258  if (auto E =
4259  DIExpression::createFragmentExpression(Expr, Start, Size))
4260  FragmentExpr = *E;
4261  else
4262  continue;
4263  }
4264  }
4265 
4266  // Remove any existing intrinsics describing the same alloca.
4267  for (DbgVariableIntrinsic *OldDII : FindDbgAddrUses(Fragment.Alloca))
4268  OldDII->eraseFromParent();
4269 
4270  DIB.insertDeclare(Fragment.Alloca, Var, FragmentExpr,
4271  DbgDeclares.front()->getDebugLoc(), &AI);
4272  }
4273  }
4274  return Changed;
4275 }
4276 
4277 /// Clobber a use with undef, deleting the used value if it becomes dead.
4278 void SROA::clobberUse(Use &U) {
4279  Value *OldV = U;
4280  // Replace the use with an undef value.
4281  U = UndefValue::get(OldV->getType());
4282 
4283  // Check for this making an instruction dead. We have to garbage collect
4284  // all the dead instructions to ensure the uses of any alloca end up being
4285  // minimal.
4286  if (Instruction *OldI = dyn_cast<Instruction>(OldV))
4287  if (isInstructionTriviallyDead(OldI)) {
4288  DeadInsts.insert(OldI);
4289  }
4290 }
4291 
4292 /// Analyze an alloca for SROA.
4293 ///
4294 /// This analyzes the alloca to ensure we can reason about it, builds
4295 /// the slices of the alloca, and then hands it off to be split and
4296 /// rewritten as needed.
4297 bool SROA::runOnAlloca(AllocaInst &AI) {
4298  LLVM_DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
4299  ++NumAllocasAnalyzed;
4300 
4301  // Special case dead allocas, as they're trivial.
4302  if (AI.use_empty()) {
4303  AI.eraseFromParent();
4304  return true;
4305  }
4306  const DataLayout &DL = AI.getModule()->getDataLayout();
4307 
4308  // Skip alloca forms that this analysis can't handle.
4309  if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
4310  DL.getTypeAllocSize(AI.getAllocatedType()) == 0)
4311  return false;
4312 
4313  bool Changed = false;
4314 
4315  // First, split any FCA loads and stores touching this alloca to promote
4316  // better splitting and promotion opportunities.
4317  AggLoadStoreRewriter AggRewriter;
4318  Changed |= AggRewriter.rewrite(AI);
4319 
4320  // Build the slices using a recursive instruction-visiting builder.
4321  AllocaSlices AS(DL, AI);
4322  LLVM_DEBUG(AS.print(dbgs()));
4323  if (AS.isEscaped())
4324  return Changed;
4325 
4326  // Delete all the dead users of this alloca before splitting and rewriting it.
4327  for (Instruction *DeadUser : AS.getDeadUsers()) {
4328  // Free up everything used by this instruction.
4329  for (Use &DeadOp : DeadUser->operands())
4330  clobberUse(DeadOp);
4331 
4332  // Now replace the uses of this instruction.
4333  DeadUser->replaceAllUsesWith(UndefValue::get(DeadUser->getType()));
4334 
4335  // And mark it for deletion.
4336  DeadInsts.insert(DeadUser);
4337  Changed = true;
4338  }
4339  for (Use *DeadOp : AS.getDeadOperands()) {
4340  clobberUse(*DeadOp);
4341  Changed = true;
4342  }
4343 
4344  // No slices to split. Leave the dead alloca for a later pass to clean up.
4345  if (AS.begin() == AS.end())
4346  return Changed;
4347 
4348  Changed |= splitAlloca(AI, AS);
4349 
4350  LLVM_DEBUG(dbgs() << " Speculating PHIs\n");
4351  while (!SpeculatablePHIs.empty())
4352  speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
4353 
4354  LLVM_DEBUG(dbgs() << " Speculating Selects\n");
4355  while (!SpeculatableSelects.empty())
4356  speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
4357 
4358  return Changed;
4359 }
4360 
4361 /// Delete the dead instructions accumulated in this run.
4362 ///
4363 /// Recursively deletes the dead instructions we've accumulated. This is done
4364 /// at the very end to maximize locality of the recursive delete and to
4365 /// minimize the problems of invalidated instruction pointers as such pointers
4366 /// are used heavily in the intermediate stages of the algorithm.
4367 ///
4368 /// We also record the alloca instructions deleted here so that they aren't
4369 /// subsequently handed to mem2reg to promote.
4370 bool SROA::deleteDeadInstructions(
4371  SmallPtrSetImpl<AllocaInst *> &DeletedAllocas) {
4372  bool Changed = false;
4373  while (!DeadInsts.empty()) {
4374  Instruction *I = DeadInsts.pop_back_val();
4375  LLVM_DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
4376 
4377  // If the instruction is an alloca, find the possible dbg.declare connected
4378  // to it, and remove it too. We must do this before calling RAUW or we will
4379  // not be able to find it.
4380  if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
4381  DeletedAllocas.insert(AI);
4382  for (DbgVariableIntrinsic *OldDII : FindDbgAddrUses(AI))
4383  OldDII->eraseFromParent();
4384  }
4385 
4387 
4388  for (Use &Operand : I->operands())
4389  if (Instruction *U = dyn_cast<Instruction>(Operand)) {
4390  // Zero out the operand and see if it becomes trivially dead.
4391  Operand = nullptr;
4393  DeadInsts.insert(U);
4394  }
4395 
4396  ++NumDeleted;
4397  I->eraseFromParent();
4398  Changed = true;
4399  }
4400  return Changed;
4401 }
4402 
4403 /// Promote the allocas, using the best available technique.
4404 ///
4405 /// This attempts to promote whatever allocas have been identified as viable in
4406 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
4407 /// This function returns whether any promotion occurred.
4408 bool SROA::promoteAllocas(Function &F) {
4409  if (PromotableAllocas.empty())
4410  return false;
4411 
4412  NumPromoted += PromotableAllocas.size();
4413 
4414  LLVM_DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
4415  PromoteMemToReg(PromotableAllocas, *DT, AC);
4416  PromotableAllocas.clear();
4417  return true;
4418 }
4419 
4420 PreservedAnalyses SROA::runImpl(Function &F, DominatorTree &RunDT,
4421  AssumptionCache &RunAC) {
4422  LLVM_DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
4423  C = &F.getContext();
4424  DT = &RunDT;
4425  AC = &RunAC;
4426 
4427  BasicBlock &EntryBB = F.getEntryBlock();
4428  for (BasicBlock::iterator I = EntryBB.begin(), E = std::prev(EntryBB.end());
4429  I != E; ++I) {
4430  if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
4431  Worklist.insert(AI);
4432  }
4433 
4434  bool Changed = false;
4435  // A set of deleted alloca instruction pointers which should be removed from
4436  // the list of promotable allocas.
4437  SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
4438 
4439  do {
4440  while (!Worklist.empty()) {
4441  Changed |= runOnAlloca(*Worklist.pop_back_val());
4442  Changed |= deleteDeadInstructions(DeletedAllocas);
4443 
4444  // Remove the deleted allocas from various lists so that we don't try to
4445  // continue processing them.
4446  if (!DeletedAllocas.empty()) {
4447  auto IsInSet = [&](AllocaInst *AI) { return DeletedAllocas.count(AI); };
4448  Worklist.remove_if(IsInSet);
4449  PostPromotionWorklist.remove_if(IsInSet);
4450  PromotableAllocas.erase(llvm::remove_if(PromotableAllocas, IsInSet),
4451  PromotableAllocas.end());
4452  DeletedAllocas.clear();
4453  }
4454  }
4455 
4456  Changed |= promoteAllocas(F);
4457 
4458  Worklist = PostPromotionWorklist;
4459  PostPromotionWorklist.clear();
4460  } while (!Worklist.empty());
4461 
4462  if (!Changed)
4463  return PreservedAnalyses::all();
4464 
4465  PreservedAnalyses PA;
4466  PA.preserveSet<CFGAnalyses>();
4467  PA.preserve<GlobalsAA>();
4468  return PA;
4469 }
4470 
4472  return runImpl(F, AM.getResult<DominatorTreeAnalysis>(F),
4473  AM.getResult<AssumptionAnalysis>(F));
4474 }
4475 
4476 /// A legacy pass for the legacy pass manager that wraps the \c SROA pass.
4477 ///
4478 /// This is in the llvm namespace purely to allow it to be a friend of the \c
4479 /// SROA pass.
4481  /// The SROA implementation.
4482  SROA Impl;
4483 
4484 public:
4485  static char ID;
4486 
4489  }
4490 
4491  bool runOnFunction(Function &F) override {
4492  if (skipFunction(F))
4493  return false;
4494 
4495  auto PA = Impl.runImpl(
4496  F, getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
4497  getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F));
4498  return !PA.areAllPreserved();
4499  }
4500 
4501  void getAnalysisUsage(AnalysisUsage &AU) const override {
4505  AU.setPreservesCFG();
4506  }
4507 
4508  StringRef getPassName() const override { return "SROA"; }
4509 };
4510 
4511 char SROALegacyPass::ID = 0;
4512 
4514 
4516  "Scalar Replacement Of Aggregates", false, false)
4520  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:1554
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:1852
iterator end() const
Definition: SROA.cpp:417
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:1440
Value * getValueOperand()
Definition: Instructions.h:399
SymbolTableList< Instruction >::iterator eraseFromParent()
This method unlinks &#39;this&#39; from the containing basic block and deletes it.
Definition: Instruction.cpp:68
void push_back(const T &Elt)
Definition: SmallVector.h:218
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:111
const_iterator end(StringRef path)
Get end iterator over path.
Definition: Path.cpp:259
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:419
An iterator over partitions of the alloca&#39;s slices.
Definition: SROA.cpp:437
void setInt(IntType IntVal)
bool isSimple() const
Definition: Instructions.h:266
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:355
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:365
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:1557
GCNRegPressure max(const GCNRegPressure &P1, const GCNRegPressure &P2)
Base class for instruction visitors.
Definition: InstVisitor.h:81
const_iterator begin(StringRef path, Style style=Style::native)
Get begin iterator over path.
Definition: Path.cpp:250
const Value * stripInBoundsOffsets() const
Strip off pointer casts and inbounds GEPs.
Definition: Value.cpp:599
uint64_t beginOffset() const
The start offset of this partition.
Definition: SROA.cpp:388
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:687
Compute iterated dominance frontiers using a linear time algorithm.
Definition: AllocatorList.h:24
PointerTy getPointer() const
Type * getElementType(unsigned N) const
Definition: DerivedTypes.h:314
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:265
iterator begin() const
Definition: ArrayRef.h:137
constexpr char IsVolatile[]
Key for Kernel::Arg::Metadata::mIsVolatile.
APInt sdiv(const APInt &RHS) const
Signed division function for APInt.
Definition: APInt.cpp:1598
LLVM_NODISCARD size_t rfind(char C, size_t From=npos) const
Search for the last character C in the string.
Definition: StringRef.h:360
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:588
This file provides the interface for LLVM&#39;s Scalar Replacement of Aggregates pass.
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:266
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:398
This class provides information about the result of a visit.
Definition: PtrUseVisitor.h:63
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:2090
This file contains the declarations for metadata subclasses.
Representation of the alloca slices.
Definition: SROA.cpp:239
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:84
SyncScope::ID getSyncScopeID() const
Returns the synchronization scope ID of this store instruction.
Definition: Instructions.h:374
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:273
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:237
Offsets
Offsets in bytes from the start of the input buffer.
Definition: SIInstrInfo.h:943
LLVMContext & getContext() const
All values hold a context through their type.
Definition: Value.cpp:714
void deleteValue()
Delete a pointer to a generic Value.
Definition: Value.cpp:99
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:183
unsigned second
Scalar Replacement Of Aggregates
Definition: SROA.cpp:4519
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:1018
unsigned getPointerSizeInBits(unsigned AS=0) const
Layout pointer size, in bits FIXME: The defaults need to be removed once all of the backends/clients ...
Definition: DataLayout.h:360
STATISTIC(NumFunctions, "Total number of functions")
Metadata node.
Definition: Metadata.h:862
Analysis pass which computes a DominatorTree.
Definition: Dominators.h:225
F(f)
unsigned getPointerAddressSpace() const
Get the address space of this pointer or pointer vector type.
Definition: DerivedTypes.h:503
#define LLVM_DUMP_METHOD
Mark debug helper function definitions like dump() that should not be stripped from debug builds...
Definition: Compiler.h:452
An instruction for reading from memory.
Definition: Instructions.h:168
APInt zextOrTrunc(unsigned width) const
Zero extend or truncate to width.
Definition: APInt.cpp:883
Hexagon Common GEP
bool isVectorTy() const
True if this is an instance of VectorType.
Definition: Type.h:230
This defines the Use class.
void reserve(size_type N)
Definition: SmallVector.h:376
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:385
bool isEscaped() const
Test whether a pointer to the allocation escapes our analysis.
Definition: SROA.cpp:248
TinyPtrVector - This class is specialized for cases where there are normally 0 or 1 element in a vect...
Definition: TinyPtrVector.h:31
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:201
op_iterator op_begin()
Definition: User.h:230
unsigned getElementContainingOffset(uint64_t Offset) const
Given a valid byte offset into the structure, returns the structure index that contains it...
Definition: DataLayout.cpp:84
unsigned getBitWidth() const
Return the number of bits in the APInt.
Definition: APInt.h:1503
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:1649
Builder for the alloca slices.
Definition: SROA.cpp:655
LLVMContext & getContext() const
Return the LLVMContext in which this type was uniqued.
Definition: Type.h:130
A templated base class for SmallPtrSet which provides the typesafe interface that is common across al...
Definition: SmallPtrSet.h:344
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:264
std::pair< iterator, bool > insert(const std::pair< KeyT, ValueT > &KV)
Definition: DenseMap.h:191
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:260
Maximum number of bits that can be specified.
Definition: DerivedTypes.h:52
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:1776
PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM)
Run the pass over the function.
Definition: SROA.cpp:4471
void * PointerTy
Definition: GenericValue.h:22
AnalysisUsage & addRequired()
Used to lazily calculate structure layout information for a target machine, based on the DataLayout s...
Definition: DataLayout.h:521
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:51
void setPointer(PointerTy PtrVal)
bool isVolatile() const
Return true if this is a load from a volatile memory location.
Definition: Instructions.h:221
This class represents the LLVM &#39;select&#39; instruction.
Type * getPointerElementType() const
Definition: Type.h:376
const DataLayout & getDataLayout() const
Get the data layout for the module&#39;s target platform.
Definition: Module.cpp:361
Twine - A lightweight data structure for efficiently representing the concatenation of temporary valu...
Definition: Twine.h:81
unsigned getAlignment() const
Return the alignment of the memory that is being allocated by the instruction.
Definition: Instructions.h:113
unsigned getPointerTypeSizeInBits(Type *) const
Layout pointer size, in bits, based on the type.
Definition: DataLayout.cpp:638
ArrayRef< T > makeArrayRef(const T &OneElt)
Construct an ArrayRef from a single element.
Definition: ArrayRef.h:451
PointerType * getType() const
Overload to return most specific pointer type.
Definition: Instructions.h:97
Class to represent struct types.
Definition: DerivedTypes.h:201
A Use represents the edge between a Value definition and its users.
Definition: Use.h:56
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:3371
PointerType * getPointerTo(unsigned AddrSpace=0) const
Return a pointer to the current type.
Definition: Type.cpp:639
ArrayRef< Slice * > splitSliceTails() const
Get the sequence of split slice tails.
Definition: SROA.cpp:425
bool isAborted() const
Did we abort the visit early?
Definition: PtrUseVisitor.h:76
This class consists of common code factored out of the SmallVector class to reduce code duplication b...
Definition: APFloat.h:42
bool isIntegerTy() const
True if this is an instance of IntegerType.
Definition: Type.h:197
This provides a uniform API for creating instructions and inserting them into a basic block: either a...
Definition: IRBuilder.h:731
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:1726
A partition of the slices.
Definition: SROA.cpp:363
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:4501
uint64_t getNumElements() const
Definition: DerivedTypes.h:359
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:336
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:4491
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:1642
unsigned getPointerAddressSpace() const
Returns the address space of the pointer operand.
Definition: Instructions.h:408
User * getUser() const LLVM_READONLY
Returns the User that contains this Use.
Definition: Use.cpp:41
bool visit(AllocaSlices::const_iterator I)
Definition: SROA.cpp:2309
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:245
bool insert(const value_type &X)
Insert a new element into the SetVector.
Definition: SetVector.h:142
void copyNonnullMetadata(const LoadInst &OldLI, MDNode *N, LoadInst &NewLI)
Copy a nonnull metadata node to a new load instruction.
Definition: Local.cpp:2472
bool isInBounds() const
Determine whether the GEP has the inbounds flag.
SmallVectorImpl< Slice >::iterator iterator
Support for iterating over the slices.
Definition: SROA.cpp:252
const_iterator end() const
Definition: SROA.cpp:262
IntType getInt() const
void setLength(Value *L)
A legacy pass for the legacy pass manager that wraps the SROA pass.
Definition: SROA.cpp:4480
Class to represent array types.
Definition: DerivedTypes.h:369
This is the common base class for debug info intrinsics for variables.
Definition: IntrinsicInst.h:88
APInt sextOrTrunc(unsigned width) const
Sign extend or truncate to width.
Definition: APInt.cpp:891
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:200
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory)...
Definition: APInt.h:33
const APInt & getValue() const
Return the constant as an APInt value reference.
Definition: Constants.h:138
ConstantFolder - Create constants with minimum, target independent, folding.
An instruction for storing to memory.
Definition: Instructions.h:310
bool isIntOrIntVectorTy() const
Return true if this is an integer type or a vector of integer types.
Definition: Type.h:203
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
Definition: Value.cpp:439
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:598
static Constant * getUDiv(Constant *C1, Constant *C2, bool isExact=false)
Definition: Constants.cpp:2219
unsigned getBitWidth() const
Get the number of bits in this IntegerType.
Definition: DerivedTypes.h:66
Type::subtype_iterator element_iterator
Definition: DerivedTypes.h:301
CRTP base class which implements the entire standard iterator facade in terms of a minimal subset of ...
Definition: iterator.h:68
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree...
Definition: Dominators.h:142
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:1675
Value * getOperand(unsigned i) const
Definition: User.h:170
Class to represent pointers.
Definition: DerivedTypes.h:467
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:1084
Type * getScalarType() const
If this is a vector type, return the element type, otherwise return &#39;this&#39;.
Definition: Type.h:304
iterator find(const_arg_type_t< KeyT > Val)
Definition: DenseMap.h:146
const BasicBlock & getEntryBlock() const
Definition: Function.h:626
constexpr uint64_t MinAlign(uint64_t A, uint64_t B)
A and B are either alignments or offsets.
Definition: MathExtras.h:610
an instruction for type-safe pointer arithmetic to access elements of arrays and structs ...
Definition: Instructions.h:841
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:742
void getAAMetadata(AAMDNodes &N, bool Merge=false) const
Fills the AAMDNodes structure with AA metadata from this instruction.
Scalar Replacement Of false
Definition: SROA.cpp:4519
iterator_range< partition_iterator > partitions()
#define P(N)
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:410
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:149
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
void dump(const SparseBitVector< ElementSize > &LHS, raw_ostream &out)
Subclasses of this class are all able to terminate a basic block.
Definition: InstrTypes.h:55
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:218
void setDebugLoc(DebugLoc Loc)
Set the debug location information for this instruction.
Definition: Instruction.h:287
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:1265
LLVM Basic Block Representation.
Definition: BasicBlock.h:59
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:46
static unsigned getPointerOperandIndex()
Definition: Instructions.h:404
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
This is an important base class in LLVM.
Definition: Constant.h:42
LLVM_NODISCARD bool empty() const
Definition: SmallPtrSet.h:92
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator begin()
Definition: SmallVector.h:129
ArrayRef< Use * > getDeadOperands() const
Access the dead operands referring to this alloca.
Definition: SROA.cpp:295
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:224
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:1308
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:250
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:371
bool mayHaveSideEffects() const
Return true if the instruction may have side effects.
Definition: Instruction.h:541
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:1514
static sys::TimePoint< std::chrono::seconds > now(bool Deterministic)
element_iterator element_end() const
Definition: DerivedTypes.h:304
DIExpression * getExpression() const
Represent the analysis usage information of a pass.
op_iterator op_end()
Definition: User.h:232
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:1025
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:4519
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:285
bool empty() const
op_range operands()
Definition: User.h:238
Value * getPointerOperand()
Definition: Instructions.h:274
unsigned getAddressSpace() const
Return the address space of the Pointer type.
Definition: DerivedTypes.h:495
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:382
Class to represent integer types.
Definition: DerivedTypes.h:40
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:2280
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:1058
const Value * getCondition() const
static Constant * getAllOnesValue(Type *Ty)
Definition: Constants.cpp:322
LLVMContext & getContext() const
getContext - Return a reference to the LLVMContext associated with this function. ...
Definition: Function.cpp:194
static UndefValue * get(Type *T)
Static factory methods - Return an &#39;undef&#39; object of the specified type.
Definition: Constants.cpp:1392
const AMDGPUAS & AS
iterator erase(const_iterator CI)
Definition: SmallVector.h:445
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:159
size_t size() const
Definition: SmallVector.h:53
Value * getIncomingValue(unsigned i) const
Return incoming value number x.
static PointerType * getInt8PtrTy(LLVMContext &C, unsigned AS=0)
Definition: Type.cpp:220
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:106
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:1396
void sort(IteratorTy Start, IteratorTy End)
Definition: STLExtras.h:969
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:3333
Intrinsic::ID getIntrinsicID() const
Return the intrinsic ID of this intrinsic.
Definition: IntrinsicInst.h:51
SmallBitVector & reset()
This provides the default implementation of the IRBuilder &#39;InsertHelper&#39; method that is called whenev...
Definition: IRBuilder.h:62
This is the superclass of the array and vector type classes.
Definition: DerivedTypes.h:343
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:227
isPodLike - This is a type trait that is used to determine whether a given type can be copied around ...
Definition: ArrayRef.h:530
A SetVector that performs no allocations if smaller than a certain size.
Definition: SetVector.h:298
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:251
This is the shared class of boolean and integer constants.
Definition: Constants.h:84
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:1488
iterator end()
Definition: BasicBlock.h:266
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:64
partition_iterator & operator++()
Definition: SROA.cpp:610
static Value * insertVector(IRBuilderTy &IRB, Value *Old, Value *V, unsigned BeginIndex, const Twine &Name)
Definition: SROA.cpp:2170
iterator end() const
Definition: ArrayRef.h:138
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:243
unsigned getABITypeAlignment(Type *Ty) const
Returns the minimum ABI-required alignment for the specified type.
Definition: DataLayout.cpp:722
A collection of metadata nodes that might be associated with a memory access used by the alias-analys...
Definition: Metadata.h:642
iterator begin() const
Definition: SROA.cpp:416
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:381
uint64_t getSizeInBytes() const
Definition: DataLayout.h:529
SmallVectorImpl< Slice >::const_iterator const_iterator
Definition: SROA.cpp:258
static IntegerType * getIntNTy(LLVMContext &C, unsigned N)
Definition: Type.cpp:180
SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &AS)
Definition: SROA.cpp:671
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:286
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:1287
void setOperand(unsigned i, Value *Val)
Definition: User.h:175
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:133
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:924
A range adaptor for a pair of iterators.
Class to represent vector types.
Definition: DerivedTypes.h:393
const Module * getModule() const
Return the module owning the function this instruction belongs to or nullptr it the function does not...
Definition: Instruction.cpp:56
Class for arbitrary precision integers.
Definition: APInt.h:70
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:1960
iterator_range< user_iterator > users()
Definition: Value.h:400
#define NDEBUG
Definition: regutils.h:48
Represents analyses that only rely on functions&#39; control flow.
Definition: PassManager.h:114
const Value * getFalseValue() const
#define LLVM_ATTRIBUTE_UNUSED
Definition: Compiler.h:147
element_iterator element_begin() const
Definition: DerivedTypes.h:303
static Value * insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old, Value *V, uint64_t Offset, const Twine &Name)
Definition: SROA.cpp:2113
bool isNonIntegralPointerType(PointerType *PT) const
Definition: DataLayout.h:346
const_iterator begin() const
Definition: SROA.cpp:261
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:1109
uint64_t getTypeSizeInBits(Type *Ty) const
Size examples:
Definition: DataLayout.h:560
static Value * foldPHINodeOrSelectInst(Instruction &I)
A helper that folds a PHI node or a select.
Definition: SROA.cpp:643
uint64_t getTypeAllocSize(Type *Ty) const
Returns the offset in bytes between successive objects of the specified type, including alignment pad...
Definition: DataLayout.h:428
Virtual Register Rewriter
Definition: VirtRegMap.cpp:222
bool operator!=(uint64_t V1, const APInt &V2)
Definition: APInt.h:1963
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:1249
static bool isZero(Value *V, const DataLayout &DL, DominatorTree *DT, AssumptionCache *AC)
Definition: Lint.cpp:546
This class wraps the llvm.memcpy/memmove intrinsics.
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator end()
Definition: SmallVector.h:133
bool isPacked() const
Definition: DerivedTypes.h:261
bool isVolatile() const
Return true if this is a store to a volatile memory location.
Definition: Instructions.h:343
const DebugLoc & getDebugLoc() const
Return the debug location for this node as a DebugLoc.
Definition: Instruction.h:290
static const size_t npos
Definition: StringRef.h:51
uint64_t getElementOffset(unsigned Idx) const
Definition: DataLayout.h:543
unsigned getAlignment() const
Return the alignment of the access that is being performed.
Definition: Instructions.h:230
unsigned getIntegerBitWidth() const
Definition: DerivedTypes.h:97
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:56
AtomicOrdering getOrdering() const
Returns the ordering constraint of this store instruction.
Definition: Instructions.h:362
StringRef getPassName() const override
getPassName - Return a nice clean name for a pass.
Definition: SROA.cpp:4508
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:482
Visitor to rewrite instructions using p particular slice of an alloca to use a new alloca...
Definition: SROA.cpp:2223
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:224
BasicBlock * getIncomingBlock(unsigned i) const
Return incoming basic block number i.
static void speculateSelectInstLoads(SelectInst &SI)
Definition: SROA.cpp:1330
SyncScope::ID getSyncScopeID() const
Returns the synchronization scope ID of this load instruction.
Definition: Instructions.h:249
#define I(x, y, z)
Definition: MD5.cpp:58
#define N
iterator end()
Definition: DenseMap.h:79
FunctionPass * createSROAPass()
Definition: SROA.cpp:4513
uint64_t endOffset() const
The end offset of this partition.
Definition: SROA.cpp:393
Instruction * getEscapingInst() const
Get the instruction causing the pointer to escape.
Definition: PtrUseVisitor.h:89
static ArrayType * get(Type *ElementType, uint64_t NumElements)
This static method is the primary way to construct an ArrayType.
Definition: Type.cpp:568
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:323
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:2144
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:141
std::string str() const
Return the twine contents as a std::string.
Definition: Twine.cpp:18
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:181
unsigned getPointerAddressSpace() const
Returns the address space of the pointer operand.
Definition: Instructions.h:280
static unsigned getPointerOperandIndex()
Definition: Instructions.h:276
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:376
unsigned getNumSuccessors() const
Return the number of successors that this terminator has.
Definition: JSON.cpp:592
Value * getArgOperand(unsigned i) const
getArgOperand/setArgOperand - Return/set the i-th call argument.
bool operator<(int64_t V1, const APSInt &V2)
Definition: APSInt.h:326
bool isSingleValueType() const
Return true if the type is a valid type for a register in codegen.
Definition: Type.h:250
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:347
LLVM Value Representation.
Definition: Value.h:73
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:1373
uint64_t getTypeStoreSize(Type *Ty) const
Returns the maximum number of bytes that may be overwritten by storing the specified type...
Definition: DataLayout.h:411
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:1176
constexpr char Size[]
Key for Kernel::Arg::Metadata::mSize.
unsigned getOpcode() const
Return the opcode for this Instruction or ConstantExpr.
Definition: Operator.h:41
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:593
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:81
void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB, BasicBlock::iterator InsertPt) const
Definition: IRBuilder.h:64
static void speculatePHINodeLoads(PHINode &PN)
Definition: SROA.cpp:1240
This class implements an extremely fast bulk output stream that can only output to a stream...
Definition: raw_ostream.h:46
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:66
Type * getElementType() const
Definition: DerivedTypes.h:360
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:398
IRTranslator LLVM IR MI
bool hasOneUse() const
Return true if there is exactly one user of this value.
Definition: Value.h:413
bool empty() const
Test whether this partition contains no slices, and merely spans a region occupied by split slices...
Definition: SROA.cpp:405
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:49
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:365
Legacy analysis pass which computes a DominatorTree.
Definition: Dominators.h:254
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:2048
bool operator==(uint64_t V1, const APInt &V2)
Definition: APInt.h:1961
bool isSimple() const
Definition: Instructions.h:391
const TerminatorInst * 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:138
This header defines various interfaces for pass management in LLVM.
bool isBigEndian() const
Definition: DataLayout.h:222
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:298
#define LLVM_DEBUG(X)
Definition: Debug.h:119
Value * getPointerOperand()
Definition: Instructions.h:402
static IntegerType * getInt8Ty(LLVMContext &C)
Definition: Type.cpp:174
Value * getRawDest() const
bool use_empty() const
Definition: Value.h:323
static Constant * get(ArrayRef< Constant *> V)
Definition: Constants.cpp:1056
Type * getElementType() const
Definition: DerivedTypes.h:486
const Use & getRawDestUse() const
void setDestAlignment(unsigned Align)
bool isArrayTy() const
True if this is an instance of ArrayType.
Definition: Type.h:221
static Optional< DIExpression * > createFragmentExpression(const DIExpression *Expr, unsigned OffsetInBits, unsigned SizeInBits)
Create a DIExpression to describe one part of an aggregate variable that is fragmented across multipl...
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:44
const BasicBlock * getParent() const
Definition: Instruction.h:67
an instruction to allocate memory on the stack
Definition: Instructions.h:60
gep_type_iterator gep_type_begin(const User *GEP)
static Value * foldSelectInst(SelectInst &SI)
Definition: SROA.cpp:630
ArrayRef< Instruction * > getDeadUsers() const
Access the dead users for this alloca.
Definition: SROA.cpp:287
bool isEscaped() const
Is the pointer escaped at some point?
Definition: PtrUseVisitor.h:79
bool operator==(const partition_iterator &RHS) const
Definition: SROA.cpp:590