LLVM  6.0.0svn
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/STLExtras.h"
28 #include "llvm/ADT/SetVector.h"
29 #include "llvm/ADT/SmallVector.h"
30 #include "llvm/ADT/Statistic.h"
33 #include "llvm/Analysis/Loads.h"
36 #include "llvm/IR/Constants.h"
37 #include "llvm/IR/DIBuilder.h"
38 #include "llvm/IR/DataLayout.h"
39 #include "llvm/IR/DebugInfo.h"
40 #include "llvm/IR/DerivedTypes.h"
41 #include "llvm/IR/IRBuilder.h"
42 #include "llvm/IR/InstVisitor.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/IntrinsicInst.h"
45 #include "llvm/IR/LLVMContext.h"
46 #include "llvm/IR/Operator.h"
47 #include "llvm/Pass.h"
48 #include "llvm/Support/Chrono.h"
50 #include "llvm/Support/Compiler.h"
51 #include "llvm/Support/Debug.h"
55 #include "llvm/Transforms/Scalar.h"
58 
59 #ifndef NDEBUG
60 // We only use this for a debug check.
61 #include <random>
62 #endif
63 
64 using namespace llvm;
65 using namespace llvm::sroa;
66 
67 #define DEBUG_TYPE "sroa"
68 
69 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
70 STATISTIC(NumAllocaPartitions, "Number of alloca partitions formed");
71 STATISTIC(MaxPartitionsPerAlloca, "Maximum number of partitions per alloca");
72 STATISTIC(NumAllocaPartitionUses, "Number of alloca partition uses rewritten");
73 STATISTIC(MaxUsesPerAllocaPartition, "Maximum number of uses of a partition");
74 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
75 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
76 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
77 STATISTIC(NumDeleted, "Number of instructions deleted");
78 STATISTIC(NumVectorized, "Number of vectorized aggregates");
79 
80 /// Hidden option to enable randomly shuffling the slices to help uncover
81 /// instability in their order.
82 static cl::opt<bool> SROARandomShuffleSlices("sroa-random-shuffle-slices",
83  cl::init(false), cl::Hidden);
84 
85 /// Hidden option to experiment with completely strict handling of inbounds
86 /// GEPs.
87 static cl::opt<bool> SROAStrictInbounds("sroa-strict-inbounds", cl::init(false),
88  cl::Hidden);
89 
90 namespace {
91 /// \brief A custom IRBuilder inserter which prefixes all names, but only in
92 /// Assert builds.
93 class IRBuilderPrefixedInserter : public IRBuilderDefaultInserter {
94  std::string Prefix;
95  const Twine getNameWithPrefix(const Twine &Name) const {
96  return Name.isTriviallyEmpty() ? Name : Prefix + Name;
97  }
98 
99 public:
100  void SetNamePrefix(const Twine &P) { Prefix = P.str(); }
101 
102 protected:
103  void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
104  BasicBlock::iterator InsertPt) const {
105  IRBuilderDefaultInserter::InsertHelper(I, getNameWithPrefix(Name), BB,
106  InsertPt);
107  }
108 };
109 
110 /// \brief Provide a typedef for IRBuilder that drops names in release builds.
112 }
113 
114 namespace {
115 /// \brief A used slice of an alloca.
116 ///
117 /// This structure represents a slice of an alloca used by some instruction. It
118 /// stores both the begin and end offsets of this use, a pointer to the use
119 /// itself, and a flag indicating whether we can classify the use as splittable
120 /// or not when forming partitions of the alloca.
121 class Slice {
122  /// \brief The beginning offset of the range.
123  uint64_t BeginOffset;
124 
125  /// \brief The ending offset, not included in the range.
126  uint64_t EndOffset;
127 
128  /// \brief Storage for both the use of this slice and whether it can be
129  /// split.
130  PointerIntPair<Use *, 1, bool> UseAndIsSplittable;
131 
132 public:
133  Slice() : BeginOffset(), EndOffset() {}
134  Slice(uint64_t BeginOffset, uint64_t EndOffset, Use *U, bool IsSplittable)
135  : BeginOffset(BeginOffset), EndOffset(EndOffset),
136  UseAndIsSplittable(U, IsSplittable) {}
137 
138  uint64_t beginOffset() const { return BeginOffset; }
139  uint64_t endOffset() const { return EndOffset; }
140 
141  bool isSplittable() const { return UseAndIsSplittable.getInt(); }
142  void makeUnsplittable() { UseAndIsSplittable.setInt(false); }
143 
144  Use *getUse() const { return UseAndIsSplittable.getPointer(); }
145 
146  bool isDead() const { return getUse() == nullptr; }
147  void kill() { UseAndIsSplittable.setPointer(nullptr); }
148 
149  /// \brief Support for ordering ranges.
150  ///
151  /// This provides an ordering over ranges such that start offsets are
152  /// always increasing, and within equal start offsets, the end offsets are
153  /// decreasing. Thus the spanning range comes first in a cluster with the
154  /// same start position.
155  bool operator<(const Slice &RHS) const {
156  if (beginOffset() < RHS.beginOffset())
157  return true;
158  if (beginOffset() > RHS.beginOffset())
159  return false;
160  if (isSplittable() != RHS.isSplittable())
161  return !isSplittable();
162  if (endOffset() > RHS.endOffset())
163  return true;
164  return false;
165  }
166 
167  /// \brief Support comparison with a single offset to allow binary searches.
168  friend LLVM_ATTRIBUTE_UNUSED bool operator<(const Slice &LHS,
169  uint64_t RHSOffset) {
170  return LHS.beginOffset() < RHSOffset;
171  }
172  friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
173  const Slice &RHS) {
174  return LHSOffset < RHS.beginOffset();
175  }
176 
177  bool operator==(const Slice &RHS) const {
178  return isSplittable() == RHS.isSplittable() &&
179  beginOffset() == RHS.beginOffset() && endOffset() == RHS.endOffset();
180  }
181  bool operator!=(const Slice &RHS) const { return !operator==(RHS); }
182 };
183 } // end anonymous namespace
184 
185 namespace llvm {
186 template <typename T> struct isPodLike;
187 template <> struct isPodLike<Slice> { static const bool value = true; };
188 }
189 
190 /// \brief Representation of the alloca slices.
191 ///
192 /// This class represents the slices of an alloca which are formed by its
193 /// various uses. If a pointer escapes, we can't fully build a representation
194 /// for the slices used and we reflect that in this structure. The uses are
195 /// stored, sorted by increasing beginning offset and with unsplittable slices
196 /// starting at a particular offset before splittable slices.
198 public:
199  /// \brief Construct the slices of a particular alloca.
200  AllocaSlices(const DataLayout &DL, AllocaInst &AI);
201 
202  /// \brief Test whether a pointer to the allocation escapes our analysis.
203  ///
204  /// If this is true, the slices are never fully built and should be
205  /// ignored.
206  bool isEscaped() const { return PointerEscapingInstr; }
207 
208  /// \brief Support for iterating over the slices.
209  /// @{
212  iterator begin() { return Slices.begin(); }
213  iterator end() { return Slices.end(); }
214 
217  const_iterator begin() const { return Slices.begin(); }
218  const_iterator end() const { return Slices.end(); }
219  /// @}
220 
221  /// \brief Erase a range of slices.
222  void erase(iterator Start, iterator Stop) { Slices.erase(Start, Stop); }
223 
224  /// \brief Insert new slices for this alloca.
225  ///
226  /// This moves the slices into the alloca's slices collection, and re-sorts
227  /// everything so that the usual ordering properties of the alloca's slices
228  /// hold.
229  void insert(ArrayRef<Slice> NewSlices) {
230  int OldSize = Slices.size();
231  Slices.append(NewSlices.begin(), NewSlices.end());
232  auto SliceI = Slices.begin() + OldSize;
233  std::sort(SliceI, Slices.end());
234  std::inplace_merge(Slices.begin(), SliceI, Slices.end());
235  }
236 
237  // Forward declare the iterator and range accessor for walking the
238  // partitions.
239  class partition_iterator;
241 
242  /// \brief Access the dead users for this alloca.
243  ArrayRef<Instruction *> getDeadUsers() const { return DeadUsers; }
244 
245  /// \brief Access the dead operands referring to this alloca.
246  ///
247  /// These are operands which have cannot actually be used to refer to the
248  /// alloca as they are outside its range and the user doesn't correct for
249  /// that. These mostly consist of PHI node inputs and the like which we just
250  /// need to replace with undef.
251  ArrayRef<Use *> getDeadOperands() const { return DeadOperands; }
252 
253 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
254  void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
255  void printSlice(raw_ostream &OS, const_iterator I,
256  StringRef Indent = " ") const;
257  void printUse(raw_ostream &OS, const_iterator I,
258  StringRef Indent = " ") const;
259  void print(raw_ostream &OS) const;
260  void dump(const_iterator I) const;
261  void dump() const;
262 #endif
263 
264 private:
265  template <typename DerivedT, typename RetT = void> class BuilderBase;
266  class SliceBuilder;
267  friend class AllocaSlices::SliceBuilder;
268 
269 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
270  /// \brief Handle to alloca instruction to simplify method interfaces.
271  AllocaInst &AI;
272 #endif
273 
274  /// \brief The instruction responsible for this alloca not having a known set
275  /// of slices.
276  ///
277  /// When an instruction (potentially) escapes the pointer to the alloca, we
278  /// store a pointer to that here and abort trying to form slices of the
279  /// alloca. This will be null if the alloca slices are analyzed successfully.
280  Instruction *PointerEscapingInstr;
281 
282  /// \brief The slices of the alloca.
283  ///
284  /// We store a vector of the slices formed by uses of the alloca here. This
285  /// vector is sorted by increasing begin offset, and then the unsplittable
286  /// slices before the splittable ones. See the Slice inner class for more
287  /// details.
288  SmallVector<Slice, 8> Slices;
289 
290  /// \brief Instructions which will become dead if we rewrite the alloca.
291  ///
292  /// Note that these are not separated by slice. This is because we expect an
293  /// alloca to be completely rewritten or not rewritten at all. If rewritten,
294  /// all these instructions can simply be removed and replaced with undef as
295  /// they come from outside of the allocated space.
297 
298  /// \brief Operands which will become dead if we rewrite the alloca.
299  ///
300  /// These are operands that in their particular use can be replaced with
301  /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
302  /// to PHI nodes and the like. They aren't entirely dead (there might be
303  /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
304  /// want to swap this particular input for undef to simplify the use lists of
305  /// the alloca.
306  SmallVector<Use *, 8> DeadOperands;
307 };
308 
309 /// \brief A partition of the slices.
310 ///
311 /// An ephemeral representation for a range of slices which can be viewed as
312 /// a partition of the alloca. This range represents a span of the alloca's
313 /// memory which cannot be split, and provides access to all of the slices
314 /// overlapping some part of the partition.
315 ///
316 /// Objects of this type are produced by traversing the alloca's slices, but
317 /// are only ephemeral and not persistent.
319 private:
320  friend class AllocaSlices;
322 
323  typedef AllocaSlices::iterator iterator;
324 
325  /// \brief The beginning and ending offsets of the alloca for this
326  /// partition.
327  uint64_t BeginOffset, EndOffset;
328 
329  /// \brief The start and end iterators of this partition.
330  iterator SI, SJ;
331 
332  /// \brief A collection of split slice tails overlapping the partition.
333  SmallVector<Slice *, 4> SplitTails;
334 
335  /// \brief Raw constructor builds an empty partition starting and ending at
336  /// the given iterator.
337  Partition(iterator SI) : SI(SI), SJ(SI) {}
338 
339 public:
340  /// \brief The start offset of this partition.
341  ///
342  /// All of the contained slices start at or after this offset.
343  uint64_t beginOffset() const { return BeginOffset; }
344 
345  /// \brief The end offset of this partition.
346  ///
347  /// All of the contained slices end at or before this offset.
348  uint64_t endOffset() const { return EndOffset; }
349 
350  /// \brief The size of the partition.
351  ///
352  /// Note that this can never be zero.
353  uint64_t size() const {
354  assert(BeginOffset < EndOffset && "Partitions must span some bytes!");
355  return EndOffset - BeginOffset;
356  }
357 
358  /// \brief Test whether this partition contains no slices, and merely spans
359  /// a region occupied by split slices.
360  bool empty() const { return SI == SJ; }
361 
362  /// \name Iterate slices that start within the partition.
363  /// These may be splittable or unsplittable. They have a begin offset >= the
364  /// partition begin offset.
365  /// @{
366  // FIXME: We should probably define a "concat_iterator" helper and use that
367  // to stitch together pointee_iterators over the split tails and the
368  // contiguous iterators of the partition. That would give a much nicer
369  // interface here. We could then additionally expose filtered iterators for
370  // split, unsplit, and unsplittable splices based on the usage patterns.
371  iterator begin() const { return SI; }
372  iterator end() const { return SJ; }
373  /// @}
374 
375  /// \brief Get the sequence of split slice tails.
376  ///
377  /// These tails are of slices which start before this partition but are
378  /// split and overlap into the partition. We accumulate these while forming
379  /// partitions.
380  ArrayRef<Slice *> splitSliceTails() const { return SplitTails; }
381 };
382 
383 /// \brief An iterator over partitions of the alloca's slices.
384 ///
385 /// This iterator implements the core algorithm for partitioning the alloca's
386 /// slices. It is a forward iterator as we don't support backtracking for
387 /// efficiency reasons, and re-use a single storage area to maintain the
388 /// current set of split slices.
389 ///
390 /// It is templated on the slice iterator type to use so that it can operate
391 /// with either const or non-const slice iterators.
393  : public iterator_facade_base<partition_iterator, std::forward_iterator_tag,
394  Partition> {
395  friend class AllocaSlices;
396 
397  /// \brief Most of the state for walking the partitions is held in a class
398  /// with a nice interface for examining them.
399  Partition P;
400 
401  /// \brief We need to keep the end of the slices to know when to stop.
403 
404  /// \brief We also need to keep track of the maximum split end offset seen.
405  /// FIXME: Do we really?
406  uint64_t MaxSplitSliceEndOffset;
407 
408  /// \brief Sets the partition to be empty at given iterator, and sets the
409  /// end iterator.
411  : P(SI), SE(SE), MaxSplitSliceEndOffset(0) {
412  // If not already at the end, advance our state to form the initial
413  // partition.
414  if (SI != SE)
415  advance();
416  }
417 
418  /// \brief Advance the iterator to the next partition.
419  ///
420  /// Requires that the iterator not be at the end of the slices.
421  void advance() {
422  assert((P.SI != SE || !P.SplitTails.empty()) &&
423  "Cannot advance past the end of the slices!");
424 
425  // Clear out any split uses which have ended.
426  if (!P.SplitTails.empty()) {
427  if (P.EndOffset >= MaxSplitSliceEndOffset) {
428  // If we've finished all splits, this is easy.
429  P.SplitTails.clear();
430  MaxSplitSliceEndOffset = 0;
431  } else {
432  // Remove the uses which have ended in the prior partition. This
433  // cannot change the max split slice end because we just checked that
434  // the prior partition ended prior to that max.
435  P.SplitTails.erase(
436  remove_if(P.SplitTails,
437  [&](Slice *S) { return S->endOffset() <= P.EndOffset; }),
438  P.SplitTails.end());
439  assert(any_of(P.SplitTails,
440  [&](Slice *S) {
441  return S->endOffset() == MaxSplitSliceEndOffset;
442  }) &&
443  "Could not find the current max split slice offset!");
444  assert(all_of(P.SplitTails,
445  [&](Slice *S) {
446  return S->endOffset() <= MaxSplitSliceEndOffset;
447  }) &&
448  "Max split slice end offset is not actually the max!");
449  }
450  }
451 
452  // If P.SI is already at the end, then we've cleared the split tail and
453  // now have an end iterator.
454  if (P.SI == SE) {
455  assert(P.SplitTails.empty() && "Failed to clear the split slices!");
456  return;
457  }
458 
459  // If we had a non-empty partition previously, set up the state for
460  // subsequent partitions.
461  if (P.SI != P.SJ) {
462  // Accumulate all the splittable slices which started in the old
463  // partition into the split list.
464  for (Slice &S : P)
465  if (S.isSplittable() && S.endOffset() > P.EndOffset) {
466  P.SplitTails.push_back(&S);
467  MaxSplitSliceEndOffset =
468  std::max(S.endOffset(), MaxSplitSliceEndOffset);
469  }
470 
471  // Start from the end of the previous partition.
472  P.SI = P.SJ;
473 
474  // If P.SI is now at the end, we at most have a tail of split slices.
475  if (P.SI == SE) {
476  P.BeginOffset = P.EndOffset;
477  P.EndOffset = MaxSplitSliceEndOffset;
478  return;
479  }
480 
481  // If the we have split slices and the next slice is after a gap and is
482  // not splittable immediately form an empty partition for the split
483  // slices up until the next slice begins.
484  if (!P.SplitTails.empty() && P.SI->beginOffset() != P.EndOffset &&
485  !P.SI->isSplittable()) {
486  P.BeginOffset = P.EndOffset;
487  P.EndOffset = P.SI->beginOffset();
488  return;
489  }
490  }
491 
492  // OK, we need to consume new slices. Set the end offset based on the
493  // current slice, and step SJ past it. The beginning offset of the
494  // partition is the beginning offset of the next slice unless we have
495  // pre-existing split slices that are continuing, in which case we begin
496  // at the prior end offset.
497  P.BeginOffset = P.SplitTails.empty() ? P.SI->beginOffset() : P.EndOffset;
498  P.EndOffset = P.SI->endOffset();
499  ++P.SJ;
500 
501  // There are two strategies to form a partition based on whether the
502  // partition starts with an unsplittable slice or a splittable slice.
503  if (!P.SI->isSplittable()) {
504  // When we're forming an unsplittable region, it must always start at
505  // the first slice and will extend through its end.
506  assert(P.BeginOffset == P.SI->beginOffset());
507 
508  // Form a partition including all of the overlapping slices with this
509  // unsplittable slice.
510  while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
511  if (!P.SJ->isSplittable())
512  P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
513  ++P.SJ;
514  }
515 
516  // We have a partition across a set of overlapping unsplittable
517  // partitions.
518  return;
519  }
520 
521  // If we're starting with a splittable slice, then we need to form
522  // a synthetic partition spanning it and any other overlapping splittable
523  // splices.
524  assert(P.SI->isSplittable() && "Forming a splittable partition!");
525 
526  // Collect all of the overlapping splittable slices.
527  while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset &&
528  P.SJ->isSplittable()) {
529  P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
530  ++P.SJ;
531  }
532 
533  // Back upiP.EndOffset if we ended the span early when encountering an
534  // unsplittable slice. This synthesizes the early end offset of
535  // a partition spanning only splittable slices.
536  if (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
537  assert(!P.SJ->isSplittable());
538  P.EndOffset = P.SJ->beginOffset();
539  }
540  }
541 
542 public:
543  bool operator==(const partition_iterator &RHS) const {
544  assert(SE == RHS.SE &&
545  "End iterators don't match between compared partition iterators!");
546 
547  // The observed positions of partitions is marked by the P.SI iterator and
548  // the emptiness of the split slices. The latter is only relevant when
549  // P.SI == SE, as the end iterator will additionally have an empty split
550  // slices list, but the prior may have the same P.SI and a tail of split
551  // slices.
552  if (P.SI == RHS.P.SI && P.SplitTails.empty() == RHS.P.SplitTails.empty()) {
553  assert(P.SJ == RHS.P.SJ &&
554  "Same set of slices formed two different sized partitions!");
555  assert(P.SplitTails.size() == RHS.P.SplitTails.size() &&
556  "Same slice position with differently sized non-empty split "
557  "slice tails!");
558  return true;
559  }
560  return false;
561  }
562 
564  advance();
565  return *this;
566  }
567 
568  Partition &operator*() { return P; }
569 };
570 
571 /// \brief A forward range over the partitions of the alloca's slices.
572 ///
573 /// This accesses an iterator range over the partitions of the alloca's
574 /// slices. It computes these partitions on the fly based on the overlapping
575 /// offsets of the slices and the ability to split them. It will visit "empty"
576 /// partitions to cover regions of the alloca only accessed via split
577 /// slices.
579  return make_range(partition_iterator(begin(), end()),
580  partition_iterator(end(), end()));
581 }
582 
584  // If the condition being selected on is a constant or the same value is
585  // being selected between, fold the select. Yes this does (rarely) happen
586  // early on.
587  if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
588  return SI.getOperand(1 + CI->isZero());
589  if (SI.getOperand(1) == SI.getOperand(2))
590  return SI.getOperand(1);
591 
592  return nullptr;
593 }
594 
595 /// \brief A helper that folds a PHI node or a select.
597  if (PHINode *PN = dyn_cast<PHINode>(&I)) {
598  // If PN merges together the same value, return that value.
599  return PN->hasConstantValue();
600  }
601  return foldSelectInst(cast<SelectInst>(I));
602 }
603 
604 /// \brief Builder for the alloca slices.
605 ///
606 /// This class builds a set of alloca slices by recursively visiting the uses
607 /// of an alloca and making a slice for each load and store at each offset.
608 class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
610  friend class InstVisitor<SliceBuilder>;
612 
613  const uint64_t AllocSize;
614  AllocaSlices &AS;
615 
616  SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
618 
619  /// \brief Set to de-duplicate dead instructions found in the use walk.
620  SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
621 
622 public:
625  AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())), AS(AS) {}
626 
627 private:
628  void markAsDead(Instruction &I) {
629  if (VisitedDeadInsts.insert(&I).second)
630  AS.DeadUsers.push_back(&I);
631  }
632 
633  void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
634  bool IsSplittable = false) {
635  // Completely skip uses which have a zero size or start either before or
636  // past the end of the allocation.
637  if (Size == 0 || Offset.uge(AllocSize)) {
638  DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
639  << " which has zero size or starts outside of the "
640  << AllocSize << " byte alloca:\n"
641  << " alloca: " << AS.AI << "\n"
642  << " use: " << I << "\n");
643  return markAsDead(I);
644  }
645 
646  uint64_t BeginOffset = Offset.getZExtValue();
647  uint64_t EndOffset = BeginOffset + Size;
648 
649  // Clamp the end offset to the end of the allocation. Note that this is
650  // formulated to handle even the case where "BeginOffset + Size" overflows.
651  // This may appear superficially to be something we could ignore entirely,
652  // but that is not so! There may be widened loads or PHI-node uses where
653  // some instructions are dead but not others. We can't completely ignore
654  // them, and so have to record at least the information here.
655  assert(AllocSize >= BeginOffset); // Established above.
656  if (Size > AllocSize - BeginOffset) {
657  DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
658  << " to remain within the " << AllocSize << " byte alloca:\n"
659  << " alloca: " << AS.AI << "\n"
660  << " use: " << I << "\n");
661  EndOffset = AllocSize;
662  }
663 
664  AS.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
665  }
666 
667  void visitBitCastInst(BitCastInst &BC) {
668  if (BC.use_empty())
669  return markAsDead(BC);
670 
671  return Base::visitBitCastInst(BC);
672  }
673 
674  void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
675  if (GEPI.use_empty())
676  return markAsDead(GEPI);
677 
678  if (SROAStrictInbounds && GEPI.isInBounds()) {
679  // FIXME: This is a manually un-factored variant of the basic code inside
680  // of GEPs with checking of the inbounds invariant specified in the
681  // langref in a very strict sense. If we ever want to enable
682  // SROAStrictInbounds, this code should be factored cleanly into
683  // PtrUseVisitor, but it is easier to experiment with SROAStrictInbounds
684  // by writing out the code here where we have the underlying allocation
685  // size readily available.
686  APInt GEPOffset = Offset;
687  const DataLayout &DL = GEPI.getModule()->getDataLayout();
688  for (gep_type_iterator GTI = gep_type_begin(GEPI),
689  GTE = gep_type_end(GEPI);
690  GTI != GTE; ++GTI) {
691  ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
692  if (!OpC)
693  break;
694 
695  // Handle a struct index, which adds its field offset to the pointer.
696  if (StructType *STy = GTI.getStructTypeOrNull()) {
697  unsigned ElementIdx = OpC->getZExtValue();
698  const StructLayout *SL = DL.getStructLayout(STy);
699  GEPOffset +=
700  APInt(Offset.getBitWidth(), SL->getElementOffset(ElementIdx));
701  } else {
702  // For array or vector indices, scale the index by the size of the
703  // type.
704  APInt Index = OpC->getValue().sextOrTrunc(Offset.getBitWidth());
705  GEPOffset += Index * APInt(Offset.getBitWidth(),
706  DL.getTypeAllocSize(GTI.getIndexedType()));
707  }
708 
709  // If this index has computed an intermediate pointer which is not
710  // inbounds, then the result of the GEP is a poison value and we can
711  // delete it and all uses.
712  if (GEPOffset.ugt(AllocSize))
713  return markAsDead(GEPI);
714  }
715  }
716 
717  return Base::visitGetElementPtrInst(GEPI);
718  }
719 
720  void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
721  uint64_t Size, bool IsVolatile) {
722  // We allow splitting of non-volatile loads and stores where the type is an
723  // integer type. These may be used to implement 'memcpy' or other "transfer
724  // of bits" patterns.
725  bool IsSplittable = Ty->isIntegerTy() && !IsVolatile;
726 
727  insertUse(I, Offset, Size, IsSplittable);
728  }
729 
730  void visitLoadInst(LoadInst &LI) {
731  assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
732  "All simple FCA loads should have been pre-split");
733 
734  if (!IsOffsetKnown)
735  return PI.setAborted(&LI);
736 
737  const DataLayout &DL = LI.getModule()->getDataLayout();
738  uint64_t Size = DL.getTypeStoreSize(LI.getType());
739  return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
740  }
741 
742  void visitStoreInst(StoreInst &SI) {
743  Value *ValOp = SI.getValueOperand();
744  if (ValOp == *U)
745  return PI.setEscapedAndAborted(&SI);
746  if (!IsOffsetKnown)
747  return PI.setAborted(&SI);
748 
749  const DataLayout &DL = SI.getModule()->getDataLayout();
750  uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
751 
752  // If this memory access can be shown to *statically* extend outside the
753  // bounds of of the allocation, it's behavior is undefined, so simply
754  // ignore it. Note that this is more strict than the generic clamping
755  // behavior of insertUse. We also try to handle cases which might run the
756  // risk of overflow.
757  // FIXME: We should instead consider the pointer to have escaped if this
758  // function is being instrumented for addressing bugs or race conditions.
759  if (Size > AllocSize || Offset.ugt(AllocSize - Size)) {
760  DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @" << Offset
761  << " which extends past the end of the " << AllocSize
762  << " byte alloca:\n"
763  << " alloca: " << AS.AI << "\n"
764  << " use: " << SI << "\n");
765  return markAsDead(SI);
766  }
767 
768  assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
769  "All simple FCA stores should have been pre-split");
770  handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
771  }
772 
773  void visitMemSetInst(MemSetInst &II) {
774  assert(II.getRawDest() == *U && "Pointer use is not the destination?");
775  ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
776  if ((Length && Length->getValue() == 0) ||
777  (IsOffsetKnown && Offset.uge(AllocSize)))
778  // Zero-length mem transfer intrinsics can be ignored entirely.
779  return markAsDead(II);
780 
781  if (!IsOffsetKnown)
782  return PI.setAborted(&II);
783 
784  insertUse(II, Offset, Length ? Length->getLimitedValue()
785  : AllocSize - Offset.getLimitedValue(),
786  (bool)Length);
787  }
788 
789  void visitMemTransferInst(MemTransferInst &II) {
790  ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
791  if (Length && Length->getValue() == 0)
792  // Zero-length mem transfer intrinsics can be ignored entirely.
793  return markAsDead(II);
794 
795  // Because we can visit these intrinsics twice, also check to see if the
796  // first time marked this instruction as dead. If so, skip it.
797  if (VisitedDeadInsts.count(&II))
798  return;
799 
800  if (!IsOffsetKnown)
801  return PI.setAborted(&II);
802 
803  // This side of the transfer is completely out-of-bounds, and so we can
804  // nuke the entire transfer. However, we also need to nuke the other side
805  // if already added to our partitions.
806  // FIXME: Yet another place we really should bypass this when
807  // instrumenting for ASan.
808  if (Offset.uge(AllocSize)) {
810  MemTransferSliceMap.find(&II);
811  if (MTPI != MemTransferSliceMap.end())
812  AS.Slices[MTPI->second].kill();
813  return markAsDead(II);
814  }
815 
816  uint64_t RawOffset = Offset.getLimitedValue();
817  uint64_t Size = Length ? Length->getLimitedValue() : AllocSize - RawOffset;
818 
819  // Check for the special case where the same exact value is used for both
820  // source and dest.
821  if (*U == II.getRawDest() && *U == II.getRawSource()) {
822  // For non-volatile transfers this is a no-op.
823  if (!II.isVolatile())
824  return markAsDead(II);
825 
826  return insertUse(II, Offset, Size, /*IsSplittable=*/false);
827  }
828 
829  // If we have seen both source and destination for a mem transfer, then
830  // they both point to the same alloca.
831  bool Inserted;
833  std::tie(MTPI, Inserted) =
834  MemTransferSliceMap.insert(std::make_pair(&II, AS.Slices.size()));
835  unsigned PrevIdx = MTPI->second;
836  if (!Inserted) {
837  Slice &PrevP = AS.Slices[PrevIdx];
838 
839  // Check if the begin offsets match and this is a non-volatile transfer.
840  // In that case, we can completely elide the transfer.
841  if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
842  PrevP.kill();
843  return markAsDead(II);
844  }
845 
846  // Otherwise we have an offset transfer within the same alloca. We can't
847  // split those.
848  PrevP.makeUnsplittable();
849  }
850 
851  // Insert the use now that we've fixed up the splittable nature.
852  insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
853 
854  // Check that we ended up with a valid index in the map.
855  assert(AS.Slices[PrevIdx].getUse()->getUser() == &II &&
856  "Map index doesn't point back to a slice with this user.");
857  }
858 
859  // Disable SRoA for any intrinsics except for lifetime invariants.
860  // FIXME: What about debug intrinsics? This matches old behavior, but
861  // doesn't make sense.
862  void visitIntrinsicInst(IntrinsicInst &II) {
863  if (!IsOffsetKnown)
864  return PI.setAborted(&II);
865 
866  if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
867  II.getIntrinsicID() == Intrinsic::lifetime_end) {
868  ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
869  uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
870  Length->getLimitedValue());
871  insertUse(II, Offset, Size, true);
872  return;
873  }
874 
875  Base::visitIntrinsicInst(II);
876  }
877 
878  Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
879  // We consider any PHI or select that results in a direct load or store of
880  // the same offset to be a viable use for slicing purposes. These uses
881  // are considered unsplittable and the size is the maximum loaded or stored
882  // size.
885  Visited.insert(Root);
886  Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
887  const DataLayout &DL = Root->getModule()->getDataLayout();
888  // If there are no loads or stores, the access is dead. We mark that as
889  // a size zero access.
890  Size = 0;
891  do {
892  Instruction *I, *UsedI;
893  std::tie(UsedI, I) = Uses.pop_back_val();
894 
895  if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
896  Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
897  continue;
898  }
899  if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
900  Value *Op = SI->getOperand(0);
901  if (Op == UsedI)
902  return SI;
903  Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
904  continue;
905  }
906 
907  if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
908  if (!GEP->hasAllZeroIndices())
909  return GEP;
910  } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
911  !isa<SelectInst>(I)) {
912  return I;
913  }
914 
915  for (User *U : I->users())
916  if (Visited.insert(cast<Instruction>(U)).second)
917  Uses.push_back(std::make_pair(I, cast<Instruction>(U)));
918  } while (!Uses.empty());
919 
920  return nullptr;
921  }
922 
923  void visitPHINodeOrSelectInst(Instruction &I) {
924  assert(isa<PHINode>(I) || isa<SelectInst>(I));
925  if (I.use_empty())
926  return markAsDead(I);
927 
928  // TODO: We could use SimplifyInstruction here to fold PHINodes and
929  // SelectInsts. However, doing so requires to change the current
930  // dead-operand-tracking mechanism. For instance, suppose neither loading
931  // from %U nor %other traps. Then "load (select undef, %U, %other)" does not
932  // trap either. However, if we simply replace %U with undef using the
933  // current dead-operand-tracking mechanism, "load (select undef, undef,
934  // %other)" may trap because the select may return the first operand
935  // "undef".
936  if (Value *Result = foldPHINodeOrSelectInst(I)) {
937  if (Result == *U)
938  // If the result of the constant fold will be the pointer, recurse
939  // through the PHI/select as if we had RAUW'ed it.
940  enqueueUsers(I);
941  else
942  // Otherwise the operand to the PHI/select is dead, and we can replace
943  // it with undef.
944  AS.DeadOperands.push_back(U);
945 
946  return;
947  }
948 
949  if (!IsOffsetKnown)
950  return PI.setAborted(&I);
951 
952  // See if we already have computed info on this node.
953  uint64_t &Size = PHIOrSelectSizes[&I];
954  if (!Size) {
955  // This is a new PHI/Select, check for an unsafe use of it.
956  if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&I, Size))
957  return PI.setAborted(UnsafeI);
958  }
959 
960  // For PHI and select operands outside the alloca, we can't nuke the entire
961  // phi or select -- the other side might still be relevant, so we special
962  // case them here and use a separate structure to track the operands
963  // themselves which should be replaced with undef.
964  // FIXME: This should instead be escaped in the event we're instrumenting
965  // for address sanitization.
966  if (Offset.uge(AllocSize)) {
967  AS.DeadOperands.push_back(U);
968  return;
969  }
970 
971  insertUse(I, Offset, Size);
972  }
973 
974  void visitPHINode(PHINode &PN) { visitPHINodeOrSelectInst(PN); }
975 
976  void visitSelectInst(SelectInst &SI) { visitPHINodeOrSelectInst(SI); }
977 
978  /// \brief Disable SROA entirely if there are unhandled users of the alloca.
979  void visitInstruction(Instruction &I) { PI.setAborted(&I); }
980 };
981 
983  :
984 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
985  AI(AI),
986 #endif
987  PointerEscapingInstr(nullptr) {
988  SliceBuilder PB(DL, AI, *this);
989  SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
990  if (PtrI.isEscaped() || PtrI.isAborted()) {
991  // FIXME: We should sink the escape vs. abort info into the caller nicely,
992  // possibly by just storing the PtrInfo in the AllocaSlices.
993  PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
994  : PtrI.getAbortingInst();
995  assert(PointerEscapingInstr && "Did not track a bad instruction");
996  return;
997  }
998 
999  Slices.erase(remove_if(Slices, [](const Slice &S) { return S.isDead(); }),
1000  Slices.end());
1001 
1002 #ifndef NDEBUG
1004  std::mt19937 MT(static_cast<unsigned>(
1005  std::chrono::system_clock::now().time_since_epoch().count()));
1006  std::shuffle(Slices.begin(), Slices.end(), MT);
1007  }
1008 #endif
1009 
1010  // Sort the uses. This arranges for the offsets to be in ascending order,
1011  // and the sizes to be in descending order.
1012  std::sort(Slices.begin(), Slices.end());
1013 }
1014 
1015 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1016 
1017 void AllocaSlices::print(raw_ostream &OS, const_iterator I,
1018  StringRef Indent) const {
1019  printSlice(OS, I, Indent);
1020  OS << "\n";
1021  printUse(OS, I, Indent);
1022 }
1023 
1024 void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
1025  StringRef Indent) const {
1026  OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
1027  << " slice #" << (I - begin())
1028  << (I->isSplittable() ? " (splittable)" : "");
1029 }
1030 
1031 void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
1032  StringRef Indent) const {
1033  OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
1034 }
1035 
1036 void AllocaSlices::print(raw_ostream &OS) const {
1037  if (PointerEscapingInstr) {
1038  OS << "Can't analyze slices for alloca: " << AI << "\n"
1039  << " A pointer to this alloca escaped by:\n"
1040  << " " << *PointerEscapingInstr << "\n";
1041  return;
1042  }
1043 
1044  OS << "Slices of alloca: " << AI << "\n";
1045  for (const_iterator I = begin(), E = end(); I != E; ++I)
1046  print(OS, I);
1047 }
1048 
1049 LLVM_DUMP_METHOD void AllocaSlices::dump(const_iterator I) const {
1050  print(dbgs(), I);
1051 }
1052 LLVM_DUMP_METHOD void AllocaSlices::dump() const { print(dbgs()); }
1053 
1054 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1055 
1056 /// Walk the range of a partitioning looking for a common type to cover this
1057 /// sequence of slices.
1060  uint64_t EndOffset) {
1061  Type *Ty = nullptr;
1062  bool TyIsCommon = true;
1063  IntegerType *ITy = nullptr;
1064 
1065  // Note that we need to look at *every* alloca slice's Use to ensure we
1066  // always get consistent results regardless of the order of slices.
1067  for (AllocaSlices::const_iterator I = B; I != E; ++I) {
1068  Use *U = I->getUse();
1069  if (isa<IntrinsicInst>(*U->getUser()))
1070  continue;
1071  if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
1072  continue;
1073 
1074  Type *UserTy = nullptr;
1075  if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1076  UserTy = LI->getType();
1077  } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1078  UserTy = SI->getValueOperand()->getType();
1079  }
1080 
1081  if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
1082  // If the type is larger than the partition, skip it. We only encounter
1083  // this for split integer operations where we want to use the type of the
1084  // entity causing the split. Also skip if the type is not a byte width
1085  // multiple.
1086  if (UserITy->getBitWidth() % 8 != 0 ||
1087  UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
1088  continue;
1089 
1090  // Track the largest bitwidth integer type used in this way in case there
1091  // is no common type.
1092  if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
1093  ITy = UserITy;
1094  }
1095 
1096  // To avoid depending on the order of slices, Ty and TyIsCommon must not
1097  // depend on types skipped above.
1098  if (!UserTy || (Ty && Ty != UserTy))
1099  TyIsCommon = false; // Give up on anything but an iN type.
1100  else
1101  Ty = UserTy;
1102  }
1103 
1104  return TyIsCommon ? Ty : ITy;
1105 }
1106 
1107 /// PHI instructions that use an alloca and are subsequently loaded can be
1108 /// rewritten to load both input pointers in the pred blocks and then PHI the
1109 /// results, allowing the load of the alloca to be promoted.
1110 /// From this:
1111 /// %P2 = phi [i32* %Alloca, i32* %Other]
1112 /// %V = load i32* %P2
1113 /// to:
1114 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1115 /// ...
1116 /// %V2 = load i32* %Other
1117 /// ...
1118 /// %V = phi [i32 %V1, i32 %V2]
1119 ///
1120 /// We can do this to a select if its only uses are loads and if the operands
1121 /// to the select can be loaded unconditionally.
1122 ///
1123 /// FIXME: This should be hoisted into a generic utility, likely in
1124 /// Transforms/Util/Local.h
1125 static bool isSafePHIToSpeculate(PHINode &PN) {
1126  // For now, we can only do this promotion if the load is in the same block
1127  // as the PHI, and if there are no stores between the phi and load.
1128  // TODO: Allow recursive phi users.
1129  // TODO: Allow stores.
1130  BasicBlock *BB = PN.getParent();
1131  unsigned MaxAlign = 0;
1132  bool HaveLoad = false;
1133  for (User *U : PN.users()) {
1134  LoadInst *LI = dyn_cast<LoadInst>(U);
1135  if (!LI || !LI->isSimple())
1136  return false;
1137 
1138  // For now we only allow loads in the same block as the PHI. This is
1139  // a common case that happens when instcombine merges two loads through
1140  // a PHI.
1141  if (LI->getParent() != BB)
1142  return false;
1143 
1144  // Ensure that there are no instructions between the PHI and the load that
1145  // could store.
1146  for (BasicBlock::iterator BBI(PN); &*BBI != LI; ++BBI)
1147  if (BBI->mayWriteToMemory())
1148  return false;
1149 
1150  MaxAlign = std::max(MaxAlign, LI->getAlignment());
1151  HaveLoad = true;
1152  }
1153 
1154  if (!HaveLoad)
1155  return false;
1156 
1157  const DataLayout &DL = PN.getModule()->getDataLayout();
1158 
1159  // We can only transform this if it is safe to push the loads into the
1160  // predecessor blocks. The only thing to watch out for is that we can't put
1161  // a possibly trapping load in the predecessor if it is a critical edge.
1162  for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1164  Value *InVal = PN.getIncomingValue(Idx);
1165 
1166  // If the value is produced by the terminator of the predecessor (an
1167  // invoke) or it has side-effects, there is no valid place to put a load
1168  // in the predecessor.
1169  if (TI == InVal || TI->mayHaveSideEffects())
1170  return false;
1171 
1172  // If the predecessor has a single successor, then the edge isn't
1173  // critical.
1174  if (TI->getNumSuccessors() == 1)
1175  continue;
1176 
1177  // If this pointer is always safe to load, or if we can prove that there
1178  // is already a load in the block, then we can move the load to the pred
1179  // block.
1180  if (isSafeToLoadUnconditionally(InVal, MaxAlign, DL, TI))
1181  continue;
1182 
1183  return false;
1184  }
1185 
1186  return true;
1187 }
1188 
1189 static void speculatePHINodeLoads(PHINode &PN) {
1190  DEBUG(dbgs() << " original: " << PN << "\n");
1191 
1192  Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1193  IRBuilderTy PHIBuilder(&PN);
1194  PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1195  PN.getName() + ".sroa.speculated");
1196 
1197  // Get the AA tags and alignment to use from one of the loads. It doesn't
1198  // matter which one we get and if any differ.
1199  LoadInst *SomeLoad = cast<LoadInst>(PN.user_back());
1200 
1201  AAMDNodes AATags;
1202  SomeLoad->getAAMetadata(AATags);
1203  unsigned Align = SomeLoad->getAlignment();
1204 
1205  // Rewrite all loads of the PN to use the new PHI.
1206  while (!PN.use_empty()) {
1207  LoadInst *LI = cast<LoadInst>(PN.user_back());
1208  LI->replaceAllUsesWith(NewPN);
1209  LI->eraseFromParent();
1210  }
1211 
1212  // Inject loads into all of the pred blocks.
1213  for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1214  BasicBlock *Pred = PN.getIncomingBlock(Idx);
1215  TerminatorInst *TI = Pred->getTerminator();
1216  Value *InVal = PN.getIncomingValue(Idx);
1217  IRBuilderTy PredBuilder(TI);
1218 
1219  LoadInst *Load = PredBuilder.CreateLoad(
1220  InVal, (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1221  ++NumLoadsSpeculated;
1222  Load->setAlignment(Align);
1223  if (AATags)
1224  Load->setAAMetadata(AATags);
1225  NewPN->addIncoming(Load, Pred);
1226  }
1227 
1228  DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1229  PN.eraseFromParent();
1230 }
1231 
1232 /// Select instructions that use an alloca and are subsequently loaded can be
1233 /// rewritten to load both input pointers and then select between the result,
1234 /// allowing the load of the alloca to be promoted.
1235 /// From this:
1236 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1237 /// %V = load i32* %P2
1238 /// to:
1239 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1240 /// %V2 = load i32* %Other
1241 /// %V = select i1 %cond, i32 %V1, i32 %V2
1242 ///
1243 /// We can do this to a select if its only uses are loads and if the operand
1244 /// to the select can be loaded unconditionally.
1246  Value *TValue = SI.getTrueValue();
1247  Value *FValue = SI.getFalseValue();
1248  const DataLayout &DL = SI.getModule()->getDataLayout();
1249 
1250  for (User *U : SI.users()) {
1251  LoadInst *LI = dyn_cast<LoadInst>(U);
1252  if (!LI || !LI->isSimple())
1253  return false;
1254 
1255  // Both operands to the select need to be dereferenceable, either
1256  // absolutely (e.g. allocas) or at this point because we can see other
1257  // accesses to it.
1258  if (!isSafeToLoadUnconditionally(TValue, LI->getAlignment(), DL, LI))
1259  return false;
1260  if (!isSafeToLoadUnconditionally(FValue, LI->getAlignment(), DL, LI))
1261  return false;
1262  }
1263 
1264  return true;
1265 }
1266 
1268  DEBUG(dbgs() << " original: " << SI << "\n");
1269 
1270  IRBuilderTy IRB(&SI);
1271  Value *TV = SI.getTrueValue();
1272  Value *FV = SI.getFalseValue();
1273  // Replace the loads of the select with a select of two loads.
1274  while (!SI.use_empty()) {
1275  LoadInst *LI = cast<LoadInst>(SI.user_back());
1276  assert(LI->isSimple() && "We only speculate simple loads");
1277 
1278  IRB.SetInsertPoint(LI);
1279  LoadInst *TL =
1280  IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1281  LoadInst *FL =
1282  IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1283  NumLoadsSpeculated += 2;
1284 
1285  // Transfer alignment and AA info if present.
1286  TL->setAlignment(LI->getAlignment());
1287  FL->setAlignment(LI->getAlignment());
1288 
1289  AAMDNodes Tags;
1290  LI->getAAMetadata(Tags);
1291  if (Tags) {
1292  TL->setAAMetadata(Tags);
1293  FL->setAAMetadata(Tags);
1294  }
1295 
1296  Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1297  LI->getName() + ".sroa.speculated");
1298 
1299  DEBUG(dbgs() << " speculated to: " << *V << "\n");
1300  LI->replaceAllUsesWith(V);
1301  LI->eraseFromParent();
1302  }
1303  SI.eraseFromParent();
1304 }
1305 
1306 /// \brief Build a GEP out of a base pointer and indices.
1307 ///
1308 /// This will return the BasePtr if that is valid, or build a new GEP
1309 /// instruction using the IRBuilder if GEP-ing is needed.
1310 static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1311  SmallVectorImpl<Value *> &Indices, Twine NamePrefix) {
1312  if (Indices.empty())
1313  return BasePtr;
1314 
1315  // A single zero index is a no-op, so check for this and avoid building a GEP
1316  // in that case.
1317  if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1318  return BasePtr;
1319 
1320  return IRB.CreateInBoundsGEP(nullptr, BasePtr, Indices,
1321  NamePrefix + "sroa_idx");
1322 }
1323 
1324 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1325 /// TargetTy without changing the offset of the pointer.
1326 ///
1327 /// This routine assumes we've already established a properly offset GEP with
1328 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1329 /// zero-indices down through type layers until we find one the same as
1330 /// TargetTy. If we can't find one with the same type, we at least try to use
1331 /// one with the same size. If none of that works, we just produce the GEP as
1332 /// indicated by Indices to have the correct offset.
1333 static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1334  Value *BasePtr, Type *Ty, Type *TargetTy,
1335  SmallVectorImpl<Value *> &Indices,
1336  Twine NamePrefix) {
1337  if (Ty == TargetTy)
1338  return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1339 
1340  // Pointer size to use for the indices.
1341  unsigned PtrSize = DL.getPointerTypeSizeInBits(BasePtr->getType());
1342 
1343  // See if we can descend into a struct and locate a field with the correct
1344  // type.
1345  unsigned NumLayers = 0;
1346  Type *ElementTy = Ty;
1347  do {
1348  if (ElementTy->isPointerTy())
1349  break;
1350 
1351  if (ArrayType *ArrayTy = dyn_cast<ArrayType>(ElementTy)) {
1352  ElementTy = ArrayTy->getElementType();
1353  Indices.push_back(IRB.getIntN(PtrSize, 0));
1354  } else if (VectorType *VectorTy = dyn_cast<VectorType>(ElementTy)) {
1355  ElementTy = VectorTy->getElementType();
1356  Indices.push_back(IRB.getInt32(0));
1357  } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1358  if (STy->element_begin() == STy->element_end())
1359  break; // Nothing left to descend into.
1360  ElementTy = *STy->element_begin();
1361  Indices.push_back(IRB.getInt32(0));
1362  } else {
1363  break;
1364  }
1365  ++NumLayers;
1366  } while (ElementTy != TargetTy);
1367  if (ElementTy != TargetTy)
1368  Indices.erase(Indices.end() - NumLayers, Indices.end());
1369 
1370  return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1371 }
1372 
1373 /// \brief Recursively compute indices for a natural GEP.
1374 ///
1375 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1376 /// element types adding appropriate indices for the GEP.
1377 static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1378  Value *Ptr, Type *Ty, APInt &Offset,
1379  Type *TargetTy,
1380  SmallVectorImpl<Value *> &Indices,
1381  Twine NamePrefix) {
1382  if (Offset == 0)
1383  return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices,
1384  NamePrefix);
1385 
1386  // We can't recurse through pointer types.
1387  if (Ty->isPointerTy())
1388  return nullptr;
1389 
1390  // We try to analyze GEPs over vectors here, but note that these GEPs are
1391  // extremely poorly defined currently. The long-term goal is to remove GEPing
1392  // over a vector from the IR completely.
1393  if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1394  unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
1395  if (ElementSizeInBits % 8 != 0) {
1396  // GEPs over non-multiple of 8 size vector elements are invalid.
1397  return nullptr;
1398  }
1399  APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1400  APInt NumSkippedElements = Offset.sdiv(ElementSize);
1401  if (NumSkippedElements.ugt(VecTy->getNumElements()))
1402  return nullptr;
1403  Offset -= NumSkippedElements * ElementSize;
1404  Indices.push_back(IRB.getInt(NumSkippedElements));
1405  return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1406  Offset, TargetTy, Indices, NamePrefix);
1407  }
1408 
1409  if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1410  Type *ElementTy = ArrTy->getElementType();
1411  APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1412  APInt NumSkippedElements = Offset.sdiv(ElementSize);
1413  if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1414  return nullptr;
1415 
1416  Offset -= NumSkippedElements * ElementSize;
1417  Indices.push_back(IRB.getInt(NumSkippedElements));
1418  return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1419  Indices, NamePrefix);
1420  }
1421 
1422  StructType *STy = dyn_cast<StructType>(Ty);
1423  if (!STy)
1424  return nullptr;
1425 
1426  const StructLayout *SL = DL.getStructLayout(STy);
1427  uint64_t StructOffset = Offset.getZExtValue();
1428  if (StructOffset >= SL->getSizeInBytes())
1429  return nullptr;
1430  unsigned Index = SL->getElementContainingOffset(StructOffset);
1431  Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1432  Type *ElementTy = STy->getElementType(Index);
1433  if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
1434  return nullptr; // The offset points into alignment padding.
1435 
1436  Indices.push_back(IRB.getInt32(Index));
1437  return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1438  Indices, NamePrefix);
1439 }
1440 
1441 /// \brief Get a natural GEP from a base pointer to a particular offset and
1442 /// resulting in a particular type.
1443 ///
1444 /// The goal is to produce a "natural" looking GEP that works with the existing
1445 /// composite types to arrive at the appropriate offset and element type for
1446 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1447 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1448 /// Indices, and setting Ty to the result subtype.
1449 ///
1450 /// If no natural GEP can be constructed, this function returns null.
1451 static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1452  Value *Ptr, APInt Offset, Type *TargetTy,
1453  SmallVectorImpl<Value *> &Indices,
1454  Twine NamePrefix) {
1455  PointerType *Ty = cast<PointerType>(Ptr->getType());
1456 
1457  // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1458  // an i8.
1459  if (Ty == IRB.getInt8PtrTy(Ty->getAddressSpace()) && TargetTy->isIntegerTy(8))
1460  return nullptr;
1461 
1462  Type *ElementTy = Ty->getElementType();
1463  if (!ElementTy->isSized())
1464  return nullptr; // We can't GEP through an unsized element.
1465  APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1466  if (ElementSize == 0)
1467  return nullptr; // Zero-length arrays can't help us build a natural GEP.
1468  APInt NumSkippedElements = Offset.sdiv(ElementSize);
1469 
1470  Offset -= NumSkippedElements * ElementSize;
1471  Indices.push_back(IRB.getInt(NumSkippedElements));
1472  return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1473  Indices, NamePrefix);
1474 }
1475 
1476 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1477 /// resulting pointer has PointerTy.
1478 ///
1479 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1480 /// and produces the pointer type desired. Where it cannot, it will try to use
1481 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1482 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1483 /// bitcast to the type.
1484 ///
1485 /// The strategy for finding the more natural GEPs is to peel off layers of the
1486 /// pointer, walking back through bit casts and GEPs, searching for a base
1487 /// pointer from which we can compute a natural GEP with the desired
1488 /// properties. The algorithm tries to fold as many constant indices into
1489 /// a single GEP as possible, thus making each GEP more independent of the
1490 /// surrounding code.
1491 static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr,
1492  APInt Offset, Type *PointerTy, Twine NamePrefix) {
1493  // Even though we don't look through PHI nodes, we could be called on an
1494  // instruction in an unreachable block, which may be on a cycle.
1495  SmallPtrSet<Value *, 4> Visited;
1496  Visited.insert(Ptr);
1497  SmallVector<Value *, 4> Indices;
1498 
1499  // We may end up computing an offset pointer that has the wrong type. If we
1500  // never are able to compute one directly that has the correct type, we'll
1501  // fall back to it, so keep it and the base it was computed from around here.
1502  Value *OffsetPtr = nullptr;
1503  Value *OffsetBasePtr;
1504 
1505  // Remember any i8 pointer we come across to re-use if we need to do a raw
1506  // byte offset.
1507  Value *Int8Ptr = nullptr;
1508  APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1509 
1510  Type *TargetTy = PointerTy->getPointerElementType();
1511 
1512  do {
1513  // First fold any existing GEPs into the offset.
1514  while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1515  APInt GEPOffset(Offset.getBitWidth(), 0);
1516  if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1517  break;
1518  Offset += GEPOffset;
1519  Ptr = GEP->getPointerOperand();
1520  if (!Visited.insert(Ptr).second)
1521  break;
1522  }
1523 
1524  // See if we can perform a natural GEP here.
1525  Indices.clear();
1526  if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1527  Indices, NamePrefix)) {
1528  // If we have a new natural pointer at the offset, clear out any old
1529  // offset pointer we computed. Unless it is the base pointer or
1530  // a non-instruction, we built a GEP we don't need. Zap it.
1531  if (OffsetPtr && OffsetPtr != OffsetBasePtr)
1532  if (Instruction *I = dyn_cast<Instruction>(OffsetPtr)) {
1533  assert(I->use_empty() && "Built a GEP with uses some how!");
1534  I->eraseFromParent();
1535  }
1536  OffsetPtr = P;
1537  OffsetBasePtr = Ptr;
1538  // If we also found a pointer of the right type, we're done.
1539  if (P->getType() == PointerTy)
1540  return P;
1541  }
1542 
1543  // Stash this pointer if we've found an i8*.
1544  if (Ptr->getType()->isIntegerTy(8)) {
1545  Int8Ptr = Ptr;
1546  Int8PtrOffset = Offset;
1547  }
1548 
1549  // Peel off a layer of the pointer and update the offset appropriately.
1550  if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1551  Ptr = cast<Operator>(Ptr)->getOperand(0);
1552  } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1553  if (GA->isInterposable())
1554  break;
1555  Ptr = GA->getAliasee();
1556  } else {
1557  break;
1558  }
1559  assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1560  } while (Visited.insert(Ptr).second);
1561 
1562  if (!OffsetPtr) {
1563  if (!Int8Ptr) {
1564  Int8Ptr = IRB.CreateBitCast(
1565  Ptr, IRB.getInt8PtrTy(PointerTy->getPointerAddressSpace()),
1566  NamePrefix + "sroa_raw_cast");
1567  Int8PtrOffset = Offset;
1568  }
1569 
1570  OffsetPtr = Int8PtrOffset == 0
1571  ? Int8Ptr
1572  : IRB.CreateInBoundsGEP(IRB.getInt8Ty(), Int8Ptr,
1573  IRB.getInt(Int8PtrOffset),
1574  NamePrefix + "sroa_raw_idx");
1575  }
1576  Ptr = OffsetPtr;
1577 
1578  // On the off chance we were targeting i8*, guard the bitcast here.
1579  if (Ptr->getType() != PointerTy)
1580  Ptr = IRB.CreateBitCast(Ptr, PointerTy, NamePrefix + "sroa_cast");
1581 
1582  return Ptr;
1583 }
1584 
1585 /// \brief Compute the adjusted alignment for a load or store from an offset.
1586 static unsigned getAdjustedAlignment(Instruction *I, uint64_t Offset,
1587  const DataLayout &DL) {
1588  unsigned Alignment;
1589  Type *Ty;
1590  if (auto *LI = dyn_cast<LoadInst>(I)) {
1591  Alignment = LI->getAlignment();
1592  Ty = LI->getType();
1593  } else if (auto *SI = dyn_cast<StoreInst>(I)) {
1594  Alignment = SI->getAlignment();
1595  Ty = SI->getValueOperand()->getType();
1596  } else {
1597  llvm_unreachable("Only loads and stores are allowed!");
1598  }
1599 
1600  if (!Alignment)
1601  Alignment = DL.getABITypeAlignment(Ty);
1602 
1603  return MinAlign(Alignment, Offset);
1604 }
1605 
1606 /// \brief Test whether we can convert a value from the old to the new type.
1607 ///
1608 /// This predicate should be used to guard calls to convertValue in order to
1609 /// ensure that we only try to convert viable values. The strategy is that we
1610 /// will peel off single element struct and array wrappings to get to an
1611 /// underlying value, and convert that value.
1612 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1613  if (OldTy == NewTy)
1614  return true;
1615 
1616  // For integer types, we can't handle any bit-width differences. This would
1617  // break both vector conversions with extension and introduce endianness
1618  // issues when in conjunction with loads and stores.
1619  if (isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) {
1620  assert(cast<IntegerType>(OldTy)->getBitWidth() !=
1621  cast<IntegerType>(NewTy)->getBitWidth() &&
1622  "We can't have the same bitwidth for different int types");
1623  return false;
1624  }
1625 
1626  if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1627  return false;
1628  if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1629  return false;
1630 
1631  // We can convert pointers to integers and vice-versa. Same for vectors
1632  // of pointers and integers.
1633  OldTy = OldTy->getScalarType();
1634  NewTy = NewTy->getScalarType();
1635  if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1636  if (NewTy->isPointerTy() && OldTy->isPointerTy()) {
1637  return cast<PointerType>(NewTy)->getPointerAddressSpace() ==
1638  cast<PointerType>(OldTy)->getPointerAddressSpace();
1639  }
1640 
1641  // We can convert integers to integral pointers, but not to non-integral
1642  // pointers.
1643  if (OldTy->isIntegerTy())
1644  return !DL.isNonIntegralPointerType(NewTy);
1645 
1646  // We can convert integral pointers to integers, but non-integral pointers
1647  // need to remain pointers.
1648  if (!DL.isNonIntegralPointerType(OldTy))
1649  return NewTy->isIntegerTy();
1650 
1651  return false;
1652  }
1653 
1654  return true;
1655 }
1656 
1657 /// \brief Generic routine to convert an SSA value to a value of a different
1658 /// type.
1659 ///
1660 /// This will try various different casting techniques, such as bitcasts,
1661 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1662 /// two types for viability with this routine.
1663 static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1664  Type *NewTy) {
1665  Type *OldTy = V->getType();
1666  assert(canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type");
1667 
1668  if (OldTy == NewTy)
1669  return V;
1670 
1671  assert(!(isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) &&
1672  "Integer types must be the exact same to convert.");
1673 
1674  // See if we need inttoptr for this type pair. A cast involving both scalars
1675  // and vectors requires and additional bitcast.
1676  if (OldTy->isIntOrIntVectorTy() && NewTy->isPtrOrPtrVectorTy()) {
1677  // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1678  if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1679  return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1680  NewTy);
1681 
1682  // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1683  if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1684  return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1685  NewTy);
1686 
1687  return IRB.CreateIntToPtr(V, NewTy);
1688  }
1689 
1690  // See if we need ptrtoint for this type pair. A cast involving both scalars
1691  // and vectors requires and additional bitcast.
1692  if (OldTy->isPtrOrPtrVectorTy() && NewTy->isIntOrIntVectorTy()) {
1693  // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1694  if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1695  return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1696  NewTy);
1697 
1698  // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1699  if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1700  return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1701  NewTy);
1702 
1703  return IRB.CreatePtrToInt(V, NewTy);
1704  }
1705 
1706  return IRB.CreateBitCast(V, NewTy);
1707 }
1708 
1709 /// \brief Test whether the given slice use can be promoted to a vector.
1710 ///
1711 /// This function is called to test each entry in a partition which is slated
1712 /// for a single slice.
1713 static bool isVectorPromotionViableForSlice(Partition &P, const Slice &S,
1714  VectorType *Ty,
1715  uint64_t ElementSize,
1716  const DataLayout &DL) {
1717  // First validate the slice offsets.
1718  uint64_t BeginOffset =
1719  std::max(S.beginOffset(), P.beginOffset()) - P.beginOffset();
1720  uint64_t BeginIndex = BeginOffset / ElementSize;
1721  if (BeginIndex * ElementSize != BeginOffset ||
1722  BeginIndex >= Ty->getNumElements())
1723  return false;
1724  uint64_t EndOffset =
1725  std::min(S.endOffset(), P.endOffset()) - P.beginOffset();
1726  uint64_t EndIndex = EndOffset / ElementSize;
1727  if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
1728  return false;
1729 
1730  assert(EndIndex > BeginIndex && "Empty vector!");
1731  uint64_t NumElements = EndIndex - BeginIndex;
1732  Type *SliceTy = (NumElements == 1)
1733  ? Ty->getElementType()
1734  : VectorType::get(Ty->getElementType(), NumElements);
1735 
1736  Type *SplitIntTy =
1737  Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1738 
1739  Use *U = S.getUse();
1740 
1741  if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1742  if (MI->isVolatile())
1743  return false;
1744  if (!S.isSplittable())
1745  return false; // Skip any unsplittable intrinsics.
1746  } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1747  if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1748  II->getIntrinsicID() != Intrinsic::lifetime_end)
1749  return false;
1750  } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1751  // Disable vector promotion when there are loads or stores of an FCA.
1752  return false;
1753  } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1754  if (LI->isVolatile())
1755  return false;
1756  Type *LTy = LI->getType();
1757  if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1758  assert(LTy->isIntegerTy());
1759  LTy = SplitIntTy;
1760  }
1761  if (!canConvertValue(DL, SliceTy, LTy))
1762  return false;
1763  } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1764  if (SI->isVolatile())
1765  return false;
1766  Type *STy = SI->getValueOperand()->getType();
1767  if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1768  assert(STy->isIntegerTy());
1769  STy = SplitIntTy;
1770  }
1771  if (!canConvertValue(DL, STy, SliceTy))
1772  return false;
1773  } else {
1774  return false;
1775  }
1776 
1777  return true;
1778 }
1779 
1780 /// \brief Test whether the given alloca partitioning and range of slices can be
1781 /// promoted to a vector.
1782 ///
1783 /// This is a quick test to check whether we can rewrite a particular alloca
1784 /// partition (and its newly formed alloca) into a vector alloca with only
1785 /// whole-vector loads and stores such that it could be promoted to a vector
1786 /// SSA value. We only can ensure this for a limited set of operations, and we
1787 /// don't want to do the rewrites unless we are confident that the result will
1788 /// be promotable, so we have an early test here.
1790  // Collect the candidate types for vector-based promotion. Also track whether
1791  // we have different element types.
1792  SmallVector<VectorType *, 4> CandidateTys;
1793  Type *CommonEltTy = nullptr;
1794  bool HaveCommonEltTy = true;
1795  auto CheckCandidateType = [&](Type *Ty) {
1796  if (auto *VTy = dyn_cast<VectorType>(Ty)) {
1797  CandidateTys.push_back(VTy);
1798  if (!CommonEltTy)
1799  CommonEltTy = VTy->getElementType();
1800  else if (CommonEltTy != VTy->getElementType())
1801  HaveCommonEltTy = false;
1802  }
1803  };
1804  // Consider any loads or stores that are the exact size of the slice.
1805  for (const Slice &S : P)
1806  if (S.beginOffset() == P.beginOffset() &&
1807  S.endOffset() == P.endOffset()) {
1808  if (auto *LI = dyn_cast<LoadInst>(S.getUse()->getUser()))
1809  CheckCandidateType(LI->getType());
1810  else if (auto *SI = dyn_cast<StoreInst>(S.getUse()->getUser()))
1811  CheckCandidateType(SI->getValueOperand()->getType());
1812  }
1813 
1814  // If we didn't find a vector type, nothing to do here.
1815  if (CandidateTys.empty())
1816  return nullptr;
1817 
1818  // Remove non-integer vector types if we had multiple common element types.
1819  // FIXME: It'd be nice to replace them with integer vector types, but we can't
1820  // do that until all the backends are known to produce good code for all
1821  // integer vector types.
1822  if (!HaveCommonEltTy) {
1823  CandidateTys.erase(remove_if(CandidateTys,
1824  [](VectorType *VTy) {
1825  return !VTy->getElementType()->isIntegerTy();
1826  }),
1827  CandidateTys.end());
1828 
1829  // If there were no integer vector types, give up.
1830  if (CandidateTys.empty())
1831  return nullptr;
1832 
1833  // Rank the remaining candidate vector types. This is easy because we know
1834  // they're all integer vectors. We sort by ascending number of elements.
1835  auto RankVectorTypes = [&DL](VectorType *RHSTy, VectorType *LHSTy) {
1836  (void)DL;
1837  assert(DL.getTypeSizeInBits(RHSTy) == DL.getTypeSizeInBits(LHSTy) &&
1838  "Cannot have vector types of different sizes!");
1839  assert(RHSTy->getElementType()->isIntegerTy() &&
1840  "All non-integer types eliminated!");
1841  assert(LHSTy->getElementType()->isIntegerTy() &&
1842  "All non-integer types eliminated!");
1843  return RHSTy->getNumElements() < LHSTy->getNumElements();
1844  };
1845  std::sort(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes);
1846  CandidateTys.erase(
1847  std::unique(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes),
1848  CandidateTys.end());
1849  } else {
1850 // The only way to have the same element type in every vector type is to
1851 // have the same vector type. Check that and remove all but one.
1852 #ifndef NDEBUG
1853  for (VectorType *VTy : CandidateTys) {
1854  assert(VTy->getElementType() == CommonEltTy &&
1855  "Unaccounted for element type!");
1856  assert(VTy == CandidateTys[0] &&
1857  "Different vector types with the same element type!");
1858  }
1859 #endif
1860  CandidateTys.resize(1);
1861  }
1862 
1863  // Try each vector type, and return the one which works.
1864  auto CheckVectorTypeForPromotion = [&](VectorType *VTy) {
1865  uint64_t ElementSize = DL.getTypeSizeInBits(VTy->getElementType());
1866 
1867  // While the definition of LLVM vectors is bitpacked, we don't support sizes
1868  // that aren't byte sized.
1869  if (ElementSize % 8)
1870  return false;
1871  assert((DL.getTypeSizeInBits(VTy) % 8) == 0 &&
1872  "vector size not a multiple of element size?");
1873  ElementSize /= 8;
1874 
1875  for (const Slice &S : P)
1876  if (!isVectorPromotionViableForSlice(P, S, VTy, ElementSize, DL))
1877  return false;
1878 
1879  for (const Slice *S : P.splitSliceTails())
1880  if (!isVectorPromotionViableForSlice(P, *S, VTy, ElementSize, DL))
1881  return false;
1882 
1883  return true;
1884  };
1885  for (VectorType *VTy : CandidateTys)
1886  if (CheckVectorTypeForPromotion(VTy))
1887  return VTy;
1888 
1889  return nullptr;
1890 }
1891 
1892 /// \brief Test whether a slice of an alloca is valid for integer widening.
1893 ///
1894 /// This implements the necessary checking for the \c isIntegerWideningViable
1895 /// test below on a single slice of the alloca.
1896 static bool isIntegerWideningViableForSlice(const Slice &S,
1897  uint64_t AllocBeginOffset,
1898  Type *AllocaTy,
1899  const DataLayout &DL,
1900  bool &WholeAllocaOp) {
1901  uint64_t Size = DL.getTypeStoreSize(AllocaTy);
1902 
1903  uint64_t RelBegin = S.beginOffset() - AllocBeginOffset;
1904  uint64_t RelEnd = S.endOffset() - AllocBeginOffset;
1905 
1906  // We can't reasonably handle cases where the load or store extends past
1907  // the end of the alloca's type and into its padding.
1908  if (RelEnd > Size)
1909  return false;
1910 
1911  Use *U = S.getUse();
1912 
1913  if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1914  if (LI->isVolatile())
1915  return false;
1916  // We can't handle loads that extend past the allocated memory.
1917  if (DL.getTypeStoreSize(LI->getType()) > Size)
1918  return false;
1919  // Note that we don't count vector loads or stores as whole-alloca
1920  // operations which enable integer widening because we would prefer to use
1921  // vector widening instead.
1922  if (!isa<VectorType>(LI->getType()) && RelBegin == 0 && RelEnd == Size)
1923  WholeAllocaOp = true;
1924  if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
1925  if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1926  return false;
1927  } else if (RelBegin != 0 || RelEnd != Size ||
1928  !canConvertValue(DL, AllocaTy, LI->getType())) {
1929  // Non-integer loads need to be convertible from the alloca type so that
1930  // they are promotable.
1931  return false;
1932  }
1933  } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1934  Type *ValueTy = SI->getValueOperand()->getType();
1935  if (SI->isVolatile())
1936  return false;
1937  // We can't handle stores that extend past the allocated memory.
1938  if (DL.getTypeStoreSize(ValueTy) > Size)
1939  return false;
1940  // Note that we don't count vector loads or stores as whole-alloca
1941  // operations which enable integer widening because we would prefer to use
1942  // vector widening instead.
1943  if (!isa<VectorType>(ValueTy) && RelBegin == 0 && RelEnd == Size)
1944  WholeAllocaOp = true;
1945  if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
1946  if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1947  return false;
1948  } else if (RelBegin != 0 || RelEnd != Size ||
1949  !canConvertValue(DL, ValueTy, AllocaTy)) {
1950  // Non-integer stores need to be convertible to the alloca type so that
1951  // they are promotable.
1952  return false;
1953  }
1954  } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1955  if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
1956  return false;
1957  if (!S.isSplittable())
1958  return false; // Skip any unsplittable intrinsics.
1959  } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1960  if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1961  II->getIntrinsicID() != Intrinsic::lifetime_end)
1962  return false;
1963  } else {
1964  return false;
1965  }
1966 
1967  return true;
1968 }
1969 
1970 /// \brief Test whether the given alloca partition's integer operations can be
1971 /// widened to promotable ones.
1972 ///
1973 /// This is a quick test to check whether we can rewrite the integer loads and
1974 /// stores to a particular alloca into wider loads and stores and be able to
1975 /// promote the resulting alloca.
1976 static bool isIntegerWideningViable(Partition &P, Type *AllocaTy,
1977  const DataLayout &DL) {
1978  uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
1979  // Don't create integer types larger than the maximum bitwidth.
1980  if (SizeInBits > IntegerType::MAX_INT_BITS)
1981  return false;
1982 
1983  // Don't try to handle allocas with bit-padding.
1984  if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
1985  return false;
1986 
1987  // We need to ensure that an integer type with the appropriate bitwidth can
1988  // be converted to the alloca type, whatever that is. We don't want to force
1989  // the alloca itself to have an integer type if there is a more suitable one.
1990  Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
1991  if (!canConvertValue(DL, AllocaTy, IntTy) ||
1992  !canConvertValue(DL, IntTy, AllocaTy))
1993  return false;
1994 
1995  // While examining uses, we ensure that the alloca has a covering load or
1996  // store. We don't want to widen the integer operations only to fail to
1997  // promote due to some other unsplittable entry (which we may make splittable
1998  // later). However, if there are only splittable uses, go ahead and assume
1999  // that we cover the alloca.
2000  // FIXME: We shouldn't consider split slices that happen to start in the
2001  // partition here...
2002  bool WholeAllocaOp =
2003  P.begin() != P.end() ? false : DL.isLegalInteger(SizeInBits);
2004 
2005  for (const Slice &S : P)
2006  if (!isIntegerWideningViableForSlice(S, P.beginOffset(), AllocaTy, DL,
2007  WholeAllocaOp))
2008  return false;
2009 
2010  for (const Slice *S : P.splitSliceTails())
2011  if (!isIntegerWideningViableForSlice(*S, P.beginOffset(), AllocaTy, DL,
2012  WholeAllocaOp))
2013  return false;
2014 
2015  return WholeAllocaOp;
2016 }
2017 
2018 static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
2019  IntegerType *Ty, uint64_t Offset,
2020  const Twine &Name) {
2021  DEBUG(dbgs() << " start: " << *V << "\n");
2022  IntegerType *IntTy = cast<IntegerType>(V->getType());
2023  assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2024  "Element extends past full value");
2025  uint64_t ShAmt = 8 * Offset;
2026  if (DL.isBigEndian())
2027  ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2028  if (ShAmt) {
2029  V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
2030  DEBUG(dbgs() << " shifted: " << *V << "\n");
2031  }
2032  assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2033  "Cannot extract to a larger integer!");
2034  if (Ty != IntTy) {
2035  V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
2036  DEBUG(dbgs() << " trunced: " << *V << "\n");
2037  }
2038  return V;
2039 }
2040 
2041 static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
2042  Value *V, uint64_t Offset, const Twine &Name) {
2043  IntegerType *IntTy = cast<IntegerType>(Old->getType());
2044  IntegerType *Ty = cast<IntegerType>(V->getType());
2045  assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2046  "Cannot insert a larger integer!");
2047  DEBUG(dbgs() << " start: " << *V << "\n");
2048  if (Ty != IntTy) {
2049  V = IRB.CreateZExt(V, IntTy, Name + ".ext");
2050  DEBUG(dbgs() << " extended: " << *V << "\n");
2051  }
2052  assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2053  "Element store outside of alloca store");
2054  uint64_t ShAmt = 8 * Offset;
2055  if (DL.isBigEndian())
2056  ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2057  if (ShAmt) {
2058  V = IRB.CreateShl(V, ShAmt, Name + ".shift");
2059  DEBUG(dbgs() << " shifted: " << *V << "\n");
2060  }
2061 
2062  if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
2063  APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
2064  Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
2065  DEBUG(dbgs() << " masked: " << *Old << "\n");
2066  V = IRB.CreateOr(Old, V, Name + ".insert");
2067  DEBUG(dbgs() << " inserted: " << *V << "\n");
2068  }
2069  return V;
2070 }
2071 
2072 static Value *extractVector(IRBuilderTy &IRB, Value *V, unsigned BeginIndex,
2073  unsigned EndIndex, const Twine &Name) {
2074  VectorType *VecTy = cast<VectorType>(V->getType());
2075  unsigned NumElements = EndIndex - BeginIndex;
2076  assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2077 
2078  if (NumElements == VecTy->getNumElements())
2079  return V;
2080 
2081  if (NumElements == 1) {
2082  V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
2083  Name + ".extract");
2084  DEBUG(dbgs() << " extract: " << *V << "\n");
2085  return V;
2086  }
2087 
2089  Mask.reserve(NumElements);
2090  for (unsigned i = BeginIndex; i != EndIndex; ++i)
2091  Mask.push_back(IRB.getInt32(i));
2092  V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2093  ConstantVector::get(Mask), Name + ".extract");
2094  DEBUG(dbgs() << " shuffle: " << *V << "\n");
2095  return V;
2096 }
2097 
2098 static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
2099  unsigned BeginIndex, const Twine &Name) {
2100  VectorType *VecTy = cast<VectorType>(Old->getType());
2101  assert(VecTy && "Can only insert a vector into a vector");
2102 
2103  VectorType *Ty = dyn_cast<VectorType>(V->getType());
2104  if (!Ty) {
2105  // Single element to insert.
2106  V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
2107  Name + ".insert");
2108  DEBUG(dbgs() << " insert: " << *V << "\n");
2109  return V;
2110  }
2111 
2112  assert(Ty->getNumElements() <= VecTy->getNumElements() &&
2113  "Too many elements!");
2114  if (Ty->getNumElements() == VecTy->getNumElements()) {
2115  assert(V->getType() == VecTy && "Vector type mismatch");
2116  return V;
2117  }
2118  unsigned EndIndex = BeginIndex + Ty->getNumElements();
2119 
2120  // When inserting a smaller vector into the larger to store, we first
2121  // use a shuffle vector to widen it with undef elements, and then
2122  // a second shuffle vector to select between the loaded vector and the
2123  // incoming vector.
2125  Mask.reserve(VecTy->getNumElements());
2126  for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2127  if (i >= BeginIndex && i < EndIndex)
2128  Mask.push_back(IRB.getInt32(i - BeginIndex));
2129  else
2130  Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
2131  V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2132  ConstantVector::get(Mask), Name + ".expand");
2133  DEBUG(dbgs() << " shuffle: " << *V << "\n");
2134 
2135  Mask.clear();
2136  for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2137  Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
2138 
2139  V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
2140 
2141  DEBUG(dbgs() << " blend: " << *V << "\n");
2142  return V;
2143 }
2144 
2145 /// \brief Visitor to rewrite instructions using p particular slice of an alloca
2146 /// to use a new alloca.
2147 ///
2148 /// Also implements the rewriting to vector-based accesses when the partition
2149 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2150 /// lives here.
2152  : public InstVisitor<AllocaSliceRewriter, bool> {
2153  // Befriend the base class so it can delegate to private visit methods.
2156 
2157  const DataLayout &DL;
2158  AllocaSlices &AS;
2159  SROA &Pass;
2160  AllocaInst &OldAI, &NewAI;
2161  const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2162  Type *NewAllocaTy;
2163 
2164  // This is a convenience and flag variable that will be null unless the new
2165  // alloca's integer operations should be widened to this integer type due to
2166  // passing isIntegerWideningViable above. If it is non-null, the desired
2167  // integer type will be stored here for easy access during rewriting.
2168  IntegerType *IntTy;
2169 
2170  // If we are rewriting an alloca partition which can be written as pure
2171  // vector operations, we stash extra information here. When VecTy is
2172  // non-null, we have some strict guarantees about the rewritten alloca:
2173  // - The new alloca is exactly the size of the vector type here.
2174  // - The accesses all either map to the entire vector or to a single
2175  // element.
2176  // - The set of accessing instructions is only one of those handled above
2177  // in isVectorPromotionViable. Generally these are the same access kinds
2178  // which are promotable via mem2reg.
2179  VectorType *VecTy;
2180  Type *ElementTy;
2181  uint64_t ElementSize;
2182 
2183  // The original offset of the slice currently being rewritten relative to
2184  // the original alloca.
2185  uint64_t BeginOffset, EndOffset;
2186  // The new offsets of the slice currently being rewritten relative to the
2187  // original alloca.
2188  uint64_t NewBeginOffset, NewEndOffset;
2189 
2190  uint64_t SliceSize;
2191  bool IsSplittable;
2192  bool IsSplit;
2193  Use *OldUse;
2194  Instruction *OldPtr;
2195 
2196  // Track post-rewrite users which are PHI nodes and Selects.
2197  SmallSetVector<PHINode *, 8> &PHIUsers;
2198  SmallSetVector<SelectInst *, 8> &SelectUsers;
2199 
2200  // Utility IR builder, whose name prefix is setup for each visited use, and
2201  // the insertion point is set to point to the user.
2202  IRBuilderTy IRB;
2203 
2204 public:
2206  AllocaInst &OldAI, AllocaInst &NewAI,
2207  uint64_t NewAllocaBeginOffset,
2208  uint64_t NewAllocaEndOffset, bool IsIntegerPromotable,
2209  VectorType *PromotableVecTy,
2210  SmallSetVector<PHINode *, 8> &PHIUsers,
2211  SmallSetVector<SelectInst *, 8> &SelectUsers)
2212  : DL(DL), AS(AS), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
2213  NewAllocaBeginOffset(NewAllocaBeginOffset),
2214  NewAllocaEndOffset(NewAllocaEndOffset),
2215  NewAllocaTy(NewAI.getAllocatedType()),
2216  IntTy(IsIntegerPromotable
2217  ? Type::getIntNTy(
2218  NewAI.getContext(),
2219  DL.getTypeSizeInBits(NewAI.getAllocatedType()))
2220  : nullptr),
2221  VecTy(PromotableVecTy),
2222  ElementTy(VecTy ? VecTy->getElementType() : nullptr),
2223  ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
2224  BeginOffset(), EndOffset(), IsSplittable(), IsSplit(), OldUse(),
2225  OldPtr(), PHIUsers(PHIUsers), SelectUsers(SelectUsers),
2226  IRB(NewAI.getContext(), ConstantFolder()) {
2227  if (VecTy) {
2228  assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
2229  "Only multiple-of-8 sized vector elements are viable");
2230  ++NumVectorized;
2231  }
2232  assert((!IntTy && !VecTy) || (IntTy && !VecTy) || (!IntTy && VecTy));
2233  }
2234 
2236  bool CanSROA = true;
2237  BeginOffset = I->beginOffset();
2238  EndOffset = I->endOffset();
2239  IsSplittable = I->isSplittable();
2240  IsSplit =
2241  BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
2242  DEBUG(dbgs() << " rewriting " << (IsSplit ? "split " : ""));
2243  DEBUG(AS.printSlice(dbgs(), I, ""));
2244  DEBUG(dbgs() << "\n");
2245 
2246  // Compute the intersecting offset range.
2247  assert(BeginOffset < NewAllocaEndOffset);
2248  assert(EndOffset > NewAllocaBeginOffset);
2249  NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2250  NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2251 
2252  SliceSize = NewEndOffset - NewBeginOffset;
2253 
2254  OldUse = I->getUse();
2255  OldPtr = cast<Instruction>(OldUse->get());
2256 
2257  Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2258  IRB.SetInsertPoint(OldUserI);
2259  IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2260  IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
2261 
2262  CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2263  if (VecTy || IntTy)
2264  assert(CanSROA);
2265  return CanSROA;
2266  }
2267 
2268 private:
2269  // Make sure the other visit overloads are visible.
2270  using Base::visit;
2271 
2272  // Every instruction which can end up as a user must have a rewrite rule.
2273  bool visitInstruction(Instruction &I) {
2274  DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2275  llvm_unreachable("No rewrite rule for this instruction!");
2276  }
2277 
2278  Value *getNewAllocaSlicePtr(IRBuilderTy &IRB, Type *PointerTy) {
2279  // Note that the offset computation can use BeginOffset or NewBeginOffset
2280  // interchangeably for unsplit slices.
2281  assert(IsSplit || BeginOffset == NewBeginOffset);
2282  uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2283 
2284 #ifndef NDEBUG
2285  StringRef OldName = OldPtr->getName();
2286  // Skip through the last '.sroa.' component of the name.
2287  size_t LastSROAPrefix = OldName.rfind(".sroa.");
2288  if (LastSROAPrefix != StringRef::npos) {
2289  OldName = OldName.substr(LastSROAPrefix + strlen(".sroa."));
2290  // Look for an SROA slice index.
2291  size_t IndexEnd = OldName.find_first_not_of("0123456789");
2292  if (IndexEnd != StringRef::npos && OldName[IndexEnd] == '.') {
2293  // Strip the index and look for the offset.
2294  OldName = OldName.substr(IndexEnd + 1);
2295  size_t OffsetEnd = OldName.find_first_not_of("0123456789");
2296  if (OffsetEnd != StringRef::npos && OldName[OffsetEnd] == '.')
2297  // Strip the offset.
2298  OldName = OldName.substr(OffsetEnd + 1);
2299  }
2300  }
2301  // Strip any SROA suffixes as well.
2302  OldName = OldName.substr(0, OldName.find(".sroa_"));
2303 #endif
2304 
2305  return getAdjustedPtr(IRB, DL, &NewAI,
2306  APInt(DL.getPointerTypeSizeInBits(PointerTy), Offset),
2307  PointerTy,
2308 #ifndef NDEBUG
2309  Twine(OldName) + "."
2310 #else
2311  Twine()
2312 #endif
2313  );
2314  }
2315 
2316  /// \brief Compute suitable alignment to access this slice of the *new*
2317  /// alloca.
2318  ///
2319  /// You can optionally pass a type to this routine and if that type's ABI
2320  /// alignment is itself suitable, this will return zero.
2321  unsigned getSliceAlign(Type *Ty = nullptr) {
2322  unsigned NewAIAlign = NewAI.getAlignment();
2323  if (!NewAIAlign)
2324  NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
2325  unsigned Align =
2326  MinAlign(NewAIAlign, NewBeginOffset - NewAllocaBeginOffset);
2327  return (Ty && Align == DL.getABITypeAlignment(Ty)) ? 0 : Align;
2328  }
2329 
2330  unsigned getIndex(uint64_t Offset) {
2331  assert(VecTy && "Can only call getIndex when rewriting a vector");
2332  uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2333  assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2334  uint32_t Index = RelOffset / ElementSize;
2335  assert(Index * ElementSize == RelOffset);
2336  return Index;
2337  }
2338 
2339  void deleteIfTriviallyDead(Value *V) {
2340  Instruction *I = cast<Instruction>(V);
2342  Pass.DeadInsts.insert(I);
2343  }
2344 
2345  Value *rewriteVectorizedLoadInst() {
2346  unsigned BeginIndex = getIndex(NewBeginOffset);
2347  unsigned EndIndex = getIndex(NewEndOffset);
2348  assert(EndIndex > BeginIndex && "Empty vector!");
2349 
2350  Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2351  return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2352  }
2353 
2354  Value *rewriteIntegerLoad(LoadInst &LI) {
2355  assert(IntTy && "We cannot insert an integer to the alloca");
2356  assert(!LI.isVolatile());
2357  Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2358  V = convertValue(DL, IRB, V, IntTy);
2359  assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2360  uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2361  if (Offset > 0 || NewEndOffset < NewAllocaEndOffset) {
2362  IntegerType *ExtractTy = Type::getIntNTy(LI.getContext(), SliceSize * 8);
2363  V = extractInteger(DL, IRB, V, ExtractTy, Offset, "extract");
2364  }
2365  // It is possible that the extracted type is not the load type. This
2366  // happens if there is a load past the end of the alloca, and as
2367  // a consequence the slice is narrower but still a candidate for integer
2368  // lowering. To handle this case, we just zero extend the extracted
2369  // integer.
2370  assert(cast<IntegerType>(LI.getType())->getBitWidth() >= SliceSize * 8 &&
2371  "Can only handle an extract for an overly wide load");
2372  if (cast<IntegerType>(LI.getType())->getBitWidth() > SliceSize * 8)
2373  V = IRB.CreateZExt(V, LI.getType());
2374  return V;
2375  }
2376 
2377  bool visitLoadInst(LoadInst &LI) {
2378  DEBUG(dbgs() << " original: " << LI << "\n");
2379  Value *OldOp = LI.getOperand(0);
2380  assert(OldOp == OldPtr);
2381 
2382  unsigned AS = LI.getPointerAddressSpace();
2383 
2384  Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), SliceSize * 8)
2385  : LI.getType();
2386  const bool IsLoadPastEnd = DL.getTypeStoreSize(TargetTy) > SliceSize;
2387  bool IsPtrAdjusted = false;
2388  Value *V;
2389  if (VecTy) {
2390  V = rewriteVectorizedLoadInst();
2391  } else if (IntTy && LI.getType()->isIntegerTy()) {
2392  V = rewriteIntegerLoad(LI);
2393  } else if (NewBeginOffset == NewAllocaBeginOffset &&
2394  NewEndOffset == NewAllocaEndOffset &&
2395  (canConvertValue(DL, NewAllocaTy, TargetTy) ||
2396  (IsLoadPastEnd && NewAllocaTy->isIntegerTy() &&
2397  TargetTy->isIntegerTy()))) {
2398  LoadInst *NewLI = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2399  LI.isVolatile(), LI.getName());
2400  if (LI.isVolatile())
2401  NewLI->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
2402 
2403  // Any !nonnull metadata or !range metadata on the old load is also valid
2404  // on the new load. This is even true in some cases even when the loads
2405  // are different types, for example by mapping !nonnull metadata to
2406  // !range metadata by modeling the null pointer constant converted to the
2407  // integer type.
2408  // FIXME: Add support for range metadata here. Currently the utilities
2409  // for this don't propagate range metadata in trivial cases from one
2410  // integer load to another, don't handle non-addrspace-0 null pointers
2411  // correctly, and don't have any support for mapping ranges as the
2412  // integer type becomes winder or narrower.
2414  copyNonnullMetadata(LI, N, *NewLI);
2415 
2416  // Try to preserve nonnull metadata
2417  V = NewLI;
2418 
2419  // If this is an integer load past the end of the slice (which means the
2420  // bytes outside the slice are undef or this load is dead) just forcibly
2421  // fix the integer size with correct handling of endianness.
2422  if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2423  if (auto *TITy = dyn_cast<IntegerType>(TargetTy))
2424  if (AITy->getBitWidth() < TITy->getBitWidth()) {
2425  V = IRB.CreateZExt(V, TITy, "load.ext");
2426  if (DL.isBigEndian())
2427  V = IRB.CreateShl(V, TITy->getBitWidth() - AITy->getBitWidth(),
2428  "endian_shift");
2429  }
2430  } else {
2431  Type *LTy = TargetTy->getPointerTo(AS);
2432  LoadInst *NewLI = IRB.CreateAlignedLoad(getNewAllocaSlicePtr(IRB, LTy),
2433  getSliceAlign(TargetTy),
2434  LI.isVolatile(), LI.getName());
2435  if (LI.isVolatile())
2436  NewLI->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
2437 
2438  V = NewLI;
2439  IsPtrAdjusted = true;
2440  }
2441  V = convertValue(DL, IRB, V, TargetTy);
2442 
2443  if (IsSplit) {
2444  assert(!LI.isVolatile());
2445  assert(LI.getType()->isIntegerTy() &&
2446  "Only integer type loads and stores are split");
2447  assert(SliceSize < DL.getTypeStoreSize(LI.getType()) &&
2448  "Split load isn't smaller than original load");
2449  assert(LI.getType()->getIntegerBitWidth() ==
2450  DL.getTypeStoreSizeInBits(LI.getType()) &&
2451  "Non-byte-multiple bit width");
2452  // Move the insertion point just past the load so that we can refer to it.
2453  IRB.SetInsertPoint(&*std::next(BasicBlock::iterator(&LI)));
2454  // Create a placeholder value with the same type as LI to use as the
2455  // basis for the new value. This allows us to replace the uses of LI with
2456  // the computed value, and then replace the placeholder with LI, leaving
2457  // LI only used for this computation.
2458  Value *Placeholder =
2459  new LoadInst(UndefValue::get(LI.getType()->getPointerTo(AS)));
2460  V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset - BeginOffset,
2461  "insert");
2462  LI.replaceAllUsesWith(V);
2463  Placeholder->replaceAllUsesWith(&LI);
2464  Placeholder->deleteValue();
2465  } else {
2466  LI.replaceAllUsesWith(V);
2467  }
2468 
2469  Pass.DeadInsts.insert(&LI);
2470  deleteIfTriviallyDead(OldOp);
2471  DEBUG(dbgs() << " to: " << *V << "\n");
2472  return !LI.isVolatile() && !IsPtrAdjusted;
2473  }
2474 
2475  bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp) {
2476  if (V->getType() != VecTy) {
2477  unsigned BeginIndex = getIndex(NewBeginOffset);
2478  unsigned EndIndex = getIndex(NewEndOffset);
2479  assert(EndIndex > BeginIndex && "Empty vector!");
2480  unsigned NumElements = EndIndex - BeginIndex;
2481  assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2482  Type *SliceTy = (NumElements == 1)
2483  ? ElementTy
2484  : VectorType::get(ElementTy, NumElements);
2485  if (V->getType() != SliceTy)
2486  V = convertValue(DL, IRB, V, SliceTy);
2487 
2488  // Mix in the existing elements.
2489  Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2490  V = insertVector(IRB, Old, V, BeginIndex, "vec");
2491  }
2492  StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2493  Pass.DeadInsts.insert(&SI);
2494 
2495  (void)Store;
2496  DEBUG(dbgs() << " to: " << *Store << "\n");
2497  return true;
2498  }
2499 
2500  bool rewriteIntegerStore(Value *V, StoreInst &SI) {
2501  assert(IntTy && "We cannot extract an integer from the alloca");
2502  assert(!SI.isVolatile());
2503  if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2504  Value *Old =
2505  IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2506  Old = convertValue(DL, IRB, Old, IntTy);
2507  assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2508  uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2509  V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset, "insert");
2510  }
2511  V = convertValue(DL, IRB, V, NewAllocaTy);
2512  StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2514  Pass.DeadInsts.insert(&SI);
2515  DEBUG(dbgs() << " to: " << *Store << "\n");
2516  return true;
2517  }
2518 
2519  bool visitStoreInst(StoreInst &SI) {
2520  DEBUG(dbgs() << " original: " << SI << "\n");
2521  Value *OldOp = SI.getOperand(1);
2522  assert(OldOp == OldPtr);
2523 
2524  Value *V = SI.getValueOperand();
2525 
2526  // Strip all inbounds GEPs and pointer casts to try to dig out any root
2527  // alloca that should be re-examined after promoting this alloca.
2528  if (V->getType()->isPointerTy())
2529  if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2530  Pass.PostPromotionWorklist.insert(AI);
2531 
2532  if (SliceSize < DL.getTypeStoreSize(V->getType())) {
2533  assert(!SI.isVolatile());
2534  assert(V->getType()->isIntegerTy() &&
2535  "Only integer type loads and stores are split");
2536  assert(V->getType()->getIntegerBitWidth() ==
2537  DL.getTypeStoreSizeInBits(V->getType()) &&
2538  "Non-byte-multiple bit width");
2539  IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), SliceSize * 8);
2540  V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset - BeginOffset,
2541  "extract");
2542  }
2543 
2544  if (VecTy)
2545  return rewriteVectorizedStoreInst(V, SI, OldOp);
2546  if (IntTy && V->getType()->isIntegerTy())
2547  return rewriteIntegerStore(V, SI);
2548 
2549  const bool IsStorePastEnd = DL.getTypeStoreSize(V->getType()) > SliceSize;
2550  StoreInst *NewSI;
2551  if (NewBeginOffset == NewAllocaBeginOffset &&
2552  NewEndOffset == NewAllocaEndOffset &&
2553  (canConvertValue(DL, V->getType(), NewAllocaTy) ||
2554  (IsStorePastEnd && NewAllocaTy->isIntegerTy() &&
2555  V->getType()->isIntegerTy()))) {
2556  // If this is an integer store past the end of slice (and thus the bytes
2557  // past that point are irrelevant or this is unreachable), truncate the
2558  // value prior to storing.
2559  if (auto *VITy = dyn_cast<IntegerType>(V->getType()))
2560  if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2561  if (VITy->getBitWidth() > AITy->getBitWidth()) {
2562  if (DL.isBigEndian())
2563  V = IRB.CreateLShr(V, VITy->getBitWidth() - AITy->getBitWidth(),
2564  "endian_shift");
2565  V = IRB.CreateTrunc(V, AITy, "load.trunc");
2566  }
2567 
2568  V = convertValue(DL, IRB, V, NewAllocaTy);
2569  NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2570  SI.isVolatile());
2571  } else {
2572  unsigned AS = SI.getPointerAddressSpace();
2573  Value *NewPtr = getNewAllocaSlicePtr(IRB, V->getType()->getPointerTo(AS));
2574  NewSI = IRB.CreateAlignedStore(V, NewPtr, getSliceAlign(V->getType()),
2575  SI.isVolatile());
2576  }
2578  if (SI.isVolatile())
2579  NewSI->setAtomic(SI.getOrdering(), SI.getSyncScopeID());
2580  Pass.DeadInsts.insert(&SI);
2581  deleteIfTriviallyDead(OldOp);
2582 
2583  DEBUG(dbgs() << " to: " << *NewSI << "\n");
2584  return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2585  }
2586 
2587  /// \brief Compute an integer value from splatting an i8 across the given
2588  /// number of bytes.
2589  ///
2590  /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2591  /// call this routine.
2592  /// FIXME: Heed the advice above.
2593  ///
2594  /// \param V The i8 value to splat.
2595  /// \param Size The number of bytes in the output (assuming i8 is one byte)
2596  Value *getIntegerSplat(Value *V, unsigned Size) {
2597  assert(Size > 0 && "Expected a positive number of bytes.");
2598  IntegerType *VTy = cast<IntegerType>(V->getType());
2599  assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2600  if (Size == 1)
2601  return V;
2602 
2603  Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size * 8);
2604  V = IRB.CreateMul(
2605  IRB.CreateZExt(V, SplatIntTy, "zext"),
2607  Constant::getAllOnesValue(SplatIntTy),
2609  SplatIntTy)),
2610  "isplat");
2611  return V;
2612  }
2613 
2614  /// \brief Compute a vector splat for a given element value.
2615  Value *getVectorSplat(Value *V, unsigned NumElements) {
2616  V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2617  DEBUG(dbgs() << " splat: " << *V << "\n");
2618  return V;
2619  }
2620 
2621  bool visitMemSetInst(MemSetInst &II) {
2622  DEBUG(dbgs() << " original: " << II << "\n");
2623  assert(II.getRawDest() == OldPtr);
2624 
2625  // If the memset has a variable size, it cannot be split, just adjust the
2626  // pointer to the new alloca.
2627  if (!isa<Constant>(II.getLength())) {
2628  assert(!IsSplit);
2629  assert(NewBeginOffset == BeginOffset);
2630  II.setDest(getNewAllocaSlicePtr(IRB, OldPtr->getType()));
2631  Type *CstTy = II.getAlignmentCst()->getType();
2632  II.setAlignment(ConstantInt::get(CstTy, getSliceAlign()));
2633 
2634  deleteIfTriviallyDead(OldPtr);
2635  return false;
2636  }
2637 
2638  // Record this instruction for deletion.
2639  Pass.DeadInsts.insert(&II);
2640 
2641  Type *AllocaTy = NewAI.getAllocatedType();
2642  Type *ScalarTy = AllocaTy->getScalarType();
2643 
2644  // If this doesn't map cleanly onto the alloca type, and that type isn't
2645  // a single value type, just emit a memset.
2646  if (!VecTy && !IntTy &&
2647  (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2648  SliceSize != DL.getTypeStoreSize(AllocaTy) ||
2649  !AllocaTy->isSingleValueType() ||
2650  !DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy)) ||
2651  DL.getTypeSizeInBits(ScalarTy) % 8 != 0)) {
2652  Type *SizeTy = II.getLength()->getType();
2653  Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2654  CallInst *New = IRB.CreateMemSet(
2655  getNewAllocaSlicePtr(IRB, OldPtr->getType()), II.getValue(), Size,
2656  getSliceAlign(), II.isVolatile());
2657  (void)New;
2658  DEBUG(dbgs() << " to: " << *New << "\n");
2659  return false;
2660  }
2661 
2662  // If we can represent this as a simple value, we have to build the actual
2663  // value to store, which requires expanding the byte present in memset to
2664  // a sensible representation for the alloca type. This is essentially
2665  // splatting the byte to a sufficiently wide integer, splatting it across
2666  // any desired vector width, and bitcasting to the final type.
2667  Value *V;
2668 
2669  if (VecTy) {
2670  // If this is a memset of a vectorized alloca, insert it.
2671  assert(ElementTy == ScalarTy);
2672 
2673  unsigned BeginIndex = getIndex(NewBeginOffset);
2674  unsigned EndIndex = getIndex(NewEndOffset);
2675  assert(EndIndex > BeginIndex && "Empty vector!");
2676  unsigned NumElements = EndIndex - BeginIndex;
2677  assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2678 
2679  Value *Splat =
2680  getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
2681  Splat = convertValue(DL, IRB, Splat, ElementTy);
2682  if (NumElements > 1)
2683  Splat = getVectorSplat(Splat, NumElements);
2684 
2685  Value *Old =
2686  IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2687  V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2688  } else if (IntTy) {
2689  // If this is a memset on an alloca where we can widen stores, insert the
2690  // set integer.
2691  assert(!II.isVolatile());
2692 
2693  uint64_t Size = NewEndOffset - NewBeginOffset;
2694  V = getIntegerSplat(II.getValue(), Size);
2695 
2696  if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2697  EndOffset != NewAllocaBeginOffset)) {
2698  Value *Old =
2699  IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2700  Old = convertValue(DL, IRB, Old, IntTy);
2701  uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2702  V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2703  } else {
2704  assert(V->getType() == IntTy &&
2705  "Wrong type for an alloca wide integer!");
2706  }
2707  V = convertValue(DL, IRB, V, AllocaTy);
2708  } else {
2709  // Established these invariants above.
2710  assert(NewBeginOffset == NewAllocaBeginOffset);
2711  assert(NewEndOffset == NewAllocaEndOffset);
2712 
2713  V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
2714  if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2715  V = getVectorSplat(V, AllocaVecTy->getNumElements());
2716 
2717  V = convertValue(DL, IRB, V, AllocaTy);
2718  }
2719 
2720  Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2721  II.isVolatile());
2722  (void)New;
2723  DEBUG(dbgs() << " to: " << *New << "\n");
2724  return !II.isVolatile();
2725  }
2726 
2728  // Rewriting of memory transfer instructions can be a bit tricky. We break
2729  // them into two categories: split intrinsics and unsplit intrinsics.
2730 
2731  DEBUG(dbgs() << " original: " << II << "\n");
2732 
2733  bool IsDest = &II.getRawDestUse() == OldUse;
2734  assert((IsDest && II.getRawDest() == OldPtr) ||
2735  (!IsDest && II.getRawSource() == OldPtr));
2736 
2737  unsigned SliceAlign = getSliceAlign();
2738 
2739  // For unsplit intrinsics, we simply modify the source and destination
2740  // pointers in place. This isn't just an optimization, it is a matter of
2741  // correctness. With unsplit intrinsics we may be dealing with transfers
2742  // within a single alloca before SROA ran, or with transfers that have
2743  // a variable length. We may also be dealing with memmove instead of
2744  // memcpy, and so simply updating the pointers is the necessary for us to
2745  // update both source and dest of a single call.
2746  if (!IsSplittable) {
2747  Value *AdjustedPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2748  if (IsDest)
2749  II.setDest(AdjustedPtr);
2750  else
2751  II.setSource(AdjustedPtr);
2752 
2753  if (II.getAlignment() > SliceAlign) {
2754  Type *CstTy = II.getAlignmentCst()->getType();
2755  II.setAlignment(
2756  ConstantInt::get(CstTy, MinAlign(II.getAlignment(), SliceAlign)));
2757  }
2758 
2759  DEBUG(dbgs() << " to: " << II << "\n");
2760  deleteIfTriviallyDead(OldPtr);
2761  return false;
2762  }
2763  // For split transfer intrinsics we have an incredibly useful assurance:
2764  // the source and destination do not reside within the same alloca, and at
2765  // least one of them does not escape. This means that we can replace
2766  // memmove with memcpy, and we don't need to worry about all manner of
2767  // downsides to splitting and transforming the operations.
2768 
2769  // If this doesn't map cleanly onto the alloca type, and that type isn't
2770  // a single value type, just emit a memcpy.
2771  bool EmitMemCpy =
2772  !VecTy && !IntTy &&
2773  (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2774  SliceSize != DL.getTypeStoreSize(NewAI.getAllocatedType()) ||
2775  !NewAI.getAllocatedType()->isSingleValueType());
2776 
2777  // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2778  // size hasn't been shrunk based on analysis of the viable range, this is
2779  // a no-op.
2780  if (EmitMemCpy && &OldAI == &NewAI) {
2781  // Ensure the start lines up.
2782  assert(NewBeginOffset == BeginOffset);
2783 
2784  // Rewrite the size as needed.
2785  if (NewEndOffset != EndOffset)
2787  NewEndOffset - NewBeginOffset));
2788  return false;
2789  }
2790  // Record this instruction for deletion.
2791  Pass.DeadInsts.insert(&II);
2792 
2793  // Strip all inbounds GEPs and pointer casts to try to dig out any root
2794  // alloca that should be re-examined after rewriting this instruction.
2795  Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2796  if (AllocaInst *AI =
2797  dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
2798  assert(AI != &OldAI && AI != &NewAI &&
2799  "Splittable transfers cannot reach the same alloca on both ends.");
2800  Pass.Worklist.insert(AI);
2801  }
2802 
2803  Type *OtherPtrTy = OtherPtr->getType();
2804  unsigned OtherAS = OtherPtrTy->getPointerAddressSpace();
2805 
2806  // Compute the relative offset for the other pointer within the transfer.
2807  unsigned IntPtrWidth = DL.getPointerSizeInBits(OtherAS);
2808  APInt OtherOffset(IntPtrWidth, NewBeginOffset - BeginOffset);
2809  unsigned OtherAlign = MinAlign(II.getAlignment() ? II.getAlignment() : 1,
2810  OtherOffset.zextOrTrunc(64).getZExtValue());
2811 
2812  if (EmitMemCpy) {
2813  // Compute the other pointer, folding as much as possible to produce
2814  // a single, simple GEP in most cases.
2815  OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2816  OtherPtr->getName() + ".");
2817 
2818  Value *OurPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2819  Type *SizeTy = II.getLength()->getType();
2820  Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2821 
2822  CallInst *New = IRB.CreateMemCpy(
2823  IsDest ? OurPtr : OtherPtr, IsDest ? OtherPtr : OurPtr, Size,
2824  MinAlign(SliceAlign, OtherAlign), II.isVolatile());
2825  (void)New;
2826  DEBUG(dbgs() << " to: " << *New << "\n");
2827  return false;
2828  }
2829 
2830  bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
2831  NewEndOffset == NewAllocaEndOffset;
2832  uint64_t Size = NewEndOffset - NewBeginOffset;
2833  unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
2834  unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
2835  unsigned NumElements = EndIndex - BeginIndex;
2836  IntegerType *SubIntTy =
2837  IntTy ? Type::getIntNTy(IntTy->getContext(), Size * 8) : nullptr;
2838 
2839  // Reset the other pointer type to match the register type we're going to
2840  // use, but using the address space of the original other pointer.
2841  if (VecTy && !IsWholeAlloca) {
2842  if (NumElements == 1)
2843  OtherPtrTy = VecTy->getElementType();
2844  else
2845  OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2846 
2847  OtherPtrTy = OtherPtrTy->getPointerTo(OtherAS);
2848  } else if (IntTy && !IsWholeAlloca) {
2849  OtherPtrTy = SubIntTy->getPointerTo(OtherAS);
2850  } else {
2851  OtherPtrTy = NewAllocaTy->getPointerTo(OtherAS);
2852  }
2853 
2854  Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2855  OtherPtr->getName() + ".");
2856  unsigned SrcAlign = OtherAlign;
2857  Value *DstPtr = &NewAI;
2858  unsigned DstAlign = SliceAlign;
2859  if (!IsDest) {
2860  std::swap(SrcPtr, DstPtr);
2861  std::swap(SrcAlign, DstAlign);
2862  }
2863 
2864  Value *Src;
2865  if (VecTy && !IsWholeAlloca && !IsDest) {
2866  Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2867  Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
2868  } else if (IntTy && !IsWholeAlloca && !IsDest) {
2869  Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2870  Src = convertValue(DL, IRB, Src, IntTy);
2871  uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2872  Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
2873  } else {
2874  Src =
2875  IRB.CreateAlignedLoad(SrcPtr, SrcAlign, II.isVolatile(), "copyload");
2876  }
2877 
2878  if (VecTy && !IsWholeAlloca && IsDest) {
2879  Value *Old =
2880  IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2881  Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
2882  } else if (IntTy && !IsWholeAlloca && IsDest) {
2883  Value *Old =
2884  IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2885  Old = convertValue(DL, IRB, Old, IntTy);
2886  uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2887  Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
2888  Src = convertValue(DL, IRB, Src, NewAllocaTy);
2889  }
2890 
2891  StoreInst *Store = cast<StoreInst>(
2892  IRB.CreateAlignedStore(Src, DstPtr, DstAlign, II.isVolatile()));
2893  (void)Store;
2894  DEBUG(dbgs() << " to: " << *Store << "\n");
2895  return !II.isVolatile();
2896  }
2897 
2898  bool visitIntrinsicInst(IntrinsicInst &II) {
2899  assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2900  II.getIntrinsicID() == Intrinsic::lifetime_end);
2901  DEBUG(dbgs() << " original: " << II << "\n");
2902  assert(II.getArgOperand(1) == OldPtr);
2903 
2904  // Record this instruction for deletion.
2905  Pass.DeadInsts.insert(&II);
2906 
2907  // Lifetime intrinsics are only promotable if they cover the whole alloca.
2908  // Therefore, we drop lifetime intrinsics which don't cover the whole
2909  // alloca.
2910  // (In theory, intrinsics which partially cover an alloca could be
2911  // promoted, but PromoteMemToReg doesn't handle that case.)
2912  // FIXME: Check whether the alloca is promotable before dropping the
2913  // lifetime intrinsics?
2914  if (NewBeginOffset != NewAllocaBeginOffset ||
2915  NewEndOffset != NewAllocaEndOffset)
2916  return true;
2917 
2918  ConstantInt *Size =
2919  ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2920  NewEndOffset - NewBeginOffset);
2921  Value *Ptr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2922  Value *New;
2923  if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2924  New = IRB.CreateLifetimeStart(Ptr, Size);
2925  else
2926  New = IRB.CreateLifetimeEnd(Ptr, Size);
2927 
2928  (void)New;
2929  DEBUG(dbgs() << " to: " << *New << "\n");
2930 
2931  return true;
2932  }
2933 
2934  bool visitPHINode(PHINode &PN) {
2935  DEBUG(dbgs() << " original: " << PN << "\n");
2936  assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
2937  assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
2938 
2939  // We would like to compute a new pointer in only one place, but have it be
2940  // as local as possible to the PHI. To do that, we re-use the location of
2941  // the old pointer, which necessarily must be in the right position to
2942  // dominate the PHI.
2943  IRBuilderTy PtrBuilder(IRB);
2944  if (isa<PHINode>(OldPtr))
2945  PtrBuilder.SetInsertPoint(&*OldPtr->getParent()->getFirstInsertionPt());
2946  else
2947  PtrBuilder.SetInsertPoint(OldPtr);
2948  PtrBuilder.SetCurrentDebugLocation(OldPtr->getDebugLoc());
2949 
2950  Value *NewPtr = getNewAllocaSlicePtr(PtrBuilder, OldPtr->getType());
2951  // Replace the operands which were using the old pointer.
2952  std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
2953 
2954  DEBUG(dbgs() << " to: " << PN << "\n");
2955  deleteIfTriviallyDead(OldPtr);
2956 
2957  // PHIs can't be promoted on their own, but often can be speculated. We
2958  // check the speculation outside of the rewriter so that we see the
2959  // fully-rewritten alloca.
2960  PHIUsers.insert(&PN);
2961  return true;
2962  }
2963 
2964  bool visitSelectInst(SelectInst &SI) {
2965  DEBUG(dbgs() << " original: " << SI << "\n");
2966  assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
2967  "Pointer isn't an operand!");
2968  assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
2969  assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
2970 
2971  Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2972  // Replace the operands which were using the old pointer.
2973  if (SI.getOperand(1) == OldPtr)
2974  SI.setOperand(1, NewPtr);
2975  if (SI.getOperand(2) == OldPtr)
2976  SI.setOperand(2, NewPtr);
2977 
2978  DEBUG(dbgs() << " to: " << SI << "\n");
2979  deleteIfTriviallyDead(OldPtr);
2980 
2981  // Selects can't be promoted on their own, but often can be speculated. We
2982  // check the speculation outside of the rewriter so that we see the
2983  // fully-rewritten alloca.
2984  SelectUsers.insert(&SI);
2985  return true;
2986  }
2987 };
2988 
2989 namespace {
2990 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
2991 ///
2992 /// This pass aggressively rewrites all aggregate loads and stores on
2993 /// a particular pointer (or any pointer derived from it which we can identify)
2994 /// with scalar loads and stores.
2995 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
2996  // Befriend the base class so it can delegate to private visit methods.
2997  friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
2998 
2999  /// Queue of pointer uses to analyze and potentially rewrite.
3000  SmallVector<Use *, 8> Queue;
3001 
3002  /// Set to prevent us from cycling with phi nodes and loops.
3003  SmallPtrSet<User *, 8> Visited;
3004 
3005  /// The current pointer use being rewritten. This is used to dig up the used
3006  /// value (as opposed to the user).
3007  Use *U;
3008 
3009 public:
3010  /// Rewrite loads and stores through a pointer and all pointers derived from
3011  /// it.
3012  bool rewrite(Instruction &I) {
3013  DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
3014  enqueueUsers(I);
3015  bool Changed = false;
3016  while (!Queue.empty()) {
3017  U = Queue.pop_back_val();
3018  Changed |= visit(cast<Instruction>(U->getUser()));
3019  }
3020  return Changed;
3021  }
3022 
3023 private:
3024  /// Enqueue all the users of the given instruction for further processing.
3025  /// This uses a set to de-duplicate users.
3026  void enqueueUsers(Instruction &I) {
3027  for (Use &U : I.uses())
3028  if (Visited.insert(U.getUser()).second)
3029  Queue.push_back(&U);
3030  }
3031 
3032  // Conservative default is to not rewrite anything.
3033  bool visitInstruction(Instruction &I) { return false; }
3034 
3035  /// \brief Generic recursive split emission class.
3036  template <typename Derived> class OpSplitter {
3037  protected:
3038  /// The builder used to form new instructions.
3039  IRBuilderTy IRB;
3040  /// The indices which to be used with insert- or extractvalue to select the
3041  /// appropriate value within the aggregate.
3042  SmallVector<unsigned, 4> Indices;
3043  /// The indices to a GEP instruction which will move Ptr to the correct slot
3044  /// within the aggregate.
3045  SmallVector<Value *, 4> GEPIndices;
3046  /// The base pointer of the original op, used as a base for GEPing the
3047  /// split operations.
3048  Value *Ptr;
3049 
3050  /// Initialize the splitter with an insertion point, Ptr and start with a
3051  /// single zero GEP index.
3052  OpSplitter(Instruction *InsertionPoint, Value *Ptr)
3053  : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
3054 
3055  public:
3056  /// \brief Generic recursive split emission routine.
3057  ///
3058  /// This method recursively splits an aggregate op (load or store) into
3059  /// scalar or vector ops. It splits recursively until it hits a single value
3060  /// and emits that single value operation via the template argument.
3061  ///
3062  /// The logic of this routine relies on GEPs and insertvalue and
3063  /// extractvalue all operating with the same fundamental index list, merely
3064  /// formatted differently (GEPs need actual values).
3065  ///
3066  /// \param Ty The type being split recursively into smaller ops.
3067  /// \param Agg The aggregate value being built up or stored, depending on
3068  /// whether this is splitting a load or a store respectively.
3069  void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
3070  if (Ty->isSingleValueType())
3071  return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
3072 
3073  if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
3074  unsigned OldSize = Indices.size();
3075  (void)OldSize;
3076  for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
3077  ++Idx) {
3078  assert(Indices.size() == OldSize && "Did not return to the old size");
3079  Indices.push_back(Idx);
3080  GEPIndices.push_back(IRB.getInt32(Idx));
3081  emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
3082  GEPIndices.pop_back();
3083  Indices.pop_back();
3084  }
3085  return;
3086  }
3087 
3088  if (StructType *STy = dyn_cast<StructType>(Ty)) {
3089  unsigned OldSize = Indices.size();
3090  (void)OldSize;
3091  for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
3092  ++Idx) {
3093  assert(Indices.size() == OldSize && "Did not return to the old size");
3094  Indices.push_back(Idx);
3095  GEPIndices.push_back(IRB.getInt32(Idx));
3096  emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
3097  GEPIndices.pop_back();
3098  Indices.pop_back();
3099  }
3100  return;
3101  }
3102 
3103  llvm_unreachable("Only arrays and structs are aggregate loadable types");
3104  }
3105  };
3106 
3107  struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
3108  LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3109  : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
3110 
3111  /// Emit a leaf load of a single value. This is called at the leaves of the
3112  /// recursive emission to actually load values.
3113  void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3114  assert(Ty->isSingleValueType());
3115  // Load the single value and insert it using the indices.
3116  Value *GEP =
3117  IRB.CreateInBoundsGEP(nullptr, Ptr, GEPIndices, Name + ".gep");
3118  Value *Load = IRB.CreateLoad(GEP, Name + ".load");
3119  Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
3120  DEBUG(dbgs() << " to: " << *Load << "\n");
3121  }
3122  };
3123 
3124  bool visitLoadInst(LoadInst &LI) {
3125  assert(LI.getPointerOperand() == *U);
3126  if (!LI.isSimple() || LI.getType()->isSingleValueType())
3127  return false;
3128 
3129  // We have an aggregate being loaded, split it apart.
3130  DEBUG(dbgs() << " original: " << LI << "\n");
3131  LoadOpSplitter Splitter(&LI, *U);
3132  Value *V = UndefValue::get(LI.getType());
3133  Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
3134  LI.replaceAllUsesWith(V);
3135  LI.eraseFromParent();
3136  return true;
3137  }
3138 
3139  struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
3140  StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3141  : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
3142 
3143  /// Emit a leaf store of a single value. This is called at the leaves of the
3144  /// recursive emission to actually produce stores.
3145  void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3146  assert(Ty->isSingleValueType());
3147  // Extract the single value and store it using the indices.
3148  //
3149  // The gep and extractvalue values are factored out of the CreateStore
3150  // call to make the output independent of the argument evaluation order.
3151  Value *ExtractValue =
3152  IRB.CreateExtractValue(Agg, Indices, Name + ".extract");
3153  Value *InBoundsGEP =
3154  IRB.CreateInBoundsGEP(nullptr, Ptr, GEPIndices, Name + ".gep");
3155  Value *Store = IRB.CreateStore(ExtractValue, InBoundsGEP);
3156  (void)Store;
3157  DEBUG(dbgs() << " to: " << *Store << "\n");
3158  }
3159  };
3160 
3161  bool visitStoreInst(StoreInst &SI) {
3162  if (!SI.isSimple() || SI.getPointerOperand() != *U)
3163  return false;
3164  Value *V = SI.getValueOperand();
3165  if (V->getType()->isSingleValueType())
3166  return false;
3167 
3168  // We have an aggregate being stored, split it apart.
3169  DEBUG(dbgs() << " original: " << SI << "\n");
3170  StoreOpSplitter Splitter(&SI, *U);
3171  Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
3172  SI.eraseFromParent();
3173  return true;
3174  }
3175 
3176  bool visitBitCastInst(BitCastInst &BC) {
3177  enqueueUsers(BC);
3178  return false;
3179  }
3180 
3181  bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
3182  enqueueUsers(GEPI);
3183  return false;
3184  }
3185 
3186  bool visitPHINode(PHINode &PN) {
3187  enqueueUsers(PN);
3188  return false;
3189  }
3190 
3191  bool visitSelectInst(SelectInst &SI) {
3192  enqueueUsers(SI);
3193  return false;
3194  }
3195 };
3196 }
3197 
3198 /// \brief Strip aggregate type wrapping.
3199 ///
3200 /// This removes no-op aggregate types wrapping an underlying type. It will
3201 /// strip as many layers of types as it can without changing either the type
3202 /// size or the allocated size.
3204  if (Ty->isSingleValueType())
3205  return Ty;
3206 
3207  uint64_t AllocSize = DL.getTypeAllocSize(Ty);
3208  uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
3209 
3210  Type *InnerTy;
3211  if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
3212  InnerTy = ArrTy->getElementType();
3213  } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
3214  const StructLayout *SL = DL.getStructLayout(STy);
3215  unsigned Index = SL->getElementContainingOffset(0);
3216  InnerTy = STy->getElementType(Index);
3217  } else {
3218  return Ty;
3219  }
3220 
3221  if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
3222  TypeSize > DL.getTypeSizeInBits(InnerTy))
3223  return Ty;
3224 
3225  return stripAggregateTypeWrapping(DL, InnerTy);
3226 }
3227 
3228 /// \brief Try to find a partition of the aggregate type passed in for a given
3229 /// offset and size.
3230 ///
3231 /// This recurses through the aggregate type and tries to compute a subtype
3232 /// based on the offset and size. When the offset and size span a sub-section
3233 /// of an array, it will even compute a new array type for that sub-section,
3234 /// and the same for structs.
3235 ///
3236 /// Note that this routine is very strict and tries to find a partition of the
3237 /// type which produces the *exact* right offset and size. It is not forgiving
3238 /// when the size or offset cause either end of type-based partition to be off.
3239 /// Also, this is a best-effort routine. It is reasonable to give up and not
3240 /// return a type if necessary.
3241 static Type *getTypePartition(const DataLayout &DL, Type *Ty, uint64_t Offset,
3242  uint64_t Size) {
3243  if (Offset == 0 && DL.getTypeAllocSize(Ty) == Size)
3244  return stripAggregateTypeWrapping(DL, Ty);
3245  if (Offset > DL.getTypeAllocSize(Ty) ||
3246  (DL.getTypeAllocSize(Ty) - Offset) < Size)
3247  return nullptr;
3248 
3249  if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3250  Type *ElementTy = SeqTy->getElementType();
3251  uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3252  uint64_t NumSkippedElements = Offset / ElementSize;
3253  if (NumSkippedElements >= SeqTy->getNumElements())
3254  return nullptr;
3255  Offset -= NumSkippedElements * ElementSize;
3256 
3257  // First check if we need to recurse.
3258  if (Offset > 0 || Size < ElementSize) {
3259  // Bail if the partition ends in a different array element.
3260  if ((Offset + Size) > ElementSize)
3261  return nullptr;
3262  // Recurse through the element type trying to peel off offset bytes.
3263  return getTypePartition(DL, ElementTy, Offset, Size);
3264  }
3265  assert(Offset == 0);
3266 
3267  if (Size == ElementSize)
3268  return stripAggregateTypeWrapping(DL, ElementTy);
3269  assert(Size > ElementSize);
3270  uint64_t NumElements = Size / ElementSize;
3271  if (NumElements * ElementSize != Size)
3272  return nullptr;
3273  return ArrayType::get(ElementTy, NumElements);
3274  }
3275 
3276  StructType *STy = dyn_cast<StructType>(Ty);
3277  if (!STy)
3278  return nullptr;
3279 
3280  const StructLayout *SL = DL.getStructLayout(STy);
3281  if (Offset >= SL->getSizeInBytes())
3282  return nullptr;
3283  uint64_t EndOffset = Offset + Size;
3284  if (EndOffset > SL->getSizeInBytes())
3285  return nullptr;
3286 
3287  unsigned Index = SL->getElementContainingOffset(Offset);
3288  Offset -= SL->getElementOffset(Index);
3289 
3290  Type *ElementTy = STy->getElementType(Index);
3291  uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3292  if (Offset >= ElementSize)
3293  return nullptr; // The offset points into alignment padding.
3294 
3295  // See if any partition must be contained by the element.
3296  if (Offset > 0 || Size < ElementSize) {
3297  if ((Offset + Size) > ElementSize)
3298  return nullptr;
3299  return getTypePartition(DL, ElementTy, Offset, Size);
3300  }
3301  assert(Offset == 0);
3302 
3303  if (Size == ElementSize)
3304  return stripAggregateTypeWrapping(DL, ElementTy);
3305 
3306  StructType::element_iterator EI = STy->element_begin() + Index,
3307  EE = STy->element_end();
3308  if (EndOffset < SL->getSizeInBytes()) {
3309  unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3310  if (Index == EndIndex)
3311  return nullptr; // Within a single element and its padding.
3312 
3313  // Don't try to form "natural" types if the elements don't line up with the
3314  // expected size.
3315  // FIXME: We could potentially recurse down through the last element in the
3316  // sub-struct to find a natural end point.
3317  if (SL->getElementOffset(EndIndex) != EndOffset)
3318  return nullptr;
3319 
3320  assert(Index < EndIndex);
3321  EE = STy->element_begin() + EndIndex;
3322  }
3323 
3324  // Try to build up a sub-structure.
3325  StructType *SubTy =
3326  StructType::get(STy->getContext(), makeArrayRef(EI, EE), STy->isPacked());
3327  const StructLayout *SubSL = DL.getStructLayout(SubTy);
3328  if (Size != SubSL->getSizeInBytes())
3329  return nullptr; // The sub-struct doesn't have quite the size needed.
3330 
3331  return SubTy;
3332 }
3333 
3334 /// \brief Pre-split loads and stores to simplify rewriting.
3335 ///
3336 /// We want to break up the splittable load+store pairs as much as
3337 /// possible. This is important to do as a preprocessing step, as once we
3338 /// start rewriting the accesses to partitions of the alloca we lose the
3339 /// necessary information to correctly split apart paired loads and stores
3340 /// which both point into this alloca. The case to consider is something like
3341 /// the following:
3342 ///
3343 /// %a = alloca [12 x i8]
3344 /// %gep1 = getelementptr [12 x i8]* %a, i32 0, i32 0
3345 /// %gep2 = getelementptr [12 x i8]* %a, i32 0, i32 4
3346 /// %gep3 = getelementptr [12 x i8]* %a, i32 0, i32 8
3347 /// %iptr1 = bitcast i8* %gep1 to i64*
3348 /// %iptr2 = bitcast i8* %gep2 to i64*
3349 /// %fptr1 = bitcast i8* %gep1 to float*
3350 /// %fptr2 = bitcast i8* %gep2 to float*
3351 /// %fptr3 = bitcast i8* %gep3 to float*
3352 /// store float 0.0, float* %fptr1
3353 /// store float 1.0, float* %fptr2
3354 /// %v = load i64* %iptr1
3355 /// store i64 %v, i64* %iptr2
3356 /// %f1 = load float* %fptr2
3357 /// %f2 = load float* %fptr3
3358 ///
3359 /// Here we want to form 3 partitions of the alloca, each 4 bytes large, and
3360 /// promote everything so we recover the 2 SSA values that should have been
3361 /// there all along.
3362 ///
3363 /// \returns true if any changes are made.
3364 bool SROA::presplitLoadsAndStores(AllocaInst &AI, AllocaSlices &AS) {
3365  DEBUG(dbgs() << "Pre-splitting loads and stores\n");
3366 
3367  // Track the loads and stores which are candidates for pre-splitting here, in
3368  // the order they first appear during the partition scan. These give stable
3369  // iteration order and a basis for tracking which loads and stores we
3370  // actually split.
3373 
3374  // We need to accumulate the splits required of each load or store where we
3375  // can find them via a direct lookup. This is important to cross-check loads
3376  // and stores against each other. We also track the slice so that we can kill
3377  // all the slices that end up split.
3378  struct SplitOffsets {
3379  Slice *S;
3380  std::vector<uint64_t> Splits;
3381  };
3383 
3384  // Track loads out of this alloca which cannot, for any reason, be pre-split.
3385  // This is important as we also cannot pre-split stores of those loads!
3386  // FIXME: This is all pretty gross. It means that we can be more aggressive
3387  // in pre-splitting when the load feeding the store happens to come from
3388  // a separate alloca. Put another way, the effectiveness of SROA would be
3389  // decreased by a frontend which just concatenated all of its local allocas
3390  // into one big flat alloca. But defeating such patterns is exactly the job
3391  // SROA is tasked with! Sadly, to not have this discrepancy we would have
3392  // change store pre-splitting to actually force pre-splitting of the load
3393  // that feeds it *and all stores*. That makes pre-splitting much harder, but
3394  // maybe it would make it more principled?
3395  SmallPtrSet<LoadInst *, 8> UnsplittableLoads;
3396 
3397  DEBUG(dbgs() << " Searching for candidate loads and stores\n");
3398  for (auto &P : AS.partitions()) {
3399  for (Slice &S : P) {
3400  Instruction *I = cast<Instruction>(S.getUse()->getUser());
3401  if (!S.isSplittable() || S.endOffset() <= P.endOffset()) {
3402  // If this is a load we have to track that it can't participate in any
3403  // pre-splitting. If this is a store of a load we have to track that
3404  // that load also can't participate in any pre-splitting.
3405  if (auto *LI = dyn_cast<LoadInst>(I))
3406  UnsplittableLoads.insert(LI);
3407  else if (auto *SI = dyn_cast<StoreInst>(I))
3408  if (auto *LI = dyn_cast<LoadInst>(SI->getValueOperand()))
3409  UnsplittableLoads.insert(LI);
3410  continue;
3411  }
3412  assert(P.endOffset() > S.beginOffset() &&
3413  "Empty or backwards partition!");
3414 
3415  // Determine if this is a pre-splittable slice.
3416  if (auto *LI = dyn_cast<LoadInst>(I)) {
3417  assert(!LI->isVolatile() && "Cannot split volatile loads!");
3418 
3419  // The load must be used exclusively to store into other pointers for
3420  // us to be able to arbitrarily pre-split it. The stores must also be
3421  // simple to avoid changing semantics.
3422  auto IsLoadSimplyStored = [](LoadInst *LI) {
3423  for (User *LU : LI->users()) {
3424  auto *SI = dyn_cast<StoreInst>(LU);
3425  if (!SI || !SI->isSimple())
3426  return false;
3427  }
3428  return true;
3429  };
3430  if (!IsLoadSimplyStored(LI)) {
3431  UnsplittableLoads.insert(LI);
3432  continue;
3433  }
3434 
3435  Loads.push_back(LI);
3436  } else if (auto *SI = dyn_cast<StoreInst>(I)) {
3437  if (S.getUse() != &SI->getOperandUse(SI->getPointerOperandIndex()))
3438  // Skip stores *of* pointers. FIXME: This shouldn't even be possible!
3439  continue;
3440  auto *StoredLoad = dyn_cast<LoadInst>(SI->getValueOperand());
3441  if (!StoredLoad || !StoredLoad->isSimple())
3442  continue;
3443  assert(!SI->isVolatile() && "Cannot split volatile stores!");
3444 
3445  Stores.push_back(SI);
3446  } else {
3447  // Other uses cannot be pre-split.
3448  continue;
3449  }
3450 
3451  // Record the initial split.
3452  DEBUG(dbgs() << " Candidate: " << *I << "\n");
3453  auto &Offsets = SplitOffsetsMap[I];
3454  assert(Offsets.Splits.empty() &&
3455  "Should not have splits the first time we see an instruction!");
3456  Offsets.S = &S;
3457  Offsets.Splits.push_back(P.endOffset() - S.beginOffset());
3458  }
3459 
3460  // Now scan the already split slices, and add a split for any of them which
3461  // we're going to pre-split.
3462  for (Slice *S : P.splitSliceTails()) {
3463  auto SplitOffsetsMapI =
3464  SplitOffsetsMap.find(cast<Instruction>(S->getUse()->getUser()));
3465  if (SplitOffsetsMapI == SplitOffsetsMap.end())
3466  continue;
3467  auto &Offsets = SplitOffsetsMapI->second;
3468 
3469  assert(Offsets.S == S && "Found a mismatched slice!");
3470  assert(!Offsets.Splits.empty() &&
3471  "Cannot have an empty set of splits on the second partition!");
3472  assert(Offsets.Splits.back() ==
3473  P.beginOffset() - Offsets.S->beginOffset() &&
3474  "Previous split does not end where this one begins!");
3475 
3476  // Record each split. The last partition's end isn't needed as the size
3477  // of the slice dictates that.
3478  if (S->endOffset() > P.endOffset())
3479  Offsets.Splits.push_back(P.endOffset() - Offsets.S->beginOffset());
3480  }
3481  }
3482 
3483  // We may have split loads where some of their stores are split stores. For
3484  // such loads and stores, we can only pre-split them if their splits exactly
3485  // match relative to their starting offset. We have to verify this prior to
3486  // any rewriting.
3487  Stores.erase(
3488  remove_if(Stores,
3489  [&UnsplittableLoads, &SplitOffsetsMap](StoreInst *SI) {
3490  // Lookup the load we are storing in our map of split
3491  // offsets.
3492  auto *LI = cast<LoadInst>(SI->getValueOperand());
3493  // If it was completely unsplittable, then we're done,
3494  // and this store can't be pre-split.
3495  if (UnsplittableLoads.count(LI))
3496  return true;
3497 
3498  auto LoadOffsetsI = SplitOffsetsMap.find(LI);
3499  if (LoadOffsetsI == SplitOffsetsMap.end())
3500  return false; // Unrelated loads are definitely safe.
3501  auto &LoadOffsets = LoadOffsetsI->second;
3502 
3503  // Now lookup the store's offsets.
3504  auto &StoreOffsets = SplitOffsetsMap[SI];
3505 
3506  // If the relative offsets of each split in the load and
3507  // store match exactly, then we can split them and we
3508  // don't need to remove them here.
3509  if (LoadOffsets.Splits == StoreOffsets.Splits)
3510  return false;
3511 
3512  DEBUG(dbgs() << " Mismatched splits for load and store:\n"
3513  << " " << *LI << "\n"
3514  << " " << *SI << "\n");
3515 
3516  // We've found a store and load that we need to split
3517  // with mismatched relative splits. Just give up on them
3518  // and remove both instructions from our list of
3519  // candidates.
3520  UnsplittableLoads.insert(LI);
3521  return true;
3522  }),
3523  Stores.end());
3524  // Now we have to go *back* through all the stores, because a later store may
3525  // have caused an earlier store's load to become unsplittable and if it is
3526  // unsplittable for the later store, then we can't rely on it being split in
3527  // the earlier store either.
3528  Stores.erase(remove_if(Stores,
3529  [&UnsplittableLoads](StoreInst *SI) {
3530  auto *LI = cast<LoadInst>(SI->getValueOperand());
3531  return UnsplittableLoads.count(LI);
3532  }),
3533  Stores.end());
3534  // Once we've established all the loads that can't be split for some reason,
3535  // filter any that made it into our list out.
3536  Loads.erase(remove_if(Loads,
3537  [&UnsplittableLoads](LoadInst *LI) {
3538  return UnsplittableLoads.count(LI);
3539  }),
3540  Loads.end());
3541 
3542  // If no loads or stores are left, there is no pre-splitting to be done for
3543  // this alloca.
3544  if (Loads.empty() && Stores.empty())
3545  return false;
3546 
3547  // From here on, we can't fail and will be building new accesses, so rig up
3548  // an IR builder.
3549  IRBuilderTy IRB(&AI);
3550 
3551  // Collect the new slices which we will merge into the alloca slices.
3552  SmallVector<Slice, 4> NewSlices;
3553 
3554  // Track any allocas we end up splitting loads and stores for so we iterate
3555  // on them.
3556  SmallPtrSet<AllocaInst *, 4> ResplitPromotableAllocas;
3557 
3558  // At this point, we have collected all of the loads and stores we can
3559  // pre-split, and the specific splits needed for them. We actually do the
3560  // splitting in a specific order in order to handle when one of the loads in
3561  // the value operand to one of the stores.
3562  //
3563  // First, we rewrite all of the split loads, and just accumulate each split
3564  // load in a parallel structure. We also build the slices for them and append
3565  // them to the alloca slices.
3567  std::vector<LoadInst *> SplitLoads;
3568  const DataLayout &DL = AI.getModule()->getDataLayout();
3569  for (LoadInst *LI : Loads) {
3570  SplitLoads.clear();
3571 
3572  IntegerType *Ty = cast<IntegerType>(LI->getType());
3573  uint64_t LoadSize = Ty->getBitWidth() / 8;
3574  assert(LoadSize > 0 && "Cannot have a zero-sized integer load!");
3575 
3576  auto &Offsets = SplitOffsetsMap[LI];
3577  assert(LoadSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&
3578  "Slice size should always match load size exactly!");
3579  uint64_t BaseOffset = Offsets.S->beginOffset();
3580  assert(BaseOffset + LoadSize > BaseOffset &&
3581  "Cannot represent alloca access size using 64-bit integers!");
3582 
3583  Instruction *BasePtr = cast<Instruction>(LI->getPointerOperand());
3584  IRB.SetInsertPoint(LI);
3585 
3586  DEBUG(dbgs() << " Splitting load: " << *LI << "\n");
3587 
3588  uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
3589  int Idx = 0, Size = Offsets.Splits.size();
3590  for (;;) {
3591  auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
3592  auto AS = LI->getPointerAddressSpace();
3593  auto *PartPtrTy = PartTy->getPointerTo(AS);
3594  LoadInst *PLoad = IRB.CreateAlignedLoad(
3595  getAdjustedPtr(IRB, DL, BasePtr,
3596  APInt(DL.getPointerSizeInBits(AS), PartOffset),
3597  PartPtrTy, BasePtr->getName() + "."),
3598  getAdjustedAlignment(LI, PartOffset, DL), /*IsVolatile*/ false,
3599  LI->getName());
3601 
3602  // Append this load onto the list of split loads so we can find it later
3603  // to rewrite the stores.
3604  SplitLoads.push_back(PLoad);
3605 
3606  // Now build a new slice for the alloca.
3607  NewSlices.push_back(
3608  Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
3609  &PLoad->getOperandUse(PLoad->getPointerOperandIndex()),
3610  /*IsSplittable*/ false));
3611  DEBUG(dbgs() << " new slice [" << NewSlices.back().beginOffset()
3612  << ", " << NewSlices.back().endOffset() << "): " << *PLoad
3613  << "\n");
3614 
3615  // See if we've handled all the splits.
3616  if (Idx >= Size)
3617  break;
3618 
3619  // Setup the next partition.
3620  PartOffset = Offsets.Splits[Idx];
3621  ++Idx;
3622  PartSize = (Idx < Size ? Offsets.Splits[Idx] : LoadSize) - PartOffset;
3623  }
3624 
3625  // Now that we have the split loads, do the slow walk over all uses of the
3626  // load and rewrite them as split stores, or save the split loads to use
3627  // below if the store is going to be split there anyways.
3628  bool DeferredStores = false;
3629  for (User *LU : LI->users()) {
3630  StoreInst *SI = cast<StoreInst>(LU);
3631  if (!Stores.empty() && SplitOffsetsMap.count(SI)) {
3632  DeferredStores = true;
3633  DEBUG(dbgs() << " Deferred splitting of store: " << *SI << "\n");
3634  continue;
3635  }
3636 
3637  Value *StoreBasePtr = SI->getPointerOperand();
3638  IRB.SetInsertPoint(SI);
3639 
3640  DEBUG(dbgs() << " Splitting store of load: " << *SI << "\n");
3641 
3642  for (int Idx = 0, Size = SplitLoads.size(); Idx < Size; ++Idx) {
3643  LoadInst *PLoad = SplitLoads[Idx];
3644  uint64_t PartOffset = Idx == 0 ? 0 : Offsets.Splits[Idx - 1];
3645  auto *PartPtrTy =
3646  PLoad->getType()->getPointerTo(SI->getPointerAddressSpace());
3647 
3648  auto AS = SI->getPointerAddressSpace();
3649  StoreInst *PStore = IRB.CreateAlignedStore(
3650  PLoad,
3651  getAdjustedPtr(IRB, DL, StoreBasePtr,
3652  APInt(DL.getPointerSizeInBits(AS), PartOffset),
3653  PartPtrTy, StoreBasePtr->getName() + "."),
3654  getAdjustedAlignment(SI, PartOffset, DL), /*IsVolatile*/ false);
3656  DEBUG(dbgs() << " +" << PartOffset << ":" << *PStore << "\n");
3657  }
3658 
3659  // We want to immediately iterate on any allocas impacted by splitting
3660  // this store, and we have to track any promotable alloca (indicated by
3661  // a direct store) as needing to be resplit because it is no longer
3662  // promotable.
3663  if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(StoreBasePtr)) {
3664  ResplitPromotableAllocas.insert(OtherAI);
3665  Worklist.insert(OtherAI);
3666  } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
3667  StoreBasePtr->stripInBoundsOffsets())) {
3668  Worklist.insert(OtherAI);
3669  }
3670 
3671  // Mark the original store as dead.
3672  DeadInsts.insert(SI);
3673  }
3674 
3675  // Save the split loads if there are deferred stores among the users.
3676  if (DeferredStores)
3677  SplitLoadsMap.insert(std::make_pair(LI, std::move(SplitLoads)));
3678 
3679  // Mark the original load as dead and kill the original slice.
3680  DeadInsts.insert(LI);
3681  Offsets.S->kill();
3682  }
3683 
3684  // Second, we rewrite all of the split stores. At this point, we know that
3685  // all loads from this alloca have been split already. For stores of such
3686  // loads, we can simply look up the pre-existing split loads. For stores of
3687  // other loads, we split those loads first and then write split stores of
3688  // them.
3689  for (StoreInst *SI : Stores) {
3690  auto *LI = cast<LoadInst>(SI->getValueOperand());
3691  IntegerType *Ty = cast<IntegerType>(LI->getType());
3692  uint64_t StoreSize = Ty->getBitWidth() / 8;
3693  assert(StoreSize > 0 && "Cannot have a zero-sized integer store!");
3694 
3695  auto &Offsets = SplitOffsetsMap[SI];
3696  assert(StoreSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&
3697  "Slice size should always match load size exactly!");
3698  uint64_t BaseOffset = Offsets.S->beginOffset();
3699  assert(BaseOffset + StoreSize > BaseOffset &&
3700  "Cannot represent alloca access size using 64-bit integers!");
3701 
3702  Value *LoadBasePtr = LI->getPointerOperand();
3703  Instruction *StoreBasePtr = cast<Instruction>(SI->getPointerOperand());
3704 
3705  DEBUG(dbgs() << " Splitting store: " << *SI << "\n");
3706 
3707  // Check whether we have an already split load.
3708  auto SplitLoadsMapI = SplitLoadsMap.find(LI);
3709  std::vector<LoadInst *> *SplitLoads = nullptr;
3710  if (SplitLoadsMapI != SplitLoadsMap.end()) {
3711  SplitLoads = &SplitLoadsMapI->second;
3712  assert(SplitLoads->size() == Offsets.Splits.size() + 1 &&
3713  "Too few split loads for the number of splits in the store!");
3714  } else {
3715  DEBUG(dbgs() << " of load: " << *LI << "\n");
3716  }
3717 
3718  uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
3719  int Idx = 0, Size = Offsets.Splits.size();
3720  for (;;) {
3721  auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
3722  auto *LoadPartPtrTy = PartTy->getPointerTo(LI->getPointerAddressSpace());
3723  auto *StorePartPtrTy = PartTy->getPointerTo(SI->getPointerAddressSpace());
3724 
3725  // Either lookup a split load or create one.
3726  LoadInst *PLoad;
3727  if (SplitLoads) {
3728  PLoad = (*SplitLoads)[Idx];
3729  } else {
3730  IRB.SetInsertPoint(LI);
3731  auto AS = LI->getPointerAddressSpace();
3732  PLoad = IRB.CreateAlignedLoad(
3733  getAdjustedPtr(IRB, DL, LoadBasePtr,
3734  APInt(DL.getPointerSizeInBits(AS), PartOffset),
3735  LoadPartPtrTy, LoadBasePtr->getName() + "."),
3736  getAdjustedAlignment(LI, PartOffset, DL), /*IsVolatile*/ false,
3737  LI->getName());
3738  }
3739 
3740  // And store this partition.
3741  IRB.SetInsertPoint(SI);
3742  auto AS = SI->getPointerAddressSpace();
3743  StoreInst *PStore = IRB.CreateAlignedStore(
3744  PLoad,
3745  getAdjustedPtr(IRB, DL, StoreBasePtr,
3746  APInt(DL.getPointerSizeInBits(AS), PartOffset),
3747  StorePartPtrTy, StoreBasePtr->getName() + "."),
3748  getAdjustedAlignment(SI, PartOffset, DL), /*IsVolatile*/ false);
3749 
3750  // Now build a new slice for the alloca.
3751  NewSlices.push_back(
3752  Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
3753  &PStore->getOperandUse(PStore->getPointerOperandIndex()),
3754  /*IsSplittable*/ false));
3755  DEBUG(dbgs() << " new slice [" << NewSlices.back().beginOffset()
3756  << ", " << NewSlices.back().endOffset() << "): " << *PStore
3757  << "\n");
3758  if (!SplitLoads) {
3759  DEBUG(dbgs() << " of split load: " << *PLoad << "\n");
3760  }
3761 
3762  // See if we've finished all the splits.
3763  if (Idx >= Size)
3764  break;
3765 
3766  // Setup the next partition.
3767  PartOffset = Offsets.Splits[Idx];
3768  ++Idx;
3769  PartSize = (Idx < Size ? Offsets.Splits[Idx] : StoreSize) - PartOffset;
3770  }
3771 
3772  // We want to immediately iterate on any allocas impacted by splitting
3773  // this load, which is only relevant if it isn't a load of this alloca and
3774  // thus we didn't already split the loads above. We also have to keep track
3775  // of any promotable allocas we split loads on as they can no longer be
3776  // promoted.
3777  if (!SplitLoads) {
3778  if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(LoadBasePtr)) {
3779  assert(OtherAI != &AI && "We can't re-split our own alloca!");
3780  ResplitPromotableAllocas.insert(OtherAI);
3781  Worklist.insert(OtherAI);
3782  } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
3783  LoadBasePtr->stripInBoundsOffsets())) {
3784  assert(OtherAI != &AI && "We can't re-split our own alloca!");
3785  Worklist.insert(OtherAI);
3786  }
3787  }
3788 
3789  // Mark the original store as dead now that we've split it up and kill its
3790  // slice. Note that we leave the original load in place unless this store
3791  // was its only use. It may in turn be split up if it is an alloca load
3792  // for some other alloca, but it may be a normal load. This may introduce
3793  // redundant loads, but where those can be merged the rest of the optimizer
3794  // should handle the merging, and this uncovers SSA splits which is more
3795  // important. In practice, the original loads will almost always be fully
3796  // split and removed eventually, and the splits will be merged by any
3797  // trivial CSE, including instcombine.
3798  if (LI->hasOneUse()) {
3799  assert(*LI->user_begin() == SI && "Single use isn't this store!");
3800  DeadInsts.insert(LI);
3801  }
3802  DeadInsts.insert(SI);
3803  Offsets.S->kill();
3804  }
3805 
3806  // Remove the killed slices that have ben pre-split.
3807  AS.erase(remove_if(AS, [](const Slice &S) { return S.isDead(); }), AS.end());
3808 
3809  // Insert our new slices. This will sort and merge them into the sorted
3810  // sequence.
3811  AS.insert(NewSlices);
3812 
3813  DEBUG(dbgs() << " Pre-split slices:\n");
3814 #ifndef NDEBUG
3815  for (auto I = AS.begin(), E = AS.end(); I != E; ++I)
3816  DEBUG(AS.print(dbgs(), I, " "));
3817 #endif
3818 
3819  // Finally, don't try to promote any allocas that new require re-splitting.
3820  // They have already been added to the worklist above.
3821  PromotableAllocas.erase(
3822  remove_if(
3823  PromotableAllocas,
3824  [&](AllocaInst *AI) { return ResplitPromotableAllocas.count(AI); }),
3825  PromotableAllocas.end());
3826 
3827  return true;
3828 }
3829 
3830 /// \brief Rewrite an alloca partition's users.
3831 ///
3832 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3833 /// to rewrite uses of an alloca partition to be conducive for SSA value
3834 /// promotion. If the partition needs a new, more refined alloca, this will
3835 /// build that new alloca, preserving as much type information as possible, and
3836 /// rewrite the uses of the old alloca to point at the new one and have the
3837 /// appropriate new offsets. It also evaluates how successful the rewrite was
3838 /// at enabling promotion and if it was successful queues the alloca to be
3839 /// promoted.
3840 AllocaInst *SROA::rewritePartition(AllocaInst &AI, AllocaSlices &AS,
3841  Partition &P) {
3842  // Try to compute a friendly type for this partition of the alloca. This
3843  // won't always succeed, in which case we fall back to a legal integer type
3844  // or an i8 array of an appropriate size.
3845  Type *SliceTy = nullptr;
3846  const DataLayout &DL = AI.getModule()->getDataLayout();
3847  if (Type *CommonUseTy = findCommonType(P.begin(), P.end(), P.endOffset()))
3848  if (DL.getTypeAllocSize(CommonUseTy) >= P.size())
3849  SliceTy = CommonUseTy;
3850  if (!SliceTy)
3851  if (Type *TypePartitionTy = getTypePartition(DL, AI.getAllocatedType(),
3852  P.beginOffset(), P.size()))
3853  SliceTy = TypePartitionTy;
3854  if ((!SliceTy || (SliceTy->isArrayTy() &&
3855  SliceTy->getArrayElementType()->isIntegerTy())) &&
3856  DL.isLegalInteger(P.size() * 8))
3857  SliceTy = Type::getIntNTy(*C, P.size() * 8);
3858  if (!SliceTy)
3859  SliceTy = ArrayType::get(Type::getInt8Ty(*C), P.size());
3860  assert(DL.getTypeAllocSize(SliceTy) >= P.size());
3861 
3862  bool IsIntegerPromotable = isIntegerWideningViable(P, SliceTy, DL);
3863 
3864  VectorType *VecTy =
3865  IsIntegerPromotable ? nullptr : isVectorPromotionViable(P, DL);
3866  if (VecTy)
3867  SliceTy = VecTy;
3868 
3869  // Check for the case where we're going to rewrite to a new alloca of the
3870  // exact same type as the original, and with the same access offsets. In that
3871  // case, re-use the existing alloca, but still run through the rewriter to
3872  // perform phi and select speculation.
3873  AllocaInst *NewAI;
3874  if (SliceTy == AI.getAllocatedType()) {
3875  assert(P.beginOffset() == 0 &&
3876  "Non-zero begin offset but same alloca type");
3877  NewAI = &AI;
3878  // FIXME: We should be able to bail at this point with "nothing changed".
3879  // FIXME: We might want to defer PHI speculation until after here.
3880  // FIXME: return nullptr;
3881  } else {
3882  unsigned Alignment = AI.getAlignment();
3883  if (!Alignment) {
3884  // The minimum alignment which users can rely on when the explicit
3885  // alignment is omitted or zero is that required by the ABI for this
3886  // type.
3887  Alignment = DL.getABITypeAlignment(AI.getAllocatedType());
3888  }
3889  Alignment = MinAlign(Alignment, P.beginOffset());
3890  // If we will get at least this much alignment from the type alone, leave
3891  // the alloca's alignment unconstrained.
3892  if (Alignment <= DL.getABITypeAlignment(SliceTy))
3893  Alignment = 0;
3894  NewAI = new AllocaInst(
3895  SliceTy, AI.getType()->getAddressSpace(), nullptr, Alignment,
3896  AI.getName() + ".sroa." + Twine(P.begin() - AS.begin()), &AI);
3897  ++NumNewAllocas;
3898  }
3899 
3900  DEBUG(dbgs() << "Rewriting alloca partition "
3901  << "[" << P.beginOffset() << "," << P.endOffset()
3902  << ") to: " << *NewAI << "\n");
3903 
3904  // Track the high watermark on the worklist as it is only relevant for
3905  // promoted allocas. We will reset it to this point if the alloca is not in
3906  // fact scheduled for promotion.
3907  unsigned PPWOldSize = PostPromotionWorklist.size();
3908  unsigned NumUses = 0;
3910  SmallSetVector<SelectInst *, 8> SelectUsers;
3911 
3912  AllocaSliceRewriter Rewriter(DL, AS, *this, AI, *NewAI, P.beginOffset(),
3913  P.endOffset(), IsIntegerPromotable, VecTy,
3914  PHIUsers, SelectUsers);
3915  bool Promotable = true;
3916  for (Slice *S : P.splitSliceTails()) {
3917  Promotable &= Rewriter.visit(S);
3918  ++NumUses;
3919  }
3920  for (Slice &S : P) {
3921  Promotable &= Rewriter.visit(&S);
3922  ++NumUses;
3923  }
3924 
3925  NumAllocaPartitionUses += NumUses;
3926  MaxUsesPerAllocaPartition.updateMax(NumUses);
3927 
3928  // Now that we've processed all the slices in the new partition, check if any
3929  // PHIs or Selects would block promotion.
3930  for (PHINode *PHI : PHIUsers)
3931  if (!isSafePHIToSpeculate(*PHI)) {
3932  Promotable = false;
3933  PHIUsers.clear();
3934  SelectUsers.clear();
3935  break;
3936  }
3937 
3938  for (SelectInst *Sel : SelectUsers)
3939  if (!isSafeSelectToSpeculate(*Sel)) {
3940  Promotable = false;
3941  PHIUsers.clear();
3942  SelectUsers.clear();
3943  break;
3944  }
3945 
3946  if (Promotable) {
3947  if (PHIUsers.empty() && SelectUsers.empty()) {
3948  // Promote the alloca.
3949  PromotableAllocas.push_back(NewAI);
3950  } else {
3951  // If we have either PHIs or Selects to speculate, add them to those
3952  // worklists and re-queue the new alloca so that we promote in on the
3953  // next iteration.
3954  for (PHINode *PHIUser : PHIUsers)
3955  SpeculatablePHIs.insert(PHIUser);
3956  for (SelectInst *SelectUser : SelectUsers)
3957  SpeculatableSelects.insert(SelectUser);
3958  Worklist.insert(NewAI);
3959  }
3960  } else {
3961  // Drop any post-promotion work items if promotion didn't happen.
3962  while (PostPromotionWorklist.size() > PPWOldSize)
3963  PostPromotionWorklist.pop_back();
3964 
3965  // We couldn't promote and we didn't create a new partition, nothing
3966  // happened.
3967  if (NewAI == &AI)
3968  return nullptr;
3969 
3970  // If we can't promote the alloca, iterate on it to check for new
3971  // refinements exposed by splitting the current alloca. Don't iterate on an
3972  // alloca which didn't actually change and didn't get promoted.
3973  Worklist.insert(NewAI);
3974  }
3975 
3976  return NewAI;
3977 }
3978 
3979 /// \brief Walks the slices of an alloca and form partitions based on them,
3980 /// rewriting each of their uses.
3981 bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &AS) {
3982  if (AS.begin() == AS.end())
3983  return false;
3984 
3985  unsigned NumPartitions = 0;
3986  bool Changed = false;
3987  const DataLayout &DL = AI.getModule()->getDataLayout();
3988 
3989  // First try to pre-split loads and stores.
3990  Changed |= presplitLoadsAndStores(AI, AS);
3991 
3992  // Now that we have identified any pre-splitting opportunities, mark any
3993  // splittable (non-whole-alloca) loads and stores as unsplittable. If we fail
3994  // to split these during pre-splitting, we want to force them to be
3995  // rewritten into a partition.
3996  bool IsSorted = true;
3997  for (Slice &S : AS) {
3998  if (!S.isSplittable())
3999  continue;
4000  // FIXME: We currently leave whole-alloca splittable loads and stores. This
4001  // used to be the only splittable loads and stores and we need to be
4002  // confident that the above handling of splittable loads and stores is
4003  // completely sufficient before we forcibly disable the remaining handling.
4004  if (S.beginOffset() == 0 &&
4005  S.endOffset() >= DL.getTypeAllocSize(AI.getAllocatedType()))
4006  continue;
4007  if (isa<LoadInst>(S.getUse()->getUser()) ||
4008  isa<StoreInst>(S.getUse()->getUser())) {
4009  S.makeUnsplittable();
4010  IsSorted = false;
4011  }
4012  }
4013  if (!IsSorted)
4014  std::sort(AS.begin(), AS.end());
4015 
4016  /// Describes the allocas introduced by rewritePartition in order to migrate
4017  /// the debug info.
4018  struct Fragment {
4019  AllocaInst *Alloca;
4020  uint64_t Offset;
4021  uint64_t Size;
4022  Fragment(AllocaInst *AI, uint64_t O, uint64_t S)
4023  : Alloca(AI), Offset(O), Size(S) {}
4024  };
4025  SmallVector<Fragment, 4> Fragments;
4026 
4027  // Rewrite each partition.
4028  for (auto &P : AS.partitions()) {
4029  if (AllocaInst *NewAI = rewritePartition(AI, AS, P)) {
4030  Changed = true;
4031  if (NewAI != &AI) {
4032  uint64_t SizeOfByte = 8;
4033  uint64_t AllocaSize = DL.getTypeSizeInBits(NewAI->getAllocatedType());
4034  // Don't include any padding.
4035  uint64_t Size = std::min(AllocaSize, P.size() * SizeOfByte);
4036  Fragments.push_back(Fragment(NewAI, P.beginOffset() * SizeOfByte, Size));
4037  }
4038  }
4039  ++NumPartitions;
4040  }
4041 
4042  NumAllocaPartitions += NumPartitions;
4043  MaxPartitionsPerAlloca.updateMax(NumPartitions);
4044 
4045  // Migrate debug information from the old alloca to the new alloca(s)
4046  // and the individual partitions.
4047  if (DbgDeclareInst *DbgDecl = FindAllocaDbgDeclare(&AI)) {
4048  auto *Var = DbgDecl->getVariable();
4049  auto *Expr = DbgDecl->getExpression();
4050  DIBuilder DIB(*AI.getModule(), /*AllowUnresolved*/ false);
4051  uint64_t AllocaSize = DL.getTypeSizeInBits(AI.getAllocatedType());
4052  for (auto Fragment : Fragments) {
4053  // Create a fragment expression describing the new partition or reuse AI's
4054  // expression if there is only one partition.
4055  auto *FragmentExpr = Expr;
4056  if (Fragment.Size < AllocaSize || Expr->isFragment()) {
4057  // If this alloca is already a scalar replacement of a larger aggregate,
4058  // Fragment.Offset describes the offset inside the scalar.
4059  auto ExprFragment = Expr->getFragmentInfo();
4060  uint64_t Offset = ExprFragment ? ExprFragment->OffsetInBits : 0;
4061  uint64_t Start = Offset + Fragment.Offset;
4062  uint64_t Size = Fragment.Size;
4063  if (ExprFragment) {
4064  uint64_t AbsEnd =
4065  ExprFragment->OffsetInBits + ExprFragment->SizeInBits;
4066  if (Start >= AbsEnd)
4067  // No need to describe a SROAed padding.
4068  continue;
4069  Size = std::min(Size, AbsEnd - Start);
4070  }
4071  FragmentExpr = DIB.createFragmentExpression(Start, Size);
4072  }
4073 
4074  // Remove any existing dbg.declare intrinsic describing the same alloca.
4075  if (DbgDeclareInst *OldDDI = FindAllocaDbgDeclare(Fragment.Alloca))
4076  OldDDI->eraseFromParent();
4077 
4078  DIB.insertDeclare(Fragment.Alloca, Var, FragmentExpr,
4079  DbgDecl->getDebugLoc(), &AI);
4080  }
4081  }
4082  return Changed;
4083 }
4084 
4085 /// \brief Clobber a use with undef, deleting the used value if it becomes dead.
4086 void SROA::clobberUse(Use &U) {
4087  Value *OldV = U;
4088  // Replace the use with an undef value.
4089  U = UndefValue::get(OldV->getType());
4090 
4091  // Check for this making an instruction dead. We have to garbage collect
4092  // all the dead instructions to ensure the uses of any alloca end up being
4093  // minimal.
4094  if (Instruction *OldI = dyn_cast<Instruction>(OldV))
4095  if (isInstructionTriviallyDead(OldI)) {
4096  DeadInsts.insert(OldI);
4097  }
4098 }
4099 
4100 /// \brief Analyze an alloca for SROA.
4101 ///
4102 /// This analyzes the alloca to ensure we can reason about it, builds
4103 /// the slices of the alloca, and then hands it off to be split and
4104 /// rewritten as needed.
4105 bool SROA::runOnAlloca(AllocaInst &AI) {
4106  DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
4107  ++NumAllocasAnalyzed;
4108 
4109  // Special case dead allocas, as they're trivial.
4110  if (AI.use_empty()) {
4111  AI.eraseFromParent();
4112  return true;
4113  }
4114  const DataLayout &DL = AI.getModule()->getDataLayout();
4115 
4116  // Skip alloca forms that this analysis can't handle.
4117  if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
4118  DL.getTypeAllocSize(AI.getAllocatedType()) == 0)
4119  return false;
4120 
4121  bool Changed = false;
4122 
4123  // First, split any FCA loads and stores touching this alloca to promote
4124  // better splitting and promotion opportunities.
4125  AggLoadStoreRewriter AggRewriter;
4126  Changed |= AggRewriter.rewrite(AI);
4127 
4128  // Build the slices using a recursive instruction-visiting builder.
4129  AllocaSlices AS(DL, AI);
4130  DEBUG(AS.print(dbgs()));
4131  if (AS.isEscaped())
4132  return Changed;
4133 
4134  // Delete all the dead users of this alloca before splitting and rewriting it.
4135  for (Instruction *DeadUser : AS.getDeadUsers()) {
4136  // Free up everything used by this instruction.
4137  for (Use &DeadOp : DeadUser->operands())
4138  clobberUse(DeadOp);
4139 
4140  // Now replace the uses of this instruction.
4141  DeadUser->replaceAllUsesWith(UndefValue::get(DeadUser->getType()));
4142 
4143  // And mark it for deletion.
4144  DeadInsts.insert(DeadUser);
4145  Changed = true;
4146  }
4147  for (Use *DeadOp : AS.getDeadOperands()) {
4148  clobberUse(*DeadOp);
4149  Changed = true;
4150  }
4151 
4152  // No slices to split. Leave the dead alloca for a later pass to clean up.
4153  if (AS.begin() == AS.end())
4154  return Changed;
4155 
4156  Changed |= splitAlloca(AI, AS);
4157 
4158  DEBUG(dbgs() << " Speculating PHIs\n");
4159  while (!SpeculatablePHIs.empty())
4160  speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
4161 
4162  DEBUG(dbgs() << " Speculating Selects\n");
4163  while (!SpeculatableSelects.empty())
4164  speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
4165 
4166  return Changed;
4167 }
4168 
4169 /// \brief Delete the dead instructions accumulated in this run.
4170 ///
4171 /// Recursively deletes the dead instructions we've accumulated. This is done
4172 /// at the very end to maximize locality of the recursive delete and to
4173 /// minimize the problems of invalidated instruction pointers as such pointers
4174 /// are used heavily in the intermediate stages of the algorithm.
4175 ///
4176 /// We also record the alloca instructions deleted here so that they aren't
4177 /// subsequently handed to mem2reg to promote.
4178 void SROA::deleteDeadInstructions(
4179  SmallPtrSetImpl<AllocaInst *> &DeletedAllocas) {
4180  while (!DeadInsts.empty()) {
4181  Instruction *I = DeadInsts.pop_back_val();
4182  DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
4183 
4185 
4186  for (Use &Operand : I->operands())
4187  if (Instruction *U = dyn_cast<Instruction>(Operand)) {
4188  // Zero out the operand and see if it becomes trivially dead.
4189  Operand = nullptr;
4191  DeadInsts.insert(U);
4192  }
4193 
4194  if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
4195  DeletedAllocas.insert(AI);
4196  if (DbgDeclareInst *DbgDecl = FindAllocaDbgDeclare(AI))
4197  DbgDecl->eraseFromParent();
4198  }
4199 
4200  ++NumDeleted;
4201  I->eraseFromParent();
4202  }
4203 }
4204 
4205 /// \brief Promote the allocas, using the best available technique.
4206 ///
4207 /// This attempts to promote whatever allocas have been identified as viable in
4208 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
4209 /// This function returns whether any promotion occurred.
4210 bool SROA::promoteAllocas(Function &F) {
4211  if (PromotableAllocas.empty())
4212  return false;
4213 
4214  NumPromoted += PromotableAllocas.size();
4215 
4216  DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
4217  PromoteMemToReg(PromotableAllocas, *DT, AC);
4218  PromotableAllocas.clear();
4219  return true;
4220 }
4221 
4222 PreservedAnalyses SROA::runImpl(Function &F, DominatorTree &RunDT,
4223  AssumptionCache &RunAC) {
4224  DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
4225  C = &F.getContext();
4226  DT = &RunDT;
4227  AC = &RunAC;
4228 
4229  BasicBlock &EntryBB = F.getEntryBlock();
4230  for (BasicBlock::iterator I = EntryBB.begin(), E = std::prev(EntryBB.end());
4231  I != E; ++I) {
4232  if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
4233  Worklist.insert(AI);
4234  }
4235 
4236  bool Changed = false;
4237  // A set of deleted alloca instruction pointers which should be removed from
4238  // the list of promotable allocas.
4239  SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
4240 
4241  do {
4242  while (!Worklist.empty()) {
4243  Changed |= runOnAlloca(*Worklist.pop_back_val());
4244  deleteDeadInstructions(DeletedAllocas);
4245 
4246  // Remove the deleted allocas from various lists so that we don't try to
4247  // continue processing them.
4248  if (!DeletedAllocas.empty()) {
4249  auto IsInSet = [&](AllocaInst *AI) { return DeletedAllocas.count(AI); };
4250  Worklist.remove_if(IsInSet);
4251  PostPromotionWorklist.remove_if(IsInSet);
4252  PromotableAllocas.erase(remove_if(PromotableAllocas, IsInSet),
4253  PromotableAllocas.end());
4254  DeletedAllocas.clear();
4255  }
4256  }
4257 
4258  Changed |= promoteAllocas(F);
4259 
4260  Worklist = PostPromotionWorklist;
4261  PostPromotionWorklist.clear();
4262  } while (!Worklist.empty());
4263 
4264  if (!Changed)
4265  return PreservedAnalyses::all();
4266 
4267  PreservedAnalyses PA;
4268  PA.preserveSet<CFGAnalyses>();
4269  PA.preserve<GlobalsAA>();
4270  return PA;
4271 }
4272 
4274  return runImpl(F, AM.getResult<DominatorTreeAnalysis>(F),
4275  AM.getResult<AssumptionAnalysis>(F));
4276 }
4277 
4278 /// A legacy pass for the legacy pass manager that wraps the \c SROA pass.
4279 ///
4280 /// This is in the llvm namespace purely to allow it to be a friend of the \c
4281 /// SROA pass.
4283  /// The SROA implementation.
4284  SROA Impl;
4285 
4286 public:
4289  }
4290  bool runOnFunction(Function &F) override {
4291  if (skipFunction(F))
4292  return false;
4293 
4294  auto PA = Impl.runImpl(
4295  F, getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
4296  getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F));
4297  return !PA.areAllPreserved();
4298  }
4299  void getAnalysisUsage(AnalysisUsage &AU) const override {
4303  AU.setPreservesCFG();
4304  }
4305 
4306  StringRef getPassName() const override { return "SROA"; }
4307  static char ID;
4308 };
4309 
4310 char SROALegacyPass::ID = 0;
4311 
4313 
4315  "Scalar Replacement Of Aggregates", false, false)
4319  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:1491
RetTy visitSelectInst(SelectInst &I)
Definition: InstVisitor.h:201
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:1789
iterator end() const
Definition: SROA.cpp:372
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:1377
Value * getValueOperand()
Definition: Instructions.h:395
SymbolTableList< Instruction >::iterator eraseFromParent()
This method unlinks &#39;this&#39; from the containing basic block and deletes it.
Definition: Instruction.cpp:69
IntegerType * getType() const
getType - Specialize the getType() method to always return an IntegerType, which reduces the amount o...
Definition: Constants.h:172
void push_back(const T &Elt)
Definition: SmallVector.h:212
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:109
const_iterator end(StringRef path)
Get end iterator over path.
Definition: Path.cpp:243
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:396
An iterator over partitions of the alloca&#39;s slices.
Definition: SROA.cpp:392
RetTy visitMemSetInst(MemSetInst &I)
Definition: InstVisitor.h:217
void setInt(IntType IntVal)
bool isSimple() const
Definition: Instructions.h:262
iterator_range< use_iterator > uses()
Definition: Value.h:350
AnalysisUsage & addPreserved()
Add the specified Pass class to the set of analyses preserved by this pass.
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:1542
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:234
const Value * stripInBoundsOffsets() const
Strip off pointer casts and inbounds GEPs.
Definition: Value.cpp:587
uint64_t beginOffset() const
The start offset of this partition.
Definition: SROA.cpp:343
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
#define LLVM_DUMP_METHOD
Mark debug helper function definitions like dump() that should not be stripped from debug builds...
Definition: Compiler.h:449
Type * getElementType(unsigned N) const
Definition: DerivedTypes.h:314
This is the interface for a simple mod/ref and alias analysis over globals.
bool isSized(SmallPtrSetImpl< Type *> *Visited=nullptr) const
Return true if it makes sense to take the size of this type.
Definition: Type.h:262
iterator begin() const
Definition: ArrayRef.h:137
LLVM_ATTRIBUTE_ALWAYS_INLINE size_type size() const
Definition: SmallVector.h:136
DbgDeclareInst * FindAllocaDbgDeclare(Value *V)
Finds the llvm.dbg.declare intrinsic corresponding to an alloca, if any.
Definition: Local.cpp:1238
APInt sdiv(const APInt &RHS) const
Signed division function for APInt.
Definition: APInt.cpp:1590
auto remove_if(R &&Range, UnaryPredicate P) -> decltype(std::begin(Range))
Provide wrappers to std::remove_if which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:858
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:562
This file provides the interface for LLVM&#39;s Scalar Replacement of Aggregates pass.
constexpr char IsVolatile[]
Key for Kernel::Arg::Metadata::mIsVolatile.
void erase(iterator Start, iterator Stop)
Erase a range of slices.
Definition: SROA.cpp:222
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:408
This class provides information about the result of a visit.
Definition: PtrUseVisitor.h:49
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:2018
Representation of the alloca slices.
Definition: SROA.cpp:197
SmallVectorImpl< Slice >::iterator iterator
Support for iterating over the slices.
Definition: SROA.cpp:210
An immutable pass that tracks lazily created AssumptionCache objects.
Instruction * getAbortingInst() const
Get the instruction causing the visit to abort.
Definition: PtrUseVisitor.h:70
SyncScope::ID getSyncScopeID() const
Returns the synchronization scope ID of this store instruction.
Definition: Instructions.h:370
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:229
A cache of .assume calls within a function.
AtomicOrdering getOrdering() const
Returns the ordering constraint of this load instruction.
Definition: Instructions.h:233
Offsets
Offsets in bytes from the start of the input buffer.
Definition: SIInstrInfo.h:864
LLVMContext & getContext() const
All values hold a context through their type.
Definition: Value.cpp:697
void deleteValue()
Delete a pointer to a generic Value.
Definition: Value.cpp:93
This class wraps the llvm.memset intrinsic.
const Use & getOperandUse(unsigned i) const
Definition: User.h:167
unsigned second
Scalar Replacement Of Aggregates
Definition: SROA.cpp:4318
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:818
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:346
STATISTIC(NumFunctions, "Total number of functions")
Metadata node.
Definition: Metadata.h:862
Analysis pass which computes a DominatorTree.
Definition: Dominators.h:232
unsigned getPointerAddressSpace() const
Get the address space of this pointer or pointer vector type.
Definition: DerivedTypes.h:503
An instruction for reading from memory.
Definition: Instructions.h:164
RetTy visitPHINode(PHINode &I)
Definition: InstVisitor.h:187
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:227
void reserve(size_type N)
Definition: SmallVector.h:380
void setDest(Value *Ptr)
Set the specified arguments of the instruction.
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:381
bool isEscaped() const
Test whether a pointer to the allocation escapes our analysis.
Definition: SROA.cpp:206
void setAlignment(Constant *A)
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:214
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:1488
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:1586
Builder for the alloca slices.
Definition: SROA.cpp:608
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:345
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:252
Value * getRawSource() const
Return the arguments to the instruction.
std::pair< iterator, bool > insert(const std::pair< KeyT, ValueT > &KV)
Definition: DenseMap.h:176
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:256
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:1713
PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM)
Run the pass over the function.
Definition: SROA.cpp:4273
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:493
#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:217
This class represents the LLVM &#39;select&#39; instruction.
Type * getPointerElementType() const
Definition: Type.h:373
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:109
RetTy visitIntrinsicInst(IntrinsicInst &I)
Definition: InstVisitor.h:225
unsigned getPointerTypeSizeInBits(Type *) const
Layout pointer size, in bits, based on the type.
Definition: DataLayout.cpp:614
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:3241
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:380
bool isAborted() const
Did we abort the visit early?
Definition: PtrUseVisitor.h:62
static bool runImpl(CallGraphSCC &SCC, AARGetterT AARGetter)
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:664
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:1663
A partition of the slices.
Definition: SROA.cpp:318
void getAnalysisUsage(AnalysisUsage &AU) const override
getAnalysisUsage - This function should be overriden by passes that need analysis information to do t...
Definition: SROA.cpp:4299
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
bool runOnFunction(Function &F) override
runOnFunction - Virtual method overriden by subclasses to do the per-function processing of the pass...
Definition: SROA.cpp:4290
This file provides a collection of visitors which walk the (instruction) uses of a pointer...
void visit(Iterator Start, Iterator End)
Definition: InstVisitor.h:90
static Constant * getZExt(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:1570
unsigned getPointerAddressSpace() const
Returns the address space of the pointer operand.
Definition: Instructions.h:404
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:2235
A base class for visitors over the uses of a pointer value.
#define F(x, y, z)
Definition: MD5.cpp:55
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:1852
bool isInBounds() const
Determine whether the GEP has the inbounds flag.
const_iterator end() const
Definition: SROA.cpp:218
IntType getInt() const
A legacy pass for the legacy pass manager that wraps the SROA pass.
Definition: SROA.cpp:4282
Class to represent array types.
Definition: DerivedTypes.h:369
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:190
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:306
bool isIntOrIntVectorTy() const
Return true if this is an integer type or a vector of integer types.
Definition: Type.h:203
auto count(R &&Range, const E &Element) -> typename std::iterator_traits< decltype(std::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:880
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
Definition: Value.cpp:428
void printUse(raw_ostream &OS, const_iterator I, StringRef Indent=" ") const
LLVM_NODISCARD LLVM_ATTRIBUTE_ALWAYS_INLINE StringRef substr(size_t Start, size_t N=npos) const
Return a reference to the substring from [Start, Start + N).
Definition: StringRef.h:598
static Constant * getUDiv(Constant *C1, Constant *C2, bool isExact=false)
Definition: Constants.cpp:2148
unsigned getBitWidth() const
Get the number of bits in this IntegerType.
Definition: DerivedTypes.h:66
RetTy visitMemTransferInst(MemTransferInst &I)
Definition: InstVisitor.h:220
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:134
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:1612
const char * Name
Value * getOperand(unsigned i) const
Definition: User.h:154
Class to represent pointers.
Definition: DerivedTypes.h:467
Type * getScalarType() const
If this is a vector type, return the element type, otherwise return &#39;this&#39;.
Definition: Type.h:301
iterator find(const_arg_type_t< KeyT > Val)
Definition: DenseMap.h:131
const BasicBlock & getEntryBlock() const
Definition: Function.h:564
constexpr uint64_t MinAlign(uint64_t A, uint64_t B)
A and B are either alignments or offsets.
Definition: MathExtras.h:602
an instruction for type-safe pointer arithmetic to access elements of arrays and structs ...
Definition: Instructions.h:837
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:702
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:4318
iterator_range< partition_iterator > partitions()
SmallVectorImpl< Slice >::const_iterator const_iterator
Definition: SROA.cpp:215
#define P(N)
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:404
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
void dump(const SparseBitVector< ElementSize > &LHS, raw_ostream &out)
Subclasses of this class are all able to terminate a basic block.
Definition: InstrTypes.h:54
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:200
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:1253
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:400
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:116
ArrayRef< Use * > getDeadOperands() const
Access the dead operands referring to this alloca.
Definition: SROA.cpp:251
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:221
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:1245
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:265
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:372
bool mayHaveSideEffects() const
Return true if the instruction may have side effects.
Definition: Instruction.h:500
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:1451
static sys::TimePoint< std::chrono::seconds > now(bool Deterministic)
Value * getRawDest() const
element_iterator element_end() const
Definition: DerivedTypes.h:304
Represent the analysis usage information of a pass.
op_iterator op_end()
Definition: User.h:216
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:825
#define LLVM_ATTRIBUTE_UNUSED
Definition: Compiler.h:144
Analysis pass providing a never-invalidated alias analysis result.
void initializeSROALegacyPassPass(PassRegistry &)
sroa
Definition: SROA.cpp:4318
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:298
op_range operands()
Definition: User.h:222
Value * getPointerOperand()
Definition: Instructions.h:270
unsigned getAddressSpace() const
Return the address space of the Pointer type.
Definition: DerivedTypes.h:495
size_type count(ConstPtrType Ptr) const
count - Return 1 if the specified pointer is in the set, 0 otherwise.
Definition: SmallPtrSet.h:383
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:2205
RetTy visitLoadInst(LoadInst &I)
Definition: InstVisitor.h:181
void setAlignment(unsigned Align)
const Value * getCondition() const
static Constant * getAllOnesValue(Type *Ty)
Get the all ones value.
Definition: Constants.cpp:261
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:1320
const AMDGPUAS & AS
iterator erase(const_iterator CI)
Definition: SmallVector.h:449
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:159
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:102
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:1333
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:3203
Intrinsic::ID getIntrinsicID() const
Return the intrinsic ID of this intrinsic.
Definition: IntrinsicInst.h:51
This provides the default implementation of the IRBuilder &#39;InsertHelper&#39; method that is called whenev...
Definition: IRBuilder.h:63
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:224
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
#define E
Definition: LargeTest.cpp:27
This is the shared class of boolean and integer constants.
Definition: Constants.h:84
#define B
Definition: LargeTest.cpp:24
iterator end()
Definition: BasicBlock.h:254
AllocaSlices(const DataLayout &DL, AllocaInst &AI)
Construct the slices of a particular alloca.
Instruction * user_back()
Specialize the methods defined in Value, as we know that an instruction can only be used by other ins...
Definition: Instruction.h:63
partition_iterator & operator++()
Definition: SROA.cpp:563
static Value * insertVector(IRBuilderTy &IRB, Value *Old, Value *V, unsigned BeginIndex, const Twine &Name)
Definition: SROA.cpp:2098
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:238
unsigned getABITypeAlignment(Type *Ty) const
Returns the minimum ABI-required alignment for the specified type.
Definition: DataLayout.cpp:682
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:371
const size_t N
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:385
uint64_t getSizeInBytes() const
Definition: DataLayout.h:501
static IntegerType * getIntNTy(LLVMContext &C, unsigned N)
Definition: Type.cpp:180
SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &AS)
Definition: SROA.cpp:623
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:560
void setPreservesCFG()
This function should be called by the pass, iff they do not:
Definition: Pass.cpp:278
unsigned getNumIncomingValues() const
Return the number of incoming edges.
bool uge(const APInt &RHS) const
Unsigned greater or equal comparison.
Definition: APInt.h:1272
void setOperand(unsigned i, Value *Val)
Definition: User.h:159
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:132
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:923
A range adaptor for a pair of iterators.
Class to represent vector types.
Definition: DerivedTypes.h:393
ConstantInt * getAlignmentCst() const
const Module * getModule() const
Return the module owning the function this instruction belongs to or nullptr it the function does not...
Definition: Instruction.cpp:57
Class for arbitrary precision integers.
Definition: APInt.h:69
static bool isIntegerWideningViableForSlice(const Slice &S, uint64_t AllocBeginOffset, Type *AllocaTy, const DataLayout &DL, bool &WholeAllocaOp)
Test whether a slice of an alloca is valid for integer widening.
Definition: SROA.cpp:1896
iterator_range< user_iterator > users()
Definition: Value.h:395
RetTy visitStoreInst(StoreInst &I)
Definition: InstVisitor.h:182
#define NDEBUG
Definition: regutils.h:48
Represents analyses that only rely on functions&#39; control flow.
Definition: PassManager.h:114
const Value * getFalseValue() const
element_iterator element_begin() const
Definition: DerivedTypes.h:303
static Value * insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old, Value *V, uint64_t Offset, const Twine &Name)
Definition: SROA.cpp:2041
void setLength(Value *L)
bool isNonIntegralPointerType(PointerType *PT) const
Definition: DataLayout.h:332
constexpr char Size[]
Key for Kernel::Arg::Metadata::mSize.
const_iterator begin() const
Definition: SROA.cpp:217
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:1058
uint64_t getTypeSizeInBits(Type *Ty) const
Size examples:
Definition: DataLayout.h:532
static Value * foldPHINodeOrSelectInst(Instruction &I)
A helper that folds a PHI node or a select.
Definition: SROA.cpp:596
uint64_t getTypeAllocSize(Type *Ty) const
Returns the offset in bytes between successive objects of the specified type, including alignment pad...
Definition: DataLayout.h:405
Virtual Register Rewriter
Definition: VirtRegMap.cpp:207
bool operator!=(uint64_t V1, const APInt &V2)
Definition: APInt.h:1948
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:1234
static bool isZero(Value *V, const DataLayout &DL, DominatorTree *DT, AssumptionCache *AC)
Definition: Lint.cpp:529
This class wraps the llvm.memcpy/memmove intrinsics.
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator end()
Definition: SmallVector.h:120
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:339
const DebugLoc & getDebugLoc() const
Return the debug location for this node as a DebugLoc.
Definition: Instruction.h:280
static const size_t npos
Definition: StringRef.h:51
uint64_t getElementOffset(unsigned Idx) const
Definition: DataLayout.h:515
unsigned getAlignment() const
Return the alignment of the access that is being performed.
Definition: Instructions.h:226
unsigned getIntegerBitWidth() const
Definition: DerivedTypes.h:97
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:61
AtomicOrdering getOrdering() const
Returns the ordering constraint of this store instruction.
Definition: Instructions.h:358
StringRef getPassName() const override
getPassName - Return a nice clean name for a pass.
Definition: SROA.cpp:4306
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:475
Visitor to rewrite instructions using p particular slice of an alloca to use a new alloca...
Definition: SROA.cpp:2151
void preserveSet()
Mark an analysis set as preserved.
Definition: PassManager.h:189
Value * getArgOperand(unsigned i) const
getArgOperand/setArgOperand - Return/set the i-th call argument.
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:218
BasicBlock * getIncomingBlock(unsigned i) const
Return incoming basic block number i.
static void speculateSelectInstLoads(SelectInst &SI)
Definition: SROA.cpp:1267
SyncScope::ID getSyncScopeID() const
Returns the synchronization scope ID of this load instruction.
Definition: Instructions.h:245
#define I(x, y, z)
Definition: MD5.cpp:58
constexpr char Align[]
Key for Kernel::Arg::Metadata::mAlign.
iterator end()
Definition: DenseMap.h:73
FunctionPass * createSROAPass()
Definition: SROA.cpp:4312
void setSource(Value *Ptr)
uint64_t endOffset() const
The end offset of this partition.
Definition: SROA.cpp:348
Instruction * getEscapingInst() const
Get the instruction causing the pointer to escape.
Definition: PtrUseVisitor.h:75
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
Value * getValue() const
Return the arguments to the instruction.
void preserve()
Mark an analysis as preserved.
Definition: PassManager.h:174
static Value * extractVector(IRBuilderTy &IRB, Value *V, unsigned BeginIndex, unsigned EndIndex, const Twine &Name)
Definition: SROA.cpp:2072
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:126
unsigned getAlignment() const
std::string str() const
Return the twine contents as a std::string.
Definition: Twine.cpp:17
iterator_range< iterator > range
Definition: SROA.cpp:211
unsigned getPointerAddressSpace() const
Returns the address space of the pointer operand.
Definition: Instructions.h:276
static unsigned getPointerOperandIndex()
Definition: Instructions.h:272
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:371
unsigned getNumSuccessors() const
Return the number of successors that this terminator has.
bool operator<(int64_t V1, const APSInt &V2)
Definition: APSInt.h:326
Value * getLength() const
bool isSingleValueType() const
Return true if the type is a valid type for a register in codegen.
Definition: Type.h:247
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:293
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:1310
uint64_t getTypeStoreSize(Type *Ty) const
Returns the maximum number of bytes that may be overwritten by storing the specified type...
Definition: DataLayout.h:388
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:1125
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:65
iterator_range< const_iterator > const_range
Definition: SROA.cpp:216
static void speculatePHINodeLoads(PHINode &PN)
Definition: SROA.cpp:1189
This class implements an extremely fast bulk output stream that can only output to a stream...
Definition: raw_ostream.h:44
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...
#define DEBUG(X)
Definition: Debug.h:118
PtrInfo visitPtr(Instruction &I)
Recursively visit the uses of the given pointer.
An optimization pass providing Scalar Replacement of Aggregates.
Definition: SROA.h:58
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:353
IRTranslator LLVM IR MI
bool hasOneUse() const
Return true if there is exactly one user of this value.
Definition: Value.h:408
bool empty() const
Test whether this partition contains no slices, and merely spans a region occupied by split slices...
Definition: SROA.cpp:360
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:49
static Expected< std::string > replace(StringRef S, StringRef From, StringRef To)
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:362
Legacy analysis pass which computes a DominatorTree.
Definition: Dominators.h:261
void sort(Policy policy, RandomAccessIterator Start, RandomAccessIterator End, const Comparator &Comp=Comparator())
Definition: Parallel.h:201
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:1976
bool operator==(uint64_t V1, const APInt &V2)
Definition: APInt.h:1946
int * Ptr
bool isSimple() const
Definition: Instructions.h:387
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:120
const Use & getRawDestUse() const
bool isBigEndian() const
Definition: DataLayout.h:217
void visitInstruction(Instruction &I)
Definition: InstVisitor.h:265
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
This represents the llvm.dbg.declare instruction.
Definition: IntrinsicInst.h:89
Value * getPointerOperand()
Definition: Instructions.h:398
static IntegerType * getInt8Ty(LLVMContext &C)
Definition: Type.cpp:174
bool use_empty() const
Definition: Value.h:322
static Constant * get(ArrayRef< Constant *> V)
Definition: Constants.cpp:984
Maximum number of bits that can be specified.
Definition: DerivedTypes.h:52
Type * getElementType() const
Definition: DerivedTypes.h:486
bool isArrayTy() const
True if this is an instance of ArrayType.
Definition: Type.h:218
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:44
const BasicBlock * getParent() const
Definition: Instruction.h:66
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:583
ArrayRef< Instruction * > getDeadUsers() const
Access the dead users for this alloca.
Definition: SROA.cpp:243
bool isEscaped() const
Is the pointer escaped at some point?
Definition: PtrUseVisitor.h:65
bool operator==(const partition_iterator &RHS) const
Definition: SROA.cpp:543