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/APInt.h"
28 #include "llvm/ADT/ArrayRef.h"
29 #include "llvm/ADT/DenseMap.h"
31 #include "llvm/ADT/STLExtras.h"
32 #include "llvm/ADT/SetVector.h"
33 #include "llvm/ADT/SmallPtrSet.h"
34 #include "llvm/ADT/SmallVector.h"
35 #include "llvm/ADT/Statistic.h"
36 #include "llvm/ADT/StringRef.h"
37 #include "llvm/ADT/Twine.h"
38 #include "llvm/ADT/iterator.h"
42 #include "llvm/Analysis/Loads.h"
44 #include "llvm/IR/BasicBlock.h"
45 #include "llvm/IR/Constant.h"
46 #include "llvm/IR/ConstantFolder.h"
47 #include "llvm/IR/Constants.h"
48 #include "llvm/IR/DIBuilder.h"
49 #include "llvm/IR/DataLayout.h"
51 #include "llvm/IR/DerivedTypes.h"
52 #include "llvm/IR/Dominators.h"
53 #include "llvm/IR/Function.h"
55 #include "llvm/IR/GlobalAlias.h"
56 #include "llvm/IR/IRBuilder.h"
57 #include "llvm/IR/InstVisitor.h"
58 #include "llvm/IR/InstrTypes.h"
59 #include "llvm/IR/Instruction.h"
60 #include "llvm/IR/Instructions.h"
61 #include "llvm/IR/IntrinsicInst.h"
62 #include "llvm/IR/Intrinsics.h"
63 #include "llvm/IR/LLVMContext.h"
64 #include "llvm/IR/Metadata.h"
65 #include "llvm/IR/Module.h"
66 #include "llvm/IR/Operator.h"
67 #include "llvm/IR/PassManager.h"
68 #include "llvm/IR/Type.h"
69 #include "llvm/IR/Use.h"
70 #include "llvm/IR/User.h"
71 #include "llvm/IR/Value.h"
72 #include "llvm/Pass.h"
73 #include "llvm/Support/Casting.h"
75 #include "llvm/Support/Compiler.h"
76 #include "llvm/Support/Debug.h"
80 #include "llvm/Transforms/Scalar.h"
83 #include <algorithm>
84 #include <cassert>
85 #include <chrono>
86 #include <cstddef>
87 #include <cstdint>
88 #include <cstring>
89 #include <iterator>
90 #include <string>
91 #include <tuple>
92 #include <utility>
93 #include <vector>
94 
95 #ifndef NDEBUG
96 // We only use this for a debug check.
97 #include <random>
98 #endif
99 
100 using namespace llvm;
101 using namespace llvm::sroa;
102 
103 #define DEBUG_TYPE "sroa"
104 
105 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
106 STATISTIC(NumAllocaPartitions, "Number of alloca partitions formed");
107 STATISTIC(MaxPartitionsPerAlloca, "Maximum number of partitions per alloca");
108 STATISTIC(NumAllocaPartitionUses, "Number of alloca partition uses rewritten");
109 STATISTIC(MaxUsesPerAllocaPartition, "Maximum number of uses of a partition");
110 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
111 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
112 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
113 STATISTIC(NumDeleted, "Number of instructions deleted");
114 STATISTIC(NumVectorized, "Number of vectorized aggregates");
115 
116 /// Hidden option to enable randomly shuffling the slices to help uncover
117 /// instability in their order.
118 static cl::opt<bool> SROARandomShuffleSlices("sroa-random-shuffle-slices",
119  cl::init(false), cl::Hidden);
120 
121 /// Hidden option to experiment with completely strict handling of inbounds
122 /// GEPs.
123 static cl::opt<bool> SROAStrictInbounds("sroa-strict-inbounds", cl::init(false),
124  cl::Hidden);
125 
126 namespace {
127 
128 /// \brief A custom IRBuilder inserter which prefixes all names, but only in
129 /// Assert builds.
130 class IRBuilderPrefixedInserter : public IRBuilderDefaultInserter {
131  std::string Prefix;
132 
133  const Twine getNameWithPrefix(const Twine &Name) const {
134  return Name.isTriviallyEmpty() ? Name : Prefix + Name;
135  }
136 
137 public:
138  void SetNamePrefix(const Twine &P) { Prefix = P.str(); }
139 
140 protected:
141  void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
142  BasicBlock::iterator InsertPt) const {
143  IRBuilderDefaultInserter::InsertHelper(I, getNameWithPrefix(Name), BB,
144  InsertPt);
145  }
146 };
147 
148 /// \brief Provide a type for IRBuilder that drops names in release builds.
150 
151 /// \brief A used slice of an alloca.
152 ///
153 /// This structure represents a slice of an alloca used by some instruction. It
154 /// stores both the begin and end offsets of this use, a pointer to the use
155 /// itself, and a flag indicating whether we can classify the use as splittable
156 /// or not when forming partitions of the alloca.
157 class Slice {
158  /// \brief The beginning offset of the range.
159  uint64_t BeginOffset = 0;
160 
161  /// \brief The ending offset, not included in the range.
162  uint64_t EndOffset = 0;
163 
164  /// \brief Storage for both the use of this slice and whether it can be
165  /// split.
166  PointerIntPair<Use *, 1, bool> UseAndIsSplittable;
167 
168 public:
169  Slice() = default;
170 
171  Slice(uint64_t BeginOffset, uint64_t EndOffset, Use *U, bool IsSplittable)
172  : BeginOffset(BeginOffset), EndOffset(EndOffset),
173  UseAndIsSplittable(U, IsSplittable) {}
174 
175  uint64_t beginOffset() const { return BeginOffset; }
176  uint64_t endOffset() const { return EndOffset; }
177 
178  bool isSplittable() const { return UseAndIsSplittable.getInt(); }
179  void makeUnsplittable() { UseAndIsSplittable.setInt(false); }
180 
181  Use *getUse() const { return UseAndIsSplittable.getPointer(); }
182 
183  bool isDead() const { return getUse() == nullptr; }
184  void kill() { UseAndIsSplittable.setPointer(nullptr); }
185 
186  /// \brief Support for ordering ranges.
187  ///
188  /// This provides an ordering over ranges such that start offsets are
189  /// always increasing, and within equal start offsets, the end offsets are
190  /// decreasing. Thus the spanning range comes first in a cluster with the
191  /// same start position.
192  bool operator<(const Slice &RHS) const {
193  if (beginOffset() < RHS.beginOffset())
194  return true;
195  if (beginOffset() > RHS.beginOffset())
196  return false;
197  if (isSplittable() != RHS.isSplittable())
198  return !isSplittable();
199  if (endOffset() > RHS.endOffset())
200  return true;
201  return false;
202  }
203 
204  /// \brief Support comparison with a single offset to allow binary searches.
205  friend LLVM_ATTRIBUTE_UNUSED bool operator<(const Slice &LHS,
206  uint64_t RHSOffset) {
207  return LHS.beginOffset() < RHSOffset;
208  }
209  friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
210  const Slice &RHS) {
211  return LHSOffset < RHS.beginOffset();
212  }
213 
214  bool operator==(const Slice &RHS) const {
215  return isSplittable() == RHS.isSplittable() &&
216  beginOffset() == RHS.beginOffset() && endOffset() == RHS.endOffset();
217  }
218  bool operator!=(const Slice &RHS) const { return !operator==(RHS); }
219 };
220 
221 } // end anonymous namespace
222 
223 namespace llvm {
224 
225 template <typename T> struct isPodLike;
226 template <> struct isPodLike<Slice> { static const bool value = true; };
227 
228 } // end namespace llvm
229 
230 /// \brief Representation of the alloca slices.
231 ///
232 /// This class represents the slices of an alloca which are formed by its
233 /// various uses. If a pointer escapes, we can't fully build a representation
234 /// for the slices used and we reflect that in this structure. The uses are
235 /// stored, sorted by increasing beginning offset and with unsplittable slices
236 /// starting at a particular offset before splittable slices.
238 public:
239  /// \brief Construct the slices of a particular alloca.
240  AllocaSlices(const DataLayout &DL, AllocaInst &AI);
241 
242  /// \brief Test whether a pointer to the allocation escapes our analysis.
243  ///
244  /// If this is true, the slices are never fully built and should be
245  /// ignored.
246  bool isEscaped() const { return PointerEscapingInstr; }
247 
248  /// \brief Support for iterating over the slices.
249  /// @{
252 
253  iterator begin() { return Slices.begin(); }
254  iterator end() { return Slices.end(); }
255 
258 
259  const_iterator begin() const { return Slices.begin(); }
260  const_iterator end() const { return Slices.end(); }
261  /// @}
262 
263  /// \brief Erase a range of slices.
264  void erase(iterator Start, iterator Stop) { Slices.erase(Start, Stop); }
265 
266  /// \brief Insert new slices for this alloca.
267  ///
268  /// This moves the slices into the alloca's slices collection, and re-sorts
269  /// everything so that the usual ordering properties of the alloca's slices
270  /// hold.
271  void insert(ArrayRef<Slice> NewSlices) {
272  int OldSize = Slices.size();
273  Slices.append(NewSlices.begin(), NewSlices.end());
274  auto SliceI = Slices.begin() + OldSize;
275  std::sort(SliceI, Slices.end());
276  std::inplace_merge(Slices.begin(), SliceI, Slices.end());
277  }
278 
279  // Forward declare the iterator and range accessor for walking the
280  // partitions.
281  class partition_iterator;
283 
284  /// \brief Access the dead users for this alloca.
285  ArrayRef<Instruction *> getDeadUsers() const { return DeadUsers; }
286 
287  /// \brief Access the dead operands referring to this alloca.
288  ///
289  /// These are operands which have cannot actually be used to refer to the
290  /// alloca as they are outside its range and the user doesn't correct for
291  /// that. These mostly consist of PHI node inputs and the like which we just
292  /// need to replace with undef.
293  ArrayRef<Use *> getDeadOperands() const { return DeadOperands; }
294 
295 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
296  void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
297  void printSlice(raw_ostream &OS, const_iterator I,
298  StringRef Indent = " ") const;
299  void printUse(raw_ostream &OS, const_iterator I,
300  StringRef Indent = " ") const;
301  void print(raw_ostream &OS) const;
302  void dump(const_iterator I) const;
303  void dump() const;
304 #endif
305 
306 private:
307  template <typename DerivedT, typename RetT = void> class BuilderBase;
308  class SliceBuilder;
309 
310  friend class AllocaSlices::SliceBuilder;
311 
312 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
313  /// \brief Handle to alloca instruction to simplify method interfaces.
314  AllocaInst &AI;
315 #endif
316 
317  /// \brief The instruction responsible for this alloca not having a known set
318  /// of slices.
319  ///
320  /// When an instruction (potentially) escapes the pointer to the alloca, we
321  /// store a pointer to that here and abort trying to form slices of the
322  /// alloca. This will be null if the alloca slices are analyzed successfully.
323  Instruction *PointerEscapingInstr;
324 
325  /// \brief The slices of the alloca.
326  ///
327  /// We store a vector of the slices formed by uses of the alloca here. This
328  /// vector is sorted by increasing begin offset, and then the unsplittable
329  /// slices before the splittable ones. See the Slice inner class for more
330  /// details.
331  SmallVector<Slice, 8> Slices;
332 
333  /// \brief Instructions which will become dead if we rewrite the alloca.
334  ///
335  /// Note that these are not separated by slice. This is because we expect an
336  /// alloca to be completely rewritten or not rewritten at all. If rewritten,
337  /// all these instructions can simply be removed and replaced with undef as
338  /// they come from outside of the allocated space.
340 
341  /// \brief Operands which will become dead if we rewrite the alloca.
342  ///
343  /// These are operands that in their particular use can be replaced with
344  /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
345  /// to PHI nodes and the like. They aren't entirely dead (there might be
346  /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
347  /// want to swap this particular input for undef to simplify the use lists of
348  /// the alloca.
349  SmallVector<Use *, 8> DeadOperands;
350 };
351 
352 /// \brief A partition of the slices.
353 ///
354 /// An ephemeral representation for a range of slices which can be viewed as
355 /// a partition of the alloca. This range represents a span of the alloca's
356 /// memory which cannot be split, and provides access to all of the slices
357 /// overlapping some part of the partition.
358 ///
359 /// Objects of this type are produced by traversing the alloca's slices, but
360 /// are only ephemeral and not persistent.
362 private:
363  friend class AllocaSlices;
365 
366  using iterator = AllocaSlices::iterator;
367 
368  /// \brief The beginning and ending offsets of the alloca for this
369  /// partition.
370  uint64_t BeginOffset, EndOffset;
371 
372  /// \brief The start and end iterators of this partition.
373  iterator SI, SJ;
374 
375  /// \brief A collection of split slice tails overlapping the partition.
376  SmallVector<Slice *, 4> SplitTails;
377 
378  /// \brief Raw constructor builds an empty partition starting and ending at
379  /// the given iterator.
380  Partition(iterator SI) : SI(SI), SJ(SI) {}
381 
382 public:
383  /// \brief The start offset of this partition.
384  ///
385  /// All of the contained slices start at or after this offset.
386  uint64_t beginOffset() const { return BeginOffset; }
387 
388  /// \brief The end offset of this partition.
389  ///
390  /// All of the contained slices end at or before this offset.
391  uint64_t endOffset() const { return EndOffset; }
392 
393  /// \brief The size of the partition.
394  ///
395  /// Note that this can never be zero.
396  uint64_t size() const {
397  assert(BeginOffset < EndOffset && "Partitions must span some bytes!");
398  return EndOffset - BeginOffset;
399  }
400 
401  /// \brief Test whether this partition contains no slices, and merely spans
402  /// a region occupied by split slices.
403  bool empty() const { return SI == SJ; }
404 
405  /// \name Iterate slices that start within the partition.
406  /// These may be splittable or unsplittable. They have a begin offset >= the
407  /// partition begin offset.
408  /// @{
409  // FIXME: We should probably define a "concat_iterator" helper and use that
410  // to stitch together pointee_iterators over the split tails and the
411  // contiguous iterators of the partition. That would give a much nicer
412  // interface here. We could then additionally expose filtered iterators for
413  // split, unsplit, and unsplittable splices based on the usage patterns.
414  iterator begin() const { return SI; }
415  iterator end() const { return SJ; }
416  /// @}
417 
418  /// \brief Get the sequence of split slice tails.
419  ///
420  /// These tails are of slices which start before this partition but are
421  /// split and overlap into the partition. We accumulate these while forming
422  /// partitions.
423  ArrayRef<Slice *> splitSliceTails() const { return SplitTails; }
424 };
425 
426 /// \brief An iterator over partitions of the alloca's slices.
427 ///
428 /// This iterator implements the core algorithm for partitioning the alloca's
429 /// slices. It is a forward iterator as we don't support backtracking for
430 /// efficiency reasons, and re-use a single storage area to maintain the
431 /// current set of split slices.
432 ///
433 /// It is templated on the slice iterator type to use so that it can operate
434 /// with either const or non-const slice iterators.
436  : public iterator_facade_base<partition_iterator, std::forward_iterator_tag,
437  Partition> {
438  friend class AllocaSlices;
439 
440  /// \brief Most of the state for walking the partitions is held in a class
441  /// with a nice interface for examining them.
442  Partition P;
443 
444  /// \brief We need to keep the end of the slices to know when to stop.
446 
447  /// \brief We also need to keep track of the maximum split end offset seen.
448  /// FIXME: Do we really?
449  uint64_t MaxSplitSliceEndOffset = 0;
450 
451  /// \brief Sets the partition to be empty at given iterator, and sets the
452  /// end iterator.
454  : P(SI), SE(SE) {
455  // If not already at the end, advance our state to form the initial
456  // partition.
457  if (SI != SE)
458  advance();
459  }
460 
461  /// \brief Advance the iterator to the next partition.
462  ///
463  /// Requires that the iterator not be at the end of the slices.
464  void advance() {
465  assert((P.SI != SE || !P.SplitTails.empty()) &&
466  "Cannot advance past the end of the slices!");
467 
468  // Clear out any split uses which have ended.
469  if (!P.SplitTails.empty()) {
470  if (P.EndOffset >= MaxSplitSliceEndOffset) {
471  // If we've finished all splits, this is easy.
472  P.SplitTails.clear();
473  MaxSplitSliceEndOffset = 0;
474  } else {
475  // Remove the uses which have ended in the prior partition. This
476  // cannot change the max split slice end because we just checked that
477  // the prior partition ended prior to that max.
478  P.SplitTails.erase(llvm::remove_if(P.SplitTails,
479  [&](Slice *S) {
480  return S->endOffset() <=
481  P.EndOffset;
482  }),
483  P.SplitTails.end());
484  assert(llvm::any_of(P.SplitTails,
485  [&](Slice *S) {
486  return S->endOffset() == MaxSplitSliceEndOffset;
487  }) &&
488  "Could not find the current max split slice offset!");
489  assert(llvm::all_of(P.SplitTails,
490  [&](Slice *S) {
491  return S->endOffset() <= MaxSplitSliceEndOffset;
492  }) &&
493  "Max split slice end offset is not actually the max!");
494  }
495  }
496 
497  // If P.SI is already at the end, then we've cleared the split tail and
498  // now have an end iterator.
499  if (P.SI == SE) {
500  assert(P.SplitTails.empty() && "Failed to clear the split slices!");
501  return;
502  }
503 
504  // If we had a non-empty partition previously, set up the state for
505  // subsequent partitions.
506  if (P.SI != P.SJ) {
507  // Accumulate all the splittable slices which started in the old
508  // partition into the split list.
509  for (Slice &S : P)
510  if (S.isSplittable() && S.endOffset() > P.EndOffset) {
511  P.SplitTails.push_back(&S);
512  MaxSplitSliceEndOffset =
513  std::max(S.endOffset(), MaxSplitSliceEndOffset);
514  }
515 
516  // Start from the end of the previous partition.
517  P.SI = P.SJ;
518 
519  // If P.SI is now at the end, we at most have a tail of split slices.
520  if (P.SI == SE) {
521  P.BeginOffset = P.EndOffset;
522  P.EndOffset = MaxSplitSliceEndOffset;
523  return;
524  }
525 
526  // If the we have split slices and the next slice is after a gap and is
527  // not splittable immediately form an empty partition for the split
528  // slices up until the next slice begins.
529  if (!P.SplitTails.empty() && P.SI->beginOffset() != P.EndOffset &&
530  !P.SI->isSplittable()) {
531  P.BeginOffset = P.EndOffset;
532  P.EndOffset = P.SI->beginOffset();
533  return;
534  }
535  }
536 
537  // OK, we need to consume new slices. Set the end offset based on the
538  // current slice, and step SJ past it. The beginning offset of the
539  // partition is the beginning offset of the next slice unless we have
540  // pre-existing split slices that are continuing, in which case we begin
541  // at the prior end offset.
542  P.BeginOffset = P.SplitTails.empty() ? P.SI->beginOffset() : P.EndOffset;
543  P.EndOffset = P.SI->endOffset();
544  ++P.SJ;
545 
546  // There are two strategies to form a partition based on whether the
547  // partition starts with an unsplittable slice or a splittable slice.
548  if (!P.SI->isSplittable()) {
549  // When we're forming an unsplittable region, it must always start at
550  // the first slice and will extend through its end.
551  assert(P.BeginOffset == P.SI->beginOffset());
552 
553  // Form a partition including all of the overlapping slices with this
554  // unsplittable slice.
555  while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
556  if (!P.SJ->isSplittable())
557  P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
558  ++P.SJ;
559  }
560 
561  // We have a partition across a set of overlapping unsplittable
562  // partitions.
563  return;
564  }
565 
566  // If we're starting with a splittable slice, then we need to form
567  // a synthetic partition spanning it and any other overlapping splittable
568  // splices.
569  assert(P.SI->isSplittable() && "Forming a splittable partition!");
570 
571  // Collect all of the overlapping splittable slices.
572  while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset &&
573  P.SJ->isSplittable()) {
574  P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
575  ++P.SJ;
576  }
577 
578  // Back upiP.EndOffset if we ended the span early when encountering an
579  // unsplittable slice. This synthesizes the early end offset of
580  // a partition spanning only splittable slices.
581  if (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
582  assert(!P.SJ->isSplittable());
583  P.EndOffset = P.SJ->beginOffset();
584  }
585  }
586 
587 public:
588  bool operator==(const partition_iterator &RHS) const {
589  assert(SE == RHS.SE &&
590  "End iterators don't match between compared partition iterators!");
591 
592  // The observed positions of partitions is marked by the P.SI iterator and
593  // the emptiness of the split slices. The latter is only relevant when
594  // P.SI == SE, as the end iterator will additionally have an empty split
595  // slices list, but the prior may have the same P.SI and a tail of split
596  // slices.
597  if (P.SI == RHS.P.SI && P.SplitTails.empty() == RHS.P.SplitTails.empty()) {
598  assert(P.SJ == RHS.P.SJ &&
599  "Same set of slices formed two different sized partitions!");
600  assert(P.SplitTails.size() == RHS.P.SplitTails.size() &&
601  "Same slice position with differently sized non-empty split "
602  "slice tails!");
603  return true;
604  }
605  return false;
606  }
607 
609  advance();
610  return *this;
611  }
612 
613  Partition &operator*() { return P; }
614 };
615 
616 /// \brief A forward range over the partitions of the alloca's slices.
617 ///
618 /// This accesses an iterator range over the partitions of the alloca's
619 /// slices. It computes these partitions on the fly based on the overlapping
620 /// offsets of the slices and the ability to split them. It will visit "empty"
621 /// partitions to cover regions of the alloca only accessed via split
622 /// slices.
624  return make_range(partition_iterator(begin(), end()),
625  partition_iterator(end(), end()));
626 }
627 
629  // If the condition being selected on is a constant or the same value is
630  // being selected between, fold the select. Yes this does (rarely) happen
631  // early on.
632  if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
633  return SI.getOperand(1 + CI->isZero());
634  if (SI.getOperand(1) == SI.getOperand(2))
635  return SI.getOperand(1);
636 
637  return nullptr;
638 }
639 
640 /// \brief A helper that folds a PHI node or a select.
642  if (PHINode *PN = dyn_cast<PHINode>(&I)) {
643  // If PN merges together the same value, return that value.
644  return PN->hasConstantValue();
645  }
646  return foldSelectInst(cast<SelectInst>(I));
647 }
648 
649 /// \brief Builder for the alloca slices.
650 ///
651 /// This class builds a set of alloca slices by recursively visiting the uses
652 /// of an alloca and making a slice for each load and store at each offset.
653 class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
655  friend class InstVisitor<SliceBuilder>;
656 
658 
659  const uint64_t AllocSize;
660  AllocaSlices &AS;
661 
662  SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
664 
665  /// \brief Set to de-duplicate dead instructions found in the use walk.
666  SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
667 
668 public:
671  AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())), AS(AS) {}
672 
673 private:
674  void markAsDead(Instruction &I) {
675  if (VisitedDeadInsts.insert(&I).second)
676  AS.DeadUsers.push_back(&I);
677  }
678 
679  void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
680  bool IsSplittable = false) {
681  // Completely skip uses which have a zero size or start either before or
682  // past the end of the allocation.
683  if (Size == 0 || Offset.uge(AllocSize)) {
684  DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
685  << " which has zero size or starts outside of the "
686  << AllocSize << " byte alloca:\n"
687  << " alloca: " << AS.AI << "\n"
688  << " use: " << I << "\n");
689  return markAsDead(I);
690  }
691 
692  uint64_t BeginOffset = Offset.getZExtValue();
693  uint64_t EndOffset = BeginOffset + Size;
694 
695  // Clamp the end offset to the end of the allocation. Note that this is
696  // formulated to handle even the case where "BeginOffset + Size" overflows.
697  // This may appear superficially to be something we could ignore entirely,
698  // but that is not so! There may be widened loads or PHI-node uses where
699  // some instructions are dead but not others. We can't completely ignore
700  // them, and so have to record at least the information here.
701  assert(AllocSize >= BeginOffset); // Established above.
702  if (Size > AllocSize - BeginOffset) {
703  DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
704  << " to remain within the " << AllocSize << " byte alloca:\n"
705  << " alloca: " << AS.AI << "\n"
706  << " use: " << I << "\n");
707  EndOffset = AllocSize;
708  }
709 
710  AS.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
711  }
712 
713  void visitBitCastInst(BitCastInst &BC) {
714  if (BC.use_empty())
715  return markAsDead(BC);
716 
717  return Base::visitBitCastInst(BC);
718  }
719 
720  void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
721  if (GEPI.use_empty())
722  return markAsDead(GEPI);
723 
724  if (SROAStrictInbounds && GEPI.isInBounds()) {
725  // FIXME: This is a manually un-factored variant of the basic code inside
726  // of GEPs with checking of the inbounds invariant specified in the
727  // langref in a very strict sense. If we ever want to enable
728  // SROAStrictInbounds, this code should be factored cleanly into
729  // PtrUseVisitor, but it is easier to experiment with SROAStrictInbounds
730  // by writing out the code here where we have the underlying allocation
731  // size readily available.
732  APInt GEPOffset = Offset;
733  const DataLayout &DL = GEPI.getModule()->getDataLayout();
734  for (gep_type_iterator GTI = gep_type_begin(GEPI),
735  GTE = gep_type_end(GEPI);
736  GTI != GTE; ++GTI) {
737  ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
738  if (!OpC)
739  break;
740 
741  // Handle a struct index, which adds its field offset to the pointer.
742  if (StructType *STy = GTI.getStructTypeOrNull()) {
743  unsigned ElementIdx = OpC->getZExtValue();
744  const StructLayout *SL = DL.getStructLayout(STy);
745  GEPOffset +=
746  APInt(Offset.getBitWidth(), SL->getElementOffset(ElementIdx));
747  } else {
748  // For array or vector indices, scale the index by the size of the
749  // type.
750  APInt Index = OpC->getValue().sextOrTrunc(Offset.getBitWidth());
751  GEPOffset += Index * APInt(Offset.getBitWidth(),
752  DL.getTypeAllocSize(GTI.getIndexedType()));
753  }
754 
755  // If this index has computed an intermediate pointer which is not
756  // inbounds, then the result of the GEP is a poison value and we can
757  // delete it and all uses.
758  if (GEPOffset.ugt(AllocSize))
759  return markAsDead(GEPI);
760  }
761  }
762 
763  return Base::visitGetElementPtrInst(GEPI);
764  }
765 
766  void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
767  uint64_t Size, bool IsVolatile) {
768  // We allow splitting of non-volatile loads and stores where the type is an
769  // integer type. These may be used to implement 'memcpy' or other "transfer
770  // of bits" patterns.
771  bool IsSplittable = Ty->isIntegerTy() && !IsVolatile;
772 
773  insertUse(I, Offset, Size, IsSplittable);
774  }
775 
776  void visitLoadInst(LoadInst &LI) {
777  assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
778  "All simple FCA loads should have been pre-split");
779 
780  if (!IsOffsetKnown)
781  return PI.setAborted(&LI);
782 
783  const DataLayout &DL = LI.getModule()->getDataLayout();
784  uint64_t Size = DL.getTypeStoreSize(LI.getType());
785  return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
786  }
787 
788  void visitStoreInst(StoreInst &SI) {
789  Value *ValOp = SI.getValueOperand();
790  if (ValOp == *U)
791  return PI.setEscapedAndAborted(&SI);
792  if (!IsOffsetKnown)
793  return PI.setAborted(&SI);
794 
795  const DataLayout &DL = SI.getModule()->getDataLayout();
796  uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
797 
798  // If this memory access can be shown to *statically* extend outside the
799  // bounds of of the allocation, it's behavior is undefined, so simply
800  // ignore it. Note that this is more strict than the generic clamping
801  // behavior of insertUse. We also try to handle cases which might run the
802  // risk of overflow.
803  // FIXME: We should instead consider the pointer to have escaped if this
804  // function is being instrumented for addressing bugs or race conditions.
805  if (Size > AllocSize || Offset.ugt(AllocSize - Size)) {
806  DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @" << Offset
807  << " which extends past the end of the " << AllocSize
808  << " byte alloca:\n"
809  << " alloca: " << AS.AI << "\n"
810  << " use: " << SI << "\n");
811  return markAsDead(SI);
812  }
813 
814  assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
815  "All simple FCA stores should have been pre-split");
816  handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
817  }
818 
819  void visitMemSetInst(MemSetInst &II) {
820  assert(II.getRawDest() == *U && "Pointer use is not the destination?");
821  ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
822  if ((Length && Length->getValue() == 0) ||
823  (IsOffsetKnown && Offset.uge(AllocSize)))
824  // Zero-length mem transfer intrinsics can be ignored entirely.
825  return markAsDead(II);
826 
827  if (!IsOffsetKnown)
828  return PI.setAborted(&II);
829 
830  insertUse(II, Offset, Length ? Length->getLimitedValue()
831  : AllocSize - Offset.getLimitedValue(),
832  (bool)Length);
833  }
834 
835  void visitMemTransferInst(MemTransferInst &II) {
836  ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
837  if (Length && Length->getValue() == 0)
838  // Zero-length mem transfer intrinsics can be ignored entirely.
839  return markAsDead(II);
840 
841  // Because we can visit these intrinsics twice, also check to see if the
842  // first time marked this instruction as dead. If so, skip it.
843  if (VisitedDeadInsts.count(&II))
844  return;
845 
846  if (!IsOffsetKnown)
847  return PI.setAborted(&II);
848 
849  // This side of the transfer is completely out-of-bounds, and so we can
850  // nuke the entire transfer. However, we also need to nuke the other side
851  // if already added to our partitions.
852  // FIXME: Yet another place we really should bypass this when
853  // instrumenting for ASan.
854  if (Offset.uge(AllocSize)) {
856  MemTransferSliceMap.find(&II);
857  if (MTPI != MemTransferSliceMap.end())
858  AS.Slices[MTPI->second].kill();
859  return markAsDead(II);
860  }
861 
862  uint64_t RawOffset = Offset.getLimitedValue();
863  uint64_t Size = Length ? Length->getLimitedValue() : AllocSize - RawOffset;
864 
865  // Check for the special case where the same exact value is used for both
866  // source and dest.
867  if (*U == II.getRawDest() && *U == II.getRawSource()) {
868  // For non-volatile transfers this is a no-op.
869  if (!II.isVolatile())
870  return markAsDead(II);
871 
872  return insertUse(II, Offset, Size, /*IsSplittable=*/false);
873  }
874 
875  // If we have seen both source and destination for a mem transfer, then
876  // they both point to the same alloca.
877  bool Inserted;
879  std::tie(MTPI, Inserted) =
880  MemTransferSliceMap.insert(std::make_pair(&II, AS.Slices.size()));
881  unsigned PrevIdx = MTPI->second;
882  if (!Inserted) {
883  Slice &PrevP = AS.Slices[PrevIdx];
884 
885  // Check if the begin offsets match and this is a non-volatile transfer.
886  // In that case, we can completely elide the transfer.
887  if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
888  PrevP.kill();
889  return markAsDead(II);
890  }
891 
892  // Otherwise we have an offset transfer within the same alloca. We can't
893  // split those.
894  PrevP.makeUnsplittable();
895  }
896 
897  // Insert the use now that we've fixed up the splittable nature.
898  insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
899 
900  // Check that we ended up with a valid index in the map.
901  assert(AS.Slices[PrevIdx].getUse()->getUser() == &II &&
902  "Map index doesn't point back to a slice with this user.");
903  }
904 
905  // Disable SRoA for any intrinsics except for lifetime invariants.
906  // FIXME: What about debug intrinsics? This matches old behavior, but
907  // doesn't make sense.
908  void visitIntrinsicInst(IntrinsicInst &II) {
909  if (!IsOffsetKnown)
910  return PI.setAborted(&II);
911 
912  if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
913  II.getIntrinsicID() == Intrinsic::lifetime_end) {
914  ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
915  uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
916  Length->getLimitedValue());
917  insertUse(II, Offset, Size, true);
918  return;
919  }
920 
921  Base::visitIntrinsicInst(II);
922  }
923 
924  Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
925  // We consider any PHI or select that results in a direct load or store of
926  // the same offset to be a viable use for slicing purposes. These uses
927  // are considered unsplittable and the size is the maximum loaded or stored
928  // size.
931  Visited.insert(Root);
932  Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
933  const DataLayout &DL = Root->getModule()->getDataLayout();
934  // If there are no loads or stores, the access is dead. We mark that as
935  // a size zero access.
936  Size = 0;
937  do {
938  Instruction *I, *UsedI;
939  std::tie(UsedI, I) = Uses.pop_back_val();
940 
941  if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
942  Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
943  continue;
944  }
945  if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
946  Value *Op = SI->getOperand(0);
947  if (Op == UsedI)
948  return SI;
949  Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
950  continue;
951  }
952 
953  if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
954  if (!GEP->hasAllZeroIndices())
955  return GEP;
956  } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
957  !isa<SelectInst>(I)) {
958  return I;
959  }
960 
961  for (User *U : I->users())
962  if (Visited.insert(cast<Instruction>(U)).second)
963  Uses.push_back(std::make_pair(I, cast<Instruction>(U)));
964  } while (!Uses.empty());
965 
966  return nullptr;
967  }
968 
969  void visitPHINodeOrSelectInst(Instruction &I) {
970  assert(isa<PHINode>(I) || isa<SelectInst>(I));
971  if (I.use_empty())
972  return markAsDead(I);
973 
974  // TODO: We could use SimplifyInstruction here to fold PHINodes and
975  // SelectInsts. However, doing so requires to change the current
976  // dead-operand-tracking mechanism. For instance, suppose neither loading
977  // from %U nor %other traps. Then "load (select undef, %U, %other)" does not
978  // trap either. However, if we simply replace %U with undef using the
979  // current dead-operand-tracking mechanism, "load (select undef, undef,
980  // %other)" may trap because the select may return the first operand
981  // "undef".
982  if (Value *Result = foldPHINodeOrSelectInst(I)) {
983  if (Result == *U)
984  // If the result of the constant fold will be the pointer, recurse
985  // through the PHI/select as if we had RAUW'ed it.
986  enqueueUsers(I);
987  else
988  // Otherwise the operand to the PHI/select is dead, and we can replace
989  // it with undef.
990  AS.DeadOperands.push_back(U);
991 
992  return;
993  }
994 
995  if (!IsOffsetKnown)
996  return PI.setAborted(&I);
997 
998  // See if we already have computed info on this node.
999  uint64_t &Size = PHIOrSelectSizes[&I];
1000  if (!Size) {
1001  // This is a new PHI/Select, check for an unsafe use of it.
1002  if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&I, Size))
1003  return PI.setAborted(UnsafeI);
1004  }
1005 
1006  // For PHI and select operands outside the alloca, we can't nuke the entire
1007  // phi or select -- the other side might still be relevant, so we special
1008  // case them here and use a separate structure to track the operands
1009  // themselves which should be replaced with undef.
1010  // FIXME: This should instead be escaped in the event we're instrumenting
1011  // for address sanitization.
1012  if (Offset.uge(AllocSize)) {
1013  AS.DeadOperands.push_back(U);
1014  return;
1015  }
1016 
1017  insertUse(I, Offset, Size);
1018  }
1019 
1020  void visitPHINode(PHINode &PN) { visitPHINodeOrSelectInst(PN); }
1021 
1022  void visitSelectInst(SelectInst &SI) { visitPHINodeOrSelectInst(SI); }
1023 
1024  /// \brief Disable SROA entirely if there are unhandled users of the alloca.
1025  void visitInstruction(Instruction &I) { PI.setAborted(&I); }
1026 };
1027 
1029  :
1030 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1031  AI(AI),
1032 #endif
1033  PointerEscapingInstr(nullptr) {
1034  SliceBuilder PB(DL, AI, *this);
1035  SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
1036  if (PtrI.isEscaped() || PtrI.isAborted()) {
1037  // FIXME: We should sink the escape vs. abort info into the caller nicely,
1038  // possibly by just storing the PtrInfo in the AllocaSlices.
1039  PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
1040  : PtrI.getAbortingInst();
1041  assert(PointerEscapingInstr && "Did not track a bad instruction");
1042  return;
1043  }
1044 
1045  Slices.erase(
1046  llvm::remove_if(Slices, [](const Slice &S) { return S.isDead(); }),
1047  Slices.end());
1048 
1049 #ifndef NDEBUG
1051  std::mt19937 MT(static_cast<unsigned>(
1052  std::chrono::system_clock::now().time_since_epoch().count()));
1053  std::shuffle(Slices.begin(), Slices.end(), MT);
1054  }
1055 #endif
1056 
1057  // Sort the uses. This arranges for the offsets to be in ascending order,
1058  // and the sizes to be in descending order.
1059  std::sort(Slices.begin(), Slices.end());
1060 }
1061 
1062 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1063 
1064 void AllocaSlices::print(raw_ostream &OS, const_iterator I,
1065  StringRef Indent) const {
1066  printSlice(OS, I, Indent);
1067  OS << "\n";
1068  printUse(OS, I, Indent);
1069 }
1070 
1071 void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
1072  StringRef Indent) const {
1073  OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
1074  << " slice #" << (I - begin())
1075  << (I->isSplittable() ? " (splittable)" : "");
1076 }
1077 
1078 void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
1079  StringRef Indent) const {
1080  OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
1081 }
1082 
1083 void AllocaSlices::print(raw_ostream &OS) const {
1084  if (PointerEscapingInstr) {
1085  OS << "Can't analyze slices for alloca: " << AI << "\n"
1086  << " A pointer to this alloca escaped by:\n"
1087  << " " << *PointerEscapingInstr << "\n";
1088  return;
1089  }
1090 
1091  OS << "Slices of alloca: " << AI << "\n";
1092  for (const_iterator I = begin(), E = end(); I != E; ++I)
1093  print(OS, I);
1094 }
1095 
1096 LLVM_DUMP_METHOD void AllocaSlices::dump(const_iterator I) const {
1097  print(dbgs(), I);
1098 }
1099 LLVM_DUMP_METHOD void AllocaSlices::dump() const { print(dbgs()); }
1100 
1101 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1102 
1103 /// Walk the range of a partitioning looking for a common type to cover this
1104 /// sequence of slices.
1107  uint64_t EndOffset) {
1108  Type *Ty = nullptr;
1109  bool TyIsCommon = true;
1110  IntegerType *ITy = nullptr;
1111 
1112  // Note that we need to look at *every* alloca slice's Use to ensure we
1113  // always get consistent results regardless of the order of slices.
1114  for (AllocaSlices::const_iterator I = B; I != E; ++I) {
1115  Use *U = I->getUse();
1116  if (isa<IntrinsicInst>(*U->getUser()))
1117  continue;
1118  if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
1119  continue;
1120 
1121  Type *UserTy = nullptr;
1122  if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1123  UserTy = LI->getType();
1124  } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1125  UserTy = SI->getValueOperand()->getType();
1126  }
1127 
1128  if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
1129  // If the type is larger than the partition, skip it. We only encounter
1130  // this for split integer operations where we want to use the type of the
1131  // entity causing the split. Also skip if the type is not a byte width
1132  // multiple.
1133  if (UserITy->getBitWidth() % 8 != 0 ||
1134  UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
1135  continue;
1136 
1137  // Track the largest bitwidth integer type used in this way in case there
1138  // is no common type.
1139  if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
1140  ITy = UserITy;
1141  }
1142 
1143  // To avoid depending on the order of slices, Ty and TyIsCommon must not
1144  // depend on types skipped above.
1145  if (!UserTy || (Ty && Ty != UserTy))
1146  TyIsCommon = false; // Give up on anything but an iN type.
1147  else
1148  Ty = UserTy;
1149  }
1150 
1151  return TyIsCommon ? Ty : ITy;
1152 }
1153 
1154 /// PHI instructions that use an alloca and are subsequently loaded can be
1155 /// rewritten to load both input pointers in the pred blocks and then PHI the
1156 /// results, allowing the load of the alloca to be promoted.
1157 /// From this:
1158 /// %P2 = phi [i32* %Alloca, i32* %Other]
1159 /// %V = load i32* %P2
1160 /// to:
1161 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1162 /// ...
1163 /// %V2 = load i32* %Other
1164 /// ...
1165 /// %V = phi [i32 %V1, i32 %V2]
1166 ///
1167 /// We can do this to a select if its only uses are loads and if the operands
1168 /// to the select can be loaded unconditionally.
1169 ///
1170 /// FIXME: This should be hoisted into a generic utility, likely in
1171 /// Transforms/Util/Local.h
1172 static bool isSafePHIToSpeculate(PHINode &PN) {
1173  // For now, we can only do this promotion if the load is in the same block
1174  // as the PHI, and if there are no stores between the phi and load.
1175  // TODO: Allow recursive phi users.
1176  // TODO: Allow stores.
1177  BasicBlock *BB = PN.getParent();
1178  unsigned MaxAlign = 0;
1179  bool HaveLoad = false;
1180  for (User *U : PN.users()) {
1181  LoadInst *LI = dyn_cast<LoadInst>(U);
1182  if (!LI || !LI->isSimple())
1183  return false;
1184 
1185  // For now we only allow loads in the same block as the PHI. This is
1186  // a common case that happens when instcombine merges two loads through
1187  // a PHI.
1188  if (LI->getParent() != BB)
1189  return false;
1190 
1191  // Ensure that there are no instructions between the PHI and the load that
1192  // could store.
1193  for (BasicBlock::iterator BBI(PN); &*BBI != LI; ++BBI)
1194  if (BBI->mayWriteToMemory())
1195  return false;
1196 
1197  MaxAlign = std::max(MaxAlign, LI->getAlignment());
1198  HaveLoad = true;
1199  }
1200 
1201  if (!HaveLoad)
1202  return false;
1203 
1204  const DataLayout &DL = PN.getModule()->getDataLayout();
1205 
1206  // We can only transform this if it is safe to push the loads into the
1207  // predecessor blocks. The only thing to watch out for is that we can't put
1208  // a possibly trapping load in the predecessor if it is a critical edge.
1209  for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1211  Value *InVal = PN.getIncomingValue(Idx);
1212 
1213  // If the value is produced by the terminator of the predecessor (an
1214  // invoke) or it has side-effects, there is no valid place to put a load
1215  // in the predecessor.
1216  if (TI == InVal || TI->mayHaveSideEffects())
1217  return false;
1218 
1219  // If the predecessor has a single successor, then the edge isn't
1220  // critical.
1221  if (TI->getNumSuccessors() == 1)
1222  continue;
1223 
1224  // If this pointer is always safe to load, or if we can prove that there
1225  // is already a load in the block, then we can move the load to the pred
1226  // block.
1227  if (isSafeToLoadUnconditionally(InVal, MaxAlign, DL, TI))
1228  continue;
1229 
1230  return false;
1231  }
1232 
1233  return true;
1234 }
1235 
1236 static void speculatePHINodeLoads(PHINode &PN) {
1237  DEBUG(dbgs() << " original: " << PN << "\n");
1238 
1239  Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1240  IRBuilderTy PHIBuilder(&PN);
1241  PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1242  PN.getName() + ".sroa.speculated");
1243 
1244  // Get the AA tags and alignment to use from one of the loads. It doesn't
1245  // matter which one we get and if any differ.
1246  LoadInst *SomeLoad = cast<LoadInst>(PN.user_back());
1247 
1248  AAMDNodes AATags;
1249  SomeLoad->getAAMetadata(AATags);
1250  unsigned Align = SomeLoad->getAlignment();
1251 
1252  // Rewrite all loads of the PN to use the new PHI.
1253  while (!PN.use_empty()) {
1254  LoadInst *LI = cast<LoadInst>(PN.user_back());
1255  LI->replaceAllUsesWith(NewPN);
1256  LI->eraseFromParent();
1257  }
1258 
1259  // Inject loads into all of the pred blocks.
1260  for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1261  BasicBlock *Pred = PN.getIncomingBlock(Idx);
1262  TerminatorInst *TI = Pred->getTerminator();
1263  Value *InVal = PN.getIncomingValue(Idx);
1264  IRBuilderTy PredBuilder(TI);
1265 
1266  LoadInst *Load = PredBuilder.CreateLoad(
1267  InVal, (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1268  ++NumLoadsSpeculated;
1269  Load->setAlignment(Align);
1270  if (AATags)
1271  Load->setAAMetadata(AATags);
1272  NewPN->addIncoming(Load, Pred);
1273  }
1274 
1275  DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1276  PN.eraseFromParent();
1277 }
1278 
1279 /// Select instructions that use an alloca and are subsequently loaded can be
1280 /// rewritten to load both input pointers and then select between the result,
1281 /// allowing the load of the alloca to be promoted.
1282 /// From this:
1283 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1284 /// %V = load i32* %P2
1285 /// to:
1286 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1287 /// %V2 = load i32* %Other
1288 /// %V = select i1 %cond, i32 %V1, i32 %V2
1289 ///
1290 /// We can do this to a select if its only uses are loads and if the operand
1291 /// to the select can be loaded unconditionally.
1293  Value *TValue = SI.getTrueValue();
1294  Value *FValue = SI.getFalseValue();
1295  const DataLayout &DL = SI.getModule()->getDataLayout();
1296 
1297  for (User *U : SI.users()) {
1298  LoadInst *LI = dyn_cast<LoadInst>(U);
1299  if (!LI || !LI->isSimple())
1300  return false;
1301 
1302  // Both operands to the select need to be dereferenceable, either
1303  // absolutely (e.g. allocas) or at this point because we can see other
1304  // accesses to it.
1305  if (!isSafeToLoadUnconditionally(TValue, LI->getAlignment(), DL, LI))
1306  return false;
1307  if (!isSafeToLoadUnconditionally(FValue, LI->getAlignment(), DL, LI))
1308  return false;
1309  }
1310 
1311  return true;
1312 }
1313 
1315  DEBUG(dbgs() << " original: " << SI << "\n");
1316 
1317  IRBuilderTy IRB(&SI);
1318  Value *TV = SI.getTrueValue();
1319  Value *FV = SI.getFalseValue();
1320  // Replace the loads of the select with a select of two loads.
1321  while (!SI.use_empty()) {
1322  LoadInst *LI = cast<LoadInst>(SI.user_back());
1323  assert(LI->isSimple() && "We only speculate simple loads");
1324 
1325  IRB.SetInsertPoint(LI);
1326  LoadInst *TL =
1327  IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1328  LoadInst *FL =
1329  IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1330  NumLoadsSpeculated += 2;
1331 
1332  // Transfer alignment and AA info if present.
1333  TL->setAlignment(LI->getAlignment());
1334  FL->setAlignment(LI->getAlignment());
1335 
1336  AAMDNodes Tags;
1337  LI->getAAMetadata(Tags);
1338  if (Tags) {
1339  TL->setAAMetadata(Tags);
1340  FL->setAAMetadata(Tags);
1341  }
1342 
1343  Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1344  LI->getName() + ".sroa.speculated");
1345 
1346  DEBUG(dbgs() << " speculated to: " << *V << "\n");
1347  LI->replaceAllUsesWith(V);
1348  LI->eraseFromParent();
1349  }
1350  SI.eraseFromParent();
1351 }
1352 
1353 /// \brief Build a GEP out of a base pointer and indices.
1354 ///
1355 /// This will return the BasePtr if that is valid, or build a new GEP
1356 /// instruction using the IRBuilder if GEP-ing is needed.
1357 static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1358  SmallVectorImpl<Value *> &Indices, Twine NamePrefix) {
1359  if (Indices.empty())
1360  return BasePtr;
1361 
1362  // A single zero index is a no-op, so check for this and avoid building a GEP
1363  // in that case.
1364  if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1365  return BasePtr;
1366 
1367  return IRB.CreateInBoundsGEP(nullptr, BasePtr, Indices,
1368  NamePrefix + "sroa_idx");
1369 }
1370 
1371 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1372 /// TargetTy without changing the offset of the pointer.
1373 ///
1374 /// This routine assumes we've already established a properly offset GEP with
1375 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1376 /// zero-indices down through type layers until we find one the same as
1377 /// TargetTy. If we can't find one with the same type, we at least try to use
1378 /// one with the same size. If none of that works, we just produce the GEP as
1379 /// indicated by Indices to have the correct offset.
1380 static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1381  Value *BasePtr, Type *Ty, Type *TargetTy,
1382  SmallVectorImpl<Value *> &Indices,
1383  Twine NamePrefix) {
1384  if (Ty == TargetTy)
1385  return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1386 
1387  // Pointer size to use for the indices.
1388  unsigned PtrSize = DL.getPointerTypeSizeInBits(BasePtr->getType());
1389 
1390  // See if we can descend into a struct and locate a field with the correct
1391  // type.
1392  unsigned NumLayers = 0;
1393  Type *ElementTy = Ty;
1394  do {
1395  if (ElementTy->isPointerTy())
1396  break;
1397 
1398  if (ArrayType *ArrayTy = dyn_cast<ArrayType>(ElementTy)) {
1399  ElementTy = ArrayTy->getElementType();
1400  Indices.push_back(IRB.getIntN(PtrSize, 0));
1401  } else if (VectorType *VectorTy = dyn_cast<VectorType>(ElementTy)) {
1402  ElementTy = VectorTy->getElementType();
1403  Indices.push_back(IRB.getInt32(0));
1404  } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1405  if (STy->element_begin() == STy->element_end())
1406  break; // Nothing left to descend into.
1407  ElementTy = *STy->element_begin();
1408  Indices.push_back(IRB.getInt32(0));
1409  } else {
1410  break;
1411  }
1412  ++NumLayers;
1413  } while (ElementTy != TargetTy);
1414  if (ElementTy != TargetTy)
1415  Indices.erase(Indices.end() - NumLayers, Indices.end());
1416 
1417  return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1418 }
1419 
1420 /// \brief Recursively compute indices for a natural GEP.
1421 ///
1422 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1423 /// element types adding appropriate indices for the GEP.
1424 static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1425  Value *Ptr, Type *Ty, APInt &Offset,
1426  Type *TargetTy,
1427  SmallVectorImpl<Value *> &Indices,
1428  Twine NamePrefix) {
1429  if (Offset == 0)
1430  return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices,
1431  NamePrefix);
1432 
1433  // We can't recurse through pointer types.
1434  if (Ty->isPointerTy())
1435  return nullptr;
1436 
1437  // We try to analyze GEPs over vectors here, but note that these GEPs are
1438  // extremely poorly defined currently. The long-term goal is to remove GEPing
1439  // over a vector from the IR completely.
1440  if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1441  unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
1442  if (ElementSizeInBits % 8 != 0) {
1443  // GEPs over non-multiple of 8 size vector elements are invalid.
1444  return nullptr;
1445  }
1446  APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1447  APInt NumSkippedElements = Offset.sdiv(ElementSize);
1448  if (NumSkippedElements.ugt(VecTy->getNumElements()))
1449  return nullptr;
1450  Offset -= NumSkippedElements * ElementSize;
1451  Indices.push_back(IRB.getInt(NumSkippedElements));
1452  return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1453  Offset, TargetTy, Indices, NamePrefix);
1454  }
1455 
1456  if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1457  Type *ElementTy = ArrTy->getElementType();
1458  APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1459  APInt NumSkippedElements = Offset.sdiv(ElementSize);
1460  if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1461  return nullptr;
1462 
1463  Offset -= NumSkippedElements * ElementSize;
1464  Indices.push_back(IRB.getInt(NumSkippedElements));
1465  return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1466  Indices, NamePrefix);
1467  }
1468 
1469  StructType *STy = dyn_cast<StructType>(Ty);
1470  if (!STy)
1471  return nullptr;
1472 
1473  const StructLayout *SL = DL.getStructLayout(STy);
1474  uint64_t StructOffset = Offset.getZExtValue();
1475  if (StructOffset >= SL->getSizeInBytes())
1476  return nullptr;
1477  unsigned Index = SL->getElementContainingOffset(StructOffset);
1478  Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1479  Type *ElementTy = STy->getElementType(Index);
1480  if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
1481  return nullptr; // The offset points into alignment padding.
1482 
1483  Indices.push_back(IRB.getInt32(Index));
1484  return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1485  Indices, NamePrefix);
1486 }
1487 
1488 /// \brief Get a natural GEP from a base pointer to a particular offset and
1489 /// resulting in a particular type.
1490 ///
1491 /// The goal is to produce a "natural" looking GEP that works with the existing
1492 /// composite types to arrive at the appropriate offset and element type for
1493 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1494 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1495 /// Indices, and setting Ty to the result subtype.
1496 ///
1497 /// If no natural GEP can be constructed, this function returns null.
1498 static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1499  Value *Ptr, APInt Offset, Type *TargetTy,
1500  SmallVectorImpl<Value *> &Indices,
1501  Twine NamePrefix) {
1502  PointerType *Ty = cast<PointerType>(Ptr->getType());
1503 
1504  // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1505  // an i8.
1506  if (Ty == IRB.getInt8PtrTy(Ty->getAddressSpace()) && TargetTy->isIntegerTy(8))
1507  return nullptr;
1508 
1509  Type *ElementTy = Ty->getElementType();
1510  if (!ElementTy->isSized())
1511  return nullptr; // We can't GEP through an unsized element.
1512  APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1513  if (ElementSize == 0)
1514  return nullptr; // Zero-length arrays can't help us build a natural GEP.
1515  APInt NumSkippedElements = Offset.sdiv(ElementSize);
1516 
1517  Offset -= NumSkippedElements * ElementSize;
1518  Indices.push_back(IRB.getInt(NumSkippedElements));
1519  return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1520  Indices, NamePrefix);
1521 }
1522 
1523 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1524 /// resulting pointer has PointerTy.
1525 ///
1526 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1527 /// and produces the pointer type desired. Where it cannot, it will try to use
1528 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1529 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1530 /// bitcast to the type.
1531 ///
1532 /// The strategy for finding the more natural GEPs is to peel off layers of the
1533 /// pointer, walking back through bit casts and GEPs, searching for a base
1534 /// pointer from which we can compute a natural GEP with the desired
1535 /// properties. The algorithm tries to fold as many constant indices into
1536 /// a single GEP as possible, thus making each GEP more independent of the
1537 /// surrounding code.
1538 static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr,
1539  APInt Offset, Type *PointerTy, Twine NamePrefix) {
1540  // Even though we don't look through PHI nodes, we could be called on an
1541  // instruction in an unreachable block, which may be on a cycle.
1542  SmallPtrSet<Value *, 4> Visited;
1543  Visited.insert(Ptr);
1544  SmallVector<Value *, 4> Indices;
1545 
1546  // We may end up computing an offset pointer that has the wrong type. If we
1547  // never are able to compute one directly that has the correct type, we'll
1548  // fall back to it, so keep it and the base it was computed from around here.
1549  Value *OffsetPtr = nullptr;
1550  Value *OffsetBasePtr;
1551 
1552  // Remember any i8 pointer we come across to re-use if we need to do a raw
1553  // byte offset.
1554  Value *Int8Ptr = nullptr;
1555  APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1556 
1557  Type *TargetTy = PointerTy->getPointerElementType();
1558 
1559  do {
1560  // First fold any existing GEPs into the offset.
1561  while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1562  APInt GEPOffset(Offset.getBitWidth(), 0);
1563  if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1564  break;
1565  Offset += GEPOffset;
1566  Ptr = GEP->getPointerOperand();
1567  if (!Visited.insert(Ptr).second)
1568  break;
1569  }
1570 
1571  // See if we can perform a natural GEP here.
1572  Indices.clear();
1573  if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1574  Indices, NamePrefix)) {
1575  // If we have a new natural pointer at the offset, clear out any old
1576  // offset pointer we computed. Unless it is the base pointer or
1577  // a non-instruction, we built a GEP we don't need. Zap it.
1578  if (OffsetPtr && OffsetPtr != OffsetBasePtr)
1579  if (Instruction *I = dyn_cast<Instruction>(OffsetPtr)) {
1580  assert(I->use_empty() && "Built a GEP with uses some how!");
1581  I->eraseFromParent();
1582  }
1583  OffsetPtr = P;
1584  OffsetBasePtr = Ptr;
1585  // If we also found a pointer of the right type, we're done.
1586  if (P->getType() == PointerTy)
1587  return P;
1588  }
1589 
1590  // Stash this pointer if we've found an i8*.
1591  if (Ptr->getType()->isIntegerTy(8)) {
1592  Int8Ptr = Ptr;
1593  Int8PtrOffset = Offset;
1594  }
1595 
1596  // Peel off a layer of the pointer and update the offset appropriately.
1597  if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1598  Ptr = cast<Operator>(Ptr)->getOperand(0);
1599  } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1600  if (GA->isInterposable())
1601  break;
1602  Ptr = GA->getAliasee();
1603  } else {
1604  break;
1605  }
1606  assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1607  } while (Visited.insert(Ptr).second);
1608 
1609  if (!OffsetPtr) {
1610  if (!Int8Ptr) {
1611  Int8Ptr = IRB.CreateBitCast(
1612  Ptr, IRB.getInt8PtrTy(PointerTy->getPointerAddressSpace()),
1613  NamePrefix + "sroa_raw_cast");
1614  Int8PtrOffset = Offset;
1615  }
1616 
1617  OffsetPtr = Int8PtrOffset == 0
1618  ? Int8Ptr
1619  : IRB.CreateInBoundsGEP(IRB.getInt8Ty(), Int8Ptr,
1620  IRB.getInt(Int8PtrOffset),
1621  NamePrefix + "sroa_raw_idx");
1622  }
1623  Ptr = OffsetPtr;
1624 
1625  // On the off chance we were targeting i8*, guard the bitcast here.
1626  if (Ptr->getType() != PointerTy)
1627  Ptr = IRB.CreateBitCast(Ptr, PointerTy, NamePrefix + "sroa_cast");
1628 
1629  return Ptr;
1630 }
1631 
1632 /// \brief Compute the adjusted alignment for a load or store from an offset.
1633 static unsigned getAdjustedAlignment(Instruction *I, uint64_t Offset,
1634  const DataLayout &DL) {
1635  unsigned Alignment;
1636  Type *Ty;
1637  if (auto *LI = dyn_cast<LoadInst>(I)) {
1638  Alignment = LI->getAlignment();
1639  Ty = LI->getType();
1640  } else if (auto *SI = dyn_cast<StoreInst>(I)) {
1641  Alignment = SI->getAlignment();
1642  Ty = SI->getValueOperand()->getType();
1643  } else {
1644  llvm_unreachable("Only loads and stores are allowed!");
1645  }
1646 
1647  if (!Alignment)
1648  Alignment = DL.getABITypeAlignment(Ty);
1649 
1650  return MinAlign(Alignment, Offset);
1651 }
1652 
1653 /// \brief Test whether we can convert a value from the old to the new type.
1654 ///
1655 /// This predicate should be used to guard calls to convertValue in order to
1656 /// ensure that we only try to convert viable values. The strategy is that we
1657 /// will peel off single element struct and array wrappings to get to an
1658 /// underlying value, and convert that value.
1659 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1660  if (OldTy == NewTy)
1661  return true;
1662 
1663  // For integer types, we can't handle any bit-width differences. This would
1664  // break both vector conversions with extension and introduce endianness
1665  // issues when in conjunction with loads and stores.
1666  if (isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) {
1667  assert(cast<IntegerType>(OldTy)->getBitWidth() !=
1668  cast<IntegerType>(NewTy)->getBitWidth() &&
1669  "We can't have the same bitwidth for different int types");
1670  return false;
1671  }
1672 
1673  if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1674  return false;
1675  if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1676  return false;
1677 
1678  // We can convert pointers to integers and vice-versa. Same for vectors
1679  // of pointers and integers.
1680  OldTy = OldTy->getScalarType();
1681  NewTy = NewTy->getScalarType();
1682  if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1683  if (NewTy->isPointerTy() && OldTy->isPointerTy()) {
1684  return cast<PointerType>(NewTy)->getPointerAddressSpace() ==
1685  cast<PointerType>(OldTy)->getPointerAddressSpace();
1686  }
1687 
1688  // We can convert integers to integral pointers, but not to non-integral
1689  // pointers.
1690  if (OldTy->isIntegerTy())
1691  return !DL.isNonIntegralPointerType(NewTy);
1692 
1693  // We can convert integral pointers to integers, but non-integral pointers
1694  // need to remain pointers.
1695  if (!DL.isNonIntegralPointerType(OldTy))
1696  return NewTy->isIntegerTy();
1697 
1698  return false;
1699  }
1700 
1701  return true;
1702 }
1703 
1704 /// \brief Generic routine to convert an SSA value to a value of a different
1705 /// type.
1706 ///
1707 /// This will try various different casting techniques, such as bitcasts,
1708 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1709 /// two types for viability with this routine.
1710 static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1711  Type *NewTy) {
1712  Type *OldTy = V->getType();
1713  assert(canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type");
1714 
1715  if (OldTy == NewTy)
1716  return V;
1717 
1718  assert(!(isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) &&
1719  "Integer types must be the exact same to convert.");
1720 
1721  // See if we need inttoptr for this type pair. A cast involving both scalars
1722  // and vectors requires and additional bitcast.
1723  if (OldTy->isIntOrIntVectorTy() && NewTy->isPtrOrPtrVectorTy()) {
1724  // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1725  if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1726  return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1727  NewTy);
1728 
1729  // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1730  if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1731  return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1732  NewTy);
1733 
1734  return IRB.CreateIntToPtr(V, NewTy);
1735  }
1736 
1737  // See if we need ptrtoint for this type pair. A cast involving both scalars
1738  // and vectors requires and additional bitcast.
1739  if (OldTy->isPtrOrPtrVectorTy() && NewTy->isIntOrIntVectorTy()) {
1740  // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1741  if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1742  return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1743  NewTy);
1744 
1745  // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1746  if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1747  return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1748  NewTy);
1749 
1750  return IRB.CreatePtrToInt(V, NewTy);
1751  }
1752 
1753  return IRB.CreateBitCast(V, NewTy);
1754 }
1755 
1756 /// \brief Test whether the given slice use can be promoted to a vector.
1757 ///
1758 /// This function is called to test each entry in a partition which is slated
1759 /// for a single slice.
1760 static bool isVectorPromotionViableForSlice(Partition &P, const Slice &S,
1761  VectorType *Ty,
1762  uint64_t ElementSize,
1763  const DataLayout &DL) {
1764  // First validate the slice offsets.
1765  uint64_t BeginOffset =
1766  std::max(S.beginOffset(), P.beginOffset()) - P.beginOffset();
1767  uint64_t BeginIndex = BeginOffset / ElementSize;
1768  if (BeginIndex * ElementSize != BeginOffset ||
1769  BeginIndex >= Ty->getNumElements())
1770  return false;
1771  uint64_t EndOffset =
1772  std::min(S.endOffset(), P.endOffset()) - P.beginOffset();
1773  uint64_t EndIndex = EndOffset / ElementSize;
1774  if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
1775  return false;
1776 
1777  assert(EndIndex > BeginIndex && "Empty vector!");
1778  uint64_t NumElements = EndIndex - BeginIndex;
1779  Type *SliceTy = (NumElements == 1)
1780  ? Ty->getElementType()
1781  : VectorType::get(Ty->getElementType(), NumElements);
1782 
1783  Type *SplitIntTy =
1784  Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1785 
1786  Use *U = S.getUse();
1787 
1788  if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1789  if (MI->isVolatile())
1790  return false;
1791  if (!S.isSplittable())
1792  return false; // Skip any unsplittable intrinsics.
1793  } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1794  if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1795  II->getIntrinsicID() != Intrinsic::lifetime_end)
1796  return false;
1797  } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1798  // Disable vector promotion when there are loads or stores of an FCA.
1799  return false;
1800  } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1801  if (LI->isVolatile())
1802  return false;
1803  Type *LTy = LI->getType();
1804  if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1805  assert(LTy->isIntegerTy());
1806  LTy = SplitIntTy;
1807  }
1808  if (!canConvertValue(DL, SliceTy, LTy))
1809  return false;
1810  } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1811  if (SI->isVolatile())
1812  return false;
1813  Type *STy = SI->getValueOperand()->getType();
1814  if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1815  assert(STy->isIntegerTy());
1816  STy = SplitIntTy;
1817  }
1818  if (!canConvertValue(DL, STy, SliceTy))
1819  return false;
1820  } else {
1821  return false;
1822  }
1823 
1824  return true;
1825 }
1826 
1827 /// \brief Test whether the given alloca partitioning and range of slices can be
1828 /// promoted to a vector.
1829 ///
1830 /// This is a quick test to check whether we can rewrite a particular alloca
1831 /// partition (and its newly formed alloca) into a vector alloca with only
1832 /// whole-vector loads and stores such that it could be promoted to a vector
1833 /// SSA value. We only can ensure this for a limited set of operations, and we
1834 /// don't want to do the rewrites unless we are confident that the result will
1835 /// be promotable, so we have an early test here.
1837  // Collect the candidate types for vector-based promotion. Also track whether
1838  // we have different element types.
1839  SmallVector<VectorType *, 4> CandidateTys;
1840  Type *CommonEltTy = nullptr;
1841  bool HaveCommonEltTy = true;
1842  auto CheckCandidateType = [&](Type *Ty) {
1843  if (auto *VTy = dyn_cast<VectorType>(Ty)) {
1844  CandidateTys.push_back(VTy);
1845  if (!CommonEltTy)
1846  CommonEltTy = VTy->getElementType();
1847  else if (CommonEltTy != VTy->getElementType())
1848  HaveCommonEltTy = false;
1849  }
1850  };
1851  // Consider any loads or stores that are the exact size of the slice.
1852  for (const Slice &S : P)
1853  if (S.beginOffset() == P.beginOffset() &&
1854  S.endOffset() == P.endOffset()) {
1855  if (auto *LI = dyn_cast<LoadInst>(S.getUse()->getUser()))
1856  CheckCandidateType(LI->getType());
1857  else if (auto *SI = dyn_cast<StoreInst>(S.getUse()->getUser()))
1858  CheckCandidateType(SI->getValueOperand()->getType());
1859  }
1860 
1861  // If we didn't find a vector type, nothing to do here.
1862  if (CandidateTys.empty())
1863  return nullptr;
1864 
1865  // Remove non-integer vector types if we had multiple common element types.
1866  // FIXME: It'd be nice to replace them with integer vector types, but we can't
1867  // do that until all the backends are known to produce good code for all
1868  // integer vector types.
1869  if (!HaveCommonEltTy) {
1870  CandidateTys.erase(
1871  llvm::remove_if(CandidateTys,
1872  [](VectorType *VTy) {
1873  return !VTy->getElementType()->isIntegerTy();
1874  }),
1875  CandidateTys.end());
1876 
1877  // If there were no integer vector types, give up.
1878  if (CandidateTys.empty())
1879  return nullptr;
1880 
1881  // Rank the remaining candidate vector types. This is easy because we know
1882  // they're all integer vectors. We sort by ascending number of elements.
1883  auto RankVectorTypes = [&DL](VectorType *RHSTy, VectorType *LHSTy) {
1884  (void)DL;
1885  assert(DL.getTypeSizeInBits(RHSTy) == DL.getTypeSizeInBits(LHSTy) &&
1886  "Cannot have vector types of different sizes!");
1887  assert(RHSTy->getElementType()->isIntegerTy() &&
1888  "All non-integer types eliminated!");
1889  assert(LHSTy->getElementType()->isIntegerTy() &&
1890  "All non-integer types eliminated!");
1891  return RHSTy->getNumElements() < LHSTy->getNumElements();
1892  };
1893  std::sort(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes);
1894  CandidateTys.erase(
1895  std::unique(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes),
1896  CandidateTys.end());
1897  } else {
1898 // The only way to have the same element type in every vector type is to
1899 // have the same vector type. Check that and remove all but one.
1900 #ifndef NDEBUG
1901  for (VectorType *VTy : CandidateTys) {
1902  assert(VTy->getElementType() == CommonEltTy &&
1903  "Unaccounted for element type!");
1904  assert(VTy == CandidateTys[0] &&
1905  "Different vector types with the same element type!");
1906  }
1907 #endif
1908  CandidateTys.resize(1);
1909  }
1910 
1911  // Try each vector type, and return the one which works.
1912  auto CheckVectorTypeForPromotion = [&](VectorType *VTy) {
1913  uint64_t ElementSize = DL.getTypeSizeInBits(VTy->getElementType());
1914 
1915  // While the definition of LLVM vectors is bitpacked, we don't support sizes
1916  // that aren't byte sized.
1917  if (ElementSize % 8)
1918  return false;
1919  assert((DL.getTypeSizeInBits(VTy) % 8) == 0 &&
1920  "vector size not a multiple of element size?");
1921  ElementSize /= 8;
1922 
1923  for (const Slice &S : P)
1924  if (!isVectorPromotionViableForSlice(P, S, VTy, ElementSize, DL))
1925  return false;
1926 
1927  for (const Slice *S : P.splitSliceTails())
1928  if (!isVectorPromotionViableForSlice(P, *S, VTy, ElementSize, DL))
1929  return false;
1930 
1931  return true;
1932  };
1933  for (VectorType *VTy : CandidateTys)
1934  if (CheckVectorTypeForPromotion(VTy))
1935  return VTy;
1936 
1937  return nullptr;
1938 }
1939 
1940 /// \brief Test whether a slice of an alloca is valid for integer widening.
1941 ///
1942 /// This implements the necessary checking for the \c isIntegerWideningViable
1943 /// test below on a single slice of the alloca.
1944 static bool isIntegerWideningViableForSlice(const Slice &S,
1945  uint64_t AllocBeginOffset,
1946  Type *AllocaTy,
1947  const DataLayout &DL,
1948  bool &WholeAllocaOp) {
1949  uint64_t Size = DL.getTypeStoreSize(AllocaTy);
1950 
1951  uint64_t RelBegin = S.beginOffset() - AllocBeginOffset;
1952  uint64_t RelEnd = S.endOffset() - AllocBeginOffset;
1953 
1954  // We can't reasonably handle cases where the load or store extends past
1955  // the end of the alloca's type and into its padding.
1956  if (RelEnd > Size)
1957  return false;
1958 
1959  Use *U = S.getUse();
1960 
1961  if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1962  if (LI->isVolatile())
1963  return false;
1964  // We can't handle loads that extend past the allocated memory.
1965  if (DL.getTypeStoreSize(LI->getType()) > Size)
1966  return false;
1967  // Note that we don't count vector loads or stores as whole-alloca
1968  // operations which enable integer widening because we would prefer to use
1969  // vector widening instead.
1970  if (!isa<VectorType>(LI->getType()) && RelBegin == 0 && RelEnd == Size)
1971  WholeAllocaOp = true;
1972  if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
1973  if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1974  return false;
1975  } else if (RelBegin != 0 || RelEnd != Size ||
1976  !canConvertValue(DL, AllocaTy, LI->getType())) {
1977  // Non-integer loads need to be convertible from the alloca type so that
1978  // they are promotable.
1979  return false;
1980  }
1981  } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1982  Type *ValueTy = SI->getValueOperand()->getType();
1983  if (SI->isVolatile())
1984  return false;
1985  // We can't handle stores that extend past the allocated memory.
1986  if (DL.getTypeStoreSize(ValueTy) > Size)
1987  return false;
1988  // Note that we don't count vector loads or stores as whole-alloca
1989  // operations which enable integer widening because we would prefer to use
1990  // vector widening instead.
1991  if (!isa<VectorType>(ValueTy) && RelBegin == 0 && RelEnd == Size)
1992  WholeAllocaOp = true;
1993  if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
1994  if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1995  return false;
1996  } else if (RelBegin != 0 || RelEnd != Size ||
1997  !canConvertValue(DL, ValueTy, AllocaTy)) {
1998  // Non-integer stores need to be convertible to the alloca type so that
1999  // they are promotable.
2000  return false;
2001  }
2002  } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
2003  if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
2004  return false;
2005  if (!S.isSplittable())
2006  return false; // Skip any unsplittable intrinsics.
2007  } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
2008  if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
2009  II->getIntrinsicID() != Intrinsic::lifetime_end)
2010  return false;
2011  } else {
2012  return false;
2013  }
2014 
2015  return true;
2016 }
2017 
2018 /// \brief Test whether the given alloca partition's integer operations can be
2019 /// widened to promotable ones.
2020 ///
2021 /// This is a quick test to check whether we can rewrite the integer loads and
2022 /// stores to a particular alloca into wider loads and stores and be able to
2023 /// promote the resulting alloca.
2024 static bool isIntegerWideningViable(Partition &P, Type *AllocaTy,
2025  const DataLayout &DL) {
2026  uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
2027  // Don't create integer types larger than the maximum bitwidth.
2028  if (SizeInBits > IntegerType::MAX_INT_BITS)
2029  return false;
2030 
2031  // Don't try to handle allocas with bit-padding.
2032  if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
2033  return false;
2034 
2035  // We need to ensure that an integer type with the appropriate bitwidth can
2036  // be converted to the alloca type, whatever that is. We don't want to force
2037  // the alloca itself to have an integer type if there is a more suitable one.
2038  Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
2039  if (!canConvertValue(DL, AllocaTy, IntTy) ||
2040  !canConvertValue(DL, IntTy, AllocaTy))
2041  return false;
2042 
2043  // While examining uses, we ensure that the alloca has a covering load or
2044  // store. We don't want to widen the integer operations only to fail to
2045  // promote due to some other unsplittable entry (which we may make splittable
2046  // later). However, if there are only splittable uses, go ahead and assume
2047  // that we cover the alloca.
2048  // FIXME: We shouldn't consider split slices that happen to start in the
2049  // partition here...
2050  bool WholeAllocaOp =
2051  P.begin() != P.end() ? false : DL.isLegalInteger(SizeInBits);
2052 
2053  for (const Slice &S : P)
2054  if (!isIntegerWideningViableForSlice(S, P.beginOffset(), AllocaTy, DL,
2055  WholeAllocaOp))
2056  return false;
2057 
2058  for (const Slice *S : P.splitSliceTails())
2059  if (!isIntegerWideningViableForSlice(*S, P.beginOffset(), AllocaTy, DL,
2060  WholeAllocaOp))
2061  return false;
2062 
2063  return WholeAllocaOp;
2064 }
2065 
2066 static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
2067  IntegerType *Ty, uint64_t Offset,
2068  const Twine &Name) {
2069  DEBUG(dbgs() << " start: " << *V << "\n");
2070  IntegerType *IntTy = cast<IntegerType>(V->getType());
2071  assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2072  "Element extends past full value");
2073  uint64_t ShAmt = 8 * Offset;
2074  if (DL.isBigEndian())
2075  ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2076  if (ShAmt) {
2077  V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
2078  DEBUG(dbgs() << " shifted: " << *V << "\n");
2079  }
2080  assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2081  "Cannot extract to a larger integer!");
2082  if (Ty != IntTy) {
2083  V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
2084  DEBUG(dbgs() << " trunced: " << *V << "\n");
2085  }
2086  return V;
2087 }
2088 
2089 static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
2090  Value *V, uint64_t Offset, const Twine &Name) {
2091  IntegerType *IntTy = cast<IntegerType>(Old->getType());
2092  IntegerType *Ty = cast<IntegerType>(V->getType());
2093  assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2094  "Cannot insert a larger integer!");
2095  DEBUG(dbgs() << " start: " << *V << "\n");
2096  if (Ty != IntTy) {
2097  V = IRB.CreateZExt(V, IntTy, Name + ".ext");
2098  DEBUG(dbgs() << " extended: " << *V << "\n");
2099  }
2100  assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2101  "Element store outside of alloca store");
2102  uint64_t ShAmt = 8 * Offset;
2103  if (DL.isBigEndian())
2104  ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2105  if (ShAmt) {
2106  V = IRB.CreateShl(V, ShAmt, Name + ".shift");
2107  DEBUG(dbgs() << " shifted: " << *V << "\n");
2108  }
2109 
2110  if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
2111  APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
2112  Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
2113  DEBUG(dbgs() << " masked: " << *Old << "\n");
2114  V = IRB.CreateOr(Old, V, Name + ".insert");
2115  DEBUG(dbgs() << " inserted: " << *V << "\n");
2116  }
2117  return V;
2118 }
2119 
2120 static Value *extractVector(IRBuilderTy &IRB, Value *V, unsigned BeginIndex,
2121  unsigned EndIndex, const Twine &Name) {
2122  VectorType *VecTy = cast<VectorType>(V->getType());
2123  unsigned NumElements = EndIndex - BeginIndex;
2124  assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2125 
2126  if (NumElements == VecTy->getNumElements())
2127  return V;
2128 
2129  if (NumElements == 1) {
2130  V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
2131  Name + ".extract");
2132  DEBUG(dbgs() << " extract: " << *V << "\n");
2133  return V;
2134  }
2135 
2137  Mask.reserve(NumElements);
2138  for (unsigned i = BeginIndex; i != EndIndex; ++i)
2139  Mask.push_back(IRB.getInt32(i));
2140  V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2141  ConstantVector::get(Mask), Name + ".extract");
2142  DEBUG(dbgs() << " shuffle: " << *V << "\n");
2143  return V;
2144 }
2145 
2146 static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
2147  unsigned BeginIndex, const Twine &Name) {
2148  VectorType *VecTy = cast<VectorType>(Old->getType());
2149  assert(VecTy && "Can only insert a vector into a vector");
2150 
2151  VectorType *Ty = dyn_cast<VectorType>(V->getType());
2152  if (!Ty) {
2153  // Single element to insert.
2154  V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
2155  Name + ".insert");
2156  DEBUG(dbgs() << " insert: " << *V << "\n");
2157  return V;
2158  }
2159 
2160  assert(Ty->getNumElements() <= VecTy->getNumElements() &&
2161  "Too many elements!");
2162  if (Ty->getNumElements() == VecTy->getNumElements()) {
2163  assert(V->getType() == VecTy && "Vector type mismatch");
2164  return V;
2165  }
2166  unsigned EndIndex = BeginIndex + Ty->getNumElements();
2167 
2168  // When inserting a smaller vector into the larger to store, we first
2169  // use a shuffle vector to widen it with undef elements, and then
2170  // a second shuffle vector to select between the loaded vector and the
2171  // incoming vector.
2173  Mask.reserve(VecTy->getNumElements());
2174  for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2175  if (i >= BeginIndex && i < EndIndex)
2176  Mask.push_back(IRB.getInt32(i - BeginIndex));
2177  else
2178  Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
2179  V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2180  ConstantVector::get(Mask), Name + ".expand");
2181  DEBUG(dbgs() << " shuffle: " << *V << "\n");
2182 
2183  Mask.clear();
2184  for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2185  Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
2186 
2187  V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
2188 
2189  DEBUG(dbgs() << " blend: " << *V << "\n");
2190  return V;
2191 }
2192 
2193 /// \brief Visitor to rewrite instructions using p particular slice of an alloca
2194 /// to use a new alloca.
2195 ///
2196 /// Also implements the rewriting to vector-based accesses when the partition
2197 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2198 /// lives here.
2200  : public InstVisitor<AllocaSliceRewriter, bool> {
2201  // Befriend the base class so it can delegate to private visit methods.
2203 
2205 
2206  const DataLayout &DL;
2207  AllocaSlices &AS;
2208  SROA &Pass;
2209  AllocaInst &OldAI, &NewAI;
2210  const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2211  Type *NewAllocaTy;
2212 
2213  // This is a convenience and flag variable that will be null unless the new
2214  // alloca's integer operations should be widened to this integer type due to
2215  // passing isIntegerWideningViable above. If it is non-null, the desired
2216  // integer type will be stored here for easy access during rewriting.
2217  IntegerType *IntTy;
2218 
2219  // If we are rewriting an alloca partition which can be written as pure
2220  // vector operations, we stash extra information here. When VecTy is
2221  // non-null, we have some strict guarantees about the rewritten alloca:
2222  // - The new alloca is exactly the size of the vector type here.
2223  // - The accesses all either map to the entire vector or to a single
2224  // element.
2225  // - The set of accessing instructions is only one of those handled above
2226  // in isVectorPromotionViable. Generally these are the same access kinds
2227  // which are promotable via mem2reg.
2228  VectorType *VecTy;
2229  Type *ElementTy;
2230  uint64_t ElementSize;
2231 
2232  // The original offset of the slice currently being rewritten relative to
2233  // the original alloca.
2234  uint64_t BeginOffset = 0;
2235  uint64_t EndOffset = 0;
2236 
2237  // The new offsets of the slice currently being rewritten relative to the
2238  // original alloca.
2239  uint64_t NewBeginOffset, NewEndOffset;
2240 
2241  uint64_t SliceSize;
2242  bool IsSplittable = false;
2243  bool IsSplit = false;
2244  Use *OldUse = nullptr;
2245  Instruction *OldPtr = nullptr;
2246 
2247  // Track post-rewrite users which are PHI nodes and Selects.
2248  SmallSetVector<PHINode *, 8> &PHIUsers;
2249  SmallSetVector<SelectInst *, 8> &SelectUsers;
2250 
2251  // Utility IR builder, whose name prefix is setup for each visited use, and
2252  // the insertion point is set to point to the user.
2253  IRBuilderTy IRB;
2254 
2255 public:
2257  AllocaInst &OldAI, AllocaInst &NewAI,
2258  uint64_t NewAllocaBeginOffset,
2259  uint64_t NewAllocaEndOffset, bool IsIntegerPromotable,
2260  VectorType *PromotableVecTy,
2261  SmallSetVector<PHINode *, 8> &PHIUsers,
2262  SmallSetVector<SelectInst *, 8> &SelectUsers)
2263  : DL(DL), AS(AS), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
2264  NewAllocaBeginOffset(NewAllocaBeginOffset),
2265  NewAllocaEndOffset(NewAllocaEndOffset),
2266  NewAllocaTy(NewAI.getAllocatedType()),
2267  IntTy(IsIntegerPromotable
2268  ? Type::getIntNTy(
2269  NewAI.getContext(),
2270  DL.getTypeSizeInBits(NewAI.getAllocatedType()))
2271  : nullptr),
2272  VecTy(PromotableVecTy),
2273  ElementTy(VecTy ? VecTy->getElementType() : nullptr),
2274  ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
2275  PHIUsers(PHIUsers), SelectUsers(SelectUsers),
2276  IRB(NewAI.getContext(), ConstantFolder()) {
2277  if (VecTy) {
2278  assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
2279  "Only multiple-of-8 sized vector elements are viable");
2280  ++NumVectorized;
2281  }
2282  assert((!IntTy && !VecTy) || (IntTy && !VecTy) || (!IntTy && VecTy));
2283  }
2284 
2286  bool CanSROA = true;
2287  BeginOffset = I->beginOffset();
2288  EndOffset = I->endOffset();
2289  IsSplittable = I->isSplittable();
2290  IsSplit =
2291  BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
2292  DEBUG(dbgs() << " rewriting " << (IsSplit ? "split " : ""));
2293  DEBUG(AS.printSlice(dbgs(), I, ""));
2294  DEBUG(dbgs() << "\n");
2295 
2296  // Compute the intersecting offset range.
2297  assert(BeginOffset < NewAllocaEndOffset);
2298  assert(EndOffset > NewAllocaBeginOffset);
2299  NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2300  NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2301 
2302  SliceSize = NewEndOffset - NewBeginOffset;
2303 
2304  OldUse = I->getUse();
2305  OldPtr = cast<Instruction>(OldUse->get());
2306 
2307  Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2308  IRB.SetInsertPoint(OldUserI);
2309  IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2310  IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
2311 
2312  CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2313  if (VecTy || IntTy)
2314  assert(CanSROA);
2315  return CanSROA;
2316  }
2317 
2318 private:
2319  // Make sure the other visit overloads are visible.
2320  using Base::visit;
2321 
2322  // Every instruction which can end up as a user must have a rewrite rule.
2323  bool visitInstruction(Instruction &I) {
2324  DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2325  llvm_unreachable("No rewrite rule for this instruction!");
2326  }
2327 
2328  Value *getNewAllocaSlicePtr(IRBuilderTy &IRB, Type *PointerTy) {
2329  // Note that the offset computation can use BeginOffset or NewBeginOffset
2330  // interchangeably for unsplit slices.
2331  assert(IsSplit || BeginOffset == NewBeginOffset);
2332  uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2333 
2334 #ifndef NDEBUG
2335  StringRef OldName = OldPtr->getName();
2336  // Skip through the last '.sroa.' component of the name.
2337  size_t LastSROAPrefix = OldName.rfind(".sroa.");
2338  if (LastSROAPrefix != StringRef::npos) {
2339  OldName = OldName.substr(LastSROAPrefix + strlen(".sroa."));
2340  // Look for an SROA slice index.
2341  size_t IndexEnd = OldName.find_first_not_of("0123456789");
2342  if (IndexEnd != StringRef::npos && OldName[IndexEnd] == '.') {
2343  // Strip the index and look for the offset.
2344  OldName = OldName.substr(IndexEnd + 1);
2345  size_t OffsetEnd = OldName.find_first_not_of("0123456789");
2346  if (OffsetEnd != StringRef::npos && OldName[OffsetEnd] == '.')
2347  // Strip the offset.
2348  OldName = OldName.substr(OffsetEnd + 1);
2349  }
2350  }
2351  // Strip any SROA suffixes as well.
2352  OldName = OldName.substr(0, OldName.find(".sroa_"));
2353 #endif
2354 
2355  return getAdjustedPtr(IRB, DL, &NewAI,
2356  APInt(DL.getPointerTypeSizeInBits(PointerTy), Offset),
2357  PointerTy,
2358 #ifndef NDEBUG
2359  Twine(OldName) + "."
2360 #else
2361  Twine()
2362 #endif
2363  );
2364  }
2365 
2366  /// \brief Compute suitable alignment to access this slice of the *new*
2367  /// alloca.
2368  ///
2369  /// You can optionally pass a type to this routine and if that type's ABI
2370  /// alignment is itself suitable, this will return zero.
2371  unsigned getSliceAlign(Type *Ty = nullptr) {
2372  unsigned NewAIAlign = NewAI.getAlignment();
2373  if (!NewAIAlign)
2374  NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
2375  unsigned Align =
2376  MinAlign(NewAIAlign, NewBeginOffset - NewAllocaBeginOffset);
2377  return (Ty && Align == DL.getABITypeAlignment(Ty)) ? 0 : Align;
2378  }
2379 
2380  unsigned getIndex(uint64_t Offset) {
2381  assert(VecTy && "Can only call getIndex when rewriting a vector");
2382  uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2383  assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2384  uint32_t Index = RelOffset / ElementSize;
2385  assert(Index * ElementSize == RelOffset);
2386  return Index;
2387  }
2388 
2389  void deleteIfTriviallyDead(Value *V) {
2390  Instruction *I = cast<Instruction>(V);
2392  Pass.DeadInsts.insert(I);
2393  }
2394 
2395  Value *rewriteVectorizedLoadInst() {
2396  unsigned BeginIndex = getIndex(NewBeginOffset);
2397  unsigned EndIndex = getIndex(NewEndOffset);
2398  assert(EndIndex > BeginIndex && "Empty vector!");
2399 
2400  Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2401  return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2402  }
2403 
2404  Value *rewriteIntegerLoad(LoadInst &LI) {
2405  assert(IntTy && "We cannot insert an integer to the alloca");
2406  assert(!LI.isVolatile());
2407  Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2408  V = convertValue(DL, IRB, V, IntTy);
2409  assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2410  uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2411  if (Offset > 0 || NewEndOffset < NewAllocaEndOffset) {
2412  IntegerType *ExtractTy = Type::getIntNTy(LI.getContext(), SliceSize * 8);
2413  V = extractInteger(DL, IRB, V, ExtractTy, Offset, "extract");
2414  }
2415  // It is possible that the extracted type is not the load type. This
2416  // happens if there is a load past the end of the alloca, and as
2417  // a consequence the slice is narrower but still a candidate for integer
2418  // lowering. To handle this case, we just zero extend the extracted
2419  // integer.
2420  assert(cast<IntegerType>(LI.getType())->getBitWidth() >= SliceSize * 8 &&
2421  "Can only handle an extract for an overly wide load");
2422  if (cast<IntegerType>(LI.getType())->getBitWidth() > SliceSize * 8)
2423  V = IRB.CreateZExt(V, LI.getType());
2424  return V;
2425  }
2426 
2427  bool visitLoadInst(LoadInst &LI) {
2428  DEBUG(dbgs() << " original: " << LI << "\n");
2429  Value *OldOp = LI.getOperand(0);
2430  assert(OldOp == OldPtr);
2431 
2432  unsigned AS = LI.getPointerAddressSpace();
2433 
2434  Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), SliceSize * 8)
2435  : LI.getType();
2436  const bool IsLoadPastEnd = DL.getTypeStoreSize(TargetTy) > SliceSize;
2437  bool IsPtrAdjusted = false;
2438  Value *V;
2439  if (VecTy) {
2440  V = rewriteVectorizedLoadInst();
2441  } else if (IntTy && LI.getType()->isIntegerTy()) {
2442  V = rewriteIntegerLoad(LI);
2443  } else if (NewBeginOffset == NewAllocaBeginOffset &&
2444  NewEndOffset == NewAllocaEndOffset &&
2445  (canConvertValue(DL, NewAllocaTy, TargetTy) ||
2446  (IsLoadPastEnd && NewAllocaTy->isIntegerTy() &&
2447  TargetTy->isIntegerTy()))) {
2448  LoadInst *NewLI = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2449  LI.isVolatile(), LI.getName());
2450  if (LI.isVolatile())
2451  NewLI->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
2452 
2453  // Any !nonnull metadata or !range metadata on the old load is also valid
2454  // on the new load. This is even true in some cases even when the loads
2455  // are different types, for example by mapping !nonnull metadata to
2456  // !range metadata by modeling the null pointer constant converted to the
2457  // integer type.
2458  // FIXME: Add support for range metadata here. Currently the utilities
2459  // for this don't propagate range metadata in trivial cases from one
2460  // integer load to another, don't handle non-addrspace-0 null pointers
2461  // correctly, and don't have any support for mapping ranges as the
2462  // integer type becomes winder or narrower.
2464  copyNonnullMetadata(LI, N, *NewLI);
2465 
2466  // Try to preserve nonnull metadata
2467  V = NewLI;
2468 
2469  // If this is an integer load past the end of the slice (which means the
2470  // bytes outside the slice are undef or this load is dead) just forcibly
2471  // fix the integer size with correct handling of endianness.
2472  if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2473  if (auto *TITy = dyn_cast<IntegerType>(TargetTy))
2474  if (AITy->getBitWidth() < TITy->getBitWidth()) {
2475  V = IRB.CreateZExt(V, TITy, "load.ext");
2476  if (DL.isBigEndian())
2477  V = IRB.CreateShl(V, TITy->getBitWidth() - AITy->getBitWidth(),
2478  "endian_shift");
2479  }
2480  } else {
2481  Type *LTy = TargetTy->getPointerTo(AS);
2482  LoadInst *NewLI = IRB.CreateAlignedLoad(getNewAllocaSlicePtr(IRB, LTy),
2483  getSliceAlign(TargetTy),
2484  LI.isVolatile(), LI.getName());
2485  if (LI.isVolatile())
2486  NewLI->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
2487 
2488  V = NewLI;
2489  IsPtrAdjusted = true;
2490  }
2491  V = convertValue(DL, IRB, V, TargetTy);
2492 
2493  if (IsSplit) {
2494  assert(!LI.isVolatile());
2495  assert(LI.getType()->isIntegerTy() &&
2496  "Only integer type loads and stores are split");
2497  assert(SliceSize < DL.getTypeStoreSize(LI.getType()) &&
2498  "Split load isn't smaller than original load");
2499  assert(LI.getType()->getIntegerBitWidth() ==
2500  DL.getTypeStoreSizeInBits(LI.getType()) &&
2501  "Non-byte-multiple bit width");
2502  // Move the insertion point just past the load so that we can refer to it.
2503  IRB.SetInsertPoint(&*std::next(BasicBlock::iterator(&LI)));
2504  // Create a placeholder value with the same type as LI to use as the
2505  // basis for the new value. This allows us to replace the uses of LI with
2506  // the computed value, and then replace the placeholder with LI, leaving
2507  // LI only used for this computation.
2508  Value *Placeholder =
2509  new LoadInst(UndefValue::get(LI.getType()->getPointerTo(AS)));
2510  V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset - BeginOffset,
2511  "insert");
2512  LI.replaceAllUsesWith(V);
2513  Placeholder->replaceAllUsesWith(&LI);
2514  Placeholder->deleteValue();
2515  } else {
2516  LI.replaceAllUsesWith(V);
2517  }
2518 
2519  Pass.DeadInsts.insert(&LI);
2520  deleteIfTriviallyDead(OldOp);
2521  DEBUG(dbgs() << " to: " << *V << "\n");
2522  return !LI.isVolatile() && !IsPtrAdjusted;
2523  }
2524 
2525  bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp) {
2526  if (V->getType() != VecTy) {
2527  unsigned BeginIndex = getIndex(NewBeginOffset);
2528  unsigned EndIndex = getIndex(NewEndOffset);
2529  assert(EndIndex > BeginIndex && "Empty vector!");
2530  unsigned NumElements = EndIndex - BeginIndex;
2531  assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2532  Type *SliceTy = (NumElements == 1)
2533  ? ElementTy
2534  : VectorType::get(ElementTy, NumElements);
2535  if (V->getType() != SliceTy)
2536  V = convertValue(DL, IRB, V, SliceTy);
2537 
2538  // Mix in the existing elements.
2539  Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2540  V = insertVector(IRB, Old, V, BeginIndex, "vec");
2541  }
2542  StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2543  Pass.DeadInsts.insert(&SI);
2544 
2545  (void)Store;
2546  DEBUG(dbgs() << " to: " << *Store << "\n");
2547  return true;
2548  }
2549 
2550  bool rewriteIntegerStore(Value *V, StoreInst &SI) {
2551  assert(IntTy && "We cannot extract an integer from the alloca");
2552  assert(!SI.isVolatile());
2553  if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2554  Value *Old =
2555  IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2556  Old = convertValue(DL, IRB, Old, IntTy);
2557  assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2558  uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2559  V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset, "insert");
2560  }
2561  V = convertValue(DL, IRB, V, NewAllocaTy);
2562  StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2564  Pass.DeadInsts.insert(&SI);
2565  DEBUG(dbgs() << " to: " << *Store << "\n");
2566  return true;
2567  }
2568 
2569  bool visitStoreInst(StoreInst &SI) {
2570  DEBUG(dbgs() << " original: " << SI << "\n");
2571  Value *OldOp = SI.getOperand(1);
2572  assert(OldOp == OldPtr);
2573 
2574  Value *V = SI.getValueOperand();
2575 
2576  // Strip all inbounds GEPs and pointer casts to try to dig out any root
2577  // alloca that should be re-examined after promoting this alloca.
2578  if (V->getType()->isPointerTy())
2579  if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2580  Pass.PostPromotionWorklist.insert(AI);
2581 
2582  if (SliceSize < DL.getTypeStoreSize(V->getType())) {
2583  assert(!SI.isVolatile());
2584  assert(V->getType()->isIntegerTy() &&
2585  "Only integer type loads and stores are split");
2586  assert(V->getType()->getIntegerBitWidth() ==
2587  DL.getTypeStoreSizeInBits(V->getType()) &&
2588  "Non-byte-multiple bit width");
2589  IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), SliceSize * 8);
2590  V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset - BeginOffset,
2591  "extract");
2592  }
2593 
2594  if (VecTy)
2595  return rewriteVectorizedStoreInst(V, SI, OldOp);
2596  if (IntTy && V->getType()->isIntegerTy())
2597  return rewriteIntegerStore(V, SI);
2598 
2599  const bool IsStorePastEnd = DL.getTypeStoreSize(V->getType()) > SliceSize;
2600  StoreInst *NewSI;
2601  if (NewBeginOffset == NewAllocaBeginOffset &&
2602  NewEndOffset == NewAllocaEndOffset &&
2603  (canConvertValue(DL, V->getType(), NewAllocaTy) ||
2604  (IsStorePastEnd && NewAllocaTy->isIntegerTy() &&
2605  V->getType()->isIntegerTy()))) {
2606  // If this is an integer store past the end of slice (and thus the bytes
2607  // past that point are irrelevant or this is unreachable), truncate the
2608  // value prior to storing.
2609  if (auto *VITy = dyn_cast<IntegerType>(V->getType()))
2610  if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2611  if (VITy->getBitWidth() > AITy->getBitWidth()) {
2612  if (DL.isBigEndian())
2613  V = IRB.CreateLShr(V, VITy->getBitWidth() - AITy->getBitWidth(),
2614  "endian_shift");
2615  V = IRB.CreateTrunc(V, AITy, "load.trunc");
2616  }
2617 
2618  V = convertValue(DL, IRB, V, NewAllocaTy);
2619  NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2620  SI.isVolatile());
2621  } else {
2622  unsigned AS = SI.getPointerAddressSpace();
2623  Value *NewPtr = getNewAllocaSlicePtr(IRB, V->getType()->getPointerTo(AS));
2624  NewSI = IRB.CreateAlignedStore(V, NewPtr, getSliceAlign(V->getType()),
2625  SI.isVolatile());
2626  }
2628  if (SI.isVolatile())
2629  NewSI->setAtomic(SI.getOrdering(), SI.getSyncScopeID());
2630  Pass.DeadInsts.insert(&SI);
2631  deleteIfTriviallyDead(OldOp);
2632 
2633  DEBUG(dbgs() << " to: " << *NewSI << "\n");
2634  return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2635  }
2636 
2637  /// \brief Compute an integer value from splatting an i8 across the given
2638  /// number of bytes.
2639  ///
2640  /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2641  /// call this routine.
2642  /// FIXME: Heed the advice above.
2643  ///
2644  /// \param V The i8 value to splat.
2645  /// \param Size The number of bytes in the output (assuming i8 is one byte)
2646  Value *getIntegerSplat(Value *V, unsigned Size) {
2647  assert(Size > 0 && "Expected a positive number of bytes.");
2648  IntegerType *VTy = cast<IntegerType>(V->getType());
2649  assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2650  if (Size == 1)
2651  return V;
2652 
2653  Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size * 8);
2654  V = IRB.CreateMul(
2655  IRB.CreateZExt(V, SplatIntTy, "zext"),
2657  Constant::getAllOnesValue(SplatIntTy),
2659  SplatIntTy)),
2660  "isplat");
2661  return V;
2662  }
2663 
2664  /// \brief Compute a vector splat for a given element value.
2665  Value *getVectorSplat(Value *V, unsigned NumElements) {
2666  V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2667  DEBUG(dbgs() << " splat: " << *V << "\n");
2668  return V;
2669  }
2670 
2671  bool visitMemSetInst(MemSetInst &II) {
2672  DEBUG(dbgs() << " original: " << II << "\n");
2673  assert(II.getRawDest() == OldPtr);
2674 
2675  // If the memset has a variable size, it cannot be split, just adjust the
2676  // pointer to the new alloca.
2677  if (!isa<Constant>(II.getLength())) {
2678  assert(!IsSplit);
2679  assert(NewBeginOffset == BeginOffset);
2680  II.setDest(getNewAllocaSlicePtr(IRB, OldPtr->getType()));
2681  Type *CstTy = II.getAlignmentCst()->getType();
2682  II.setAlignment(ConstantInt::get(CstTy, getSliceAlign()));
2683 
2684  deleteIfTriviallyDead(OldPtr);
2685  return false;
2686  }
2687 
2688  // Record this instruction for deletion.
2689  Pass.DeadInsts.insert(&II);
2690 
2691  Type *AllocaTy = NewAI.getAllocatedType();
2692  Type *ScalarTy = AllocaTy->getScalarType();
2693 
2694  // If this doesn't map cleanly onto the alloca type, and that type isn't
2695  // a single value type, just emit a memset.
2696  if (!VecTy && !IntTy &&
2697  (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2698  SliceSize != DL.getTypeStoreSize(AllocaTy) ||
2699  !AllocaTy->isSingleValueType() ||
2700  !DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy)) ||
2701  DL.getTypeSizeInBits(ScalarTy) % 8 != 0)) {
2702  Type *SizeTy = II.getLength()->getType();
2703  Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2704  CallInst *New = IRB.CreateMemSet(
2705  getNewAllocaSlicePtr(IRB, OldPtr->getType()), II.getValue(), Size,
2706  getSliceAlign(), II.isVolatile());
2707  (void)New;
2708  DEBUG(dbgs() << " to: " << *New << "\n");
2709  return false;
2710  }
2711 
2712  // If we can represent this as a simple value, we have to build the actual
2713  // value to store, which requires expanding the byte present in memset to
2714  // a sensible representation for the alloca type. This is essentially
2715  // splatting the byte to a sufficiently wide integer, splatting it across
2716  // any desired vector width, and bitcasting to the final type.
2717  Value *V;
2718 
2719  if (VecTy) {
2720  // If this is a memset of a vectorized alloca, insert it.
2721  assert(ElementTy == ScalarTy);
2722 
2723  unsigned BeginIndex = getIndex(NewBeginOffset);
2724  unsigned EndIndex = getIndex(NewEndOffset);
2725  assert(EndIndex > BeginIndex && "Empty vector!");
2726  unsigned NumElements = EndIndex - BeginIndex;
2727  assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2728 
2729  Value *Splat =
2730  getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
2731  Splat = convertValue(DL, IRB, Splat, ElementTy);
2732  if (NumElements > 1)
2733  Splat = getVectorSplat(Splat, NumElements);
2734 
2735  Value *Old =
2736  IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2737  V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2738  } else if (IntTy) {
2739  // If this is a memset on an alloca where we can widen stores, insert the
2740  // set integer.
2741  assert(!II.isVolatile());
2742 
2743  uint64_t Size = NewEndOffset - NewBeginOffset;
2744  V = getIntegerSplat(II.getValue(), Size);
2745 
2746  if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2747  EndOffset != NewAllocaBeginOffset)) {
2748  Value *Old =
2749  IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2750  Old = convertValue(DL, IRB, Old, IntTy);
2751  uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2752  V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2753  } else {
2754  assert(V->getType() == IntTy &&
2755  "Wrong type for an alloca wide integer!");
2756  }
2757  V = convertValue(DL, IRB, V, AllocaTy);
2758  } else {
2759  // Established these invariants above.
2760  assert(NewBeginOffset == NewAllocaBeginOffset);
2761  assert(NewEndOffset == NewAllocaEndOffset);
2762 
2763  V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
2764  if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2765  V = getVectorSplat(V, AllocaVecTy->getNumElements());
2766 
2767  V = convertValue(DL, IRB, V, AllocaTy);
2768  }
2769 
2770  Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2771  II.isVolatile());
2772  (void)New;
2773  DEBUG(dbgs() << " to: " << *New << "\n");
2774  return !II.isVolatile();
2775  }
2776 
2777  bool visitMemTransferInst(MemTransferInst &II) {
2778  // Rewriting of memory transfer instructions can be a bit tricky. We break
2779  // them into two categories: split intrinsics and unsplit intrinsics.
2780 
2781  DEBUG(dbgs() << " original: " << II << "\n");
2782 
2783  bool IsDest = &II.getRawDestUse() == OldUse;
2784  assert((IsDest && II.getRawDest() == OldPtr) ||
2785  (!IsDest && II.getRawSource() == OldPtr));
2786 
2787  unsigned SliceAlign = getSliceAlign();
2788 
2789  // For unsplit intrinsics, we simply modify the source and destination
2790  // pointers in place. This isn't just an optimization, it is a matter of
2791  // correctness. With unsplit intrinsics we may be dealing with transfers
2792  // within a single alloca before SROA ran, or with transfers that have
2793  // a variable length. We may also be dealing with memmove instead of
2794  // memcpy, and so simply updating the pointers is the necessary for us to
2795  // update both source and dest of a single call.
2796  if (!IsSplittable) {
2797  Value *AdjustedPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2798  if (IsDest)
2799  II.setDest(AdjustedPtr);
2800  else
2801  II.setSource(AdjustedPtr);
2802 
2803  if (II.getAlignment() > SliceAlign) {
2804  Type *CstTy = II.getAlignmentCst()->getType();
2805  II.setAlignment(
2806  ConstantInt::get(CstTy, MinAlign(II.getAlignment(), SliceAlign)));
2807  }
2808 
2809  DEBUG(dbgs() << " to: " << II << "\n");
2810  deleteIfTriviallyDead(OldPtr);
2811  return false;
2812  }
2813  // For split transfer intrinsics we have an incredibly useful assurance:
2814  // the source and destination do not reside within the same alloca, and at
2815  // least one of them does not escape. This means that we can replace
2816  // memmove with memcpy, and we don't need to worry about all manner of
2817  // downsides to splitting and transforming the operations.
2818 
2819  // If this doesn't map cleanly onto the alloca type, and that type isn't
2820  // a single value type, just emit a memcpy.
2821  bool EmitMemCpy =
2822  !VecTy && !IntTy &&
2823  (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2824  SliceSize != DL.getTypeStoreSize(NewAI.getAllocatedType()) ||
2825  !NewAI.getAllocatedType()->isSingleValueType());
2826 
2827  // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2828  // size hasn't been shrunk based on analysis of the viable range, this is
2829  // a no-op.
2830  if (EmitMemCpy && &OldAI == &NewAI) {
2831  // Ensure the start lines up.
2832  assert(NewBeginOffset == BeginOffset);
2833 
2834  // Rewrite the size as needed.
2835  if (NewEndOffset != EndOffset)
2837  NewEndOffset - NewBeginOffset));
2838  return false;
2839  }
2840  // Record this instruction for deletion.
2841  Pass.DeadInsts.insert(&II);
2842 
2843  // Strip all inbounds GEPs and pointer casts to try to dig out any root
2844  // alloca that should be re-examined after rewriting this instruction.
2845  Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2846  if (AllocaInst *AI =
2847  dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
2848  assert(AI != &OldAI && AI != &NewAI &&
2849  "Splittable transfers cannot reach the same alloca on both ends.");
2850  Pass.Worklist.insert(AI);
2851  }
2852 
2853  Type *OtherPtrTy = OtherPtr->getType();
2854  unsigned OtherAS = OtherPtrTy->getPointerAddressSpace();
2855 
2856  // Compute the relative offset for the other pointer within the transfer.
2857  unsigned IntPtrWidth = DL.getPointerSizeInBits(OtherAS);
2858  APInt OtherOffset(IntPtrWidth, NewBeginOffset - BeginOffset);
2859  unsigned OtherAlign = MinAlign(II.getAlignment() ? II.getAlignment() : 1,
2860  OtherOffset.zextOrTrunc(64).getZExtValue());
2861 
2862  if (EmitMemCpy) {
2863  // Compute the other pointer, folding as much as possible to produce
2864  // a single, simple GEP in most cases.
2865  OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2866  OtherPtr->getName() + ".");
2867 
2868  Value *OurPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2869  Type *SizeTy = II.getLength()->getType();
2870  Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2871 
2872  CallInst *New = IRB.CreateMemCpy(
2873  IsDest ? OurPtr : OtherPtr, IsDest ? OtherPtr : OurPtr, Size,
2874  MinAlign(SliceAlign, OtherAlign), II.isVolatile());
2875  (void)New;
2876  DEBUG(dbgs() << " to: " << *New << "\n");
2877  return false;
2878  }
2879 
2880  bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
2881  NewEndOffset == NewAllocaEndOffset;
2882  uint64_t Size = NewEndOffset - NewBeginOffset;
2883  unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
2884  unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
2885  unsigned NumElements = EndIndex - BeginIndex;
2886  IntegerType *SubIntTy =
2887  IntTy ? Type::getIntNTy(IntTy->getContext(), Size * 8) : nullptr;
2888 
2889  // Reset the other pointer type to match the register type we're going to
2890  // use, but using the address space of the original other pointer.
2891  if (VecTy && !IsWholeAlloca) {
2892  if (NumElements == 1)
2893  OtherPtrTy = VecTy->getElementType();
2894  else
2895  OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2896 
2897  OtherPtrTy = OtherPtrTy->getPointerTo(OtherAS);
2898  } else if (IntTy && !IsWholeAlloca) {
2899  OtherPtrTy = SubIntTy->getPointerTo(OtherAS);
2900  } else {
2901  OtherPtrTy = NewAllocaTy->getPointerTo(OtherAS);
2902  }
2903 
2904  Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2905  OtherPtr->getName() + ".");
2906  unsigned SrcAlign = OtherAlign;
2907  Value *DstPtr = &NewAI;
2908  unsigned DstAlign = SliceAlign;
2909  if (!IsDest) {
2910  std::swap(SrcPtr, DstPtr);
2911  std::swap(SrcAlign, DstAlign);
2912  }
2913 
2914  Value *Src;
2915  if (VecTy && !IsWholeAlloca && !IsDest) {
2916  Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2917  Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
2918  } else if (IntTy && !IsWholeAlloca && !IsDest) {
2919  Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2920  Src = convertValue(DL, IRB, Src, IntTy);
2921  uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2922  Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
2923  } else {
2924  Src =
2925  IRB.CreateAlignedLoad(SrcPtr, SrcAlign, II.isVolatile(), "copyload");
2926  }
2927 
2928  if (VecTy && !IsWholeAlloca && IsDest) {
2929  Value *Old =
2930  IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2931  Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
2932  } else if (IntTy && !IsWholeAlloca && IsDest) {
2933  Value *Old =
2934  IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2935  Old = convertValue(DL, IRB, Old, IntTy);
2936  uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2937  Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
2938  Src = convertValue(DL, IRB, Src, NewAllocaTy);
2939  }
2940 
2941  StoreInst *Store = cast<StoreInst>(
2942  IRB.CreateAlignedStore(Src, DstPtr, DstAlign, II.isVolatile()));
2943  (void)Store;
2944  DEBUG(dbgs() << " to: " << *Store << "\n");
2945  return !II.isVolatile();
2946  }
2947 
2948  bool visitIntrinsicInst(IntrinsicInst &II) {
2949  assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2950  II.getIntrinsicID() == Intrinsic::lifetime_end);
2951  DEBUG(dbgs() << " original: " << II << "\n");
2952  assert(II.getArgOperand(1) == OldPtr);
2953 
2954  // Record this instruction for deletion.
2955  Pass.DeadInsts.insert(&II);
2956 
2957  // Lifetime intrinsics are only promotable if they cover the whole alloca.
2958  // Therefore, we drop lifetime intrinsics which don't cover the whole
2959  // alloca.
2960  // (In theory, intrinsics which partially cover an alloca could be
2961  // promoted, but PromoteMemToReg doesn't handle that case.)
2962  // FIXME: Check whether the alloca is promotable before dropping the
2963  // lifetime intrinsics?
2964  if (NewBeginOffset != NewAllocaBeginOffset ||
2965  NewEndOffset != NewAllocaEndOffset)
2966  return true;
2967 
2968  ConstantInt *Size =
2969  ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2970  NewEndOffset - NewBeginOffset);
2971  Value *Ptr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2972  Value *New;
2973  if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2974  New = IRB.CreateLifetimeStart(Ptr, Size);
2975  else
2976  New = IRB.CreateLifetimeEnd(Ptr, Size);
2977 
2978  (void)New;
2979  DEBUG(dbgs() << " to: " << *New << "\n");
2980 
2981  return true;
2982  }
2983 
2984  bool visitPHINode(PHINode &PN) {
2985  DEBUG(dbgs() << " original: " << PN << "\n");
2986  assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
2987  assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
2988 
2989  // We would like to compute a new pointer in only one place, but have it be
2990  // as local as possible to the PHI. To do that, we re-use the location of
2991  // the old pointer, which necessarily must be in the right position to
2992  // dominate the PHI.
2993  IRBuilderTy PtrBuilder(IRB);
2994  if (isa<PHINode>(OldPtr))
2995  PtrBuilder.SetInsertPoint(&*OldPtr->getParent()->getFirstInsertionPt());
2996  else
2997  PtrBuilder.SetInsertPoint(OldPtr);
2998  PtrBuilder.SetCurrentDebugLocation(OldPtr->getDebugLoc());
2999 
3000  Value *NewPtr = getNewAllocaSlicePtr(PtrBuilder, OldPtr->getType());
3001  // Replace the operands which were using the old pointer.
3002  std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
3003 
3004  DEBUG(dbgs() << " to: " << PN << "\n");
3005  deleteIfTriviallyDead(OldPtr);
3006 
3007  // PHIs can't be promoted on their own, but often can be speculated. We
3008  // check the speculation outside of the rewriter so that we see the
3009  // fully-rewritten alloca.
3010  PHIUsers.insert(&PN);
3011  return true;
3012  }
3013 
3014  bool visitSelectInst(SelectInst &SI) {
3015  DEBUG(dbgs() << " original: " << SI << "\n");
3016  assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
3017  "Pointer isn't an operand!");
3018  assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
3019  assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
3020 
3021  Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
3022  // Replace the operands which were using the old pointer.
3023  if (SI.getOperand(1) == OldPtr)
3024  SI.setOperand(1, NewPtr);
3025  if (SI.getOperand(2) == OldPtr)
3026  SI.setOperand(2, NewPtr);
3027 
3028  DEBUG(dbgs() << " to: " << SI << "\n");
3029  deleteIfTriviallyDead(OldPtr);
3030 
3031  // Selects can't be promoted on their own, but often can be speculated. We
3032  // check the speculation outside of the rewriter so that we see the
3033  // fully-rewritten alloca.
3034  SelectUsers.insert(&SI);
3035  return true;
3036  }
3037 };
3038 
3039 namespace {
3040 
3041 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
3042 ///
3043 /// This pass aggressively rewrites all aggregate loads and stores on
3044 /// a particular pointer (or any pointer derived from it which we can identify)
3045 /// with scalar loads and stores.
3046 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
3047  // Befriend the base class so it can delegate to private visit methods.
3048  friend class InstVisitor<AggLoadStoreRewriter, bool>;
3049 
3050  /// Queue of pointer uses to analyze and potentially rewrite.
3051  SmallVector<Use *, 8> Queue;
3052 
3053  /// Set to prevent us from cycling with phi nodes and loops.
3054  SmallPtrSet<User *, 8> Visited;
3055 
3056  /// The current pointer use being rewritten. This is used to dig up the used
3057  /// value (as opposed to the user).
3058  Use *U;
3059 
3060 public:
3061  /// Rewrite loads and stores through a pointer and all pointers derived from
3062  /// it.
3063  bool rewrite(Instruction &I) {
3064  DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
3065  enqueueUsers(I);
3066  bool Changed = false;
3067  while (!Queue.empty()) {
3068  U = Queue.pop_back_val();
3069  Changed |= visit(cast<Instruction>(U->getUser()));
3070  }
3071  return Changed;
3072  }
3073 
3074 private:
3075  /// Enqueue all the users of the given instruction for further processing.
3076  /// This uses a set to de-duplicate users.
3077  void enqueueUsers(Instruction &I) {
3078  for (Use &U : I.uses())
3079  if (Visited.insert(U.getUser()).second)
3080  Queue.push_back(&U);
3081  }
3082 
3083  // Conservative default is to not rewrite anything.
3084  bool visitInstruction(Instruction &I) { return false; }
3085 
3086  /// \brief Generic recursive split emission class.
3087  template <typename Derived> class OpSplitter {
3088  protected:
3089  /// The builder used to form new instructions.
3090  IRBuilderTy IRB;
3091 
3092  /// The indices which to be used with insert- or extractvalue to select the
3093  /// appropriate value within the aggregate.
3094  SmallVector<unsigned, 4> Indices;
3095 
3096  /// The indices to a GEP instruction which will move Ptr to the correct slot
3097  /// within the aggregate.
3098  SmallVector<Value *, 4> GEPIndices;
3099 
3100  /// The base pointer of the original op, used as a base for GEPing the
3101  /// split operations.
3102  Value *Ptr;
3103 
3104  /// Initialize the splitter with an insertion point, Ptr and start with a
3105  /// single zero GEP index.
3106  OpSplitter(Instruction *InsertionPoint, Value *Ptr)
3107  : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
3108 
3109  public:
3110  /// \brief Generic recursive split emission routine.
3111  ///
3112  /// This method recursively splits an aggregate op (load or store) into
3113  /// scalar or vector ops. It splits recursively until it hits a single value
3114  /// and emits that single value operation via the template argument.
3115  ///
3116  /// The logic of this routine relies on GEPs and insertvalue and
3117  /// extractvalue all operating with the same fundamental index list, merely
3118  /// formatted differently (GEPs need actual values).
3119  ///
3120  /// \param Ty The type being split recursively into smaller ops.
3121  /// \param Agg The aggregate value being built up or stored, depending on
3122  /// whether this is splitting a load or a store respectively.
3123  void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
3124  if (Ty->isSingleValueType())
3125  return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
3126 
3127  if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
3128  unsigned OldSize = Indices.size();
3129  (void)OldSize;
3130  for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
3131  ++Idx) {
3132  assert(Indices.size() == OldSize && "Did not return to the old size");
3133  Indices.push_back(Idx);
3134  GEPIndices.push_back(IRB.getInt32(Idx));
3135  emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
3136  GEPIndices.pop_back();
3137  Indices.pop_back();
3138  }
3139  return;
3140  }
3141 
3142  if (StructType *STy = dyn_cast<StructType>(Ty)) {
3143  unsigned OldSize = Indices.size();
3144  (void)OldSize;
3145  for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
3146  ++Idx) {
3147  assert(Indices.size() == OldSize && "Did not return to the old size");
3148  Indices.push_back(Idx);
3149  GEPIndices.push_back(IRB.getInt32(Idx));
3150  emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
3151  GEPIndices.pop_back();
3152  Indices.pop_back();
3153  }
3154  return;
3155  }
3156 
3157  llvm_unreachable("Only arrays and structs are aggregate loadable types");
3158  }
3159  };
3160 
3161  struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
3162  LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3163  : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
3164 
3165  /// Emit a leaf load of a single value. This is called at the leaves of the
3166  /// recursive emission to actually load values.
3167  void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3168  assert(Ty->isSingleValueType());
3169  // Load the single value and insert it using the indices.
3170  Value *GEP =
3171  IRB.CreateInBoundsGEP(nullptr, Ptr, GEPIndices, Name + ".gep");
3172  Value *Load = IRB.CreateLoad(GEP, Name + ".load");
3173  Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
3174  DEBUG(dbgs() << " to: " << *Load << "\n");
3175  }
3176  };
3177 
3178  bool visitLoadInst(LoadInst &LI) {
3179  assert(LI.getPointerOperand() == *U);
3180  if (!LI.isSimple() || LI.getType()->isSingleValueType())
3181  return false;
3182 
3183  // We have an aggregate being loaded, split it apart.
3184  DEBUG(dbgs() << " original: " << LI << "\n");
3185  LoadOpSplitter Splitter(&LI, *U);
3186  Value *V = UndefValue::get(LI.getType());
3187  Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
3188  LI.replaceAllUsesWith(V);
3189  LI.eraseFromParent();
3190  return true;
3191  }
3192 
3193  struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
3194  StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3195  : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
3196 
3197  /// Emit a leaf store of a single value. This is called at the leaves of the
3198  /// recursive emission to actually produce stores.
3199  void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3200  assert(Ty->isSingleValueType());
3201  // Extract the single value and store it using the indices.
3202  //
3203  // The gep and extractvalue values are factored out of the CreateStore
3204  // call to make the output independent of the argument evaluation order.
3205  Value *ExtractValue =
3206  IRB.CreateExtractValue(Agg, Indices, Name + ".extract");
3207  Value *InBoundsGEP =
3208  IRB.CreateInBoundsGEP(nullptr, Ptr, GEPIndices, Name + ".gep");
3209  Value *Store = IRB.CreateStore(ExtractValue, InBoundsGEP);
3210  (void)Store;
3211  DEBUG(dbgs() << " to: " << *Store << "\n");
3212  }
3213  };
3214 
3215  bool visitStoreInst(StoreInst &SI) {
3216  if (!SI.isSimple() || SI.getPointerOperand() != *U)
3217  return false;
3218  Value *V = SI.getValueOperand();
3219  if (V->getType()->isSingleValueType())
3220  return false;
3221 
3222  // We have an aggregate being stored, split it apart.
3223  DEBUG(dbgs() << " original: " << SI << "\n");
3224  StoreOpSplitter Splitter(&SI, *U);
3225  Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
3226  SI.eraseFromParent();
3227  return true;
3228  }
3229 
3230  bool visitBitCastInst(BitCastInst &BC) {
3231  enqueueUsers(BC);
3232  return false;
3233  }
3234 
3235  bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
3236  enqueueUsers(GEPI);
3237  return false;
3238  }
3239 
3240  bool visitPHINode(PHINode &PN) {
3241  enqueueUsers(PN);
3242  return false;
3243  }
3244 
3245  bool visitSelectInst(SelectInst &SI) {
3246  enqueueUsers(SI);
3247  return false;
3248  }
3249 };
3250 
3251 } // end anonymous namespace
3252 
3253 /// \brief Strip aggregate type wrapping.
3254 ///
3255 /// This removes no-op aggregate types wrapping an underlying type. It will
3256 /// strip as many layers of types as it can without changing either the type
3257 /// size or the allocated size.
3259  if (Ty->isSingleValueType())
3260  return Ty;
3261 
3262  uint64_t AllocSize = DL.getTypeAllocSize(Ty);
3263  uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
3264 
3265  Type *InnerTy;
3266  if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
3267  InnerTy = ArrTy->getElementType();
3268  } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
3269  const StructLayout *SL = DL.getStructLayout(STy);
3270  unsigned Index = SL->getElementContainingOffset(0);
3271  InnerTy = STy->getElementType(Index);
3272  } else {
3273  return Ty;
3274  }
3275 
3276  if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
3277  TypeSize > DL.getTypeSizeInBits(InnerTy))
3278  return Ty;
3279 
3280  return stripAggregateTypeWrapping(DL, InnerTy);
3281 }
3282 
3283 /// \brief Try to find a partition of the aggregate type passed in for a given
3284 /// offset and size.
3285 ///
3286 /// This recurses through the aggregate type and tries to compute a subtype
3287 /// based on the offset and size. When the offset and size span a sub-section
3288 /// of an array, it will even compute a new array type for that sub-section,
3289 /// and the same for structs.
3290 ///
3291 /// Note that this routine is very strict and tries to find a partition of the
3292 /// type which produces the *exact* right offset and size. It is not forgiving
3293 /// when the size or offset cause either end of type-based partition to be off.
3294 /// Also, this is a best-effort routine. It is reasonable to give up and not
3295 /// return a type if necessary.
3296 static Type *getTypePartition(const DataLayout &DL, Type *Ty, uint64_t Offset,
3297  uint64_t Size) {
3298  if (Offset == 0 && DL.getTypeAllocSize(Ty) == Size)
3299  return stripAggregateTypeWrapping(DL, Ty);
3300  if (Offset > DL.getTypeAllocSize(Ty) ||
3301  (DL.getTypeAllocSize(Ty) - Offset) < Size)
3302  return nullptr;
3303 
3304  if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3305  Type *ElementTy = SeqTy->getElementType();
3306  uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3307  uint64_t NumSkippedElements = Offset / ElementSize;
3308  if (NumSkippedElements >= SeqTy->getNumElements())
3309  return nullptr;
3310  Offset -= NumSkippedElements * ElementSize;
3311 
3312  // First check if we need to recurse.
3313  if (Offset > 0 || Size < ElementSize) {
3314  // Bail if the partition ends in a different array element.
3315  if ((Offset + Size) > ElementSize)
3316  return nullptr;
3317  // Recurse through the element type trying to peel off offset bytes.
3318  return getTypePartition(DL, ElementTy, Offset, Size);
3319  }
3320  assert(Offset == 0);
3321 
3322  if (Size == ElementSize)
3323  return stripAggregateTypeWrapping(DL, ElementTy);
3324  assert(Size > ElementSize);
3325  uint64_t NumElements = Size / ElementSize;
3326  if (NumElements * ElementSize != Size)
3327  return nullptr;
3328  return ArrayType::get(ElementTy, NumElements);
3329  }
3330 
3331  StructType *STy = dyn_cast<StructType>(Ty);
3332  if (!STy)
3333  return nullptr;
3334 
3335  const StructLayout *SL = DL.getStructLayout(STy);
3336  if (Offset >= SL->getSizeInBytes())
3337  return nullptr;
3338  uint64_t EndOffset = Offset + Size;
3339  if (EndOffset > SL->getSizeInBytes())
3340  return nullptr;
3341 
3342  unsigned Index = SL->getElementContainingOffset(Offset);
3343  Offset -= SL->getElementOffset(Index);
3344 
3345  Type *ElementTy = STy->getElementType(Index);
3346  uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3347  if (Offset >= ElementSize)
3348  return nullptr; // The offset points into alignment padding.
3349 
3350  // See if any partition must be contained by the element.
3351  if (Offset > 0 || Size < ElementSize) {
3352  if ((Offset + Size) > ElementSize)
3353  return nullptr;
3354  return getTypePartition(DL, ElementTy, Offset, Size);
3355  }
3356  assert(Offset == 0);
3357 
3358  if (Size == ElementSize)
3359  return stripAggregateTypeWrapping(DL, ElementTy);
3360 
3361  StructType::element_iterator EI = STy->element_begin() + Index,
3362  EE = STy->element_end();
3363  if (EndOffset < SL->getSizeInBytes()) {
3364  unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3365  if (Index == EndIndex)
3366  return nullptr; // Within a single element and its padding.
3367 
3368  // Don't try to form "natural" types if the elements don't line up with the
3369  // expected size.
3370  // FIXME: We could potentially recurse down through the last element in the
3371  // sub-struct to find a natural end point.
3372  if (SL->getElementOffset(EndIndex) != EndOffset)
3373  return nullptr;
3374 
3375  assert(Index < EndIndex);
3376  EE = STy->element_begin() + EndIndex;
3377  }
3378 
3379  // Try to build up a sub-structure.
3380  StructType *SubTy =
3381  StructType::get(STy->getContext(), makeArrayRef(EI, EE), STy->isPacked());
3382  const StructLayout *SubSL = DL.getStructLayout(SubTy);
3383  if (Size != SubSL->getSizeInBytes())
3384  return nullptr; // The sub-struct doesn't have quite the size needed.
3385 
3386  return SubTy;
3387 }
3388 
3389 /// \brief Pre-split loads and stores to simplify rewriting.
3390 ///
3391 /// We want to break up the splittable load+store pairs as much as
3392 /// possible. This is important to do as a preprocessing step, as once we
3393 /// start rewriting the accesses to partitions of the alloca we lose the
3394 /// necessary information to correctly split apart paired loads and stores
3395 /// which both point into this alloca. The case to consider is something like
3396 /// the following:
3397 ///
3398 /// %a = alloca [12 x i8]
3399 /// %gep1 = getelementptr [12 x i8]* %a, i32 0, i32 0
3400 /// %gep2 = getelementptr [12 x i8]* %a, i32 0, i32 4
3401 /// %gep3 = getelementptr [12 x i8]* %a, i32 0, i32 8
3402 /// %iptr1 = bitcast i8* %gep1 to i64*
3403 /// %iptr2 = bitcast i8* %gep2 to i64*
3404 /// %fptr1 = bitcast i8* %gep1 to float*
3405 /// %fptr2 = bitcast i8* %gep2 to float*
3406 /// %fptr3 = bitcast i8* %gep3 to float*
3407 /// store float 0.0, float* %fptr1
3408 /// store float 1.0, float* %fptr2
3409 /// %v = load i64* %iptr1
3410 /// store i64 %v, i64* %iptr2
3411 /// %f1 = load float* %fptr2
3412 /// %f2 = load float* %fptr3
3413 ///
3414 /// Here we want to form 3 partitions of the alloca, each 4 bytes large, and
3415 /// promote everything so we recover the 2 SSA values that should have been
3416 /// there all along.
3417 ///
3418 /// \returns true if any changes are made.
3419 bool SROA::presplitLoadsAndStores(AllocaInst &AI, AllocaSlices &AS) {
3420  DEBUG(dbgs() << "Pre-splitting loads and stores\n");
3421 
3422  // Track the loads and stores which are candidates for pre-splitting here, in
3423  // the order they first appear during the partition scan. These give stable
3424  // iteration order and a basis for tracking which loads and stores we
3425  // actually split.
3428 
3429  // We need to accumulate the splits required of each load or store where we
3430  // can find them via a direct lookup. This is important to cross-check loads
3431  // and stores against each other. We also track the slice so that we can kill
3432  // all the slices that end up split.
3433  struct SplitOffsets {
3434  Slice *S;
3435  std::vector<uint64_t> Splits;
3436  };
3438 
3439  // Track loads out of this alloca which cannot, for any reason, be pre-split.
3440  // This is important as we also cannot pre-split stores of those loads!
3441  // FIXME: This is all pretty gross. It means that we can be more aggressive
3442  // in pre-splitting when the load feeding the store happens to come from
3443  // a separate alloca. Put another way, the effectiveness of SROA would be
3444  // decreased by a frontend which just concatenated all of its local allocas
3445  // into one big flat alloca. But defeating such patterns is exactly the job
3446  // SROA is tasked with! Sadly, to not have this discrepancy we would have
3447  // change store pre-splitting to actually force pre-splitting of the load
3448  // that feeds it *and all stores*. That makes pre-splitting much harder, but
3449  // maybe it would make it more principled?
3450  SmallPtrSet<LoadInst *, 8> UnsplittableLoads;
3451 
3452  DEBUG(dbgs() << " Searching for candidate loads and stores\n");
3453  for (auto &P : AS.partitions()) {
3454  for (Slice &S : P) {
3455  Instruction *I = cast<Instruction>(S.getUse()->getUser());
3456  if (!S.isSplittable() || S.endOffset() <= P.endOffset()) {
3457  // If this is a load we have to track that it can't participate in any
3458  // pre-splitting. If this is a store of a load we have to track that
3459  // that load also can't participate in any pre-splitting.
3460  if (auto *LI = dyn_cast<LoadInst>(I))
3461  UnsplittableLoads.insert(LI);
3462  else if (auto *SI = dyn_cast<StoreInst>(I))
3463  if (auto *LI = dyn_cast<LoadInst>(SI->getValueOperand()))
3464  UnsplittableLoads.insert(LI);
3465  continue;
3466  }
3467  assert(P.endOffset() > S.beginOffset() &&
3468  "Empty or backwards partition!");
3469 
3470  // Determine if this is a pre-splittable slice.
3471  if (auto *LI = dyn_cast<LoadInst>(I)) {
3472  assert(!LI->isVolatile() && "Cannot split volatile loads!");
3473 
3474  // The load must be used exclusively to store into other pointers for
3475  // us to be able to arbitrarily pre-split it. The stores must also be
3476  // simple to avoid changing semantics.
3477  auto IsLoadSimplyStored = [](LoadInst *LI) {
3478  for (User *LU : LI->users()) {
3479  auto *SI = dyn_cast<StoreInst>(LU);
3480  if (!SI || !SI->isSimple())
3481  return false;
3482  }
3483  return true;
3484  };
3485  if (!IsLoadSimplyStored(LI)) {
3486  UnsplittableLoads.insert(LI);
3487  continue;
3488  }
3489 
3490  Loads.push_back(LI);
3491  } else if (auto *SI = dyn_cast<StoreInst>(I)) {
3492  if (S.getUse() != &SI->getOperandUse(SI->getPointerOperandIndex()))
3493  // Skip stores *of* pointers. FIXME: This shouldn't even be possible!
3494  continue;
3495  auto *StoredLoad = dyn_cast<LoadInst>(SI->getValueOperand());
3496  if (!StoredLoad || !StoredLoad->isSimple())
3497  continue;
3498  assert(!SI->isVolatile() && "Cannot split volatile stores!");
3499 
3500  Stores.push_back(SI);
3501  } else {
3502  // Other uses cannot be pre-split.
3503  continue;
3504  }
3505 
3506  // Record the initial split.
3507  DEBUG(dbgs() << " Candidate: " << *I << "\n");
3508  auto &Offsets = SplitOffsetsMap[I];
3509  assert(Offsets.Splits.empty() &&
3510  "Should not have splits the first time we see an instruction!");
3511  Offsets.S = &S;
3512  Offsets.Splits.push_back(P.endOffset() - S.beginOffset());
3513  }
3514 
3515  // Now scan the already split slices, and add a split for any of them which
3516  // we're going to pre-split.
3517  for (Slice *S : P.splitSliceTails()) {
3518  auto SplitOffsetsMapI =
3519  SplitOffsetsMap.find(cast<Instruction>(S->getUse()->getUser()));
3520  if (SplitOffsetsMapI == SplitOffsetsMap.end())
3521  continue;
3522  auto &Offsets = SplitOffsetsMapI->second;
3523 
3524  assert(Offsets.S == S && "Found a mismatched slice!");
3525  assert(!Offsets.Splits.empty() &&
3526  "Cannot have an empty set of splits on the second partition!");
3527  assert(Offsets.Splits.back() ==
3528  P.beginOffset() - Offsets.S->beginOffset() &&
3529  "Previous split does not end where this one begins!");
3530 
3531  // Record each split. The last partition's end isn't needed as the size
3532  // of the slice dictates that.
3533  if (S->endOffset() > P.endOffset())
3534  Offsets.Splits.push_back(P.endOffset() - Offsets.S->beginOffset());
3535  }
3536  }
3537 
3538  // We may have split loads where some of their stores are split stores. For
3539  // such loads and stores, we can only pre-split them if their splits exactly
3540  // match relative to their starting offset. We have to verify this prior to
3541  // any rewriting.
3542  Stores.erase(
3543  llvm::remove_if(Stores,
3544  [&UnsplittableLoads, &SplitOffsetsMap](StoreInst *SI) {
3545  // Lookup the load we are storing in our map of split
3546  // offsets.
3547  auto *LI = cast<LoadInst>(SI->getValueOperand());
3548  // If it was completely unsplittable, then we're done,
3549  // and this store can't be pre-split.
3550  if (UnsplittableLoads.count(LI))
3551  return true;
3552 
3553  auto LoadOffsetsI = SplitOffsetsMap.find(LI);
3554  if (LoadOffsetsI == SplitOffsetsMap.end())
3555  return false; // Unrelated loads are definitely safe.
3556  auto &LoadOffsets = LoadOffsetsI->second;
3557 
3558  // Now lookup the store's offsets.
3559  auto &StoreOffsets = SplitOffsetsMap[SI];
3560 
3561  // If the relative offsets of each split in the load and
3562  // store match exactly, then we can split them and we
3563  // don't need to remove them here.
3564  if (LoadOffsets.Splits == StoreOffsets.Splits)
3565  return false;
3566 
3567  DEBUG(dbgs()
3568  << " Mismatched splits for load and store:\n"
3569  << " " << *LI << "\n"
3570  << " " << *SI << "\n");
3571 
3572  // We've found a store and load that we need to split
3573  // with mismatched relative splits. Just give up on them
3574  // and remove both instructions from our list of
3575  // candidates.
3576  UnsplittableLoads.insert(LI);
3577  return true;
3578  }),
3579  Stores.end());
3580  // Now we have to go *back* through all the stores, because a later store may
3581  // have caused an earlier store's load to become unsplittable and if it is
3582  // unsplittable for the later store, then we can't rely on it being split in
3583  // the earlier store either.
3584  Stores.erase(llvm::remove_if(Stores,
3585  [&UnsplittableLoads](StoreInst *SI) {
3586  auto *LI =
3587  cast<LoadInst>(SI->getValueOperand());
3588  return UnsplittableLoads.count(LI);
3589  }),
3590  Stores.end());
3591  // Once we've established all the loads that can't be split for some reason,
3592  // filter any that made it into our list out.
3593  Loads.erase(llvm::remove_if(Loads,
3594  [&UnsplittableLoads](LoadInst *LI) {
3595  return UnsplittableLoads.count(LI);
3596  }),
3597  Loads.end());
3598 
3599  // If no loads or stores are left, there is no pre-splitting to be done for
3600  // this alloca.
3601  if (Loads.empty() && Stores.empty())
3602  return false;
3603 
3604  // From here on, we can't fail and will be building new accesses, so rig up
3605  // an IR builder.
3606  IRBuilderTy IRB(&AI);
3607 
3608  // Collect the new slices which we will merge into the alloca slices.
3609  SmallVector<Slice, 4> NewSlices;
3610 
3611  // Track any allocas we end up splitting loads and stores for so we iterate
3612  // on them.
3613  SmallPtrSet<AllocaInst *, 4> ResplitPromotableAllocas;
3614 
3615  // At this point, we have collected all of the loads and stores we can
3616  // pre-split, and the specific splits needed for them. We actually do the
3617  // splitting in a specific order in order to handle when one of the loads in
3618  // the value operand to one of the stores.
3619  //
3620  // First, we rewrite all of the split loads, and just accumulate each split
3621  // load in a parallel structure. We also build the slices for them and append
3622  // them to the alloca slices.
3624  std::vector<LoadInst *> SplitLoads;
3625  const DataLayout &DL = AI.getModule()->getDataLayout();
3626  for (LoadInst *LI : Loads) {
3627  SplitLoads.clear();
3628 
3629  IntegerType *Ty = cast<IntegerType>(LI->getType());
3630  uint64_t LoadSize = Ty->getBitWidth() / 8;
3631  assert(LoadSize > 0 && "Cannot have a zero-sized integer load!");
3632 
3633  auto &Offsets = SplitOffsetsMap[LI];
3634  assert(LoadSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&
3635  "Slice size should always match load size exactly!");
3636  uint64_t BaseOffset = Offsets.S->beginOffset();
3637  assert(BaseOffset + LoadSize > BaseOffset &&
3638  "Cannot represent alloca access size using 64-bit integers!");
3639 
3640  Instruction *BasePtr = cast<Instruction>(LI->getPointerOperand());
3641  IRB.SetInsertPoint(LI);
3642 
3643  DEBUG(dbgs() << " Splitting load: " << *LI << "\n");
3644 
3645  uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
3646  int Idx = 0, Size = Offsets.Splits.size();
3647  for (;;) {
3648  auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
3649  auto AS = LI->getPointerAddressSpace();
3650  auto *PartPtrTy = PartTy->getPointerTo(AS);
3651  LoadInst *PLoad = IRB.CreateAlignedLoad(
3652  getAdjustedPtr(IRB, DL, BasePtr,
3653  APInt(DL.getPointerSizeInBits(AS), PartOffset),
3654  PartPtrTy, BasePtr->getName() + "."),
3655  getAdjustedAlignment(LI, PartOffset, DL), /*IsVolatile*/ false,
3656  LI->getName());
3658 
3659  // Append this load onto the list of split loads so we can find it later
3660  // to rewrite the stores.
3661  SplitLoads.push_back(PLoad);
3662 
3663  // Now build a new slice for the alloca.
3664  NewSlices.push_back(
3665  Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
3666  &PLoad->getOperandUse(PLoad->getPointerOperandIndex()),
3667  /*IsSplittable*/ false));
3668  DEBUG(dbgs() << " new slice [" << NewSlices.back().beginOffset()
3669  << ", " << NewSlices.back().endOffset() << "): " << *PLoad
3670  << "\n");
3671 
3672  // See if we've handled all the splits.
3673  if (Idx >= Size)
3674  break;
3675 
3676  // Setup the next partition.
3677  PartOffset = Offsets.Splits[Idx];
3678  ++Idx;
3679  PartSize = (Idx < Size ? Offsets.Splits[Idx] : LoadSize) - PartOffset;
3680  }
3681 
3682  // Now that we have the split loads, do the slow walk over all uses of the
3683  // load and rewrite them as split stores, or save the split loads to use
3684  // below if the store is going to be split there anyways.
3685  bool DeferredStores = false;
3686  for (User *LU : LI->users()) {
3687  StoreInst *SI = cast<StoreInst>(LU);
3688  if (!Stores.empty() && SplitOffsetsMap.count(SI)) {
3689  DeferredStores = true;
3690  DEBUG(dbgs() << " Deferred splitting of store: " << *SI << "\n");
3691  continue;
3692  }
3693 
3694  Value *StoreBasePtr = SI->getPointerOperand();
3695  IRB.SetInsertPoint(SI);
3696 
3697  DEBUG(dbgs() << " Splitting store of load: " << *SI << "\n");
3698 
3699  for (int Idx = 0, Size = SplitLoads.size(); Idx < Size; ++Idx) {
3700  LoadInst *PLoad = SplitLoads[Idx];
3701  uint64_t PartOffset = Idx == 0 ? 0 : Offsets.Splits[Idx - 1];
3702  auto *PartPtrTy =
3703  PLoad->getType()->getPointerTo(SI->getPointerAddressSpace());
3704 
3705  auto AS = SI->getPointerAddressSpace();
3706  StoreInst *PStore = IRB.CreateAlignedStore(
3707  PLoad,
3708  getAdjustedPtr(IRB, DL, StoreBasePtr,
3709  APInt(DL.getPointerSizeInBits(AS), PartOffset),
3710  PartPtrTy, StoreBasePtr->getName() + "."),
3711  getAdjustedAlignment(SI, PartOffset, DL), /*IsVolatile*/ false);
3713  DEBUG(dbgs() << " +" << PartOffset << ":" << *PStore << "\n");
3714  }
3715 
3716  // We want to immediately iterate on any allocas impacted by splitting
3717  // this store, and we have to track any promotable alloca (indicated by
3718  // a direct store) as needing to be resplit because it is no longer
3719  // promotable.
3720  if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(StoreBasePtr)) {
3721  ResplitPromotableAllocas.insert(OtherAI);
3722  Worklist.insert(OtherAI);
3723  } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
3724  StoreBasePtr->stripInBoundsOffsets())) {
3725  Worklist.insert(OtherAI);
3726  }
3727 
3728  // Mark the original store as dead.
3729  DeadInsts.insert(SI);
3730  }
3731 
3732  // Save the split loads if there are deferred stores among the users.
3733  if (DeferredStores)
3734  SplitLoadsMap.insert(std::make_pair(LI, std::move(SplitLoads)));
3735 
3736  // Mark the original load as dead and kill the original slice.
3737  DeadInsts.insert(LI);
3738  Offsets.S->kill();
3739  }
3740 
3741  // Second, we rewrite all of the split stores. At this point, we know that
3742  // all loads from this alloca have been split already. For stores of such
3743  // loads, we can simply look up the pre-existing split loads. For stores of
3744  // other loads, we split those loads first and then write split stores of
3745  // them.
3746  for (StoreInst *SI : Stores) {
3747  auto *LI = cast<LoadInst>(SI->getValueOperand());
3748  IntegerType *Ty = cast<IntegerType>(LI->getType());
3749  uint64_t StoreSize = Ty->getBitWidth() / 8;
3750  assert(StoreSize > 0 && "Cannot have a zero-sized integer store!");
3751 
3752  auto &Offsets = SplitOffsetsMap[SI];
3753  assert(StoreSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&
3754  "Slice size should always match load size exactly!");
3755  uint64_t BaseOffset = Offsets.S->beginOffset();
3756  assert(BaseOffset + StoreSize > BaseOffset &&
3757  "Cannot represent alloca access size using 64-bit integers!");
3758 
3759  Value *LoadBasePtr = LI->getPointerOperand();
3760  Instruction *StoreBasePtr = cast<Instruction>(SI->getPointerOperand());
3761 
3762  DEBUG(dbgs() << " Splitting store: " << *SI << "\n");
3763 
3764  // Check whether we have an already split load.
3765  auto SplitLoadsMapI = SplitLoadsMap.find(LI);
3766  std::vector<LoadInst *> *SplitLoads = nullptr;
3767  if (SplitLoadsMapI != SplitLoadsMap.end()) {
3768  SplitLoads = &SplitLoadsMapI->second;
3769  assert(SplitLoads->size() == Offsets.Splits.size() + 1 &&
3770  "Too few split loads for the number of splits in the store!");
3771  } else {
3772  DEBUG(dbgs() << " of load: " << *LI << "\n");
3773  }
3774 
3775  uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
3776  int Idx = 0, Size = Offsets.Splits.size();
3777  for (;;) {
3778  auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
3779  auto *LoadPartPtrTy = PartTy->getPointerTo(LI->getPointerAddressSpace());
3780  auto *StorePartPtrTy = PartTy->getPointerTo(SI->getPointerAddressSpace());
3781 
3782  // Either lookup a split load or create one.
3783  LoadInst *PLoad;
3784  if (SplitLoads) {
3785  PLoad = (*SplitLoads)[Idx];
3786  } else {
3787  IRB.SetInsertPoint(LI);
3788  auto AS = LI->getPointerAddressSpace();
3789  PLoad = IRB.CreateAlignedLoad(
3790  getAdjustedPtr(IRB, DL, LoadBasePtr,
3791  APInt(DL.getPointerSizeInBits(AS), PartOffset),
3792  LoadPartPtrTy, LoadBasePtr->getName() + "."),
3793  getAdjustedAlignment(LI, PartOffset, DL), /*IsVolatile*/ false,
3794  LI->getName());
3795  }
3796 
3797  // And store this partition.
3798  IRB.SetInsertPoint(SI);
3799  auto AS = SI->getPointerAddressSpace();
3800  StoreInst *PStore = IRB.CreateAlignedStore(
3801  PLoad,
3802  getAdjustedPtr(IRB, DL, StoreBasePtr,
3803  APInt(DL.getPointerSizeInBits(AS), PartOffset),
3804  StorePartPtrTy, StoreBasePtr->getName() + "."),
3805  getAdjustedAlignment(SI, PartOffset, DL), /*IsVolatile*/ false);
3806 
3807  // Now build a new slice for the alloca.
3808  NewSlices.push_back(
3809  Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
3810  &PStore->getOperandUse(PStore->getPointerOperandIndex()),
3811  /*IsSplittable*/ false));
3812  DEBUG(dbgs() << " new slice [" << NewSlices.back().beginOffset()
3813  << ", " << NewSlices.back().endOffset() << "): " << *PStore
3814  << "\n");
3815  if (!SplitLoads) {
3816  DEBUG(dbgs() << " of split load: " << *PLoad << "\n");
3817  }
3818 
3819  // See if we've finished all the splits.
3820  if (Idx >= Size)
3821  break;
3822 
3823  // Setup the next partition.
3824  PartOffset = Offsets.Splits[Idx];
3825  ++Idx;
3826  PartSize = (Idx < Size ? Offsets.Splits[Idx] : StoreSize) - PartOffset;
3827  }
3828 
3829  // We want to immediately iterate on any allocas impacted by splitting
3830  // this load, which is only relevant if it isn't a load of this alloca and
3831  // thus we didn't already split the loads above. We also have to keep track
3832  // of any promotable allocas we split loads on as they can no longer be
3833  // promoted.
3834  if (!SplitLoads) {
3835  if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(LoadBasePtr)) {
3836  assert(OtherAI != &AI && "We can't re-split our own alloca!");
3837  ResplitPromotableAllocas.insert(OtherAI);
3838  Worklist.insert(OtherAI);
3839  } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
3840  LoadBasePtr->stripInBoundsOffsets())) {
3841  assert(OtherAI != &AI && "We can't re-split our own alloca!");
3842  Worklist.insert(OtherAI);
3843  }
3844  }
3845 
3846  // Mark the original store as dead now that we've split it up and kill its
3847  // slice. Note that we leave the original load in place unless this store
3848  // was its only use. It may in turn be split up if it is an alloca load
3849  // for some other alloca, but it may be a normal load. This may introduce
3850  // redundant loads, but where those can be merged the rest of the optimizer
3851  // should handle the merging, and this uncovers SSA splits which is more
3852  // important. In practice, the original loads will almost always be fully
3853  // split and removed eventually, and the splits will be merged by any
3854  // trivial CSE, including instcombine.
3855  if (LI->hasOneUse()) {
3856  assert(*LI->user_begin() == SI && "Single use isn't this store!");
3857  DeadInsts.insert(LI);
3858  }
3859  DeadInsts.insert(SI);
3860  Offsets.S->kill();
3861  }
3862 
3863  // Remove the killed slices that have ben pre-split.
3864  AS.erase(llvm::remove_if(AS, [](const Slice &S) { return S.isDead(); }),
3865  AS.end());
3866 
3867  // Insert our new slices. This will sort and merge them into the sorted
3868  // sequence.
3869  AS.insert(NewSlices);
3870 
3871  DEBUG(dbgs() << " Pre-split slices:\n");
3872 #ifndef NDEBUG
3873  for (auto I = AS.begin(), E = AS.end(); I != E; ++I)
3874  DEBUG(AS.print(dbgs(), I, " "));
3875 #endif
3876 
3877  // Finally, don't try to promote any allocas that new require re-splitting.
3878  // They have already been added to the worklist above.
3879  PromotableAllocas.erase(
3881  PromotableAllocas,
3882  [&](AllocaInst *AI) { return ResplitPromotableAllocas.count(AI); }),
3883  PromotableAllocas.end());
3884 
3885  return true;
3886 }
3887 
3888 /// \brief Rewrite an alloca partition's users.
3889 ///
3890 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3891 /// to rewrite uses of an alloca partition to be conducive for SSA value
3892 /// promotion. If the partition needs a new, more refined alloca, this will
3893 /// build that new alloca, preserving as much type information as possible, and
3894 /// rewrite the uses of the old alloca to point at the new one and have the
3895 /// appropriate new offsets. It also evaluates how successful the rewrite was
3896 /// at enabling promotion and if it was successful queues the alloca to be
3897 /// promoted.
3898 AllocaInst *SROA::rewritePartition(AllocaInst &AI, AllocaSlices &AS,
3899  Partition &P) {
3900  // Try to compute a friendly type for this partition of the alloca. This
3901  // won't always succeed, in which case we fall back to a legal integer type
3902  // or an i8 array of an appropriate size.
3903  Type *SliceTy = nullptr;
3904  const DataLayout &DL = AI.getModule()->getDataLayout();
3905  if (Type *CommonUseTy = findCommonType(P.begin(), P.end(), P.endOffset()))
3906  if (DL.getTypeAllocSize(CommonUseTy) >= P.size())
3907  SliceTy = CommonUseTy;
3908  if (!SliceTy)
3909  if (Type *TypePartitionTy = getTypePartition(DL, AI.getAllocatedType(),
3910  P.beginOffset(), P.size()))
3911  SliceTy = TypePartitionTy;
3912  if ((!SliceTy || (SliceTy->isArrayTy() &&
3913  SliceTy->getArrayElementType()->isIntegerTy())) &&
3914  DL.isLegalInteger(P.size() * 8))
3915  SliceTy = Type::getIntNTy(*C, P.size() * 8);
3916  if (!SliceTy)
3917  SliceTy = ArrayType::get(Type::getInt8Ty(*C), P.size());
3918  assert(DL.getTypeAllocSize(SliceTy) >= P.size());
3919 
3920  bool IsIntegerPromotable = isIntegerWideningViable(P, SliceTy, DL);
3921 
3922  VectorType *VecTy =
3923  IsIntegerPromotable ? nullptr : isVectorPromotionViable(P, DL);
3924  if (VecTy)
3925  SliceTy = VecTy;
3926 
3927  // Check for the case where we're going to rewrite to a new alloca of the
3928  // exact same type as the original, and with the same access offsets. In that
3929  // case, re-use the existing alloca, but still run through the rewriter to
3930  // perform phi and select speculation.
3931  AllocaInst *NewAI;
3932  if (SliceTy == AI.getAllocatedType()) {
3933  assert(P.beginOffset() == 0 &&
3934  "Non-zero begin offset but same alloca type");
3935  NewAI = &AI;
3936  // FIXME: We should be able to bail at this point with "nothing changed".
3937  // FIXME: We might want to defer PHI speculation until after here.
3938  // FIXME: return nullptr;
3939  } else {
3940  unsigned Alignment = AI.getAlignment();
3941  if (!Alignment) {
3942  // The minimum alignment which users can rely on when the explicit
3943  // alignment is omitted or zero is that required by the ABI for this
3944  // type.
3945  Alignment = DL.getABITypeAlignment(AI.getAllocatedType());
3946  }
3947  Alignment = MinAlign(Alignment, P.beginOffset());
3948  // If we will get at least this much alignment from the type alone, leave
3949  // the alloca's alignment unconstrained.
3950  if (Alignment <= DL.getABITypeAlignment(SliceTy))
3951  Alignment = 0;
3952  NewAI = new AllocaInst(
3953  SliceTy, AI.getType()->getAddressSpace(), nullptr, Alignment,
3954  AI.getName() + ".sroa." + Twine(P.begin() - AS.begin()), &AI);
3955  ++NumNewAllocas;
3956  }
3957 
3958  DEBUG(dbgs() << "Rewriting alloca partition "
3959  << "[" << P.beginOffset() << "," << P.endOffset()
3960  << ") to: " << *NewAI << "\n");
3961 
3962  // Track the high watermark on the worklist as it is only relevant for
3963  // promoted allocas. We will reset it to this point if the alloca is not in
3964  // fact scheduled for promotion.
3965  unsigned PPWOldSize = PostPromotionWorklist.size();
3966  unsigned NumUses = 0;
3968  SmallSetVector<SelectInst *, 8> SelectUsers;
3969 
3970  AllocaSliceRewriter Rewriter(DL, AS, *this, AI, *NewAI, P.beginOffset(),
3971  P.endOffset(), IsIntegerPromotable, VecTy,
3972  PHIUsers, SelectUsers);
3973  bool Promotable = true;
3974  for (Slice *S : P.splitSliceTails()) {
3975  Promotable &= Rewriter.visit(S);
3976  ++NumUses;
3977  }
3978  for (Slice &S : P) {
3979  Promotable &= Rewriter.visit(&S);
3980  ++NumUses;
3981  }
3982 
3983  NumAllocaPartitionUses += NumUses;
3984  MaxUsesPerAllocaPartition.updateMax(NumUses);
3985 
3986  // Now that we've processed all the slices in the new partition, check if any
3987  // PHIs or Selects would block promotion.
3988  for (PHINode *PHI : PHIUsers)
3989  if (!isSafePHIToSpeculate(*PHI)) {
3990  Promotable = false;
3991  PHIUsers.clear();
3992  SelectUsers.clear();
3993  break;
3994  }
3995 
3996  for (SelectInst *Sel : SelectUsers)
3997  if (!isSafeSelectToSpeculate(*Sel)) {
3998  Promotable = false;
3999  PHIUsers.clear();
4000  SelectUsers.clear();
4001  break;
4002  }
4003 
4004  if (Promotable) {
4005  if (PHIUsers.empty() && SelectUsers.empty()) {
4006  // Promote the alloca.
4007  PromotableAllocas.push_back(NewAI);
4008  } else {
4009  // If we have either PHIs or Selects to speculate, add them to those
4010  // worklists and re-queue the new alloca so that we promote in on the
4011  // next iteration.
4012  for (PHINode *PHIUser : PHIUsers)
4013  SpeculatablePHIs.insert(PHIUser);
4014  for (SelectInst *SelectUser : SelectUsers)
4015  SpeculatableSelects.insert(SelectUser);
4016  Worklist.insert(NewAI);
4017  }
4018  } else {
4019  // Drop any post-promotion work items if promotion didn't happen.
4020  while (PostPromotionWorklist.size() > PPWOldSize)
4021  PostPromotionWorklist.pop_back();
4022 
4023  // We couldn't promote and we didn't create a new partition, nothing
4024  // happened.
4025  if (NewAI == &AI)
4026  return nullptr;
4027 
4028  // If we can't promote the alloca, iterate on it to check for new
4029  // refinements exposed by splitting the current alloca. Don't iterate on an
4030  // alloca which didn't actually change and didn't get promoted.
4031  Worklist.insert(NewAI);
4032  }
4033 
4034  return NewAI;
4035 }
4036 
4037 /// \brief Walks the slices of an alloca and form partitions based on them,
4038 /// rewriting each of their uses.
4039 bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &AS) {
4040  if (AS.begin() == AS.end())
4041  return false;
4042 
4043  unsigned NumPartitions = 0;
4044  bool Changed = false;
4045  const DataLayout &DL = AI.getModule()->getDataLayout();
4046 
4047  // First try to pre-split loads and stores.
4048  Changed |= presplitLoadsAndStores(AI, AS);
4049 
4050  // Now that we have identified any pre-splitting opportunities, mark any
4051  // splittable (non-whole-alloca) loads and stores as unsplittable. If we fail
4052  // to split these during pre-splitting, we want to force them to be
4053  // rewritten into a partition.
4054  bool IsSorted = true;
4055  for (Slice &S : AS) {
4056  if (!S.isSplittable())
4057  continue;
4058  // FIXME: We currently leave whole-alloca splittable loads and stores. This
4059  // used to be the only splittable loads and stores and we need to be
4060  // confident that the above handling of splittable loads and stores is
4061  // completely sufficient before we forcibly disable the remaining handling.
4062  if (S.beginOffset() == 0 &&
4063  S.endOffset() >= DL.getTypeAllocSize(AI.getAllocatedType()))
4064  continue;
4065  if (isa<LoadInst>(S.getUse()->getUser()) ||
4066  isa<StoreInst>(S.getUse()->getUser())) {
4067  S.makeUnsplittable();
4068  IsSorted = false;
4069  }
4070  }
4071  if (!IsSorted)
4072  std::sort(AS.begin(), AS.end());
4073 
4074  /// Describes the allocas introduced by rewritePartition in order to migrate
4075  /// the debug info.
4076  struct Fragment {
4077  AllocaInst *Alloca;
4078  uint64_t Offset;
4079  uint64_t Size;
4080  Fragment(AllocaInst *AI, uint64_t O, uint64_t S)
4081  : Alloca(AI), Offset(O), Size(S) {}
4082  };
4083  SmallVector<Fragment, 4> Fragments;
4084 
4085  // Rewrite each partition.
4086  for (auto &P : AS.partitions()) {
4087  if (AllocaInst *NewAI = rewritePartition(AI, AS, P)) {
4088  Changed = true;
4089  if (NewAI != &AI) {
4090  uint64_t SizeOfByte = 8;
4091  uint64_t AllocaSize = DL.getTypeSizeInBits(NewAI->getAllocatedType());
4092  // Don't include any padding.
4093  uint64_t Size = std::min(AllocaSize, P.size() * SizeOfByte);
4094  Fragments.push_back(Fragment(NewAI, P.beginOffset() * SizeOfByte, Size));
4095  }
4096  }
4097  ++NumPartitions;
4098  }
4099 
4100  NumAllocaPartitions += NumPartitions;
4101  MaxPartitionsPerAlloca.updateMax(NumPartitions);
4102 
4103  // Migrate debug information from the old alloca to the new alloca(s)
4104  // and the individual partitions.
4106  if (!DbgDeclares.empty()) {
4107  auto *Var = DbgDeclares.front()->getVariable();
4108  auto *Expr = DbgDeclares.front()->getExpression();
4109  DIBuilder DIB(*AI.getModule(), /*AllowUnresolved*/ false);
4110  uint64_t AllocaSize = DL.getTypeSizeInBits(AI.getAllocatedType());
4111  for (auto Fragment : Fragments) {
4112  // Create a fragment expression describing the new partition or reuse AI's
4113  // expression if there is only one partition.
4114  auto *FragmentExpr = Expr;
4115  if (Fragment.Size < AllocaSize || Expr->isFragment()) {
4116  // If this alloca is already a scalar replacement of a larger aggregate,
4117  // Fragment.Offset describes the offset inside the scalar.
4118  auto ExprFragment = Expr->getFragmentInfo();
4119  uint64_t Offset = ExprFragment ? ExprFragment->OffsetInBits : 0;
4120  uint64_t Start = Offset + Fragment.Offset;
4121  uint64_t Size = Fragment.Size;
4122  if (ExprFragment) {
4123  uint64_t AbsEnd =
4124  ExprFragment->OffsetInBits + ExprFragment->SizeInBits;
4125  if (Start >= AbsEnd)
4126  // No need to describe a SROAed padding.
4127  continue;
4128  Size = std::min(Size, AbsEnd - Start);
4129  }
4130  // The new, smaller fragment is stenciled out from the old fragment.
4131  if (auto OrigFragment = FragmentExpr->getFragmentInfo()) {
4132  assert(Start >= OrigFragment->OffsetInBits &&
4133  "new fragment is outside of original fragment");
4134  Start -= OrigFragment->OffsetInBits;
4135  }
4136  FragmentExpr =
4137  DIExpression::createFragmentExpression(Expr, Start, Size);
4138  }
4139 
4140  // Remove any existing intrinsics describing the same alloca.
4141  for (DbgInfoIntrinsic *OldDII : FindDbgAddrUses(Fragment.Alloca))
4142  OldDII->eraseFromParent();
4143 
4144  DIB.insertDeclare(Fragment.Alloca, Var, FragmentExpr,
4145  DbgDeclares.front()->getDebugLoc(), &AI);
4146  }
4147  }
4148  return Changed;
4149 }
4150 
4151 /// \brief Clobber a use with undef, deleting the used value if it becomes dead.
4152 void SROA::clobberUse(Use &U) {
4153  Value *OldV = U;
4154  // Replace the use with an undef value.
4155  U = UndefValue::get(OldV->getType());
4156 
4157  // Check for this making an instruction dead. We have to garbage collect
4158  // all the dead instructions to ensure the uses of any alloca end up being
4159  // minimal.
4160  if (Instruction *OldI = dyn_cast<Instruction>(OldV))
4161  if (isInstructionTriviallyDead(OldI)) {
4162  DeadInsts.insert(OldI);
4163  }
4164 }
4165 
4166 /// \brief Analyze an alloca for SROA.
4167 ///
4168 /// This analyzes the alloca to ensure we can reason about it, builds
4169 /// the slices of the alloca, and then hands it off to be split and
4170 /// rewritten as needed.
4171 bool SROA::runOnAlloca(AllocaInst &AI) {
4172  DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
4173  ++NumAllocasAnalyzed;
4174 
4175  // Special case dead allocas, as they're trivial.
4176  if (AI.use_empty()) {
4177  AI.eraseFromParent();
4178  return true;
4179  }
4180  const DataLayout &DL = AI.getModule()->getDataLayout();
4181 
4182  // Skip alloca forms that this analysis can't handle.
4183  if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
4184  DL.getTypeAllocSize(AI.getAllocatedType()) == 0)
4185  return false;
4186 
4187  bool Changed = false;
4188 
4189  // First, split any FCA loads and stores touching this alloca to promote
4190  // better splitting and promotion opportunities.
4191  AggLoadStoreRewriter AggRewriter;
4192  Changed |= AggRewriter.rewrite(AI);
4193 
4194  // Build the slices using a recursive instruction-visiting builder.
4195  AllocaSlices AS(DL, AI);
4196  DEBUG(AS.print(dbgs()));
4197  if (AS.isEscaped())
4198  return Changed;
4199 
4200  // Delete all the dead users of this alloca before splitting and rewriting it.
4201  for (Instruction *DeadUser : AS.getDeadUsers()) {
4202  // Free up everything used by this instruction.
4203  for (Use &DeadOp : DeadUser->operands())
4204  clobberUse(DeadOp);
4205 
4206  // Now replace the uses of this instruction.
4207  DeadUser->replaceAllUsesWith(UndefValue::get(DeadUser->getType()));
4208 
4209  // And mark it for deletion.
4210  DeadInsts.insert(DeadUser);
4211  Changed = true;
4212  }
4213  for (Use *DeadOp : AS.getDeadOperands()) {
4214  clobberUse(*DeadOp);
4215  Changed = true;
4216  }
4217 
4218  // No slices to split. Leave the dead alloca for a later pass to clean up.
4219  if (AS.begin() == AS.end())
4220  return Changed;
4221 
4222  Changed |= splitAlloca(AI, AS);
4223 
4224  DEBUG(dbgs() << " Speculating PHIs\n");
4225  while (!SpeculatablePHIs.empty())
4226  speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
4227 
4228  DEBUG(dbgs() << " Speculating Selects\n");
4229  while (!SpeculatableSelects.empty())
4230  speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
4231 
4232  return Changed;
4233 }
4234 
4235 /// \brief Delete the dead instructions accumulated in this run.
4236 ///
4237 /// Recursively deletes the dead instructions we've accumulated. This is done
4238 /// at the very end to maximize locality of the recursive delete and to
4239 /// minimize the problems of invalidated instruction pointers as such pointers
4240 /// are used heavily in the intermediate stages of the algorithm.
4241 ///
4242 /// We also record the alloca instructions deleted here so that they aren't
4243 /// subsequently handed to mem2reg to promote.
4244 void SROA::deleteDeadInstructions(
4245  SmallPtrSetImpl<AllocaInst *> &DeletedAllocas) {
4246  while (!DeadInsts.empty()) {
4247  Instruction *I = DeadInsts.pop_back_val();
4248  DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
4249 
4250  // If the instruction is an alloca, find the possible dbg.declare connected
4251  // to it, and remove it too. We must do this before calling RAUW or we will
4252  // not be able to find it.
4253  if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
4254  DeletedAllocas.insert(AI);
4255  for (DbgInfoIntrinsic *OldDII : FindDbgAddrUses(AI))
4256  OldDII->eraseFromParent();
4257  }
4258 
4260 
4261  for (Use &Operand : I->operands())
4262  if (Instruction *U = dyn_cast<Instruction>(Operand)) {
4263  // Zero out the operand and see if it becomes trivially dead.
4264  Operand = nullptr;
4266  DeadInsts.insert(U);
4267  }
4268 
4269  ++NumDeleted;
4270  I->eraseFromParent();
4271  }
4272 }
4273 
4274 /// \brief Promote the allocas, using the best available technique.
4275 ///
4276 /// This attempts to promote whatever allocas have been identified as viable in
4277 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
4278 /// This function returns whether any promotion occurred.
4279 bool SROA::promoteAllocas(Function &F) {
4280  if (PromotableAllocas.empty())
4281  return false;
4282 
4283  NumPromoted += PromotableAllocas.size();
4284 
4285  DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
4286  PromoteMemToReg(PromotableAllocas, *DT, AC);
4287  PromotableAllocas.clear();
4288  return true;
4289 }
4290 
4291 PreservedAnalyses SROA::runImpl(Function &F, DominatorTree &RunDT,
4292  AssumptionCache &RunAC) {
4293  DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
4294  C = &F.getContext();
4295  DT = &RunDT;
4296  AC = &RunAC;
4297 
4298  BasicBlock &EntryBB = F.getEntryBlock();
4299  for (BasicBlock::iterator I = EntryBB.begin(), E = std::prev(EntryBB.end());
4300  I != E; ++I) {
4301  if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
4302  Worklist.insert(AI);
4303  }
4304 
4305  bool Changed = false;
4306  // A set of deleted alloca instruction pointers which should be removed from
4307  // the list of promotable allocas.
4308  SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
4309 
4310  do {
4311  while (!Worklist.empty()) {
4312  Changed |= runOnAlloca(*Worklist.pop_back_val());
4313  deleteDeadInstructions(DeletedAllocas);
4314 
4315  // Remove the deleted allocas from various lists so that we don't try to
4316  // continue processing them.
4317  if (!DeletedAllocas.empty()) {
4318  auto IsInSet = [&](AllocaInst *AI) { return DeletedAllocas.count(AI); };
4319  Worklist.remove_if(IsInSet);
4320  PostPromotionWorklist.remove_if(IsInSet);
4321  PromotableAllocas.erase(llvm::remove_if(PromotableAllocas, IsInSet),
4322  PromotableAllocas.end());
4323  DeletedAllocas.clear();
4324  }
4325  }
4326 
4327  Changed |= promoteAllocas(F);
4328 
4329  Worklist = PostPromotionWorklist;
4330  PostPromotionWorklist.clear();
4331  } while (!Worklist.empty());
4332 
4333  if (!Changed)
4334  return PreservedAnalyses::all();
4335 
4336  PreservedAnalyses PA;
4337  PA.preserveSet<CFGAnalyses>();
4338  PA.preserve<GlobalsAA>();
4339  return PA;
4340 }
4341 
4343  return runImpl(F, AM.getResult<DominatorTreeAnalysis>(F),
4344  AM.getResult<AssumptionAnalysis>(F));
4345 }
4346 
4347 /// A legacy pass for the legacy pass manager that wraps the \c SROA pass.
4348 ///
4349 /// This is in the llvm namespace purely to allow it to be a friend of the \c
4350 /// SROA pass.
4352  /// The SROA implementation.
4353  SROA Impl;
4354 
4355 public:
4356  static char ID;
4357 
4360  }
4361 
4362  bool runOnFunction(Function &F) override {
4363  if (skipFunction(F))
4364  return false;
4365 
4366  auto PA = Impl.runImpl(
4367  F, getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
4368  getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F));
4369  return !PA.areAllPreserved();
4370  }
4371 
4372  void getAnalysisUsage(AnalysisUsage &AU) const override {
4376  AU.setPreservesCFG();
4377  }
4378 
4379  StringRef getPassName() const override { return "SROA"; }
4380 };
4381 
4382 char SROALegacyPass::ID = 0;
4383 
4385 
4387  "Scalar Replacement Of Aggregates", false, false)
4391  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:1538
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:1836
iterator end() const
Definition: SROA.cpp:415
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:1424
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:435
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
static DIExpression * createFragmentExpression(const DIExpression *Exp, unsigned OffsetInBits, unsigned SizeInBits)
Create a DIExpression to describe one part of an aggregate variable that is fragmented across multipl...
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:386
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.
EltTy front() const
bool isSized(SmallPtrSetImpl< Type *> *Visited=nullptr) const
Return true if it makes sense to take the size of this type.
Definition: Type.h:262
iterator begin() const
Definition: ArrayRef.h:137
LLVM_ATTRIBUTE_ALWAYS_INLINE size_type size() const
Definition: SmallVector.h:136
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:856
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.
This provides a very simple, boring adaptor for a begin and end iterator into a range type...
void erase(iterator Start, iterator Stop)
Erase a range of slices.
Definition: SROA.cpp:264
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:63
This class represents a function call, abstracting a target machine&#39;s calling convention.
static Value * extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V, IntegerType *Ty, uint64_t Offset, const Twine &Name)
Definition: SROA.cpp:2066
This file contains the declarations for metadata subclasses.
Representation of the alloca slices.
Definition: SROA.cpp:237
An immutable pass that tracks lazily created AssumptionCache objects.
Instruction * getAbortingInst() const
Get the instruction causing the visit to abort.
Definition: PtrUseVisitor.h:84
SyncScope::ID getSyncScopeID() const
Returns the synchronization scope ID of this store instruction.
Definition: Instructions.h: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:271
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:910
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:4390
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:816
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:238
F(f)
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
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
This defines the Use class.
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:246
TinyPtrVector - This class is specialized for cases where there are normally 0 or 1 element in a vect...
Definition: TinyPtrVector.h:31
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:1633
Builder for the alloca slices.
Definition: SROA.cpp:653
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:336
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:191
void setAtomic(AtomicOrdering Ordering, SyncScope::ID SSID=SyncScope::System)
Sets the ordering constraint and the synchronization scope ID of this load instruction.
Definition: Instructions.h: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:1760
PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM)
Run the pass over the function.
Definition: SROA.cpp:4342
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
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:3296
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:423
bool isAborted() const
Did we abort the visit early?
Definition: PtrUseVisitor.h:76
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:1710
A partition of the slices.
Definition: SROA.cpp:361
void getAnalysisUsage(AnalysisUsage &AU) const override
getAnalysisUsage - This function should be overriden by passes that need analysis information to do t...
Definition: SROA.cpp:4372
uint64_t getNumElements() const
Definition: DerivedTypes.h:359
static StructType * get(LLVMContext &Context, ArrayRef< Type *> Elements, bool isPacked=false)
This static method is the primary way to create a literal StructType.
Definition: Type.cpp:336
This file implements a class to represent arbitrary precision integral constant values and operations...
bool runOnFunction(Function &F) override
runOnFunction - Virtual method overriden by subclasses to do the per-function processing of the pass...
Definition: SROA.cpp:4362
This file provides a collection of visitors which walk the (instruction) uses of a pointer...
static Constant * getZExt(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp: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:2285
A base class for visitors over the uses of a pointer value.
Maximum number of bits that can be specified.
Definition: DerivedTypes.h:52
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:1869
bool isInBounds() const
Determine whether the GEP has the inbounds flag.
SmallVectorImpl< Slice >::iterator iterator
Support for iterating over the slices.
Definition: SROA.cpp:250
const_iterator end() const
Definition: SROA.cpp:260
IntType getInt() const
A legacy pass for the legacy pass manager that wraps the SROA pass.
Definition: SROA.cpp:4351
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:194
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:878
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
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:140
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:1659
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:146
const BasicBlock & getEntryBlock() const
Definition: Function.h:572
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:4390
iterator_range< partition_iterator > partitions()
#define P(N)
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:406
uint64_t getZExtValue() const
Return the constant as a 64-bit unsigned integer value after it has been zero extended as appropriate...
Definition: Constants.h:149
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
void dump(const SparseBitVector< ElementSize > &LHS, raw_ostream &out)
Subclasses of this class are all able to terminate a basic block.
Definition: InstrTypes.h: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
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
This is an important base class in LLVM.
Definition: Constant.h:42
LLVM_NODISCARD bool empty() const
Definition: SmallPtrSet.h:91
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:293
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:1292
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:363
bool mayHaveSideEffects() const
Return true if the instruction may have side effects.
Definition: Instruction.h:504
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:1498
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.
DILocalVariable * getVariable() const
Definition: IntrinsicInst.h:80
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:823
#define LLVM_ATTRIBUTE_UNUSED
Definition: Compiler.h:144
Analysis pass providing a never-invalidated alias analysis result.
void initializeSROALegacyPassPass(PassRegistry &)
sroa
Definition: SROA.cpp:4390
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:285
bool empty() const
DIExpression * getExpression() const
Definition: IntrinsicInst.h:84
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:374
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:2256
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:1380
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:3258
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
This is the shared class of boolean and integer constants.
Definition: Constants.h:84
iterator end()
Definition: BasicBlock.h:254
AllocaSlices(const DataLayout &DL, AllocaInst &AI)
Construct the slices of a particular alloca.
Module.h This file contains the declarations for the Module class.
Instruction * user_back()
Specialize the methods defined in Value, as we know that an instruction can only be used by other ins...
Definition: Instruction.h:63
partition_iterator & operator++()
Definition: SROA.cpp:608
static Value * insertVector(IRBuilderTy &IRB, Value *Old, Value *V, unsigned BeginIndex, const Twine &Name)
Definition: SROA.cpp:2146
This is the common base class for debug info intrinsics.
Definition: IntrinsicInst.h:67
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:414
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:385
uint64_t getSizeInBytes() const
Definition: DataLayout.h:501
SmallVectorImpl< Slice >::const_iterator const_iterator
Definition: SROA.cpp:256
static IntegerType * getIntNTy(LLVMContext &C, unsigned N)
Definition: Type.cpp:180
SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &AS)
Definition: SROA.cpp:669
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:285
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:1944
iterator_range< user_iterator > users()
Definition: Value.h:395
TinyPtrVector< DbgInfoIntrinsic * > FindDbgAddrUses(Value *V)
Finds all intrinsics declaring local variables as living in the memory that &#39;V&#39; points to...
Definition: Local.cpp:1238
#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:2089
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:259
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:1105
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:641
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:221
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:284
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:4379
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:2199
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:1314
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
#define N
constexpr char Align[]
Key for Kernel::Arg::Metadata::mAlign.
iterator end()
Definition: DenseMap.h:79
FunctionPass * createSROAPass()
Definition: SROA.cpp:4384
void setSource(Value *Ptr)
uint64_t endOffset() const
The end offset of this partition.
Definition: SROA.cpp:391
Instruction * getEscapingInst() const
Get the instruction causing the pointer to escape.
Definition: PtrUseVisitor.h:89
static ArrayType * get(Type *ElementType, uint64_t NumElements)
This static method is the primary way to construct an ArrayType.
Definition: Type.cpp:568
LLVM_NODISCARD std::enable_if<!is_simple_type< Y >::value, typename cast_retty< X, const Y >::ret_type >::type dyn_cast(const Y &Val)
Definition: Casting.h:323
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:2120
size_type count(const_arg_type_t< KeyT > Val) const
Return 1 if the specified key is in the map, 0 otherwise.
Definition: DenseMap.h:141
unsigned getAlignment() const
std::string str() const
Return the twine contents as a std::string.
Definition: Twine.cpp:17
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:1357
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:1172
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
static void speculatePHINodeLoads(PHINode &PN)
Definition: SROA.cpp:1236
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:66
Type * getElementType() const
Definition: DerivedTypes.h:360
static cl::opt< bool > SROARandomShuffleSlices("sroa-random-shuffle-slices", cl::init(false), cl::Hidden)
Hidden option to enable randomly shuffling the slices to help uncover instability in their order...
uint64_t size() const
The size of the partition.
Definition: SROA.cpp:396
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:403
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:267
void sort(Policy policy, RandomAccessIterator Start, RandomAccessIterator End, const Comparator &Comp=Comparator())
Definition: Parallel.h:199
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:2024
bool operator==(uint64_t V1, const APInt &V2)
Definition: APInt.h:1946
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
This header defines various interfaces for pass management in LLVM.
bool isBigEndian() const
Definition: DataLayout.h:217
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
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
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:628
ArrayRef< Instruction * > getDeadUsers() const
Access the dead users for this alloca.
Definition: SROA.cpp:285
bool isEscaped() const
Is the pointer escaped at some point?
Definition: PtrUseVisitor.h:79
bool operator==(const partition_iterator &RHS) const
Definition: SROA.cpp:588