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
26 #include "llvm/Transforms/Scalar/SROA.h"
27 #include "llvm/ADT/STLExtras.h"
28 #include "llvm/ADT/SmallVector.h"
29 #include "llvm/ADT/Statistic.h"
30 #include "llvm/Analysis/AssumptionCache.h"
31 #include "llvm/Analysis/GlobalsModRef.h"
32 #include "llvm/Analysis/Loads.h"
33 #include "llvm/Analysis/PtrUseVisitor.h"
34 #include "llvm/Analysis/ValueTracking.h"
35 #include "llvm/IR/Constants.h"
36 #include "llvm/IR/DIBuilder.h"
37 #include "llvm/IR/DataLayout.h"
38 #include "llvm/IR/DebugInfo.h"
39 #include "llvm/IR/DerivedTypes.h"
40 #include "llvm/IR/IRBuilder.h"
41 #include "llvm/IR/InstVisitor.h"
42 #include "llvm/IR/Instructions.h"
43 #include "llvm/IR/IntrinsicInst.h"
44 #include "llvm/IR/LLVMContext.h"
45 #include "llvm/IR/Operator.h"
46 #include "llvm/Pass.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/Compiler.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/ErrorHandling.h"
51 #include "llvm/Support/MathExtras.h"
52 #include "llvm/Support/TimeValue.h"
53 #include "llvm/Support/raw_ostream.h"
54 #include "llvm/Transforms/Scalar.h"
55 #include "llvm/Transforms/Utils/Local.h"
56 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
57
58 #ifndef NDEBUG
59 // We only use this for a debug check.
60 #include <random>
61 #endif
62
63 using namespace llvm;
64 using namespace llvm::sroa;
65
66 #define DEBUG_TYPE "sroa"
67
68 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
69 STATISTIC(NumAllocaPartitions, "Number of alloca partitions formed");
70 STATISTIC(MaxPartitionsPerAlloca, "Maximum number of partitions per alloca");
71 STATISTIC(NumAllocaPartitionUses, "Number of alloca partition uses rewritten");
72 STATISTIC(MaxUsesPerAllocaPartition, "Maximum number of uses of a partition");
73 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
74 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
75 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
76 STATISTIC(NumDeleted, "Number of instructions deleted");
77 STATISTIC(NumVectorized, "Number of vectorized aggregates");
78
79 /// Hidden option to enable randomly shuffling the slices to help uncover
80 /// instability in their order.
81 static cl::opt<bool> SROARandomShuffleSlices("sroa-random-shuffle-slices",
82 cl::init(false), cl::Hidden);
83
84 /// Hidden option to experiment with completely strict handling of inbounds
85 /// GEPs.
86 static cl::opt<bool> SROAStrictInbounds("sroa-strict-inbounds", cl::init(false),
87 cl::Hidden);
88
89 namespace {
90 /// \brief A custom IRBuilder inserter which prefixes all names, but only in
91 /// Assert builds.
92 class IRBuilderPrefixedInserter : public IRBuilderDefaultInserter {
93 std::string Prefix;
getNameWithPrefix(const Twine & Name) const94 const Twine getNameWithPrefix(const Twine &Name) const {
95 return Name.isTriviallyEmpty() ? Name : Prefix + Name;
96 }
97
98 public:
SetNamePrefix(const Twine & P)99 void SetNamePrefix(const Twine &P) { Prefix = P.str(); }
100
101 protected:
InsertHelper(Instruction * I,const Twine & Name,BasicBlock * BB,BasicBlock::iterator InsertPt) const102 void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
103 BasicBlock::iterator InsertPt) const {
104 IRBuilderDefaultInserter::InsertHelper(I, getNameWithPrefix(Name), BB,
105 InsertPt);
106 }
107 };
108
109 /// \brief Provide a typedef for IRBuilder that drops names in release builds.
110 using IRBuilderTy = llvm::IRBuilder<ConstantFolder, IRBuilderPrefixedInserter>;
111 }
112
113 namespace {
114 /// \brief A used slice of an alloca.
115 ///
116 /// This structure represents a slice of an alloca used by some instruction. It
117 /// stores both the begin and end offsets of this use, a pointer to the use
118 /// itself, and a flag indicating whether we can classify the use as splittable
119 /// or not when forming partitions of the alloca.
120 class Slice {
121 /// \brief The beginning offset of the range.
122 uint64_t BeginOffset;
123
124 /// \brief The ending offset, not included in the range.
125 uint64_t EndOffset;
126
127 /// \brief Storage for both the use of this slice and whether it can be
128 /// split.
129 PointerIntPair<Use *, 1, bool> UseAndIsSplittable;
130
131 public:
Slice()132 Slice() : BeginOffset(), EndOffset() {}
Slice(uint64_t BeginOffset,uint64_t EndOffset,Use * U,bool IsSplittable)133 Slice(uint64_t BeginOffset, uint64_t EndOffset, Use *U, bool IsSplittable)
134 : BeginOffset(BeginOffset), EndOffset(EndOffset),
135 UseAndIsSplittable(U, IsSplittable) {}
136
beginOffset() const137 uint64_t beginOffset() const { return BeginOffset; }
endOffset() const138 uint64_t endOffset() const { return EndOffset; }
139
isSplittable() const140 bool isSplittable() const { return UseAndIsSplittable.getInt(); }
makeUnsplittable()141 void makeUnsplittable() { UseAndIsSplittable.setInt(false); }
142
getUse() const143 Use *getUse() const { return UseAndIsSplittable.getPointer(); }
144
isDead() const145 bool isDead() const { return getUse() == nullptr; }
kill()146 void kill() { UseAndIsSplittable.setPointer(nullptr); }
147
148 /// \brief Support for ordering ranges.
149 ///
150 /// This provides an ordering over ranges such that start offsets are
151 /// always increasing, and within equal start offsets, the end offsets are
152 /// decreasing. Thus the spanning range comes first in a cluster with the
153 /// same start position.
operator <(const Slice & RHS) const154 bool operator<(const Slice &RHS) const {
155 if (beginOffset() < RHS.beginOffset())
156 return true;
157 if (beginOffset() > RHS.beginOffset())
158 return false;
159 if (isSplittable() != RHS.isSplittable())
160 return !isSplittable();
161 if (endOffset() > RHS.endOffset())
162 return true;
163 return false;
164 }
165
166 /// \brief Support comparison with a single offset to allow binary searches.
operator <(const Slice & LHS,uint64_t RHSOffset)167 friend LLVM_ATTRIBUTE_UNUSED bool operator<(const Slice &LHS,
168 uint64_t RHSOffset) {
169 return LHS.beginOffset() < RHSOffset;
170 }
operator <(uint64_t LHSOffset,const Slice & RHS)171 friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
172 const Slice &RHS) {
173 return LHSOffset < RHS.beginOffset();
174 }
175
operator ==(const Slice & RHS) const176 bool operator==(const Slice &RHS) const {
177 return isSplittable() == RHS.isSplittable() &&
178 beginOffset() == RHS.beginOffset() && endOffset() == RHS.endOffset();
179 }
operator !=(const Slice & RHS) const180 bool operator!=(const Slice &RHS) const { return !operator==(RHS); }
181 };
182 } // end anonymous namespace
183
184 namespace llvm {
185 template <typename T> struct isPodLike;
186 template <> struct isPodLike<Slice> { static const bool value = true; };
187 }
188
189 /// \brief Representation of the alloca slices.
190 ///
191 /// This class represents the slices of an alloca which are formed by its
192 /// various uses. If a pointer escapes, we can't fully build a representation
193 /// for the slices used and we reflect that in this structure. The uses are
194 /// stored, sorted by increasing beginning offset and with unsplittable slices
195 /// starting at a particular offset before splittable slices.
196 class llvm::sroa::AllocaSlices {
197 public:
198 /// \brief Construct the slices of a particular alloca.
199 AllocaSlices(const DataLayout &DL, AllocaInst &AI);
200
201 /// \brief Test whether a pointer to the allocation escapes our analysis.
202 ///
203 /// If this is true, the slices are never fully built and should be
204 /// ignored.
isEscaped() const205 bool isEscaped() const { return PointerEscapingInstr; }
206
207 /// \brief Support for iterating over the slices.
208 /// @{
209 typedef SmallVectorImpl<Slice>::iterator iterator;
210 typedef iterator_range<iterator> range;
begin()211 iterator begin() { return Slices.begin(); }
end()212 iterator end() { return Slices.end(); }
213
214 typedef SmallVectorImpl<Slice>::const_iterator const_iterator;
215 typedef iterator_range<const_iterator> const_range;
begin() const216 const_iterator begin() const { return Slices.begin(); }
end() const217 const_iterator end() const { return Slices.end(); }
218 /// @}
219
220 /// \brief Erase a range of slices.
erase(iterator Start,iterator Stop)221 void erase(iterator Start, iterator Stop) { Slices.erase(Start, Stop); }
222
223 /// \brief Insert new slices for this alloca.
224 ///
225 /// This moves the slices into the alloca's slices collection, and re-sorts
226 /// everything so that the usual ordering properties of the alloca's slices
227 /// hold.
insert(ArrayRef<Slice> NewSlices)228 void insert(ArrayRef<Slice> NewSlices) {
229 int OldSize = Slices.size();
230 Slices.append(NewSlices.begin(), NewSlices.end());
231 auto SliceI = Slices.begin() + OldSize;
232 std::sort(SliceI, Slices.end());
233 std::inplace_merge(Slices.begin(), SliceI, Slices.end());
234 }
235
236 // Forward declare the iterator and range accessor for walking the
237 // partitions.
238 class partition_iterator;
239 iterator_range<partition_iterator> partitions();
240
241 /// \brief Access the dead users for this alloca.
getDeadUsers() const242 ArrayRef<Instruction *> getDeadUsers() const { return DeadUsers; }
243
244 /// \brief Access the dead operands referring to this alloca.
245 ///
246 /// These are operands which have cannot actually be used to refer to the
247 /// alloca as they are outside its range and the user doesn't correct for
248 /// that. These mostly consist of PHI node inputs and the like which we just
249 /// need to replace with undef.
getDeadOperands() const250 ArrayRef<Use *> getDeadOperands() const { return DeadOperands; }
251
252 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
253 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
254 void printSlice(raw_ostream &OS, const_iterator I,
255 StringRef Indent = " ") const;
256 void printUse(raw_ostream &OS, const_iterator I,
257 StringRef Indent = " ") const;
258 void print(raw_ostream &OS) const;
259 void dump(const_iterator I) const;
260 void dump() const;
261 #endif
262
263 private:
264 template <typename DerivedT, typename RetT = void> class BuilderBase;
265 class SliceBuilder;
266 friend class AllocaSlices::SliceBuilder;
267
268 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
269 /// \brief Handle to alloca instruction to simplify method interfaces.
270 AllocaInst &AI;
271 #endif
272
273 /// \brief The instruction responsible for this alloca not having a known set
274 /// of slices.
275 ///
276 /// When an instruction (potentially) escapes the pointer to the alloca, we
277 /// store a pointer to that here and abort trying to form slices of the
278 /// alloca. This will be null if the alloca slices are analyzed successfully.
279 Instruction *PointerEscapingInstr;
280
281 /// \brief The slices of the alloca.
282 ///
283 /// We store a vector of the slices formed by uses of the alloca here. This
284 /// vector is sorted by increasing begin offset, and then the unsplittable
285 /// slices before the splittable ones. See the Slice inner class for more
286 /// details.
287 SmallVector<Slice, 8> Slices;
288
289 /// \brief Instructions which will become dead if we rewrite the alloca.
290 ///
291 /// Note that these are not separated by slice. This is because we expect an
292 /// alloca to be completely rewritten or not rewritten at all. If rewritten,
293 /// all these instructions can simply be removed and replaced with undef as
294 /// they come from outside of the allocated space.
295 SmallVector<Instruction *, 8> DeadUsers;
296
297 /// \brief Operands which will become dead if we rewrite the alloca.
298 ///
299 /// These are operands that in their particular use can be replaced with
300 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
301 /// to PHI nodes and the like. They aren't entirely dead (there might be
302 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
303 /// want to swap this particular input for undef to simplify the use lists of
304 /// the alloca.
305 SmallVector<Use *, 8> DeadOperands;
306 };
307
308 /// \brief A partition of the slices.
309 ///
310 /// An ephemeral representation for a range of slices which can be viewed as
311 /// a partition of the alloca. This range represents a span of the alloca's
312 /// memory which cannot be split, and provides access to all of the slices
313 /// overlapping some part of the partition.
314 ///
315 /// Objects of this type are produced by traversing the alloca's slices, but
316 /// are only ephemeral and not persistent.
317 class llvm::sroa::Partition {
318 private:
319 friend class AllocaSlices;
320 friend class AllocaSlices::partition_iterator;
321
322 typedef AllocaSlices::iterator iterator;
323
324 /// \brief The beginning and ending offsets of the alloca for this
325 /// partition.
326 uint64_t BeginOffset, EndOffset;
327
328 /// \brief The start end end iterators of this partition.
329 iterator SI, SJ;
330
331 /// \brief A collection of split slice tails overlapping the partition.
332 SmallVector<Slice *, 4> SplitTails;
333
334 /// \brief Raw constructor builds an empty partition starting and ending at
335 /// the given iterator.
Partition(iterator SI)336 Partition(iterator SI) : SI(SI), SJ(SI) {}
337
338 public:
339 /// \brief The start offset of this partition.
340 ///
341 /// All of the contained slices start at or after this offset.
beginOffset() const342 uint64_t beginOffset() const { return BeginOffset; }
343
344 /// \brief The end offset of this partition.
345 ///
346 /// All of the contained slices end at or before this offset.
endOffset() const347 uint64_t endOffset() const { return EndOffset; }
348
349 /// \brief The size of the partition.
350 ///
351 /// Note that this can never be zero.
size() const352 uint64_t size() const {
353 assert(BeginOffset < EndOffset && "Partitions must span some bytes!");
354 return EndOffset - BeginOffset;
355 }
356
357 /// \brief Test whether this partition contains no slices, and merely spans
358 /// a region occupied by split slices.
empty() const359 bool empty() const { return SI == SJ; }
360
361 /// \name Iterate slices that start within the partition.
362 /// These may be splittable or unsplittable. They have a begin offset >= the
363 /// partition begin offset.
364 /// @{
365 // FIXME: We should probably define a "concat_iterator" helper and use that
366 // to stitch together pointee_iterators over the split tails and the
367 // contiguous iterators of the partition. That would give a much nicer
368 // interface here. We could then additionally expose filtered iterators for
369 // split, unsplit, and unsplittable splices based on the usage patterns.
begin() const370 iterator begin() const { return SI; }
end() const371 iterator end() const { return SJ; }
372 /// @}
373
374 /// \brief Get the sequence of split slice tails.
375 ///
376 /// These tails are of slices which start before this partition but are
377 /// split and overlap into the partition. We accumulate these while forming
378 /// partitions.
splitSliceTails() const379 ArrayRef<Slice *> splitSliceTails() const { return SplitTails; }
380 };
381
382 /// \brief An iterator over partitions of the alloca's slices.
383 ///
384 /// This iterator implements the core algorithm for partitioning the alloca's
385 /// slices. It is a forward iterator as we don't support backtracking for
386 /// efficiency reasons, and re-use a single storage area to maintain the
387 /// current set of split slices.
388 ///
389 /// It is templated on the slice iterator type to use so that it can operate
390 /// with either const or non-const slice iterators.
391 class AllocaSlices::partition_iterator
392 : public iterator_facade_base<partition_iterator, std::forward_iterator_tag,
393 Partition> {
394 friend class AllocaSlices;
395
396 /// \brief Most of the state for walking the partitions is held in a class
397 /// with a nice interface for examining them.
398 Partition P;
399
400 /// \brief We need to keep the end of the slices to know when to stop.
401 AllocaSlices::iterator SE;
402
403 /// \brief We also need to keep track of the maximum split end offset seen.
404 /// FIXME: Do we really?
405 uint64_t MaxSplitSliceEndOffset;
406
407 /// \brief Sets the partition to be empty at given iterator, and sets the
408 /// end iterator.
partition_iterator(AllocaSlices::iterator SI,AllocaSlices::iterator SE)409 partition_iterator(AllocaSlices::iterator SI, AllocaSlices::iterator SE)
410 : P(SI), SE(SE), MaxSplitSliceEndOffset(0) {
411 // If not already at the end, advance our state to form the initial
412 // partition.
413 if (SI != SE)
414 advance();
415 }
416
417 /// \brief Advance the iterator to the next partition.
418 ///
419 /// Requires that the iterator not be at the end of the slices.
advance()420 void advance() {
421 assert((P.SI != SE || !P.SplitTails.empty()) &&
422 "Cannot advance past the end of the slices!");
423
424 // Clear out any split uses which have ended.
425 if (!P.SplitTails.empty()) {
426 if (P.EndOffset >= MaxSplitSliceEndOffset) {
427 // If we've finished all splits, this is easy.
428 P.SplitTails.clear();
429 MaxSplitSliceEndOffset = 0;
430 } else {
431 // Remove the uses which have ended in the prior partition. This
432 // cannot change the max split slice end because we just checked that
433 // the prior partition ended prior to that max.
434 P.SplitTails.erase(
435 std::remove_if(
436 P.SplitTails.begin(), P.SplitTails.end(),
437 [&](Slice *S) { return S->endOffset() <= P.EndOffset; }),
438 P.SplitTails.end());
439 assert(std::any_of(P.SplitTails.begin(), P.SplitTails.end(),
440 [&](Slice *S) {
441 return S->endOffset() == MaxSplitSliceEndOffset;
442 }) &&
443 "Could not find the current max split slice offset!");
444 assert(std::all_of(P.SplitTails.begin(), P.SplitTails.end(),
445 [&](Slice *S) {
446 return S->endOffset() <= MaxSplitSliceEndOffset;
447 }) &&
448 "Max split slice end offset is not actually the max!");
449 }
450 }
451
452 // If P.SI is already at the end, then we've cleared the split tail and
453 // now have an end iterator.
454 if (P.SI == SE) {
455 assert(P.SplitTails.empty() && "Failed to clear the split slices!");
456 return;
457 }
458
459 // If we had a non-empty partition previously, set up the state for
460 // subsequent partitions.
461 if (P.SI != P.SJ) {
462 // Accumulate all the splittable slices which started in the old
463 // partition into the split list.
464 for (Slice &S : P)
465 if (S.isSplittable() && S.endOffset() > P.EndOffset) {
466 P.SplitTails.push_back(&S);
467 MaxSplitSliceEndOffset =
468 std::max(S.endOffset(), MaxSplitSliceEndOffset);
469 }
470
471 // Start from the end of the previous partition.
472 P.SI = P.SJ;
473
474 // If P.SI is now at the end, we at most have a tail of split slices.
475 if (P.SI == SE) {
476 P.BeginOffset = P.EndOffset;
477 P.EndOffset = MaxSplitSliceEndOffset;
478 return;
479 }
480
481 // If the we have split slices and the next slice is after a gap and is
482 // not splittable immediately form an empty partition for the split
483 // slices up until the next slice begins.
484 if (!P.SplitTails.empty() && P.SI->beginOffset() != P.EndOffset &&
485 !P.SI->isSplittable()) {
486 P.BeginOffset = P.EndOffset;
487 P.EndOffset = P.SI->beginOffset();
488 return;
489 }
490 }
491
492 // OK, we need to consume new slices. Set the end offset based on the
493 // current slice, and step SJ past it. The beginning offset of the
494 // partition is the beginning offset of the next slice unless we have
495 // pre-existing split slices that are continuing, in which case we begin
496 // at the prior end offset.
497 P.BeginOffset = P.SplitTails.empty() ? P.SI->beginOffset() : P.EndOffset;
498 P.EndOffset = P.SI->endOffset();
499 ++P.SJ;
500
501 // There are two strategies to form a partition based on whether the
502 // partition starts with an unsplittable slice or a splittable slice.
503 if (!P.SI->isSplittable()) {
504 // When we're forming an unsplittable region, it must always start at
505 // the first slice and will extend through its end.
506 assert(P.BeginOffset == P.SI->beginOffset());
507
508 // Form a partition including all of the overlapping slices with this
509 // unsplittable slice.
510 while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
511 if (!P.SJ->isSplittable())
512 P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
513 ++P.SJ;
514 }
515
516 // We have a partition across a set of overlapping unsplittable
517 // partitions.
518 return;
519 }
520
521 // If we're starting with a splittable slice, then we need to form
522 // a synthetic partition spanning it and any other overlapping splittable
523 // splices.
524 assert(P.SI->isSplittable() && "Forming a splittable partition!");
525
526 // Collect all of the overlapping splittable slices.
527 while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset &&
528 P.SJ->isSplittable()) {
529 P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
530 ++P.SJ;
531 }
532
533 // Back upiP.EndOffset if we ended the span early when encountering an
534 // unsplittable slice. This synthesizes the early end offset of
535 // a partition spanning only splittable slices.
536 if (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
537 assert(!P.SJ->isSplittable());
538 P.EndOffset = P.SJ->beginOffset();
539 }
540 }
541
542 public:
operator ==(const partition_iterator & RHS) const543 bool operator==(const partition_iterator &RHS) const {
544 assert(SE == RHS.SE &&
545 "End iterators don't match between compared partition iterators!");
546
547 // The observed positions of partitions is marked by the P.SI iterator and
548 // the emptiness of the split slices. The latter is only relevant when
549 // P.SI == SE, as the end iterator will additionally have an empty split
550 // slices list, but the prior may have the same P.SI and a tail of split
551 // slices.
552 if (P.SI == RHS.P.SI && P.SplitTails.empty() == RHS.P.SplitTails.empty()) {
553 assert(P.SJ == RHS.P.SJ &&
554 "Same set of slices formed two different sized partitions!");
555 assert(P.SplitTails.size() == RHS.P.SplitTails.size() &&
556 "Same slice position with differently sized non-empty split "
557 "slice tails!");
558 return true;
559 }
560 return false;
561 }
562
operator ++()563 partition_iterator &operator++() {
564 advance();
565 return *this;
566 }
567
operator *()568 Partition &operator*() { return P; }
569 };
570
571 /// \brief A forward range over the partitions of the alloca's slices.
572 ///
573 /// This accesses an iterator range over the partitions of the alloca's
574 /// slices. It computes these partitions on the fly based on the overlapping
575 /// offsets of the slices and the ability to split them. It will visit "empty"
576 /// partitions to cover regions of the alloca only accessed via split
577 /// slices.
partitions()578 iterator_range<AllocaSlices::partition_iterator> AllocaSlices::partitions() {
579 return make_range(partition_iterator(begin(), end()),
580 partition_iterator(end(), end()));
581 }
582
foldSelectInst(SelectInst & SI)583 static Value *foldSelectInst(SelectInst &SI) {
584 // If the condition being selected on is a constant or the same value is
585 // being selected between, fold the select. Yes this does (rarely) happen
586 // early on.
587 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
588 return SI.getOperand(1 + CI->isZero());
589 if (SI.getOperand(1) == SI.getOperand(2))
590 return SI.getOperand(1);
591
592 return nullptr;
593 }
594
595 /// \brief A helper that folds a PHI node or a select.
foldPHINodeOrSelectInst(Instruction & I)596 static Value *foldPHINodeOrSelectInst(Instruction &I) {
597 if (PHINode *PN = dyn_cast<PHINode>(&I)) {
598 // If PN merges together the same value, return that value.
599 return PN->hasConstantValue();
600 }
601 return foldSelectInst(cast<SelectInst>(I));
602 }
603
604 /// \brief Builder for the alloca slices.
605 ///
606 /// This class builds a set of alloca slices by recursively visiting the uses
607 /// of an alloca and making a slice for each load and store at each offset.
608 class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
609 friend class PtrUseVisitor<SliceBuilder>;
610 friend class InstVisitor<SliceBuilder>;
611 typedef PtrUseVisitor<SliceBuilder> Base;
612
613 const uint64_t AllocSize;
614 AllocaSlices &AS;
615
616 SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
617 SmallDenseMap<Instruction *, uint64_t> PHIOrSelectSizes;
618
619 /// \brief Set to de-duplicate dead instructions found in the use walk.
620 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
621
622 public:
SliceBuilder(const DataLayout & DL,AllocaInst & AI,AllocaSlices & AS)623 SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &AS)
624 : PtrUseVisitor<SliceBuilder>(DL),
625 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())), AS(AS) {}
626
627 private:
markAsDead(Instruction & I)628 void markAsDead(Instruction &I) {
629 if (VisitedDeadInsts.insert(&I).second)
630 AS.DeadUsers.push_back(&I);
631 }
632
insertUse(Instruction & I,const APInt & Offset,uint64_t Size,bool IsSplittable=false)633 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
634 bool IsSplittable = false) {
635 // Completely skip uses which have a zero size or start either before or
636 // past the end of the allocation.
637 if (Size == 0 || Offset.uge(AllocSize)) {
638 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
639 << " which has zero size or starts outside of the "
640 << AllocSize << " byte alloca:\n"
641 << " alloca: " << AS.AI << "\n"
642 << " use: " << I << "\n");
643 return markAsDead(I);
644 }
645
646 uint64_t BeginOffset = Offset.getZExtValue();
647 uint64_t EndOffset = BeginOffset + Size;
648
649 // Clamp the end offset to the end of the allocation. Note that this is
650 // formulated to handle even the case where "BeginOffset + Size" overflows.
651 // This may appear superficially to be something we could ignore entirely,
652 // but that is not so! There may be widened loads or PHI-node uses where
653 // some instructions are dead but not others. We can't completely ignore
654 // them, and so have to record at least the information here.
655 assert(AllocSize >= BeginOffset); // Established above.
656 if (Size > AllocSize - BeginOffset) {
657 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
658 << " to remain within the " << AllocSize << " byte alloca:\n"
659 << " alloca: " << AS.AI << "\n"
660 << " use: " << I << "\n");
661 EndOffset = AllocSize;
662 }
663
664 AS.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
665 }
666
visitBitCastInst(BitCastInst & BC)667 void visitBitCastInst(BitCastInst &BC) {
668 if (BC.use_empty())
669 return markAsDead(BC);
670
671 return Base::visitBitCastInst(BC);
672 }
673
visitGetElementPtrInst(GetElementPtrInst & GEPI)674 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
675 if (GEPI.use_empty())
676 return markAsDead(GEPI);
677
678 if (SROAStrictInbounds && GEPI.isInBounds()) {
679 // FIXME: This is a manually un-factored variant of the basic code inside
680 // of GEPs with checking of the inbounds invariant specified in the
681 // langref in a very strict sense. If we ever want to enable
682 // SROAStrictInbounds, this code should be factored cleanly into
683 // PtrUseVisitor, but it is easier to experiment with SROAStrictInbounds
684 // by writing out the code here where we have the underlying allocation
685 // size readily available.
686 APInt GEPOffset = Offset;
687 const DataLayout &DL = GEPI.getModule()->getDataLayout();
688 for (gep_type_iterator GTI = gep_type_begin(GEPI),
689 GTE = gep_type_end(GEPI);
690 GTI != GTE; ++GTI) {
691 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
692 if (!OpC)
693 break;
694
695 // Handle a struct index, which adds its field offset to the pointer.
696 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
697 unsigned ElementIdx = OpC->getZExtValue();
698 const StructLayout *SL = DL.getStructLayout(STy);
699 GEPOffset +=
700 APInt(Offset.getBitWidth(), SL->getElementOffset(ElementIdx));
701 } else {
702 // For array or vector indices, scale the index by the size of the
703 // type.
704 APInt Index = OpC->getValue().sextOrTrunc(Offset.getBitWidth());
705 GEPOffset += Index * APInt(Offset.getBitWidth(),
706 DL.getTypeAllocSize(GTI.getIndexedType()));
707 }
708
709 // If this index has computed an intermediate pointer which is not
710 // inbounds, then the result of the GEP is a poison value and we can
711 // delete it and all uses.
712 if (GEPOffset.ugt(AllocSize))
713 return markAsDead(GEPI);
714 }
715 }
716
717 return Base::visitGetElementPtrInst(GEPI);
718 }
719
handleLoadOrStore(Type * Ty,Instruction & I,const APInt & Offset,uint64_t Size,bool IsVolatile)720 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
721 uint64_t Size, bool IsVolatile) {
722 // We allow splitting of non-volatile loads and stores where the type is an
723 // integer type. These may be used to implement 'memcpy' or other "transfer
724 // of bits" patterns.
725 bool IsSplittable = Ty->isIntegerTy() && !IsVolatile;
726
727 insertUse(I, Offset, Size, IsSplittable);
728 }
729
visitLoadInst(LoadInst & LI)730 void visitLoadInst(LoadInst &LI) {
731 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
732 "All simple FCA loads should have been pre-split");
733
734 if (!IsOffsetKnown)
735 return PI.setAborted(&LI);
736
737 const DataLayout &DL = LI.getModule()->getDataLayout();
738 uint64_t Size = DL.getTypeStoreSize(LI.getType());
739 return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
740 }
741
visitStoreInst(StoreInst & SI)742 void visitStoreInst(StoreInst &SI) {
743 Value *ValOp = SI.getValueOperand();
744 if (ValOp == *U)
745 return PI.setEscapedAndAborted(&SI);
746 if (!IsOffsetKnown)
747 return PI.setAborted(&SI);
748
749 const DataLayout &DL = SI.getModule()->getDataLayout();
750 uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
751
752 // If this memory access can be shown to *statically* extend outside the
753 // bounds of of the allocation, it's behavior is undefined, so simply
754 // ignore it. Note that this is more strict than the generic clamping
755 // behavior of insertUse. We also try to handle cases which might run the
756 // risk of overflow.
757 // FIXME: We should instead consider the pointer to have escaped if this
758 // function is being instrumented for addressing bugs or race conditions.
759 if (Size > AllocSize || Offset.ugt(AllocSize - Size)) {
760 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @" << Offset
761 << " which extends past the end of the " << AllocSize
762 << " byte alloca:\n"
763 << " alloca: " << AS.AI << "\n"
764 << " use: " << SI << "\n");
765 return markAsDead(SI);
766 }
767
768 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
769 "All simple FCA stores should have been pre-split");
770 handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
771 }
772
visitMemSetInst(MemSetInst & II)773 void visitMemSetInst(MemSetInst &II) {
774 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
775 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
776 if ((Length && Length->getValue() == 0) ||
777 (IsOffsetKnown && Offset.uge(AllocSize)))
778 // Zero-length mem transfer intrinsics can be ignored entirely.
779 return markAsDead(II);
780
781 if (!IsOffsetKnown)
782 return PI.setAborted(&II);
783
784 insertUse(II, Offset, Length ? Length->getLimitedValue()
785 : AllocSize - Offset.getLimitedValue(),
786 (bool)Length);
787 }
788
visitMemTransferInst(MemTransferInst & II)789 void visitMemTransferInst(MemTransferInst &II) {
790 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
791 if (Length && Length->getValue() == 0)
792 // Zero-length mem transfer intrinsics can be ignored entirely.
793 return markAsDead(II);
794
795 // Because we can visit these intrinsics twice, also check to see if the
796 // first time marked this instruction as dead. If so, skip it.
797 if (VisitedDeadInsts.count(&II))
798 return;
799
800 if (!IsOffsetKnown)
801 return PI.setAborted(&II);
802
803 // This side of the transfer is completely out-of-bounds, and so we can
804 // nuke the entire transfer. However, we also need to nuke the other side
805 // if already added to our partitions.
806 // FIXME: Yet another place we really should bypass this when
807 // instrumenting for ASan.
808 if (Offset.uge(AllocSize)) {
809 SmallDenseMap<Instruction *, unsigned>::iterator MTPI =
810 MemTransferSliceMap.find(&II);
811 if (MTPI != MemTransferSliceMap.end())
812 AS.Slices[MTPI->second].kill();
813 return markAsDead(II);
814 }
815
816 uint64_t RawOffset = Offset.getLimitedValue();
817 uint64_t Size = Length ? Length->getLimitedValue() : AllocSize - RawOffset;
818
819 // Check for the special case where the same exact value is used for both
820 // source and dest.
821 if (*U == II.getRawDest() && *U == II.getRawSource()) {
822 // For non-volatile transfers this is a no-op.
823 if (!II.isVolatile())
824 return markAsDead(II);
825
826 return insertUse(II, Offset, Size, /*IsSplittable=*/false);
827 }
828
829 // If we have seen both source and destination for a mem transfer, then
830 // they both point to the same alloca.
831 bool Inserted;
832 SmallDenseMap<Instruction *, unsigned>::iterator MTPI;
833 std::tie(MTPI, Inserted) =
834 MemTransferSliceMap.insert(std::make_pair(&II, AS.Slices.size()));
835 unsigned PrevIdx = MTPI->second;
836 if (!Inserted) {
837 Slice &PrevP = AS.Slices[PrevIdx];
838
839 // Check if the begin offsets match and this is a non-volatile transfer.
840 // In that case, we can completely elide the transfer.
841 if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
842 PrevP.kill();
843 return markAsDead(II);
844 }
845
846 // Otherwise we have an offset transfer within the same alloca. We can't
847 // split those.
848 PrevP.makeUnsplittable();
849 }
850
851 // Insert the use now that we've fixed up the splittable nature.
852 insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
853
854 // Check that we ended up with a valid index in the map.
855 assert(AS.Slices[PrevIdx].getUse()->getUser() == &II &&
856 "Map index doesn't point back to a slice with this user.");
857 }
858
859 // Disable SRoA for any intrinsics except for lifetime invariants.
860 // FIXME: What about debug intrinsics? This matches old behavior, but
861 // doesn't make sense.
visitIntrinsicInst(IntrinsicInst & II)862 void visitIntrinsicInst(IntrinsicInst &II) {
863 if (!IsOffsetKnown)
864 return PI.setAborted(&II);
865
866 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
867 II.getIntrinsicID() == Intrinsic::lifetime_end) {
868 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
869 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
870 Length->getLimitedValue());
871 insertUse(II, Offset, Size, true);
872 return;
873 }
874
875 Base::visitIntrinsicInst(II);
876 }
877
hasUnsafePHIOrSelectUse(Instruction * Root,uint64_t & Size)878 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
879 // We consider any PHI or select that results in a direct load or store of
880 // the same offset to be a viable use for slicing purposes. These uses
881 // are considered unsplittable and the size is the maximum loaded or stored
882 // size.
883 SmallPtrSet<Instruction *, 4> Visited;
884 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
885 Visited.insert(Root);
886 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
887 const DataLayout &DL = Root->getModule()->getDataLayout();
888 // If there are no loads or stores, the access is dead. We mark that as
889 // a size zero access.
890 Size = 0;
891 do {
892 Instruction *I, *UsedI;
893 std::tie(UsedI, I) = Uses.pop_back_val();
894
895 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
896 Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
897 continue;
898 }
899 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
900 Value *Op = SI->getOperand(0);
901 if (Op == UsedI)
902 return SI;
903 Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
904 continue;
905 }
906
907 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
908 if (!GEP->hasAllZeroIndices())
909 return GEP;
910 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
911 !isa<SelectInst>(I)) {
912 return I;
913 }
914
915 for (User *U : I->users())
916 if (Visited.insert(cast<Instruction>(U)).second)
917 Uses.push_back(std::make_pair(I, cast<Instruction>(U)));
918 } while (!Uses.empty());
919
920 return nullptr;
921 }
922
visitPHINodeOrSelectInst(Instruction & I)923 void visitPHINodeOrSelectInst(Instruction &I) {
924 assert(isa<PHINode>(I) || isa<SelectInst>(I));
925 if (I.use_empty())
926 return markAsDead(I);
927
928 // TODO: We could use SimplifyInstruction here to fold PHINodes and
929 // SelectInsts. However, doing so requires to change the current
930 // dead-operand-tracking mechanism. For instance, suppose neither loading
931 // from %U nor %other traps. Then "load (select undef, %U, %other)" does not
932 // trap either. However, if we simply replace %U with undef using the
933 // current dead-operand-tracking mechanism, "load (select undef, undef,
934 // %other)" may trap because the select may return the first operand
935 // "undef".
936 if (Value *Result = foldPHINodeOrSelectInst(I)) {
937 if (Result == *U)
938 // If the result of the constant fold will be the pointer, recurse
939 // through the PHI/select as if we had RAUW'ed it.
940 enqueueUsers(I);
941 else
942 // Otherwise the operand to the PHI/select is dead, and we can replace
943 // it with undef.
944 AS.DeadOperands.push_back(U);
945
946 return;
947 }
948
949 if (!IsOffsetKnown)
950 return PI.setAborted(&I);
951
952 // See if we already have computed info on this node.
953 uint64_t &Size = PHIOrSelectSizes[&I];
954 if (!Size) {
955 // This is a new PHI/Select, check for an unsafe use of it.
956 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&I, Size))
957 return PI.setAborted(UnsafeI);
958 }
959
960 // For PHI and select operands outside the alloca, we can't nuke the entire
961 // phi or select -- the other side might still be relevant, so we special
962 // case them here and use a separate structure to track the operands
963 // themselves which should be replaced with undef.
964 // FIXME: This should instead be escaped in the event we're instrumenting
965 // for address sanitization.
966 if (Offset.uge(AllocSize)) {
967 AS.DeadOperands.push_back(U);
968 return;
969 }
970
971 insertUse(I, Offset, Size);
972 }
973
visitPHINode(PHINode & PN)974 void visitPHINode(PHINode &PN) { visitPHINodeOrSelectInst(PN); }
975
visitSelectInst(SelectInst & SI)976 void visitSelectInst(SelectInst &SI) { visitPHINodeOrSelectInst(SI); }
977
978 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
visitInstruction(Instruction & I)979 void visitInstruction(Instruction &I) { PI.setAborted(&I); }
980 };
981
AllocaSlices(const DataLayout & DL,AllocaInst & AI)982 AllocaSlices::AllocaSlices(const DataLayout &DL, AllocaInst &AI)
983 :
984 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
985 AI(AI),
986 #endif
987 PointerEscapingInstr(nullptr) {
988 SliceBuilder PB(DL, AI, *this);
989 SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
990 if (PtrI.isEscaped() || PtrI.isAborted()) {
991 // FIXME: We should sink the escape vs. abort info into the caller nicely,
992 // possibly by just storing the PtrInfo in the AllocaSlices.
993 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
994 : PtrI.getAbortingInst();
995 assert(PointerEscapingInstr && "Did not track a bad instruction");
996 return;
997 }
998
999 Slices.erase(std::remove_if(Slices.begin(), Slices.end(),
1000 [](const Slice &S) {
1001 return S.isDead();
1002 }),
1003 Slices.end());
1004
1005 #ifndef NDEBUG
1006 if (SROARandomShuffleSlices) {
1007 std::mt19937 MT(static_cast<unsigned>(sys::TimeValue::now().msec()));
1008 std::shuffle(Slices.begin(), Slices.end(), MT);
1009 }
1010 #endif
1011
1012 // Sort the uses. This arranges for the offsets to be in ascending order,
1013 // and the sizes to be in descending order.
1014 std::sort(Slices.begin(), Slices.end());
1015 }
1016
1017 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1018
print(raw_ostream & OS,const_iterator I,StringRef Indent) const1019 void AllocaSlices::print(raw_ostream &OS, const_iterator I,
1020 StringRef Indent) const {
1021 printSlice(OS, I, Indent);
1022 OS << "\n";
1023 printUse(OS, I, Indent);
1024 }
1025
printSlice(raw_ostream & OS,const_iterator I,StringRef Indent) const1026 void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
1027 StringRef Indent) const {
1028 OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
1029 << " slice #" << (I - begin())
1030 << (I->isSplittable() ? " (splittable)" : "");
1031 }
1032
printUse(raw_ostream & OS,const_iterator I,StringRef Indent) const1033 void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
1034 StringRef Indent) const {
1035 OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
1036 }
1037
print(raw_ostream & OS) const1038 void AllocaSlices::print(raw_ostream &OS) const {
1039 if (PointerEscapingInstr) {
1040 OS << "Can't analyze slices for alloca: " << AI << "\n"
1041 << " A pointer to this alloca escaped by:\n"
1042 << " " << *PointerEscapingInstr << "\n";
1043 return;
1044 }
1045
1046 OS << "Slices of alloca: " << AI << "\n";
1047 for (const_iterator I = begin(), E = end(); I != E; ++I)
1048 print(OS, I);
1049 }
1050
dump(const_iterator I) const1051 LLVM_DUMP_METHOD void AllocaSlices::dump(const_iterator I) const {
1052 print(dbgs(), I);
1053 }
dump() const1054 LLVM_DUMP_METHOD void AllocaSlices::dump() const { print(dbgs()); }
1055
1056 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1057
1058 /// Walk the range of a partitioning looking for a common type to cover this
1059 /// sequence of slices.
findCommonType(AllocaSlices::const_iterator B,AllocaSlices::const_iterator E,uint64_t EndOffset)1060 static Type *findCommonType(AllocaSlices::const_iterator B,
1061 AllocaSlices::const_iterator E,
1062 uint64_t EndOffset) {
1063 Type *Ty = nullptr;
1064 bool TyIsCommon = true;
1065 IntegerType *ITy = nullptr;
1066
1067 // Note that we need to look at *every* alloca slice's Use to ensure we
1068 // always get consistent results regardless of the order of slices.
1069 for (AllocaSlices::const_iterator I = B; I != E; ++I) {
1070 Use *U = I->getUse();
1071 if (isa<IntrinsicInst>(*U->getUser()))
1072 continue;
1073 if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
1074 continue;
1075
1076 Type *UserTy = nullptr;
1077 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1078 UserTy = LI->getType();
1079 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1080 UserTy = SI->getValueOperand()->getType();
1081 }
1082
1083 if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
1084 // If the type is larger than the partition, skip it. We only encounter
1085 // this for split integer operations where we want to use the type of the
1086 // entity causing the split. Also skip if the type is not a byte width
1087 // multiple.
1088 if (UserITy->getBitWidth() % 8 != 0 ||
1089 UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
1090 continue;
1091
1092 // Track the largest bitwidth integer type used in this way in case there
1093 // is no common type.
1094 if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
1095 ITy = UserITy;
1096 }
1097
1098 // To avoid depending on the order of slices, Ty and TyIsCommon must not
1099 // depend on types skipped above.
1100 if (!UserTy || (Ty && Ty != UserTy))
1101 TyIsCommon = false; // Give up on anything but an iN type.
1102 else
1103 Ty = UserTy;
1104 }
1105
1106 return TyIsCommon ? Ty : ITy;
1107 }
1108
1109 /// PHI instructions that use an alloca and are subsequently loaded can be
1110 /// rewritten to load both input pointers in the pred blocks and then PHI the
1111 /// results, allowing the load of the alloca to be promoted.
1112 /// From this:
1113 /// %P2 = phi [i32* %Alloca, i32* %Other]
1114 /// %V = load i32* %P2
1115 /// to:
1116 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1117 /// ...
1118 /// %V2 = load i32* %Other
1119 /// ...
1120 /// %V = phi [i32 %V1, i32 %V2]
1121 ///
1122 /// We can do this to a select if its only uses are loads and if the operands
1123 /// to the select can be loaded unconditionally.
1124 ///
1125 /// FIXME: This should be hoisted into a generic utility, likely in
1126 /// Transforms/Util/Local.h
isSafePHIToSpeculate(PHINode & PN)1127 static bool isSafePHIToSpeculate(PHINode &PN) {
1128 // For now, we can only do this promotion if the load is in the same block
1129 // as the PHI, and if there are no stores between the phi and load.
1130 // TODO: Allow recursive phi users.
1131 // TODO: Allow stores.
1132 BasicBlock *BB = PN.getParent();
1133 unsigned MaxAlign = 0;
1134 bool HaveLoad = false;
1135 for (User *U : PN.users()) {
1136 LoadInst *LI = dyn_cast<LoadInst>(U);
1137 if (!LI || !LI->isSimple())
1138 return false;
1139
1140 // For now we only allow loads in the same block as the PHI. This is
1141 // a common case that happens when instcombine merges two loads through
1142 // a PHI.
1143 if (LI->getParent() != BB)
1144 return false;
1145
1146 // Ensure that there are no instructions between the PHI and the load that
1147 // could store.
1148 for (BasicBlock::iterator BBI(PN); &*BBI != LI; ++BBI)
1149 if (BBI->mayWriteToMemory())
1150 return false;
1151
1152 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1153 HaveLoad = true;
1154 }
1155
1156 if (!HaveLoad)
1157 return false;
1158
1159 const DataLayout &DL = PN.getModule()->getDataLayout();
1160
1161 // We can only transform this if it is safe to push the loads into the
1162 // predecessor blocks. The only thing to watch out for is that we can't put
1163 // a possibly trapping load in the predecessor if it is a critical edge.
1164 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1165 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1166 Value *InVal = PN.getIncomingValue(Idx);
1167
1168 // If the value is produced by the terminator of the predecessor (an
1169 // invoke) or it has side-effects, there is no valid place to put a load
1170 // in the predecessor.
1171 if (TI == InVal || TI->mayHaveSideEffects())
1172 return false;
1173
1174 // If the predecessor has a single successor, then the edge isn't
1175 // critical.
1176 if (TI->getNumSuccessors() == 1)
1177 continue;
1178
1179 // If this pointer is always safe to load, or if we can prove that there
1180 // is already a load in the block, then we can move the load to the pred
1181 // block.
1182 if (isSafeToLoadUnconditionally(InVal, MaxAlign, DL, TI))
1183 continue;
1184
1185 return false;
1186 }
1187
1188 return true;
1189 }
1190
speculatePHINodeLoads(PHINode & PN)1191 static void speculatePHINodeLoads(PHINode &PN) {
1192 DEBUG(dbgs() << " original: " << PN << "\n");
1193
1194 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1195 IRBuilderTy PHIBuilder(&PN);
1196 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1197 PN.getName() + ".sroa.speculated");
1198
1199 // Get the AA tags and alignment to use from one of the loads. It doesn't
1200 // matter which one we get and if any differ.
1201 LoadInst *SomeLoad = cast<LoadInst>(PN.user_back());
1202
1203 AAMDNodes AATags;
1204 SomeLoad->getAAMetadata(AATags);
1205 unsigned Align = SomeLoad->getAlignment();
1206
1207 // Rewrite all loads of the PN to use the new PHI.
1208 while (!PN.use_empty()) {
1209 LoadInst *LI = cast<LoadInst>(PN.user_back());
1210 LI->replaceAllUsesWith(NewPN);
1211 LI->eraseFromParent();
1212 }
1213
1214 // Inject loads into all of the pred blocks.
1215 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1216 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1217 TerminatorInst *TI = Pred->getTerminator();
1218 Value *InVal = PN.getIncomingValue(Idx);
1219 IRBuilderTy PredBuilder(TI);
1220
1221 LoadInst *Load = PredBuilder.CreateLoad(
1222 InVal, (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1223 ++NumLoadsSpeculated;
1224 Load->setAlignment(Align);
1225 if (AATags)
1226 Load->setAAMetadata(AATags);
1227 NewPN->addIncoming(Load, Pred);
1228 }
1229
1230 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1231 PN.eraseFromParent();
1232 }
1233
1234 /// Select instructions that use an alloca and are subsequently loaded can be
1235 /// rewritten to load both input pointers and then select between the result,
1236 /// allowing the load of the alloca to be promoted.
1237 /// From this:
1238 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1239 /// %V = load i32* %P2
1240 /// to:
1241 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1242 /// %V2 = load i32* %Other
1243 /// %V = select i1 %cond, i32 %V1, i32 %V2
1244 ///
1245 /// We can do this to a select if its only uses are loads and if the operand
1246 /// to the select can be loaded unconditionally.
isSafeSelectToSpeculate(SelectInst & SI)1247 static bool isSafeSelectToSpeculate(SelectInst &SI) {
1248 Value *TValue = SI.getTrueValue();
1249 Value *FValue = SI.getFalseValue();
1250 const DataLayout &DL = SI.getModule()->getDataLayout();
1251
1252 for (User *U : SI.users()) {
1253 LoadInst *LI = dyn_cast<LoadInst>(U);
1254 if (!LI || !LI->isSimple())
1255 return false;
1256
1257 // Both operands to the select need to be dereferencable, either
1258 // absolutely (e.g. allocas) or at this point because we can see other
1259 // accesses to it.
1260 if (!isSafeToLoadUnconditionally(TValue, LI->getAlignment(), DL, LI))
1261 return false;
1262 if (!isSafeToLoadUnconditionally(FValue, LI->getAlignment(), DL, LI))
1263 return false;
1264 }
1265
1266 return true;
1267 }
1268
speculateSelectInstLoads(SelectInst & SI)1269 static void speculateSelectInstLoads(SelectInst &SI) {
1270 DEBUG(dbgs() << " original: " << SI << "\n");
1271
1272 IRBuilderTy IRB(&SI);
1273 Value *TV = SI.getTrueValue();
1274 Value *FV = SI.getFalseValue();
1275 // Replace the loads of the select with a select of two loads.
1276 while (!SI.use_empty()) {
1277 LoadInst *LI = cast<LoadInst>(SI.user_back());
1278 assert(LI->isSimple() && "We only speculate simple loads");
1279
1280 IRB.SetInsertPoint(LI);
1281 LoadInst *TL =
1282 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1283 LoadInst *FL =
1284 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1285 NumLoadsSpeculated += 2;
1286
1287 // Transfer alignment and AA info if present.
1288 TL->setAlignment(LI->getAlignment());
1289 FL->setAlignment(LI->getAlignment());
1290
1291 AAMDNodes Tags;
1292 LI->getAAMetadata(Tags);
1293 if (Tags) {
1294 TL->setAAMetadata(Tags);
1295 FL->setAAMetadata(Tags);
1296 }
1297
1298 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1299 LI->getName() + ".sroa.speculated");
1300
1301 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1302 LI->replaceAllUsesWith(V);
1303 LI->eraseFromParent();
1304 }
1305 SI.eraseFromParent();
1306 }
1307
1308 /// \brief Build a GEP out of a base pointer and indices.
1309 ///
1310 /// This will return the BasePtr if that is valid, or build a new GEP
1311 /// instruction using the IRBuilder if GEP-ing is needed.
buildGEP(IRBuilderTy & IRB,Value * BasePtr,SmallVectorImpl<Value * > & Indices,Twine NamePrefix)1312 static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1313 SmallVectorImpl<Value *> &Indices, Twine NamePrefix) {
1314 if (Indices.empty())
1315 return BasePtr;
1316
1317 // A single zero index is a no-op, so check for this and avoid building a GEP
1318 // in that case.
1319 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1320 return BasePtr;
1321
1322 return IRB.CreateInBoundsGEP(nullptr, BasePtr, Indices,
1323 NamePrefix + "sroa_idx");
1324 }
1325
1326 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1327 /// TargetTy without changing the offset of the pointer.
1328 ///
1329 /// This routine assumes we've already established a properly offset GEP with
1330 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1331 /// zero-indices down through type layers until we find one the same as
1332 /// TargetTy. If we can't find one with the same type, we at least try to use
1333 /// one with the same size. If none of that works, we just produce the GEP as
1334 /// indicated by Indices to have the correct offset.
getNaturalGEPWithType(IRBuilderTy & IRB,const DataLayout & DL,Value * BasePtr,Type * Ty,Type * TargetTy,SmallVectorImpl<Value * > & Indices,Twine NamePrefix)1335 static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1336 Value *BasePtr, Type *Ty, Type *TargetTy,
1337 SmallVectorImpl<Value *> &Indices,
1338 Twine NamePrefix) {
1339 if (Ty == TargetTy)
1340 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1341
1342 // Pointer size to use for the indices.
1343 unsigned PtrSize = DL.getPointerTypeSizeInBits(BasePtr->getType());
1344
1345 // See if we can descend into a struct and locate a field with the correct
1346 // type.
1347 unsigned NumLayers = 0;
1348 Type *ElementTy = Ty;
1349 do {
1350 if (ElementTy->isPointerTy())
1351 break;
1352
1353 if (ArrayType *ArrayTy = dyn_cast<ArrayType>(ElementTy)) {
1354 ElementTy = ArrayTy->getElementType();
1355 Indices.push_back(IRB.getIntN(PtrSize, 0));
1356 } else if (VectorType *VectorTy = dyn_cast<VectorType>(ElementTy)) {
1357 ElementTy = VectorTy->getElementType();
1358 Indices.push_back(IRB.getInt32(0));
1359 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1360 if (STy->element_begin() == STy->element_end())
1361 break; // Nothing left to descend into.
1362 ElementTy = *STy->element_begin();
1363 Indices.push_back(IRB.getInt32(0));
1364 } else {
1365 break;
1366 }
1367 ++NumLayers;
1368 } while (ElementTy != TargetTy);
1369 if (ElementTy != TargetTy)
1370 Indices.erase(Indices.end() - NumLayers, Indices.end());
1371
1372 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1373 }
1374
1375 /// \brief Recursively compute indices for a natural GEP.
1376 ///
1377 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1378 /// element types adding appropriate indices for the GEP.
getNaturalGEPRecursively(IRBuilderTy & IRB,const DataLayout & DL,Value * Ptr,Type * Ty,APInt & Offset,Type * TargetTy,SmallVectorImpl<Value * > & Indices,Twine NamePrefix)1379 static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1380 Value *Ptr, Type *Ty, APInt &Offset,
1381 Type *TargetTy,
1382 SmallVectorImpl<Value *> &Indices,
1383 Twine NamePrefix) {
1384 if (Offset == 0)
1385 return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices,
1386 NamePrefix);
1387
1388 // We can't recurse through pointer types.
1389 if (Ty->isPointerTy())
1390 return nullptr;
1391
1392 // We try to analyze GEPs over vectors here, but note that these GEPs are
1393 // extremely poorly defined currently. The long-term goal is to remove GEPing
1394 // over a vector from the IR completely.
1395 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1396 unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
1397 if (ElementSizeInBits % 8 != 0) {
1398 // GEPs over non-multiple of 8 size vector elements are invalid.
1399 return nullptr;
1400 }
1401 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1402 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1403 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1404 return nullptr;
1405 Offset -= NumSkippedElements * ElementSize;
1406 Indices.push_back(IRB.getInt(NumSkippedElements));
1407 return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1408 Offset, TargetTy, Indices, NamePrefix);
1409 }
1410
1411 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1412 Type *ElementTy = ArrTy->getElementType();
1413 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1414 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1415 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1416 return nullptr;
1417
1418 Offset -= NumSkippedElements * ElementSize;
1419 Indices.push_back(IRB.getInt(NumSkippedElements));
1420 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1421 Indices, NamePrefix);
1422 }
1423
1424 StructType *STy = dyn_cast<StructType>(Ty);
1425 if (!STy)
1426 return nullptr;
1427
1428 const StructLayout *SL = DL.getStructLayout(STy);
1429 uint64_t StructOffset = Offset.getZExtValue();
1430 if (StructOffset >= SL->getSizeInBytes())
1431 return nullptr;
1432 unsigned Index = SL->getElementContainingOffset(StructOffset);
1433 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1434 Type *ElementTy = STy->getElementType(Index);
1435 if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
1436 return nullptr; // The offset points into alignment padding.
1437
1438 Indices.push_back(IRB.getInt32(Index));
1439 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1440 Indices, NamePrefix);
1441 }
1442
1443 /// \brief Get a natural GEP from a base pointer to a particular offset and
1444 /// resulting in a particular type.
1445 ///
1446 /// The goal is to produce a "natural" looking GEP that works with the existing
1447 /// composite types to arrive at the appropriate offset and element type for
1448 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1449 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1450 /// Indices, and setting Ty to the result subtype.
1451 ///
1452 /// If no natural GEP can be constructed, this function returns null.
getNaturalGEPWithOffset(IRBuilderTy & IRB,const DataLayout & DL,Value * Ptr,APInt Offset,Type * TargetTy,SmallVectorImpl<Value * > & Indices,Twine NamePrefix)1453 static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1454 Value *Ptr, APInt Offset, Type *TargetTy,
1455 SmallVectorImpl<Value *> &Indices,
1456 Twine NamePrefix) {
1457 PointerType *Ty = cast<PointerType>(Ptr->getType());
1458
1459 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1460 // an i8.
1461 if (Ty == IRB.getInt8PtrTy(Ty->getAddressSpace()) && TargetTy->isIntegerTy(8))
1462 return nullptr;
1463
1464 Type *ElementTy = Ty->getElementType();
1465 if (!ElementTy->isSized())
1466 return nullptr; // We can't GEP through an unsized element.
1467 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1468 if (ElementSize == 0)
1469 return nullptr; // Zero-length arrays can't help us build a natural GEP.
1470 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1471
1472 Offset -= NumSkippedElements * ElementSize;
1473 Indices.push_back(IRB.getInt(NumSkippedElements));
1474 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1475 Indices, NamePrefix);
1476 }
1477
1478 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1479 /// resulting pointer has PointerTy.
1480 ///
1481 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1482 /// and produces the pointer type desired. Where it cannot, it will try to use
1483 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1484 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1485 /// bitcast to the type.
1486 ///
1487 /// The strategy for finding the more natural GEPs is to peel off layers of the
1488 /// pointer, walking back through bit casts and GEPs, searching for a base
1489 /// pointer from which we can compute a natural GEP with the desired
1490 /// properties. The algorithm tries to fold as many constant indices into
1491 /// a single GEP as possible, thus making each GEP more independent of the
1492 /// surrounding code.
getAdjustedPtr(IRBuilderTy & IRB,const DataLayout & DL,Value * Ptr,APInt Offset,Type * PointerTy,Twine NamePrefix)1493 static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr,
1494 APInt Offset, Type *PointerTy, Twine NamePrefix) {
1495 // Even though we don't look through PHI nodes, we could be called on an
1496 // instruction in an unreachable block, which may be on a cycle.
1497 SmallPtrSet<Value *, 4> Visited;
1498 Visited.insert(Ptr);
1499 SmallVector<Value *, 4> Indices;
1500
1501 // We may end up computing an offset pointer that has the wrong type. If we
1502 // never are able to compute one directly that has the correct type, we'll
1503 // fall back to it, so keep it and the base it was computed from around here.
1504 Value *OffsetPtr = nullptr;
1505 Value *OffsetBasePtr;
1506
1507 // Remember any i8 pointer we come across to re-use if we need to do a raw
1508 // byte offset.
1509 Value *Int8Ptr = nullptr;
1510 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1511
1512 Type *TargetTy = PointerTy->getPointerElementType();
1513
1514 do {
1515 // First fold any existing GEPs into the offset.
1516 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1517 APInt GEPOffset(Offset.getBitWidth(), 0);
1518 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1519 break;
1520 Offset += GEPOffset;
1521 Ptr = GEP->getPointerOperand();
1522 if (!Visited.insert(Ptr).second)
1523 break;
1524 }
1525
1526 // See if we can perform a natural GEP here.
1527 Indices.clear();
1528 if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1529 Indices, NamePrefix)) {
1530 // If we have a new natural pointer at the offset, clear out any old
1531 // offset pointer we computed. Unless it is the base pointer or
1532 // a non-instruction, we built a GEP we don't need. Zap it.
1533 if (OffsetPtr && OffsetPtr != OffsetBasePtr)
1534 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr)) {
1535 assert(I->use_empty() && "Built a GEP with uses some how!");
1536 I->eraseFromParent();
1537 }
1538 OffsetPtr = P;
1539 OffsetBasePtr = Ptr;
1540 // If we also found a pointer of the right type, we're done.
1541 if (P->getType() == PointerTy)
1542 return P;
1543 }
1544
1545 // Stash this pointer if we've found an i8*.
1546 if (Ptr->getType()->isIntegerTy(8)) {
1547 Int8Ptr = Ptr;
1548 Int8PtrOffset = Offset;
1549 }
1550
1551 // Peel off a layer of the pointer and update the offset appropriately.
1552 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1553 Ptr = cast<Operator>(Ptr)->getOperand(0);
1554 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1555 if (GA->isInterposable())
1556 break;
1557 Ptr = GA->getAliasee();
1558 } else {
1559 break;
1560 }
1561 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1562 } while (Visited.insert(Ptr).second);
1563
1564 if (!OffsetPtr) {
1565 if (!Int8Ptr) {
1566 Int8Ptr = IRB.CreateBitCast(
1567 Ptr, IRB.getInt8PtrTy(PointerTy->getPointerAddressSpace()),
1568 NamePrefix + "sroa_raw_cast");
1569 Int8PtrOffset = Offset;
1570 }
1571
1572 OffsetPtr = Int8PtrOffset == 0
1573 ? Int8Ptr
1574 : IRB.CreateInBoundsGEP(IRB.getInt8Ty(), Int8Ptr,
1575 IRB.getInt(Int8PtrOffset),
1576 NamePrefix + "sroa_raw_idx");
1577 }
1578 Ptr = OffsetPtr;
1579
1580 // On the off chance we were targeting i8*, guard the bitcast here.
1581 if (Ptr->getType() != PointerTy)
1582 Ptr = IRB.CreateBitCast(Ptr, PointerTy, NamePrefix + "sroa_cast");
1583
1584 return Ptr;
1585 }
1586
1587 /// \brief Compute the adjusted alignment for a load or store from an offset.
getAdjustedAlignment(Instruction * I,uint64_t Offset,const DataLayout & DL)1588 static unsigned getAdjustedAlignment(Instruction *I, uint64_t Offset,
1589 const DataLayout &DL) {
1590 unsigned Alignment;
1591 Type *Ty;
1592 if (auto *LI = dyn_cast<LoadInst>(I)) {
1593 Alignment = LI->getAlignment();
1594 Ty = LI->getType();
1595 } else if (auto *SI = dyn_cast<StoreInst>(I)) {
1596 Alignment = SI->getAlignment();
1597 Ty = SI->getValueOperand()->getType();
1598 } else {
1599 llvm_unreachable("Only loads and stores are allowed!");
1600 }
1601
1602 if (!Alignment)
1603 Alignment = DL.getABITypeAlignment(Ty);
1604
1605 return MinAlign(Alignment, Offset);
1606 }
1607
1608 /// \brief Test whether we can convert a value from the old to the new type.
1609 ///
1610 /// This predicate should be used to guard calls to convertValue in order to
1611 /// ensure that we only try to convert viable values. The strategy is that we
1612 /// will peel off single element struct and array wrappings to get to an
1613 /// underlying value, and convert that value.
canConvertValue(const DataLayout & DL,Type * OldTy,Type * NewTy)1614 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1615 if (OldTy == NewTy)
1616 return true;
1617
1618 // For integer types, we can't handle any bit-width differences. This would
1619 // break both vector conversions with extension and introduce endianness
1620 // issues when in conjunction with loads and stores.
1621 if (isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) {
1622 assert(cast<IntegerType>(OldTy)->getBitWidth() !=
1623 cast<IntegerType>(NewTy)->getBitWidth() &&
1624 "We can't have the same bitwidth for different int types");
1625 return false;
1626 }
1627
1628 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1629 return false;
1630 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1631 return false;
1632
1633 // We can convert pointers to integers and vice-versa. Same for vectors
1634 // of pointers and integers.
1635 OldTy = OldTy->getScalarType();
1636 NewTy = NewTy->getScalarType();
1637 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1638 if (NewTy->isPointerTy() && OldTy->isPointerTy()) {
1639 return cast<PointerType>(NewTy)->getPointerAddressSpace() ==
1640 cast<PointerType>(OldTy)->getPointerAddressSpace();
1641 }
1642 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
1643 return true;
1644 return false;
1645 }
1646
1647 return true;
1648 }
1649
1650 /// \brief Generic routine to convert an SSA value to a value of a different
1651 /// type.
1652 ///
1653 /// This will try various different casting techniques, such as bitcasts,
1654 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1655 /// two types for viability with this routine.
convertValue(const DataLayout & DL,IRBuilderTy & IRB,Value * V,Type * NewTy)1656 static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1657 Type *NewTy) {
1658 Type *OldTy = V->getType();
1659 assert(canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type");
1660
1661 if (OldTy == NewTy)
1662 return V;
1663
1664 assert(!(isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) &&
1665 "Integer types must be the exact same to convert.");
1666
1667 // See if we need inttoptr for this type pair. A cast involving both scalars
1668 // and vectors requires and additional bitcast.
1669 if (OldTy->getScalarType()->isIntegerTy() &&
1670 NewTy->getScalarType()->isPointerTy()) {
1671 // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1672 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1673 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1674 NewTy);
1675
1676 // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1677 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1678 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1679 NewTy);
1680
1681 return IRB.CreateIntToPtr(V, NewTy);
1682 }
1683
1684 // See if we need ptrtoint for this type pair. A cast involving both scalars
1685 // and vectors requires and additional bitcast.
1686 if (OldTy->getScalarType()->isPointerTy() &&
1687 NewTy->getScalarType()->isIntegerTy()) {
1688 // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1689 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1690 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1691 NewTy);
1692
1693 // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1694 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1695 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1696 NewTy);
1697
1698 return IRB.CreatePtrToInt(V, NewTy);
1699 }
1700
1701 return IRB.CreateBitCast(V, NewTy);
1702 }
1703
1704 /// \brief Test whether the given slice use can be promoted to a vector.
1705 ///
1706 /// This function is called to test each entry in a partition which is slated
1707 /// for a single slice.
isVectorPromotionViableForSlice(Partition & P,const Slice & S,VectorType * Ty,uint64_t ElementSize,const DataLayout & DL)1708 static bool isVectorPromotionViableForSlice(Partition &P, const Slice &S,
1709 VectorType *Ty,
1710 uint64_t ElementSize,
1711 const DataLayout &DL) {
1712 // First validate the slice offsets.
1713 uint64_t BeginOffset =
1714 std::max(S.beginOffset(), P.beginOffset()) - P.beginOffset();
1715 uint64_t BeginIndex = BeginOffset / ElementSize;
1716 if (BeginIndex * ElementSize != BeginOffset ||
1717 BeginIndex >= Ty->getNumElements())
1718 return false;
1719 uint64_t EndOffset =
1720 std::min(S.endOffset(), P.endOffset()) - P.beginOffset();
1721 uint64_t EndIndex = EndOffset / ElementSize;
1722 if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
1723 return false;
1724
1725 assert(EndIndex > BeginIndex && "Empty vector!");
1726 uint64_t NumElements = EndIndex - BeginIndex;
1727 Type *SliceTy = (NumElements == 1)
1728 ? Ty->getElementType()
1729 : VectorType::get(Ty->getElementType(), NumElements);
1730
1731 Type *SplitIntTy =
1732 Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1733
1734 Use *U = S.getUse();
1735
1736 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1737 if (MI->isVolatile())
1738 return false;
1739 if (!S.isSplittable())
1740 return false; // Skip any unsplittable intrinsics.
1741 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1742 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1743 II->getIntrinsicID() != Intrinsic::lifetime_end)
1744 return false;
1745 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1746 // Disable vector promotion when there are loads or stores of an FCA.
1747 return false;
1748 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1749 if (LI->isVolatile())
1750 return false;
1751 Type *LTy = LI->getType();
1752 if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1753 assert(LTy->isIntegerTy());
1754 LTy = SplitIntTy;
1755 }
1756 if (!canConvertValue(DL, SliceTy, LTy))
1757 return false;
1758 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1759 if (SI->isVolatile())
1760 return false;
1761 Type *STy = SI->getValueOperand()->getType();
1762 if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1763 assert(STy->isIntegerTy());
1764 STy = SplitIntTy;
1765 }
1766 if (!canConvertValue(DL, STy, SliceTy))
1767 return false;
1768 } else {
1769 return false;
1770 }
1771
1772 return true;
1773 }
1774
1775 /// \brief Test whether the given alloca partitioning and range of slices can be
1776 /// promoted to a vector.
1777 ///
1778 /// This is a quick test to check whether we can rewrite a particular alloca
1779 /// partition (and its newly formed alloca) into a vector alloca with only
1780 /// whole-vector loads and stores such that it could be promoted to a vector
1781 /// SSA value. We only can ensure this for a limited set of operations, and we
1782 /// don't want to do the rewrites unless we are confident that the result will
1783 /// be promotable, so we have an early test here.
isVectorPromotionViable(Partition & P,const DataLayout & DL)1784 static VectorType *isVectorPromotionViable(Partition &P, const DataLayout &DL) {
1785 // Collect the candidate types for vector-based promotion. Also track whether
1786 // we have different element types.
1787 SmallVector<VectorType *, 4> CandidateTys;
1788 Type *CommonEltTy = nullptr;
1789 bool HaveCommonEltTy = true;
1790 auto CheckCandidateType = [&](Type *Ty) {
1791 if (auto *VTy = dyn_cast<VectorType>(Ty)) {
1792 CandidateTys.push_back(VTy);
1793 if (!CommonEltTy)
1794 CommonEltTy = VTy->getElementType();
1795 else if (CommonEltTy != VTy->getElementType())
1796 HaveCommonEltTy = false;
1797 }
1798 };
1799 // Consider any loads or stores that are the exact size of the slice.
1800 for (const Slice &S : P)
1801 if (S.beginOffset() == P.beginOffset() &&
1802 S.endOffset() == P.endOffset()) {
1803 if (auto *LI = dyn_cast<LoadInst>(S.getUse()->getUser()))
1804 CheckCandidateType(LI->getType());
1805 else if (auto *SI = dyn_cast<StoreInst>(S.getUse()->getUser()))
1806 CheckCandidateType(SI->getValueOperand()->getType());
1807 }
1808
1809 // If we didn't find a vector type, nothing to do here.
1810 if (CandidateTys.empty())
1811 return nullptr;
1812
1813 // Remove non-integer vector types if we had multiple common element types.
1814 // FIXME: It'd be nice to replace them with integer vector types, but we can't
1815 // do that until all the backends are known to produce good code for all
1816 // integer vector types.
1817 if (!HaveCommonEltTy) {
1818 CandidateTys.erase(std::remove_if(CandidateTys.begin(), CandidateTys.end(),
1819 [](VectorType *VTy) {
1820 return !VTy->getElementType()->isIntegerTy();
1821 }),
1822 CandidateTys.end());
1823
1824 // If there were no integer vector types, give up.
1825 if (CandidateTys.empty())
1826 return nullptr;
1827
1828 // Rank the remaining candidate vector types. This is easy because we know
1829 // they're all integer vectors. We sort by ascending number of elements.
1830 auto RankVectorTypes = [&DL](VectorType *RHSTy, VectorType *LHSTy) {
1831 assert(DL.getTypeSizeInBits(RHSTy) == DL.getTypeSizeInBits(LHSTy) &&
1832 "Cannot have vector types of different sizes!");
1833 assert(RHSTy->getElementType()->isIntegerTy() &&
1834 "All non-integer types eliminated!");
1835 assert(LHSTy->getElementType()->isIntegerTy() &&
1836 "All non-integer types eliminated!");
1837 return RHSTy->getNumElements() < LHSTy->getNumElements();
1838 };
1839 std::sort(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes);
1840 CandidateTys.erase(
1841 std::unique(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes),
1842 CandidateTys.end());
1843 } else {
1844 // The only way to have the same element type in every vector type is to
1845 // have the same vector type. Check that and remove all but one.
1846 #ifndef NDEBUG
1847 for (VectorType *VTy : CandidateTys) {
1848 assert(VTy->getElementType() == CommonEltTy &&
1849 "Unaccounted for element type!");
1850 assert(VTy == CandidateTys[0] &&
1851 "Different vector types with the same element type!");
1852 }
1853 #endif
1854 CandidateTys.resize(1);
1855 }
1856
1857 // Try each vector type, and return the one which works.
1858 auto CheckVectorTypeForPromotion = [&](VectorType *VTy) {
1859 uint64_t ElementSize = DL.getTypeSizeInBits(VTy->getElementType());
1860
1861 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1862 // that aren't byte sized.
1863 if (ElementSize % 8)
1864 return false;
1865 assert((DL.getTypeSizeInBits(VTy) % 8) == 0 &&
1866 "vector size not a multiple of element size?");
1867 ElementSize /= 8;
1868
1869 for (const Slice &S : P)
1870 if (!isVectorPromotionViableForSlice(P, S, VTy, ElementSize, DL))
1871 return false;
1872
1873 for (const Slice *S : P.splitSliceTails())
1874 if (!isVectorPromotionViableForSlice(P, *S, VTy, ElementSize, DL))
1875 return false;
1876
1877 return true;
1878 };
1879 for (VectorType *VTy : CandidateTys)
1880 if (CheckVectorTypeForPromotion(VTy))
1881 return VTy;
1882
1883 return nullptr;
1884 }
1885
1886 /// \brief Test whether a slice of an alloca is valid for integer widening.
1887 ///
1888 /// This implements the necessary checking for the \c isIntegerWideningViable
1889 /// test below on a single slice of the alloca.
isIntegerWideningViableForSlice(const Slice & S,uint64_t AllocBeginOffset,Type * AllocaTy,const DataLayout & DL,bool & WholeAllocaOp)1890 static bool isIntegerWideningViableForSlice(const Slice &S,
1891 uint64_t AllocBeginOffset,
1892 Type *AllocaTy,
1893 const DataLayout &DL,
1894 bool &WholeAllocaOp) {
1895 uint64_t Size = DL.getTypeStoreSize(AllocaTy);
1896
1897 uint64_t RelBegin = S.beginOffset() - AllocBeginOffset;
1898 uint64_t RelEnd = S.endOffset() - AllocBeginOffset;
1899
1900 // We can't reasonably handle cases where the load or store extends past
1901 // the end of the alloca's type and into its padding.
1902 if (RelEnd > Size)
1903 return false;
1904
1905 Use *U = S.getUse();
1906
1907 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1908 if (LI->isVolatile())
1909 return false;
1910 // We can't handle loads that extend past the allocated memory.
1911 if (DL.getTypeStoreSize(LI->getType()) > Size)
1912 return false;
1913 // Note that we don't count vector loads or stores as whole-alloca
1914 // operations which enable integer widening because we would prefer to use
1915 // vector widening instead.
1916 if (!isa<VectorType>(LI->getType()) && RelBegin == 0 && RelEnd == Size)
1917 WholeAllocaOp = true;
1918 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
1919 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1920 return false;
1921 } else if (RelBegin != 0 || RelEnd != Size ||
1922 !canConvertValue(DL, AllocaTy, LI->getType())) {
1923 // Non-integer loads need to be convertible from the alloca type so that
1924 // they are promotable.
1925 return false;
1926 }
1927 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1928 Type *ValueTy = SI->getValueOperand()->getType();
1929 if (SI->isVolatile())
1930 return false;
1931 // We can't handle stores that extend past the allocated memory.
1932 if (DL.getTypeStoreSize(ValueTy) > Size)
1933 return false;
1934 // Note that we don't count vector loads or stores as whole-alloca
1935 // operations which enable integer widening because we would prefer to use
1936 // vector widening instead.
1937 if (!isa<VectorType>(ValueTy) && RelBegin == 0 && RelEnd == Size)
1938 WholeAllocaOp = true;
1939 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
1940 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1941 return false;
1942 } else if (RelBegin != 0 || RelEnd != Size ||
1943 !canConvertValue(DL, ValueTy, AllocaTy)) {
1944 // Non-integer stores need to be convertible to the alloca type so that
1945 // they are promotable.
1946 return false;
1947 }
1948 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1949 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
1950 return false;
1951 if (!S.isSplittable())
1952 return false; // Skip any unsplittable intrinsics.
1953 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1954 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1955 II->getIntrinsicID() != Intrinsic::lifetime_end)
1956 return false;
1957 } else {
1958 return false;
1959 }
1960
1961 return true;
1962 }
1963
1964 /// \brief Test whether the given alloca partition's integer operations can be
1965 /// widened to promotable ones.
1966 ///
1967 /// This is a quick test to check whether we can rewrite the integer loads and
1968 /// stores to a particular alloca into wider loads and stores and be able to
1969 /// promote the resulting alloca.
isIntegerWideningViable(Partition & P,Type * AllocaTy,const DataLayout & DL)1970 static bool isIntegerWideningViable(Partition &P, Type *AllocaTy,
1971 const DataLayout &DL) {
1972 uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
1973 // Don't create integer types larger than the maximum bitwidth.
1974 if (SizeInBits > IntegerType::MAX_INT_BITS)
1975 return false;
1976
1977 // Don't try to handle allocas with bit-padding.
1978 if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
1979 return false;
1980
1981 // We need to ensure that an integer type with the appropriate bitwidth can
1982 // be converted to the alloca type, whatever that is. We don't want to force
1983 // the alloca itself to have an integer type if there is a more suitable one.
1984 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
1985 if (!canConvertValue(DL, AllocaTy, IntTy) ||
1986 !canConvertValue(DL, IntTy, AllocaTy))
1987 return false;
1988
1989 // While examining uses, we ensure that the alloca has a covering load or
1990 // store. We don't want to widen the integer operations only to fail to
1991 // promote due to some other unsplittable entry (which we may make splittable
1992 // later). However, if there are only splittable uses, go ahead and assume
1993 // that we cover the alloca.
1994 // FIXME: We shouldn't consider split slices that happen to start in the
1995 // partition here...
1996 bool WholeAllocaOp =
1997 P.begin() != P.end() ? false : DL.isLegalInteger(SizeInBits);
1998
1999 for (const Slice &S : P)
2000 if (!isIntegerWideningViableForSlice(S, P.beginOffset(), AllocaTy, DL,
2001 WholeAllocaOp))
2002 return false;
2003
2004 for (const Slice *S : P.splitSliceTails())
2005 if (!isIntegerWideningViableForSlice(*S, P.beginOffset(), AllocaTy, DL,
2006 WholeAllocaOp))
2007 return false;
2008
2009 return WholeAllocaOp;
2010 }
2011
extractInteger(const DataLayout & DL,IRBuilderTy & IRB,Value * V,IntegerType * Ty,uint64_t Offset,const Twine & Name)2012 static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
2013 IntegerType *Ty, uint64_t Offset,
2014 const Twine &Name) {
2015 DEBUG(dbgs() << " start: " << *V << "\n");
2016 IntegerType *IntTy = cast<IntegerType>(V->getType());
2017 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2018 "Element extends past full value");
2019 uint64_t ShAmt = 8 * Offset;
2020 if (DL.isBigEndian())
2021 ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2022 if (ShAmt) {
2023 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
2024 DEBUG(dbgs() << " shifted: " << *V << "\n");
2025 }
2026 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2027 "Cannot extract to a larger integer!");
2028 if (Ty != IntTy) {
2029 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
2030 DEBUG(dbgs() << " trunced: " << *V << "\n");
2031 }
2032 return V;
2033 }
2034
insertInteger(const DataLayout & DL,IRBuilderTy & IRB,Value * Old,Value * V,uint64_t Offset,const Twine & Name)2035 static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
2036 Value *V, uint64_t Offset, const Twine &Name) {
2037 IntegerType *IntTy = cast<IntegerType>(Old->getType());
2038 IntegerType *Ty = cast<IntegerType>(V->getType());
2039 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2040 "Cannot insert a larger integer!");
2041 DEBUG(dbgs() << " start: " << *V << "\n");
2042 if (Ty != IntTy) {
2043 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
2044 DEBUG(dbgs() << " extended: " << *V << "\n");
2045 }
2046 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2047 "Element store outside of alloca store");
2048 uint64_t ShAmt = 8 * Offset;
2049 if (DL.isBigEndian())
2050 ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2051 if (ShAmt) {
2052 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
2053 DEBUG(dbgs() << " shifted: " << *V << "\n");
2054 }
2055
2056 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
2057 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
2058 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
2059 DEBUG(dbgs() << " masked: " << *Old << "\n");
2060 V = IRB.CreateOr(Old, V, Name + ".insert");
2061 DEBUG(dbgs() << " inserted: " << *V << "\n");
2062 }
2063 return V;
2064 }
2065
extractVector(IRBuilderTy & IRB,Value * V,unsigned BeginIndex,unsigned EndIndex,const Twine & Name)2066 static Value *extractVector(IRBuilderTy &IRB, Value *V, unsigned BeginIndex,
2067 unsigned EndIndex, const Twine &Name) {
2068 VectorType *VecTy = cast<VectorType>(V->getType());
2069 unsigned NumElements = EndIndex - BeginIndex;
2070 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2071
2072 if (NumElements == VecTy->getNumElements())
2073 return V;
2074
2075 if (NumElements == 1) {
2076 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
2077 Name + ".extract");
2078 DEBUG(dbgs() << " extract: " << *V << "\n");
2079 return V;
2080 }
2081
2082 SmallVector<Constant *, 8> Mask;
2083 Mask.reserve(NumElements);
2084 for (unsigned i = BeginIndex; i != EndIndex; ++i)
2085 Mask.push_back(IRB.getInt32(i));
2086 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2087 ConstantVector::get(Mask), Name + ".extract");
2088 DEBUG(dbgs() << " shuffle: " << *V << "\n");
2089 return V;
2090 }
2091
insertVector(IRBuilderTy & IRB,Value * Old,Value * V,unsigned BeginIndex,const Twine & Name)2092 static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
2093 unsigned BeginIndex, const Twine &Name) {
2094 VectorType *VecTy = cast<VectorType>(Old->getType());
2095 assert(VecTy && "Can only insert a vector into a vector");
2096
2097 VectorType *Ty = dyn_cast<VectorType>(V->getType());
2098 if (!Ty) {
2099 // Single element to insert.
2100 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
2101 Name + ".insert");
2102 DEBUG(dbgs() << " insert: " << *V << "\n");
2103 return V;
2104 }
2105
2106 assert(Ty->getNumElements() <= VecTy->getNumElements() &&
2107 "Too many elements!");
2108 if (Ty->getNumElements() == VecTy->getNumElements()) {
2109 assert(V->getType() == VecTy && "Vector type mismatch");
2110 return V;
2111 }
2112 unsigned EndIndex = BeginIndex + Ty->getNumElements();
2113
2114 // When inserting a smaller vector into the larger to store, we first
2115 // use a shuffle vector to widen it with undef elements, and then
2116 // a second shuffle vector to select between the loaded vector and the
2117 // incoming vector.
2118 SmallVector<Constant *, 8> Mask;
2119 Mask.reserve(VecTy->getNumElements());
2120 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2121 if (i >= BeginIndex && i < EndIndex)
2122 Mask.push_back(IRB.getInt32(i - BeginIndex));
2123 else
2124 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
2125 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2126 ConstantVector::get(Mask), Name + ".expand");
2127 DEBUG(dbgs() << " shuffle: " << *V << "\n");
2128
2129 Mask.clear();
2130 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2131 Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
2132
2133 V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
2134
2135 DEBUG(dbgs() << " blend: " << *V << "\n");
2136 return V;
2137 }
2138
2139 /// \brief Visitor to rewrite instructions using p particular slice of an alloca
2140 /// to use a new alloca.
2141 ///
2142 /// Also implements the rewriting to vector-based accesses when the partition
2143 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2144 /// lives here.
2145 class llvm::sroa::AllocaSliceRewriter
2146 : public InstVisitor<AllocaSliceRewriter, bool> {
2147 // Befriend the base class so it can delegate to private visit methods.
2148 friend class llvm::InstVisitor<AllocaSliceRewriter, bool>;
2149 typedef llvm::InstVisitor<AllocaSliceRewriter, bool> Base;
2150
2151 const DataLayout &DL;
2152 AllocaSlices &AS;
2153 SROA &Pass;
2154 AllocaInst &OldAI, &NewAI;
2155 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2156 Type *NewAllocaTy;
2157
2158 // This is a convenience and flag variable that will be null unless the new
2159 // alloca's integer operations should be widened to this integer type due to
2160 // passing isIntegerWideningViable above. If it is non-null, the desired
2161 // integer type will be stored here for easy access during rewriting.
2162 IntegerType *IntTy;
2163
2164 // If we are rewriting an alloca partition which can be written as pure
2165 // vector operations, we stash extra information here. When VecTy is
2166 // non-null, we have some strict guarantees about the rewritten alloca:
2167 // - The new alloca is exactly the size of the vector type here.
2168 // - The accesses all either map to the entire vector or to a single
2169 // element.
2170 // - The set of accessing instructions is only one of those handled above
2171 // in isVectorPromotionViable. Generally these are the same access kinds
2172 // which are promotable via mem2reg.
2173 VectorType *VecTy;
2174 Type *ElementTy;
2175 uint64_t ElementSize;
2176
2177 // The original offset of the slice currently being rewritten relative to
2178 // the original alloca.
2179 uint64_t BeginOffset, EndOffset;
2180 // The new offsets of the slice currently being rewritten relative to the
2181 // original alloca.
2182 uint64_t NewBeginOffset, NewEndOffset;
2183
2184 uint64_t SliceSize;
2185 bool IsSplittable;
2186 bool IsSplit;
2187 Use *OldUse;
2188 Instruction *OldPtr;
2189
2190 // Track post-rewrite users which are PHI nodes and Selects.
2191 SmallPtrSetImpl<PHINode *> &PHIUsers;
2192 SmallPtrSetImpl<SelectInst *> &SelectUsers;
2193
2194 // Utility IR builder, whose name prefix is setup for each visited use, and
2195 // the insertion point is set to point to the user.
2196 IRBuilderTy IRB;
2197
2198 public:
AllocaSliceRewriter(const DataLayout & DL,AllocaSlices & AS,SROA & Pass,AllocaInst & OldAI,AllocaInst & NewAI,uint64_t NewAllocaBeginOffset,uint64_t NewAllocaEndOffset,bool IsIntegerPromotable,VectorType * PromotableVecTy,SmallPtrSetImpl<PHINode * > & PHIUsers,SmallPtrSetImpl<SelectInst * > & SelectUsers)2199 AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &AS, SROA &Pass,
2200 AllocaInst &OldAI, AllocaInst &NewAI,
2201 uint64_t NewAllocaBeginOffset,
2202 uint64_t NewAllocaEndOffset, bool IsIntegerPromotable,
2203 VectorType *PromotableVecTy,
2204 SmallPtrSetImpl<PHINode *> &PHIUsers,
2205 SmallPtrSetImpl<SelectInst *> &SelectUsers)
2206 : DL(DL), AS(AS), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
2207 NewAllocaBeginOffset(NewAllocaBeginOffset),
2208 NewAllocaEndOffset(NewAllocaEndOffset),
2209 NewAllocaTy(NewAI.getAllocatedType()),
2210 IntTy(IsIntegerPromotable
2211 ? Type::getIntNTy(
2212 NewAI.getContext(),
2213 DL.getTypeSizeInBits(NewAI.getAllocatedType()))
2214 : nullptr),
2215 VecTy(PromotableVecTy),
2216 ElementTy(VecTy ? VecTy->getElementType() : nullptr),
2217 ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
2218 BeginOffset(), EndOffset(), IsSplittable(), IsSplit(), OldUse(),
2219 OldPtr(), PHIUsers(PHIUsers), SelectUsers(SelectUsers),
2220 IRB(NewAI.getContext(), ConstantFolder()) {
2221 if (VecTy) {
2222 assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
2223 "Only multiple-of-8 sized vector elements are viable");
2224 ++NumVectorized;
2225 }
2226 assert((!IntTy && !VecTy) || (IntTy && !VecTy) || (!IntTy && VecTy));
2227 }
2228
visit(AllocaSlices::const_iterator I)2229 bool visit(AllocaSlices::const_iterator I) {
2230 bool CanSROA = true;
2231 BeginOffset = I->beginOffset();
2232 EndOffset = I->endOffset();
2233 IsSplittable = I->isSplittable();
2234 IsSplit =
2235 BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
2236 DEBUG(dbgs() << " rewriting " << (IsSplit ? "split " : ""));
2237 DEBUG(AS.printSlice(dbgs(), I, ""));
2238 DEBUG(dbgs() << "\n");
2239
2240 // Compute the intersecting offset range.
2241 assert(BeginOffset < NewAllocaEndOffset);
2242 assert(EndOffset > NewAllocaBeginOffset);
2243 NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2244 NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2245
2246 SliceSize = NewEndOffset - NewBeginOffset;
2247
2248 OldUse = I->getUse();
2249 OldPtr = cast<Instruction>(OldUse->get());
2250
2251 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2252 IRB.SetInsertPoint(OldUserI);
2253 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2254 IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
2255
2256 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2257 if (VecTy || IntTy)
2258 assert(CanSROA);
2259 return CanSROA;
2260 }
2261
2262 private:
2263 // Make sure the other visit overloads are visible.
2264 using Base::visit;
2265
2266 // Every instruction which can end up as a user must have a rewrite rule.
visitInstruction(Instruction & I)2267 bool visitInstruction(Instruction &I) {
2268 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2269 llvm_unreachable("No rewrite rule for this instruction!");
2270 }
2271
getNewAllocaSlicePtr(IRBuilderTy & IRB,Type * PointerTy)2272 Value *getNewAllocaSlicePtr(IRBuilderTy &IRB, Type *PointerTy) {
2273 // Note that the offset computation can use BeginOffset or NewBeginOffset
2274 // interchangeably for unsplit slices.
2275 assert(IsSplit || BeginOffset == NewBeginOffset);
2276 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2277
2278 #ifndef NDEBUG
2279 StringRef OldName = OldPtr->getName();
2280 // Skip through the last '.sroa.' component of the name.
2281 size_t LastSROAPrefix = OldName.rfind(".sroa.");
2282 if (LastSROAPrefix != StringRef::npos) {
2283 OldName = OldName.substr(LastSROAPrefix + strlen(".sroa."));
2284 // Look for an SROA slice index.
2285 size_t IndexEnd = OldName.find_first_not_of("0123456789");
2286 if (IndexEnd != StringRef::npos && OldName[IndexEnd] == '.') {
2287 // Strip the index and look for the offset.
2288 OldName = OldName.substr(IndexEnd + 1);
2289 size_t OffsetEnd = OldName.find_first_not_of("0123456789");
2290 if (OffsetEnd != StringRef::npos && OldName[OffsetEnd] == '.')
2291 // Strip the offset.
2292 OldName = OldName.substr(OffsetEnd + 1);
2293 }
2294 }
2295 // Strip any SROA suffixes as well.
2296 OldName = OldName.substr(0, OldName.find(".sroa_"));
2297 #endif
2298
2299 return getAdjustedPtr(IRB, DL, &NewAI,
2300 APInt(DL.getPointerSizeInBits(), Offset), PointerTy,
2301 #ifndef NDEBUG
2302 Twine(OldName) + "."
2303 #else
2304 Twine()
2305 #endif
2306 );
2307 }
2308
2309 /// \brief Compute suitable alignment to access this slice of the *new*
2310 /// alloca.
2311 ///
2312 /// You can optionally pass a type to this routine and if that type's ABI
2313 /// alignment is itself suitable, this will return zero.
getSliceAlign(Type * Ty=nullptr)2314 unsigned getSliceAlign(Type *Ty = nullptr) {
2315 unsigned NewAIAlign = NewAI.getAlignment();
2316 if (!NewAIAlign)
2317 NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
2318 unsigned Align =
2319 MinAlign(NewAIAlign, NewBeginOffset - NewAllocaBeginOffset);
2320 return (Ty && Align == DL.getABITypeAlignment(Ty)) ? 0 : Align;
2321 }
2322
getIndex(uint64_t Offset)2323 unsigned getIndex(uint64_t Offset) {
2324 assert(VecTy && "Can only call getIndex when rewriting a vector");
2325 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2326 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2327 uint32_t Index = RelOffset / ElementSize;
2328 assert(Index * ElementSize == RelOffset);
2329 return Index;
2330 }
2331
deleteIfTriviallyDead(Value * V)2332 void deleteIfTriviallyDead(Value *V) {
2333 Instruction *I = cast<Instruction>(V);
2334 if (isInstructionTriviallyDead(I))
2335 Pass.DeadInsts.insert(I);
2336 }
2337
rewriteVectorizedLoadInst()2338 Value *rewriteVectorizedLoadInst() {
2339 unsigned BeginIndex = getIndex(NewBeginOffset);
2340 unsigned EndIndex = getIndex(NewEndOffset);
2341 assert(EndIndex > BeginIndex && "Empty vector!");
2342
2343 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2344 return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2345 }
2346
rewriteIntegerLoad(LoadInst & LI)2347 Value *rewriteIntegerLoad(LoadInst &LI) {
2348 assert(IntTy && "We cannot insert an integer to the alloca");
2349 assert(!LI.isVolatile());
2350 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2351 V = convertValue(DL, IRB, V, IntTy);
2352 assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2353 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2354 if (Offset > 0 || NewEndOffset < NewAllocaEndOffset) {
2355 IntegerType *ExtractTy = Type::getIntNTy(LI.getContext(), SliceSize * 8);
2356 V = extractInteger(DL, IRB, V, ExtractTy, Offset, "extract");
2357 }
2358 // It is possible that the extracted type is not the load type. This
2359 // happens if there is a load past the end of the alloca, and as
2360 // a consequence the slice is narrower but still a candidate for integer
2361 // lowering. To handle this case, we just zero extend the extracted
2362 // integer.
2363 assert(cast<IntegerType>(LI.getType())->getBitWidth() >= SliceSize * 8 &&
2364 "Can only handle an extract for an overly wide load");
2365 if (cast<IntegerType>(LI.getType())->getBitWidth() > SliceSize * 8)
2366 V = IRB.CreateZExt(V, LI.getType());
2367 return V;
2368 }
2369
visitLoadInst(LoadInst & LI)2370 bool visitLoadInst(LoadInst &LI) {
2371 DEBUG(dbgs() << " original: " << LI << "\n");
2372 Value *OldOp = LI.getOperand(0);
2373 assert(OldOp == OldPtr);
2374
2375 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), SliceSize * 8)
2376 : LI.getType();
2377 const bool IsLoadPastEnd = DL.getTypeStoreSize(TargetTy) > SliceSize;
2378 bool IsPtrAdjusted = false;
2379 Value *V;
2380 if (VecTy) {
2381 V = rewriteVectorizedLoadInst();
2382 } else if (IntTy && LI.getType()->isIntegerTy()) {
2383 V = rewriteIntegerLoad(LI);
2384 } else if (NewBeginOffset == NewAllocaBeginOffset &&
2385 NewEndOffset == NewAllocaEndOffset &&
2386 (canConvertValue(DL, NewAllocaTy, TargetTy) ||
2387 (IsLoadPastEnd && NewAllocaTy->isIntegerTy() &&
2388 TargetTy->isIntegerTy()))) {
2389 LoadInst *NewLI = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2390 LI.isVolatile(), LI.getName());
2391 if (LI.isVolatile())
2392 NewLI->setAtomic(LI.getOrdering(), LI.getSynchScope());
2393 V = NewLI;
2394
2395 // If this is an integer load past the end of the slice (which means the
2396 // bytes outside the slice are undef or this load is dead) just forcibly
2397 // fix the integer size with correct handling of endianness.
2398 if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2399 if (auto *TITy = dyn_cast<IntegerType>(TargetTy))
2400 if (AITy->getBitWidth() < TITy->getBitWidth()) {
2401 V = IRB.CreateZExt(V, TITy, "load.ext");
2402 if (DL.isBigEndian())
2403 V = IRB.CreateShl(V, TITy->getBitWidth() - AITy->getBitWidth(),
2404 "endian_shift");
2405 }
2406 } else {
2407 Type *LTy = TargetTy->getPointerTo();
2408 LoadInst *NewLI = IRB.CreateAlignedLoad(getNewAllocaSlicePtr(IRB, LTy),
2409 getSliceAlign(TargetTy),
2410 LI.isVolatile(), LI.getName());
2411 if (LI.isVolatile())
2412 NewLI->setAtomic(LI.getOrdering(), LI.getSynchScope());
2413
2414 V = NewLI;
2415 IsPtrAdjusted = true;
2416 }
2417 V = convertValue(DL, IRB, V, TargetTy);
2418
2419 if (IsSplit) {
2420 assert(!LI.isVolatile());
2421 assert(LI.getType()->isIntegerTy() &&
2422 "Only integer type loads and stores are split");
2423 assert(SliceSize < DL.getTypeStoreSize(LI.getType()) &&
2424 "Split load isn't smaller than original load");
2425 assert(LI.getType()->getIntegerBitWidth() ==
2426 DL.getTypeStoreSizeInBits(LI.getType()) &&
2427 "Non-byte-multiple bit width");
2428 // Move the insertion point just past the load so that we can refer to it.
2429 IRB.SetInsertPoint(&*std::next(BasicBlock::iterator(&LI)));
2430 // Create a placeholder value with the same type as LI to use as the
2431 // basis for the new value. This allows us to replace the uses of LI with
2432 // the computed value, and then replace the placeholder with LI, leaving
2433 // LI only used for this computation.
2434 Value *Placeholder =
2435 new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
2436 V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset - BeginOffset,
2437 "insert");
2438 LI.replaceAllUsesWith(V);
2439 Placeholder->replaceAllUsesWith(&LI);
2440 delete Placeholder;
2441 } else {
2442 LI.replaceAllUsesWith(V);
2443 }
2444
2445 Pass.DeadInsts.insert(&LI);
2446 deleteIfTriviallyDead(OldOp);
2447 DEBUG(dbgs() << " to: " << *V << "\n");
2448 return !LI.isVolatile() && !IsPtrAdjusted;
2449 }
2450
rewriteVectorizedStoreInst(Value * V,StoreInst & SI,Value * OldOp)2451 bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp) {
2452 if (V->getType() != VecTy) {
2453 unsigned BeginIndex = getIndex(NewBeginOffset);
2454 unsigned EndIndex = getIndex(NewEndOffset);
2455 assert(EndIndex > BeginIndex && "Empty vector!");
2456 unsigned NumElements = EndIndex - BeginIndex;
2457 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2458 Type *SliceTy = (NumElements == 1)
2459 ? ElementTy
2460 : VectorType::get(ElementTy, NumElements);
2461 if (V->getType() != SliceTy)
2462 V = convertValue(DL, IRB, V, SliceTy);
2463
2464 // Mix in the existing elements.
2465 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2466 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2467 }
2468 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2469 Pass.DeadInsts.insert(&SI);
2470
2471 (void)Store;
2472 DEBUG(dbgs() << " to: " << *Store << "\n");
2473 return true;
2474 }
2475
rewriteIntegerStore(Value * V,StoreInst & SI)2476 bool rewriteIntegerStore(Value *V, StoreInst &SI) {
2477 assert(IntTy && "We cannot extract an integer from the alloca");
2478 assert(!SI.isVolatile());
2479 if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2480 Value *Old =
2481 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2482 Old = convertValue(DL, IRB, Old, IntTy);
2483 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2484 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2485 V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset, "insert");
2486 }
2487 V = convertValue(DL, IRB, V, NewAllocaTy);
2488 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2489 Pass.DeadInsts.insert(&SI);
2490 (void)Store;
2491 DEBUG(dbgs() << " to: " << *Store << "\n");
2492 return true;
2493 }
2494
visitStoreInst(StoreInst & SI)2495 bool visitStoreInst(StoreInst &SI) {
2496 DEBUG(dbgs() << " original: " << SI << "\n");
2497 Value *OldOp = SI.getOperand(1);
2498 assert(OldOp == OldPtr);
2499
2500 Value *V = SI.getValueOperand();
2501
2502 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2503 // alloca that should be re-examined after promoting this alloca.
2504 if (V->getType()->isPointerTy())
2505 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2506 Pass.PostPromotionWorklist.insert(AI);
2507
2508 if (SliceSize < DL.getTypeStoreSize(V->getType())) {
2509 assert(!SI.isVolatile());
2510 assert(V->getType()->isIntegerTy() &&
2511 "Only integer type loads and stores are split");
2512 assert(V->getType()->getIntegerBitWidth() ==
2513 DL.getTypeStoreSizeInBits(V->getType()) &&
2514 "Non-byte-multiple bit width");
2515 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), SliceSize * 8);
2516 V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset - BeginOffset,
2517 "extract");
2518 }
2519
2520 if (VecTy)
2521 return rewriteVectorizedStoreInst(V, SI, OldOp);
2522 if (IntTy && V->getType()->isIntegerTy())
2523 return rewriteIntegerStore(V, SI);
2524
2525 const bool IsStorePastEnd = DL.getTypeStoreSize(V->getType()) > SliceSize;
2526 StoreInst *NewSI;
2527 if (NewBeginOffset == NewAllocaBeginOffset &&
2528 NewEndOffset == NewAllocaEndOffset &&
2529 (canConvertValue(DL, V->getType(), NewAllocaTy) ||
2530 (IsStorePastEnd && NewAllocaTy->isIntegerTy() &&
2531 V->getType()->isIntegerTy()))) {
2532 // If this is an integer store past the end of slice (and thus the bytes
2533 // past that point are irrelevant or this is unreachable), truncate the
2534 // value prior to storing.
2535 if (auto *VITy = dyn_cast<IntegerType>(V->getType()))
2536 if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2537 if (VITy->getBitWidth() > AITy->getBitWidth()) {
2538 if (DL.isBigEndian())
2539 V = IRB.CreateLShr(V, VITy->getBitWidth() - AITy->getBitWidth(),
2540 "endian_shift");
2541 V = IRB.CreateTrunc(V, AITy, "load.trunc");
2542 }
2543
2544 V = convertValue(DL, IRB, V, NewAllocaTy);
2545 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2546 SI.isVolatile());
2547 } else {
2548 Value *NewPtr = getNewAllocaSlicePtr(IRB, V->getType()->getPointerTo());
2549 NewSI = IRB.CreateAlignedStore(V, NewPtr, getSliceAlign(V->getType()),
2550 SI.isVolatile());
2551 }
2552 if (SI.isVolatile())
2553 NewSI->setAtomic(SI.getOrdering(), SI.getSynchScope());
2554 Pass.DeadInsts.insert(&SI);
2555 deleteIfTriviallyDead(OldOp);
2556
2557 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2558 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2559 }
2560
2561 /// \brief Compute an integer value from splatting an i8 across the given
2562 /// number of bytes.
2563 ///
2564 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2565 /// call this routine.
2566 /// FIXME: Heed the advice above.
2567 ///
2568 /// \param V The i8 value to splat.
2569 /// \param Size The number of bytes in the output (assuming i8 is one byte)
getIntegerSplat(Value * V,unsigned Size)2570 Value *getIntegerSplat(Value *V, unsigned Size) {
2571 assert(Size > 0 && "Expected a positive number of bytes.");
2572 IntegerType *VTy = cast<IntegerType>(V->getType());
2573 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2574 if (Size == 1)
2575 return V;
2576
2577 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size * 8);
2578 V = IRB.CreateMul(
2579 IRB.CreateZExt(V, SplatIntTy, "zext"),
2580 ConstantExpr::getUDiv(
2581 Constant::getAllOnesValue(SplatIntTy),
2582 ConstantExpr::getZExt(Constant::getAllOnesValue(V->getType()),
2583 SplatIntTy)),
2584 "isplat");
2585 return V;
2586 }
2587
2588 /// \brief Compute a vector splat for a given element value.
getVectorSplat(Value * V,unsigned NumElements)2589 Value *getVectorSplat(Value *V, unsigned NumElements) {
2590 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2591 DEBUG(dbgs() << " splat: " << *V << "\n");
2592 return V;
2593 }
2594
visitMemSetInst(MemSetInst & II)2595 bool visitMemSetInst(MemSetInst &II) {
2596 DEBUG(dbgs() << " original: " << II << "\n");
2597 assert(II.getRawDest() == OldPtr);
2598
2599 // If the memset has a variable size, it cannot be split, just adjust the
2600 // pointer to the new alloca.
2601 if (!isa<Constant>(II.getLength())) {
2602 assert(!IsSplit);
2603 assert(NewBeginOffset == BeginOffset);
2604 II.setDest(getNewAllocaSlicePtr(IRB, OldPtr->getType()));
2605 Type *CstTy = II.getAlignmentCst()->getType();
2606 II.setAlignment(ConstantInt::get(CstTy, getSliceAlign()));
2607
2608 deleteIfTriviallyDead(OldPtr);
2609 return false;
2610 }
2611
2612 // Record this instruction for deletion.
2613 Pass.DeadInsts.insert(&II);
2614
2615 Type *AllocaTy = NewAI.getAllocatedType();
2616 Type *ScalarTy = AllocaTy->getScalarType();
2617
2618 // If this doesn't map cleanly onto the alloca type, and that type isn't
2619 // a single value type, just emit a memset.
2620 if (!VecTy && !IntTy &&
2621 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2622 SliceSize != DL.getTypeStoreSize(AllocaTy) ||
2623 !AllocaTy->isSingleValueType() ||
2624 !DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy)) ||
2625 DL.getTypeSizeInBits(ScalarTy) % 8 != 0)) {
2626 Type *SizeTy = II.getLength()->getType();
2627 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2628 CallInst *New = IRB.CreateMemSet(
2629 getNewAllocaSlicePtr(IRB, OldPtr->getType()), II.getValue(), Size,
2630 getSliceAlign(), II.isVolatile());
2631 (void)New;
2632 DEBUG(dbgs() << " to: " << *New << "\n");
2633 return false;
2634 }
2635
2636 // If we can represent this as a simple value, we have to build the actual
2637 // value to store, which requires expanding the byte present in memset to
2638 // a sensible representation for the alloca type. This is essentially
2639 // splatting the byte to a sufficiently wide integer, splatting it across
2640 // any desired vector width, and bitcasting to the final type.
2641 Value *V;
2642
2643 if (VecTy) {
2644 // If this is a memset of a vectorized alloca, insert it.
2645 assert(ElementTy == ScalarTy);
2646
2647 unsigned BeginIndex = getIndex(NewBeginOffset);
2648 unsigned EndIndex = getIndex(NewEndOffset);
2649 assert(EndIndex > BeginIndex && "Empty vector!");
2650 unsigned NumElements = EndIndex - BeginIndex;
2651 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2652
2653 Value *Splat =
2654 getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
2655 Splat = convertValue(DL, IRB, Splat, ElementTy);
2656 if (NumElements > 1)
2657 Splat = getVectorSplat(Splat, NumElements);
2658
2659 Value *Old =
2660 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2661 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2662 } else if (IntTy) {
2663 // If this is a memset on an alloca where we can widen stores, insert the
2664 // set integer.
2665 assert(!II.isVolatile());
2666
2667 uint64_t Size = NewEndOffset - NewBeginOffset;
2668 V = getIntegerSplat(II.getValue(), Size);
2669
2670 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2671 EndOffset != NewAllocaBeginOffset)) {
2672 Value *Old =
2673 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2674 Old = convertValue(DL, IRB, Old, IntTy);
2675 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2676 V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2677 } else {
2678 assert(V->getType() == IntTy &&
2679 "Wrong type for an alloca wide integer!");
2680 }
2681 V = convertValue(DL, IRB, V, AllocaTy);
2682 } else {
2683 // Established these invariants above.
2684 assert(NewBeginOffset == NewAllocaBeginOffset);
2685 assert(NewEndOffset == NewAllocaEndOffset);
2686
2687 V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
2688 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2689 V = getVectorSplat(V, AllocaVecTy->getNumElements());
2690
2691 V = convertValue(DL, IRB, V, AllocaTy);
2692 }
2693
2694 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2695 II.isVolatile());
2696 (void)New;
2697 DEBUG(dbgs() << " to: " << *New << "\n");
2698 return !II.isVolatile();
2699 }
2700
visitMemTransferInst(MemTransferInst & II)2701 bool visitMemTransferInst(MemTransferInst &II) {
2702 // Rewriting of memory transfer instructions can be a bit tricky. We break
2703 // them into two categories: split intrinsics and unsplit intrinsics.
2704
2705 DEBUG(dbgs() << " original: " << II << "\n");
2706
2707 bool IsDest = &II.getRawDestUse() == OldUse;
2708 assert((IsDest && II.getRawDest() == OldPtr) ||
2709 (!IsDest && II.getRawSource() == OldPtr));
2710
2711 unsigned SliceAlign = getSliceAlign();
2712
2713 // For unsplit intrinsics, we simply modify the source and destination
2714 // pointers in place. This isn't just an optimization, it is a matter of
2715 // correctness. With unsplit intrinsics we may be dealing with transfers
2716 // within a single alloca before SROA ran, or with transfers that have
2717 // a variable length. We may also be dealing with memmove instead of
2718 // memcpy, and so simply updating the pointers is the necessary for us to
2719 // update both source and dest of a single call.
2720 if (!IsSplittable) {
2721 Value *AdjustedPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2722 if (IsDest)
2723 II.setDest(AdjustedPtr);
2724 else
2725 II.setSource(AdjustedPtr);
2726
2727 if (II.getAlignment() > SliceAlign) {
2728 Type *CstTy = II.getAlignmentCst()->getType();
2729 II.setAlignment(
2730 ConstantInt::get(CstTy, MinAlign(II.getAlignment(), SliceAlign)));
2731 }
2732
2733 DEBUG(dbgs() << " to: " << II << "\n");
2734 deleteIfTriviallyDead(OldPtr);
2735 return false;
2736 }
2737 // For split transfer intrinsics we have an incredibly useful assurance:
2738 // the source and destination do not reside within the same alloca, and at
2739 // least one of them does not escape. This means that we can replace
2740 // memmove with memcpy, and we don't need to worry about all manner of
2741 // downsides to splitting and transforming the operations.
2742
2743 // If this doesn't map cleanly onto the alloca type, and that type isn't
2744 // a single value type, just emit a memcpy.
2745 bool EmitMemCpy =
2746 !VecTy && !IntTy &&
2747 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2748 SliceSize != DL.getTypeStoreSize(NewAI.getAllocatedType()) ||
2749 !NewAI.getAllocatedType()->isSingleValueType());
2750
2751 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2752 // size hasn't been shrunk based on analysis of the viable range, this is
2753 // a no-op.
2754 if (EmitMemCpy && &OldAI == &NewAI) {
2755 // Ensure the start lines up.
2756 assert(NewBeginOffset == BeginOffset);
2757
2758 // Rewrite the size as needed.
2759 if (NewEndOffset != EndOffset)
2760 II.setLength(ConstantInt::get(II.getLength()->getType(),
2761 NewEndOffset - NewBeginOffset));
2762 return false;
2763 }
2764 // Record this instruction for deletion.
2765 Pass.DeadInsts.insert(&II);
2766
2767 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2768 // alloca that should be re-examined after rewriting this instruction.
2769 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2770 if (AllocaInst *AI =
2771 dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
2772 assert(AI != &OldAI && AI != &NewAI &&
2773 "Splittable transfers cannot reach the same alloca on both ends.");
2774 Pass.Worklist.insert(AI);
2775 }
2776
2777 Type *OtherPtrTy = OtherPtr->getType();
2778 unsigned OtherAS = OtherPtrTy->getPointerAddressSpace();
2779
2780 // Compute the relative offset for the other pointer within the transfer.
2781 unsigned IntPtrWidth = DL.getPointerSizeInBits(OtherAS);
2782 APInt OtherOffset(IntPtrWidth, NewBeginOffset - BeginOffset);
2783 unsigned OtherAlign = MinAlign(II.getAlignment() ? II.getAlignment() : 1,
2784 OtherOffset.zextOrTrunc(64).getZExtValue());
2785
2786 if (EmitMemCpy) {
2787 // Compute the other pointer, folding as much as possible to produce
2788 // a single, simple GEP in most cases.
2789 OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2790 OtherPtr->getName() + ".");
2791
2792 Value *OurPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2793 Type *SizeTy = II.getLength()->getType();
2794 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2795
2796 CallInst *New = IRB.CreateMemCpy(
2797 IsDest ? OurPtr : OtherPtr, IsDest ? OtherPtr : OurPtr, Size,
2798 MinAlign(SliceAlign, OtherAlign), II.isVolatile());
2799 (void)New;
2800 DEBUG(dbgs() << " to: " << *New << "\n");
2801 return false;
2802 }
2803
2804 bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
2805 NewEndOffset == NewAllocaEndOffset;
2806 uint64_t Size = NewEndOffset - NewBeginOffset;
2807 unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
2808 unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
2809 unsigned NumElements = EndIndex - BeginIndex;
2810 IntegerType *SubIntTy =
2811 IntTy ? Type::getIntNTy(IntTy->getContext(), Size * 8) : nullptr;
2812
2813 // Reset the other pointer type to match the register type we're going to
2814 // use, but using the address space of the original other pointer.
2815 if (VecTy && !IsWholeAlloca) {
2816 if (NumElements == 1)
2817 OtherPtrTy = VecTy->getElementType();
2818 else
2819 OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2820
2821 OtherPtrTy = OtherPtrTy->getPointerTo(OtherAS);
2822 } else if (IntTy && !IsWholeAlloca) {
2823 OtherPtrTy = SubIntTy->getPointerTo(OtherAS);
2824 } else {
2825 OtherPtrTy = NewAllocaTy->getPointerTo(OtherAS);
2826 }
2827
2828 Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2829 OtherPtr->getName() + ".");
2830 unsigned SrcAlign = OtherAlign;
2831 Value *DstPtr = &NewAI;
2832 unsigned DstAlign = SliceAlign;
2833 if (!IsDest) {
2834 std::swap(SrcPtr, DstPtr);
2835 std::swap(SrcAlign, DstAlign);
2836 }
2837
2838 Value *Src;
2839 if (VecTy && !IsWholeAlloca && !IsDest) {
2840 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2841 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
2842 } else if (IntTy && !IsWholeAlloca && !IsDest) {
2843 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2844 Src = convertValue(DL, IRB, Src, IntTy);
2845 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2846 Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
2847 } else {
2848 Src =
2849 IRB.CreateAlignedLoad(SrcPtr, SrcAlign, II.isVolatile(), "copyload");
2850 }
2851
2852 if (VecTy && !IsWholeAlloca && IsDest) {
2853 Value *Old =
2854 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2855 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
2856 } else if (IntTy && !IsWholeAlloca && IsDest) {
2857 Value *Old =
2858 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2859 Old = convertValue(DL, IRB, Old, IntTy);
2860 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2861 Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
2862 Src = convertValue(DL, IRB, Src, NewAllocaTy);
2863 }
2864
2865 StoreInst *Store = cast<StoreInst>(
2866 IRB.CreateAlignedStore(Src, DstPtr, DstAlign, II.isVolatile()));
2867 (void)Store;
2868 DEBUG(dbgs() << " to: " << *Store << "\n");
2869 return !II.isVolatile();
2870 }
2871
visitIntrinsicInst(IntrinsicInst & II)2872 bool visitIntrinsicInst(IntrinsicInst &II) {
2873 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2874 II.getIntrinsicID() == Intrinsic::lifetime_end);
2875 DEBUG(dbgs() << " original: " << II << "\n");
2876 assert(II.getArgOperand(1) == OldPtr);
2877
2878 // Record this instruction for deletion.
2879 Pass.DeadInsts.insert(&II);
2880
2881 ConstantInt *Size =
2882 ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2883 NewEndOffset - NewBeginOffset);
2884 Value *Ptr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2885 Value *New;
2886 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2887 New = IRB.CreateLifetimeStart(Ptr, Size);
2888 else
2889 New = IRB.CreateLifetimeEnd(Ptr, Size);
2890
2891 (void)New;
2892 DEBUG(dbgs() << " to: " << *New << "\n");
2893 return true;
2894 }
2895
visitPHINode(PHINode & PN)2896 bool visitPHINode(PHINode &PN) {
2897 DEBUG(dbgs() << " original: " << PN << "\n");
2898 assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
2899 assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
2900
2901 // We would like to compute a new pointer in only one place, but have it be
2902 // as local as possible to the PHI. To do that, we re-use the location of
2903 // the old pointer, which necessarily must be in the right position to
2904 // dominate the PHI.
2905 IRBuilderTy PtrBuilder(IRB);
2906 if (isa<PHINode>(OldPtr))
2907 PtrBuilder.SetInsertPoint(&*OldPtr->getParent()->getFirstInsertionPt());
2908 else
2909 PtrBuilder.SetInsertPoint(OldPtr);
2910 PtrBuilder.SetCurrentDebugLocation(OldPtr->getDebugLoc());
2911
2912 Value *NewPtr = getNewAllocaSlicePtr(PtrBuilder, OldPtr->getType());
2913 // Replace the operands which were using the old pointer.
2914 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
2915
2916 DEBUG(dbgs() << " to: " << PN << "\n");
2917 deleteIfTriviallyDead(OldPtr);
2918
2919 // PHIs can't be promoted on their own, but often can be speculated. We
2920 // check the speculation outside of the rewriter so that we see the
2921 // fully-rewritten alloca.
2922 PHIUsers.insert(&PN);
2923 return true;
2924 }
2925
visitSelectInst(SelectInst & SI)2926 bool visitSelectInst(SelectInst &SI) {
2927 DEBUG(dbgs() << " original: " << SI << "\n");
2928 assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
2929 "Pointer isn't an operand!");
2930 assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
2931 assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
2932
2933 Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2934 // Replace the operands which were using the old pointer.
2935 if (SI.getOperand(1) == OldPtr)
2936 SI.setOperand(1, NewPtr);
2937 if (SI.getOperand(2) == OldPtr)
2938 SI.setOperand(2, NewPtr);
2939
2940 DEBUG(dbgs() << " to: " << SI << "\n");
2941 deleteIfTriviallyDead(OldPtr);
2942
2943 // Selects can't be promoted on their own, but often can be speculated. We
2944 // check the speculation outside of the rewriter so that we see the
2945 // fully-rewritten alloca.
2946 SelectUsers.insert(&SI);
2947 return true;
2948 }
2949 };
2950
2951 namespace {
2952 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
2953 ///
2954 /// This pass aggressively rewrites all aggregate loads and stores on
2955 /// a particular pointer (or any pointer derived from it which we can identify)
2956 /// with scalar loads and stores.
2957 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
2958 // Befriend the base class so it can delegate to private visit methods.
2959 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
2960
2961 /// Queue of pointer uses to analyze and potentially rewrite.
2962 SmallVector<Use *, 8> Queue;
2963
2964 /// Set to prevent us from cycling with phi nodes and loops.
2965 SmallPtrSet<User *, 8> Visited;
2966
2967 /// The current pointer use being rewritten. This is used to dig up the used
2968 /// value (as opposed to the user).
2969 Use *U;
2970
2971 public:
2972 /// Rewrite loads and stores through a pointer and all pointers derived from
2973 /// it.
rewrite(Instruction & I)2974 bool rewrite(Instruction &I) {
2975 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
2976 enqueueUsers(I);
2977 bool Changed = false;
2978 while (!Queue.empty()) {
2979 U = Queue.pop_back_val();
2980 Changed |= visit(cast<Instruction>(U->getUser()));
2981 }
2982 return Changed;
2983 }
2984
2985 private:
2986 /// Enqueue all the users of the given instruction for further processing.
2987 /// This uses a set to de-duplicate users.
enqueueUsers(Instruction & I)2988 void enqueueUsers(Instruction &I) {
2989 for (Use &U : I.uses())
2990 if (Visited.insert(U.getUser()).second)
2991 Queue.push_back(&U);
2992 }
2993
2994 // Conservative default is to not rewrite anything.
visitInstruction(Instruction & I)2995 bool visitInstruction(Instruction &I) { return false; }
2996
2997 /// \brief Generic recursive split emission class.
2998 template <typename Derived> class OpSplitter {
2999 protected:
3000 /// The builder used to form new instructions.
3001 IRBuilderTy IRB;
3002 /// The indices which to be used with insert- or extractvalue to select the
3003 /// appropriate value within the aggregate.
3004 SmallVector<unsigned, 4> Indices;
3005 /// The indices to a GEP instruction which will move Ptr to the correct slot
3006 /// within the aggregate.
3007 SmallVector<Value *, 4> GEPIndices;
3008 /// The base pointer of the original op, used as a base for GEPing the
3009 /// split operations.
3010 Value *Ptr;
3011
3012 /// Initialize the splitter with an insertion point, Ptr and start with a
3013 /// single zero GEP index.
OpSplitter(Instruction * InsertionPoint,Value * Ptr)3014 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
3015 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
3016
3017 public:
3018 /// \brief Generic recursive split emission routine.
3019 ///
3020 /// This method recursively splits an aggregate op (load or store) into
3021 /// scalar or vector ops. It splits recursively until it hits a single value
3022 /// and emits that single value operation via the template argument.
3023 ///
3024 /// The logic of this routine relies on GEPs and insertvalue and
3025 /// extractvalue all operating with the same fundamental index list, merely
3026 /// formatted differently (GEPs need actual values).
3027 ///
3028 /// \param Ty The type being split recursively into smaller ops.
3029 /// \param Agg The aggregate value being built up or stored, depending on
3030 /// whether this is splitting a load or a store respectively.
emitSplitOps(Type * Ty,Value * & Agg,const Twine & Name)3031 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
3032 if (Ty->isSingleValueType())
3033 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
3034
3035 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
3036 unsigned OldSize = Indices.size();
3037 (void)OldSize;
3038 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
3039 ++Idx) {
3040 assert(Indices.size() == OldSize && "Did not return to the old size");
3041 Indices.push_back(Idx);
3042 GEPIndices.push_back(IRB.getInt32(Idx));
3043 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
3044 GEPIndices.pop_back();
3045 Indices.pop_back();
3046 }
3047 return;
3048 }
3049
3050 if (StructType *STy = dyn_cast<StructType>(Ty)) {
3051 unsigned OldSize = Indices.size();
3052 (void)OldSize;
3053 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
3054 ++Idx) {
3055 assert(Indices.size() == OldSize && "Did not return to the old size");
3056 Indices.push_back(Idx);
3057 GEPIndices.push_back(IRB.getInt32(Idx));
3058 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
3059 GEPIndices.pop_back();
3060 Indices.pop_back();
3061 }
3062 return;
3063 }
3064
3065 llvm_unreachable("Only arrays and structs are aggregate loadable types");
3066 }
3067 };
3068
3069 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
LoadOpSplitter__anond316f5690b11::AggLoadStoreRewriter::LoadOpSplitter3070 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3071 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
3072
3073 /// Emit a leaf load of a single value. This is called at the leaves of the
3074 /// recursive emission to actually load values.
emitFunc__anond316f5690b11::AggLoadStoreRewriter::LoadOpSplitter3075 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3076 assert(Ty->isSingleValueType());
3077 // Load the single value and insert it using the indices.
3078 Value *GEP =
3079 IRB.CreateInBoundsGEP(nullptr, Ptr, GEPIndices, Name + ".gep");
3080 Value *Load = IRB.CreateLoad(GEP, Name + ".load");
3081 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
3082 DEBUG(dbgs() << " to: " << *Load << "\n");
3083 }
3084 };
3085
visitLoadInst(LoadInst & LI)3086 bool visitLoadInst(LoadInst &LI) {
3087 assert(LI.getPointerOperand() == *U);
3088 if (!LI.isSimple() || LI.getType()->isSingleValueType())
3089 return false;
3090
3091 // We have an aggregate being loaded, split it apart.
3092 DEBUG(dbgs() << " original: " << LI << "\n");
3093 LoadOpSplitter Splitter(&LI, *U);
3094 Value *V = UndefValue::get(LI.getType());
3095 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
3096 LI.replaceAllUsesWith(V);
3097 LI.eraseFromParent();
3098 return true;
3099 }
3100
3101 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
StoreOpSplitter__anond316f5690b11::AggLoadStoreRewriter::StoreOpSplitter3102 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3103 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
3104
3105 /// Emit a leaf store of a single value. This is called at the leaves of the
3106 /// recursive emission to actually produce stores.
emitFunc__anond316f5690b11::AggLoadStoreRewriter::StoreOpSplitter3107 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3108 assert(Ty->isSingleValueType());
3109 // Extract the single value and store it using the indices.
3110 //
3111 // The gep and extractvalue values are factored out of the CreateStore
3112 // call to make the output independent of the argument evaluation order.
3113 Value *ExtractValue =
3114 IRB.CreateExtractValue(Agg, Indices, Name + ".extract");
3115 Value *InBoundsGEP =
3116 IRB.CreateInBoundsGEP(nullptr, Ptr, GEPIndices, Name + ".gep");
3117 Value *Store = IRB.CreateStore(ExtractValue, InBoundsGEP);
3118 (void)Store;
3119 DEBUG(dbgs() << " to: " << *Store << "\n");
3120 }
3121 };
3122
visitStoreInst(StoreInst & SI)3123 bool visitStoreInst(StoreInst &SI) {
3124 if (!SI.isSimple() || SI.getPointerOperand() != *U)
3125 return false;
3126 Value *V = SI.getValueOperand();
3127 if (V->getType()->isSingleValueType())
3128 return false;
3129
3130 // We have an aggregate being stored, split it apart.
3131 DEBUG(dbgs() << " original: " << SI << "\n");
3132 StoreOpSplitter Splitter(&SI, *U);
3133 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
3134 SI.eraseFromParent();
3135 return true;
3136 }
3137
visitBitCastInst(BitCastInst & BC)3138 bool visitBitCastInst(BitCastInst &BC) {
3139 enqueueUsers(BC);
3140 return false;
3141 }
3142
visitGetElementPtrInst(GetElementPtrInst & GEPI)3143 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
3144 enqueueUsers(GEPI);
3145 return false;
3146 }
3147
visitPHINode(PHINode & PN)3148 bool visitPHINode(PHINode &PN) {
3149 enqueueUsers(PN);
3150 return false;
3151 }
3152
visitSelectInst(SelectInst & SI)3153 bool visitSelectInst(SelectInst &SI) {
3154 enqueueUsers(SI);
3155 return false;
3156 }
3157 };
3158 }
3159
3160 /// \brief Strip aggregate type wrapping.
3161 ///
3162 /// This removes no-op aggregate types wrapping an underlying type. It will
3163 /// strip as many layers of types as it can without changing either the type
3164 /// size or the allocated size.
stripAggregateTypeWrapping(const DataLayout & DL,Type * Ty)3165 static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
3166 if (Ty->isSingleValueType())
3167 return Ty;
3168
3169 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
3170 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
3171
3172 Type *InnerTy;
3173 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
3174 InnerTy = ArrTy->getElementType();
3175 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
3176 const StructLayout *SL = DL.getStructLayout(STy);
3177 unsigned Index = SL->getElementContainingOffset(0);
3178 InnerTy = STy->getElementType(Index);
3179 } else {
3180 return Ty;
3181 }
3182
3183 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
3184 TypeSize > DL.getTypeSizeInBits(InnerTy))
3185 return Ty;
3186
3187 return stripAggregateTypeWrapping(DL, InnerTy);
3188 }
3189
3190 /// \brief Try to find a partition of the aggregate type passed in for a given
3191 /// offset and size.
3192 ///
3193 /// This recurses through the aggregate type and tries to compute a subtype
3194 /// based on the offset and size. When the offset and size span a sub-section
3195 /// of an array, it will even compute a new array type for that sub-section,
3196 /// and the same for structs.
3197 ///
3198 /// Note that this routine is very strict and tries to find a partition of the
3199 /// type which produces the *exact* right offset and size. It is not forgiving
3200 /// when the size or offset cause either end of type-based partition to be off.
3201 /// Also, this is a best-effort routine. It is reasonable to give up and not
3202 /// return a type if necessary.
getTypePartition(const DataLayout & DL,Type * Ty,uint64_t Offset,uint64_t Size)3203 static Type *getTypePartition(const DataLayout &DL, Type *Ty, uint64_t Offset,
3204 uint64_t Size) {
3205 if (Offset == 0 && DL.getTypeAllocSize(Ty) == Size)
3206 return stripAggregateTypeWrapping(DL, Ty);
3207 if (Offset > DL.getTypeAllocSize(Ty) ||
3208 (DL.getTypeAllocSize(Ty) - Offset) < Size)
3209 return nullptr;
3210
3211 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3212 // We can't partition pointers...
3213 if (SeqTy->isPointerTy())
3214 return nullptr;
3215
3216 Type *ElementTy = SeqTy->getElementType();
3217 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3218 uint64_t NumSkippedElements = Offset / ElementSize;
3219 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy)) {
3220 if (NumSkippedElements >= ArrTy->getNumElements())
3221 return nullptr;
3222 } else if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy)) {
3223 if (NumSkippedElements >= VecTy->getNumElements())
3224 return nullptr;
3225 }
3226 Offset -= NumSkippedElements * ElementSize;
3227
3228 // First check if we need to recurse.
3229 if (Offset > 0 || Size < ElementSize) {
3230 // Bail if the partition ends in a different array element.
3231 if ((Offset + Size) > ElementSize)
3232 return nullptr;
3233 // Recurse through the element type trying to peel off offset bytes.
3234 return getTypePartition(DL, ElementTy, Offset, Size);
3235 }
3236 assert(Offset == 0);
3237
3238 if (Size == ElementSize)
3239 return stripAggregateTypeWrapping(DL, ElementTy);
3240 assert(Size > ElementSize);
3241 uint64_t NumElements = Size / ElementSize;
3242 if (NumElements * ElementSize != Size)
3243 return nullptr;
3244 return ArrayType::get(ElementTy, NumElements);
3245 }
3246
3247 StructType *STy = dyn_cast<StructType>(Ty);
3248 if (!STy)
3249 return nullptr;
3250
3251 const StructLayout *SL = DL.getStructLayout(STy);
3252 if (Offset >= SL->getSizeInBytes())
3253 return nullptr;
3254 uint64_t EndOffset = Offset + Size;
3255 if (EndOffset > SL->getSizeInBytes())
3256 return nullptr;
3257
3258 unsigned Index = SL->getElementContainingOffset(Offset);
3259 Offset -= SL->getElementOffset(Index);
3260
3261 Type *ElementTy = STy->getElementType(Index);
3262 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3263 if (Offset >= ElementSize)
3264 return nullptr; // The offset points into alignment padding.
3265
3266 // See if any partition must be contained by the element.
3267 if (Offset > 0 || Size < ElementSize) {
3268 if ((Offset + Size) > ElementSize)
3269 return nullptr;
3270 return getTypePartition(DL, ElementTy, Offset, Size);
3271 }
3272 assert(Offset == 0);
3273
3274 if (Size == ElementSize)
3275 return stripAggregateTypeWrapping(DL, ElementTy);
3276
3277 StructType::element_iterator EI = STy->element_begin() + Index,
3278 EE = STy->element_end();
3279 if (EndOffset < SL->getSizeInBytes()) {
3280 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3281 if (Index == EndIndex)
3282 return nullptr; // Within a single element and its padding.
3283
3284 // Don't try to form "natural" types if the elements don't line up with the
3285 // expected size.
3286 // FIXME: We could potentially recurse down through the last element in the
3287 // sub-struct to find a natural end point.
3288 if (SL->getElementOffset(EndIndex) != EndOffset)
3289 return nullptr;
3290
3291 assert(Index < EndIndex);
3292 EE = STy->element_begin() + EndIndex;
3293 }
3294
3295 // Try to build up a sub-structure.
3296 StructType *SubTy =
3297 StructType::get(STy->getContext(), makeArrayRef(EI, EE), STy->isPacked());
3298 const StructLayout *SubSL = DL.getStructLayout(SubTy);
3299 if (Size != SubSL->getSizeInBytes())
3300 return nullptr; // The sub-struct doesn't have quite the size needed.
3301
3302 return SubTy;
3303 }
3304
3305 /// \brief Pre-split loads and stores to simplify rewriting.
3306 ///
3307 /// We want to break up the splittable load+store pairs as much as
3308 /// possible. This is important to do as a preprocessing step, as once we
3309 /// start rewriting the accesses to partitions of the alloca we lose the
3310 /// necessary information to correctly split apart paired loads and stores
3311 /// which both point into this alloca. The case to consider is something like
3312 /// the following:
3313 ///
3314 /// %a = alloca [12 x i8]
3315 /// %gep1 = getelementptr [12 x i8]* %a, i32 0, i32 0
3316 /// %gep2 = getelementptr [12 x i8]* %a, i32 0, i32 4
3317 /// %gep3 = getelementptr [12 x i8]* %a, i32 0, i32 8
3318 /// %iptr1 = bitcast i8* %gep1 to i64*
3319 /// %iptr2 = bitcast i8* %gep2 to i64*
3320 /// %fptr1 = bitcast i8* %gep1 to float*
3321 /// %fptr2 = bitcast i8* %gep2 to float*
3322 /// %fptr3 = bitcast i8* %gep3 to float*
3323 /// store float 0.0, float* %fptr1
3324 /// store float 1.0, float* %fptr2
3325 /// %v = load i64* %iptr1
3326 /// store i64 %v, i64* %iptr2
3327 /// %f1 = load float* %fptr2
3328 /// %f2 = load float* %fptr3
3329 ///
3330 /// Here we want to form 3 partitions of the alloca, each 4 bytes large, and
3331 /// promote everything so we recover the 2 SSA values that should have been
3332 /// there all along.
3333 ///
3334 /// \returns true if any changes are made.
presplitLoadsAndStores(AllocaInst & AI,AllocaSlices & AS)3335 bool SROA::presplitLoadsAndStores(AllocaInst &AI, AllocaSlices &AS) {
3336 DEBUG(dbgs() << "Pre-splitting loads and stores\n");
3337
3338 // Track the loads and stores which are candidates for pre-splitting here, in
3339 // the order they first appear during the partition scan. These give stable
3340 // iteration order and a basis for tracking which loads and stores we
3341 // actually split.
3342 SmallVector<LoadInst *, 4> Loads;
3343 SmallVector<StoreInst *, 4> Stores;
3344
3345 // We need to accumulate the splits required of each load or store where we
3346 // can find them via a direct lookup. This is important to cross-check loads
3347 // and stores against each other. We also track the slice so that we can kill
3348 // all the slices that end up split.
3349 struct SplitOffsets {
3350 Slice *S;
3351 std::vector<uint64_t> Splits;
3352 };
3353 SmallDenseMap<Instruction *, SplitOffsets, 8> SplitOffsetsMap;
3354
3355 // Track loads out of this alloca which cannot, for any reason, be pre-split.
3356 // This is important as we also cannot pre-split stores of those loads!
3357 // FIXME: This is all pretty gross. It means that we can be more aggressive
3358 // in pre-splitting when the load feeding the store happens to come from
3359 // a separate alloca. Put another way, the effectiveness of SROA would be
3360 // decreased by a frontend which just concatenated all of its local allocas
3361 // into one big flat alloca. But defeating such patterns is exactly the job
3362 // SROA is tasked with! Sadly, to not have this discrepancy we would have
3363 // change store pre-splitting to actually force pre-splitting of the load
3364 // that feeds it *and all stores*. That makes pre-splitting much harder, but
3365 // maybe it would make it more principled?
3366 SmallPtrSet<LoadInst *, 8> UnsplittableLoads;
3367
3368 DEBUG(dbgs() << " Searching for candidate loads and stores\n");
3369 for (auto &P : AS.partitions()) {
3370 for (Slice &S : P) {
3371 Instruction *I = cast<Instruction>(S.getUse()->getUser());
3372 if (!S.isSplittable() || S.endOffset() <= P.endOffset()) {
3373 // If this is a load we have to track that it can't participate in any
3374 // pre-splitting. If this is a store of a load we have to track that
3375 // that load also can't participate in any pre-splitting.
3376 if (auto *LI = dyn_cast<LoadInst>(I))
3377 UnsplittableLoads.insert(LI);
3378 else if (auto *SI = dyn_cast<StoreInst>(I))
3379 if (auto *LI = dyn_cast<LoadInst>(SI->getValueOperand()))
3380 UnsplittableLoads.insert(LI);
3381 continue;
3382 }
3383 assert(P.endOffset() > S.beginOffset() &&
3384 "Empty or backwards partition!");
3385
3386 // Determine if this is a pre-splittable slice.
3387 if (auto *LI = dyn_cast<LoadInst>(I)) {
3388 assert(!LI->isVolatile() && "Cannot split volatile loads!");
3389
3390 // The load must be used exclusively to store into other pointers for
3391 // us to be able to arbitrarily pre-split it. The stores must also be
3392 // simple to avoid changing semantics.
3393 auto IsLoadSimplyStored = [](LoadInst *LI) {
3394 for (User *LU : LI->users()) {
3395 auto *SI = dyn_cast<StoreInst>(LU);
3396 if (!SI || !SI->isSimple())
3397 return false;
3398 }
3399 return true;
3400 };
3401 if (!IsLoadSimplyStored(LI)) {
3402 UnsplittableLoads.insert(LI);
3403 continue;
3404 }
3405
3406 Loads.push_back(LI);
3407 } else if (auto *SI = dyn_cast<StoreInst>(I)) {
3408 if (S.getUse() != &SI->getOperandUse(SI->getPointerOperandIndex()))
3409 // Skip stores *of* pointers. FIXME: This shouldn't even be possible!
3410 continue;
3411 auto *StoredLoad = dyn_cast<LoadInst>(SI->getValueOperand());
3412 if (!StoredLoad || !StoredLoad->isSimple())
3413 continue;
3414 assert(!SI->isVolatile() && "Cannot split volatile stores!");
3415
3416 Stores.push_back(SI);
3417 } else {
3418 // Other uses cannot be pre-split.
3419 continue;
3420 }
3421
3422 // Record the initial split.
3423 DEBUG(dbgs() << " Candidate: " << *I << "\n");
3424 auto &Offsets = SplitOffsetsMap[I];
3425 assert(Offsets.Splits.empty() &&
3426 "Should not have splits the first time we see an instruction!");
3427 Offsets.S = &S;
3428 Offsets.Splits.push_back(P.endOffset() - S.beginOffset());
3429 }
3430
3431 // Now scan the already split slices, and add a split for any of them which
3432 // we're going to pre-split.
3433 for (Slice *S : P.splitSliceTails()) {
3434 auto SplitOffsetsMapI =
3435 SplitOffsetsMap.find(cast<Instruction>(S->getUse()->getUser()));
3436 if (SplitOffsetsMapI == SplitOffsetsMap.end())
3437 continue;
3438 auto &Offsets = SplitOffsetsMapI->second;
3439
3440 assert(Offsets.S == S && "Found a mismatched slice!");
3441 assert(!Offsets.Splits.empty() &&
3442 "Cannot have an empty set of splits on the second partition!");
3443 assert(Offsets.Splits.back() ==
3444 P.beginOffset() - Offsets.S->beginOffset() &&
3445 "Previous split does not end where this one begins!");
3446
3447 // Record each split. The last partition's end isn't needed as the size
3448 // of the slice dictates that.
3449 if (S->endOffset() > P.endOffset())
3450 Offsets.Splits.push_back(P.endOffset() - Offsets.S->beginOffset());
3451 }
3452 }
3453
3454 // We may have split loads where some of their stores are split stores. For
3455 // such loads and stores, we can only pre-split them if their splits exactly
3456 // match relative to their starting offset. We have to verify this prior to
3457 // any rewriting.
3458 Stores.erase(
3459 std::remove_if(Stores.begin(), Stores.end(),
3460 [&UnsplittableLoads, &SplitOffsetsMap](StoreInst *SI) {
3461 // Lookup the load we are storing in our map of split
3462 // offsets.
3463 auto *LI = cast<LoadInst>(SI->getValueOperand());
3464 // If it was completely unsplittable, then we're done,
3465 // and this store can't be pre-split.
3466 if (UnsplittableLoads.count(LI))
3467 return true;
3468
3469 auto LoadOffsetsI = SplitOffsetsMap.find(LI);
3470 if (LoadOffsetsI == SplitOffsetsMap.end())
3471 return false; // Unrelated loads are definitely safe.
3472 auto &LoadOffsets = LoadOffsetsI->second;
3473
3474 // Now lookup the store's offsets.
3475 auto &StoreOffsets = SplitOffsetsMap[SI];
3476
3477 // If the relative offsets of each split in the load and
3478 // store match exactly, then we can split them and we
3479 // don't need to remove them here.
3480 if (LoadOffsets.Splits == StoreOffsets.Splits)
3481 return false;
3482
3483 DEBUG(dbgs()
3484 << " Mismatched splits for load and store:\n"
3485 << " " << *LI << "\n"
3486 << " " << *SI << "\n");
3487
3488 // We've found a store and load that we need to split
3489 // with mismatched relative splits. Just give up on them
3490 // and remove both instructions from our list of
3491 // candidates.
3492 UnsplittableLoads.insert(LI);
3493 return true;
3494 }),
3495 Stores.end());
3496 // Now we have to go *back* through all the stores, because a later store may
3497 // have caused an earlier store's load to become unsplittable and if it is
3498 // unsplittable for the later store, then we can't rely on it being split in
3499 // the earlier store either.
3500 Stores.erase(std::remove_if(Stores.begin(), Stores.end(),
3501 [&UnsplittableLoads](StoreInst *SI) {
3502 auto *LI =
3503 cast<LoadInst>(SI->getValueOperand());
3504 return UnsplittableLoads.count(LI);
3505 }),
3506 Stores.end());
3507 // Once we've established all the loads that can't be split for some reason,
3508 // filter any that made it into our list out.
3509 Loads.erase(std::remove_if(Loads.begin(), Loads.end(),
3510 [&UnsplittableLoads](LoadInst *LI) {
3511 return UnsplittableLoads.count(LI);
3512 }),
3513 Loads.end());
3514
3515
3516 // If no loads or stores are left, there is no pre-splitting to be done for
3517 // this alloca.
3518 if (Loads.empty() && Stores.empty())
3519 return false;
3520
3521 // From here on, we can't fail and will be building new accesses, so rig up
3522 // an IR builder.
3523 IRBuilderTy IRB(&AI);
3524
3525 // Collect the new slices which we will merge into the alloca slices.
3526 SmallVector<Slice, 4> NewSlices;
3527
3528 // Track any allocas we end up splitting loads and stores for so we iterate
3529 // on them.
3530 SmallPtrSet<AllocaInst *, 4> ResplitPromotableAllocas;
3531
3532 // At this point, we have collected all of the loads and stores we can
3533 // pre-split, and the specific splits needed for them. We actually do the
3534 // splitting in a specific order in order to handle when one of the loads in
3535 // the value operand to one of the stores.
3536 //
3537 // First, we rewrite all of the split loads, and just accumulate each split
3538 // load in a parallel structure. We also build the slices for them and append
3539 // them to the alloca slices.
3540 SmallDenseMap<LoadInst *, std::vector<LoadInst *>, 1> SplitLoadsMap;
3541 std::vector<LoadInst *> SplitLoads;
3542 const DataLayout &DL = AI.getModule()->getDataLayout();
3543 for (LoadInst *LI : Loads) {
3544 SplitLoads.clear();
3545
3546 IntegerType *Ty = cast<IntegerType>(LI->getType());
3547 uint64_t LoadSize = Ty->getBitWidth() / 8;
3548 assert(LoadSize > 0 && "Cannot have a zero-sized integer load!");
3549
3550 auto &Offsets = SplitOffsetsMap[LI];
3551 assert(LoadSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&
3552 "Slice size should always match load size exactly!");
3553 uint64_t BaseOffset = Offsets.S->beginOffset();
3554 assert(BaseOffset + LoadSize > BaseOffset &&
3555 "Cannot represent alloca access size using 64-bit integers!");
3556
3557 Instruction *BasePtr = cast<Instruction>(LI->getPointerOperand());
3558 IRB.SetInsertPoint(LI);
3559
3560 DEBUG(dbgs() << " Splitting load: " << *LI << "\n");
3561
3562 uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
3563 int Idx = 0, Size = Offsets.Splits.size();
3564 for (;;) {
3565 auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
3566 auto *PartPtrTy = PartTy->getPointerTo(LI->getPointerAddressSpace());
3567 LoadInst *PLoad = IRB.CreateAlignedLoad(
3568 getAdjustedPtr(IRB, DL, BasePtr,
3569 APInt(DL.getPointerSizeInBits(), PartOffset),
3570 PartPtrTy, BasePtr->getName() + "."),
3571 getAdjustedAlignment(LI, PartOffset, DL), /*IsVolatile*/ false,
3572 LI->getName());
3573
3574 // Append this load onto the list of split loads so we can find it later
3575 // to rewrite the stores.
3576 SplitLoads.push_back(PLoad);
3577
3578 // Now build a new slice for the alloca.
3579 NewSlices.push_back(
3580 Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
3581 &PLoad->getOperandUse(PLoad->getPointerOperandIndex()),
3582 /*IsSplittable*/ false));
3583 DEBUG(dbgs() << " new slice [" << NewSlices.back().beginOffset()
3584 << ", " << NewSlices.back().endOffset() << "): " << *PLoad
3585 << "\n");
3586
3587 // See if we've handled all the splits.
3588 if (Idx >= Size)
3589 break;
3590
3591 // Setup the next partition.
3592 PartOffset = Offsets.Splits[Idx];
3593 ++Idx;
3594 PartSize = (Idx < Size ? Offsets.Splits[Idx] : LoadSize) - PartOffset;
3595 }
3596
3597 // Now that we have the split loads, do the slow walk over all uses of the
3598 // load and rewrite them as split stores, or save the split loads to use
3599 // below if the store is going to be split there anyways.
3600 bool DeferredStores = false;
3601 for (User *LU : LI->users()) {
3602 StoreInst *SI = cast<StoreInst>(LU);
3603 if (!Stores.empty() && SplitOffsetsMap.count(SI)) {
3604 DeferredStores = true;
3605 DEBUG(dbgs() << " Deferred splitting of store: " << *SI << "\n");
3606 continue;
3607 }
3608
3609 Value *StoreBasePtr = SI->getPointerOperand();
3610 IRB.SetInsertPoint(SI);
3611
3612 DEBUG(dbgs() << " Splitting store of load: " << *SI << "\n");
3613
3614 for (int Idx = 0, Size = SplitLoads.size(); Idx < Size; ++Idx) {
3615 LoadInst *PLoad = SplitLoads[Idx];
3616 uint64_t PartOffset = Idx == 0 ? 0 : Offsets.Splits[Idx - 1];
3617 auto *PartPtrTy =
3618 PLoad->getType()->getPointerTo(SI->getPointerAddressSpace());
3619
3620 StoreInst *PStore = IRB.CreateAlignedStore(
3621 PLoad, getAdjustedPtr(IRB, DL, StoreBasePtr,
3622 APInt(DL.getPointerSizeInBits(), PartOffset),
3623 PartPtrTy, StoreBasePtr->getName() + "."),
3624 getAdjustedAlignment(SI, PartOffset, DL), /*IsVolatile*/ false);
3625 (void)PStore;
3626 DEBUG(dbgs() << " +" << PartOffset << ":" << *PStore << "\n");
3627 }
3628
3629 // We want to immediately iterate on any allocas impacted by splitting
3630 // this store, and we have to track any promotable alloca (indicated by
3631 // a direct store) as needing to be resplit because it is no longer
3632 // promotable.
3633 if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(StoreBasePtr)) {
3634 ResplitPromotableAllocas.insert(OtherAI);
3635 Worklist.insert(OtherAI);
3636 } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
3637 StoreBasePtr->stripInBoundsOffsets())) {
3638 Worklist.insert(OtherAI);
3639 }
3640
3641 // Mark the original store as dead.
3642 DeadInsts.insert(SI);
3643 }
3644
3645 // Save the split loads if there are deferred stores among the users.
3646 if (DeferredStores)
3647 SplitLoadsMap.insert(std::make_pair(LI, std::move(SplitLoads)));
3648
3649 // Mark the original load as dead and kill the original slice.
3650 DeadInsts.insert(LI);
3651 Offsets.S->kill();
3652 }
3653
3654 // Second, we rewrite all of the split stores. At this point, we know that
3655 // all loads from this alloca have been split already. For stores of such
3656 // loads, we can simply look up the pre-existing split loads. For stores of
3657 // other loads, we split those loads first and then write split stores of
3658 // them.
3659 for (StoreInst *SI : Stores) {
3660 auto *LI = cast<LoadInst>(SI->getValueOperand());
3661 IntegerType *Ty = cast<IntegerType>(LI->getType());
3662 uint64_t StoreSize = Ty->getBitWidth() / 8;
3663 assert(StoreSize > 0 && "Cannot have a zero-sized integer store!");
3664
3665 auto &Offsets = SplitOffsetsMap[SI];
3666 assert(StoreSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&
3667 "Slice size should always match load size exactly!");
3668 uint64_t BaseOffset = Offsets.S->beginOffset();
3669 assert(BaseOffset + StoreSize > BaseOffset &&
3670 "Cannot represent alloca access size using 64-bit integers!");
3671
3672 Value *LoadBasePtr = LI->getPointerOperand();
3673 Instruction *StoreBasePtr = cast<Instruction>(SI->getPointerOperand());
3674
3675 DEBUG(dbgs() << " Splitting store: " << *SI << "\n");
3676
3677 // Check whether we have an already split load.
3678 auto SplitLoadsMapI = SplitLoadsMap.find(LI);
3679 std::vector<LoadInst *> *SplitLoads = nullptr;
3680 if (SplitLoadsMapI != SplitLoadsMap.end()) {
3681 SplitLoads = &SplitLoadsMapI->second;
3682 assert(SplitLoads->size() == Offsets.Splits.size() + 1 &&
3683 "Too few split loads for the number of splits in the store!");
3684 } else {
3685 DEBUG(dbgs() << " of load: " << *LI << "\n");
3686 }
3687
3688 uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
3689 int Idx = 0, Size = Offsets.Splits.size();
3690 for (;;) {
3691 auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
3692 auto *PartPtrTy = PartTy->getPointerTo(SI->getPointerAddressSpace());
3693
3694 // Either lookup a split load or create one.
3695 LoadInst *PLoad;
3696 if (SplitLoads) {
3697 PLoad = (*SplitLoads)[Idx];
3698 } else {
3699 IRB.SetInsertPoint(LI);
3700 PLoad = IRB.CreateAlignedLoad(
3701 getAdjustedPtr(IRB, DL, LoadBasePtr,
3702 APInt(DL.getPointerSizeInBits(), PartOffset),
3703 PartPtrTy, LoadBasePtr->getName() + "."),
3704 getAdjustedAlignment(LI, PartOffset, DL), /*IsVolatile*/ false,
3705 LI->getName());
3706 }
3707
3708 // And store this partition.
3709 IRB.SetInsertPoint(SI);
3710 StoreInst *PStore = IRB.CreateAlignedStore(
3711 PLoad, getAdjustedPtr(IRB, DL, StoreBasePtr,
3712 APInt(DL.getPointerSizeInBits(), PartOffset),
3713 PartPtrTy, StoreBasePtr->getName() + "."),
3714 getAdjustedAlignment(SI, PartOffset, DL), /*IsVolatile*/ false);
3715
3716 // Now build a new slice for the alloca.
3717 NewSlices.push_back(
3718 Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
3719 &PStore->getOperandUse(PStore->getPointerOperandIndex()),
3720 /*IsSplittable*/ false));
3721 DEBUG(dbgs() << " new slice [" << NewSlices.back().beginOffset()
3722 << ", " << NewSlices.back().endOffset() << "): " << *PStore
3723 << "\n");
3724 if (!SplitLoads) {
3725 DEBUG(dbgs() << " of split load: " << *PLoad << "\n");
3726 }
3727
3728 // See if we've finished all the splits.
3729 if (Idx >= Size)
3730 break;
3731
3732 // Setup the next partition.
3733 PartOffset = Offsets.Splits[Idx];
3734 ++Idx;
3735 PartSize = (Idx < Size ? Offsets.Splits[Idx] : StoreSize) - PartOffset;
3736 }
3737
3738 // We want to immediately iterate on any allocas impacted by splitting
3739 // this load, which is only relevant if it isn't a load of this alloca and
3740 // thus we didn't already split the loads above. We also have to keep track
3741 // of any promotable allocas we split loads on as they can no longer be
3742 // promoted.
3743 if (!SplitLoads) {
3744 if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(LoadBasePtr)) {
3745 assert(OtherAI != &AI && "We can't re-split our own alloca!");
3746 ResplitPromotableAllocas.insert(OtherAI);
3747 Worklist.insert(OtherAI);
3748 } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
3749 LoadBasePtr->stripInBoundsOffsets())) {
3750 assert(OtherAI != &AI && "We can't re-split our own alloca!");
3751 Worklist.insert(OtherAI);
3752 }
3753 }
3754
3755 // Mark the original store as dead now that we've split it up and kill its
3756 // slice. Note that we leave the original load in place unless this store
3757 // was its only use. It may in turn be split up if it is an alloca load
3758 // for some other alloca, but it may be a normal load. This may introduce
3759 // redundant loads, but where those can be merged the rest of the optimizer
3760 // should handle the merging, and this uncovers SSA splits which is more
3761 // important. In practice, the original loads will almost always be fully
3762 // split and removed eventually, and the splits will be merged by any
3763 // trivial CSE, including instcombine.
3764 if (LI->hasOneUse()) {
3765 assert(*LI->user_begin() == SI && "Single use isn't this store!");
3766 DeadInsts.insert(LI);
3767 }
3768 DeadInsts.insert(SI);
3769 Offsets.S->kill();
3770 }
3771
3772 // Remove the killed slices that have ben pre-split.
3773 AS.erase(std::remove_if(AS.begin(), AS.end(), [](const Slice &S) {
3774 return S.isDead();
3775 }), AS.end());
3776
3777 // Insert our new slices. This will sort and merge them into the sorted
3778 // sequence.
3779 AS.insert(NewSlices);
3780
3781 DEBUG(dbgs() << " Pre-split slices:\n");
3782 #ifndef NDEBUG
3783 for (auto I = AS.begin(), E = AS.end(); I != E; ++I)
3784 DEBUG(AS.print(dbgs(), I, " "));
3785 #endif
3786
3787 // Finally, don't try to promote any allocas that new require re-splitting.
3788 // They have already been added to the worklist above.
3789 PromotableAllocas.erase(
3790 std::remove_if(
3791 PromotableAllocas.begin(), PromotableAllocas.end(),
3792 [&](AllocaInst *AI) { return ResplitPromotableAllocas.count(AI); }),
3793 PromotableAllocas.end());
3794
3795 return true;
3796 }
3797
3798 /// \brief Rewrite an alloca partition's users.
3799 ///
3800 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3801 /// to rewrite uses of an alloca partition to be conducive for SSA value
3802 /// promotion. If the partition needs a new, more refined alloca, this will
3803 /// build that new alloca, preserving as much type information as possible, and
3804 /// rewrite the uses of the old alloca to point at the new one and have the
3805 /// appropriate new offsets. It also evaluates how successful the rewrite was
3806 /// at enabling promotion and if it was successful queues the alloca to be
3807 /// promoted.
rewritePartition(AllocaInst & AI,AllocaSlices & AS,Partition & P)3808 AllocaInst *SROA::rewritePartition(AllocaInst &AI, AllocaSlices &AS,
3809 Partition &P) {
3810 // Try to compute a friendly type for this partition of the alloca. This
3811 // won't always succeed, in which case we fall back to a legal integer type
3812 // or an i8 array of an appropriate size.
3813 Type *SliceTy = nullptr;
3814 const DataLayout &DL = AI.getModule()->getDataLayout();
3815 if (Type *CommonUseTy = findCommonType(P.begin(), P.end(), P.endOffset()))
3816 if (DL.getTypeAllocSize(CommonUseTy) >= P.size())
3817 SliceTy = CommonUseTy;
3818 if (!SliceTy)
3819 if (Type *TypePartitionTy = getTypePartition(DL, AI.getAllocatedType(),
3820 P.beginOffset(), P.size()))
3821 SliceTy = TypePartitionTy;
3822 if ((!SliceTy || (SliceTy->isArrayTy() &&
3823 SliceTy->getArrayElementType()->isIntegerTy())) &&
3824 DL.isLegalInteger(P.size() * 8))
3825 SliceTy = Type::getIntNTy(*C, P.size() * 8);
3826 if (!SliceTy)
3827 SliceTy = ArrayType::get(Type::getInt8Ty(*C), P.size());
3828 assert(DL.getTypeAllocSize(SliceTy) >= P.size());
3829
3830 bool IsIntegerPromotable = isIntegerWideningViable(P, SliceTy, DL);
3831
3832 VectorType *VecTy =
3833 IsIntegerPromotable ? nullptr : isVectorPromotionViable(P, DL);
3834 if (VecTy)
3835 SliceTy = VecTy;
3836
3837 // Check for the case where we're going to rewrite to a new alloca of the
3838 // exact same type as the original, and with the same access offsets. In that
3839 // case, re-use the existing alloca, but still run through the rewriter to
3840 // perform phi and select speculation.
3841 AllocaInst *NewAI;
3842 if (SliceTy == AI.getAllocatedType()) {
3843 assert(P.beginOffset() == 0 &&
3844 "Non-zero begin offset but same alloca type");
3845 NewAI = &AI;
3846 // FIXME: We should be able to bail at this point with "nothing changed".
3847 // FIXME: We might want to defer PHI speculation until after here.
3848 // FIXME: return nullptr;
3849 } else {
3850 unsigned Alignment = AI.getAlignment();
3851 if (!Alignment) {
3852 // The minimum alignment which users can rely on when the explicit
3853 // alignment is omitted or zero is that required by the ABI for this
3854 // type.
3855 Alignment = DL.getABITypeAlignment(AI.getAllocatedType());
3856 }
3857 Alignment = MinAlign(Alignment, P.beginOffset());
3858 // If we will get at least this much alignment from the type alone, leave
3859 // the alloca's alignment unconstrained.
3860 if (Alignment <= DL.getABITypeAlignment(SliceTy))
3861 Alignment = 0;
3862 NewAI = new AllocaInst(
3863 SliceTy, nullptr, Alignment,
3864 AI.getName() + ".sroa." + Twine(P.begin() - AS.begin()), &AI);
3865 ++NumNewAllocas;
3866 }
3867
3868 DEBUG(dbgs() << "Rewriting alloca partition "
3869 << "[" << P.beginOffset() << "," << P.endOffset()
3870 << ") to: " << *NewAI << "\n");
3871
3872 // Track the high watermark on the worklist as it is only relevant for
3873 // promoted allocas. We will reset it to this point if the alloca is not in
3874 // fact scheduled for promotion.
3875 unsigned PPWOldSize = PostPromotionWorklist.size();
3876 unsigned NumUses = 0;
3877 SmallPtrSet<PHINode *, 8> PHIUsers;
3878 SmallPtrSet<SelectInst *, 8> SelectUsers;
3879
3880 AllocaSliceRewriter Rewriter(DL, AS, *this, AI, *NewAI, P.beginOffset(),
3881 P.endOffset(), IsIntegerPromotable, VecTy,
3882 PHIUsers, SelectUsers);
3883 bool Promotable = true;
3884 for (Slice *S : P.splitSliceTails()) {
3885 Promotable &= Rewriter.visit(S);
3886 ++NumUses;
3887 }
3888 for (Slice &S : P) {
3889 Promotable &= Rewriter.visit(&S);
3890 ++NumUses;
3891 }
3892
3893 NumAllocaPartitionUses += NumUses;
3894 MaxUsesPerAllocaPartition =
3895 std::max<unsigned>(NumUses, MaxUsesPerAllocaPartition);
3896
3897 // Now that we've processed all the slices in the new partition, check if any
3898 // PHIs or Selects would block promotion.
3899 for (SmallPtrSetImpl<PHINode *>::iterator I = PHIUsers.begin(),
3900 E = PHIUsers.end();
3901 I != E; ++I)
3902 if (!isSafePHIToSpeculate(**I)) {
3903 Promotable = false;
3904 PHIUsers.clear();
3905 SelectUsers.clear();
3906 break;
3907 }
3908 for (SmallPtrSetImpl<SelectInst *>::iterator I = SelectUsers.begin(),
3909 E = SelectUsers.end();
3910 I != E; ++I)
3911 if (!isSafeSelectToSpeculate(**I)) {
3912 Promotable = false;
3913 PHIUsers.clear();
3914 SelectUsers.clear();
3915 break;
3916 }
3917
3918 if (Promotable) {
3919 if (PHIUsers.empty() && SelectUsers.empty()) {
3920 // Promote the alloca.
3921 PromotableAllocas.push_back(NewAI);
3922 } else {
3923 // If we have either PHIs or Selects to speculate, add them to those
3924 // worklists and re-queue the new alloca so that we promote in on the
3925 // next iteration.
3926 for (PHINode *PHIUser : PHIUsers)
3927 SpeculatablePHIs.insert(PHIUser);
3928 for (SelectInst *SelectUser : SelectUsers)
3929 SpeculatableSelects.insert(SelectUser);
3930 Worklist.insert(NewAI);
3931 }
3932 } else {
3933 // Drop any post-promotion work items if promotion didn't happen.
3934 while (PostPromotionWorklist.size() > PPWOldSize)
3935 PostPromotionWorklist.pop_back();
3936
3937 // We couldn't promote and we didn't create a new partition, nothing
3938 // happened.
3939 if (NewAI == &AI)
3940 return nullptr;
3941
3942 // If we can't promote the alloca, iterate on it to check for new
3943 // refinements exposed by splitting the current alloca. Don't iterate on an
3944 // alloca which didn't actually change and didn't get promoted.
3945 Worklist.insert(NewAI);
3946 }
3947
3948 return NewAI;
3949 }
3950
3951 /// \brief Walks the slices of an alloca and form partitions based on them,
3952 /// rewriting each of their uses.
splitAlloca(AllocaInst & AI,AllocaSlices & AS)3953 bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &AS) {
3954 if (AS.begin() == AS.end())
3955 return false;
3956
3957 unsigned NumPartitions = 0;
3958 bool Changed = false;
3959 const DataLayout &DL = AI.getModule()->getDataLayout();
3960
3961 // First try to pre-split loads and stores.
3962 Changed |= presplitLoadsAndStores(AI, AS);
3963
3964 // Now that we have identified any pre-splitting opportunities, mark any
3965 // splittable (non-whole-alloca) loads and stores as unsplittable. If we fail
3966 // to split these during pre-splitting, we want to force them to be
3967 // rewritten into a partition.
3968 bool IsSorted = true;
3969 for (Slice &S : AS) {
3970 if (!S.isSplittable())
3971 continue;
3972 // FIXME: We currently leave whole-alloca splittable loads and stores. This
3973 // used to be the only splittable loads and stores and we need to be
3974 // confident that the above handling of splittable loads and stores is
3975 // completely sufficient before we forcibly disable the remaining handling.
3976 if (S.beginOffset() == 0 &&
3977 S.endOffset() >= DL.getTypeAllocSize(AI.getAllocatedType()))
3978 continue;
3979 if (isa<LoadInst>(S.getUse()->getUser()) ||
3980 isa<StoreInst>(S.getUse()->getUser())) {
3981 S.makeUnsplittable();
3982 IsSorted = false;
3983 }
3984 }
3985 if (!IsSorted)
3986 std::sort(AS.begin(), AS.end());
3987
3988 /// \brief Describes the allocas introduced by rewritePartition
3989 /// in order to migrate the debug info.
3990 struct Piece {
3991 AllocaInst *Alloca;
3992 uint64_t Offset;
3993 uint64_t Size;
3994 Piece(AllocaInst *AI, uint64_t O, uint64_t S)
3995 : Alloca(AI), Offset(O), Size(S) {}
3996 };
3997 SmallVector<Piece, 4> Pieces;
3998
3999 // Rewrite each partition.
4000 for (auto &P : AS.partitions()) {
4001 if (AllocaInst *NewAI = rewritePartition(AI, AS, P)) {
4002 Changed = true;
4003 if (NewAI != &AI) {
4004 uint64_t SizeOfByte = 8;
4005 uint64_t AllocaSize = DL.getTypeSizeInBits(NewAI->getAllocatedType());
4006 // Don't include any padding.
4007 uint64_t Size = std::min(AllocaSize, P.size() * SizeOfByte);
4008 Pieces.push_back(Piece(NewAI, P.beginOffset() * SizeOfByte, Size));
4009 }
4010 }
4011 ++NumPartitions;
4012 }
4013
4014 NumAllocaPartitions += NumPartitions;
4015 MaxPartitionsPerAlloca =
4016 std::max<unsigned>(NumPartitions, MaxPartitionsPerAlloca);
4017
4018 // Migrate debug information from the old alloca to the new alloca(s)
4019 // and the individual partitions.
4020 if (DbgDeclareInst *DbgDecl = FindAllocaDbgDeclare(&AI)) {
4021 auto *Var = DbgDecl->getVariable();
4022 auto *Expr = DbgDecl->getExpression();
4023 DIBuilder DIB(*AI.getModule(), /*AllowUnresolved*/ false);
4024 uint64_t AllocaSize = DL.getTypeSizeInBits(AI.getAllocatedType());
4025 for (auto Piece : Pieces) {
4026 // Create a piece expression describing the new partition or reuse AI's
4027 // expression if there is only one partition.
4028 auto *PieceExpr = Expr;
4029 if (Piece.Size < AllocaSize || Expr->isBitPiece()) {
4030 // If this alloca is already a scalar replacement of a larger aggregate,
4031 // Piece.Offset describes the offset inside the scalar.
4032 uint64_t Offset = Expr->isBitPiece() ? Expr->getBitPieceOffset() : 0;
4033 uint64_t Start = Offset + Piece.Offset;
4034 uint64_t Size = Piece.Size;
4035 if (Expr->isBitPiece()) {
4036 uint64_t AbsEnd = Expr->getBitPieceOffset() + Expr->getBitPieceSize();
4037 if (Start >= AbsEnd)
4038 // No need to describe a SROAed padding.
4039 continue;
4040 Size = std::min(Size, AbsEnd - Start);
4041 }
4042 PieceExpr = DIB.createBitPieceExpression(Start, Size);
4043 } else {
4044 assert(Pieces.size() == 1 &&
4045 "partition is as large as original alloca");
4046 }
4047
4048 // Remove any existing dbg.declare intrinsic describing the same alloca.
4049 if (DbgDeclareInst *OldDDI = FindAllocaDbgDeclare(Piece.Alloca))
4050 OldDDI->eraseFromParent();
4051
4052 DIB.insertDeclare(Piece.Alloca, Var, PieceExpr, DbgDecl->getDebugLoc(),
4053 &AI);
4054 }
4055 }
4056 return Changed;
4057 }
4058
4059 /// \brief Clobber a use with undef, deleting the used value if it becomes dead.
clobberUse(Use & U)4060 void SROA::clobberUse(Use &U) {
4061 Value *OldV = U;
4062 // Replace the use with an undef value.
4063 U = UndefValue::get(OldV->getType());
4064
4065 // Check for this making an instruction dead. We have to garbage collect
4066 // all the dead instructions to ensure the uses of any alloca end up being
4067 // minimal.
4068 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
4069 if (isInstructionTriviallyDead(OldI)) {
4070 DeadInsts.insert(OldI);
4071 }
4072 }
4073
4074 /// \brief Analyze an alloca for SROA.
4075 ///
4076 /// This analyzes the alloca to ensure we can reason about it, builds
4077 /// the slices of the alloca, and then hands it off to be split and
4078 /// rewritten as needed.
runOnAlloca(AllocaInst & AI)4079 bool SROA::runOnAlloca(AllocaInst &AI) {
4080 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
4081 ++NumAllocasAnalyzed;
4082
4083 // Special case dead allocas, as they're trivial.
4084 if (AI.use_empty()) {
4085 AI.eraseFromParent();
4086 return true;
4087 }
4088 const DataLayout &DL = AI.getModule()->getDataLayout();
4089
4090 // Skip alloca forms that this analysis can't handle.
4091 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
4092 DL.getTypeAllocSize(AI.getAllocatedType()) == 0)
4093 return false;
4094
4095 bool Changed = false;
4096
4097 // First, split any FCA loads and stores touching this alloca to promote
4098 // better splitting and promotion opportunities.
4099 AggLoadStoreRewriter AggRewriter;
4100 Changed |= AggRewriter.rewrite(AI);
4101
4102 // Build the slices using a recursive instruction-visiting builder.
4103 AllocaSlices AS(DL, AI);
4104 DEBUG(AS.print(dbgs()));
4105 if (AS.isEscaped())
4106 return Changed;
4107
4108 // Delete all the dead users of this alloca before splitting and rewriting it.
4109 for (Instruction *DeadUser : AS.getDeadUsers()) {
4110 // Free up everything used by this instruction.
4111 for (Use &DeadOp : DeadUser->operands())
4112 clobberUse(DeadOp);
4113
4114 // Now replace the uses of this instruction.
4115 DeadUser->replaceAllUsesWith(UndefValue::get(DeadUser->getType()));
4116
4117 // And mark it for deletion.
4118 DeadInsts.insert(DeadUser);
4119 Changed = true;
4120 }
4121 for (Use *DeadOp : AS.getDeadOperands()) {
4122 clobberUse(*DeadOp);
4123 Changed = true;
4124 }
4125
4126 // No slices to split. Leave the dead alloca for a later pass to clean up.
4127 if (AS.begin() == AS.end())
4128 return Changed;
4129
4130 Changed |= splitAlloca(AI, AS);
4131
4132 DEBUG(dbgs() << " Speculating PHIs\n");
4133 while (!SpeculatablePHIs.empty())
4134 speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
4135
4136 DEBUG(dbgs() << " Speculating Selects\n");
4137 while (!SpeculatableSelects.empty())
4138 speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
4139
4140 return Changed;
4141 }
4142
4143 /// \brief Delete the dead instructions accumulated in this run.
4144 ///
4145 /// Recursively deletes the dead instructions we've accumulated. This is done
4146 /// at the very end to maximize locality of the recursive delete and to
4147 /// minimize the problems of invalidated instruction pointers as such pointers
4148 /// are used heavily in the intermediate stages of the algorithm.
4149 ///
4150 /// We also record the alloca instructions deleted here so that they aren't
4151 /// subsequently handed to mem2reg to promote.
deleteDeadInstructions(SmallPtrSetImpl<AllocaInst * > & DeletedAllocas)4152 void SROA::deleteDeadInstructions(
4153 SmallPtrSetImpl<AllocaInst *> &DeletedAllocas) {
4154 while (!DeadInsts.empty()) {
4155 Instruction *I = DeadInsts.pop_back_val();
4156 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
4157
4158 I->replaceAllUsesWith(UndefValue::get(I->getType()));
4159
4160 for (Use &Operand : I->operands())
4161 if (Instruction *U = dyn_cast<Instruction>(Operand)) {
4162 // Zero out the operand and see if it becomes trivially dead.
4163 Operand = nullptr;
4164 if (isInstructionTriviallyDead(U))
4165 DeadInsts.insert(U);
4166 }
4167
4168 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
4169 DeletedAllocas.insert(AI);
4170 if (DbgDeclareInst *DbgDecl = FindAllocaDbgDeclare(AI))
4171 DbgDecl->eraseFromParent();
4172 }
4173
4174 ++NumDeleted;
4175 I->eraseFromParent();
4176 }
4177 }
4178
4179 /// \brief Promote the allocas, using the best available technique.
4180 ///
4181 /// This attempts to promote whatever allocas have been identified as viable in
4182 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
4183 /// This function returns whether any promotion occurred.
promoteAllocas(Function & F)4184 bool SROA::promoteAllocas(Function &F) {
4185 if (PromotableAllocas.empty())
4186 return false;
4187
4188 NumPromoted += PromotableAllocas.size();
4189
4190 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
4191 PromoteMemToReg(PromotableAllocas, *DT, nullptr, AC);
4192 PromotableAllocas.clear();
4193 return true;
4194 }
4195
runImpl(Function & F,DominatorTree & RunDT,AssumptionCache & RunAC)4196 PreservedAnalyses SROA::runImpl(Function &F, DominatorTree &RunDT,
4197 AssumptionCache &RunAC) {
4198 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
4199 C = &F.getContext();
4200 DT = &RunDT;
4201 AC = &RunAC;
4202
4203 BasicBlock &EntryBB = F.getEntryBlock();
4204 for (BasicBlock::iterator I = EntryBB.begin(), E = std::prev(EntryBB.end());
4205 I != E; ++I) {
4206 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
4207 Worklist.insert(AI);
4208 }
4209
4210 bool Changed = false;
4211 // A set of deleted alloca instruction pointers which should be removed from
4212 // the list of promotable allocas.
4213 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
4214
4215 do {
4216 while (!Worklist.empty()) {
4217 Changed |= runOnAlloca(*Worklist.pop_back_val());
4218 deleteDeadInstructions(DeletedAllocas);
4219
4220 // Remove the deleted allocas from various lists so that we don't try to
4221 // continue processing them.
4222 if (!DeletedAllocas.empty()) {
4223 auto IsInSet = [&](AllocaInst *AI) { return DeletedAllocas.count(AI); };
4224 Worklist.remove_if(IsInSet);
4225 PostPromotionWorklist.remove_if(IsInSet);
4226 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
4227 PromotableAllocas.end(),
4228 IsInSet),
4229 PromotableAllocas.end());
4230 DeletedAllocas.clear();
4231 }
4232 }
4233
4234 Changed |= promoteAllocas(F);
4235
4236 Worklist = PostPromotionWorklist;
4237 PostPromotionWorklist.clear();
4238 } while (!Worklist.empty());
4239
4240 if (!Changed)
4241 return PreservedAnalyses::all();
4242
4243 // FIXME: Even when promoting allocas we should preserve some abstract set of
4244 // CFG-specific analyses.
4245 PreservedAnalyses PA;
4246 PA.preserve<GlobalsAA>();
4247 return PA;
4248 }
4249
run(Function & F,AnalysisManager<Function> & AM)4250 PreservedAnalyses SROA::run(Function &F, AnalysisManager<Function> &AM) {
4251 return runImpl(F, AM.getResult<DominatorTreeAnalysis>(F),
4252 AM.getResult<AssumptionAnalysis>(F));
4253 }
4254
4255 /// A legacy pass for the legacy pass manager that wraps the \c SROA pass.
4256 ///
4257 /// This is in the llvm namespace purely to allow it to be a friend of the \c
4258 /// SROA pass.
4259 class llvm::sroa::SROALegacyPass : public FunctionPass {
4260 /// The SROA implementation.
4261 SROA Impl;
4262
4263 public:
SROALegacyPass()4264 SROALegacyPass() : FunctionPass(ID) {
4265 initializeSROALegacyPassPass(*PassRegistry::getPassRegistry());
4266 }
runOnFunction(Function & F)4267 bool runOnFunction(Function &F) override {
4268 if (skipFunction(F))
4269 return false;
4270
4271 auto PA = Impl.runImpl(
4272 F, getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
4273 getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F));
4274 return !PA.areAllPreserved();
4275 }
getAnalysisUsage(AnalysisUsage & AU) const4276 void getAnalysisUsage(AnalysisUsage &AU) const override {
4277 AU.addRequired<AssumptionCacheTracker>();
4278 AU.addRequired<DominatorTreeWrapperPass>();
4279 AU.addPreserved<GlobalsAAWrapperPass>();
4280 AU.setPreservesCFG();
4281 }
4282
getPassName() const4283 const char *getPassName() const override { return "SROA"; }
4284 static char ID;
4285 };
4286
4287 char SROALegacyPass::ID = 0;
4288
createSROAPass()4289 FunctionPass *llvm::createSROAPass() { return new SROALegacyPass(); }
4290
4291 INITIALIZE_PASS_BEGIN(SROALegacyPass, "sroa",
4292 "Scalar Replacement Of Aggregates", false, false)
4293 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
4294 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
4295 INITIALIZE_PASS_END(SROALegacyPass, "sroa", "Scalar Replacement Of Aggregates",
4296 false, false)
4297