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.h"
27#include "llvm/ADT/STLExtras.h"
28#include "llvm/ADT/SetVector.h"
29#include "llvm/ADT/SmallVector.h"
30#include "llvm/ADT/Statistic.h"
31#include "llvm/Analysis/Loads.h"
32#include "llvm/Analysis/PtrUseVisitor.h"
33#include "llvm/Analysis/ValueTracking.h"
34#include "llvm/IR/Constants.h"
35#include "llvm/IR/DIBuilder.h"
36#include "llvm/IR/DataLayout.h"
37#include "llvm/IR/DebugInfo.h"
38#include "llvm/IR/DerivedTypes.h"
39#include "llvm/IR/Dominators.h"
40#include "llvm/IR/Function.h"
41#include "llvm/IR/IRBuilder.h"
42#include "llvm/IR/InstVisitor.h"
43#include "llvm/IR/Instructions.h"
44#include "llvm/IR/IntrinsicInst.h"
45#include "llvm/IR/LLVMContext.h"
46#include "llvm/IR/Operator.h"
47#include "llvm/Pass.h"
48#include "llvm/Support/CommandLine.h"
49#include "llvm/Support/Compiler.h"
50#include "llvm/Support/Debug.h"
51#include "llvm/Support/ErrorHandling.h"
52#include "llvm/Support/MathExtras.h"
53#include "llvm/Support/TimeValue.h"
54#include "llvm/Support/raw_ostream.h"
55#include "llvm/Transforms/Utils/Local.h"
56#include "llvm/Transforms/Utils/PromoteMemToReg.h"
57#include "llvm/Transforms/Utils/SSAUpdater.h"
58
59#if __cplusplus >= 201103L && !defined(NDEBUG)
60// We only use this for a debug check in C++11
61#include <random>
62#endif
63
64using namespace llvm;
65
66#define DEBUG_TYPE "sroa"
67
68STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
69STATISTIC(NumAllocaPartitions, "Number of alloca partitions formed");
70STATISTIC(MaxPartitionsPerAlloca, "Maximum number of partitions per alloca");
71STATISTIC(NumAllocaPartitionUses, "Number of alloca partition uses rewritten");
72STATISTIC(MaxUsesPerAllocaPartition, "Maximum number of uses of a partition");
73STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
74STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
75STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
76STATISTIC(NumDeleted, "Number of instructions deleted");
77STATISTIC(NumVectorized, "Number of vectorized aggregates");
78
79/// Hidden option to force the pass to not use DomTree and mem2reg, instead
80/// forming SSA values through the SSAUpdater infrastructure.
81static cl::opt<bool>
82ForceSSAUpdater("force-ssa-updater", cl::init(false), cl::Hidden);
83
84/// Hidden option to enable randomly shuffling the slices to help uncover
85/// instability in their order.
86static cl::opt<bool> SROARandomShuffleSlices("sroa-random-shuffle-slices",
87 cl::init(false), cl::Hidden);
88
89/// Hidden option to experiment with completely strict handling of inbounds
90/// GEPs.
91static cl::opt<bool> SROAStrictInbounds("sroa-strict-inbounds",
92 cl::init(false), cl::Hidden);
93
94namespace {
95/// \brief A custom IRBuilder inserter which prefixes all names if they are
96/// preserved.
97template <bool preserveNames = true>
98class IRBuilderPrefixedInserter :
99 public IRBuilderDefaultInserter<preserveNames> {
100 std::string Prefix;
101
102public:
103 void SetNamePrefix(const Twine &P) { Prefix = P.str(); }
104
105protected:
106 void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
107 BasicBlock::iterator InsertPt) const {
108 IRBuilderDefaultInserter<preserveNames>::InsertHelper(
109 I, Name.isTriviallyEmpty() ? Name : Prefix + Name, BB, InsertPt);
110 }
111};
112
113// Specialization for not preserving the name is trivial.
114template <>
115class IRBuilderPrefixedInserter<false> :
116 public IRBuilderDefaultInserter<false> {
117public:
118 void SetNamePrefix(const Twine &P) {}
119};
120
121/// \brief Provide a typedef for IRBuilder that drops names in release builds.
122#ifndef NDEBUG
123typedef llvm::IRBuilder<true, ConstantFolder,
124 IRBuilderPrefixedInserter<true> > IRBuilderTy;
125#else
126typedef llvm::IRBuilder<false, ConstantFolder,
127 IRBuilderPrefixedInserter<false> > IRBuilderTy;
128#endif
129}
130
131namespace {
132/// \brief A used slice of an alloca.
133///
134/// This structure represents a slice of an alloca used by some instruction. It
135/// stores both the begin and end offsets of this use, a pointer to the use
136/// itself, and a flag indicating whether we can classify the use as splittable
137/// or not when forming partitions of the alloca.
138class Slice {
139 /// \brief The beginning offset of the range.
140 uint64_t BeginOffset;
141
142 /// \brief The ending offset, not included in the range.
143 uint64_t EndOffset;
144
145 /// \brief Storage for both the use of this slice and whether it can be
146 /// split.
147 PointerIntPair<Use *, 1, bool> UseAndIsSplittable;
148
149public:
150 Slice() : BeginOffset(), EndOffset() {}
151 Slice(uint64_t BeginOffset, uint64_t EndOffset, Use *U, bool IsSplittable)
152 : BeginOffset(BeginOffset), EndOffset(EndOffset),
153 UseAndIsSplittable(U, IsSplittable) {}
154
155 uint64_t beginOffset() const { return BeginOffset; }
156 uint64_t endOffset() const { return EndOffset; }
157
158 bool isSplittable() const { return UseAndIsSplittable.getInt(); }
159 void makeUnsplittable() { UseAndIsSplittable.setInt(false); }
160
161 Use *getUse() const { return UseAndIsSplittable.getPointer(); }
162
163 bool isDead() const { return getUse() == nullptr; }
164 void kill() { UseAndIsSplittable.setPointer(nullptr); }
165
166 /// \brief Support for ordering ranges.
167 ///
168 /// This provides an ordering over ranges such that start offsets are
169 /// always increasing, and within equal start offsets, the end offsets are
170 /// decreasing. Thus the spanning range comes first in a cluster with the
171 /// same start position.
172 bool operator<(const Slice &RHS) const {
173 if (beginOffset() < RHS.beginOffset()) return true;
174 if (beginOffset() > RHS.beginOffset()) return false;
175 if (isSplittable() != RHS.isSplittable()) return !isSplittable();
176 if (endOffset() > RHS.endOffset()) return true;
177 return false;
178 }
179
180 /// \brief Support comparison with a single offset to allow binary searches.
181 friend LLVM_ATTRIBUTE_UNUSED bool operator<(const Slice &LHS,
182 uint64_t RHSOffset) {
183 return LHS.beginOffset() < RHSOffset;
184 }
185 friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
186 const Slice &RHS) {
187 return LHSOffset < RHS.beginOffset();
188 }
189
190 bool operator==(const Slice &RHS) const {
191 return isSplittable() == RHS.isSplittable() &&
192 beginOffset() == RHS.beginOffset() && endOffset() == RHS.endOffset();
193 }
194 bool operator!=(const Slice &RHS) const { return !operator==(RHS); }
195};
196} // end anonymous namespace
197
198namespace llvm {
199template <typename T> struct isPodLike;
200template <> struct isPodLike<Slice> {
201 static const bool value = true;
202};
203}
204
205namespace {
206/// \brief Representation of the alloca slices.
207///
208/// This class represents the slices of an alloca which are formed by its
209/// various uses. If a pointer escapes, we can't fully build a representation
210/// for the slices used and we reflect that in this structure. The uses are
211/// stored, sorted by increasing beginning offset and with unsplittable slices
212/// starting at a particular offset before splittable slices.
213class AllocaSlices {
214public:
215 /// \brief Construct the slices of a particular alloca.
216 AllocaSlices(const DataLayout &DL, AllocaInst &AI);
217
218 /// \brief Test whether a pointer to the allocation escapes our analysis.
219 ///
220 /// If this is true, the slices are never fully built and should be
221 /// ignored.
222 bool isEscaped() const { return PointerEscapingInstr; }
223
224 /// \brief Support for iterating over the slices.
225 /// @{
226 typedef SmallVectorImpl<Slice>::iterator iterator;
227 iterator begin() { return Slices.begin(); }
228 iterator end() { return Slices.end(); }
229
230 typedef SmallVectorImpl<Slice>::const_iterator const_iterator;
231 const_iterator begin() const { return Slices.begin(); }
232 const_iterator end() const { return Slices.end(); }
233 /// @}
234
235 /// \brief Allow iterating the dead users for this alloca.
236 ///
237 /// These are instructions which will never actually use the alloca as they
238 /// are outside the allocated range. They are safe to replace with undef and
239 /// delete.
240 /// @{
241 typedef SmallVectorImpl<Instruction *>::const_iterator dead_user_iterator;
242 dead_user_iterator dead_user_begin() const { return DeadUsers.begin(); }
243 dead_user_iterator dead_user_end() const { return DeadUsers.end(); }
244 /// @}
245
246 /// \brief Allow iterating the dead expressions referring to this alloca.
247 ///
248 /// These are operands which have cannot actually be used to refer to the
249 /// alloca as they are outside its range and the user doesn't correct for
250 /// that. These mostly consist of PHI node inputs and the like which we just
251 /// need to replace with undef.
252 /// @{
253 typedef SmallVectorImpl<Use *>::const_iterator dead_op_iterator;
254 dead_op_iterator dead_op_begin() const { return DeadOperands.begin(); }
255 dead_op_iterator dead_op_end() const { return DeadOperands.end(); }
256 /// @}
257
258#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
259 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
260 void printSlice(raw_ostream &OS, const_iterator I,
261 StringRef Indent = " ") const;
262 void printUse(raw_ostream &OS, const_iterator I,
263 StringRef Indent = " ") const;
264 void print(raw_ostream &OS) const;
265 void dump(const_iterator I) const;
266 void dump() const;
267#endif
268
269private:
270 template <typename DerivedT, typename RetT = void> class BuilderBase;
271 class SliceBuilder;
272 friend class AllocaSlices::SliceBuilder;
273
274#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
275 /// \brief Handle to alloca instruction to simplify method interfaces.
276 AllocaInst &AI;
277#endif
278
279 /// \brief The instruction responsible for this alloca not having a known set
280 /// of slices.
281 ///
282 /// When an instruction (potentially) escapes the pointer to the alloca, we
283 /// store a pointer to that here and abort trying to form slices of the
284 /// alloca. This will be null if the alloca slices are analyzed successfully.
285 Instruction *PointerEscapingInstr;
286
287 /// \brief The slices of the alloca.
288 ///
289 /// We store a vector of the slices formed by uses of the alloca here. This
290 /// vector is sorted by increasing begin offset, and then the unsplittable
291 /// slices before the splittable ones. See the Slice inner class for more
292 /// details.
293 SmallVector<Slice, 8> Slices;
294
295 /// \brief Instructions which will become dead if we rewrite the alloca.
296 ///
297 /// Note that these are not separated by slice. This is because we expect an
298 /// alloca to be completely rewritten or not rewritten at all. If rewritten,
299 /// all these instructions can simply be removed and replaced with undef as
300 /// they come from outside of the allocated space.
301 SmallVector<Instruction *, 8> DeadUsers;
302
303 /// \brief Operands which will become dead if we rewrite the alloca.
304 ///
305 /// These are operands that in their particular use can be replaced with
306 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
307 /// to PHI nodes and the like. They aren't entirely dead (there might be
308 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
309 /// want to swap this particular input for undef to simplify the use lists of
310 /// the alloca.
311 SmallVector<Use *, 8> DeadOperands;
312};
313}
314
315static Value *foldSelectInst(SelectInst &SI) {
316 // If the condition being selected on is a constant or the same value is
317 // being selected between, fold the select. Yes this does (rarely) happen
318 // early on.
319 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
320 return SI.getOperand(1+CI->isZero());
321 if (SI.getOperand(1) == SI.getOperand(2))
322 return SI.getOperand(1);
323
324 return nullptr;
325}
326
327/// \brief Builder for the alloca slices.
328///
329/// This class builds a set of alloca slices by recursively visiting the uses
330/// of an alloca and making a slice for each load and store at each offset.
331class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
332 friend class PtrUseVisitor<SliceBuilder>;
333 friend class InstVisitor<SliceBuilder>;
334 typedef PtrUseVisitor<SliceBuilder> Base;
335
336 const uint64_t AllocSize;
337 AllocaSlices &S;
338
339 SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
340 SmallDenseMap<Instruction *, uint64_t> PHIOrSelectSizes;
341
342 /// \brief Set to de-duplicate dead instructions found in the use walk.
343 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
344
345public:
346 SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &S)
347 : PtrUseVisitor<SliceBuilder>(DL),
348 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())), S(S) {}
349
350private:
351 void markAsDead(Instruction &I) {
352 if (VisitedDeadInsts.insert(&I))
353 S.DeadUsers.push_back(&I);
354 }
355
356 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
357 bool IsSplittable = false) {
358 // Completely skip uses which have a zero size or start either before or
359 // past the end of the allocation.
360 if (Size == 0 || Offset.uge(AllocSize)) {
361 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
362 << " which has zero size or starts outside of the "
363 << AllocSize << " byte alloca:\n"
364 << " alloca: " << S.AI << "\n"
365 << " use: " << I << "\n");
366 return markAsDead(I);
367 }
368
369 uint64_t BeginOffset = Offset.getZExtValue();
370 uint64_t EndOffset = BeginOffset + Size;
371
372 // Clamp the end offset to the end of the allocation. Note that this is
373 // formulated to handle even the case where "BeginOffset + Size" overflows.
374 // This may appear superficially to be something we could ignore entirely,
375 // but that is not so! There may be widened loads or PHI-node uses where
376 // some instructions are dead but not others. We can't completely ignore
377 // them, and so have to record at least the information here.
378 assert(AllocSize >= BeginOffset); // Established above.
379 if (Size > AllocSize - BeginOffset) {
380 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
381 << " to remain within the " << AllocSize << " byte alloca:\n"
382 << " alloca: " << S.AI << "\n"
383 << " use: " << I << "\n");
384 EndOffset = AllocSize;
385 }
386
387 S.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
388 }
389
390 void visitBitCastInst(BitCastInst &BC) {
391 if (BC.use_empty())
392 return markAsDead(BC);
393
394 return Base::visitBitCastInst(BC);
395 }
396
397 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
398 if (GEPI.use_empty())
399 return markAsDead(GEPI);
400
401 if (SROAStrictInbounds && GEPI.isInBounds()) {
402 // FIXME: This is a manually un-factored variant of the basic code inside
403 // of GEPs with checking of the inbounds invariant specified in the
404 // langref in a very strict sense. If we ever want to enable
405 // SROAStrictInbounds, this code should be factored cleanly into
406 // PtrUseVisitor, but it is easier to experiment with SROAStrictInbounds
407 // by writing out the code here where we have tho underlying allocation
408 // size readily available.
409 APInt GEPOffset = Offset;
410 for (gep_type_iterator GTI = gep_type_begin(GEPI),
411 GTE = gep_type_end(GEPI);
412 GTI != GTE; ++GTI) {
413 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
414 if (!OpC)
415 break;
416
417 // Handle a struct index, which adds its field offset to the pointer.
418 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
419 unsigned ElementIdx = OpC->getZExtValue();
420 const StructLayout *SL = DL.getStructLayout(STy);
421 GEPOffset +=
422 APInt(Offset.getBitWidth(), SL->getElementOffset(ElementIdx));
423 } else {
424 // For array or vector indices, scale the index by the size of the type.
425 APInt Index = OpC->getValue().sextOrTrunc(Offset.getBitWidth());
426 GEPOffset += Index * APInt(Offset.getBitWidth(),
427 DL.getTypeAllocSize(GTI.getIndexedType()));
428 }
429
430 // If this index has computed an intermediate pointer which is not
431 // inbounds, then the result of the GEP is a poison value and we can
432 // delete it and all uses.
433 if (GEPOffset.ugt(AllocSize))
434 return markAsDead(GEPI);
435 }
436 }
437
438 return Base::visitGetElementPtrInst(GEPI);
439 }
440
441 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
442 uint64_t Size, bool IsVolatile) {
443 // We allow splitting of loads and stores where the type is an integer type
444 // and cover the entire alloca. This prevents us from splitting over
445 // eagerly.
446 // FIXME: In the great blue eventually, we should eagerly split all integer
447 // loads and stores, and then have a separate step that merges adjacent
448 // alloca partitions into a single partition suitable for integer widening.
449 // Or we should skip the merge step and rely on GVN and other passes to
450 // merge adjacent loads and stores that survive mem2reg.
451 bool IsSplittable =
452 Ty->isIntegerTy() && !IsVolatile && Offset == 0 && Size >= AllocSize;
453
454 insertUse(I, Offset, Size, IsSplittable);
455 }
456
457 void visitLoadInst(LoadInst &LI) {
458 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
459 "All simple FCA loads should have been pre-split");
460
461 if (!IsOffsetKnown)
462 return PI.setAborted(&LI);
463
464 uint64_t Size = DL.getTypeStoreSize(LI.getType());
465 return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
466 }
467
468 void visitStoreInst(StoreInst &SI) {
469 Value *ValOp = SI.getValueOperand();
470 if (ValOp == *U)
471 return PI.setEscapedAndAborted(&SI);
472 if (!IsOffsetKnown)
473 return PI.setAborted(&SI);
474
475 uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
476
477 // If this memory access can be shown to *statically* extend outside the
478 // bounds of of the allocation, it's behavior is undefined, so simply
479 // ignore it. Note that this is more strict than the generic clamping
480 // behavior of insertUse. We also try to handle cases which might run the
481 // risk of overflow.
482 // FIXME: We should instead consider the pointer to have escaped if this
483 // function is being instrumented for addressing bugs or race conditions.
484 if (Size > AllocSize || Offset.ugt(AllocSize - Size)) {
485 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @" << Offset
486 << " which extends past the end of the " << AllocSize
487 << " byte alloca:\n"
488 << " alloca: " << S.AI << "\n"
489 << " use: " << SI << "\n");
490 return markAsDead(SI);
491 }
492
493 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
494 "All simple FCA stores should have been pre-split");
495 handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
496 }
497
498
499 void visitMemSetInst(MemSetInst &II) {
500 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
501 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
502 if ((Length && Length->getValue() == 0) ||
503 (IsOffsetKnown && Offset.uge(AllocSize)))
504 // Zero-length mem transfer intrinsics can be ignored entirely.
505 return markAsDead(II);
506
507 if (!IsOffsetKnown)
508 return PI.setAborted(&II);
509
510 insertUse(II, Offset,
511 Length ? Length->getLimitedValue()
512 : AllocSize - Offset.getLimitedValue(),
513 (bool)Length);
514 }
515
516 void visitMemTransferInst(MemTransferInst &II) {
517 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
518 if (Length && Length->getValue() == 0)
519 // Zero-length mem transfer intrinsics can be ignored entirely.
520 return markAsDead(II);
521
522 // Because we can visit these intrinsics twice, also check to see if the
523 // first time marked this instruction as dead. If so, skip it.
524 if (VisitedDeadInsts.count(&II))
525 return;
526
527 if (!IsOffsetKnown)
528 return PI.setAborted(&II);
529
530 // This side of the transfer is completely out-of-bounds, and so we can
531 // nuke the entire transfer. However, we also need to nuke the other side
532 // if already added to our partitions.
533 // FIXME: Yet another place we really should bypass this when
534 // instrumenting for ASan.
535 if (Offset.uge(AllocSize)) {
536 SmallDenseMap<Instruction *, unsigned>::iterator MTPI = MemTransferSliceMap.find(&II);
537 if (MTPI != MemTransferSliceMap.end())
538 S.Slices[MTPI->second].kill();
539 return markAsDead(II);
540 }
541
542 uint64_t RawOffset = Offset.getLimitedValue();
543 uint64_t Size = Length ? Length->getLimitedValue()
544 : AllocSize - RawOffset;
545
546 // Check for the special case where the same exact value is used for both
547 // source and dest.
548 if (*U == II.getRawDest() && *U == II.getRawSource()) {
549 // For non-volatile transfers this is a no-op.
550 if (!II.isVolatile())
551 return markAsDead(II);
552
553 return insertUse(II, Offset, Size, /*IsSplittable=*/false);
554 }
555
556 // If we have seen both source and destination for a mem transfer, then
557 // they both point to the same alloca.
558 bool Inserted;
559 SmallDenseMap<Instruction *, unsigned>::iterator MTPI;
560 std::tie(MTPI, Inserted) =
561 MemTransferSliceMap.insert(std::make_pair(&II, S.Slices.size()));
562 unsigned PrevIdx = MTPI->second;
563 if (!Inserted) {
564 Slice &PrevP = S.Slices[PrevIdx];
565
566 // Check if the begin offsets match and this is a non-volatile transfer.
567 // In that case, we can completely elide the transfer.
568 if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
569 PrevP.kill();
570 return markAsDead(II);
571 }
572
573 // Otherwise we have an offset transfer within the same alloca. We can't
574 // split those.
575 PrevP.makeUnsplittable();
576 }
577
578 // Insert the use now that we've fixed up the splittable nature.
579 insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
580
581 // Check that we ended up with a valid index in the map.
582 assert(S.Slices[PrevIdx].getUse()->getUser() == &II &&
583 "Map index doesn't point back to a slice with this user.");
584 }
585
586 // Disable SRoA for any intrinsics except for lifetime invariants.
587 // FIXME: What about debug intrinsics? This matches old behavior, but
588 // doesn't make sense.
589 void visitIntrinsicInst(IntrinsicInst &II) {
590 if (!IsOffsetKnown)
591 return PI.setAborted(&II);
592
593 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
594 II.getIntrinsicID() == Intrinsic::lifetime_end) {
595 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
596 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
597 Length->getLimitedValue());
598 insertUse(II, Offset, Size, true);
599 return;
600 }
601
602 Base::visitIntrinsicInst(II);
603 }
604
605 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
606 // We consider any PHI or select that results in a direct load or store of
607 // the same offset to be a viable use for slicing purposes. These uses
608 // are considered unsplittable and the size is the maximum loaded or stored
609 // size.
610 SmallPtrSet<Instruction *, 4> Visited;
611 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
612 Visited.insert(Root);
613 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
614 // If there are no loads or stores, the access is dead. We mark that as
615 // a size zero access.
616 Size = 0;
617 do {
618 Instruction *I, *UsedI;
619 std::tie(UsedI, I) = Uses.pop_back_val();
620
621 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
622 Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
623 continue;
624 }
625 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
626 Value *Op = SI->getOperand(0);
627 if (Op == UsedI)
628 return SI;
629 Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
630 continue;
631 }
632
633 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
634 if (!GEP->hasAllZeroIndices())
635 return GEP;
636 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
637 !isa<SelectInst>(I)) {
638 return I;
639 }
640
641 for (User *U : I->users())
642 if (Visited.insert(cast<Instruction>(U)))
643 Uses.push_back(std::make_pair(I, cast<Instruction>(U)));
644 } while (!Uses.empty());
645
646 return nullptr;
647 }
648
649 void visitPHINode(PHINode &PN) {
650 if (PN.use_empty())
651 return markAsDead(PN);
652 if (!IsOffsetKnown)
653 return PI.setAborted(&PN);
654
655 // See if we already have computed info on this node.
656 uint64_t &PHISize = PHIOrSelectSizes[&PN];
657 if (!PHISize) {
658 // This is a new PHI node, check for an unsafe use of the PHI node.
659 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&PN, PHISize))
660 return PI.setAborted(UnsafeI);
661 }
662
663 // For PHI and select operands outside the alloca, we can't nuke the entire
664 // phi or select -- the other side might still be relevant, so we special
665 // case them here and use a separate structure to track the operands
666 // themselves which should be replaced with undef.
667 // FIXME: This should instead be escaped in the event we're instrumenting
668 // for address sanitization.
669 if (Offset.uge(AllocSize)) {
670 S.DeadOperands.push_back(U);
671 return;
672 }
673
674 insertUse(PN, Offset, PHISize);
675 }
676
677 void visitSelectInst(SelectInst &SI) {
678 if (SI.use_empty())
679 return markAsDead(SI);
680 if (Value *Result = foldSelectInst(SI)) {
681 if (Result == *U)
682 // If the result of the constant fold will be the pointer, recurse
683 // through the select as if we had RAUW'ed it.
684 enqueueUsers(SI);
685 else
686 // Otherwise the operand to the select is dead, and we can replace it
687 // with undef.
688 S.DeadOperands.push_back(U);
689
690 return;
691 }
692 if (!IsOffsetKnown)
693 return PI.setAborted(&SI);
694
695 // See if we already have computed info on this node.
696 uint64_t &SelectSize = PHIOrSelectSizes[&SI];
697 if (!SelectSize) {
698 // This is a new Select, check for an unsafe use of it.
699 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&SI, SelectSize))
700 return PI.setAborted(UnsafeI);
701 }
702
703 // For PHI and select operands outside the alloca, we can't nuke the entire
704 // phi or select -- the other side might still be relevant, so we special
705 // case them here and use a separate structure to track the operands
706 // themselves which should be replaced with undef.
707 // FIXME: This should instead be escaped in the event we're instrumenting
708 // for address sanitization.
709 if (Offset.uge(AllocSize)) {
710 S.DeadOperands.push_back(U);
711 return;
712 }
713
714 insertUse(SI, Offset, SelectSize);
715 }
716
717 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
718 void visitInstruction(Instruction &I) {
719 PI.setAborted(&I);
720 }
721};
722
723AllocaSlices::AllocaSlices(const DataLayout &DL, AllocaInst &AI)
724 :
725#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
726 AI(AI),
727#endif
728 PointerEscapingInstr(nullptr) {
729 SliceBuilder PB(DL, AI, *this);
730 SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
731 if (PtrI.isEscaped() || PtrI.isAborted()) {
732 // FIXME: We should sink the escape vs. abort info into the caller nicely,
733 // possibly by just storing the PtrInfo in the AllocaSlices.
734 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
735 : PtrI.getAbortingInst();
736 assert(PointerEscapingInstr && "Did not track a bad instruction");
737 return;
738 }
739
740 Slices.erase(std::remove_if(Slices.begin(), Slices.end(),
741 std::mem_fun_ref(&Slice::isDead)),
742 Slices.end());
743
744#if __cplusplus >= 201103L && !defined(NDEBUG)
745 if (SROARandomShuffleSlices) {
746 std::mt19937 MT(static_cast<unsigned>(sys::TimeValue::now().msec()));
747 std::shuffle(Slices.begin(), Slices.end(), MT);
748 }
749#endif
750
751 // Sort the uses. This arranges for the offsets to be in ascending order,
752 // and the sizes to be in descending order.
753 std::sort(Slices.begin(), Slices.end());
754}
755
756#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
757
758void AllocaSlices::print(raw_ostream &OS, const_iterator I,
759 StringRef Indent) const {
760 printSlice(OS, I, Indent);
761 printUse(OS, I, Indent);
762}
763
764void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
765 StringRef Indent) const {
766 OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
767 << " slice #" << (I - begin())
768 << (I->isSplittable() ? " (splittable)" : "") << "\n";
769}
770
771void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
772 StringRef Indent) const {
773 OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
774}
775
776void AllocaSlices::print(raw_ostream &OS) const {
777 if (PointerEscapingInstr) {
778 OS << "Can't analyze slices for alloca: " << AI << "\n"
779 << " A pointer to this alloca escaped by:\n"
780 << " " << *PointerEscapingInstr << "\n";
781 return;
782 }
783
784 OS << "Slices of alloca: " << AI << "\n";
785 for (const_iterator I = begin(), E = end(); I != E; ++I)
786 print(OS, I);
787}
788
789LLVM_DUMP_METHOD void AllocaSlices::dump(const_iterator I) const {
790 print(dbgs(), I);
791}
792LLVM_DUMP_METHOD void AllocaSlices::dump() const { print(dbgs()); }
793
794#endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
795
796namespace {
797/// \brief Implementation of LoadAndStorePromoter for promoting allocas.
798///
799/// This subclass of LoadAndStorePromoter adds overrides to handle promoting
800/// the loads and stores of an alloca instruction, as well as updating its
801/// debug information. This is used when a domtree is unavailable and thus
802/// mem2reg in its full form can't be used to handle promotion of allocas to
803/// scalar values.
804class AllocaPromoter : public LoadAndStorePromoter {
805 AllocaInst &AI;
806 DIBuilder &DIB;
807
808 SmallVector<DbgDeclareInst *, 4> DDIs;
809 SmallVector<DbgValueInst *, 4> DVIs;
810
811public:
812 AllocaPromoter(const SmallVectorImpl<Instruction *> &Insts, SSAUpdater &S,
813 AllocaInst &AI, DIBuilder &DIB)
814 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
815
816 void run(const SmallVectorImpl<Instruction*> &Insts) {
817 // Retain the debug information attached to the alloca for use when
818 // rewriting loads and stores.
819 if (MDNode *DebugNode = MDNode::getIfExists(AI.getContext(), &AI)) {
820 for (User *U : DebugNode->users())
821 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U))
822 DDIs.push_back(DDI);
823 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(U))
824 DVIs.push_back(DVI);
825 }
826
827 LoadAndStorePromoter::run(Insts);
828
829 // While we have the debug information, clear it off of the alloca. The
830 // caller takes care of deleting the alloca.
831 while (!DDIs.empty())
832 DDIs.pop_back_val()->eraseFromParent();
833 while (!DVIs.empty())
834 DVIs.pop_back_val()->eraseFromParent();
835 }
836
837 bool isInstInList(Instruction *I,
838 const SmallVectorImpl<Instruction*> &Insts) const override {
839 Value *Ptr;
840 if (LoadInst *LI = dyn_cast<LoadInst>(I))
841 Ptr = LI->getOperand(0);
842 else
843 Ptr = cast<StoreInst>(I)->getPointerOperand();
844
845 // Only used to detect cycles, which will be rare and quickly found as
846 // we're walking up a chain of defs rather than down through uses.
847 SmallPtrSet<Value *, 4> Visited;
848
849 do {
850 if (Ptr == &AI)
851 return true;
852
853 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr))
854 Ptr = BCI->getOperand(0);
855 else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Ptr))
856 Ptr = GEPI->getPointerOperand();
857 else
858 return false;
859
860 } while (Visited.insert(Ptr));
861
862 return false;
863 }
864
865 void updateDebugInfo(Instruction *Inst) const override {
866 for (SmallVectorImpl<DbgDeclareInst *>::const_iterator I = DDIs.begin(),
867 E = DDIs.end(); I != E; ++I) {
868 DbgDeclareInst *DDI = *I;
869 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
870 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
871 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
872 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
873 }
874 for (SmallVectorImpl<DbgValueInst *>::const_iterator I = DVIs.begin(),
875 E = DVIs.end(); I != E; ++I) {
876 DbgValueInst *DVI = *I;
877 Value *Arg = nullptr;
878 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
879 // If an argument is zero extended then use argument directly. The ZExt
880 // may be zapped by an optimization pass in future.
881 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
882 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
883 else if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
884 Arg = dyn_cast<Argument>(SExt->getOperand(0));
885 if (!Arg)
886 Arg = SI->getValueOperand();
887 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
888 Arg = LI->getPointerOperand();
889 } else {
890 continue;
891 }
892 Instruction *DbgVal =
893 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
894 Inst);
895 DbgVal->setDebugLoc(DVI->getDebugLoc());
896 }
897 }
898};
899} // end anon namespace
900
901
902namespace {
903/// \brief An optimization pass providing Scalar Replacement of Aggregates.
904///
905/// This pass takes allocations which can be completely analyzed (that is, they
906/// don't escape) and tries to turn them into scalar SSA values. There are
907/// a few steps to this process.
908///
909/// 1) It takes allocations of aggregates and analyzes the ways in which they
910/// are used to try to split them into smaller allocations, ideally of
911/// a single scalar data type. It will split up memcpy and memset accesses
912/// as necessary and try to isolate individual scalar accesses.
913/// 2) It will transform accesses into forms which are suitable for SSA value
914/// promotion. This can be replacing a memset with a scalar store of an
915/// integer value, or it can involve speculating operations on a PHI or
916/// select to be a PHI or select of the results.
917/// 3) Finally, this will try to detect a pattern of accesses which map cleanly
918/// onto insert and extract operations on a vector value, and convert them to
919/// this form. By doing so, it will enable promotion of vector aggregates to
920/// SSA vector values.
921class SROA : public FunctionPass {
922 const bool RequiresDomTree;
923
924 LLVMContext *C;
925 const DataLayout *DL;
926 DominatorTree *DT;
927
928 /// \brief Worklist of alloca instructions to simplify.
929 ///
930 /// Each alloca in the function is added to this. Each new alloca formed gets
931 /// added to it as well to recursively simplify unless that alloca can be
932 /// directly promoted. Finally, each time we rewrite a use of an alloca other
933 /// the one being actively rewritten, we add it back onto the list if not
934 /// already present to ensure it is re-visited.
935 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > Worklist;
936
937 /// \brief A collection of instructions to delete.
938 /// We try to batch deletions to simplify code and make things a bit more
939 /// efficient.
940 SetVector<Instruction *, SmallVector<Instruction *, 8> > DeadInsts;
941
942 /// \brief Post-promotion worklist.
943 ///
944 /// Sometimes we discover an alloca which has a high probability of becoming
945 /// viable for SROA after a round of promotion takes place. In those cases,
946 /// the alloca is enqueued here for re-processing.
947 ///
948 /// Note that we have to be very careful to clear allocas out of this list in
949 /// the event they are deleted.
950 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > PostPromotionWorklist;
951
952 /// \brief A collection of alloca instructions we can directly promote.
953 std::vector<AllocaInst *> PromotableAllocas;
954
955 /// \brief A worklist of PHIs to speculate prior to promoting allocas.
956 ///
957 /// All of these PHIs have been checked for the safety of speculation and by
958 /// being speculated will allow promoting allocas currently in the promotable
959 /// queue.
960 SetVector<PHINode *, SmallVector<PHINode *, 2> > SpeculatablePHIs;
961
962 /// \brief A worklist of select instructions to speculate prior to promoting
963 /// allocas.
964 ///
965 /// All of these select instructions have been checked for the safety of
966 /// speculation and by being speculated will allow promoting allocas
967 /// currently in the promotable queue.
968 SetVector<SelectInst *, SmallVector<SelectInst *, 2> > SpeculatableSelects;
969
970public:
971 SROA(bool RequiresDomTree = true)
972 : FunctionPass(ID), RequiresDomTree(RequiresDomTree),
973 C(nullptr), DL(nullptr), DT(nullptr) {
974 initializeSROAPass(*PassRegistry::getPassRegistry());
975 }
976 bool runOnFunction(Function &F) override;
977 void getAnalysisUsage(AnalysisUsage &AU) const override;
978
979 const char *getPassName() const override { return "SROA"; }
980 static char ID;
981
982private:
983 friend class PHIOrSelectSpeculator;
984 friend class AllocaSliceRewriter;
985
986 bool rewritePartition(AllocaInst &AI, AllocaSlices &S,
987 AllocaSlices::iterator B, AllocaSlices::iterator E,
988 int64_t BeginOffset, int64_t EndOffset,
989 ArrayRef<AllocaSlices::iterator> SplitUses);
990 bool splitAlloca(AllocaInst &AI, AllocaSlices &S);
991 bool runOnAlloca(AllocaInst &AI);
992 void clobberUse(Use &U);
993 void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas);
994 bool promoteAllocas(Function &F);
995};
996}
997
998char SROA::ID = 0;
999
1000FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
1001 return new SROA(RequiresDomTree);
1002}
1003
1004INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates",
1005 false, false)
1006INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1007INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
1008 false, false)
1009
1010/// Walk the range of a partitioning looking for a common type to cover this
1011/// sequence of slices.
1012static Type *findCommonType(AllocaSlices::const_iterator B,
1013 AllocaSlices::const_iterator E,
1014 uint64_t EndOffset) {
1015 Type *Ty = nullptr;
1016 bool TyIsCommon = true;
1017 IntegerType *ITy = nullptr;
1018
1019 // Note that we need to look at *every* alloca slice's Use to ensure we
1020 // always get consistent results regardless of the order of slices.
1021 for (AllocaSlices::const_iterator I = B; I != E; ++I) {
1022 Use *U = I->getUse();
1023 if (isa<IntrinsicInst>(*U->getUser()))
1024 continue;
1025 if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
1026 continue;
1027
1028 Type *UserTy = nullptr;
1029 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1030 UserTy = LI->getType();
1031 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1032 UserTy = SI->getValueOperand()->getType();
1033 }
1034
1035 if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
1036 // If the type is larger than the partition, skip it. We only encounter
1037 // this for split integer operations where we want to use the type of the
1038 // entity causing the split. Also skip if the type is not a byte width
1039 // multiple.
1040 if (UserITy->getBitWidth() % 8 != 0 ||
1041 UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
1042 continue;
1043
1044 // Track the largest bitwidth integer type used in this way in case there
1045 // is no common type.
1046 if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
1047 ITy = UserITy;
1048 }
1049
1050 // To avoid depending on the order of slices, Ty and TyIsCommon must not
1051 // depend on types skipped above.
1052 if (!UserTy || (Ty && Ty != UserTy))
1053 TyIsCommon = false; // Give up on anything but an iN type.
1054 else
1055 Ty = UserTy;
1056 }
1057
1058 return TyIsCommon ? Ty : ITy;
1059}
1060
1061/// PHI instructions that use an alloca and are subsequently loaded can be
1062/// rewritten to load both input pointers in the pred blocks and then PHI the
1063/// results, allowing the load of the alloca to be promoted.
1064/// From this:
1065/// %P2 = phi [i32* %Alloca, i32* %Other]
1066/// %V = load i32* %P2
1067/// to:
1068/// %V1 = load i32* %Alloca -> will be mem2reg'd
1069/// ...
1070/// %V2 = load i32* %Other
1071/// ...
1072/// %V = phi [i32 %V1, i32 %V2]
1073///
1074/// We can do this to a select if its only uses are loads and if the operands
1075/// to the select can be loaded unconditionally.
1076///
1077/// FIXME: This should be hoisted into a generic utility, likely in
1078/// Transforms/Util/Local.h
1079static bool isSafePHIToSpeculate(PHINode &PN,
1080 const DataLayout *DL = nullptr) {
1081 // For now, we can only do this promotion if the load is in the same block
1082 // as the PHI, and if there are no stores between the phi and load.
1083 // TODO: Allow recursive phi users.
1084 // TODO: Allow stores.
1085 BasicBlock *BB = PN.getParent();
1086 unsigned MaxAlign = 0;
1087 bool HaveLoad = false;
1088 for (User *U : PN.users()) {
1089 LoadInst *LI = dyn_cast<LoadInst>(U);
1090 if (!LI || !LI->isSimple())
1091 return false;
1092
1093 // For now we only allow loads in the same block as the PHI. This is
1094 // a common case that happens when instcombine merges two loads through
1095 // a PHI.
1096 if (LI->getParent() != BB)
1097 return false;
1098
1099 // Ensure that there are no instructions between the PHI and the load that
1100 // could store.
1101 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1102 if (BBI->mayWriteToMemory())
1103 return false;
1104
1105 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1106 HaveLoad = true;
1107 }
1108
1109 if (!HaveLoad)
1110 return false;
1111
1112 // We can only transform this if it is safe to push the loads into the
1113 // predecessor blocks. The only thing to watch out for is that we can't put
1114 // a possibly trapping load in the predecessor if it is a critical edge.
1115 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1116 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1117 Value *InVal = PN.getIncomingValue(Idx);
1118
1119 // If the value is produced by the terminator of the predecessor (an
1120 // invoke) or it has side-effects, there is no valid place to put a load
1121 // in the predecessor.
1122 if (TI == InVal || TI->mayHaveSideEffects())
1123 return false;
1124
1125 // If the predecessor has a single successor, then the edge isn't
1126 // critical.
1127 if (TI->getNumSuccessors() == 1)
1128 continue;
1129
1130 // If this pointer is always safe to load, or if we can prove that there
1131 // is already a load in the block, then we can move the load to the pred
1132 // block.
1133 if (InVal->isDereferenceablePointer(DL) ||
1134 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, DL))
1135 continue;
1136
1137 return false;
1138 }
1139
1140 return true;
1141}
1142
1143static void speculatePHINodeLoads(PHINode &PN) {
1144 DEBUG(dbgs() << " original: " << PN << "\n");
1145
1146 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1147 IRBuilderTy PHIBuilder(&PN);
1148 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1149 PN.getName() + ".sroa.speculated");
1150
1151 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1152 // matter which one we get and if any differ.
1153 LoadInst *SomeLoad = cast<LoadInst>(PN.user_back());
1154 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1155 unsigned Align = SomeLoad->getAlignment();
1156
1157 // Rewrite all loads of the PN to use the new PHI.
1158 while (!PN.use_empty()) {
1159 LoadInst *LI = cast<LoadInst>(PN.user_back());
1160 LI->replaceAllUsesWith(NewPN);
1161 LI->eraseFromParent();
1162 }
1163
1164 // Inject loads into all of the pred blocks.
1165 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1166 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1167 TerminatorInst *TI = Pred->getTerminator();
1168 Value *InVal = PN.getIncomingValue(Idx);
1169 IRBuilderTy PredBuilder(TI);
1170
1171 LoadInst *Load = PredBuilder.CreateLoad(
1172 InVal, (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1173 ++NumLoadsSpeculated;
1174 Load->setAlignment(Align);
1175 if (TBAATag)
1176 Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1177 NewPN->addIncoming(Load, Pred);
1178 }
1179
1180 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1181 PN.eraseFromParent();
1182}
1183
1184/// Select instructions that use an alloca and are subsequently loaded can be
1185/// rewritten to load both input pointers and then select between the result,
1186/// allowing the load of the alloca to be promoted.
1187/// From this:
1188/// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1189/// %V = load i32* %P2
1190/// to:
1191/// %V1 = load i32* %Alloca -> will be mem2reg'd
1192/// %V2 = load i32* %Other
1193/// %V = select i1 %cond, i32 %V1, i32 %V2
1194///
1195/// We can do this to a select if its only uses are loads and if the operand
1196/// to the select can be loaded unconditionally.
1197static bool isSafeSelectToSpeculate(SelectInst &SI,
1198 const DataLayout *DL = nullptr) {
1199 Value *TValue = SI.getTrueValue();
1200 Value *FValue = SI.getFalseValue();
1201 bool TDerefable = TValue->isDereferenceablePointer(DL);
1202 bool FDerefable = FValue->isDereferenceablePointer(DL);
1203
1204 for (User *U : SI.users()) {
1205 LoadInst *LI = dyn_cast<LoadInst>(U);
1206 if (!LI || !LI->isSimple())
1207 return false;
1208
1209 // Both operands to the select need to be dereferencable, either
1210 // absolutely (e.g. allocas) or at this point because we can see other
1211 // accesses to it.
1212 if (!TDerefable &&
1213 !isSafeToLoadUnconditionally(TValue, LI, LI->getAlignment(), DL))
1214 return false;
1215 if (!FDerefable &&
1216 !isSafeToLoadUnconditionally(FValue, LI, LI->getAlignment(), DL))
1217 return false;
1218 }
1219
1220 return true;
1221}
1222
1223static void speculateSelectInstLoads(SelectInst &SI) {
1224 DEBUG(dbgs() << " original: " << SI << "\n");
1225
1226 IRBuilderTy IRB(&SI);
1227 Value *TV = SI.getTrueValue();
1228 Value *FV = SI.getFalseValue();
1229 // Replace the loads of the select with a select of two loads.
1230 while (!SI.use_empty()) {
1231 LoadInst *LI = cast<LoadInst>(SI.user_back());
1232 assert(LI->isSimple() && "We only speculate simple loads");
1233
1234 IRB.SetInsertPoint(LI);
1235 LoadInst *TL =
1236 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1237 LoadInst *FL =
1238 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1239 NumLoadsSpeculated += 2;
1240
1241 // Transfer alignment and TBAA info if present.
1242 TL->setAlignment(LI->getAlignment());
1243 FL->setAlignment(LI->getAlignment());
1244 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
1245 TL->setMetadata(LLVMContext::MD_tbaa, Tag);
1246 FL->setMetadata(LLVMContext::MD_tbaa, Tag);
1247 }
1248
1249 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1250 LI->getName() + ".sroa.speculated");
1251
1252 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1253 LI->replaceAllUsesWith(V);
1254 LI->eraseFromParent();
1255 }
1256 SI.eraseFromParent();
1257}
1258
1259/// \brief Build a GEP out of a base pointer and indices.
1260///
1261/// This will return the BasePtr if that is valid, or build a new GEP
1262/// instruction using the IRBuilder if GEP-ing is needed.
1263static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1264 SmallVectorImpl<Value *> &Indices, Twine NamePrefix) {
1265 if (Indices.empty())
1266 return BasePtr;
1267
1268 // A single zero index is a no-op, so check for this and avoid building a GEP
1269 // in that case.
1270 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1271 return BasePtr;
1272
1273 return IRB.CreateInBoundsGEP(BasePtr, Indices, NamePrefix + "sroa_idx");
1274}
1275
1276/// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1277/// TargetTy without changing the offset of the pointer.
1278///
1279/// This routine assumes we've already established a properly offset GEP with
1280/// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1281/// zero-indices down through type layers until we find one the same as
1282/// TargetTy. If we can't find one with the same type, we at least try to use
1283/// one with the same size. If none of that works, we just produce the GEP as
1284/// indicated by Indices to have the correct offset.
1285static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1286 Value *BasePtr, Type *Ty, Type *TargetTy,
1287 SmallVectorImpl<Value *> &Indices,
1288 Twine NamePrefix) {
1289 if (Ty == TargetTy)
1290 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1291
1292 // Pointer size to use for the indices.
1293 unsigned PtrSize = DL.getPointerTypeSizeInBits(BasePtr->getType());
1294
1295 // See if we can descend into a struct and locate a field with the correct
1296 // type.
1297 unsigned NumLayers = 0;
1298 Type *ElementTy = Ty;
1299 do {
1300 if (ElementTy->isPointerTy())
1301 break;
1302
1303 if (ArrayType *ArrayTy = dyn_cast<ArrayType>(ElementTy)) {
1304 ElementTy = ArrayTy->getElementType();
1305 Indices.push_back(IRB.getIntN(PtrSize, 0));
1306 } else if (VectorType *VectorTy = dyn_cast<VectorType>(ElementTy)) {
1307 ElementTy = VectorTy->getElementType();
1308 Indices.push_back(IRB.getInt32(0));
1309 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1310 if (STy->element_begin() == STy->element_end())
1311 break; // Nothing left to descend into.
1312 ElementTy = *STy->element_begin();
1313 Indices.push_back(IRB.getInt32(0));
1314 } else {
1315 break;
1316 }
1317 ++NumLayers;
1318 } while (ElementTy != TargetTy);
1319 if (ElementTy != TargetTy)
1320 Indices.erase(Indices.end() - NumLayers, Indices.end());
1321
1322 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1323}
1324
1325/// \brief Recursively compute indices for a natural GEP.
1326///
1327/// This is the recursive step for getNaturalGEPWithOffset that walks down the
1328/// element types adding appropriate indices for the GEP.
1329static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1330 Value *Ptr, Type *Ty, APInt &Offset,
1331 Type *TargetTy,
1332 SmallVectorImpl<Value *> &Indices,
1333 Twine NamePrefix) {
1334 if (Offset == 0)
1335 return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices, NamePrefix);
1336
1337 // We can't recurse through pointer types.
1338 if (Ty->isPointerTy())
1339 return nullptr;
1340
1341 // We try to analyze GEPs over vectors here, but note that these GEPs are
1342 // extremely poorly defined currently. The long-term goal is to remove GEPing
1343 // over a vector from the IR completely.
1344 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1345 unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
1346 if (ElementSizeInBits % 8 != 0) {
1347 // GEPs over non-multiple of 8 size vector elements are invalid.
1348 return nullptr;
1349 }
1350 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1351 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1352 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1353 return nullptr;
1354 Offset -= NumSkippedElements * ElementSize;
1355 Indices.push_back(IRB.getInt(NumSkippedElements));
1356 return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1357 Offset, TargetTy, Indices, NamePrefix);
1358 }
1359
1360 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1361 Type *ElementTy = ArrTy->getElementType();
1362 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1363 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1364 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1365 return nullptr;
1366
1367 Offset -= NumSkippedElements * ElementSize;
1368 Indices.push_back(IRB.getInt(NumSkippedElements));
1369 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1370 Indices, NamePrefix);
1371 }
1372
1373 StructType *STy = dyn_cast<StructType>(Ty);
1374 if (!STy)
1375 return nullptr;
1376
1377 const StructLayout *SL = DL.getStructLayout(STy);
1378 uint64_t StructOffset = Offset.getZExtValue();
1379 if (StructOffset >= SL->getSizeInBytes())
1380 return nullptr;
1381 unsigned Index = SL->getElementContainingOffset(StructOffset);
1382 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1383 Type *ElementTy = STy->getElementType(Index);
1384 if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
1385 return nullptr; // The offset points into alignment padding.
1386
1387 Indices.push_back(IRB.getInt32(Index));
1388 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1389 Indices, NamePrefix);
1390}
1391
1392/// \brief Get a natural GEP from a base pointer to a particular offset and
1393/// resulting in a particular type.
1394///
1395/// The goal is to produce a "natural" looking GEP that works with the existing
1396/// composite types to arrive at the appropriate offset and element type for
1397/// a pointer. TargetTy is the element type the returned GEP should point-to if
1398/// possible. We recurse by decreasing Offset, adding the appropriate index to
1399/// Indices, and setting Ty to the result subtype.
1400///
1401/// If no natural GEP can be constructed, this function returns null.
1402static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1403 Value *Ptr, APInt Offset, Type *TargetTy,
1404 SmallVectorImpl<Value *> &Indices,
1405 Twine NamePrefix) {
1406 PointerType *Ty = cast<PointerType>(Ptr->getType());
1407
1408 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1409 // an i8.
1410 if (Ty == IRB.getInt8PtrTy(Ty->getAddressSpace()) && TargetTy->isIntegerTy(8))
1411 return nullptr;
1412
1413 Type *ElementTy = Ty->getElementType();
1414 if (!ElementTy->isSized())
1415 return nullptr; // We can't GEP through an unsized element.
1416 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1417 if (ElementSize == 0)
1418 return nullptr; // Zero-length arrays can't help us build a natural GEP.
1419 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1420
1421 Offset -= NumSkippedElements * ElementSize;
1422 Indices.push_back(IRB.getInt(NumSkippedElements));
1423 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1424 Indices, NamePrefix);
1425}
1426
1427/// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1428/// resulting pointer has PointerTy.
1429///
1430/// This tries very hard to compute a "natural" GEP which arrives at the offset
1431/// and produces the pointer type desired. Where it cannot, it will try to use
1432/// the natural GEP to arrive at the offset and bitcast to the type. Where that
1433/// fails, it will try to use an existing i8* and GEP to the byte offset and
1434/// bitcast to the type.
1435///
1436/// The strategy for finding the more natural GEPs is to peel off layers of the
1437/// pointer, walking back through bit casts and GEPs, searching for a base
1438/// pointer from which we can compute a natural GEP with the desired
1439/// properties. The algorithm tries to fold as many constant indices into
1440/// a single GEP as possible, thus making each GEP more independent of the
1441/// surrounding code.
1442static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr,
1443 APInt Offset, Type *PointerTy,
1444 Twine NamePrefix) {
1445 // Even though we don't look through PHI nodes, we could be called on an
1446 // instruction in an unreachable block, which may be on a cycle.
1447 SmallPtrSet<Value *, 4> Visited;
1448 Visited.insert(Ptr);
1449 SmallVector<Value *, 4> Indices;
1450
1451 // We may end up computing an offset pointer that has the wrong type. If we
1452 // never are able to compute one directly that has the correct type, we'll
1453 // fall back to it, so keep it around here.
1454 Value *OffsetPtr = nullptr;
1455
1456 // Remember any i8 pointer we come across to re-use if we need to do a raw
1457 // byte offset.
1458 Value *Int8Ptr = nullptr;
1459 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1460
1461 Type *TargetTy = PointerTy->getPointerElementType();
1462
1463 do {
1464 // First fold any existing GEPs into the offset.
1465 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1466 APInt GEPOffset(Offset.getBitWidth(), 0);
1467 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1468 break;
1469 Offset += GEPOffset;
1470 Ptr = GEP->getPointerOperand();
1471 if (!Visited.insert(Ptr))
1472 break;
1473 }
1474
1475 // See if we can perform a natural GEP here.
1476 Indices.clear();
1477 if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1478 Indices, NamePrefix)) {
1479 if (P->getType() == PointerTy) {
1480 // Zap any offset pointer that we ended up computing in previous rounds.
1481 if (OffsetPtr && OffsetPtr->use_empty())
1482 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1483 I->eraseFromParent();
1484 return P;
1485 }
1486 if (!OffsetPtr) {
1487 OffsetPtr = P;
1488 }
1489 }
1490
1491 // Stash this pointer if we've found an i8*.
1492 if (Ptr->getType()->isIntegerTy(8)) {
1493 Int8Ptr = Ptr;
1494 Int8PtrOffset = Offset;
1495 }
1496
1497 // Peel off a layer of the pointer and update the offset appropriately.
1498 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1499 Ptr = cast<Operator>(Ptr)->getOperand(0);
1500 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1501 if (GA->mayBeOverridden())
1502 break;
1503 Ptr = GA->getAliasee();
1504 } else {
1505 break;
1506 }
1507 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1508 } while (Visited.insert(Ptr));
1509
1510 if (!OffsetPtr) {
1511 if (!Int8Ptr) {
1512 Int8Ptr = IRB.CreateBitCast(
1513 Ptr, IRB.getInt8PtrTy(PointerTy->getPointerAddressSpace()),
1514 NamePrefix + "sroa_raw_cast");
1515 Int8PtrOffset = Offset;
1516 }
1517
1518 OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
1519 IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1520 NamePrefix + "sroa_raw_idx");
1521 }
1522 Ptr = OffsetPtr;
1523
1524 // On the off chance we were targeting i8*, guard the bitcast here.
1525 if (Ptr->getType() != PointerTy)
1526 Ptr = IRB.CreateBitCast(Ptr, PointerTy, NamePrefix + "sroa_cast");
1527
1528 return Ptr;
1529}
1530
1531/// \brief Test whether we can convert a value from the old to the new type.
1532///
1533/// This predicate should be used to guard calls to convertValue in order to
1534/// ensure that we only try to convert viable values. The strategy is that we
1535/// will peel off single element struct and array wrappings to get to an
1536/// underlying value, and convert that value.
1537static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1538 if (OldTy == NewTy)
1539 return true;
1540 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1541 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1542 if (NewITy->getBitWidth() >= OldITy->getBitWidth())
1543 return true;
1544 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1545 return false;
1546 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1547 return false;
1548
1549 // We can convert pointers to integers and vice-versa. Same for vectors
1550 // of pointers and integers.
1551 OldTy = OldTy->getScalarType();
1552 NewTy = NewTy->getScalarType();
1553 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1554 if (NewTy->isPointerTy() && OldTy->isPointerTy())
1555 return true;
1556 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
1557 return true;
1558 return false;
1559 }
1560
1561 return true;
1562}
1563
1564/// \brief Generic routine to convert an SSA value to a value of a different
1565/// type.
1566///
1567/// This will try various different casting techniques, such as bitcasts,
1568/// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1569/// two types for viability with this routine.
1570static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1571 Type *NewTy) {
1572 Type *OldTy = V->getType();
1573 assert(canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type");
1574
1575 if (OldTy == NewTy)
1576 return V;
1577
1578 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1579 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1580 if (NewITy->getBitWidth() > OldITy->getBitWidth())
1581 return IRB.CreateZExt(V, NewITy);
1582
1583 // See if we need inttoptr for this type pair. A cast involving both scalars
1584 // and vectors requires and additional bitcast.
1585 if (OldTy->getScalarType()->isIntegerTy() &&
1586 NewTy->getScalarType()->isPointerTy()) {
1587 // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1588 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1589 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1590 NewTy);
1591
1592 // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1593 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1594 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1595 NewTy);
1596
1597 return IRB.CreateIntToPtr(V, NewTy);
1598 }
1599
1600 // See if we need ptrtoint for this type pair. A cast involving both scalars
1601 // and vectors requires and additional bitcast.
1602 if (OldTy->getScalarType()->isPointerTy() &&
1603 NewTy->getScalarType()->isIntegerTy()) {
1604 // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1605 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1606 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1607 NewTy);
1608
1609 // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1610 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1611 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1612 NewTy);
1613
1614 return IRB.CreatePtrToInt(V, NewTy);
1615 }
1616
1617 return IRB.CreateBitCast(V, NewTy);
1618}
1619
1620/// \brief Test whether the given slice use can be promoted to a vector.
1621///
1622/// This function is called to test each entry in a partioning which is slated
1623/// for a single slice.
1624static bool isVectorPromotionViableForSlice(
1625 const DataLayout &DL, AllocaSlices &S, uint64_t SliceBeginOffset,
1626 uint64_t SliceEndOffset, VectorType *Ty, uint64_t ElementSize,
1627 AllocaSlices::const_iterator I) {
1628 // First validate the slice offsets.
1629 uint64_t BeginOffset =
1630 std::max(I->beginOffset(), SliceBeginOffset) - SliceBeginOffset;
1631 uint64_t BeginIndex = BeginOffset / ElementSize;
1632 if (BeginIndex * ElementSize != BeginOffset ||
1633 BeginIndex >= Ty->getNumElements())
1634 return false;
1635 uint64_t EndOffset =
1636 std::min(I->endOffset(), SliceEndOffset) - SliceBeginOffset;
1637 uint64_t EndIndex = EndOffset / ElementSize;
1638 if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
1639 return false;
1640
1641 assert(EndIndex > BeginIndex && "Empty vector!");
1642 uint64_t NumElements = EndIndex - BeginIndex;
1643 Type *SliceTy =
1644 (NumElements == 1) ? Ty->getElementType()
1645 : VectorType::get(Ty->getElementType(), NumElements);
1646
1647 Type *SplitIntTy =
1648 Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1649
1650 Use *U = I->getUse();
1651
1652 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1653 if (MI->isVolatile())
1654 return false;
1655 if (!I->isSplittable())
1656 return false; // Skip any unsplittable intrinsics.
1657 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1658 // Disable vector promotion when there are loads or stores of an FCA.
1659 return false;
1660 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1661 if (LI->isVolatile())
1662 return false;
1663 Type *LTy = LI->getType();
1664 if (SliceBeginOffset > I->beginOffset() ||
1665 SliceEndOffset < I->endOffset()) {
1666 assert(LTy->isIntegerTy());
1667 LTy = SplitIntTy;
1668 }
1669 if (!canConvertValue(DL, SliceTy, LTy))
1670 return false;
1671 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1672 if (SI->isVolatile())
1673 return false;
1674 Type *STy = SI->getValueOperand()->getType();
1675 if (SliceBeginOffset > I->beginOffset() ||
1676 SliceEndOffset < I->endOffset()) {
1677 assert(STy->isIntegerTy());
1678 STy = SplitIntTy;
1679 }
1680 if (!canConvertValue(DL, STy, SliceTy))
1681 return false;
1682 } else {
1683 return false;
1684 }
1685
1686 return true;
1687}
1688
1689/// \brief Test whether the given alloca partitioning and range of slices can be
1690/// promoted to a vector.
1691///
1692/// This is a quick test to check whether we can rewrite a particular alloca
1693/// partition (and its newly formed alloca) into a vector alloca with only
1694/// whole-vector loads and stores such that it could be promoted to a vector
1695/// SSA value. We only can ensure this for a limited set of operations, and we
1696/// don't want to do the rewrites unless we are confident that the result will
1697/// be promotable, so we have an early test here.
1698static bool
1699isVectorPromotionViable(const DataLayout &DL, Type *AllocaTy, AllocaSlices &S,
1700 uint64_t SliceBeginOffset, uint64_t SliceEndOffset,
1701 AllocaSlices::const_iterator I,
1702 AllocaSlices::const_iterator E,
1703 ArrayRef<AllocaSlices::iterator> SplitUses) {
1704 VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
1705 if (!Ty)
1706 return false;
1707
1708 uint64_t ElementSize = DL.getTypeSizeInBits(Ty->getScalarType());
1709
1710 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1711 // that aren't byte sized.
1712 if (ElementSize % 8)
1713 return false;
1714 assert((DL.getTypeSizeInBits(Ty) % 8) == 0 &&
1715 "vector size not a multiple of element size?");
1716 ElementSize /= 8;
1717
1718 for (; I != E; ++I)
1719 if (!isVectorPromotionViableForSlice(DL, S, SliceBeginOffset,
1720 SliceEndOffset, Ty, ElementSize, I))
1721 return false;
1722
1723 for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
1724 SUE = SplitUses.end();
1725 SUI != SUE; ++SUI)
1726 if (!isVectorPromotionViableForSlice(DL, S, SliceBeginOffset,
1727 SliceEndOffset, Ty, ElementSize, *SUI))
1728 return false;
1729
1730 return true;
1731}
1732
1733/// \brief Test whether a slice of an alloca is valid for integer widening.
1734///
1735/// This implements the necessary checking for the \c isIntegerWideningViable
1736/// test below on a single slice of the alloca.
1737static bool isIntegerWideningViableForSlice(const DataLayout &DL,
1738 Type *AllocaTy,
1739 uint64_t AllocBeginOffset,
1740 uint64_t Size, AllocaSlices &S,
1741 AllocaSlices::const_iterator I,
1742 bool &WholeAllocaOp) {
1743 uint64_t RelBegin = I->beginOffset() - AllocBeginOffset;
1744 uint64_t RelEnd = I->endOffset() - AllocBeginOffset;
1745
1746 // We can't reasonably handle cases where the load or store extends past
1747 // the end of the aloca's type and into its padding.
1748 if (RelEnd > Size)
1749 return false;
1750
1751 Use *U = I->getUse();
1752
1753 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1754 if (LI->isVolatile())
1755 return false;
1756 if (RelBegin == 0 && RelEnd == Size)
1757 WholeAllocaOp = true;
1758 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
1759 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1760 return false;
1761 } else if (RelBegin != 0 || RelEnd != Size ||
1762 !canConvertValue(DL, AllocaTy, LI->getType())) {
1763 // Non-integer loads need to be convertible from the alloca type so that
1764 // they are promotable.
1765 return false;
1766 }
1767 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1768 Type *ValueTy = SI->getValueOperand()->getType();
1769 if (SI->isVolatile())
1770 return false;
1771 if (RelBegin == 0 && RelEnd == Size)
1772 WholeAllocaOp = true;
1773 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
1774 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1775 return false;
1776 } else if (RelBegin != 0 || RelEnd != Size ||
1777 !canConvertValue(DL, ValueTy, AllocaTy)) {
1778 // Non-integer stores need to be convertible to the alloca type so that
1779 // they are promotable.
1780 return false;
1781 }
1782 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1783 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
1784 return false;
1785 if (!I->isSplittable())
1786 return false; // Skip any unsplittable intrinsics.
1787 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1788 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1789 II->getIntrinsicID() != Intrinsic::lifetime_end)
1790 return false;
1791 } else {
1792 return false;
1793 }
1794
1795 return true;
1796}
1797
1798/// \brief Test whether the given alloca partition's integer operations can be
1799/// widened to promotable ones.
1800///
1801/// This is a quick test to check whether we can rewrite the integer loads and
1802/// stores to a particular alloca into wider loads and stores and be able to
1803/// promote the resulting alloca.
1804static bool
1805isIntegerWideningViable(const DataLayout &DL, Type *AllocaTy,
1806 uint64_t AllocBeginOffset, AllocaSlices &S,
1807 AllocaSlices::const_iterator I,
1808 AllocaSlices::const_iterator E,
1809 ArrayRef<AllocaSlices::iterator> SplitUses) {
1810 uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
1811 // Don't create integer types larger than the maximum bitwidth.
1812 if (SizeInBits > IntegerType::MAX_INT_BITS)
1813 return false;
1814
1815 // Don't try to handle allocas with bit-padding.
1816 if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
1817 return false;
1818
1819 // We need to ensure that an integer type with the appropriate bitwidth can
1820 // be converted to the alloca type, whatever that is. We don't want to force
1821 // the alloca itself to have an integer type if there is a more suitable one.
1822 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
1823 if (!canConvertValue(DL, AllocaTy, IntTy) ||
1824 !canConvertValue(DL, IntTy, AllocaTy))
1825 return false;
1826
1827 uint64_t Size = DL.getTypeStoreSize(AllocaTy);
1828
1829 // While examining uses, we ensure that the alloca has a covering load or
1830 // store. We don't want to widen the integer operations only to fail to
1831 // promote due to some other unsplittable entry (which we may make splittable
1832 // later). However, if there are only splittable uses, go ahead and assume
1833 // that we cover the alloca.
1834 bool WholeAllocaOp = (I != E) ? false : DL.isLegalInteger(SizeInBits);
1835
1836 for (; I != E; ++I)
1837 if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
1838 S, I, WholeAllocaOp))
1839 return false;
1840
1841 for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
1842 SUE = SplitUses.end();
1843 SUI != SUE; ++SUI)
1844 if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
1845 S, *SUI, WholeAllocaOp))
1846 return false;
1847
1848 return WholeAllocaOp;
1849}
1850
1851static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1852 IntegerType *Ty, uint64_t Offset,
1853 const Twine &Name) {
1854 DEBUG(dbgs() << " start: " << *V << "\n");
1855 IntegerType *IntTy = cast<IntegerType>(V->getType());
1856 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
1857 "Element extends past full value");
1858 uint64_t ShAmt = 8*Offset;
1859 if (DL.isBigEndian())
1860 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
1861 if (ShAmt) {
1862 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
1863 DEBUG(dbgs() << " shifted: " << *V << "\n");
1864 }
1865 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
1866 "Cannot extract to a larger integer!");
1867 if (Ty != IntTy) {
1868 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
1869 DEBUG(dbgs() << " trunced: " << *V << "\n");
1870 }
1871 return V;
1872}
1873
1874static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
1875 Value *V, uint64_t Offset, const Twine &Name) {
1876 IntegerType *IntTy = cast<IntegerType>(Old->getType());
1877 IntegerType *Ty = cast<IntegerType>(V->getType());
1878 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
1879 "Cannot insert a larger integer!");
1880 DEBUG(dbgs() << " start: " << *V << "\n");
1881 if (Ty != IntTy) {
1882 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
1883 DEBUG(dbgs() << " extended: " << *V << "\n");
1884 }
1885 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
1886 "Element store outside of alloca store");
1887 uint64_t ShAmt = 8*Offset;
1888 if (DL.isBigEndian())
1889 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
1890 if (ShAmt) {
1891 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
1892 DEBUG(dbgs() << " shifted: " << *V << "\n");
1893 }
1894
1895 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
1896 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
1897 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
1898 DEBUG(dbgs() << " masked: " << *Old << "\n");
1899 V = IRB.CreateOr(Old, V, Name + ".insert");
1900 DEBUG(dbgs() << " inserted: " << *V << "\n");
1901 }
1902 return V;
1903}
1904
1905static Value *extractVector(IRBuilderTy &IRB, Value *V,
1906 unsigned BeginIndex, unsigned EndIndex,
1907 const Twine &Name) {
1908 VectorType *VecTy = cast<VectorType>(V->getType());
1909 unsigned NumElements = EndIndex - BeginIndex;
1910 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
1911
1912 if (NumElements == VecTy->getNumElements())
1913 return V;
1914
1915 if (NumElements == 1) {
1916 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
1917 Name + ".extract");
1918 DEBUG(dbgs() << " extract: " << *V << "\n");
1919 return V;
1920 }
1921
1922 SmallVector<Constant*, 8> Mask;
1923 Mask.reserve(NumElements);
1924 for (unsigned i = BeginIndex; i != EndIndex; ++i)
1925 Mask.push_back(IRB.getInt32(i));
1926 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
1927 ConstantVector::get(Mask),
1928 Name + ".extract");
1929 DEBUG(dbgs() << " shuffle: " << *V << "\n");
1930 return V;
1931}
1932
1933static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
1934 unsigned BeginIndex, const Twine &Name) {
1935 VectorType *VecTy = cast<VectorType>(Old->getType());
1936 assert(VecTy && "Can only insert a vector into a vector");
1937
1938 VectorType *Ty = dyn_cast<VectorType>(V->getType());
1939 if (!Ty) {
1940 // Single element to insert.
1941 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
1942 Name + ".insert");
1943 DEBUG(dbgs() << " insert: " << *V << "\n");
1944 return V;
1945 }
1946
1947 assert(Ty->getNumElements() <= VecTy->getNumElements() &&
1948 "Too many elements!");
1949 if (Ty->getNumElements() == VecTy->getNumElements()) {
1950 assert(V->getType() == VecTy && "Vector type mismatch");
1951 return V;
1952 }
1953 unsigned EndIndex = BeginIndex + Ty->getNumElements();
1954
1955 // When inserting a smaller vector into the larger to store, we first
1956 // use a shuffle vector to widen it with undef elements, and then
1957 // a second shuffle vector to select between the loaded vector and the
1958 // incoming vector.
1959 SmallVector<Constant*, 8> Mask;
1960 Mask.reserve(VecTy->getNumElements());
1961 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
1962 if (i >= BeginIndex && i < EndIndex)
1963 Mask.push_back(IRB.getInt32(i - BeginIndex));
1964 else
1965 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
1966 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
1967 ConstantVector::get(Mask),
1968 Name + ".expand");
1969 DEBUG(dbgs() << " shuffle: " << *V << "\n");
1970
1971 Mask.clear();
1972 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
1973 Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
1974
1975 V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
1976
1977 DEBUG(dbgs() << " blend: " << *V << "\n");
1978 return V;
1979}
1980
1981namespace {
1982/// \brief Visitor to rewrite instructions using p particular slice of an alloca
1983/// to use a new alloca.
1984///
1985/// Also implements the rewriting to vector-based accesses when the partition
1986/// passes the isVectorPromotionViable predicate. Most of the rewriting logic
1987/// lives here.
1988class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
1989 // Befriend the base class so it can delegate to private visit methods.
1990 friend class llvm::InstVisitor<AllocaSliceRewriter, bool>;
1991 typedef llvm::InstVisitor<AllocaSliceRewriter, bool> Base;
1992
1993 const DataLayout &DL;
1994 AllocaSlices &S;
1995 SROA &Pass;
1996 AllocaInst &OldAI, &NewAI;
1997 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
1998 Type *NewAllocaTy;
1999
2000 // If we are rewriting an alloca partition which can be written as pure
2001 // vector operations, we stash extra information here. When VecTy is
2002 // non-null, we have some strict guarantees about the rewritten alloca:
2003 // - The new alloca is exactly the size of the vector type here.
2004 // - The accesses all either map to the entire vector or to a single
2005 // element.
2006 // - The set of accessing instructions is only one of those handled above
2007 // in isVectorPromotionViable. Generally these are the same access kinds
2008 // which are promotable via mem2reg.
2009 VectorType *VecTy;
2010 Type *ElementTy;
2011 uint64_t ElementSize;
2012
2013 // This is a convenience and flag variable that will be null unless the new
2014 // alloca's integer operations should be widened to this integer type due to
2015 // passing isIntegerWideningViable above. If it is non-null, the desired
2016 // integer type will be stored here for easy access during rewriting.
2017 IntegerType *IntTy;
2018
2019 // The original offset of the slice currently being rewritten relative to
2020 // the original alloca.
2021 uint64_t BeginOffset, EndOffset;
2022 // The new offsets of the slice currently being rewritten relative to the
2023 // original alloca.
2024 uint64_t NewBeginOffset, NewEndOffset;
2025
2026 uint64_t SliceSize;
2027 bool IsSplittable;
2028 bool IsSplit;
2029 Use *OldUse;
2030 Instruction *OldPtr;
2031
2032 // Track post-rewrite users which are PHI nodes and Selects.
2033 SmallPtrSetImpl<PHINode *> &PHIUsers;
2034 SmallPtrSetImpl<SelectInst *> &SelectUsers;
2035
2036 // Utility IR builder, whose name prefix is setup for each visited use, and
2037 // the insertion point is set to point to the user.
2038 IRBuilderTy IRB;
2039
2040public:
2041 AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &S, SROA &Pass,
2042 AllocaInst &OldAI, AllocaInst &NewAI,
2043 uint64_t NewAllocaBeginOffset,
2044 uint64_t NewAllocaEndOffset, bool IsVectorPromotable,
2045 bool IsIntegerPromotable,
2046 SmallPtrSetImpl<PHINode *> &PHIUsers,
2047 SmallPtrSetImpl<SelectInst *> &SelectUsers)
2048 : DL(DL), S(S), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
2049 NewAllocaBeginOffset(NewAllocaBeginOffset),
2050 NewAllocaEndOffset(NewAllocaEndOffset),
2051 NewAllocaTy(NewAI.getAllocatedType()),
2052 VecTy(IsVectorPromotable ? cast<VectorType>(NewAllocaTy) : nullptr),
2053 ElementTy(VecTy ? VecTy->getElementType() : nullptr),
2054 ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
2055 IntTy(IsIntegerPromotable
2056 ? Type::getIntNTy(
2057 NewAI.getContext(),
2058 DL.getTypeSizeInBits(NewAI.getAllocatedType()))
2059 : nullptr),
2060 BeginOffset(), EndOffset(), IsSplittable(), IsSplit(), OldUse(),
2061 OldPtr(), PHIUsers(PHIUsers), SelectUsers(SelectUsers),
2062 IRB(NewAI.getContext(), ConstantFolder()) {
2063 if (VecTy) {
2064 assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
2065 "Only multiple-of-8 sized vector elements are viable");
2066 ++NumVectorized;
2067 }
2068 assert((!IsVectorPromotable && !IsIntegerPromotable) ||
2069 IsVectorPromotable != IsIntegerPromotable);
2070 }
2071
2072 bool visit(AllocaSlices::const_iterator I) {
2073 bool CanSROA = true;
2074 BeginOffset = I->beginOffset();
2075 EndOffset = I->endOffset();
2076 IsSplittable = I->isSplittable();
2077 IsSplit =
2078 BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
2079
2080 // Compute the intersecting offset range.
2081 assert(BeginOffset < NewAllocaEndOffset);
2082 assert(EndOffset > NewAllocaBeginOffset);
2083 NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2084 NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2085
2086 SliceSize = NewEndOffset - NewBeginOffset;
2087
2088 OldUse = I->getUse();
2089 OldPtr = cast<Instruction>(OldUse->get());
2090
2091 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2092 IRB.SetInsertPoint(OldUserI);
2093 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2094 IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
2095
2096 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2097 if (VecTy || IntTy)
2098 assert(CanSROA);
2099 return CanSROA;
2100 }
2101
2102private:
2103 // Make sure the other visit overloads are visible.
2104 using Base::visit;
2105
2106 // Every instruction which can end up as a user must have a rewrite rule.
2107 bool visitInstruction(Instruction &I) {
2108 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2109 llvm_unreachable("No rewrite rule for this instruction!");
2110 }
2111
2112 Value *getNewAllocaSlicePtr(IRBuilderTy &IRB, Type *PointerTy) {
2113 // Note that the offset computation can use BeginOffset or NewBeginOffset
2114 // interchangeably for unsplit slices.
2115 assert(IsSplit || BeginOffset == NewBeginOffset);
2116 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2117
2118#ifndef NDEBUG
2119 StringRef OldName = OldPtr->getName();
2120 // Skip through the last '.sroa.' component of the name.
2121 size_t LastSROAPrefix = OldName.rfind(".sroa.");
2122 if (LastSROAPrefix != StringRef::npos) {
2123 OldName = OldName.substr(LastSROAPrefix + strlen(".sroa."));
2124 // Look for an SROA slice index.
2125 size_t IndexEnd = OldName.find_first_not_of("0123456789");
2126 if (IndexEnd != StringRef::npos && OldName[IndexEnd] == '.') {
2127 // Strip the index and look for the offset.
2128 OldName = OldName.substr(IndexEnd + 1);
2129 size_t OffsetEnd = OldName.find_first_not_of("0123456789");
2130 if (OffsetEnd != StringRef::npos && OldName[OffsetEnd] == '.')
2131 // Strip the offset.
2132 OldName = OldName.substr(OffsetEnd + 1);
2133 }
2134 }
2135 // Strip any SROA suffixes as well.
2136 OldName = OldName.substr(0, OldName.find(".sroa_"));
2137#endif
2138
2139 return getAdjustedPtr(IRB, DL, &NewAI,
2140 APInt(DL.getPointerSizeInBits(), Offset), PointerTy,
2141#ifndef NDEBUG
2142 Twine(OldName) + "."
2143#else
2144 Twine()
2145#endif
2146 );
2147 }
2148
2149 /// \brief Compute suitable alignment to access this slice of the *new* alloca.
2150 ///
2151 /// You can optionally pass a type to this routine and if that type's ABI
2152 /// alignment is itself suitable, this will return zero.
2153 unsigned getSliceAlign(Type *Ty = nullptr) {
2154 unsigned NewAIAlign = NewAI.getAlignment();
2155 if (!NewAIAlign)
2156 NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
2157 unsigned Align = MinAlign(NewAIAlign, NewBeginOffset - NewAllocaBeginOffset);
2158 return (Ty && Align == DL.getABITypeAlignment(Ty)) ? 0 : Align;
2159 }
2160
2161 unsigned getIndex(uint64_t Offset) {
2162 assert(VecTy && "Can only call getIndex when rewriting a vector");
2163 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2164 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2165 uint32_t Index = RelOffset / ElementSize;
2166 assert(Index * ElementSize == RelOffset);
2167 return Index;
2168 }
2169
2170 void deleteIfTriviallyDead(Value *V) {
2171 Instruction *I = cast<Instruction>(V);
2172 if (isInstructionTriviallyDead(I))
2173 Pass.DeadInsts.insert(I);
2174 }
2175
2176 Value *rewriteVectorizedLoadInst() {
2177 unsigned BeginIndex = getIndex(NewBeginOffset);
2178 unsigned EndIndex = getIndex(NewEndOffset);
2179 assert(EndIndex > BeginIndex && "Empty vector!");
2180
2181 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2182 "load");
2183 return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2184 }
2185
2186 Value *rewriteIntegerLoad(LoadInst &LI) {
2187 assert(IntTy && "We cannot insert an integer to the alloca");
2188 assert(!LI.isVolatile());
2189 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2190 "load");
2191 V = convertValue(DL, IRB, V, IntTy);
2192 assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2193 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2194 if (Offset > 0 || NewEndOffset < NewAllocaEndOffset)
2195 V = extractInteger(DL, IRB, V, cast<IntegerType>(LI.getType()), Offset,
2196 "extract");
2197 return V;
2198 }
2199
2200 bool visitLoadInst(LoadInst &LI) {
2201 DEBUG(dbgs() << " original: " << LI << "\n");
2202 Value *OldOp = LI.getOperand(0);
2203 assert(OldOp == OldPtr);
2204
2205 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), SliceSize * 8)
2206 : LI.getType();
2207 bool IsPtrAdjusted = false;
2208 Value *V;
2209 if (VecTy) {
2210 V = rewriteVectorizedLoadInst();
2211 } else if (IntTy && LI.getType()->isIntegerTy()) {
2212 V = rewriteIntegerLoad(LI);
2213 } else if (NewBeginOffset == NewAllocaBeginOffset &&
2214 canConvertValue(DL, NewAllocaTy, LI.getType())) {
2215 V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2216 LI.isVolatile(), LI.getName());
2217 } else {
2218 Type *LTy = TargetTy->getPointerTo();
2219 V = IRB.CreateAlignedLoad(getNewAllocaSlicePtr(IRB, LTy),
2220 getSliceAlign(TargetTy), LI.isVolatile(),
2221 LI.getName());
2222 IsPtrAdjusted = true;
2223 }
2224 V = convertValue(DL, IRB, V, TargetTy);
2225
2226 if (IsSplit) {
2227 assert(!LI.isVolatile());
2228 assert(LI.getType()->isIntegerTy() &&
2229 "Only integer type loads and stores are split");
2230 assert(SliceSize < DL.getTypeStoreSize(LI.getType()) &&
2231 "Split load isn't smaller than original load");
2232 assert(LI.getType()->getIntegerBitWidth() ==
2233 DL.getTypeStoreSizeInBits(LI.getType()) &&
2234 "Non-byte-multiple bit width");
2235 // Move the insertion point just past the load so that we can refer to it.
2236 IRB.SetInsertPoint(std::next(BasicBlock::iterator(&LI)));
2237 // Create a placeholder value with the same type as LI to use as the
2238 // basis for the new value. This allows us to replace the uses of LI with
2239 // the computed value, and then replace the placeholder with LI, leaving
2240 // LI only used for this computation.
2241 Value *Placeholder
2242 = new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
2243 V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset,
2244 "insert");
2245 LI.replaceAllUsesWith(V);
2246 Placeholder->replaceAllUsesWith(&LI);
2247 delete Placeholder;
2248 } else {
2249 LI.replaceAllUsesWith(V);
2250 }
2251
2252 Pass.DeadInsts.insert(&LI);
2253 deleteIfTriviallyDead(OldOp);
2254 DEBUG(dbgs() << " to: " << *V << "\n");
2255 return !LI.isVolatile() && !IsPtrAdjusted;
2256 }
2257
2258 bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp) {
2259 if (V->getType() != VecTy) {
2260 unsigned BeginIndex = getIndex(NewBeginOffset);
2261 unsigned EndIndex = getIndex(NewEndOffset);
2262 assert(EndIndex > BeginIndex && "Empty vector!");
2263 unsigned NumElements = EndIndex - BeginIndex;
2264 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2265 Type *SliceTy =
2266 (NumElements == 1) ? ElementTy
2267 : VectorType::get(ElementTy, NumElements);
2268 if (V->getType() != SliceTy)
2269 V = convertValue(DL, IRB, V, SliceTy);
2270
2271 // Mix in the existing elements.
2272 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2273 "load");
2274 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2275 }
2276 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2277 Pass.DeadInsts.insert(&SI);
2278
2279 (void)Store;
2280 DEBUG(dbgs() << " to: " << *Store << "\n");
2281 return true;
2282 }
2283
2284 bool rewriteIntegerStore(Value *V, StoreInst &SI) {
2285 assert(IntTy && "We cannot extract an integer from the alloca");
2286 assert(!SI.isVolatile());
2287 if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2288 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2289 "oldload");
2290 Old = convertValue(DL, IRB, Old, IntTy);
2291 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2292 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2293 V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset,
2294 "insert");
2295 }
2296 V = convertValue(DL, IRB, V, NewAllocaTy);
2297 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2298 Pass.DeadInsts.insert(&SI);
2299 (void)Store;
2300 DEBUG(dbgs() << " to: " << *Store << "\n");
2301 return true;
2302 }
2303
2304 bool visitStoreInst(StoreInst &SI) {
2305 DEBUG(dbgs() << " original: " << SI << "\n");
2306 Value *OldOp = SI.getOperand(1);
2307 assert(OldOp == OldPtr);
2308
2309 Value *V = SI.getValueOperand();
2310
2311 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2312 // alloca that should be re-examined after promoting this alloca.
2313 if (V->getType()->isPointerTy())
2314 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2315 Pass.PostPromotionWorklist.insert(AI);
2316
2317 if (SliceSize < DL.getTypeStoreSize(V->getType())) {
2318 assert(!SI.isVolatile());
2319 assert(V->getType()->isIntegerTy() &&
2320 "Only integer type loads and stores are split");
2321 assert(V->getType()->getIntegerBitWidth() ==
2322 DL.getTypeStoreSizeInBits(V->getType()) &&
2323 "Non-byte-multiple bit width");
2324 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), SliceSize * 8);
2325 V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset,
2326 "extract");
2327 }
2328
2329 if (VecTy)
2330 return rewriteVectorizedStoreInst(V, SI, OldOp);
2331 if (IntTy && V->getType()->isIntegerTy())
2332 return rewriteIntegerStore(V, SI);
2333
2334 StoreInst *NewSI;
2335 if (NewBeginOffset == NewAllocaBeginOffset &&
2336 NewEndOffset == NewAllocaEndOffset &&
2337 canConvertValue(DL, V->getType(), NewAllocaTy)) {
2338 V = convertValue(DL, IRB, V, NewAllocaTy);
2339 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2340 SI.isVolatile());
2341 } else {
2342 Value *NewPtr = getNewAllocaSlicePtr(IRB, V->getType()->getPointerTo());
2343 NewSI = IRB.CreateAlignedStore(V, NewPtr, getSliceAlign(V->getType()),
2344 SI.isVolatile());
2345 }
2346 (void)NewSI;
2347 Pass.DeadInsts.insert(&SI);
2348 deleteIfTriviallyDead(OldOp);
2349
2350 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2351 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2352 }
2353
2354 /// \brief Compute an integer value from splatting an i8 across the given
2355 /// number of bytes.
2356 ///
2357 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2358 /// call this routine.
2359 /// FIXME: Heed the advice above.
2360 ///
2361 /// \param V The i8 value to splat.
2362 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2363 Value *getIntegerSplat(Value *V, unsigned Size) {
2364 assert(Size > 0 && "Expected a positive number of bytes.");
2365 IntegerType *VTy = cast<IntegerType>(V->getType());
2366 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2367 if (Size == 1)
2368 return V;
2369
2370 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size*8);
2371 V = IRB.CreateMul(IRB.CreateZExt(V, SplatIntTy, "zext"),
2372 ConstantExpr::getUDiv(
2373 Constant::getAllOnesValue(SplatIntTy),
2374 ConstantExpr::getZExt(
2375 Constant::getAllOnesValue(V->getType()),
2376 SplatIntTy)),
2377 "isplat");
2378 return V;
2379 }
2380
2381 /// \brief Compute a vector splat for a given element value.
2382 Value *getVectorSplat(Value *V, unsigned NumElements) {
2383 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2384 DEBUG(dbgs() << " splat: " << *V << "\n");
2385 return V;
2386 }
2387
2388 bool visitMemSetInst(MemSetInst &II) {
2389 DEBUG(dbgs() << " original: " << II << "\n");
2390 assert(II.getRawDest() == OldPtr);
2391
2392 // If the memset has a variable size, it cannot be split, just adjust the
2393 // pointer to the new alloca.
2394 if (!isa<Constant>(II.getLength())) {
2395 assert(!IsSplit);
2396 assert(NewBeginOffset == BeginOffset);
2397 II.setDest(getNewAllocaSlicePtr(IRB, OldPtr->getType()));
2398 Type *CstTy = II.getAlignmentCst()->getType();
2399 II.setAlignment(ConstantInt::get(CstTy, getSliceAlign()));
2400
2401 deleteIfTriviallyDead(OldPtr);
2402 return false;
2403 }
2404
2405 // Record this instruction for deletion.
2406 Pass.DeadInsts.insert(&II);
2407
2408 Type *AllocaTy = NewAI.getAllocatedType();
2409 Type *ScalarTy = AllocaTy->getScalarType();
2410
2411 // If this doesn't map cleanly onto the alloca type, and that type isn't
2412 // a single value type, just emit a memset.
2413 if (!VecTy && !IntTy &&
2414 (BeginOffset > NewAllocaBeginOffset ||
2415 EndOffset < NewAllocaEndOffset ||
2416 !AllocaTy->isSingleValueType() ||
2417 !DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy)) ||
2418 DL.getTypeSizeInBits(ScalarTy)%8 != 0)) {
2419 Type *SizeTy = II.getLength()->getType();
2420 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2421 CallInst *New = IRB.CreateMemSet(
2422 getNewAllocaSlicePtr(IRB, OldPtr->getType()), II.getValue(), Size,
2423 getSliceAlign(), II.isVolatile());
2424 (void)New;
2425 DEBUG(dbgs() << " to: " << *New << "\n");
2426 return false;
2427 }
2428
2429 // If we can represent this as a simple value, we have to build the actual
2430 // value to store, which requires expanding the byte present in memset to
2431 // a sensible representation for the alloca type. This is essentially
2432 // splatting the byte to a sufficiently wide integer, splatting it across
2433 // any desired vector width, and bitcasting to the final type.
2434 Value *V;
2435
2436 if (VecTy) {
2437 // If this is a memset of a vectorized alloca, insert it.
2438 assert(ElementTy == ScalarTy);
2439
2440 unsigned BeginIndex = getIndex(NewBeginOffset);
2441 unsigned EndIndex = getIndex(NewEndOffset);
2442 assert(EndIndex > BeginIndex && "Empty vector!");
2443 unsigned NumElements = EndIndex - BeginIndex;
2444 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2445
2446 Value *Splat =
2447 getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
2448 Splat = convertValue(DL, IRB, Splat, ElementTy);
2449 if (NumElements > 1)
2450 Splat = getVectorSplat(Splat, NumElements);
2451
2452 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2453 "oldload");
2454 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2455 } else if (IntTy) {
2456 // If this is a memset on an alloca where we can widen stores, insert the
2457 // set integer.
2458 assert(!II.isVolatile());
2459
2460 uint64_t Size = NewEndOffset - NewBeginOffset;
2461 V = getIntegerSplat(II.getValue(), Size);
2462
2463 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2464 EndOffset != NewAllocaBeginOffset)) {
2465 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2466 "oldload");
2467 Old = convertValue(DL, IRB, Old, IntTy);
2468 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2469 V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2470 } else {
2471 assert(V->getType() == IntTy &&
2472 "Wrong type for an alloca wide integer!");
2473 }
2474 V = convertValue(DL, IRB, V, AllocaTy);
2475 } else {
2476 // Established these invariants above.
2477 assert(NewBeginOffset == NewAllocaBeginOffset);
2478 assert(NewEndOffset == NewAllocaEndOffset);
2479
2480 V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
2481 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2482 V = getVectorSplat(V, AllocaVecTy->getNumElements());
2483
2484 V = convertValue(DL, IRB, V, AllocaTy);
2485 }
2486
2487 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2488 II.isVolatile());
2489 (void)New;
2490 DEBUG(dbgs() << " to: " << *New << "\n");
2491 return !II.isVolatile();
2492 }
2493
2494 bool visitMemTransferInst(MemTransferInst &II) {
2495 // Rewriting of memory transfer instructions can be a bit tricky. We break
2496 // them into two categories: split intrinsics and unsplit intrinsics.
2497
2498 DEBUG(dbgs() << " original: " << II << "\n");
2499
2500 bool IsDest = &II.getRawDestUse() == OldUse;
2501 assert((IsDest && II.getRawDest() == OldPtr) ||
2502 (!IsDest && II.getRawSource() == OldPtr));
2503
2504 unsigned SliceAlign = getSliceAlign();
2505
2506 // For unsplit intrinsics, we simply modify the source and destination
2507 // pointers in place. This isn't just an optimization, it is a matter of
2508 // correctness. With unsplit intrinsics we may be dealing with transfers
2509 // within a single alloca before SROA ran, or with transfers that have
2510 // a variable length. We may also be dealing with memmove instead of
2511 // memcpy, and so simply updating the pointers is the necessary for us to
2512 // update both source and dest of a single call.
2513 if (!IsSplittable) {
2514 Value *AdjustedPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2515 if (IsDest)
2516 II.setDest(AdjustedPtr);
2517 else
2518 II.setSource(AdjustedPtr);
2519
2520 if (II.getAlignment() > SliceAlign) {
2521 Type *CstTy = II.getAlignmentCst()->getType();
2522 II.setAlignment(
2523 ConstantInt::get(CstTy, MinAlign(II.getAlignment(), SliceAlign)));
2524 }
2525
2526 DEBUG(dbgs() << " to: " << II << "\n");
2527 deleteIfTriviallyDead(OldPtr);
2528 return false;
2529 }
2530 // For split transfer intrinsics we have an incredibly useful assurance:
2531 // the source and destination do not reside within the same alloca, and at
2532 // least one of them does not escape. This means that we can replace
2533 // memmove with memcpy, and we don't need to worry about all manner of
2534 // downsides to splitting and transforming the operations.
2535
2536 // If this doesn't map cleanly onto the alloca type, and that type isn't
2537 // a single value type, just emit a memcpy.
2538 bool EmitMemCpy
2539 = !VecTy && !IntTy && (BeginOffset > NewAllocaBeginOffset ||
2540 EndOffset < NewAllocaEndOffset ||
2541 !NewAI.getAllocatedType()->isSingleValueType());
2542
2543 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2544 // size hasn't been shrunk based on analysis of the viable range, this is
2545 // a no-op.
2546 if (EmitMemCpy && &OldAI == &NewAI) {
2547 // Ensure the start lines up.
2548 assert(NewBeginOffset == BeginOffset);
2549
2550 // Rewrite the size as needed.
2551 if (NewEndOffset != EndOffset)
2552 II.setLength(ConstantInt::get(II.getLength()->getType(),
2553 NewEndOffset - NewBeginOffset));
2554 return false;
2555 }
2556 // Record this instruction for deletion.
2557 Pass.DeadInsts.insert(&II);
2558
2559 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2560 // alloca that should be re-examined after rewriting this instruction.
2561 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2562 if (AllocaInst *AI
2563 = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
2564 assert(AI != &OldAI && AI != &NewAI &&
2565 "Splittable transfers cannot reach the same alloca on both ends.");
2566 Pass.Worklist.insert(AI);
2567 }
2568
2569 Type *OtherPtrTy = OtherPtr->getType();
2570 unsigned OtherAS = OtherPtrTy->getPointerAddressSpace();
2571
2572 // Compute the relative offset for the other pointer within the transfer.
2573 unsigned IntPtrWidth = DL.getPointerSizeInBits(OtherAS);
2574 APInt OtherOffset(IntPtrWidth, NewBeginOffset - BeginOffset);
2575 unsigned OtherAlign = MinAlign(II.getAlignment() ? II.getAlignment() : 1,
2576 OtherOffset.zextOrTrunc(64).getZExtValue());
2577
2578 if (EmitMemCpy) {
2579 // Compute the other pointer, folding as much as possible to produce
2580 // a single, simple GEP in most cases.
2581 OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2582 OtherPtr->getName() + ".");
2583
2584 Value *OurPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2585 Type *SizeTy = II.getLength()->getType();
2586 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2587
2588 CallInst *New = IRB.CreateMemCpy(
2589 IsDest ? OurPtr : OtherPtr, IsDest ? OtherPtr : OurPtr, Size,
2590 MinAlign(SliceAlign, OtherAlign), II.isVolatile());
2591 (void)New;
2592 DEBUG(dbgs() << " to: " << *New << "\n");
2593 return false;
2594 }
2595
2596 bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
2597 NewEndOffset == NewAllocaEndOffset;
2598 uint64_t Size = NewEndOffset - NewBeginOffset;
2599 unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
2600 unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
2601 unsigned NumElements = EndIndex - BeginIndex;
2602 IntegerType *SubIntTy
2603 = IntTy ? Type::getIntNTy(IntTy->getContext(), Size*8) : nullptr;
2604
2605 // Reset the other pointer type to match the register type we're going to
2606 // use, but using the address space of the original other pointer.
2607 if (VecTy && !IsWholeAlloca) {
2608 if (NumElements == 1)
2609 OtherPtrTy = VecTy->getElementType();
2610 else
2611 OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2612
2613 OtherPtrTy = OtherPtrTy->getPointerTo(OtherAS);
2614 } else if (IntTy && !IsWholeAlloca) {
2615 OtherPtrTy = SubIntTy->getPointerTo(OtherAS);
2616 } else {
2617 OtherPtrTy = NewAllocaTy->getPointerTo(OtherAS);
2618 }
2619
2620 Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2621 OtherPtr->getName() + ".");
2622 unsigned SrcAlign = OtherAlign;
2623 Value *DstPtr = &NewAI;
2624 unsigned DstAlign = SliceAlign;
2625 if (!IsDest) {
2626 std::swap(SrcPtr, DstPtr);
2627 std::swap(SrcAlign, DstAlign);
2628 }
2629
2630 Value *Src;
2631 if (VecTy && !IsWholeAlloca && !IsDest) {
2632 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2633 "load");
2634 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
2635 } else if (IntTy && !IsWholeAlloca && !IsDest) {
2636 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2637 "load");
2638 Src = convertValue(DL, IRB, Src, IntTy);
2639 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2640 Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
2641 } else {
2642 Src = IRB.CreateAlignedLoad(SrcPtr, SrcAlign, II.isVolatile(),
2643 "copyload");
2644 }
2645
2646 if (VecTy && !IsWholeAlloca && IsDest) {
2647 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2648 "oldload");
2649 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
2650 } else if (IntTy && !IsWholeAlloca && IsDest) {
2651 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2652 "oldload");
2653 Old = convertValue(DL, IRB, Old, IntTy);
2654 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2655 Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
2656 Src = convertValue(DL, IRB, Src, NewAllocaTy);
2657 }
2658
2659 StoreInst *Store = cast<StoreInst>(
2660 IRB.CreateAlignedStore(Src, DstPtr, DstAlign, II.isVolatile()));
2661 (void)Store;
2662 DEBUG(dbgs() << " to: " << *Store << "\n");
2663 return !II.isVolatile();
2664 }
2665
2666 bool visitIntrinsicInst(IntrinsicInst &II) {
2667 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2668 II.getIntrinsicID() == Intrinsic::lifetime_end);
2669 DEBUG(dbgs() << " original: " << II << "\n");
2670 assert(II.getArgOperand(1) == OldPtr);
2671
2672 // Record this instruction for deletion.
2673 Pass.DeadInsts.insert(&II);
2674
2675 ConstantInt *Size
2676 = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2677 NewEndOffset - NewBeginOffset);
2678 Value *Ptr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2679 Value *New;
2680 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2681 New = IRB.CreateLifetimeStart(Ptr, Size);
2682 else
2683 New = IRB.CreateLifetimeEnd(Ptr, Size);
2684
2685 (void)New;
2686 DEBUG(dbgs() << " to: " << *New << "\n");
2687 return true;
2688 }
2689
2690 bool visitPHINode(PHINode &PN) {
2691 DEBUG(dbgs() << " original: " << PN << "\n");
2692 assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
2693 assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
2694
2695 // We would like to compute a new pointer in only one place, but have it be
2696 // as local as possible to the PHI. To do that, we re-use the location of
2697 // the old pointer, which necessarily must be in the right position to
2698 // dominate the PHI.
2699 IRBuilderTy PtrBuilder(IRB);
| 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.h"
27#include "llvm/ADT/STLExtras.h"
28#include "llvm/ADT/SetVector.h"
29#include "llvm/ADT/SmallVector.h"
30#include "llvm/ADT/Statistic.h"
31#include "llvm/Analysis/Loads.h"
32#include "llvm/Analysis/PtrUseVisitor.h"
33#include "llvm/Analysis/ValueTracking.h"
34#include "llvm/IR/Constants.h"
35#include "llvm/IR/DIBuilder.h"
36#include "llvm/IR/DataLayout.h"
37#include "llvm/IR/DebugInfo.h"
38#include "llvm/IR/DerivedTypes.h"
39#include "llvm/IR/Dominators.h"
40#include "llvm/IR/Function.h"
41#include "llvm/IR/IRBuilder.h"
42#include "llvm/IR/InstVisitor.h"
43#include "llvm/IR/Instructions.h"
44#include "llvm/IR/IntrinsicInst.h"
45#include "llvm/IR/LLVMContext.h"
46#include "llvm/IR/Operator.h"
47#include "llvm/Pass.h"
48#include "llvm/Support/CommandLine.h"
49#include "llvm/Support/Compiler.h"
50#include "llvm/Support/Debug.h"
51#include "llvm/Support/ErrorHandling.h"
52#include "llvm/Support/MathExtras.h"
53#include "llvm/Support/TimeValue.h"
54#include "llvm/Support/raw_ostream.h"
55#include "llvm/Transforms/Utils/Local.h"
56#include "llvm/Transforms/Utils/PromoteMemToReg.h"
57#include "llvm/Transforms/Utils/SSAUpdater.h"
58
59#if __cplusplus >= 201103L && !defined(NDEBUG)
60// We only use this for a debug check in C++11
61#include <random>
62#endif
63
64using namespace llvm;
65
66#define DEBUG_TYPE "sroa"
67
68STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
69STATISTIC(NumAllocaPartitions, "Number of alloca partitions formed");
70STATISTIC(MaxPartitionsPerAlloca, "Maximum number of partitions per alloca");
71STATISTIC(NumAllocaPartitionUses, "Number of alloca partition uses rewritten");
72STATISTIC(MaxUsesPerAllocaPartition, "Maximum number of uses of a partition");
73STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
74STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
75STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
76STATISTIC(NumDeleted, "Number of instructions deleted");
77STATISTIC(NumVectorized, "Number of vectorized aggregates");
78
79/// Hidden option to force the pass to not use DomTree and mem2reg, instead
80/// forming SSA values through the SSAUpdater infrastructure.
81static cl::opt<bool>
82ForceSSAUpdater("force-ssa-updater", cl::init(false), cl::Hidden);
83
84/// Hidden option to enable randomly shuffling the slices to help uncover
85/// instability in their order.
86static cl::opt<bool> SROARandomShuffleSlices("sroa-random-shuffle-slices",
87 cl::init(false), cl::Hidden);
88
89/// Hidden option to experiment with completely strict handling of inbounds
90/// GEPs.
91static cl::opt<bool> SROAStrictInbounds("sroa-strict-inbounds",
92 cl::init(false), cl::Hidden);
93
94namespace {
95/// \brief A custom IRBuilder inserter which prefixes all names if they are
96/// preserved.
97template <bool preserveNames = true>
98class IRBuilderPrefixedInserter :
99 public IRBuilderDefaultInserter<preserveNames> {
100 std::string Prefix;
101
102public:
103 void SetNamePrefix(const Twine &P) { Prefix = P.str(); }
104
105protected:
106 void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
107 BasicBlock::iterator InsertPt) const {
108 IRBuilderDefaultInserter<preserveNames>::InsertHelper(
109 I, Name.isTriviallyEmpty() ? Name : Prefix + Name, BB, InsertPt);
110 }
111};
112
113// Specialization for not preserving the name is trivial.
114template <>
115class IRBuilderPrefixedInserter<false> :
116 public IRBuilderDefaultInserter<false> {
117public:
118 void SetNamePrefix(const Twine &P) {}
119};
120
121/// \brief Provide a typedef for IRBuilder that drops names in release builds.
122#ifndef NDEBUG
123typedef llvm::IRBuilder<true, ConstantFolder,
124 IRBuilderPrefixedInserter<true> > IRBuilderTy;
125#else
126typedef llvm::IRBuilder<false, ConstantFolder,
127 IRBuilderPrefixedInserter<false> > IRBuilderTy;
128#endif
129}
130
131namespace {
132/// \brief A used slice of an alloca.
133///
134/// This structure represents a slice of an alloca used by some instruction. It
135/// stores both the begin and end offsets of this use, a pointer to the use
136/// itself, and a flag indicating whether we can classify the use as splittable
137/// or not when forming partitions of the alloca.
138class Slice {
139 /// \brief The beginning offset of the range.
140 uint64_t BeginOffset;
141
142 /// \brief The ending offset, not included in the range.
143 uint64_t EndOffset;
144
145 /// \brief Storage for both the use of this slice and whether it can be
146 /// split.
147 PointerIntPair<Use *, 1, bool> UseAndIsSplittable;
148
149public:
150 Slice() : BeginOffset(), EndOffset() {}
151 Slice(uint64_t BeginOffset, uint64_t EndOffset, Use *U, bool IsSplittable)
152 : BeginOffset(BeginOffset), EndOffset(EndOffset),
153 UseAndIsSplittable(U, IsSplittable) {}
154
155 uint64_t beginOffset() const { return BeginOffset; }
156 uint64_t endOffset() const { return EndOffset; }
157
158 bool isSplittable() const { return UseAndIsSplittable.getInt(); }
159 void makeUnsplittable() { UseAndIsSplittable.setInt(false); }
160
161 Use *getUse() const { return UseAndIsSplittable.getPointer(); }
162
163 bool isDead() const { return getUse() == nullptr; }
164 void kill() { UseAndIsSplittable.setPointer(nullptr); }
165
166 /// \brief Support for ordering ranges.
167 ///
168 /// This provides an ordering over ranges such that start offsets are
169 /// always increasing, and within equal start offsets, the end offsets are
170 /// decreasing. Thus the spanning range comes first in a cluster with the
171 /// same start position.
172 bool operator<(const Slice &RHS) const {
173 if (beginOffset() < RHS.beginOffset()) return true;
174 if (beginOffset() > RHS.beginOffset()) return false;
175 if (isSplittable() != RHS.isSplittable()) return !isSplittable();
176 if (endOffset() > RHS.endOffset()) return true;
177 return false;
178 }
179
180 /// \brief Support comparison with a single offset to allow binary searches.
181 friend LLVM_ATTRIBUTE_UNUSED bool operator<(const Slice &LHS,
182 uint64_t RHSOffset) {
183 return LHS.beginOffset() < RHSOffset;
184 }
185 friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
186 const Slice &RHS) {
187 return LHSOffset < RHS.beginOffset();
188 }
189
190 bool operator==(const Slice &RHS) const {
191 return isSplittable() == RHS.isSplittable() &&
192 beginOffset() == RHS.beginOffset() && endOffset() == RHS.endOffset();
193 }
194 bool operator!=(const Slice &RHS) const { return !operator==(RHS); }
195};
196} // end anonymous namespace
197
198namespace llvm {
199template <typename T> struct isPodLike;
200template <> struct isPodLike<Slice> {
201 static const bool value = true;
202};
203}
204
205namespace {
206/// \brief Representation of the alloca slices.
207///
208/// This class represents the slices of an alloca which are formed by its
209/// various uses. If a pointer escapes, we can't fully build a representation
210/// for the slices used and we reflect that in this structure. The uses are
211/// stored, sorted by increasing beginning offset and with unsplittable slices
212/// starting at a particular offset before splittable slices.
213class AllocaSlices {
214public:
215 /// \brief Construct the slices of a particular alloca.
216 AllocaSlices(const DataLayout &DL, AllocaInst &AI);
217
218 /// \brief Test whether a pointer to the allocation escapes our analysis.
219 ///
220 /// If this is true, the slices are never fully built and should be
221 /// ignored.
222 bool isEscaped() const { return PointerEscapingInstr; }
223
224 /// \brief Support for iterating over the slices.
225 /// @{
226 typedef SmallVectorImpl<Slice>::iterator iterator;
227 iterator begin() { return Slices.begin(); }
228 iterator end() { return Slices.end(); }
229
230 typedef SmallVectorImpl<Slice>::const_iterator const_iterator;
231 const_iterator begin() const { return Slices.begin(); }
232 const_iterator end() const { return Slices.end(); }
233 /// @}
234
235 /// \brief Allow iterating the dead users for this alloca.
236 ///
237 /// These are instructions which will never actually use the alloca as they
238 /// are outside the allocated range. They are safe to replace with undef and
239 /// delete.
240 /// @{
241 typedef SmallVectorImpl<Instruction *>::const_iterator dead_user_iterator;
242 dead_user_iterator dead_user_begin() const { return DeadUsers.begin(); }
243 dead_user_iterator dead_user_end() const { return DeadUsers.end(); }
244 /// @}
245
246 /// \brief Allow iterating the dead expressions referring to this alloca.
247 ///
248 /// These are operands which have cannot actually be used to refer to the
249 /// alloca as they are outside its range and the user doesn't correct for
250 /// that. These mostly consist of PHI node inputs and the like which we just
251 /// need to replace with undef.
252 /// @{
253 typedef SmallVectorImpl<Use *>::const_iterator dead_op_iterator;
254 dead_op_iterator dead_op_begin() const { return DeadOperands.begin(); }
255 dead_op_iterator dead_op_end() const { return DeadOperands.end(); }
256 /// @}
257
258#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
259 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
260 void printSlice(raw_ostream &OS, const_iterator I,
261 StringRef Indent = " ") const;
262 void printUse(raw_ostream &OS, const_iterator I,
263 StringRef Indent = " ") const;
264 void print(raw_ostream &OS) const;
265 void dump(const_iterator I) const;
266 void dump() const;
267#endif
268
269private:
270 template <typename DerivedT, typename RetT = void> class BuilderBase;
271 class SliceBuilder;
272 friend class AllocaSlices::SliceBuilder;
273
274#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
275 /// \brief Handle to alloca instruction to simplify method interfaces.
276 AllocaInst &AI;
277#endif
278
279 /// \brief The instruction responsible for this alloca not having a known set
280 /// of slices.
281 ///
282 /// When an instruction (potentially) escapes the pointer to the alloca, we
283 /// store a pointer to that here and abort trying to form slices of the
284 /// alloca. This will be null if the alloca slices are analyzed successfully.
285 Instruction *PointerEscapingInstr;
286
287 /// \brief The slices of the alloca.
288 ///
289 /// We store a vector of the slices formed by uses of the alloca here. This
290 /// vector is sorted by increasing begin offset, and then the unsplittable
291 /// slices before the splittable ones. See the Slice inner class for more
292 /// details.
293 SmallVector<Slice, 8> Slices;
294
295 /// \brief Instructions which will become dead if we rewrite the alloca.
296 ///
297 /// Note that these are not separated by slice. This is because we expect an
298 /// alloca to be completely rewritten or not rewritten at all. If rewritten,
299 /// all these instructions can simply be removed and replaced with undef as
300 /// they come from outside of the allocated space.
301 SmallVector<Instruction *, 8> DeadUsers;
302
303 /// \brief Operands which will become dead if we rewrite the alloca.
304 ///
305 /// These are operands that in their particular use can be replaced with
306 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
307 /// to PHI nodes and the like. They aren't entirely dead (there might be
308 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
309 /// want to swap this particular input for undef to simplify the use lists of
310 /// the alloca.
311 SmallVector<Use *, 8> DeadOperands;
312};
313}
314
315static Value *foldSelectInst(SelectInst &SI) {
316 // If the condition being selected on is a constant or the same value is
317 // being selected between, fold the select. Yes this does (rarely) happen
318 // early on.
319 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
320 return SI.getOperand(1+CI->isZero());
321 if (SI.getOperand(1) == SI.getOperand(2))
322 return SI.getOperand(1);
323
324 return nullptr;
325}
326
327/// \brief Builder for the alloca slices.
328///
329/// This class builds a set of alloca slices by recursively visiting the uses
330/// of an alloca and making a slice for each load and store at each offset.
331class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
332 friend class PtrUseVisitor<SliceBuilder>;
333 friend class InstVisitor<SliceBuilder>;
334 typedef PtrUseVisitor<SliceBuilder> Base;
335
336 const uint64_t AllocSize;
337 AllocaSlices &S;
338
339 SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
340 SmallDenseMap<Instruction *, uint64_t> PHIOrSelectSizes;
341
342 /// \brief Set to de-duplicate dead instructions found in the use walk.
343 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
344
345public:
346 SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &S)
347 : PtrUseVisitor<SliceBuilder>(DL),
348 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())), S(S) {}
349
350private:
351 void markAsDead(Instruction &I) {
352 if (VisitedDeadInsts.insert(&I))
353 S.DeadUsers.push_back(&I);
354 }
355
356 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
357 bool IsSplittable = false) {
358 // Completely skip uses which have a zero size or start either before or
359 // past the end of the allocation.
360 if (Size == 0 || Offset.uge(AllocSize)) {
361 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
362 << " which has zero size or starts outside of the "
363 << AllocSize << " byte alloca:\n"
364 << " alloca: " << S.AI << "\n"
365 << " use: " << I << "\n");
366 return markAsDead(I);
367 }
368
369 uint64_t BeginOffset = Offset.getZExtValue();
370 uint64_t EndOffset = BeginOffset + Size;
371
372 // Clamp the end offset to the end of the allocation. Note that this is
373 // formulated to handle even the case where "BeginOffset + Size" overflows.
374 // This may appear superficially to be something we could ignore entirely,
375 // but that is not so! There may be widened loads or PHI-node uses where
376 // some instructions are dead but not others. We can't completely ignore
377 // them, and so have to record at least the information here.
378 assert(AllocSize >= BeginOffset); // Established above.
379 if (Size > AllocSize - BeginOffset) {
380 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
381 << " to remain within the " << AllocSize << " byte alloca:\n"
382 << " alloca: " << S.AI << "\n"
383 << " use: " << I << "\n");
384 EndOffset = AllocSize;
385 }
386
387 S.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
388 }
389
390 void visitBitCastInst(BitCastInst &BC) {
391 if (BC.use_empty())
392 return markAsDead(BC);
393
394 return Base::visitBitCastInst(BC);
395 }
396
397 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
398 if (GEPI.use_empty())
399 return markAsDead(GEPI);
400
401 if (SROAStrictInbounds && GEPI.isInBounds()) {
402 // FIXME: This is a manually un-factored variant of the basic code inside
403 // of GEPs with checking of the inbounds invariant specified in the
404 // langref in a very strict sense. If we ever want to enable
405 // SROAStrictInbounds, this code should be factored cleanly into
406 // PtrUseVisitor, but it is easier to experiment with SROAStrictInbounds
407 // by writing out the code here where we have tho underlying allocation
408 // size readily available.
409 APInt GEPOffset = Offset;
410 for (gep_type_iterator GTI = gep_type_begin(GEPI),
411 GTE = gep_type_end(GEPI);
412 GTI != GTE; ++GTI) {
413 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
414 if (!OpC)
415 break;
416
417 // Handle a struct index, which adds its field offset to the pointer.
418 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
419 unsigned ElementIdx = OpC->getZExtValue();
420 const StructLayout *SL = DL.getStructLayout(STy);
421 GEPOffset +=
422 APInt(Offset.getBitWidth(), SL->getElementOffset(ElementIdx));
423 } else {
424 // For array or vector indices, scale the index by the size of the type.
425 APInt Index = OpC->getValue().sextOrTrunc(Offset.getBitWidth());
426 GEPOffset += Index * APInt(Offset.getBitWidth(),
427 DL.getTypeAllocSize(GTI.getIndexedType()));
428 }
429
430 // If this index has computed an intermediate pointer which is not
431 // inbounds, then the result of the GEP is a poison value and we can
432 // delete it and all uses.
433 if (GEPOffset.ugt(AllocSize))
434 return markAsDead(GEPI);
435 }
436 }
437
438 return Base::visitGetElementPtrInst(GEPI);
439 }
440
441 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
442 uint64_t Size, bool IsVolatile) {
443 // We allow splitting of loads and stores where the type is an integer type
444 // and cover the entire alloca. This prevents us from splitting over
445 // eagerly.
446 // FIXME: In the great blue eventually, we should eagerly split all integer
447 // loads and stores, and then have a separate step that merges adjacent
448 // alloca partitions into a single partition suitable for integer widening.
449 // Or we should skip the merge step and rely on GVN and other passes to
450 // merge adjacent loads and stores that survive mem2reg.
451 bool IsSplittable =
452 Ty->isIntegerTy() && !IsVolatile && Offset == 0 && Size >= AllocSize;
453
454 insertUse(I, Offset, Size, IsSplittable);
455 }
456
457 void visitLoadInst(LoadInst &LI) {
458 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
459 "All simple FCA loads should have been pre-split");
460
461 if (!IsOffsetKnown)
462 return PI.setAborted(&LI);
463
464 uint64_t Size = DL.getTypeStoreSize(LI.getType());
465 return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
466 }
467
468 void visitStoreInst(StoreInst &SI) {
469 Value *ValOp = SI.getValueOperand();
470 if (ValOp == *U)
471 return PI.setEscapedAndAborted(&SI);
472 if (!IsOffsetKnown)
473 return PI.setAborted(&SI);
474
475 uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
476
477 // If this memory access can be shown to *statically* extend outside the
478 // bounds of of the allocation, it's behavior is undefined, so simply
479 // ignore it. Note that this is more strict than the generic clamping
480 // behavior of insertUse. We also try to handle cases which might run the
481 // risk of overflow.
482 // FIXME: We should instead consider the pointer to have escaped if this
483 // function is being instrumented for addressing bugs or race conditions.
484 if (Size > AllocSize || Offset.ugt(AllocSize - Size)) {
485 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @" << Offset
486 << " which extends past the end of the " << AllocSize
487 << " byte alloca:\n"
488 << " alloca: " << S.AI << "\n"
489 << " use: " << SI << "\n");
490 return markAsDead(SI);
491 }
492
493 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
494 "All simple FCA stores should have been pre-split");
495 handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
496 }
497
498
499 void visitMemSetInst(MemSetInst &II) {
500 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
501 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
502 if ((Length && Length->getValue() == 0) ||
503 (IsOffsetKnown && Offset.uge(AllocSize)))
504 // Zero-length mem transfer intrinsics can be ignored entirely.
505 return markAsDead(II);
506
507 if (!IsOffsetKnown)
508 return PI.setAborted(&II);
509
510 insertUse(II, Offset,
511 Length ? Length->getLimitedValue()
512 : AllocSize - Offset.getLimitedValue(),
513 (bool)Length);
514 }
515
516 void visitMemTransferInst(MemTransferInst &II) {
517 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
518 if (Length && Length->getValue() == 0)
519 // Zero-length mem transfer intrinsics can be ignored entirely.
520 return markAsDead(II);
521
522 // Because we can visit these intrinsics twice, also check to see if the
523 // first time marked this instruction as dead. If so, skip it.
524 if (VisitedDeadInsts.count(&II))
525 return;
526
527 if (!IsOffsetKnown)
528 return PI.setAborted(&II);
529
530 // This side of the transfer is completely out-of-bounds, and so we can
531 // nuke the entire transfer. However, we also need to nuke the other side
532 // if already added to our partitions.
533 // FIXME: Yet another place we really should bypass this when
534 // instrumenting for ASan.
535 if (Offset.uge(AllocSize)) {
536 SmallDenseMap<Instruction *, unsigned>::iterator MTPI = MemTransferSliceMap.find(&II);
537 if (MTPI != MemTransferSliceMap.end())
538 S.Slices[MTPI->second].kill();
539 return markAsDead(II);
540 }
541
542 uint64_t RawOffset = Offset.getLimitedValue();
543 uint64_t Size = Length ? Length->getLimitedValue()
544 : AllocSize - RawOffset;
545
546 // Check for the special case where the same exact value is used for both
547 // source and dest.
548 if (*U == II.getRawDest() && *U == II.getRawSource()) {
549 // For non-volatile transfers this is a no-op.
550 if (!II.isVolatile())
551 return markAsDead(II);
552
553 return insertUse(II, Offset, Size, /*IsSplittable=*/false);
554 }
555
556 // If we have seen both source and destination for a mem transfer, then
557 // they both point to the same alloca.
558 bool Inserted;
559 SmallDenseMap<Instruction *, unsigned>::iterator MTPI;
560 std::tie(MTPI, Inserted) =
561 MemTransferSliceMap.insert(std::make_pair(&II, S.Slices.size()));
562 unsigned PrevIdx = MTPI->second;
563 if (!Inserted) {
564 Slice &PrevP = S.Slices[PrevIdx];
565
566 // Check if the begin offsets match and this is a non-volatile transfer.
567 // In that case, we can completely elide the transfer.
568 if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
569 PrevP.kill();
570 return markAsDead(II);
571 }
572
573 // Otherwise we have an offset transfer within the same alloca. We can't
574 // split those.
575 PrevP.makeUnsplittable();
576 }
577
578 // Insert the use now that we've fixed up the splittable nature.
579 insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
580
581 // Check that we ended up with a valid index in the map.
582 assert(S.Slices[PrevIdx].getUse()->getUser() == &II &&
583 "Map index doesn't point back to a slice with this user.");
584 }
585
586 // Disable SRoA for any intrinsics except for lifetime invariants.
587 // FIXME: What about debug intrinsics? This matches old behavior, but
588 // doesn't make sense.
589 void visitIntrinsicInst(IntrinsicInst &II) {
590 if (!IsOffsetKnown)
591 return PI.setAborted(&II);
592
593 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
594 II.getIntrinsicID() == Intrinsic::lifetime_end) {
595 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
596 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
597 Length->getLimitedValue());
598 insertUse(II, Offset, Size, true);
599 return;
600 }
601
602 Base::visitIntrinsicInst(II);
603 }
604
605 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
606 // We consider any PHI or select that results in a direct load or store of
607 // the same offset to be a viable use for slicing purposes. These uses
608 // are considered unsplittable and the size is the maximum loaded or stored
609 // size.
610 SmallPtrSet<Instruction *, 4> Visited;
611 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
612 Visited.insert(Root);
613 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
614 // If there are no loads or stores, the access is dead. We mark that as
615 // a size zero access.
616 Size = 0;
617 do {
618 Instruction *I, *UsedI;
619 std::tie(UsedI, I) = Uses.pop_back_val();
620
621 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
622 Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
623 continue;
624 }
625 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
626 Value *Op = SI->getOperand(0);
627 if (Op == UsedI)
628 return SI;
629 Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
630 continue;
631 }
632
633 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
634 if (!GEP->hasAllZeroIndices())
635 return GEP;
636 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
637 !isa<SelectInst>(I)) {
638 return I;
639 }
640
641 for (User *U : I->users())
642 if (Visited.insert(cast<Instruction>(U)))
643 Uses.push_back(std::make_pair(I, cast<Instruction>(U)));
644 } while (!Uses.empty());
645
646 return nullptr;
647 }
648
649 void visitPHINode(PHINode &PN) {
650 if (PN.use_empty())
651 return markAsDead(PN);
652 if (!IsOffsetKnown)
653 return PI.setAborted(&PN);
654
655 // See if we already have computed info on this node.
656 uint64_t &PHISize = PHIOrSelectSizes[&PN];
657 if (!PHISize) {
658 // This is a new PHI node, check for an unsafe use of the PHI node.
659 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&PN, PHISize))
660 return PI.setAborted(UnsafeI);
661 }
662
663 // For PHI and select operands outside the alloca, we can't nuke the entire
664 // phi or select -- the other side might still be relevant, so we special
665 // case them here and use a separate structure to track the operands
666 // themselves which should be replaced with undef.
667 // FIXME: This should instead be escaped in the event we're instrumenting
668 // for address sanitization.
669 if (Offset.uge(AllocSize)) {
670 S.DeadOperands.push_back(U);
671 return;
672 }
673
674 insertUse(PN, Offset, PHISize);
675 }
676
677 void visitSelectInst(SelectInst &SI) {
678 if (SI.use_empty())
679 return markAsDead(SI);
680 if (Value *Result = foldSelectInst(SI)) {
681 if (Result == *U)
682 // If the result of the constant fold will be the pointer, recurse
683 // through the select as if we had RAUW'ed it.
684 enqueueUsers(SI);
685 else
686 // Otherwise the operand to the select is dead, and we can replace it
687 // with undef.
688 S.DeadOperands.push_back(U);
689
690 return;
691 }
692 if (!IsOffsetKnown)
693 return PI.setAborted(&SI);
694
695 // See if we already have computed info on this node.
696 uint64_t &SelectSize = PHIOrSelectSizes[&SI];
697 if (!SelectSize) {
698 // This is a new Select, check for an unsafe use of it.
699 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&SI, SelectSize))
700 return PI.setAborted(UnsafeI);
701 }
702
703 // For PHI and select operands outside the alloca, we can't nuke the entire
704 // phi or select -- the other side might still be relevant, so we special
705 // case them here and use a separate structure to track the operands
706 // themselves which should be replaced with undef.
707 // FIXME: This should instead be escaped in the event we're instrumenting
708 // for address sanitization.
709 if (Offset.uge(AllocSize)) {
710 S.DeadOperands.push_back(U);
711 return;
712 }
713
714 insertUse(SI, Offset, SelectSize);
715 }
716
717 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
718 void visitInstruction(Instruction &I) {
719 PI.setAborted(&I);
720 }
721};
722
723AllocaSlices::AllocaSlices(const DataLayout &DL, AllocaInst &AI)
724 :
725#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
726 AI(AI),
727#endif
728 PointerEscapingInstr(nullptr) {
729 SliceBuilder PB(DL, AI, *this);
730 SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
731 if (PtrI.isEscaped() || PtrI.isAborted()) {
732 // FIXME: We should sink the escape vs. abort info into the caller nicely,
733 // possibly by just storing the PtrInfo in the AllocaSlices.
734 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
735 : PtrI.getAbortingInst();
736 assert(PointerEscapingInstr && "Did not track a bad instruction");
737 return;
738 }
739
740 Slices.erase(std::remove_if(Slices.begin(), Slices.end(),
741 std::mem_fun_ref(&Slice::isDead)),
742 Slices.end());
743
744#if __cplusplus >= 201103L && !defined(NDEBUG)
745 if (SROARandomShuffleSlices) {
746 std::mt19937 MT(static_cast<unsigned>(sys::TimeValue::now().msec()));
747 std::shuffle(Slices.begin(), Slices.end(), MT);
748 }
749#endif
750
751 // Sort the uses. This arranges for the offsets to be in ascending order,
752 // and the sizes to be in descending order.
753 std::sort(Slices.begin(), Slices.end());
754}
755
756#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
757
758void AllocaSlices::print(raw_ostream &OS, const_iterator I,
759 StringRef Indent) const {
760 printSlice(OS, I, Indent);
761 printUse(OS, I, Indent);
762}
763
764void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
765 StringRef Indent) const {
766 OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
767 << " slice #" << (I - begin())
768 << (I->isSplittable() ? " (splittable)" : "") << "\n";
769}
770
771void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
772 StringRef Indent) const {
773 OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
774}
775
776void AllocaSlices::print(raw_ostream &OS) const {
777 if (PointerEscapingInstr) {
778 OS << "Can't analyze slices for alloca: " << AI << "\n"
779 << " A pointer to this alloca escaped by:\n"
780 << " " << *PointerEscapingInstr << "\n";
781 return;
782 }
783
784 OS << "Slices of alloca: " << AI << "\n";
785 for (const_iterator I = begin(), E = end(); I != E; ++I)
786 print(OS, I);
787}
788
789LLVM_DUMP_METHOD void AllocaSlices::dump(const_iterator I) const {
790 print(dbgs(), I);
791}
792LLVM_DUMP_METHOD void AllocaSlices::dump() const { print(dbgs()); }
793
794#endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
795
796namespace {
797/// \brief Implementation of LoadAndStorePromoter for promoting allocas.
798///
799/// This subclass of LoadAndStorePromoter adds overrides to handle promoting
800/// the loads and stores of an alloca instruction, as well as updating its
801/// debug information. This is used when a domtree is unavailable and thus
802/// mem2reg in its full form can't be used to handle promotion of allocas to
803/// scalar values.
804class AllocaPromoter : public LoadAndStorePromoter {
805 AllocaInst &AI;
806 DIBuilder &DIB;
807
808 SmallVector<DbgDeclareInst *, 4> DDIs;
809 SmallVector<DbgValueInst *, 4> DVIs;
810
811public:
812 AllocaPromoter(const SmallVectorImpl<Instruction *> &Insts, SSAUpdater &S,
813 AllocaInst &AI, DIBuilder &DIB)
814 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
815
816 void run(const SmallVectorImpl<Instruction*> &Insts) {
817 // Retain the debug information attached to the alloca for use when
818 // rewriting loads and stores.
819 if (MDNode *DebugNode = MDNode::getIfExists(AI.getContext(), &AI)) {
820 for (User *U : DebugNode->users())
821 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U))
822 DDIs.push_back(DDI);
823 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(U))
824 DVIs.push_back(DVI);
825 }
826
827 LoadAndStorePromoter::run(Insts);
828
829 // While we have the debug information, clear it off of the alloca. The
830 // caller takes care of deleting the alloca.
831 while (!DDIs.empty())
832 DDIs.pop_back_val()->eraseFromParent();
833 while (!DVIs.empty())
834 DVIs.pop_back_val()->eraseFromParent();
835 }
836
837 bool isInstInList(Instruction *I,
838 const SmallVectorImpl<Instruction*> &Insts) const override {
839 Value *Ptr;
840 if (LoadInst *LI = dyn_cast<LoadInst>(I))
841 Ptr = LI->getOperand(0);
842 else
843 Ptr = cast<StoreInst>(I)->getPointerOperand();
844
845 // Only used to detect cycles, which will be rare and quickly found as
846 // we're walking up a chain of defs rather than down through uses.
847 SmallPtrSet<Value *, 4> Visited;
848
849 do {
850 if (Ptr == &AI)
851 return true;
852
853 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr))
854 Ptr = BCI->getOperand(0);
855 else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Ptr))
856 Ptr = GEPI->getPointerOperand();
857 else
858 return false;
859
860 } while (Visited.insert(Ptr));
861
862 return false;
863 }
864
865 void updateDebugInfo(Instruction *Inst) const override {
866 for (SmallVectorImpl<DbgDeclareInst *>::const_iterator I = DDIs.begin(),
867 E = DDIs.end(); I != E; ++I) {
868 DbgDeclareInst *DDI = *I;
869 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
870 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
871 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
872 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
873 }
874 for (SmallVectorImpl<DbgValueInst *>::const_iterator I = DVIs.begin(),
875 E = DVIs.end(); I != E; ++I) {
876 DbgValueInst *DVI = *I;
877 Value *Arg = nullptr;
878 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
879 // If an argument is zero extended then use argument directly. The ZExt
880 // may be zapped by an optimization pass in future.
881 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
882 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
883 else if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
884 Arg = dyn_cast<Argument>(SExt->getOperand(0));
885 if (!Arg)
886 Arg = SI->getValueOperand();
887 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
888 Arg = LI->getPointerOperand();
889 } else {
890 continue;
891 }
892 Instruction *DbgVal =
893 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
894 Inst);
895 DbgVal->setDebugLoc(DVI->getDebugLoc());
896 }
897 }
898};
899} // end anon namespace
900
901
902namespace {
903/// \brief An optimization pass providing Scalar Replacement of Aggregates.
904///
905/// This pass takes allocations which can be completely analyzed (that is, they
906/// don't escape) and tries to turn them into scalar SSA values. There are
907/// a few steps to this process.
908///
909/// 1) It takes allocations of aggregates and analyzes the ways in which they
910/// are used to try to split them into smaller allocations, ideally of
911/// a single scalar data type. It will split up memcpy and memset accesses
912/// as necessary and try to isolate individual scalar accesses.
913/// 2) It will transform accesses into forms which are suitable for SSA value
914/// promotion. This can be replacing a memset with a scalar store of an
915/// integer value, or it can involve speculating operations on a PHI or
916/// select to be a PHI or select of the results.
917/// 3) Finally, this will try to detect a pattern of accesses which map cleanly
918/// onto insert and extract operations on a vector value, and convert them to
919/// this form. By doing so, it will enable promotion of vector aggregates to
920/// SSA vector values.
921class SROA : public FunctionPass {
922 const bool RequiresDomTree;
923
924 LLVMContext *C;
925 const DataLayout *DL;
926 DominatorTree *DT;
927
928 /// \brief Worklist of alloca instructions to simplify.
929 ///
930 /// Each alloca in the function is added to this. Each new alloca formed gets
931 /// added to it as well to recursively simplify unless that alloca can be
932 /// directly promoted. Finally, each time we rewrite a use of an alloca other
933 /// the one being actively rewritten, we add it back onto the list if not
934 /// already present to ensure it is re-visited.
935 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > Worklist;
936
937 /// \brief A collection of instructions to delete.
938 /// We try to batch deletions to simplify code and make things a bit more
939 /// efficient.
940 SetVector<Instruction *, SmallVector<Instruction *, 8> > DeadInsts;
941
942 /// \brief Post-promotion worklist.
943 ///
944 /// Sometimes we discover an alloca which has a high probability of becoming
945 /// viable for SROA after a round of promotion takes place. In those cases,
946 /// the alloca is enqueued here for re-processing.
947 ///
948 /// Note that we have to be very careful to clear allocas out of this list in
949 /// the event they are deleted.
950 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > PostPromotionWorklist;
951
952 /// \brief A collection of alloca instructions we can directly promote.
953 std::vector<AllocaInst *> PromotableAllocas;
954
955 /// \brief A worklist of PHIs to speculate prior to promoting allocas.
956 ///
957 /// All of these PHIs have been checked for the safety of speculation and by
958 /// being speculated will allow promoting allocas currently in the promotable
959 /// queue.
960 SetVector<PHINode *, SmallVector<PHINode *, 2> > SpeculatablePHIs;
961
962 /// \brief A worklist of select instructions to speculate prior to promoting
963 /// allocas.
964 ///
965 /// All of these select instructions have been checked for the safety of
966 /// speculation and by being speculated will allow promoting allocas
967 /// currently in the promotable queue.
968 SetVector<SelectInst *, SmallVector<SelectInst *, 2> > SpeculatableSelects;
969
970public:
971 SROA(bool RequiresDomTree = true)
972 : FunctionPass(ID), RequiresDomTree(RequiresDomTree),
973 C(nullptr), DL(nullptr), DT(nullptr) {
974 initializeSROAPass(*PassRegistry::getPassRegistry());
975 }
976 bool runOnFunction(Function &F) override;
977 void getAnalysisUsage(AnalysisUsage &AU) const override;
978
979 const char *getPassName() const override { return "SROA"; }
980 static char ID;
981
982private:
983 friend class PHIOrSelectSpeculator;
984 friend class AllocaSliceRewriter;
985
986 bool rewritePartition(AllocaInst &AI, AllocaSlices &S,
987 AllocaSlices::iterator B, AllocaSlices::iterator E,
988 int64_t BeginOffset, int64_t EndOffset,
989 ArrayRef<AllocaSlices::iterator> SplitUses);
990 bool splitAlloca(AllocaInst &AI, AllocaSlices &S);
991 bool runOnAlloca(AllocaInst &AI);
992 void clobberUse(Use &U);
993 void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas);
994 bool promoteAllocas(Function &F);
995};
996}
997
998char SROA::ID = 0;
999
1000FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
1001 return new SROA(RequiresDomTree);
1002}
1003
1004INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates",
1005 false, false)
1006INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1007INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
1008 false, false)
1009
1010/// Walk the range of a partitioning looking for a common type to cover this
1011/// sequence of slices.
1012static Type *findCommonType(AllocaSlices::const_iterator B,
1013 AllocaSlices::const_iterator E,
1014 uint64_t EndOffset) {
1015 Type *Ty = nullptr;
1016 bool TyIsCommon = true;
1017 IntegerType *ITy = nullptr;
1018
1019 // Note that we need to look at *every* alloca slice's Use to ensure we
1020 // always get consistent results regardless of the order of slices.
1021 for (AllocaSlices::const_iterator I = B; I != E; ++I) {
1022 Use *U = I->getUse();
1023 if (isa<IntrinsicInst>(*U->getUser()))
1024 continue;
1025 if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
1026 continue;
1027
1028 Type *UserTy = nullptr;
1029 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1030 UserTy = LI->getType();
1031 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1032 UserTy = SI->getValueOperand()->getType();
1033 }
1034
1035 if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
1036 // If the type is larger than the partition, skip it. We only encounter
1037 // this for split integer operations where we want to use the type of the
1038 // entity causing the split. Also skip if the type is not a byte width
1039 // multiple.
1040 if (UserITy->getBitWidth() % 8 != 0 ||
1041 UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
1042 continue;
1043
1044 // Track the largest bitwidth integer type used in this way in case there
1045 // is no common type.
1046 if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
1047 ITy = UserITy;
1048 }
1049
1050 // To avoid depending on the order of slices, Ty and TyIsCommon must not
1051 // depend on types skipped above.
1052 if (!UserTy || (Ty && Ty != UserTy))
1053 TyIsCommon = false; // Give up on anything but an iN type.
1054 else
1055 Ty = UserTy;
1056 }
1057
1058 return TyIsCommon ? Ty : ITy;
1059}
1060
1061/// PHI instructions that use an alloca and are subsequently loaded can be
1062/// rewritten to load both input pointers in the pred blocks and then PHI the
1063/// results, allowing the load of the alloca to be promoted.
1064/// From this:
1065/// %P2 = phi [i32* %Alloca, i32* %Other]
1066/// %V = load i32* %P2
1067/// to:
1068/// %V1 = load i32* %Alloca -> will be mem2reg'd
1069/// ...
1070/// %V2 = load i32* %Other
1071/// ...
1072/// %V = phi [i32 %V1, i32 %V2]
1073///
1074/// We can do this to a select if its only uses are loads and if the operands
1075/// to the select can be loaded unconditionally.
1076///
1077/// FIXME: This should be hoisted into a generic utility, likely in
1078/// Transforms/Util/Local.h
1079static bool isSafePHIToSpeculate(PHINode &PN,
1080 const DataLayout *DL = nullptr) {
1081 // For now, we can only do this promotion if the load is in the same block
1082 // as the PHI, and if there are no stores between the phi and load.
1083 // TODO: Allow recursive phi users.
1084 // TODO: Allow stores.
1085 BasicBlock *BB = PN.getParent();
1086 unsigned MaxAlign = 0;
1087 bool HaveLoad = false;
1088 for (User *U : PN.users()) {
1089 LoadInst *LI = dyn_cast<LoadInst>(U);
1090 if (!LI || !LI->isSimple())
1091 return false;
1092
1093 // For now we only allow loads in the same block as the PHI. This is
1094 // a common case that happens when instcombine merges two loads through
1095 // a PHI.
1096 if (LI->getParent() != BB)
1097 return false;
1098
1099 // Ensure that there are no instructions between the PHI and the load that
1100 // could store.
1101 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1102 if (BBI->mayWriteToMemory())
1103 return false;
1104
1105 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1106 HaveLoad = true;
1107 }
1108
1109 if (!HaveLoad)
1110 return false;
1111
1112 // We can only transform this if it is safe to push the loads into the
1113 // predecessor blocks. The only thing to watch out for is that we can't put
1114 // a possibly trapping load in the predecessor if it is a critical edge.
1115 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1116 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1117 Value *InVal = PN.getIncomingValue(Idx);
1118
1119 // If the value is produced by the terminator of the predecessor (an
1120 // invoke) or it has side-effects, there is no valid place to put a load
1121 // in the predecessor.
1122 if (TI == InVal || TI->mayHaveSideEffects())
1123 return false;
1124
1125 // If the predecessor has a single successor, then the edge isn't
1126 // critical.
1127 if (TI->getNumSuccessors() == 1)
1128 continue;
1129
1130 // If this pointer is always safe to load, or if we can prove that there
1131 // is already a load in the block, then we can move the load to the pred
1132 // block.
1133 if (InVal->isDereferenceablePointer(DL) ||
1134 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, DL))
1135 continue;
1136
1137 return false;
1138 }
1139
1140 return true;
1141}
1142
1143static void speculatePHINodeLoads(PHINode &PN) {
1144 DEBUG(dbgs() << " original: " << PN << "\n");
1145
1146 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1147 IRBuilderTy PHIBuilder(&PN);
1148 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1149 PN.getName() + ".sroa.speculated");
1150
1151 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1152 // matter which one we get and if any differ.
1153 LoadInst *SomeLoad = cast<LoadInst>(PN.user_back());
1154 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1155 unsigned Align = SomeLoad->getAlignment();
1156
1157 // Rewrite all loads of the PN to use the new PHI.
1158 while (!PN.use_empty()) {
1159 LoadInst *LI = cast<LoadInst>(PN.user_back());
1160 LI->replaceAllUsesWith(NewPN);
1161 LI->eraseFromParent();
1162 }
1163
1164 // Inject loads into all of the pred blocks.
1165 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1166 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1167 TerminatorInst *TI = Pred->getTerminator();
1168 Value *InVal = PN.getIncomingValue(Idx);
1169 IRBuilderTy PredBuilder(TI);
1170
1171 LoadInst *Load = PredBuilder.CreateLoad(
1172 InVal, (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1173 ++NumLoadsSpeculated;
1174 Load->setAlignment(Align);
1175 if (TBAATag)
1176 Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1177 NewPN->addIncoming(Load, Pred);
1178 }
1179
1180 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1181 PN.eraseFromParent();
1182}
1183
1184/// Select instructions that use an alloca and are subsequently loaded can be
1185/// rewritten to load both input pointers and then select between the result,
1186/// allowing the load of the alloca to be promoted.
1187/// From this:
1188/// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1189/// %V = load i32* %P2
1190/// to:
1191/// %V1 = load i32* %Alloca -> will be mem2reg'd
1192/// %V2 = load i32* %Other
1193/// %V = select i1 %cond, i32 %V1, i32 %V2
1194///
1195/// We can do this to a select if its only uses are loads and if the operand
1196/// to the select can be loaded unconditionally.
1197static bool isSafeSelectToSpeculate(SelectInst &SI,
1198 const DataLayout *DL = nullptr) {
1199 Value *TValue = SI.getTrueValue();
1200 Value *FValue = SI.getFalseValue();
1201 bool TDerefable = TValue->isDereferenceablePointer(DL);
1202 bool FDerefable = FValue->isDereferenceablePointer(DL);
1203
1204 for (User *U : SI.users()) {
1205 LoadInst *LI = dyn_cast<LoadInst>(U);
1206 if (!LI || !LI->isSimple())
1207 return false;
1208
1209 // Both operands to the select need to be dereferencable, either
1210 // absolutely (e.g. allocas) or at this point because we can see other
1211 // accesses to it.
1212 if (!TDerefable &&
1213 !isSafeToLoadUnconditionally(TValue, LI, LI->getAlignment(), DL))
1214 return false;
1215 if (!FDerefable &&
1216 !isSafeToLoadUnconditionally(FValue, LI, LI->getAlignment(), DL))
1217 return false;
1218 }
1219
1220 return true;
1221}
1222
1223static void speculateSelectInstLoads(SelectInst &SI) {
1224 DEBUG(dbgs() << " original: " << SI << "\n");
1225
1226 IRBuilderTy IRB(&SI);
1227 Value *TV = SI.getTrueValue();
1228 Value *FV = SI.getFalseValue();
1229 // Replace the loads of the select with a select of two loads.
1230 while (!SI.use_empty()) {
1231 LoadInst *LI = cast<LoadInst>(SI.user_back());
1232 assert(LI->isSimple() && "We only speculate simple loads");
1233
1234 IRB.SetInsertPoint(LI);
1235 LoadInst *TL =
1236 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1237 LoadInst *FL =
1238 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1239 NumLoadsSpeculated += 2;
1240
1241 // Transfer alignment and TBAA info if present.
1242 TL->setAlignment(LI->getAlignment());
1243 FL->setAlignment(LI->getAlignment());
1244 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
1245 TL->setMetadata(LLVMContext::MD_tbaa, Tag);
1246 FL->setMetadata(LLVMContext::MD_tbaa, Tag);
1247 }
1248
1249 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1250 LI->getName() + ".sroa.speculated");
1251
1252 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1253 LI->replaceAllUsesWith(V);
1254 LI->eraseFromParent();
1255 }
1256 SI.eraseFromParent();
1257}
1258
1259/// \brief Build a GEP out of a base pointer and indices.
1260///
1261/// This will return the BasePtr if that is valid, or build a new GEP
1262/// instruction using the IRBuilder if GEP-ing is needed.
1263static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1264 SmallVectorImpl<Value *> &Indices, Twine NamePrefix) {
1265 if (Indices.empty())
1266 return BasePtr;
1267
1268 // A single zero index is a no-op, so check for this and avoid building a GEP
1269 // in that case.
1270 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1271 return BasePtr;
1272
1273 return IRB.CreateInBoundsGEP(BasePtr, Indices, NamePrefix + "sroa_idx");
1274}
1275
1276/// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1277/// TargetTy without changing the offset of the pointer.
1278///
1279/// This routine assumes we've already established a properly offset GEP with
1280/// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1281/// zero-indices down through type layers until we find one the same as
1282/// TargetTy. If we can't find one with the same type, we at least try to use
1283/// one with the same size. If none of that works, we just produce the GEP as
1284/// indicated by Indices to have the correct offset.
1285static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1286 Value *BasePtr, Type *Ty, Type *TargetTy,
1287 SmallVectorImpl<Value *> &Indices,
1288 Twine NamePrefix) {
1289 if (Ty == TargetTy)
1290 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1291
1292 // Pointer size to use for the indices.
1293 unsigned PtrSize = DL.getPointerTypeSizeInBits(BasePtr->getType());
1294
1295 // See if we can descend into a struct and locate a field with the correct
1296 // type.
1297 unsigned NumLayers = 0;
1298 Type *ElementTy = Ty;
1299 do {
1300 if (ElementTy->isPointerTy())
1301 break;
1302
1303 if (ArrayType *ArrayTy = dyn_cast<ArrayType>(ElementTy)) {
1304 ElementTy = ArrayTy->getElementType();
1305 Indices.push_back(IRB.getIntN(PtrSize, 0));
1306 } else if (VectorType *VectorTy = dyn_cast<VectorType>(ElementTy)) {
1307 ElementTy = VectorTy->getElementType();
1308 Indices.push_back(IRB.getInt32(0));
1309 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1310 if (STy->element_begin() == STy->element_end())
1311 break; // Nothing left to descend into.
1312 ElementTy = *STy->element_begin();
1313 Indices.push_back(IRB.getInt32(0));
1314 } else {
1315 break;
1316 }
1317 ++NumLayers;
1318 } while (ElementTy != TargetTy);
1319 if (ElementTy != TargetTy)
1320 Indices.erase(Indices.end() - NumLayers, Indices.end());
1321
1322 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1323}
1324
1325/// \brief Recursively compute indices for a natural GEP.
1326///
1327/// This is the recursive step for getNaturalGEPWithOffset that walks down the
1328/// element types adding appropriate indices for the GEP.
1329static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1330 Value *Ptr, Type *Ty, APInt &Offset,
1331 Type *TargetTy,
1332 SmallVectorImpl<Value *> &Indices,
1333 Twine NamePrefix) {
1334 if (Offset == 0)
1335 return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices, NamePrefix);
1336
1337 // We can't recurse through pointer types.
1338 if (Ty->isPointerTy())
1339 return nullptr;
1340
1341 // We try to analyze GEPs over vectors here, but note that these GEPs are
1342 // extremely poorly defined currently. The long-term goal is to remove GEPing
1343 // over a vector from the IR completely.
1344 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1345 unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
1346 if (ElementSizeInBits % 8 != 0) {
1347 // GEPs over non-multiple of 8 size vector elements are invalid.
1348 return nullptr;
1349 }
1350 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1351 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1352 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1353 return nullptr;
1354 Offset -= NumSkippedElements * ElementSize;
1355 Indices.push_back(IRB.getInt(NumSkippedElements));
1356 return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1357 Offset, TargetTy, Indices, NamePrefix);
1358 }
1359
1360 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1361 Type *ElementTy = ArrTy->getElementType();
1362 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1363 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1364 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1365 return nullptr;
1366
1367 Offset -= NumSkippedElements * ElementSize;
1368 Indices.push_back(IRB.getInt(NumSkippedElements));
1369 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1370 Indices, NamePrefix);
1371 }
1372
1373 StructType *STy = dyn_cast<StructType>(Ty);
1374 if (!STy)
1375 return nullptr;
1376
1377 const StructLayout *SL = DL.getStructLayout(STy);
1378 uint64_t StructOffset = Offset.getZExtValue();
1379 if (StructOffset >= SL->getSizeInBytes())
1380 return nullptr;
1381 unsigned Index = SL->getElementContainingOffset(StructOffset);
1382 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1383 Type *ElementTy = STy->getElementType(Index);
1384 if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
1385 return nullptr; // The offset points into alignment padding.
1386
1387 Indices.push_back(IRB.getInt32(Index));
1388 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1389 Indices, NamePrefix);
1390}
1391
1392/// \brief Get a natural GEP from a base pointer to a particular offset and
1393/// resulting in a particular type.
1394///
1395/// The goal is to produce a "natural" looking GEP that works with the existing
1396/// composite types to arrive at the appropriate offset and element type for
1397/// a pointer. TargetTy is the element type the returned GEP should point-to if
1398/// possible. We recurse by decreasing Offset, adding the appropriate index to
1399/// Indices, and setting Ty to the result subtype.
1400///
1401/// If no natural GEP can be constructed, this function returns null.
1402static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1403 Value *Ptr, APInt Offset, Type *TargetTy,
1404 SmallVectorImpl<Value *> &Indices,
1405 Twine NamePrefix) {
1406 PointerType *Ty = cast<PointerType>(Ptr->getType());
1407
1408 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1409 // an i8.
1410 if (Ty == IRB.getInt8PtrTy(Ty->getAddressSpace()) && TargetTy->isIntegerTy(8))
1411 return nullptr;
1412
1413 Type *ElementTy = Ty->getElementType();
1414 if (!ElementTy->isSized())
1415 return nullptr; // We can't GEP through an unsized element.
1416 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1417 if (ElementSize == 0)
1418 return nullptr; // Zero-length arrays can't help us build a natural GEP.
1419 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1420
1421 Offset -= NumSkippedElements * ElementSize;
1422 Indices.push_back(IRB.getInt(NumSkippedElements));
1423 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1424 Indices, NamePrefix);
1425}
1426
1427/// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1428/// resulting pointer has PointerTy.
1429///
1430/// This tries very hard to compute a "natural" GEP which arrives at the offset
1431/// and produces the pointer type desired. Where it cannot, it will try to use
1432/// the natural GEP to arrive at the offset and bitcast to the type. Where that
1433/// fails, it will try to use an existing i8* and GEP to the byte offset and
1434/// bitcast to the type.
1435///
1436/// The strategy for finding the more natural GEPs is to peel off layers of the
1437/// pointer, walking back through bit casts and GEPs, searching for a base
1438/// pointer from which we can compute a natural GEP with the desired
1439/// properties. The algorithm tries to fold as many constant indices into
1440/// a single GEP as possible, thus making each GEP more independent of the
1441/// surrounding code.
1442static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr,
1443 APInt Offset, Type *PointerTy,
1444 Twine NamePrefix) {
1445 // Even though we don't look through PHI nodes, we could be called on an
1446 // instruction in an unreachable block, which may be on a cycle.
1447 SmallPtrSet<Value *, 4> Visited;
1448 Visited.insert(Ptr);
1449 SmallVector<Value *, 4> Indices;
1450
1451 // We may end up computing an offset pointer that has the wrong type. If we
1452 // never are able to compute one directly that has the correct type, we'll
1453 // fall back to it, so keep it around here.
1454 Value *OffsetPtr = nullptr;
1455
1456 // Remember any i8 pointer we come across to re-use if we need to do a raw
1457 // byte offset.
1458 Value *Int8Ptr = nullptr;
1459 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1460
1461 Type *TargetTy = PointerTy->getPointerElementType();
1462
1463 do {
1464 // First fold any existing GEPs into the offset.
1465 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1466 APInt GEPOffset(Offset.getBitWidth(), 0);
1467 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1468 break;
1469 Offset += GEPOffset;
1470 Ptr = GEP->getPointerOperand();
1471 if (!Visited.insert(Ptr))
1472 break;
1473 }
1474
1475 // See if we can perform a natural GEP here.
1476 Indices.clear();
1477 if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1478 Indices, NamePrefix)) {
1479 if (P->getType() == PointerTy) {
1480 // Zap any offset pointer that we ended up computing in previous rounds.
1481 if (OffsetPtr && OffsetPtr->use_empty())
1482 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1483 I->eraseFromParent();
1484 return P;
1485 }
1486 if (!OffsetPtr) {
1487 OffsetPtr = P;
1488 }
1489 }
1490
1491 // Stash this pointer if we've found an i8*.
1492 if (Ptr->getType()->isIntegerTy(8)) {
1493 Int8Ptr = Ptr;
1494 Int8PtrOffset = Offset;
1495 }
1496
1497 // Peel off a layer of the pointer and update the offset appropriately.
1498 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1499 Ptr = cast<Operator>(Ptr)->getOperand(0);
1500 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1501 if (GA->mayBeOverridden())
1502 break;
1503 Ptr = GA->getAliasee();
1504 } else {
1505 break;
1506 }
1507 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1508 } while (Visited.insert(Ptr));
1509
1510 if (!OffsetPtr) {
1511 if (!Int8Ptr) {
1512 Int8Ptr = IRB.CreateBitCast(
1513 Ptr, IRB.getInt8PtrTy(PointerTy->getPointerAddressSpace()),
1514 NamePrefix + "sroa_raw_cast");
1515 Int8PtrOffset = Offset;
1516 }
1517
1518 OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
1519 IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1520 NamePrefix + "sroa_raw_idx");
1521 }
1522 Ptr = OffsetPtr;
1523
1524 // On the off chance we were targeting i8*, guard the bitcast here.
1525 if (Ptr->getType() != PointerTy)
1526 Ptr = IRB.CreateBitCast(Ptr, PointerTy, NamePrefix + "sroa_cast");
1527
1528 return Ptr;
1529}
1530
1531/// \brief Test whether we can convert a value from the old to the new type.
1532///
1533/// This predicate should be used to guard calls to convertValue in order to
1534/// ensure that we only try to convert viable values. The strategy is that we
1535/// will peel off single element struct and array wrappings to get to an
1536/// underlying value, and convert that value.
1537static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1538 if (OldTy == NewTy)
1539 return true;
1540 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1541 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1542 if (NewITy->getBitWidth() >= OldITy->getBitWidth())
1543 return true;
1544 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1545 return false;
1546 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1547 return false;
1548
1549 // We can convert pointers to integers and vice-versa. Same for vectors
1550 // of pointers and integers.
1551 OldTy = OldTy->getScalarType();
1552 NewTy = NewTy->getScalarType();
1553 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1554 if (NewTy->isPointerTy() && OldTy->isPointerTy())
1555 return true;
1556 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
1557 return true;
1558 return false;
1559 }
1560
1561 return true;
1562}
1563
1564/// \brief Generic routine to convert an SSA value to a value of a different
1565/// type.
1566///
1567/// This will try various different casting techniques, such as bitcasts,
1568/// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1569/// two types for viability with this routine.
1570static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1571 Type *NewTy) {
1572 Type *OldTy = V->getType();
1573 assert(canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type");
1574
1575 if (OldTy == NewTy)
1576 return V;
1577
1578 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1579 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1580 if (NewITy->getBitWidth() > OldITy->getBitWidth())
1581 return IRB.CreateZExt(V, NewITy);
1582
1583 // See if we need inttoptr for this type pair. A cast involving both scalars
1584 // and vectors requires and additional bitcast.
1585 if (OldTy->getScalarType()->isIntegerTy() &&
1586 NewTy->getScalarType()->isPointerTy()) {
1587 // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1588 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1589 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1590 NewTy);
1591
1592 // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1593 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1594 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1595 NewTy);
1596
1597 return IRB.CreateIntToPtr(V, NewTy);
1598 }
1599
1600 // See if we need ptrtoint for this type pair. A cast involving both scalars
1601 // and vectors requires and additional bitcast.
1602 if (OldTy->getScalarType()->isPointerTy() &&
1603 NewTy->getScalarType()->isIntegerTy()) {
1604 // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1605 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1606 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1607 NewTy);
1608
1609 // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1610 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1611 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1612 NewTy);
1613
1614 return IRB.CreatePtrToInt(V, NewTy);
1615 }
1616
1617 return IRB.CreateBitCast(V, NewTy);
1618}
1619
1620/// \brief Test whether the given slice use can be promoted to a vector.
1621///
1622/// This function is called to test each entry in a partioning which is slated
1623/// for a single slice.
1624static bool isVectorPromotionViableForSlice(
1625 const DataLayout &DL, AllocaSlices &S, uint64_t SliceBeginOffset,
1626 uint64_t SliceEndOffset, VectorType *Ty, uint64_t ElementSize,
1627 AllocaSlices::const_iterator I) {
1628 // First validate the slice offsets.
1629 uint64_t BeginOffset =
1630 std::max(I->beginOffset(), SliceBeginOffset) - SliceBeginOffset;
1631 uint64_t BeginIndex = BeginOffset / ElementSize;
1632 if (BeginIndex * ElementSize != BeginOffset ||
1633 BeginIndex >= Ty->getNumElements())
1634 return false;
1635 uint64_t EndOffset =
1636 std::min(I->endOffset(), SliceEndOffset) - SliceBeginOffset;
1637 uint64_t EndIndex = EndOffset / ElementSize;
1638 if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
1639 return false;
1640
1641 assert(EndIndex > BeginIndex && "Empty vector!");
1642 uint64_t NumElements = EndIndex - BeginIndex;
1643 Type *SliceTy =
1644 (NumElements == 1) ? Ty->getElementType()
1645 : VectorType::get(Ty->getElementType(), NumElements);
1646
1647 Type *SplitIntTy =
1648 Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1649
1650 Use *U = I->getUse();
1651
1652 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1653 if (MI->isVolatile())
1654 return false;
1655 if (!I->isSplittable())
1656 return false; // Skip any unsplittable intrinsics.
1657 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1658 // Disable vector promotion when there are loads or stores of an FCA.
1659 return false;
1660 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1661 if (LI->isVolatile())
1662 return false;
1663 Type *LTy = LI->getType();
1664 if (SliceBeginOffset > I->beginOffset() ||
1665 SliceEndOffset < I->endOffset()) {
1666 assert(LTy->isIntegerTy());
1667 LTy = SplitIntTy;
1668 }
1669 if (!canConvertValue(DL, SliceTy, LTy))
1670 return false;
1671 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1672 if (SI->isVolatile())
1673 return false;
1674 Type *STy = SI->getValueOperand()->getType();
1675 if (SliceBeginOffset > I->beginOffset() ||
1676 SliceEndOffset < I->endOffset()) {
1677 assert(STy->isIntegerTy());
1678 STy = SplitIntTy;
1679 }
1680 if (!canConvertValue(DL, STy, SliceTy))
1681 return false;
1682 } else {
1683 return false;
1684 }
1685
1686 return true;
1687}
1688
1689/// \brief Test whether the given alloca partitioning and range of slices can be
1690/// promoted to a vector.
1691///
1692/// This is a quick test to check whether we can rewrite a particular alloca
1693/// partition (and its newly formed alloca) into a vector alloca with only
1694/// whole-vector loads and stores such that it could be promoted to a vector
1695/// SSA value. We only can ensure this for a limited set of operations, and we
1696/// don't want to do the rewrites unless we are confident that the result will
1697/// be promotable, so we have an early test here.
1698static bool
1699isVectorPromotionViable(const DataLayout &DL, Type *AllocaTy, AllocaSlices &S,
1700 uint64_t SliceBeginOffset, uint64_t SliceEndOffset,
1701 AllocaSlices::const_iterator I,
1702 AllocaSlices::const_iterator E,
1703 ArrayRef<AllocaSlices::iterator> SplitUses) {
1704 VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
1705 if (!Ty)
1706 return false;
1707
1708 uint64_t ElementSize = DL.getTypeSizeInBits(Ty->getScalarType());
1709
1710 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1711 // that aren't byte sized.
1712 if (ElementSize % 8)
1713 return false;
1714 assert((DL.getTypeSizeInBits(Ty) % 8) == 0 &&
1715 "vector size not a multiple of element size?");
1716 ElementSize /= 8;
1717
1718 for (; I != E; ++I)
1719 if (!isVectorPromotionViableForSlice(DL, S, SliceBeginOffset,
1720 SliceEndOffset, Ty, ElementSize, I))
1721 return false;
1722
1723 for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
1724 SUE = SplitUses.end();
1725 SUI != SUE; ++SUI)
1726 if (!isVectorPromotionViableForSlice(DL, S, SliceBeginOffset,
1727 SliceEndOffset, Ty, ElementSize, *SUI))
1728 return false;
1729
1730 return true;
1731}
1732
1733/// \brief Test whether a slice of an alloca is valid for integer widening.
1734///
1735/// This implements the necessary checking for the \c isIntegerWideningViable
1736/// test below on a single slice of the alloca.
1737static bool isIntegerWideningViableForSlice(const DataLayout &DL,
1738 Type *AllocaTy,
1739 uint64_t AllocBeginOffset,
1740 uint64_t Size, AllocaSlices &S,
1741 AllocaSlices::const_iterator I,
1742 bool &WholeAllocaOp) {
1743 uint64_t RelBegin = I->beginOffset() - AllocBeginOffset;
1744 uint64_t RelEnd = I->endOffset() - AllocBeginOffset;
1745
1746 // We can't reasonably handle cases where the load or store extends past
1747 // the end of the aloca's type and into its padding.
1748 if (RelEnd > Size)
1749 return false;
1750
1751 Use *U = I->getUse();
1752
1753 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1754 if (LI->isVolatile())
1755 return false;
1756 if (RelBegin == 0 && RelEnd == Size)
1757 WholeAllocaOp = true;
1758 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
1759 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1760 return false;
1761 } else if (RelBegin != 0 || RelEnd != Size ||
1762 !canConvertValue(DL, AllocaTy, LI->getType())) {
1763 // Non-integer loads need to be convertible from the alloca type so that
1764 // they are promotable.
1765 return false;
1766 }
1767 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1768 Type *ValueTy = SI->getValueOperand()->getType();
1769 if (SI->isVolatile())
1770 return false;
1771 if (RelBegin == 0 && RelEnd == Size)
1772 WholeAllocaOp = true;
1773 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
1774 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1775 return false;
1776 } else if (RelBegin != 0 || RelEnd != Size ||
1777 !canConvertValue(DL, ValueTy, AllocaTy)) {
1778 // Non-integer stores need to be convertible to the alloca type so that
1779 // they are promotable.
1780 return false;
1781 }
1782 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1783 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
1784 return false;
1785 if (!I->isSplittable())
1786 return false; // Skip any unsplittable intrinsics.
1787 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1788 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1789 II->getIntrinsicID() != Intrinsic::lifetime_end)
1790 return false;
1791 } else {
1792 return false;
1793 }
1794
1795 return true;
1796}
1797
1798/// \brief Test whether the given alloca partition's integer operations can be
1799/// widened to promotable ones.
1800///
1801/// This is a quick test to check whether we can rewrite the integer loads and
1802/// stores to a particular alloca into wider loads and stores and be able to
1803/// promote the resulting alloca.
1804static bool
1805isIntegerWideningViable(const DataLayout &DL, Type *AllocaTy,
1806 uint64_t AllocBeginOffset, AllocaSlices &S,
1807 AllocaSlices::const_iterator I,
1808 AllocaSlices::const_iterator E,
1809 ArrayRef<AllocaSlices::iterator> SplitUses) {
1810 uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
1811 // Don't create integer types larger than the maximum bitwidth.
1812 if (SizeInBits > IntegerType::MAX_INT_BITS)
1813 return false;
1814
1815 // Don't try to handle allocas with bit-padding.
1816 if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
1817 return false;
1818
1819 // We need to ensure that an integer type with the appropriate bitwidth can
1820 // be converted to the alloca type, whatever that is. We don't want to force
1821 // the alloca itself to have an integer type if there is a more suitable one.
1822 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
1823 if (!canConvertValue(DL, AllocaTy, IntTy) ||
1824 !canConvertValue(DL, IntTy, AllocaTy))
1825 return false;
1826
1827 uint64_t Size = DL.getTypeStoreSize(AllocaTy);
1828
1829 // While examining uses, we ensure that the alloca has a covering load or
1830 // store. We don't want to widen the integer operations only to fail to
1831 // promote due to some other unsplittable entry (which we may make splittable
1832 // later). However, if there are only splittable uses, go ahead and assume
1833 // that we cover the alloca.
1834 bool WholeAllocaOp = (I != E) ? false : DL.isLegalInteger(SizeInBits);
1835
1836 for (; I != E; ++I)
1837 if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
1838 S, I, WholeAllocaOp))
1839 return false;
1840
1841 for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
1842 SUE = SplitUses.end();
1843 SUI != SUE; ++SUI)
1844 if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
1845 S, *SUI, WholeAllocaOp))
1846 return false;
1847
1848 return WholeAllocaOp;
1849}
1850
1851static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1852 IntegerType *Ty, uint64_t Offset,
1853 const Twine &Name) {
1854 DEBUG(dbgs() << " start: " << *V << "\n");
1855 IntegerType *IntTy = cast<IntegerType>(V->getType());
1856 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
1857 "Element extends past full value");
1858 uint64_t ShAmt = 8*Offset;
1859 if (DL.isBigEndian())
1860 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
1861 if (ShAmt) {
1862 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
1863 DEBUG(dbgs() << " shifted: " << *V << "\n");
1864 }
1865 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
1866 "Cannot extract to a larger integer!");
1867 if (Ty != IntTy) {
1868 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
1869 DEBUG(dbgs() << " trunced: " << *V << "\n");
1870 }
1871 return V;
1872}
1873
1874static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
1875 Value *V, uint64_t Offset, const Twine &Name) {
1876 IntegerType *IntTy = cast<IntegerType>(Old->getType());
1877 IntegerType *Ty = cast<IntegerType>(V->getType());
1878 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
1879 "Cannot insert a larger integer!");
1880 DEBUG(dbgs() << " start: " << *V << "\n");
1881 if (Ty != IntTy) {
1882 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
1883 DEBUG(dbgs() << " extended: " << *V << "\n");
1884 }
1885 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
1886 "Element store outside of alloca store");
1887 uint64_t ShAmt = 8*Offset;
1888 if (DL.isBigEndian())
1889 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
1890 if (ShAmt) {
1891 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
1892 DEBUG(dbgs() << " shifted: " << *V << "\n");
1893 }
1894
1895 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
1896 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
1897 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
1898 DEBUG(dbgs() << " masked: " << *Old << "\n");
1899 V = IRB.CreateOr(Old, V, Name + ".insert");
1900 DEBUG(dbgs() << " inserted: " << *V << "\n");
1901 }
1902 return V;
1903}
1904
1905static Value *extractVector(IRBuilderTy &IRB, Value *V,
1906 unsigned BeginIndex, unsigned EndIndex,
1907 const Twine &Name) {
1908 VectorType *VecTy = cast<VectorType>(V->getType());
1909 unsigned NumElements = EndIndex - BeginIndex;
1910 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
1911
1912 if (NumElements == VecTy->getNumElements())
1913 return V;
1914
1915 if (NumElements == 1) {
1916 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
1917 Name + ".extract");
1918 DEBUG(dbgs() << " extract: " << *V << "\n");
1919 return V;
1920 }
1921
1922 SmallVector<Constant*, 8> Mask;
1923 Mask.reserve(NumElements);
1924 for (unsigned i = BeginIndex; i != EndIndex; ++i)
1925 Mask.push_back(IRB.getInt32(i));
1926 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
1927 ConstantVector::get(Mask),
1928 Name + ".extract");
1929 DEBUG(dbgs() << " shuffle: " << *V << "\n");
1930 return V;
1931}
1932
1933static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
1934 unsigned BeginIndex, const Twine &Name) {
1935 VectorType *VecTy = cast<VectorType>(Old->getType());
1936 assert(VecTy && "Can only insert a vector into a vector");
1937
1938 VectorType *Ty = dyn_cast<VectorType>(V->getType());
1939 if (!Ty) {
1940 // Single element to insert.
1941 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
1942 Name + ".insert");
1943 DEBUG(dbgs() << " insert: " << *V << "\n");
1944 return V;
1945 }
1946
1947 assert(Ty->getNumElements() <= VecTy->getNumElements() &&
1948 "Too many elements!");
1949 if (Ty->getNumElements() == VecTy->getNumElements()) {
1950 assert(V->getType() == VecTy && "Vector type mismatch");
1951 return V;
1952 }
1953 unsigned EndIndex = BeginIndex + Ty->getNumElements();
1954
1955 // When inserting a smaller vector into the larger to store, we first
1956 // use a shuffle vector to widen it with undef elements, and then
1957 // a second shuffle vector to select between the loaded vector and the
1958 // incoming vector.
1959 SmallVector<Constant*, 8> Mask;
1960 Mask.reserve(VecTy->getNumElements());
1961 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
1962 if (i >= BeginIndex && i < EndIndex)
1963 Mask.push_back(IRB.getInt32(i - BeginIndex));
1964 else
1965 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
1966 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
1967 ConstantVector::get(Mask),
1968 Name + ".expand");
1969 DEBUG(dbgs() << " shuffle: " << *V << "\n");
1970
1971 Mask.clear();
1972 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
1973 Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
1974
1975 V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
1976
1977 DEBUG(dbgs() << " blend: " << *V << "\n");
1978 return V;
1979}
1980
1981namespace {
1982/// \brief Visitor to rewrite instructions using p particular slice of an alloca
1983/// to use a new alloca.
1984///
1985/// Also implements the rewriting to vector-based accesses when the partition
1986/// passes the isVectorPromotionViable predicate. Most of the rewriting logic
1987/// lives here.
1988class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
1989 // Befriend the base class so it can delegate to private visit methods.
1990 friend class llvm::InstVisitor<AllocaSliceRewriter, bool>;
1991 typedef llvm::InstVisitor<AllocaSliceRewriter, bool> Base;
1992
1993 const DataLayout &DL;
1994 AllocaSlices &S;
1995 SROA &Pass;
1996 AllocaInst &OldAI, &NewAI;
1997 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
1998 Type *NewAllocaTy;
1999
2000 // If we are rewriting an alloca partition which can be written as pure
2001 // vector operations, we stash extra information here. When VecTy is
2002 // non-null, we have some strict guarantees about the rewritten alloca:
2003 // - The new alloca is exactly the size of the vector type here.
2004 // - The accesses all either map to the entire vector or to a single
2005 // element.
2006 // - The set of accessing instructions is only one of those handled above
2007 // in isVectorPromotionViable. Generally these are the same access kinds
2008 // which are promotable via mem2reg.
2009 VectorType *VecTy;
2010 Type *ElementTy;
2011 uint64_t ElementSize;
2012
2013 // This is a convenience and flag variable that will be null unless the new
2014 // alloca's integer operations should be widened to this integer type due to
2015 // passing isIntegerWideningViable above. If it is non-null, the desired
2016 // integer type will be stored here for easy access during rewriting.
2017 IntegerType *IntTy;
2018
2019 // The original offset of the slice currently being rewritten relative to
2020 // the original alloca.
2021 uint64_t BeginOffset, EndOffset;
2022 // The new offsets of the slice currently being rewritten relative to the
2023 // original alloca.
2024 uint64_t NewBeginOffset, NewEndOffset;
2025
2026 uint64_t SliceSize;
2027 bool IsSplittable;
2028 bool IsSplit;
2029 Use *OldUse;
2030 Instruction *OldPtr;
2031
2032 // Track post-rewrite users which are PHI nodes and Selects.
2033 SmallPtrSetImpl<PHINode *> &PHIUsers;
2034 SmallPtrSetImpl<SelectInst *> &SelectUsers;
2035
2036 // Utility IR builder, whose name prefix is setup for each visited use, and
2037 // the insertion point is set to point to the user.
2038 IRBuilderTy IRB;
2039
2040public:
2041 AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &S, SROA &Pass,
2042 AllocaInst &OldAI, AllocaInst &NewAI,
2043 uint64_t NewAllocaBeginOffset,
2044 uint64_t NewAllocaEndOffset, bool IsVectorPromotable,
2045 bool IsIntegerPromotable,
2046 SmallPtrSetImpl<PHINode *> &PHIUsers,
2047 SmallPtrSetImpl<SelectInst *> &SelectUsers)
2048 : DL(DL), S(S), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
2049 NewAllocaBeginOffset(NewAllocaBeginOffset),
2050 NewAllocaEndOffset(NewAllocaEndOffset),
2051 NewAllocaTy(NewAI.getAllocatedType()),
2052 VecTy(IsVectorPromotable ? cast<VectorType>(NewAllocaTy) : nullptr),
2053 ElementTy(VecTy ? VecTy->getElementType() : nullptr),
2054 ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
2055 IntTy(IsIntegerPromotable
2056 ? Type::getIntNTy(
2057 NewAI.getContext(),
2058 DL.getTypeSizeInBits(NewAI.getAllocatedType()))
2059 : nullptr),
2060 BeginOffset(), EndOffset(), IsSplittable(), IsSplit(), OldUse(),
2061 OldPtr(), PHIUsers(PHIUsers), SelectUsers(SelectUsers),
2062 IRB(NewAI.getContext(), ConstantFolder()) {
2063 if (VecTy) {
2064 assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
2065 "Only multiple-of-8 sized vector elements are viable");
2066 ++NumVectorized;
2067 }
2068 assert((!IsVectorPromotable && !IsIntegerPromotable) ||
2069 IsVectorPromotable != IsIntegerPromotable);
2070 }
2071
2072 bool visit(AllocaSlices::const_iterator I) {
2073 bool CanSROA = true;
2074 BeginOffset = I->beginOffset();
2075 EndOffset = I->endOffset();
2076 IsSplittable = I->isSplittable();
2077 IsSplit =
2078 BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
2079
2080 // Compute the intersecting offset range.
2081 assert(BeginOffset < NewAllocaEndOffset);
2082 assert(EndOffset > NewAllocaBeginOffset);
2083 NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2084 NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2085
2086 SliceSize = NewEndOffset - NewBeginOffset;
2087
2088 OldUse = I->getUse();
2089 OldPtr = cast<Instruction>(OldUse->get());
2090
2091 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2092 IRB.SetInsertPoint(OldUserI);
2093 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2094 IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
2095
2096 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2097 if (VecTy || IntTy)
2098 assert(CanSROA);
2099 return CanSROA;
2100 }
2101
2102private:
2103 // Make sure the other visit overloads are visible.
2104 using Base::visit;
2105
2106 // Every instruction which can end up as a user must have a rewrite rule.
2107 bool visitInstruction(Instruction &I) {
2108 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2109 llvm_unreachable("No rewrite rule for this instruction!");
2110 }
2111
2112 Value *getNewAllocaSlicePtr(IRBuilderTy &IRB, Type *PointerTy) {
2113 // Note that the offset computation can use BeginOffset or NewBeginOffset
2114 // interchangeably for unsplit slices.
2115 assert(IsSplit || BeginOffset == NewBeginOffset);
2116 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2117
2118#ifndef NDEBUG
2119 StringRef OldName = OldPtr->getName();
2120 // Skip through the last '.sroa.' component of the name.
2121 size_t LastSROAPrefix = OldName.rfind(".sroa.");
2122 if (LastSROAPrefix != StringRef::npos) {
2123 OldName = OldName.substr(LastSROAPrefix + strlen(".sroa."));
2124 // Look for an SROA slice index.
2125 size_t IndexEnd = OldName.find_first_not_of("0123456789");
2126 if (IndexEnd != StringRef::npos && OldName[IndexEnd] == '.') {
2127 // Strip the index and look for the offset.
2128 OldName = OldName.substr(IndexEnd + 1);
2129 size_t OffsetEnd = OldName.find_first_not_of("0123456789");
2130 if (OffsetEnd != StringRef::npos && OldName[OffsetEnd] == '.')
2131 // Strip the offset.
2132 OldName = OldName.substr(OffsetEnd + 1);
2133 }
2134 }
2135 // Strip any SROA suffixes as well.
2136 OldName = OldName.substr(0, OldName.find(".sroa_"));
2137#endif
2138
2139 return getAdjustedPtr(IRB, DL, &NewAI,
2140 APInt(DL.getPointerSizeInBits(), Offset), PointerTy,
2141#ifndef NDEBUG
2142 Twine(OldName) + "."
2143#else
2144 Twine()
2145#endif
2146 );
2147 }
2148
2149 /// \brief Compute suitable alignment to access this slice of the *new* alloca.
2150 ///
2151 /// You can optionally pass a type to this routine and if that type's ABI
2152 /// alignment is itself suitable, this will return zero.
2153 unsigned getSliceAlign(Type *Ty = nullptr) {
2154 unsigned NewAIAlign = NewAI.getAlignment();
2155 if (!NewAIAlign)
2156 NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
2157 unsigned Align = MinAlign(NewAIAlign, NewBeginOffset - NewAllocaBeginOffset);
2158 return (Ty && Align == DL.getABITypeAlignment(Ty)) ? 0 : Align;
2159 }
2160
2161 unsigned getIndex(uint64_t Offset) {
2162 assert(VecTy && "Can only call getIndex when rewriting a vector");
2163 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2164 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2165 uint32_t Index = RelOffset / ElementSize;
2166 assert(Index * ElementSize == RelOffset);
2167 return Index;
2168 }
2169
2170 void deleteIfTriviallyDead(Value *V) {
2171 Instruction *I = cast<Instruction>(V);
2172 if (isInstructionTriviallyDead(I))
2173 Pass.DeadInsts.insert(I);
2174 }
2175
2176 Value *rewriteVectorizedLoadInst() {
2177 unsigned BeginIndex = getIndex(NewBeginOffset);
2178 unsigned EndIndex = getIndex(NewEndOffset);
2179 assert(EndIndex > BeginIndex && "Empty vector!");
2180
2181 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2182 "load");
2183 return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2184 }
2185
2186 Value *rewriteIntegerLoad(LoadInst &LI) {
2187 assert(IntTy && "We cannot insert an integer to the alloca");
2188 assert(!LI.isVolatile());
2189 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2190 "load");
2191 V = convertValue(DL, IRB, V, IntTy);
2192 assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2193 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2194 if (Offset > 0 || NewEndOffset < NewAllocaEndOffset)
2195 V = extractInteger(DL, IRB, V, cast<IntegerType>(LI.getType()), Offset,
2196 "extract");
2197 return V;
2198 }
2199
2200 bool visitLoadInst(LoadInst &LI) {
2201 DEBUG(dbgs() << " original: " << LI << "\n");
2202 Value *OldOp = LI.getOperand(0);
2203 assert(OldOp == OldPtr);
2204
2205 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), SliceSize * 8)
2206 : LI.getType();
2207 bool IsPtrAdjusted = false;
2208 Value *V;
2209 if (VecTy) {
2210 V = rewriteVectorizedLoadInst();
2211 } else if (IntTy && LI.getType()->isIntegerTy()) {
2212 V = rewriteIntegerLoad(LI);
2213 } else if (NewBeginOffset == NewAllocaBeginOffset &&
2214 canConvertValue(DL, NewAllocaTy, LI.getType())) {
2215 V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2216 LI.isVolatile(), LI.getName());
2217 } else {
2218 Type *LTy = TargetTy->getPointerTo();
2219 V = IRB.CreateAlignedLoad(getNewAllocaSlicePtr(IRB, LTy),
2220 getSliceAlign(TargetTy), LI.isVolatile(),
2221 LI.getName());
2222 IsPtrAdjusted = true;
2223 }
2224 V = convertValue(DL, IRB, V, TargetTy);
2225
2226 if (IsSplit) {
2227 assert(!LI.isVolatile());
2228 assert(LI.getType()->isIntegerTy() &&
2229 "Only integer type loads and stores are split");
2230 assert(SliceSize < DL.getTypeStoreSize(LI.getType()) &&
2231 "Split load isn't smaller than original load");
2232 assert(LI.getType()->getIntegerBitWidth() ==
2233 DL.getTypeStoreSizeInBits(LI.getType()) &&
2234 "Non-byte-multiple bit width");
2235 // Move the insertion point just past the load so that we can refer to it.
2236 IRB.SetInsertPoint(std::next(BasicBlock::iterator(&LI)));
2237 // Create a placeholder value with the same type as LI to use as the
2238 // basis for the new value. This allows us to replace the uses of LI with
2239 // the computed value, and then replace the placeholder with LI, leaving
2240 // LI only used for this computation.
2241 Value *Placeholder
2242 = new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
2243 V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset,
2244 "insert");
2245 LI.replaceAllUsesWith(V);
2246 Placeholder->replaceAllUsesWith(&LI);
2247 delete Placeholder;
2248 } else {
2249 LI.replaceAllUsesWith(V);
2250 }
2251
2252 Pass.DeadInsts.insert(&LI);
2253 deleteIfTriviallyDead(OldOp);
2254 DEBUG(dbgs() << " to: " << *V << "\n");
2255 return !LI.isVolatile() && !IsPtrAdjusted;
2256 }
2257
2258 bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp) {
2259 if (V->getType() != VecTy) {
2260 unsigned BeginIndex = getIndex(NewBeginOffset);
2261 unsigned EndIndex = getIndex(NewEndOffset);
2262 assert(EndIndex > BeginIndex && "Empty vector!");
2263 unsigned NumElements = EndIndex - BeginIndex;
2264 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2265 Type *SliceTy =
2266 (NumElements == 1) ? ElementTy
2267 : VectorType::get(ElementTy, NumElements);
2268 if (V->getType() != SliceTy)
2269 V = convertValue(DL, IRB, V, SliceTy);
2270
2271 // Mix in the existing elements.
2272 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2273 "load");
2274 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2275 }
2276 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2277 Pass.DeadInsts.insert(&SI);
2278
2279 (void)Store;
2280 DEBUG(dbgs() << " to: " << *Store << "\n");
2281 return true;
2282 }
2283
2284 bool rewriteIntegerStore(Value *V, StoreInst &SI) {
2285 assert(IntTy && "We cannot extract an integer from the alloca");
2286 assert(!SI.isVolatile());
2287 if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2288 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2289 "oldload");
2290 Old = convertValue(DL, IRB, Old, IntTy);
2291 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2292 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2293 V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset,
2294 "insert");
2295 }
2296 V = convertValue(DL, IRB, V, NewAllocaTy);
2297 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2298 Pass.DeadInsts.insert(&SI);
2299 (void)Store;
2300 DEBUG(dbgs() << " to: " << *Store << "\n");
2301 return true;
2302 }
2303
2304 bool visitStoreInst(StoreInst &SI) {
2305 DEBUG(dbgs() << " original: " << SI << "\n");
2306 Value *OldOp = SI.getOperand(1);
2307 assert(OldOp == OldPtr);
2308
2309 Value *V = SI.getValueOperand();
2310
2311 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2312 // alloca that should be re-examined after promoting this alloca.
2313 if (V->getType()->isPointerTy())
2314 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2315 Pass.PostPromotionWorklist.insert(AI);
2316
2317 if (SliceSize < DL.getTypeStoreSize(V->getType())) {
2318 assert(!SI.isVolatile());
2319 assert(V->getType()->isIntegerTy() &&
2320 "Only integer type loads and stores are split");
2321 assert(V->getType()->getIntegerBitWidth() ==
2322 DL.getTypeStoreSizeInBits(V->getType()) &&
2323 "Non-byte-multiple bit width");
2324 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), SliceSize * 8);
2325 V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset,
2326 "extract");
2327 }
2328
2329 if (VecTy)
2330 return rewriteVectorizedStoreInst(V, SI, OldOp);
2331 if (IntTy && V->getType()->isIntegerTy())
2332 return rewriteIntegerStore(V, SI);
2333
2334 StoreInst *NewSI;
2335 if (NewBeginOffset == NewAllocaBeginOffset &&
2336 NewEndOffset == NewAllocaEndOffset &&
2337 canConvertValue(DL, V->getType(), NewAllocaTy)) {
2338 V = convertValue(DL, IRB, V, NewAllocaTy);
2339 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2340 SI.isVolatile());
2341 } else {
2342 Value *NewPtr = getNewAllocaSlicePtr(IRB, V->getType()->getPointerTo());
2343 NewSI = IRB.CreateAlignedStore(V, NewPtr, getSliceAlign(V->getType()),
2344 SI.isVolatile());
2345 }
2346 (void)NewSI;
2347 Pass.DeadInsts.insert(&SI);
2348 deleteIfTriviallyDead(OldOp);
2349
2350 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2351 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2352 }
2353
2354 /// \brief Compute an integer value from splatting an i8 across the given
2355 /// number of bytes.
2356 ///
2357 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2358 /// call this routine.
2359 /// FIXME: Heed the advice above.
2360 ///
2361 /// \param V The i8 value to splat.
2362 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2363 Value *getIntegerSplat(Value *V, unsigned Size) {
2364 assert(Size > 0 && "Expected a positive number of bytes.");
2365 IntegerType *VTy = cast<IntegerType>(V->getType());
2366 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2367 if (Size == 1)
2368 return V;
2369
2370 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size*8);
2371 V = IRB.CreateMul(IRB.CreateZExt(V, SplatIntTy, "zext"),
2372 ConstantExpr::getUDiv(
2373 Constant::getAllOnesValue(SplatIntTy),
2374 ConstantExpr::getZExt(
2375 Constant::getAllOnesValue(V->getType()),
2376 SplatIntTy)),
2377 "isplat");
2378 return V;
2379 }
2380
2381 /// \brief Compute a vector splat for a given element value.
2382 Value *getVectorSplat(Value *V, unsigned NumElements) {
2383 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2384 DEBUG(dbgs() << " splat: " << *V << "\n");
2385 return V;
2386 }
2387
2388 bool visitMemSetInst(MemSetInst &II) {
2389 DEBUG(dbgs() << " original: " << II << "\n");
2390 assert(II.getRawDest() == OldPtr);
2391
2392 // If the memset has a variable size, it cannot be split, just adjust the
2393 // pointer to the new alloca.
2394 if (!isa<Constant>(II.getLength())) {
2395 assert(!IsSplit);
2396 assert(NewBeginOffset == BeginOffset);
2397 II.setDest(getNewAllocaSlicePtr(IRB, OldPtr->getType()));
2398 Type *CstTy = II.getAlignmentCst()->getType();
2399 II.setAlignment(ConstantInt::get(CstTy, getSliceAlign()));
2400
2401 deleteIfTriviallyDead(OldPtr);
2402 return false;
2403 }
2404
2405 // Record this instruction for deletion.
2406 Pass.DeadInsts.insert(&II);
2407
2408 Type *AllocaTy = NewAI.getAllocatedType();
2409 Type *ScalarTy = AllocaTy->getScalarType();
2410
2411 // If this doesn't map cleanly onto the alloca type, and that type isn't
2412 // a single value type, just emit a memset.
2413 if (!VecTy && !IntTy &&
2414 (BeginOffset > NewAllocaBeginOffset ||
2415 EndOffset < NewAllocaEndOffset ||
2416 !AllocaTy->isSingleValueType() ||
2417 !DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy)) ||
2418 DL.getTypeSizeInBits(ScalarTy)%8 != 0)) {
2419 Type *SizeTy = II.getLength()->getType();
2420 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2421 CallInst *New = IRB.CreateMemSet(
2422 getNewAllocaSlicePtr(IRB, OldPtr->getType()), II.getValue(), Size,
2423 getSliceAlign(), II.isVolatile());
2424 (void)New;
2425 DEBUG(dbgs() << " to: " << *New << "\n");
2426 return false;
2427 }
2428
2429 // If we can represent this as a simple value, we have to build the actual
2430 // value to store, which requires expanding the byte present in memset to
2431 // a sensible representation for the alloca type. This is essentially
2432 // splatting the byte to a sufficiently wide integer, splatting it across
2433 // any desired vector width, and bitcasting to the final type.
2434 Value *V;
2435
2436 if (VecTy) {
2437 // If this is a memset of a vectorized alloca, insert it.
2438 assert(ElementTy == ScalarTy);
2439
2440 unsigned BeginIndex = getIndex(NewBeginOffset);
2441 unsigned EndIndex = getIndex(NewEndOffset);
2442 assert(EndIndex > BeginIndex && "Empty vector!");
2443 unsigned NumElements = EndIndex - BeginIndex;
2444 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2445
2446 Value *Splat =
2447 getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
2448 Splat = convertValue(DL, IRB, Splat, ElementTy);
2449 if (NumElements > 1)
2450 Splat = getVectorSplat(Splat, NumElements);
2451
2452 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2453 "oldload");
2454 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2455 } else if (IntTy) {
2456 // If this is a memset on an alloca where we can widen stores, insert the
2457 // set integer.
2458 assert(!II.isVolatile());
2459
2460 uint64_t Size = NewEndOffset - NewBeginOffset;
2461 V = getIntegerSplat(II.getValue(), Size);
2462
2463 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2464 EndOffset != NewAllocaBeginOffset)) {
2465 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2466 "oldload");
2467 Old = convertValue(DL, IRB, Old, IntTy);
2468 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2469 V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2470 } else {
2471 assert(V->getType() == IntTy &&
2472 "Wrong type for an alloca wide integer!");
2473 }
2474 V = convertValue(DL, IRB, V, AllocaTy);
2475 } else {
2476 // Established these invariants above.
2477 assert(NewBeginOffset == NewAllocaBeginOffset);
2478 assert(NewEndOffset == NewAllocaEndOffset);
2479
2480 V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
2481 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2482 V = getVectorSplat(V, AllocaVecTy->getNumElements());
2483
2484 V = convertValue(DL, IRB, V, AllocaTy);
2485 }
2486
2487 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2488 II.isVolatile());
2489 (void)New;
2490 DEBUG(dbgs() << " to: " << *New << "\n");
2491 return !II.isVolatile();
2492 }
2493
2494 bool visitMemTransferInst(MemTransferInst &II) {
2495 // Rewriting of memory transfer instructions can be a bit tricky. We break
2496 // them into two categories: split intrinsics and unsplit intrinsics.
2497
2498 DEBUG(dbgs() << " original: " << II << "\n");
2499
2500 bool IsDest = &II.getRawDestUse() == OldUse;
2501 assert((IsDest && II.getRawDest() == OldPtr) ||
2502 (!IsDest && II.getRawSource() == OldPtr));
2503
2504 unsigned SliceAlign = getSliceAlign();
2505
2506 // For unsplit intrinsics, we simply modify the source and destination
2507 // pointers in place. This isn't just an optimization, it is a matter of
2508 // correctness. With unsplit intrinsics we may be dealing with transfers
2509 // within a single alloca before SROA ran, or with transfers that have
2510 // a variable length. We may also be dealing with memmove instead of
2511 // memcpy, and so simply updating the pointers is the necessary for us to
2512 // update both source and dest of a single call.
2513 if (!IsSplittable) {
2514 Value *AdjustedPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2515 if (IsDest)
2516 II.setDest(AdjustedPtr);
2517 else
2518 II.setSource(AdjustedPtr);
2519
2520 if (II.getAlignment() > SliceAlign) {
2521 Type *CstTy = II.getAlignmentCst()->getType();
2522 II.setAlignment(
2523 ConstantInt::get(CstTy, MinAlign(II.getAlignment(), SliceAlign)));
2524 }
2525
2526 DEBUG(dbgs() << " to: " << II << "\n");
2527 deleteIfTriviallyDead(OldPtr);
2528 return false;
2529 }
2530 // For split transfer intrinsics we have an incredibly useful assurance:
2531 // the source and destination do not reside within the same alloca, and at
2532 // least one of them does not escape. This means that we can replace
2533 // memmove with memcpy, and we don't need to worry about all manner of
2534 // downsides to splitting and transforming the operations.
2535
2536 // If this doesn't map cleanly onto the alloca type, and that type isn't
2537 // a single value type, just emit a memcpy.
2538 bool EmitMemCpy
2539 = !VecTy && !IntTy && (BeginOffset > NewAllocaBeginOffset ||
2540 EndOffset < NewAllocaEndOffset ||
2541 !NewAI.getAllocatedType()->isSingleValueType());
2542
2543 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2544 // size hasn't been shrunk based on analysis of the viable range, this is
2545 // a no-op.
2546 if (EmitMemCpy && &OldAI == &NewAI) {
2547 // Ensure the start lines up.
2548 assert(NewBeginOffset == BeginOffset);
2549
2550 // Rewrite the size as needed.
2551 if (NewEndOffset != EndOffset)
2552 II.setLength(ConstantInt::get(II.getLength()->getType(),
2553 NewEndOffset - NewBeginOffset));
2554 return false;
2555 }
2556 // Record this instruction for deletion.
2557 Pass.DeadInsts.insert(&II);
2558
2559 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2560 // alloca that should be re-examined after rewriting this instruction.
2561 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2562 if (AllocaInst *AI
2563 = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
2564 assert(AI != &OldAI && AI != &NewAI &&
2565 "Splittable transfers cannot reach the same alloca on both ends.");
2566 Pass.Worklist.insert(AI);
2567 }
2568
2569 Type *OtherPtrTy = OtherPtr->getType();
2570 unsigned OtherAS = OtherPtrTy->getPointerAddressSpace();
2571
2572 // Compute the relative offset for the other pointer within the transfer.
2573 unsigned IntPtrWidth = DL.getPointerSizeInBits(OtherAS);
2574 APInt OtherOffset(IntPtrWidth, NewBeginOffset - BeginOffset);
2575 unsigned OtherAlign = MinAlign(II.getAlignment() ? II.getAlignment() : 1,
2576 OtherOffset.zextOrTrunc(64).getZExtValue());
2577
2578 if (EmitMemCpy) {
2579 // Compute the other pointer, folding as much as possible to produce
2580 // a single, simple GEP in most cases.
2581 OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2582 OtherPtr->getName() + ".");
2583
2584 Value *OurPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2585 Type *SizeTy = II.getLength()->getType();
2586 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2587
2588 CallInst *New = IRB.CreateMemCpy(
2589 IsDest ? OurPtr : OtherPtr, IsDest ? OtherPtr : OurPtr, Size,
2590 MinAlign(SliceAlign, OtherAlign), II.isVolatile());
2591 (void)New;
2592 DEBUG(dbgs() << " to: " << *New << "\n");
2593 return false;
2594 }
2595
2596 bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
2597 NewEndOffset == NewAllocaEndOffset;
2598 uint64_t Size = NewEndOffset - NewBeginOffset;
2599 unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
2600 unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
2601 unsigned NumElements = EndIndex - BeginIndex;
2602 IntegerType *SubIntTy
2603 = IntTy ? Type::getIntNTy(IntTy->getContext(), Size*8) : nullptr;
2604
2605 // Reset the other pointer type to match the register type we're going to
2606 // use, but using the address space of the original other pointer.
2607 if (VecTy && !IsWholeAlloca) {
2608 if (NumElements == 1)
2609 OtherPtrTy = VecTy->getElementType();
2610 else
2611 OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2612
2613 OtherPtrTy = OtherPtrTy->getPointerTo(OtherAS);
2614 } else if (IntTy && !IsWholeAlloca) {
2615 OtherPtrTy = SubIntTy->getPointerTo(OtherAS);
2616 } else {
2617 OtherPtrTy = NewAllocaTy->getPointerTo(OtherAS);
2618 }
2619
2620 Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2621 OtherPtr->getName() + ".");
2622 unsigned SrcAlign = OtherAlign;
2623 Value *DstPtr = &NewAI;
2624 unsigned DstAlign = SliceAlign;
2625 if (!IsDest) {
2626 std::swap(SrcPtr, DstPtr);
2627 std::swap(SrcAlign, DstAlign);
2628 }
2629
2630 Value *Src;
2631 if (VecTy && !IsWholeAlloca && !IsDest) {
2632 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2633 "load");
2634 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
2635 } else if (IntTy && !IsWholeAlloca && !IsDest) {
2636 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2637 "load");
2638 Src = convertValue(DL, IRB, Src, IntTy);
2639 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2640 Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
2641 } else {
2642 Src = IRB.CreateAlignedLoad(SrcPtr, SrcAlign, II.isVolatile(),
2643 "copyload");
2644 }
2645
2646 if (VecTy && !IsWholeAlloca && IsDest) {
2647 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2648 "oldload");
2649 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
2650 } else if (IntTy && !IsWholeAlloca && IsDest) {
2651 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2652 "oldload");
2653 Old = convertValue(DL, IRB, Old, IntTy);
2654 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2655 Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
2656 Src = convertValue(DL, IRB, Src, NewAllocaTy);
2657 }
2658
2659 StoreInst *Store = cast<StoreInst>(
2660 IRB.CreateAlignedStore(Src, DstPtr, DstAlign, II.isVolatile()));
2661 (void)Store;
2662 DEBUG(dbgs() << " to: " << *Store << "\n");
2663 return !II.isVolatile();
2664 }
2665
2666 bool visitIntrinsicInst(IntrinsicInst &II) {
2667 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2668 II.getIntrinsicID() == Intrinsic::lifetime_end);
2669 DEBUG(dbgs() << " original: " << II << "\n");
2670 assert(II.getArgOperand(1) == OldPtr);
2671
2672 // Record this instruction for deletion.
2673 Pass.DeadInsts.insert(&II);
2674
2675 ConstantInt *Size
2676 = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2677 NewEndOffset - NewBeginOffset);
2678 Value *Ptr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2679 Value *New;
2680 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2681 New = IRB.CreateLifetimeStart(Ptr, Size);
2682 else
2683 New = IRB.CreateLifetimeEnd(Ptr, Size);
2684
2685 (void)New;
2686 DEBUG(dbgs() << " to: " << *New << "\n");
2687 return true;
2688 }
2689
2690 bool visitPHINode(PHINode &PN) {
2691 DEBUG(dbgs() << " original: " << PN << "\n");
2692 assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
2693 assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
2694
2695 // We would like to compute a new pointer in only one place, but have it be
2696 // as local as possible to the PHI. To do that, we re-use the location of
2697 // the old pointer, which necessarily must be in the right position to
2698 // dominate the PHI.
2699 IRBuilderTy PtrBuilder(IRB);
|
2701 PtrBuilder.SetCurrentDebugLocation(OldPtr->getDebugLoc());
2702
2703 Value *NewPtr = getNewAllocaSlicePtr(PtrBuilder, OldPtr->getType());
2704 // Replace the operands which were using the old pointer.
2705 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
2706
2707 DEBUG(dbgs() << " to: " << PN << "\n");
2708 deleteIfTriviallyDead(OldPtr);
2709
2710 // PHIs can't be promoted on their own, but often can be speculated. We
2711 // check the speculation outside of the rewriter so that we see the
2712 // fully-rewritten alloca.
2713 PHIUsers.insert(&PN);
2714 return true;
2715 }
2716
2717 bool visitSelectInst(SelectInst &SI) {
2718 DEBUG(dbgs() << " original: " << SI << "\n");
2719 assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
2720 "Pointer isn't an operand!");
2721 assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
2722 assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
2723
2724 Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2725 // Replace the operands which were using the old pointer.
2726 if (SI.getOperand(1) == OldPtr)
2727 SI.setOperand(1, NewPtr);
2728 if (SI.getOperand(2) == OldPtr)
2729 SI.setOperand(2, NewPtr);
2730
2731 DEBUG(dbgs() << " to: " << SI << "\n");
2732 deleteIfTriviallyDead(OldPtr);
2733
2734 // Selects can't be promoted on their own, but often can be speculated. We
2735 // check the speculation outside of the rewriter so that we see the
2736 // fully-rewritten alloca.
2737 SelectUsers.insert(&SI);
2738 return true;
2739 }
2740
2741};
2742}
2743
2744namespace {
2745/// \brief Visitor to rewrite aggregate loads and stores as scalar.
2746///
2747/// This pass aggressively rewrites all aggregate loads and stores on
2748/// a particular pointer (or any pointer derived from it which we can identify)
2749/// with scalar loads and stores.
2750class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
2751 // Befriend the base class so it can delegate to private visit methods.
2752 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
2753
2754 const DataLayout &DL;
2755
2756 /// Queue of pointer uses to analyze and potentially rewrite.
2757 SmallVector<Use *, 8> Queue;
2758
2759 /// Set to prevent us from cycling with phi nodes and loops.
2760 SmallPtrSet<User *, 8> Visited;
2761
2762 /// The current pointer use being rewritten. This is used to dig up the used
2763 /// value (as opposed to the user).
2764 Use *U;
2765
2766public:
2767 AggLoadStoreRewriter(const DataLayout &DL) : DL(DL) {}
2768
2769 /// Rewrite loads and stores through a pointer and all pointers derived from
2770 /// it.
2771 bool rewrite(Instruction &I) {
2772 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
2773 enqueueUsers(I);
2774 bool Changed = false;
2775 while (!Queue.empty()) {
2776 U = Queue.pop_back_val();
2777 Changed |= visit(cast<Instruction>(U->getUser()));
2778 }
2779 return Changed;
2780 }
2781
2782private:
2783 /// Enqueue all the users of the given instruction for further processing.
2784 /// This uses a set to de-duplicate users.
2785 void enqueueUsers(Instruction &I) {
2786 for (Use &U : I.uses())
2787 if (Visited.insert(U.getUser()))
2788 Queue.push_back(&U);
2789 }
2790
2791 // Conservative default is to not rewrite anything.
2792 bool visitInstruction(Instruction &I) { return false; }
2793
2794 /// \brief Generic recursive split emission class.
2795 template <typename Derived>
2796 class OpSplitter {
2797 protected:
2798 /// The builder used to form new instructions.
2799 IRBuilderTy IRB;
2800 /// The indices which to be used with insert- or extractvalue to select the
2801 /// appropriate value within the aggregate.
2802 SmallVector<unsigned, 4> Indices;
2803 /// The indices to a GEP instruction which will move Ptr to the correct slot
2804 /// within the aggregate.
2805 SmallVector<Value *, 4> GEPIndices;
2806 /// The base pointer of the original op, used as a base for GEPing the
2807 /// split operations.
2808 Value *Ptr;
2809
2810 /// Initialize the splitter with an insertion point, Ptr and start with a
2811 /// single zero GEP index.
2812 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
2813 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
2814
2815 public:
2816 /// \brief Generic recursive split emission routine.
2817 ///
2818 /// This method recursively splits an aggregate op (load or store) into
2819 /// scalar or vector ops. It splits recursively until it hits a single value
2820 /// and emits that single value operation via the template argument.
2821 ///
2822 /// The logic of this routine relies on GEPs and insertvalue and
2823 /// extractvalue all operating with the same fundamental index list, merely
2824 /// formatted differently (GEPs need actual values).
2825 ///
2826 /// \param Ty The type being split recursively into smaller ops.
2827 /// \param Agg The aggregate value being built up or stored, depending on
2828 /// whether this is splitting a load or a store respectively.
2829 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
2830 if (Ty->isSingleValueType())
2831 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
2832
2833 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2834 unsigned OldSize = Indices.size();
2835 (void)OldSize;
2836 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
2837 ++Idx) {
2838 assert(Indices.size() == OldSize && "Did not return to the old size");
2839 Indices.push_back(Idx);
2840 GEPIndices.push_back(IRB.getInt32(Idx));
2841 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
2842 GEPIndices.pop_back();
2843 Indices.pop_back();
2844 }
2845 return;
2846 }
2847
2848 if (StructType *STy = dyn_cast<StructType>(Ty)) {
2849 unsigned OldSize = Indices.size();
2850 (void)OldSize;
2851 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
2852 ++Idx) {
2853 assert(Indices.size() == OldSize && "Did not return to the old size");
2854 Indices.push_back(Idx);
2855 GEPIndices.push_back(IRB.getInt32(Idx));
2856 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
2857 GEPIndices.pop_back();
2858 Indices.pop_back();
2859 }
2860 return;
2861 }
2862
2863 llvm_unreachable("Only arrays and structs are aggregate loadable types");
2864 }
2865 };
2866
2867 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
2868 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2869 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
2870
2871 /// Emit a leaf load of a single value. This is called at the leaves of the
2872 /// recursive emission to actually load values.
2873 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2874 assert(Ty->isSingleValueType());
2875 // Load the single value and insert it using the indices.
2876 Value *GEP = IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep");
2877 Value *Load = IRB.CreateLoad(GEP, Name + ".load");
2878 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
2879 DEBUG(dbgs() << " to: " << *Load << "\n");
2880 }
2881 };
2882
2883 bool visitLoadInst(LoadInst &LI) {
2884 assert(LI.getPointerOperand() == *U);
2885 if (!LI.isSimple() || LI.getType()->isSingleValueType())
2886 return false;
2887
2888 // We have an aggregate being loaded, split it apart.
2889 DEBUG(dbgs() << " original: " << LI << "\n");
2890 LoadOpSplitter Splitter(&LI, *U);
2891 Value *V = UndefValue::get(LI.getType());
2892 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
2893 LI.replaceAllUsesWith(V);
2894 LI.eraseFromParent();
2895 return true;
2896 }
2897
2898 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
2899 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2900 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
2901
2902 /// Emit a leaf store of a single value. This is called at the leaves of the
2903 /// recursive emission to actually produce stores.
2904 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2905 assert(Ty->isSingleValueType());
2906 // Extract the single value and store it using the indices.
2907 Value *Store = IRB.CreateStore(
2908 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
2909 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
2910 (void)Store;
2911 DEBUG(dbgs() << " to: " << *Store << "\n");
2912 }
2913 };
2914
2915 bool visitStoreInst(StoreInst &SI) {
2916 if (!SI.isSimple() || SI.getPointerOperand() != *U)
2917 return false;
2918 Value *V = SI.getValueOperand();
2919 if (V->getType()->isSingleValueType())
2920 return false;
2921
2922 // We have an aggregate being stored, split it apart.
2923 DEBUG(dbgs() << " original: " << SI << "\n");
2924 StoreOpSplitter Splitter(&SI, *U);
2925 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
2926 SI.eraseFromParent();
2927 return true;
2928 }
2929
2930 bool visitBitCastInst(BitCastInst &BC) {
2931 enqueueUsers(BC);
2932 return false;
2933 }
2934
2935 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
2936 enqueueUsers(GEPI);
2937 return false;
2938 }
2939
2940 bool visitPHINode(PHINode &PN) {
2941 enqueueUsers(PN);
2942 return false;
2943 }
2944
2945 bool visitSelectInst(SelectInst &SI) {
2946 enqueueUsers(SI);
2947 return false;
2948 }
2949};
2950}
2951
2952/// \brief Strip aggregate type wrapping.
2953///
2954/// This removes no-op aggregate types wrapping an underlying type. It will
2955/// strip as many layers of types as it can without changing either the type
2956/// size or the allocated size.
2957static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
2958 if (Ty->isSingleValueType())
2959 return Ty;
2960
2961 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
2962 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
2963
2964 Type *InnerTy;
2965 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
2966 InnerTy = ArrTy->getElementType();
2967 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
2968 const StructLayout *SL = DL.getStructLayout(STy);
2969 unsigned Index = SL->getElementContainingOffset(0);
2970 InnerTy = STy->getElementType(Index);
2971 } else {
2972 return Ty;
2973 }
2974
2975 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
2976 TypeSize > DL.getTypeSizeInBits(InnerTy))
2977 return Ty;
2978
2979 return stripAggregateTypeWrapping(DL, InnerTy);
2980}
2981
2982/// \brief Try to find a partition of the aggregate type passed in for a given
2983/// offset and size.
2984///
2985/// This recurses through the aggregate type and tries to compute a subtype
2986/// based on the offset and size. When the offset and size span a sub-section
2987/// of an array, it will even compute a new array type for that sub-section,
2988/// and the same for structs.
2989///
2990/// Note that this routine is very strict and tries to find a partition of the
2991/// type which produces the *exact* right offset and size. It is not forgiving
2992/// when the size or offset cause either end of type-based partition to be off.
2993/// Also, this is a best-effort routine. It is reasonable to give up and not
2994/// return a type if necessary.
2995static Type *getTypePartition(const DataLayout &DL, Type *Ty,
2996 uint64_t Offset, uint64_t Size) {
2997 if (Offset == 0 && DL.getTypeAllocSize(Ty) == Size)
2998 return stripAggregateTypeWrapping(DL, Ty);
2999 if (Offset > DL.getTypeAllocSize(Ty) ||
3000 (DL.getTypeAllocSize(Ty) - Offset) < Size)
3001 return nullptr;
3002
3003 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3004 // We can't partition pointers...
3005 if (SeqTy->isPointerTy())
3006 return nullptr;
3007
3008 Type *ElementTy = SeqTy->getElementType();
3009 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3010 uint64_t NumSkippedElements = Offset / ElementSize;
3011 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy)) {
3012 if (NumSkippedElements >= ArrTy->getNumElements())
3013 return nullptr;
3014 } else if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy)) {
3015 if (NumSkippedElements >= VecTy->getNumElements())
3016 return nullptr;
3017 }
3018 Offset -= NumSkippedElements * ElementSize;
3019
3020 // First check if we need to recurse.
3021 if (Offset > 0 || Size < ElementSize) {
3022 // Bail if the partition ends in a different array element.
3023 if ((Offset + Size) > ElementSize)
3024 return nullptr;
3025 // Recurse through the element type trying to peel off offset bytes.
3026 return getTypePartition(DL, ElementTy, Offset, Size);
3027 }
3028 assert(Offset == 0);
3029
3030 if (Size == ElementSize)
3031 return stripAggregateTypeWrapping(DL, ElementTy);
3032 assert(Size > ElementSize);
3033 uint64_t NumElements = Size / ElementSize;
3034 if (NumElements * ElementSize != Size)
3035 return nullptr;
3036 return ArrayType::get(ElementTy, NumElements);
3037 }
3038
3039 StructType *STy = dyn_cast<StructType>(Ty);
3040 if (!STy)
3041 return nullptr;
3042
3043 const StructLayout *SL = DL.getStructLayout(STy);
3044 if (Offset >= SL->getSizeInBytes())
3045 return nullptr;
3046 uint64_t EndOffset = Offset + Size;
3047 if (EndOffset > SL->getSizeInBytes())
3048 return nullptr;
3049
3050 unsigned Index = SL->getElementContainingOffset(Offset);
3051 Offset -= SL->getElementOffset(Index);
3052
3053 Type *ElementTy = STy->getElementType(Index);
3054 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3055 if (Offset >= ElementSize)
3056 return nullptr; // The offset points into alignment padding.
3057
3058 // See if any partition must be contained by the element.
3059 if (Offset > 0 || Size < ElementSize) {
3060 if ((Offset + Size) > ElementSize)
3061 return nullptr;
3062 return getTypePartition(DL, ElementTy, Offset, Size);
3063 }
3064 assert(Offset == 0);
3065
3066 if (Size == ElementSize)
3067 return stripAggregateTypeWrapping(DL, ElementTy);
3068
3069 StructType::element_iterator EI = STy->element_begin() + Index,
3070 EE = STy->element_end();
3071 if (EndOffset < SL->getSizeInBytes()) {
3072 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3073 if (Index == EndIndex)
3074 return nullptr; // Within a single element and its padding.
3075
3076 // Don't try to form "natural" types if the elements don't line up with the
3077 // expected size.
3078 // FIXME: We could potentially recurse down through the last element in the
3079 // sub-struct to find a natural end point.
3080 if (SL->getElementOffset(EndIndex) != EndOffset)
3081 return nullptr;
3082
3083 assert(Index < EndIndex);
3084 EE = STy->element_begin() + EndIndex;
3085 }
3086
3087 // Try to build up a sub-structure.
3088 StructType *SubTy = StructType::get(STy->getContext(), makeArrayRef(EI, EE),
3089 STy->isPacked());
3090 const StructLayout *SubSL = DL.getStructLayout(SubTy);
3091 if (Size != SubSL->getSizeInBytes())
3092 return nullptr; // The sub-struct doesn't have quite the size needed.
3093
3094 return SubTy;
3095}
3096
3097/// \brief Rewrite an alloca partition's users.
3098///
3099/// This routine drives both of the rewriting goals of the SROA pass. It tries
3100/// to rewrite uses of an alloca partition to be conducive for SSA value
3101/// promotion. If the partition needs a new, more refined alloca, this will
3102/// build that new alloca, preserving as much type information as possible, and
3103/// rewrite the uses of the old alloca to point at the new one and have the
3104/// appropriate new offsets. It also evaluates how successful the rewrite was
3105/// at enabling promotion and if it was successful queues the alloca to be
3106/// promoted.
3107bool SROA::rewritePartition(AllocaInst &AI, AllocaSlices &S,
3108 AllocaSlices::iterator B, AllocaSlices::iterator E,
3109 int64_t BeginOffset, int64_t EndOffset,
3110 ArrayRef<AllocaSlices::iterator> SplitUses) {
3111 assert(BeginOffset < EndOffset);
3112 uint64_t SliceSize = EndOffset - BeginOffset;
3113
3114 // Try to compute a friendly type for this partition of the alloca. This
3115 // won't always succeed, in which case we fall back to a legal integer type
3116 // or an i8 array of an appropriate size.
3117 Type *SliceTy = nullptr;
3118 if (Type *CommonUseTy = findCommonType(B, E, EndOffset))
3119 if (DL->getTypeAllocSize(CommonUseTy) >= SliceSize)
3120 SliceTy = CommonUseTy;
3121 if (!SliceTy)
3122 if (Type *TypePartitionTy = getTypePartition(*DL, AI.getAllocatedType(),
3123 BeginOffset, SliceSize))
3124 SliceTy = TypePartitionTy;
3125 if ((!SliceTy || (SliceTy->isArrayTy() &&
3126 SliceTy->getArrayElementType()->isIntegerTy())) &&
3127 DL->isLegalInteger(SliceSize * 8))
3128 SliceTy = Type::getIntNTy(*C, SliceSize * 8);
3129 if (!SliceTy)
3130 SliceTy = ArrayType::get(Type::getInt8Ty(*C), SliceSize);
3131 assert(DL->getTypeAllocSize(SliceTy) >= SliceSize);
3132
3133 bool IsVectorPromotable = isVectorPromotionViable(
3134 *DL, SliceTy, S, BeginOffset, EndOffset, B, E, SplitUses);
3135
3136 bool IsIntegerPromotable =
3137 !IsVectorPromotable &&
3138 isIntegerWideningViable(*DL, SliceTy, BeginOffset, S, B, E, SplitUses);
3139
3140 // Check for the case where we're going to rewrite to a new alloca of the
3141 // exact same type as the original, and with the same access offsets. In that
3142 // case, re-use the existing alloca, but still run through the rewriter to
3143 // perform phi and select speculation.
3144 AllocaInst *NewAI;
3145 if (SliceTy == AI.getAllocatedType()) {
3146 assert(BeginOffset == 0 &&
3147 "Non-zero begin offset but same alloca type");
3148 NewAI = &AI;
3149 // FIXME: We should be able to bail at this point with "nothing changed".
3150 // FIXME: We might want to defer PHI speculation until after here.
3151 } else {
3152 unsigned Alignment = AI.getAlignment();
3153 if (!Alignment) {
3154 // The minimum alignment which users can rely on when the explicit
3155 // alignment is omitted or zero is that required by the ABI for this
3156 // type.
3157 Alignment = DL->getABITypeAlignment(AI.getAllocatedType());
3158 }
3159 Alignment = MinAlign(Alignment, BeginOffset);
3160 // If we will get at least this much alignment from the type alone, leave
3161 // the alloca's alignment unconstrained.
3162 if (Alignment <= DL->getABITypeAlignment(SliceTy))
3163 Alignment = 0;
3164 NewAI = new AllocaInst(SliceTy, nullptr, Alignment,
3165 AI.getName() + ".sroa." + Twine(B - S.begin()), &AI);
3166 ++NumNewAllocas;
3167 }
3168
3169 DEBUG(dbgs() << "Rewriting alloca partition "
3170 << "[" << BeginOffset << "," << EndOffset << ") to: " << *NewAI
3171 << "\n");
3172
3173 // Track the high watermark on the worklist as it is only relevant for
3174 // promoted allocas. We will reset it to this point if the alloca is not in
3175 // fact scheduled for promotion.
3176 unsigned PPWOldSize = PostPromotionWorklist.size();
3177 unsigned NumUses = 0;
3178 SmallPtrSet<PHINode *, 8> PHIUsers;
3179 SmallPtrSet<SelectInst *, 8> SelectUsers;
3180
3181 AllocaSliceRewriter Rewriter(*DL, S, *this, AI, *NewAI, BeginOffset,
3182 EndOffset, IsVectorPromotable,
3183 IsIntegerPromotable, PHIUsers, SelectUsers);
3184 bool Promotable = true;
3185 for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
3186 SUE = SplitUses.end();
3187 SUI != SUE; ++SUI) {
3188 DEBUG(dbgs() << " rewriting split ");
3189 DEBUG(S.printSlice(dbgs(), *SUI, ""));
3190 Promotable &= Rewriter.visit(*SUI);
3191 ++NumUses;
3192 }
3193 for (AllocaSlices::iterator I = B; I != E; ++I) {
3194 DEBUG(dbgs() << " rewriting ");
3195 DEBUG(S.printSlice(dbgs(), I, ""));
3196 Promotable &= Rewriter.visit(I);
3197 ++NumUses;
3198 }
3199
3200 NumAllocaPartitionUses += NumUses;
3201 MaxUsesPerAllocaPartition =
3202 std::max<unsigned>(NumUses, MaxUsesPerAllocaPartition);
3203
3204 // Now that we've processed all the slices in the new partition, check if any
3205 // PHIs or Selects would block promotion.
3206 for (SmallPtrSetImpl<PHINode *>::iterator I = PHIUsers.begin(),
3207 E = PHIUsers.end();
3208 I != E; ++I)
3209 if (!isSafePHIToSpeculate(**I, DL)) {
3210 Promotable = false;
3211 PHIUsers.clear();
3212 SelectUsers.clear();
3213 break;
3214 }
3215 for (SmallPtrSetImpl<SelectInst *>::iterator I = SelectUsers.begin(),
3216 E = SelectUsers.end();
3217 I != E; ++I)
3218 if (!isSafeSelectToSpeculate(**I, DL)) {
3219 Promotable = false;
3220 PHIUsers.clear();
3221 SelectUsers.clear();
3222 break;
3223 }
3224
3225 if (Promotable) {
3226 if (PHIUsers.empty() && SelectUsers.empty()) {
3227 // Promote the alloca.
3228 PromotableAllocas.push_back(NewAI);
3229 } else {
3230 // If we have either PHIs or Selects to speculate, add them to those
3231 // worklists and re-queue the new alloca so that we promote in on the
3232 // next iteration.
3233 for (SmallPtrSetImpl<PHINode *>::iterator I = PHIUsers.begin(),
3234 E = PHIUsers.end();
3235 I != E; ++I)
3236 SpeculatablePHIs.insert(*I);
3237 for (SmallPtrSetImpl<SelectInst *>::iterator I = SelectUsers.begin(),
3238 E = SelectUsers.end();
3239 I != E; ++I)
3240 SpeculatableSelects.insert(*I);
3241 Worklist.insert(NewAI);
3242 }
3243 } else {
3244 // If we can't promote the alloca, iterate on it to check for new
3245 // refinements exposed by splitting the current alloca. Don't iterate on an
3246 // alloca which didn't actually change and didn't get promoted.
3247 if (NewAI != &AI)
3248 Worklist.insert(NewAI);
3249
3250 // Drop any post-promotion work items if promotion didn't happen.
3251 while (PostPromotionWorklist.size() > PPWOldSize)
3252 PostPromotionWorklist.pop_back();
3253 }
3254
3255 return true;
3256}
3257
3258static void
3259removeFinishedSplitUses(SmallVectorImpl<AllocaSlices::iterator> &SplitUses,
3260 uint64_t &MaxSplitUseEndOffset, uint64_t Offset) {
3261 if (Offset >= MaxSplitUseEndOffset) {
3262 SplitUses.clear();
3263 MaxSplitUseEndOffset = 0;
3264 return;
3265 }
3266
3267 size_t SplitUsesOldSize = SplitUses.size();
3268 SplitUses.erase(std::remove_if(SplitUses.begin(), SplitUses.end(),
3269 [Offset](const AllocaSlices::iterator &I) {
3270 return I->endOffset() <= Offset;
3271 }),
3272 SplitUses.end());
3273 if (SplitUsesOldSize == SplitUses.size())
3274 return;
3275
3276 // Recompute the max. While this is linear, so is remove_if.
3277 MaxSplitUseEndOffset = 0;
3278 for (SmallVectorImpl<AllocaSlices::iterator>::iterator
3279 SUI = SplitUses.begin(),
3280 SUE = SplitUses.end();
3281 SUI != SUE; ++SUI)
3282 MaxSplitUseEndOffset = std::max((*SUI)->endOffset(), MaxSplitUseEndOffset);
3283}
3284
3285/// \brief Walks the slices of an alloca and form partitions based on them,
3286/// rewriting each of their uses.
3287bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &S) {
3288 if (S.begin() == S.end())
3289 return false;
3290
3291 unsigned NumPartitions = 0;
3292 bool Changed = false;
3293 SmallVector<AllocaSlices::iterator, 4> SplitUses;
3294 uint64_t MaxSplitUseEndOffset = 0;
3295
3296 uint64_t BeginOffset = S.begin()->beginOffset();
3297
3298 for (AllocaSlices::iterator SI = S.begin(), SJ = std::next(SI), SE = S.end();
3299 SI != SE; SI = SJ) {
3300 uint64_t MaxEndOffset = SI->endOffset();
3301
3302 if (!SI->isSplittable()) {
3303 // When we're forming an unsplittable region, it must always start at the
3304 // first slice and will extend through its end.
3305 assert(BeginOffset == SI->beginOffset());
3306
3307 // Form a partition including all of the overlapping slices with this
3308 // unsplittable slice.
3309 while (SJ != SE && SJ->beginOffset() < MaxEndOffset) {
3310 if (!SJ->isSplittable())
3311 MaxEndOffset = std::max(MaxEndOffset, SJ->endOffset());
3312 ++SJ;
3313 }
3314 } else {
3315 assert(SI->isSplittable()); // Established above.
3316
3317 // Collect all of the overlapping splittable slices.
3318 while (SJ != SE && SJ->beginOffset() < MaxEndOffset &&
3319 SJ->isSplittable()) {
3320 MaxEndOffset = std::max(MaxEndOffset, SJ->endOffset());
3321 ++SJ;
3322 }
3323
3324 // Back up MaxEndOffset and SJ if we ended the span early when
3325 // encountering an unsplittable slice.
3326 if (SJ != SE && SJ->beginOffset() < MaxEndOffset) {
3327 assert(!SJ->isSplittable());
3328 MaxEndOffset = SJ->beginOffset();
3329 }
3330 }
3331
3332 // Check if we have managed to move the end offset forward yet. If so,
3333 // we'll have to rewrite uses and erase old split uses.
3334 if (BeginOffset < MaxEndOffset) {
3335 // Rewrite a sequence of overlapping slices.
3336 Changed |=
3337 rewritePartition(AI, S, SI, SJ, BeginOffset, MaxEndOffset, SplitUses);
3338 ++NumPartitions;
3339
3340 removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset, MaxEndOffset);
3341 }
3342
3343 // Accumulate all the splittable slices from the [SI,SJ) region which
3344 // overlap going forward.
3345 for (AllocaSlices::iterator SK = SI; SK != SJ; ++SK)
3346 if (SK->isSplittable() && SK->endOffset() > MaxEndOffset) {
3347 SplitUses.push_back(SK);
3348 MaxSplitUseEndOffset = std::max(SK->endOffset(), MaxSplitUseEndOffset);
3349 }
3350
3351 // If we're already at the end and we have no split uses, we're done.
3352 if (SJ == SE && SplitUses.empty())
3353 break;
3354
3355 // If we have no split uses or no gap in offsets, we're ready to move to
3356 // the next slice.
3357 if (SplitUses.empty() || (SJ != SE && MaxEndOffset == SJ->beginOffset())) {
3358 BeginOffset = SJ->beginOffset();
3359 continue;
3360 }
3361
3362 // Even if we have split slices, if the next slice is splittable and the
3363 // split slices reach it, we can simply set up the beginning offset of the
3364 // next iteration to bridge between them.
3365 if (SJ != SE && SJ->isSplittable() &&
3366 MaxSplitUseEndOffset > SJ->beginOffset()) {
3367 BeginOffset = MaxEndOffset;
3368 continue;
3369 }
3370
3371 // Otherwise, we have a tail of split slices. Rewrite them with an empty
3372 // range of slices.
3373 uint64_t PostSplitEndOffset =
3374 SJ == SE ? MaxSplitUseEndOffset : SJ->beginOffset();
3375
3376 Changed |= rewritePartition(AI, S, SJ, SJ, MaxEndOffset, PostSplitEndOffset,
3377 SplitUses);
3378 ++NumPartitions;
3379
3380 if (SJ == SE)
3381 break; // Skip the rest, we don't need to do any cleanup.
3382
3383 removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset,
3384 PostSplitEndOffset);
3385
3386 // Now just reset the begin offset for the next iteration.
3387 BeginOffset = SJ->beginOffset();
3388 }
3389
3390 NumAllocaPartitions += NumPartitions;
3391 MaxPartitionsPerAlloca =
3392 std::max<unsigned>(NumPartitions, MaxPartitionsPerAlloca);
3393
3394 return Changed;
3395}
3396
3397/// \brief Clobber a use with undef, deleting the used value if it becomes dead.
3398void SROA::clobberUse(Use &U) {
3399 Value *OldV = U;
3400 // Replace the use with an undef value.
3401 U = UndefValue::get(OldV->getType());
3402
3403 // Check for this making an instruction dead. We have to garbage collect
3404 // all the dead instructions to ensure the uses of any alloca end up being
3405 // minimal.
3406 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3407 if (isInstructionTriviallyDead(OldI)) {
3408 DeadInsts.insert(OldI);
3409 }
3410}
3411
3412/// \brief Analyze an alloca for SROA.
3413///
3414/// This analyzes the alloca to ensure we can reason about it, builds
3415/// the slices of the alloca, and then hands it off to be split and
3416/// rewritten as needed.
3417bool SROA::runOnAlloca(AllocaInst &AI) {
3418 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3419 ++NumAllocasAnalyzed;
3420
3421 // Special case dead allocas, as they're trivial.
3422 if (AI.use_empty()) {
3423 AI.eraseFromParent();
3424 return true;
3425 }
3426
3427 // Skip alloca forms that this analysis can't handle.
3428 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3429 DL->getTypeAllocSize(AI.getAllocatedType()) == 0)
3430 return false;
3431
3432 bool Changed = false;
3433
3434 // First, split any FCA loads and stores touching this alloca to promote
3435 // better splitting and promotion opportunities.
3436 AggLoadStoreRewriter AggRewriter(*DL);
3437 Changed |= AggRewriter.rewrite(AI);
3438
3439 // Build the slices using a recursive instruction-visiting builder.
3440 AllocaSlices S(*DL, AI);
3441 DEBUG(S.print(dbgs()));
3442 if (S.isEscaped())
3443 return Changed;
3444
3445 // Delete all the dead users of this alloca before splitting and rewriting it.
3446 for (AllocaSlices::dead_user_iterator DI = S.dead_user_begin(),
3447 DE = S.dead_user_end();
3448 DI != DE; ++DI) {
3449 // Free up everything used by this instruction.
3450 for (Use &DeadOp : (*DI)->operands())
3451 clobberUse(DeadOp);
3452
3453 // Now replace the uses of this instruction.
3454 (*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType()));
3455
3456 // And mark it for deletion.
3457 DeadInsts.insert(*DI);
3458 Changed = true;
3459 }
3460 for (AllocaSlices::dead_op_iterator DO = S.dead_op_begin(),
3461 DE = S.dead_op_end();
3462 DO != DE; ++DO) {
3463 clobberUse(**DO);
3464 Changed = true;
3465 }
3466
3467 // No slices to split. Leave the dead alloca for a later pass to clean up.
3468 if (S.begin() == S.end())
3469 return Changed;
3470
3471 Changed |= splitAlloca(AI, S);
3472
3473 DEBUG(dbgs() << " Speculating PHIs\n");
3474 while (!SpeculatablePHIs.empty())
3475 speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
3476
3477 DEBUG(dbgs() << " Speculating Selects\n");
3478 while (!SpeculatableSelects.empty())
3479 speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
3480
3481 return Changed;
3482}
3483
3484/// \brief Delete the dead instructions accumulated in this run.
3485///
3486/// Recursively deletes the dead instructions we've accumulated. This is done
3487/// at the very end to maximize locality of the recursive delete and to
3488/// minimize the problems of invalidated instruction pointers as such pointers
3489/// are used heavily in the intermediate stages of the algorithm.
3490///
3491/// We also record the alloca instructions deleted here so that they aren't
3492/// subsequently handed to mem2reg to promote.
3493void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
3494 while (!DeadInsts.empty()) {
3495 Instruction *I = DeadInsts.pop_back_val();
3496 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3497
3498 I->replaceAllUsesWith(UndefValue::get(I->getType()));
3499
3500 for (Use &Operand : I->operands())
3501 if (Instruction *U = dyn_cast<Instruction>(Operand)) {
3502 // Zero out the operand and see if it becomes trivially dead.
3503 Operand = nullptr;
3504 if (isInstructionTriviallyDead(U))
3505 DeadInsts.insert(U);
3506 }
3507
3508 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3509 DeletedAllocas.insert(AI);
3510
3511 ++NumDeleted;
3512 I->eraseFromParent();
3513 }
3514}
3515
3516static void enqueueUsersInWorklist(Instruction &I,
3517 SmallVectorImpl<Instruction *> &Worklist,
3518 SmallPtrSet<Instruction *, 8> &Visited) {
3519 for (User *U : I.users())
3520 if (Visited.insert(cast<Instruction>(U)))
3521 Worklist.push_back(cast<Instruction>(U));
3522}
3523
3524/// \brief Promote the allocas, using the best available technique.
3525///
3526/// This attempts to promote whatever allocas have been identified as viable in
3527/// the PromotableAllocas list. If that list is empty, there is nothing to do.
3528/// If there is a domtree available, we attempt to promote using the full power
3529/// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3530/// based on the SSAUpdater utilities. This function returns whether any
3531/// promotion occurred.
3532bool SROA::promoteAllocas(Function &F) {
3533 if (PromotableAllocas.empty())
3534 return false;
3535
3536 NumPromoted += PromotableAllocas.size();
3537
3538 if (DT && !ForceSSAUpdater) {
3539 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3540 PromoteMemToReg(PromotableAllocas, *DT);
3541 PromotableAllocas.clear();
3542 return true;
3543 }
3544
3545 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3546 SSAUpdater SSA;
3547 DIBuilder DIB(*F.getParent());
3548 SmallVector<Instruction *, 64> Insts;
3549
3550 // We need a worklist to walk the uses of each alloca.
3551 SmallVector<Instruction *, 8> Worklist;
3552 SmallPtrSet<Instruction *, 8> Visited;
3553 SmallVector<Instruction *, 32> DeadInsts;
3554
3555 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3556 AllocaInst *AI = PromotableAllocas[Idx];
3557 Insts.clear();
3558 Worklist.clear();
3559 Visited.clear();
3560
3561 enqueueUsersInWorklist(*AI, Worklist, Visited);
3562
3563 while (!Worklist.empty()) {
3564 Instruction *I = Worklist.pop_back_val();
3565
3566 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3567 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3568 // leading to them) here. Eventually it should use them to optimize the
3569 // scalar values produced.
3570 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3571 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3572 II->getIntrinsicID() == Intrinsic::lifetime_end);
3573 II->eraseFromParent();
3574 continue;
3575 }
3576
3577 // Push the loads and stores we find onto the list. SROA will already
3578 // have validated that all loads and stores are viable candidates for
3579 // promotion.
3580 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
3581 assert(LI->getType() == AI->getAllocatedType());
3582 Insts.push_back(LI);
3583 continue;
3584 }
3585 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
3586 assert(SI->getValueOperand()->getType() == AI->getAllocatedType());
3587 Insts.push_back(SI);
3588 continue;
3589 }
3590
3591 // For everything else, we know that only no-op bitcasts and GEPs will
3592 // make it this far, just recurse through them and recall them for later
3593 // removal.
3594 DeadInsts.push_back(I);
3595 enqueueUsersInWorklist(*I, Worklist, Visited);
3596 }
3597 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3598 while (!DeadInsts.empty())
3599 DeadInsts.pop_back_val()->eraseFromParent();
3600 AI->eraseFromParent();
3601 }
3602
3603 PromotableAllocas.clear();
3604 return true;
3605}
3606
3607bool SROA::runOnFunction(Function &F) {
3608 if (skipOptnoneFunction(F))
3609 return false;
3610
3611 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3612 C = &F.getContext();
3613 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
3614 if (!DLP) {
3615 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3616 return false;
3617 }
3618 DL = &DLP->getDataLayout();
3619 DominatorTreeWrapperPass *DTWP =
3620 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
3621 DT = DTWP ? &DTWP->getDomTree() : nullptr;
3622
3623 BasicBlock &EntryBB = F.getEntryBlock();
3624 for (BasicBlock::iterator I = EntryBB.begin(), E = std::prev(EntryBB.end());
3625 I != E; ++I)
3626 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3627 Worklist.insert(AI);
3628
3629 bool Changed = false;
3630 // A set of deleted alloca instruction pointers which should be removed from
3631 // the list of promotable allocas.
3632 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3633
3634 do {
3635 while (!Worklist.empty()) {
3636 Changed |= runOnAlloca(*Worklist.pop_back_val());
3637 deleteDeadInstructions(DeletedAllocas);
3638
3639 // Remove the deleted allocas from various lists so that we don't try to
3640 // continue processing them.
3641 if (!DeletedAllocas.empty()) {
3642 auto IsInSet = [&](AllocaInst *AI) {
3643 return DeletedAllocas.count(AI);
3644 };
3645 Worklist.remove_if(IsInSet);
3646 PostPromotionWorklist.remove_if(IsInSet);
3647 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3648 PromotableAllocas.end(),
3649 IsInSet),
3650 PromotableAllocas.end());
3651 DeletedAllocas.clear();
3652 }
3653 }
3654
3655 Changed |= promoteAllocas(F);
3656
3657 Worklist = PostPromotionWorklist;
3658 PostPromotionWorklist.clear();
3659 } while (!Worklist.empty());
3660
3661 return Changed;
3662}
3663
3664void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3665 if (RequiresDomTree)
3666 AU.addRequired<DominatorTreeWrapperPass>();
3667 AU.setPreservesCFG();
3668}
| 2704 PtrBuilder.SetCurrentDebugLocation(OldPtr->getDebugLoc());
2705
2706 Value *NewPtr = getNewAllocaSlicePtr(PtrBuilder, OldPtr->getType());
2707 // Replace the operands which were using the old pointer.
2708 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
2709
2710 DEBUG(dbgs() << " to: " << PN << "\n");
2711 deleteIfTriviallyDead(OldPtr);
2712
2713 // PHIs can't be promoted on their own, but often can be speculated. We
2714 // check the speculation outside of the rewriter so that we see the
2715 // fully-rewritten alloca.
2716 PHIUsers.insert(&PN);
2717 return true;
2718 }
2719
2720 bool visitSelectInst(SelectInst &SI) {
2721 DEBUG(dbgs() << " original: " << SI << "\n");
2722 assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
2723 "Pointer isn't an operand!");
2724 assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
2725 assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
2726
2727 Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2728 // Replace the operands which were using the old pointer.
2729 if (SI.getOperand(1) == OldPtr)
2730 SI.setOperand(1, NewPtr);
2731 if (SI.getOperand(2) == OldPtr)
2732 SI.setOperand(2, NewPtr);
2733
2734 DEBUG(dbgs() << " to: " << SI << "\n");
2735 deleteIfTriviallyDead(OldPtr);
2736
2737 // Selects can't be promoted on their own, but often can be speculated. We
2738 // check the speculation outside of the rewriter so that we see the
2739 // fully-rewritten alloca.
2740 SelectUsers.insert(&SI);
2741 return true;
2742 }
2743
2744};
2745}
2746
2747namespace {
2748/// \brief Visitor to rewrite aggregate loads and stores as scalar.
2749///
2750/// This pass aggressively rewrites all aggregate loads and stores on
2751/// a particular pointer (or any pointer derived from it which we can identify)
2752/// with scalar loads and stores.
2753class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
2754 // Befriend the base class so it can delegate to private visit methods.
2755 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
2756
2757 const DataLayout &DL;
2758
2759 /// Queue of pointer uses to analyze and potentially rewrite.
2760 SmallVector<Use *, 8> Queue;
2761
2762 /// Set to prevent us from cycling with phi nodes and loops.
2763 SmallPtrSet<User *, 8> Visited;
2764
2765 /// The current pointer use being rewritten. This is used to dig up the used
2766 /// value (as opposed to the user).
2767 Use *U;
2768
2769public:
2770 AggLoadStoreRewriter(const DataLayout &DL) : DL(DL) {}
2771
2772 /// Rewrite loads and stores through a pointer and all pointers derived from
2773 /// it.
2774 bool rewrite(Instruction &I) {
2775 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
2776 enqueueUsers(I);
2777 bool Changed = false;
2778 while (!Queue.empty()) {
2779 U = Queue.pop_back_val();
2780 Changed |= visit(cast<Instruction>(U->getUser()));
2781 }
2782 return Changed;
2783 }
2784
2785private:
2786 /// Enqueue all the users of the given instruction for further processing.
2787 /// This uses a set to de-duplicate users.
2788 void enqueueUsers(Instruction &I) {
2789 for (Use &U : I.uses())
2790 if (Visited.insert(U.getUser()))
2791 Queue.push_back(&U);
2792 }
2793
2794 // Conservative default is to not rewrite anything.
2795 bool visitInstruction(Instruction &I) { return false; }
2796
2797 /// \brief Generic recursive split emission class.
2798 template <typename Derived>
2799 class OpSplitter {
2800 protected:
2801 /// The builder used to form new instructions.
2802 IRBuilderTy IRB;
2803 /// The indices which to be used with insert- or extractvalue to select the
2804 /// appropriate value within the aggregate.
2805 SmallVector<unsigned, 4> Indices;
2806 /// The indices to a GEP instruction which will move Ptr to the correct slot
2807 /// within the aggregate.
2808 SmallVector<Value *, 4> GEPIndices;
2809 /// The base pointer of the original op, used as a base for GEPing the
2810 /// split operations.
2811 Value *Ptr;
2812
2813 /// Initialize the splitter with an insertion point, Ptr and start with a
2814 /// single zero GEP index.
2815 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
2816 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
2817
2818 public:
2819 /// \brief Generic recursive split emission routine.
2820 ///
2821 /// This method recursively splits an aggregate op (load or store) into
2822 /// scalar or vector ops. It splits recursively until it hits a single value
2823 /// and emits that single value operation via the template argument.
2824 ///
2825 /// The logic of this routine relies on GEPs and insertvalue and
2826 /// extractvalue all operating with the same fundamental index list, merely
2827 /// formatted differently (GEPs need actual values).
2828 ///
2829 /// \param Ty The type being split recursively into smaller ops.
2830 /// \param Agg The aggregate value being built up or stored, depending on
2831 /// whether this is splitting a load or a store respectively.
2832 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
2833 if (Ty->isSingleValueType())
2834 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
2835
2836 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2837 unsigned OldSize = Indices.size();
2838 (void)OldSize;
2839 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
2840 ++Idx) {
2841 assert(Indices.size() == OldSize && "Did not return to the old size");
2842 Indices.push_back(Idx);
2843 GEPIndices.push_back(IRB.getInt32(Idx));
2844 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
2845 GEPIndices.pop_back();
2846 Indices.pop_back();
2847 }
2848 return;
2849 }
2850
2851 if (StructType *STy = dyn_cast<StructType>(Ty)) {
2852 unsigned OldSize = Indices.size();
2853 (void)OldSize;
2854 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
2855 ++Idx) {
2856 assert(Indices.size() == OldSize && "Did not return to the old size");
2857 Indices.push_back(Idx);
2858 GEPIndices.push_back(IRB.getInt32(Idx));
2859 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
2860 GEPIndices.pop_back();
2861 Indices.pop_back();
2862 }
2863 return;
2864 }
2865
2866 llvm_unreachable("Only arrays and structs are aggregate loadable types");
2867 }
2868 };
2869
2870 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
2871 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2872 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
2873
2874 /// Emit a leaf load of a single value. This is called at the leaves of the
2875 /// recursive emission to actually load values.
2876 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2877 assert(Ty->isSingleValueType());
2878 // Load the single value and insert it using the indices.
2879 Value *GEP = IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep");
2880 Value *Load = IRB.CreateLoad(GEP, Name + ".load");
2881 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
2882 DEBUG(dbgs() << " to: " << *Load << "\n");
2883 }
2884 };
2885
2886 bool visitLoadInst(LoadInst &LI) {
2887 assert(LI.getPointerOperand() == *U);
2888 if (!LI.isSimple() || LI.getType()->isSingleValueType())
2889 return false;
2890
2891 // We have an aggregate being loaded, split it apart.
2892 DEBUG(dbgs() << " original: " << LI << "\n");
2893 LoadOpSplitter Splitter(&LI, *U);
2894 Value *V = UndefValue::get(LI.getType());
2895 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
2896 LI.replaceAllUsesWith(V);
2897 LI.eraseFromParent();
2898 return true;
2899 }
2900
2901 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
2902 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2903 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
2904
2905 /// Emit a leaf store of a single value. This is called at the leaves of the
2906 /// recursive emission to actually produce stores.
2907 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2908 assert(Ty->isSingleValueType());
2909 // Extract the single value and store it using the indices.
2910 Value *Store = IRB.CreateStore(
2911 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
2912 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
2913 (void)Store;
2914 DEBUG(dbgs() << " to: " << *Store << "\n");
2915 }
2916 };
2917
2918 bool visitStoreInst(StoreInst &SI) {
2919 if (!SI.isSimple() || SI.getPointerOperand() != *U)
2920 return false;
2921 Value *V = SI.getValueOperand();
2922 if (V->getType()->isSingleValueType())
2923 return false;
2924
2925 // We have an aggregate being stored, split it apart.
2926 DEBUG(dbgs() << " original: " << SI << "\n");
2927 StoreOpSplitter Splitter(&SI, *U);
2928 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
2929 SI.eraseFromParent();
2930 return true;
2931 }
2932
2933 bool visitBitCastInst(BitCastInst &BC) {
2934 enqueueUsers(BC);
2935 return false;
2936 }
2937
2938 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
2939 enqueueUsers(GEPI);
2940 return false;
2941 }
2942
2943 bool visitPHINode(PHINode &PN) {
2944 enqueueUsers(PN);
2945 return false;
2946 }
2947
2948 bool visitSelectInst(SelectInst &SI) {
2949 enqueueUsers(SI);
2950 return false;
2951 }
2952};
2953}
2954
2955/// \brief Strip aggregate type wrapping.
2956///
2957/// This removes no-op aggregate types wrapping an underlying type. It will
2958/// strip as many layers of types as it can without changing either the type
2959/// size or the allocated size.
2960static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
2961 if (Ty->isSingleValueType())
2962 return Ty;
2963
2964 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
2965 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
2966
2967 Type *InnerTy;
2968 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
2969 InnerTy = ArrTy->getElementType();
2970 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
2971 const StructLayout *SL = DL.getStructLayout(STy);
2972 unsigned Index = SL->getElementContainingOffset(0);
2973 InnerTy = STy->getElementType(Index);
2974 } else {
2975 return Ty;
2976 }
2977
2978 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
2979 TypeSize > DL.getTypeSizeInBits(InnerTy))
2980 return Ty;
2981
2982 return stripAggregateTypeWrapping(DL, InnerTy);
2983}
2984
2985/// \brief Try to find a partition of the aggregate type passed in for a given
2986/// offset and size.
2987///
2988/// This recurses through the aggregate type and tries to compute a subtype
2989/// based on the offset and size. When the offset and size span a sub-section
2990/// of an array, it will even compute a new array type for that sub-section,
2991/// and the same for structs.
2992///
2993/// Note that this routine is very strict and tries to find a partition of the
2994/// type which produces the *exact* right offset and size. It is not forgiving
2995/// when the size or offset cause either end of type-based partition to be off.
2996/// Also, this is a best-effort routine. It is reasonable to give up and not
2997/// return a type if necessary.
2998static Type *getTypePartition(const DataLayout &DL, Type *Ty,
2999 uint64_t Offset, uint64_t Size) {
3000 if (Offset == 0 && DL.getTypeAllocSize(Ty) == Size)
3001 return stripAggregateTypeWrapping(DL, Ty);
3002 if (Offset > DL.getTypeAllocSize(Ty) ||
3003 (DL.getTypeAllocSize(Ty) - Offset) < Size)
3004 return nullptr;
3005
3006 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3007 // We can't partition pointers...
3008 if (SeqTy->isPointerTy())
3009 return nullptr;
3010
3011 Type *ElementTy = SeqTy->getElementType();
3012 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3013 uint64_t NumSkippedElements = Offset / ElementSize;
3014 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy)) {
3015 if (NumSkippedElements >= ArrTy->getNumElements())
3016 return nullptr;
3017 } else if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy)) {
3018 if (NumSkippedElements >= VecTy->getNumElements())
3019 return nullptr;
3020 }
3021 Offset -= NumSkippedElements * ElementSize;
3022
3023 // First check if we need to recurse.
3024 if (Offset > 0 || Size < ElementSize) {
3025 // Bail if the partition ends in a different array element.
3026 if ((Offset + Size) > ElementSize)
3027 return nullptr;
3028 // Recurse through the element type trying to peel off offset bytes.
3029 return getTypePartition(DL, ElementTy, Offset, Size);
3030 }
3031 assert(Offset == 0);
3032
3033 if (Size == ElementSize)
3034 return stripAggregateTypeWrapping(DL, ElementTy);
3035 assert(Size > ElementSize);
3036 uint64_t NumElements = Size / ElementSize;
3037 if (NumElements * ElementSize != Size)
3038 return nullptr;
3039 return ArrayType::get(ElementTy, NumElements);
3040 }
3041
3042 StructType *STy = dyn_cast<StructType>(Ty);
3043 if (!STy)
3044 return nullptr;
3045
3046 const StructLayout *SL = DL.getStructLayout(STy);
3047 if (Offset >= SL->getSizeInBytes())
3048 return nullptr;
3049 uint64_t EndOffset = Offset + Size;
3050 if (EndOffset > SL->getSizeInBytes())
3051 return nullptr;
3052
3053 unsigned Index = SL->getElementContainingOffset(Offset);
3054 Offset -= SL->getElementOffset(Index);
3055
3056 Type *ElementTy = STy->getElementType(Index);
3057 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3058 if (Offset >= ElementSize)
3059 return nullptr; // The offset points into alignment padding.
3060
3061 // See if any partition must be contained by the element.
3062 if (Offset > 0 || Size < ElementSize) {
3063 if ((Offset + Size) > ElementSize)
3064 return nullptr;
3065 return getTypePartition(DL, ElementTy, Offset, Size);
3066 }
3067 assert(Offset == 0);
3068
3069 if (Size == ElementSize)
3070 return stripAggregateTypeWrapping(DL, ElementTy);
3071
3072 StructType::element_iterator EI = STy->element_begin() + Index,
3073 EE = STy->element_end();
3074 if (EndOffset < SL->getSizeInBytes()) {
3075 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3076 if (Index == EndIndex)
3077 return nullptr; // Within a single element and its padding.
3078
3079 // Don't try to form "natural" types if the elements don't line up with the
3080 // expected size.
3081 // FIXME: We could potentially recurse down through the last element in the
3082 // sub-struct to find a natural end point.
3083 if (SL->getElementOffset(EndIndex) != EndOffset)
3084 return nullptr;
3085
3086 assert(Index < EndIndex);
3087 EE = STy->element_begin() + EndIndex;
3088 }
3089
3090 // Try to build up a sub-structure.
3091 StructType *SubTy = StructType::get(STy->getContext(), makeArrayRef(EI, EE),
3092 STy->isPacked());
3093 const StructLayout *SubSL = DL.getStructLayout(SubTy);
3094 if (Size != SubSL->getSizeInBytes())
3095 return nullptr; // The sub-struct doesn't have quite the size needed.
3096
3097 return SubTy;
3098}
3099
3100/// \brief Rewrite an alloca partition's users.
3101///
3102/// This routine drives both of the rewriting goals of the SROA pass. It tries
3103/// to rewrite uses of an alloca partition to be conducive for SSA value
3104/// promotion. If the partition needs a new, more refined alloca, this will
3105/// build that new alloca, preserving as much type information as possible, and
3106/// rewrite the uses of the old alloca to point at the new one and have the
3107/// appropriate new offsets. It also evaluates how successful the rewrite was
3108/// at enabling promotion and if it was successful queues the alloca to be
3109/// promoted.
3110bool SROA::rewritePartition(AllocaInst &AI, AllocaSlices &S,
3111 AllocaSlices::iterator B, AllocaSlices::iterator E,
3112 int64_t BeginOffset, int64_t EndOffset,
3113 ArrayRef<AllocaSlices::iterator> SplitUses) {
3114 assert(BeginOffset < EndOffset);
3115 uint64_t SliceSize = EndOffset - BeginOffset;
3116
3117 // Try to compute a friendly type for this partition of the alloca. This
3118 // won't always succeed, in which case we fall back to a legal integer type
3119 // or an i8 array of an appropriate size.
3120 Type *SliceTy = nullptr;
3121 if (Type *CommonUseTy = findCommonType(B, E, EndOffset))
3122 if (DL->getTypeAllocSize(CommonUseTy) >= SliceSize)
3123 SliceTy = CommonUseTy;
3124 if (!SliceTy)
3125 if (Type *TypePartitionTy = getTypePartition(*DL, AI.getAllocatedType(),
3126 BeginOffset, SliceSize))
3127 SliceTy = TypePartitionTy;
3128 if ((!SliceTy || (SliceTy->isArrayTy() &&
3129 SliceTy->getArrayElementType()->isIntegerTy())) &&
3130 DL->isLegalInteger(SliceSize * 8))
3131 SliceTy = Type::getIntNTy(*C, SliceSize * 8);
3132 if (!SliceTy)
3133 SliceTy = ArrayType::get(Type::getInt8Ty(*C), SliceSize);
3134 assert(DL->getTypeAllocSize(SliceTy) >= SliceSize);
3135
3136 bool IsVectorPromotable = isVectorPromotionViable(
3137 *DL, SliceTy, S, BeginOffset, EndOffset, B, E, SplitUses);
3138
3139 bool IsIntegerPromotable =
3140 !IsVectorPromotable &&
3141 isIntegerWideningViable(*DL, SliceTy, BeginOffset, S, B, E, SplitUses);
3142
3143 // Check for the case where we're going to rewrite to a new alloca of the
3144 // exact same type as the original, and with the same access offsets. In that
3145 // case, re-use the existing alloca, but still run through the rewriter to
3146 // perform phi and select speculation.
3147 AllocaInst *NewAI;
3148 if (SliceTy == AI.getAllocatedType()) {
3149 assert(BeginOffset == 0 &&
3150 "Non-zero begin offset but same alloca type");
3151 NewAI = &AI;
3152 // FIXME: We should be able to bail at this point with "nothing changed".
3153 // FIXME: We might want to defer PHI speculation until after here.
3154 } else {
3155 unsigned Alignment = AI.getAlignment();
3156 if (!Alignment) {
3157 // The minimum alignment which users can rely on when the explicit
3158 // alignment is omitted or zero is that required by the ABI for this
3159 // type.
3160 Alignment = DL->getABITypeAlignment(AI.getAllocatedType());
3161 }
3162 Alignment = MinAlign(Alignment, BeginOffset);
3163 // If we will get at least this much alignment from the type alone, leave
3164 // the alloca's alignment unconstrained.
3165 if (Alignment <= DL->getABITypeAlignment(SliceTy))
3166 Alignment = 0;
3167 NewAI = new AllocaInst(SliceTy, nullptr, Alignment,
3168 AI.getName() + ".sroa." + Twine(B - S.begin()), &AI);
3169 ++NumNewAllocas;
3170 }
3171
3172 DEBUG(dbgs() << "Rewriting alloca partition "
3173 << "[" << BeginOffset << "," << EndOffset << ") to: " << *NewAI
3174 << "\n");
3175
3176 // Track the high watermark on the worklist as it is only relevant for
3177 // promoted allocas. We will reset it to this point if the alloca is not in
3178 // fact scheduled for promotion.
3179 unsigned PPWOldSize = PostPromotionWorklist.size();
3180 unsigned NumUses = 0;
3181 SmallPtrSet<PHINode *, 8> PHIUsers;
3182 SmallPtrSet<SelectInst *, 8> SelectUsers;
3183
3184 AllocaSliceRewriter Rewriter(*DL, S, *this, AI, *NewAI, BeginOffset,
3185 EndOffset, IsVectorPromotable,
3186 IsIntegerPromotable, PHIUsers, SelectUsers);
3187 bool Promotable = true;
3188 for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
3189 SUE = SplitUses.end();
3190 SUI != SUE; ++SUI) {
3191 DEBUG(dbgs() << " rewriting split ");
3192 DEBUG(S.printSlice(dbgs(), *SUI, ""));
3193 Promotable &= Rewriter.visit(*SUI);
3194 ++NumUses;
3195 }
3196 for (AllocaSlices::iterator I = B; I != E; ++I) {
3197 DEBUG(dbgs() << " rewriting ");
3198 DEBUG(S.printSlice(dbgs(), I, ""));
3199 Promotable &= Rewriter.visit(I);
3200 ++NumUses;
3201 }
3202
3203 NumAllocaPartitionUses += NumUses;
3204 MaxUsesPerAllocaPartition =
3205 std::max<unsigned>(NumUses, MaxUsesPerAllocaPartition);
3206
3207 // Now that we've processed all the slices in the new partition, check if any
3208 // PHIs or Selects would block promotion.
3209 for (SmallPtrSetImpl<PHINode *>::iterator I = PHIUsers.begin(),
3210 E = PHIUsers.end();
3211 I != E; ++I)
3212 if (!isSafePHIToSpeculate(**I, DL)) {
3213 Promotable = false;
3214 PHIUsers.clear();
3215 SelectUsers.clear();
3216 break;
3217 }
3218 for (SmallPtrSetImpl<SelectInst *>::iterator I = SelectUsers.begin(),
3219 E = SelectUsers.end();
3220 I != E; ++I)
3221 if (!isSafeSelectToSpeculate(**I, DL)) {
3222 Promotable = false;
3223 PHIUsers.clear();
3224 SelectUsers.clear();
3225 break;
3226 }
3227
3228 if (Promotable) {
3229 if (PHIUsers.empty() && SelectUsers.empty()) {
3230 // Promote the alloca.
3231 PromotableAllocas.push_back(NewAI);
3232 } else {
3233 // If we have either PHIs or Selects to speculate, add them to those
3234 // worklists and re-queue the new alloca so that we promote in on the
3235 // next iteration.
3236 for (SmallPtrSetImpl<PHINode *>::iterator I = PHIUsers.begin(),
3237 E = PHIUsers.end();
3238 I != E; ++I)
3239 SpeculatablePHIs.insert(*I);
3240 for (SmallPtrSetImpl<SelectInst *>::iterator I = SelectUsers.begin(),
3241 E = SelectUsers.end();
3242 I != E; ++I)
3243 SpeculatableSelects.insert(*I);
3244 Worklist.insert(NewAI);
3245 }
3246 } else {
3247 // If we can't promote the alloca, iterate on it to check for new
3248 // refinements exposed by splitting the current alloca. Don't iterate on an
3249 // alloca which didn't actually change and didn't get promoted.
3250 if (NewAI != &AI)
3251 Worklist.insert(NewAI);
3252
3253 // Drop any post-promotion work items if promotion didn't happen.
3254 while (PostPromotionWorklist.size() > PPWOldSize)
3255 PostPromotionWorklist.pop_back();
3256 }
3257
3258 return true;
3259}
3260
3261static void
3262removeFinishedSplitUses(SmallVectorImpl<AllocaSlices::iterator> &SplitUses,
3263 uint64_t &MaxSplitUseEndOffset, uint64_t Offset) {
3264 if (Offset >= MaxSplitUseEndOffset) {
3265 SplitUses.clear();
3266 MaxSplitUseEndOffset = 0;
3267 return;
3268 }
3269
3270 size_t SplitUsesOldSize = SplitUses.size();
3271 SplitUses.erase(std::remove_if(SplitUses.begin(), SplitUses.end(),
3272 [Offset](const AllocaSlices::iterator &I) {
3273 return I->endOffset() <= Offset;
3274 }),
3275 SplitUses.end());
3276 if (SplitUsesOldSize == SplitUses.size())
3277 return;
3278
3279 // Recompute the max. While this is linear, so is remove_if.
3280 MaxSplitUseEndOffset = 0;
3281 for (SmallVectorImpl<AllocaSlices::iterator>::iterator
3282 SUI = SplitUses.begin(),
3283 SUE = SplitUses.end();
3284 SUI != SUE; ++SUI)
3285 MaxSplitUseEndOffset = std::max((*SUI)->endOffset(), MaxSplitUseEndOffset);
3286}
3287
3288/// \brief Walks the slices of an alloca and form partitions based on them,
3289/// rewriting each of their uses.
3290bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &S) {
3291 if (S.begin() == S.end())
3292 return false;
3293
3294 unsigned NumPartitions = 0;
3295 bool Changed = false;
3296 SmallVector<AllocaSlices::iterator, 4> SplitUses;
3297 uint64_t MaxSplitUseEndOffset = 0;
3298
3299 uint64_t BeginOffset = S.begin()->beginOffset();
3300
3301 for (AllocaSlices::iterator SI = S.begin(), SJ = std::next(SI), SE = S.end();
3302 SI != SE; SI = SJ) {
3303 uint64_t MaxEndOffset = SI->endOffset();
3304
3305 if (!SI->isSplittable()) {
3306 // When we're forming an unsplittable region, it must always start at the
3307 // first slice and will extend through its end.
3308 assert(BeginOffset == SI->beginOffset());
3309
3310 // Form a partition including all of the overlapping slices with this
3311 // unsplittable slice.
3312 while (SJ != SE && SJ->beginOffset() < MaxEndOffset) {
3313 if (!SJ->isSplittable())
3314 MaxEndOffset = std::max(MaxEndOffset, SJ->endOffset());
3315 ++SJ;
3316 }
3317 } else {
3318 assert(SI->isSplittable()); // Established above.
3319
3320 // Collect all of the overlapping splittable slices.
3321 while (SJ != SE && SJ->beginOffset() < MaxEndOffset &&
3322 SJ->isSplittable()) {
3323 MaxEndOffset = std::max(MaxEndOffset, SJ->endOffset());
3324 ++SJ;
3325 }
3326
3327 // Back up MaxEndOffset and SJ if we ended the span early when
3328 // encountering an unsplittable slice.
3329 if (SJ != SE && SJ->beginOffset() < MaxEndOffset) {
3330 assert(!SJ->isSplittable());
3331 MaxEndOffset = SJ->beginOffset();
3332 }
3333 }
3334
3335 // Check if we have managed to move the end offset forward yet. If so,
3336 // we'll have to rewrite uses and erase old split uses.
3337 if (BeginOffset < MaxEndOffset) {
3338 // Rewrite a sequence of overlapping slices.
3339 Changed |=
3340 rewritePartition(AI, S, SI, SJ, BeginOffset, MaxEndOffset, SplitUses);
3341 ++NumPartitions;
3342
3343 removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset, MaxEndOffset);
3344 }
3345
3346 // Accumulate all the splittable slices from the [SI,SJ) region which
3347 // overlap going forward.
3348 for (AllocaSlices::iterator SK = SI; SK != SJ; ++SK)
3349 if (SK->isSplittable() && SK->endOffset() > MaxEndOffset) {
3350 SplitUses.push_back(SK);
3351 MaxSplitUseEndOffset = std::max(SK->endOffset(), MaxSplitUseEndOffset);
3352 }
3353
3354 // If we're already at the end and we have no split uses, we're done.
3355 if (SJ == SE && SplitUses.empty())
3356 break;
3357
3358 // If we have no split uses or no gap in offsets, we're ready to move to
3359 // the next slice.
3360 if (SplitUses.empty() || (SJ != SE && MaxEndOffset == SJ->beginOffset())) {
3361 BeginOffset = SJ->beginOffset();
3362 continue;
3363 }
3364
3365 // Even if we have split slices, if the next slice is splittable and the
3366 // split slices reach it, we can simply set up the beginning offset of the
3367 // next iteration to bridge between them.
3368 if (SJ != SE && SJ->isSplittable() &&
3369 MaxSplitUseEndOffset > SJ->beginOffset()) {
3370 BeginOffset = MaxEndOffset;
3371 continue;
3372 }
3373
3374 // Otherwise, we have a tail of split slices. Rewrite them with an empty
3375 // range of slices.
3376 uint64_t PostSplitEndOffset =
3377 SJ == SE ? MaxSplitUseEndOffset : SJ->beginOffset();
3378
3379 Changed |= rewritePartition(AI, S, SJ, SJ, MaxEndOffset, PostSplitEndOffset,
3380 SplitUses);
3381 ++NumPartitions;
3382
3383 if (SJ == SE)
3384 break; // Skip the rest, we don't need to do any cleanup.
3385
3386 removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset,
3387 PostSplitEndOffset);
3388
3389 // Now just reset the begin offset for the next iteration.
3390 BeginOffset = SJ->beginOffset();
3391 }
3392
3393 NumAllocaPartitions += NumPartitions;
3394 MaxPartitionsPerAlloca =
3395 std::max<unsigned>(NumPartitions, MaxPartitionsPerAlloca);
3396
3397 return Changed;
3398}
3399
3400/// \brief Clobber a use with undef, deleting the used value if it becomes dead.
3401void SROA::clobberUse(Use &U) {
3402 Value *OldV = U;
3403 // Replace the use with an undef value.
3404 U = UndefValue::get(OldV->getType());
3405
3406 // Check for this making an instruction dead. We have to garbage collect
3407 // all the dead instructions to ensure the uses of any alloca end up being
3408 // minimal.
3409 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3410 if (isInstructionTriviallyDead(OldI)) {
3411 DeadInsts.insert(OldI);
3412 }
3413}
3414
3415/// \brief Analyze an alloca for SROA.
3416///
3417/// This analyzes the alloca to ensure we can reason about it, builds
3418/// the slices of the alloca, and then hands it off to be split and
3419/// rewritten as needed.
3420bool SROA::runOnAlloca(AllocaInst &AI) {
3421 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3422 ++NumAllocasAnalyzed;
3423
3424 // Special case dead allocas, as they're trivial.
3425 if (AI.use_empty()) {
3426 AI.eraseFromParent();
3427 return true;
3428 }
3429
3430 // Skip alloca forms that this analysis can't handle.
3431 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3432 DL->getTypeAllocSize(AI.getAllocatedType()) == 0)
3433 return false;
3434
3435 bool Changed = false;
3436
3437 // First, split any FCA loads and stores touching this alloca to promote
3438 // better splitting and promotion opportunities.
3439 AggLoadStoreRewriter AggRewriter(*DL);
3440 Changed |= AggRewriter.rewrite(AI);
3441
3442 // Build the slices using a recursive instruction-visiting builder.
3443 AllocaSlices S(*DL, AI);
3444 DEBUG(S.print(dbgs()));
3445 if (S.isEscaped())
3446 return Changed;
3447
3448 // Delete all the dead users of this alloca before splitting and rewriting it.
3449 for (AllocaSlices::dead_user_iterator DI = S.dead_user_begin(),
3450 DE = S.dead_user_end();
3451 DI != DE; ++DI) {
3452 // Free up everything used by this instruction.
3453 for (Use &DeadOp : (*DI)->operands())
3454 clobberUse(DeadOp);
3455
3456 // Now replace the uses of this instruction.
3457 (*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType()));
3458
3459 // And mark it for deletion.
3460 DeadInsts.insert(*DI);
3461 Changed = true;
3462 }
3463 for (AllocaSlices::dead_op_iterator DO = S.dead_op_begin(),
3464 DE = S.dead_op_end();
3465 DO != DE; ++DO) {
3466 clobberUse(**DO);
3467 Changed = true;
3468 }
3469
3470 // No slices to split. Leave the dead alloca for a later pass to clean up.
3471 if (S.begin() == S.end())
3472 return Changed;
3473
3474 Changed |= splitAlloca(AI, S);
3475
3476 DEBUG(dbgs() << " Speculating PHIs\n");
3477 while (!SpeculatablePHIs.empty())
3478 speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
3479
3480 DEBUG(dbgs() << " Speculating Selects\n");
3481 while (!SpeculatableSelects.empty())
3482 speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
3483
3484 return Changed;
3485}
3486
3487/// \brief Delete the dead instructions accumulated in this run.
3488///
3489/// Recursively deletes the dead instructions we've accumulated. This is done
3490/// at the very end to maximize locality of the recursive delete and to
3491/// minimize the problems of invalidated instruction pointers as such pointers
3492/// are used heavily in the intermediate stages of the algorithm.
3493///
3494/// We also record the alloca instructions deleted here so that they aren't
3495/// subsequently handed to mem2reg to promote.
3496void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
3497 while (!DeadInsts.empty()) {
3498 Instruction *I = DeadInsts.pop_back_val();
3499 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3500
3501 I->replaceAllUsesWith(UndefValue::get(I->getType()));
3502
3503 for (Use &Operand : I->operands())
3504 if (Instruction *U = dyn_cast<Instruction>(Operand)) {
3505 // Zero out the operand and see if it becomes trivially dead.
3506 Operand = nullptr;
3507 if (isInstructionTriviallyDead(U))
3508 DeadInsts.insert(U);
3509 }
3510
3511 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3512 DeletedAllocas.insert(AI);
3513
3514 ++NumDeleted;
3515 I->eraseFromParent();
3516 }
3517}
3518
3519static void enqueueUsersInWorklist(Instruction &I,
3520 SmallVectorImpl<Instruction *> &Worklist,
3521 SmallPtrSet<Instruction *, 8> &Visited) {
3522 for (User *U : I.users())
3523 if (Visited.insert(cast<Instruction>(U)))
3524 Worklist.push_back(cast<Instruction>(U));
3525}
3526
3527/// \brief Promote the allocas, using the best available technique.
3528///
3529/// This attempts to promote whatever allocas have been identified as viable in
3530/// the PromotableAllocas list. If that list is empty, there is nothing to do.
3531/// If there is a domtree available, we attempt to promote using the full power
3532/// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3533/// based on the SSAUpdater utilities. This function returns whether any
3534/// promotion occurred.
3535bool SROA::promoteAllocas(Function &F) {
3536 if (PromotableAllocas.empty())
3537 return false;
3538
3539 NumPromoted += PromotableAllocas.size();
3540
3541 if (DT && !ForceSSAUpdater) {
3542 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3543 PromoteMemToReg(PromotableAllocas, *DT);
3544 PromotableAllocas.clear();
3545 return true;
3546 }
3547
3548 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3549 SSAUpdater SSA;
3550 DIBuilder DIB(*F.getParent());
3551 SmallVector<Instruction *, 64> Insts;
3552
3553 // We need a worklist to walk the uses of each alloca.
3554 SmallVector<Instruction *, 8> Worklist;
3555 SmallPtrSet<Instruction *, 8> Visited;
3556 SmallVector<Instruction *, 32> DeadInsts;
3557
3558 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3559 AllocaInst *AI = PromotableAllocas[Idx];
3560 Insts.clear();
3561 Worklist.clear();
3562 Visited.clear();
3563
3564 enqueueUsersInWorklist(*AI, Worklist, Visited);
3565
3566 while (!Worklist.empty()) {
3567 Instruction *I = Worklist.pop_back_val();
3568
3569 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3570 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3571 // leading to them) here. Eventually it should use them to optimize the
3572 // scalar values produced.
3573 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3574 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3575 II->getIntrinsicID() == Intrinsic::lifetime_end);
3576 II->eraseFromParent();
3577 continue;
3578 }
3579
3580 // Push the loads and stores we find onto the list. SROA will already
3581 // have validated that all loads and stores are viable candidates for
3582 // promotion.
3583 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
3584 assert(LI->getType() == AI->getAllocatedType());
3585 Insts.push_back(LI);
3586 continue;
3587 }
3588 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
3589 assert(SI->getValueOperand()->getType() == AI->getAllocatedType());
3590 Insts.push_back(SI);
3591 continue;
3592 }
3593
3594 // For everything else, we know that only no-op bitcasts and GEPs will
3595 // make it this far, just recurse through them and recall them for later
3596 // removal.
3597 DeadInsts.push_back(I);
3598 enqueueUsersInWorklist(*I, Worklist, Visited);
3599 }
3600 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3601 while (!DeadInsts.empty())
3602 DeadInsts.pop_back_val()->eraseFromParent();
3603 AI->eraseFromParent();
3604 }
3605
3606 PromotableAllocas.clear();
3607 return true;
3608}
3609
3610bool SROA::runOnFunction(Function &F) {
3611 if (skipOptnoneFunction(F))
3612 return false;
3613
3614 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3615 C = &F.getContext();
3616 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
3617 if (!DLP) {
3618 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3619 return false;
3620 }
3621 DL = &DLP->getDataLayout();
3622 DominatorTreeWrapperPass *DTWP =
3623 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
3624 DT = DTWP ? &DTWP->getDomTree() : nullptr;
3625
3626 BasicBlock &EntryBB = F.getEntryBlock();
3627 for (BasicBlock::iterator I = EntryBB.begin(), E = std::prev(EntryBB.end());
3628 I != E; ++I)
3629 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3630 Worklist.insert(AI);
3631
3632 bool Changed = false;
3633 // A set of deleted alloca instruction pointers which should be removed from
3634 // the list of promotable allocas.
3635 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3636
3637 do {
3638 while (!Worklist.empty()) {
3639 Changed |= runOnAlloca(*Worklist.pop_back_val());
3640 deleteDeadInstructions(DeletedAllocas);
3641
3642 // Remove the deleted allocas from various lists so that we don't try to
3643 // continue processing them.
3644 if (!DeletedAllocas.empty()) {
3645 auto IsInSet = [&](AllocaInst *AI) {
3646 return DeletedAllocas.count(AI);
3647 };
3648 Worklist.remove_if(IsInSet);
3649 PostPromotionWorklist.remove_if(IsInSet);
3650 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3651 PromotableAllocas.end(),
3652 IsInSet),
3653 PromotableAllocas.end());
3654 DeletedAllocas.clear();
3655 }
3656 }
3657
3658 Changed |= promoteAllocas(F);
3659
3660 Worklist = PostPromotionWorklist;
3661 PostPromotionWorklist.clear();
3662 } while (!Worklist.empty());
3663
3664 return Changed;
3665}
3666
3667void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3668 if (RequiresDomTree)
3669 AU.addRequired<DominatorTreeWrapperPass>();
3670 AU.setPreservesCFG();
3671}
|