1 //===- SROA.cpp - Scalar Replacement Of Aggregates ------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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.
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.
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.
24 //===----------------------------------------------------------------------===//
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/AssumptionTracker.h"
32 #include "llvm/Analysis/Loads.h"
33 #include "llvm/Analysis/PtrUseVisitor.h"
34 #include "llvm/Analysis/ValueTracking.h"
35 #include "llvm/IR/Constants.h"
36 #include "llvm/IR/DIBuilder.h"
37 #include "llvm/IR/DataLayout.h"
38 #include "llvm/IR/DebugInfo.h"
39 #include "llvm/IR/DerivedTypes.h"
40 #include "llvm/IR/Dominators.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/IRBuilder.h"
43 #include "llvm/IR/InstVisitor.h"
44 #include "llvm/IR/Instructions.h"
45 #include "llvm/IR/IntrinsicInst.h"
46 #include "llvm/IR/LLVMContext.h"
47 #include "llvm/IR/Operator.h"
48 #include "llvm/Pass.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Compiler.h"
51 #include "llvm/Support/Debug.h"
52 #include "llvm/Support/ErrorHandling.h"
53 #include "llvm/Support/MathExtras.h"
54 #include "llvm/Support/TimeValue.h"
55 #include "llvm/Support/raw_ostream.h"
56 #include "llvm/Transforms/Utils/Local.h"
57 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
58 #include "llvm/Transforms/Utils/SSAUpdater.h"
60 #if __cplusplus >= 201103L && !defined(NDEBUG)
61 // We only use this for a debug check in C++11
67 #define DEBUG_TYPE "sroa"
69 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
70 STATISTIC(NumAllocaPartitions, "Number of alloca partitions formed");
71 STATISTIC(MaxPartitionsPerAlloca, "Maximum number of partitions per alloca");
72 STATISTIC(NumAllocaPartitionUses, "Number of alloca partition uses rewritten");
73 STATISTIC(MaxUsesPerAllocaPartition, "Maximum number of uses of a partition");
74 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
75 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
76 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
77 STATISTIC(NumDeleted, "Number of instructions deleted");
78 STATISTIC(NumVectorized, "Number of vectorized aggregates");
80 /// Hidden option to force the pass to not use DomTree and mem2reg, instead
81 /// forming SSA values through the SSAUpdater infrastructure.
83 ForceSSAUpdater("force-ssa-updater", cl::init(false), cl::Hidden);
85 /// Hidden option to enable randomly shuffling the slices to help uncover
86 /// instability in their order.
87 static cl::opt<bool> SROARandomShuffleSlices("sroa-random-shuffle-slices",
88 cl::init(false), cl::Hidden);
90 /// Hidden option to experiment with completely strict handling of inbounds
92 static cl::opt<bool> SROAStrictInbounds("sroa-strict-inbounds",
93 cl::init(false), cl::Hidden);
96 /// \brief A custom IRBuilder inserter which prefixes all names if they are
98 template <bool preserveNames = true>
99 class IRBuilderPrefixedInserter :
100 public IRBuilderDefaultInserter<preserveNames> {
104 void SetNamePrefix(const Twine &P) { Prefix = P.str(); }
107 void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
108 BasicBlock::iterator InsertPt) const {
109 IRBuilderDefaultInserter<preserveNames>::InsertHelper(
110 I, Name.isTriviallyEmpty() ? Name : Prefix + Name, BB, InsertPt);
114 // Specialization for not preserving the name is trivial.
116 class IRBuilderPrefixedInserter<false> :
117 public IRBuilderDefaultInserter<false> {
119 void SetNamePrefix(const Twine &P) {}
122 /// \brief Provide a typedef for IRBuilder that drops names in release builds.
124 typedef llvm::IRBuilder<true, ConstantFolder,
125 IRBuilderPrefixedInserter<true> > IRBuilderTy;
127 typedef llvm::IRBuilder<false, ConstantFolder,
128 IRBuilderPrefixedInserter<false> > IRBuilderTy;
133 /// \brief A used slice of an alloca.
135 /// This structure represents a slice of an alloca used by some instruction. It
136 /// stores both the begin and end offsets of this use, a pointer to the use
137 /// itself, and a flag indicating whether we can classify the use as splittable
138 /// or not when forming partitions of the alloca.
140 /// \brief The beginning offset of the range.
141 uint64_t BeginOffset;
143 /// \brief The ending offset, not included in the range.
146 /// \brief Storage for both the use of this slice and whether it can be
148 PointerIntPair<Use *, 1, bool> UseAndIsSplittable;
151 Slice() : BeginOffset(), EndOffset() {}
152 Slice(uint64_t BeginOffset, uint64_t EndOffset, Use *U, bool IsSplittable)
153 : BeginOffset(BeginOffset), EndOffset(EndOffset),
154 UseAndIsSplittable(U, IsSplittable) {}
156 uint64_t beginOffset() const { return BeginOffset; }
157 uint64_t endOffset() const { return EndOffset; }
159 bool isSplittable() const { return UseAndIsSplittable.getInt(); }
160 void makeUnsplittable() { UseAndIsSplittable.setInt(false); }
162 Use *getUse() const { return UseAndIsSplittable.getPointer(); }
164 bool isDead() const { return getUse() == nullptr; }
165 void kill() { UseAndIsSplittable.setPointer(nullptr); }
167 /// \brief Support for ordering ranges.
169 /// This provides an ordering over ranges such that start offsets are
170 /// always increasing, and within equal start offsets, the end offsets are
171 /// decreasing. Thus the spanning range comes first in a cluster with the
172 /// same start position.
173 bool operator<(const Slice &RHS) const {
174 if (beginOffset() < RHS.beginOffset()) return true;
175 if (beginOffset() > RHS.beginOffset()) return false;
176 if (isSplittable() != RHS.isSplittable()) return !isSplittable();
177 if (endOffset() > RHS.endOffset()) return true;
181 /// \brief Support comparison with a single offset to allow binary searches.
182 friend LLVM_ATTRIBUTE_UNUSED bool operator<(const Slice &LHS,
183 uint64_t RHSOffset) {
184 return LHS.beginOffset() < RHSOffset;
186 friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
188 return LHSOffset < RHS.beginOffset();
191 bool operator==(const Slice &RHS) const {
192 return isSplittable() == RHS.isSplittable() &&
193 beginOffset() == RHS.beginOffset() && endOffset() == RHS.endOffset();
195 bool operator!=(const Slice &RHS) const { return !operator==(RHS); }
197 } // end anonymous namespace
200 template <typename T> struct isPodLike;
201 template <> struct isPodLike<Slice> {
202 static const bool value = true;
207 /// \brief Representation of the alloca slices.
209 /// This class represents the slices of an alloca which are formed by its
210 /// various uses. If a pointer escapes, we can't fully build a representation
211 /// for the slices used and we reflect that in this structure. The uses are
212 /// stored, sorted by increasing beginning offset and with unsplittable slices
213 /// starting at a particular offset before splittable slices.
216 /// \brief Construct the slices of a particular alloca.
217 AllocaSlices(const DataLayout &DL, AllocaInst &AI);
219 /// \brief Test whether a pointer to the allocation escapes our analysis.
221 /// If this is true, the slices are never fully built and should be
223 bool isEscaped() const { return PointerEscapingInstr; }
225 /// \brief Support for iterating over the slices.
227 typedef SmallVectorImpl<Slice>::iterator iterator;
228 typedef iterator_range<iterator> range;
229 iterator begin() { return Slices.begin(); }
230 iterator end() { return Slices.end(); }
232 typedef SmallVectorImpl<Slice>::const_iterator const_iterator;
233 typedef iterator_range<const_iterator> const_range;
234 const_iterator begin() const { return Slices.begin(); }
235 const_iterator end() const { return Slices.end(); }
238 /// \brief Access the dead users for this alloca.
239 ArrayRef<Instruction *> getDeadUsers() const { return DeadUsers; }
241 /// \brief Access the dead operands referring to this alloca.
243 /// These are operands which have cannot actually be used to refer to the
244 /// alloca as they are outside its range and the user doesn't correct for
245 /// that. These mostly consist of PHI node inputs and the like which we just
246 /// need to replace with undef.
247 ArrayRef<Use *> getDeadOperands() const { return DeadOperands; }
249 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
250 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
251 void printSlice(raw_ostream &OS, const_iterator I,
252 StringRef Indent = " ") const;
253 void printUse(raw_ostream &OS, const_iterator I,
254 StringRef Indent = " ") const;
255 void print(raw_ostream &OS) const;
256 void dump(const_iterator I) const;
261 template <typename DerivedT, typename RetT = void> class BuilderBase;
263 friend class AllocaSlices::SliceBuilder;
265 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
266 /// \brief Handle to alloca instruction to simplify method interfaces.
270 /// \brief The instruction responsible for this alloca not having a known set
273 /// When an instruction (potentially) escapes the pointer to the alloca, we
274 /// store a pointer to that here and abort trying to form slices of the
275 /// alloca. This will be null if the alloca slices are analyzed successfully.
276 Instruction *PointerEscapingInstr;
278 /// \brief The slices of the alloca.
280 /// We store a vector of the slices formed by uses of the alloca here. This
281 /// vector is sorted by increasing begin offset, and then the unsplittable
282 /// slices before the splittable ones. See the Slice inner class for more
284 SmallVector<Slice, 8> Slices;
286 /// \brief Instructions which will become dead if we rewrite the alloca.
288 /// Note that these are not separated by slice. This is because we expect an
289 /// alloca to be completely rewritten or not rewritten at all. If rewritten,
290 /// all these instructions can simply be removed and replaced with undef as
291 /// they come from outside of the allocated space.
292 SmallVector<Instruction *, 8> DeadUsers;
294 /// \brief Operands which will become dead if we rewrite the alloca.
296 /// These are operands that in their particular use can be replaced with
297 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
298 /// to PHI nodes and the like. They aren't entirely dead (there might be
299 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
300 /// want to swap this particular input for undef to simplify the use lists of
302 SmallVector<Use *, 8> DeadOperands;
306 static Value *foldSelectInst(SelectInst &SI) {
307 // If the condition being selected on is a constant or the same value is
308 // being selected between, fold the select. Yes this does (rarely) happen
310 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
311 return SI.getOperand(1+CI->isZero());
312 if (SI.getOperand(1) == SI.getOperand(2))
313 return SI.getOperand(1);
318 /// \brief A helper that folds a PHI node or a select.
319 static Value *foldPHINodeOrSelectInst(Instruction &I) {
320 if (PHINode *PN = dyn_cast<PHINode>(&I)) {
321 // If PN merges together the same value, return that value.
322 return PN->hasConstantValue();
324 return foldSelectInst(cast<SelectInst>(I));
327 /// \brief Builder for the alloca slices.
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.
331 class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
332 friend class PtrUseVisitor<SliceBuilder>;
333 friend class InstVisitor<SliceBuilder>;
334 typedef PtrUseVisitor<SliceBuilder> Base;
336 const uint64_t AllocSize;
339 SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
340 SmallDenseMap<Instruction *, uint64_t> PHIOrSelectSizes;
342 /// \brief Set to de-duplicate dead instructions found in the use walk.
343 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
346 SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &S)
347 : PtrUseVisitor<SliceBuilder>(DL),
348 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())), S(S) {}
351 void markAsDead(Instruction &I) {
352 if (VisitedDeadInsts.insert(&I))
353 S.DeadUsers.push_back(&I);
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);
369 uint64_t BeginOffset = Offset.getZExtValue();
370 uint64_t EndOffset = BeginOffset + Size;
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;
387 S.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
390 void visitBitCastInst(BitCastInst &BC) {
392 return markAsDead(BC);
394 return Base::visitBitCastInst(BC);
397 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
398 if (GEPI.use_empty())
399 return markAsDead(GEPI);
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);
413 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
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);
422 APInt(Offset.getBitWidth(), SL->getElementOffset(ElementIdx));
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()));
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);
438 return Base::visitGetElementPtrInst(GEPI);
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
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.
452 Ty->isIntegerTy() && !IsVolatile && Offset == 0 && Size >= AllocSize;
454 insertUse(I, Offset, Size, IsSplittable);
457 void visitLoadInst(LoadInst &LI) {
458 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
459 "All simple FCA loads should have been pre-split");
462 return PI.setAborted(&LI);
464 uint64_t Size = DL.getTypeStoreSize(LI.getType());
465 return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
468 void visitStoreInst(StoreInst &SI) {
469 Value *ValOp = SI.getValueOperand();
471 return PI.setEscapedAndAborted(&SI);
473 return PI.setAborted(&SI);
475 uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
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
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
488 << " alloca: " << S.AI << "\n"
489 << " use: " << SI << "\n");
490 return markAsDead(SI);
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());
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);
508 return PI.setAborted(&II);
510 insertUse(II, Offset,
511 Length ? Length->getLimitedValue()
512 : AllocSize - Offset.getLimitedValue(),
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);
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))
528 return PI.setAborted(&II);
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);
542 uint64_t RawOffset = Offset.getLimitedValue();
543 uint64_t Size = Length ? Length->getLimitedValue()
544 : AllocSize - RawOffset;
546 // Check for the special case where the same exact value is used for both
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);
553 return insertUse(II, Offset, Size, /*IsSplittable=*/false);
556 // If we have seen both source and destination for a mem transfer, then
557 // they both point to the same alloca.
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;
564 Slice &PrevP = S.Slices[PrevIdx];
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) {
570 return markAsDead(II);
573 // Otherwise we have an offset transfer within the same alloca. We can't
575 PrevP.makeUnsplittable();
578 // Insert the use now that we've fixed up the splittable nature.
579 insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
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.");
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) {
591 return PI.setAborted(&II);
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);
602 Base::visitIntrinsicInst(II);
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
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.
618 Instruction *I, *UsedI;
619 std::tie(UsedI, I) = Uses.pop_back_val();
621 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
622 Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
625 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
626 Value *Op = SI->getOperand(0);
629 Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
633 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
634 if (!GEP->hasAllZeroIndices())
636 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
637 !isa<SelectInst>(I)) {
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());
649 void visitPHINodeOrSelectInst(Instruction &I) {
650 assert(isa<PHINode>(I) || isa<SelectInst>(I));
652 return markAsDead(I);
654 // TODO: We could use SimplifyInstruction here to fold PHINodes and
655 // SelectInsts. However, doing so requires to change the current
656 // dead-operand-tracking mechanism. For instance, suppose neither loading
657 // from %U nor %other traps. Then "load (select undef, %U, %other)" does not
658 // trap either. However, if we simply replace %U with undef using the
659 // current dead-operand-tracking mechanism, "load (select undef, undef,
660 // %other)" may trap because the select may return the first operand
662 if (Value *Result = foldPHINodeOrSelectInst(I)) {
664 // If the result of the constant fold will be the pointer, recurse
665 // through the PHI/select as if we had RAUW'ed it.
668 // Otherwise the operand to the PHI/select is dead, and we can replace
670 S.DeadOperands.push_back(U);
676 return PI.setAborted(&I);
678 // See if we already have computed info on this node.
679 uint64_t &Size = PHIOrSelectSizes[&I];
681 // This is a new PHI/Select, check for an unsafe use of it.
682 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&I, Size))
683 return PI.setAborted(UnsafeI);
686 // For PHI and select operands outside the alloca, we can't nuke the entire
687 // phi or select -- the other side might still be relevant, so we special
688 // case them here and use a separate structure to track the operands
689 // themselves which should be replaced with undef.
690 // FIXME: This should instead be escaped in the event we're instrumenting
691 // for address sanitization.
692 if (Offset.uge(AllocSize)) {
693 S.DeadOperands.push_back(U);
697 insertUse(I, Offset, Size);
700 void visitPHINode(PHINode &PN) {
701 visitPHINodeOrSelectInst(PN);
704 void visitSelectInst(SelectInst &SI) {
705 visitPHINodeOrSelectInst(SI);
708 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
709 void visitInstruction(Instruction &I) {
714 AllocaSlices::AllocaSlices(const DataLayout &DL, AllocaInst &AI)
716 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
719 PointerEscapingInstr(nullptr) {
720 SliceBuilder PB(DL, AI, *this);
721 SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
722 if (PtrI.isEscaped() || PtrI.isAborted()) {
723 // FIXME: We should sink the escape vs. abort info into the caller nicely,
724 // possibly by just storing the PtrInfo in the AllocaSlices.
725 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
726 : PtrI.getAbortingInst();
727 assert(PointerEscapingInstr && "Did not track a bad instruction");
731 Slices.erase(std::remove_if(Slices.begin(), Slices.end(),
732 std::mem_fun_ref(&Slice::isDead)),
735 #if __cplusplus >= 201103L && !defined(NDEBUG)
736 if (SROARandomShuffleSlices) {
737 std::mt19937 MT(static_cast<unsigned>(sys::TimeValue::now().msec()));
738 std::shuffle(Slices.begin(), Slices.end(), MT);
742 // Sort the uses. This arranges for the offsets to be in ascending order,
743 // and the sizes to be in descending order.
744 std::sort(Slices.begin(), Slices.end());
747 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
749 void AllocaSlices::print(raw_ostream &OS, const_iterator I,
750 StringRef Indent) const {
751 printSlice(OS, I, Indent);
752 printUse(OS, I, Indent);
755 void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
756 StringRef Indent) const {
757 OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
758 << " slice #" << (I - begin())
759 << (I->isSplittable() ? " (splittable)" : "") << "\n";
762 void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
763 StringRef Indent) const {
764 OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
767 void AllocaSlices::print(raw_ostream &OS) const {
768 if (PointerEscapingInstr) {
769 OS << "Can't analyze slices for alloca: " << AI << "\n"
770 << " A pointer to this alloca escaped by:\n"
771 << " " << *PointerEscapingInstr << "\n";
775 OS << "Slices of alloca: " << AI << "\n";
776 for (const_iterator I = begin(), E = end(); I != E; ++I)
780 LLVM_DUMP_METHOD void AllocaSlices::dump(const_iterator I) const {
783 LLVM_DUMP_METHOD void AllocaSlices::dump() const { print(dbgs()); }
785 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
788 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
790 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
791 /// the loads and stores of an alloca instruction, as well as updating its
792 /// debug information. This is used when a domtree is unavailable and thus
793 /// mem2reg in its full form can't be used to handle promotion of allocas to
795 class AllocaPromoter : public LoadAndStorePromoter {
799 SmallVector<DbgDeclareInst *, 4> DDIs;
800 SmallVector<DbgValueInst *, 4> DVIs;
803 AllocaPromoter(const SmallVectorImpl<Instruction *> &Insts, SSAUpdater &S,
804 AllocaInst &AI, DIBuilder &DIB)
805 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
807 void run(const SmallVectorImpl<Instruction*> &Insts) {
808 // Retain the debug information attached to the alloca for use when
809 // rewriting loads and stores.
810 if (MDNode *DebugNode = MDNode::getIfExists(AI.getContext(), &AI)) {
811 for (User *U : DebugNode->users())
812 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U))
814 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(U))
818 LoadAndStorePromoter::run(Insts);
820 // While we have the debug information, clear it off of the alloca. The
821 // caller takes care of deleting the alloca.
822 while (!DDIs.empty())
823 DDIs.pop_back_val()->eraseFromParent();
824 while (!DVIs.empty())
825 DVIs.pop_back_val()->eraseFromParent();
828 bool isInstInList(Instruction *I,
829 const SmallVectorImpl<Instruction*> &Insts) const override {
831 if (LoadInst *LI = dyn_cast<LoadInst>(I))
832 Ptr = LI->getOperand(0);
834 Ptr = cast<StoreInst>(I)->getPointerOperand();
836 // Only used to detect cycles, which will be rare and quickly found as
837 // we're walking up a chain of defs rather than down through uses.
838 SmallPtrSet<Value *, 4> Visited;
844 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr))
845 Ptr = BCI->getOperand(0);
846 else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Ptr))
847 Ptr = GEPI->getPointerOperand();
851 } while (Visited.insert(Ptr));
856 void updateDebugInfo(Instruction *Inst) const override {
857 for (DbgDeclareInst *DDI : DDIs)
858 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
859 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
860 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
861 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
862 for (DbgValueInst *DVI : DVIs) {
863 Value *Arg = nullptr;
864 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
865 // If an argument is zero extended then use argument directly. The ZExt
866 // may be zapped by an optimization pass in future.
867 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
868 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
869 else if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
870 Arg = dyn_cast<Argument>(SExt->getOperand(0));
872 Arg = SI->getValueOperand();
873 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
874 Arg = LI->getPointerOperand();
878 Instruction *DbgVal =
879 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
880 DIExpression(DVI->getExpression()), Inst);
881 DbgVal->setDebugLoc(DVI->getDebugLoc());
885 } // end anon namespace
889 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
891 /// This pass takes allocations which can be completely analyzed (that is, they
892 /// don't escape) and tries to turn them into scalar SSA values. There are
893 /// a few steps to this process.
895 /// 1) It takes allocations of aggregates and analyzes the ways in which they
896 /// are used to try to split them into smaller allocations, ideally of
897 /// a single scalar data type. It will split up memcpy and memset accesses
898 /// as necessary and try to isolate individual scalar accesses.
899 /// 2) It will transform accesses into forms which are suitable for SSA value
900 /// promotion. This can be replacing a memset with a scalar store of an
901 /// integer value, or it can involve speculating operations on a PHI or
902 /// select to be a PHI or select of the results.
903 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
904 /// onto insert and extract operations on a vector value, and convert them to
905 /// this form. By doing so, it will enable promotion of vector aggregates to
906 /// SSA vector values.
907 class SROA : public FunctionPass {
908 const bool RequiresDomTree;
911 const DataLayout *DL;
913 AssumptionTracker *AT;
915 /// \brief Worklist of alloca instructions to simplify.
917 /// Each alloca in the function is added to this. Each new alloca formed gets
918 /// added to it as well to recursively simplify unless that alloca can be
919 /// directly promoted. Finally, each time we rewrite a use of an alloca other
920 /// the one being actively rewritten, we add it back onto the list if not
921 /// already present to ensure it is re-visited.
922 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > Worklist;
924 /// \brief A collection of instructions to delete.
925 /// We try to batch deletions to simplify code and make things a bit more
927 SetVector<Instruction *, SmallVector<Instruction *, 8> > DeadInsts;
929 /// \brief Post-promotion worklist.
931 /// Sometimes we discover an alloca which has a high probability of becoming
932 /// viable for SROA after a round of promotion takes place. In those cases,
933 /// the alloca is enqueued here for re-processing.
935 /// Note that we have to be very careful to clear allocas out of this list in
936 /// the event they are deleted.
937 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > PostPromotionWorklist;
939 /// \brief A collection of alloca instructions we can directly promote.
940 std::vector<AllocaInst *> PromotableAllocas;
942 /// \brief A worklist of PHIs to speculate prior to promoting allocas.
944 /// All of these PHIs have been checked for the safety of speculation and by
945 /// being speculated will allow promoting allocas currently in the promotable
947 SetVector<PHINode *, SmallVector<PHINode *, 2> > SpeculatablePHIs;
949 /// \brief A worklist of select instructions to speculate prior to promoting
952 /// All of these select instructions have been checked for the safety of
953 /// speculation and by being speculated will allow promoting allocas
954 /// currently in the promotable queue.
955 SetVector<SelectInst *, SmallVector<SelectInst *, 2> > SpeculatableSelects;
958 SROA(bool RequiresDomTree = true)
959 : FunctionPass(ID), RequiresDomTree(RequiresDomTree),
960 C(nullptr), DL(nullptr), DT(nullptr) {
961 initializeSROAPass(*PassRegistry::getPassRegistry());
963 bool runOnFunction(Function &F) override;
964 void getAnalysisUsage(AnalysisUsage &AU) const override;
966 const char *getPassName() const override { return "SROA"; }
970 friend class PHIOrSelectSpeculator;
971 friend class AllocaSliceRewriter;
973 bool rewritePartition(AllocaInst &AI, AllocaSlices &S,
974 AllocaSlices::iterator B, AllocaSlices::iterator E,
975 int64_t BeginOffset, int64_t EndOffset,
976 ArrayRef<AllocaSlices::iterator> SplitUses);
977 bool splitAlloca(AllocaInst &AI, AllocaSlices &S);
978 bool runOnAlloca(AllocaInst &AI);
979 void clobberUse(Use &U);
980 void deleteDeadInstructions(SmallPtrSetImpl<AllocaInst *> &DeletedAllocas);
981 bool promoteAllocas(Function &F);
987 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
988 return new SROA(RequiresDomTree);
991 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates",
993 INITIALIZE_PASS_DEPENDENCY(AssumptionTracker)
994 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
995 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
998 /// Walk the range of a partitioning looking for a common type to cover this
999 /// sequence of slices.
1000 static Type *findCommonType(AllocaSlices::const_iterator B,
1001 AllocaSlices::const_iterator E,
1002 uint64_t EndOffset) {
1004 bool TyIsCommon = true;
1005 IntegerType *ITy = nullptr;
1007 // Note that we need to look at *every* alloca slice's Use to ensure we
1008 // always get consistent results regardless of the order of slices.
1009 for (AllocaSlices::const_iterator I = B; I != E; ++I) {
1010 Use *U = I->getUse();
1011 if (isa<IntrinsicInst>(*U->getUser()))
1013 if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
1016 Type *UserTy = nullptr;
1017 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1018 UserTy = LI->getType();
1019 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1020 UserTy = SI->getValueOperand()->getType();
1023 if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
1024 // If the type is larger than the partition, skip it. We only encounter
1025 // this for split integer operations where we want to use the type of the
1026 // entity causing the split. Also skip if the type is not a byte width
1028 if (UserITy->getBitWidth() % 8 != 0 ||
1029 UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
1032 // Track the largest bitwidth integer type used in this way in case there
1033 // is no common type.
1034 if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
1038 // To avoid depending on the order of slices, Ty and TyIsCommon must not
1039 // depend on types skipped above.
1040 if (!UserTy || (Ty && Ty != UserTy))
1041 TyIsCommon = false; // Give up on anything but an iN type.
1046 return TyIsCommon ? Ty : ITy;
1049 /// PHI instructions that use an alloca and are subsequently loaded can be
1050 /// rewritten to load both input pointers in the pred blocks and then PHI the
1051 /// results, allowing the load of the alloca to be promoted.
1053 /// %P2 = phi [i32* %Alloca, i32* %Other]
1054 /// %V = load i32* %P2
1056 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1058 /// %V2 = load i32* %Other
1060 /// %V = phi [i32 %V1, i32 %V2]
1062 /// We can do this to a select if its only uses are loads and if the operands
1063 /// to the select can be loaded unconditionally.
1065 /// FIXME: This should be hoisted into a generic utility, likely in
1066 /// Transforms/Util/Local.h
1067 static bool isSafePHIToSpeculate(PHINode &PN,
1068 const DataLayout *DL = nullptr) {
1069 // For now, we can only do this promotion if the load is in the same block
1070 // as the PHI, and if there are no stores between the phi and load.
1071 // TODO: Allow recursive phi users.
1072 // TODO: Allow stores.
1073 BasicBlock *BB = PN.getParent();
1074 unsigned MaxAlign = 0;
1075 bool HaveLoad = false;
1076 for (User *U : PN.users()) {
1077 LoadInst *LI = dyn_cast<LoadInst>(U);
1078 if (!LI || !LI->isSimple())
1081 // For now we only allow loads in the same block as the PHI. This is
1082 // a common case that happens when instcombine merges two loads through
1084 if (LI->getParent() != BB)
1087 // Ensure that there are no instructions between the PHI and the load that
1089 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1090 if (BBI->mayWriteToMemory())
1093 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1100 // We can only transform this if it is safe to push the loads into the
1101 // predecessor blocks. The only thing to watch out for is that we can't put
1102 // a possibly trapping load in the predecessor if it is a critical edge.
1103 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1104 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1105 Value *InVal = PN.getIncomingValue(Idx);
1107 // If the value is produced by the terminator of the predecessor (an
1108 // invoke) or it has side-effects, there is no valid place to put a load
1109 // in the predecessor.
1110 if (TI == InVal || TI->mayHaveSideEffects())
1113 // If the predecessor has a single successor, then the edge isn't
1115 if (TI->getNumSuccessors() == 1)
1118 // If this pointer is always safe to load, or if we can prove that there
1119 // is already a load in the block, then we can move the load to the pred
1121 if (InVal->isDereferenceablePointer(DL) ||
1122 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, DL))
1131 static void speculatePHINodeLoads(PHINode &PN) {
1132 DEBUG(dbgs() << " original: " << PN << "\n");
1134 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1135 IRBuilderTy PHIBuilder(&PN);
1136 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1137 PN.getName() + ".sroa.speculated");
1139 // Get the AA tags and alignment to use from one of the loads. It doesn't
1140 // matter which one we get and if any differ.
1141 LoadInst *SomeLoad = cast<LoadInst>(PN.user_back());
1144 SomeLoad->getAAMetadata(AATags);
1145 unsigned Align = SomeLoad->getAlignment();
1147 // Rewrite all loads of the PN to use the new PHI.
1148 while (!PN.use_empty()) {
1149 LoadInst *LI = cast<LoadInst>(PN.user_back());
1150 LI->replaceAllUsesWith(NewPN);
1151 LI->eraseFromParent();
1154 // Inject loads into all of the pred blocks.
1155 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1156 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1157 TerminatorInst *TI = Pred->getTerminator();
1158 Value *InVal = PN.getIncomingValue(Idx);
1159 IRBuilderTy PredBuilder(TI);
1161 LoadInst *Load = PredBuilder.CreateLoad(
1162 InVal, (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1163 ++NumLoadsSpeculated;
1164 Load->setAlignment(Align);
1166 Load->setAAMetadata(AATags);
1167 NewPN->addIncoming(Load, Pred);
1170 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1171 PN.eraseFromParent();
1174 /// Select instructions that use an alloca and are subsequently loaded can be
1175 /// rewritten to load both input pointers and then select between the result,
1176 /// allowing the load of the alloca to be promoted.
1178 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1179 /// %V = load i32* %P2
1181 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1182 /// %V2 = load i32* %Other
1183 /// %V = select i1 %cond, i32 %V1, i32 %V2
1185 /// We can do this to a select if its only uses are loads and if the operand
1186 /// to the select can be loaded unconditionally.
1187 static bool isSafeSelectToSpeculate(SelectInst &SI,
1188 const DataLayout *DL = nullptr) {
1189 Value *TValue = SI.getTrueValue();
1190 Value *FValue = SI.getFalseValue();
1191 bool TDerefable = TValue->isDereferenceablePointer(DL);
1192 bool FDerefable = FValue->isDereferenceablePointer(DL);
1194 for (User *U : SI.users()) {
1195 LoadInst *LI = dyn_cast<LoadInst>(U);
1196 if (!LI || !LI->isSimple())
1199 // Both operands to the select need to be dereferencable, either
1200 // absolutely (e.g. allocas) or at this point because we can see other
1203 !isSafeToLoadUnconditionally(TValue, LI, LI->getAlignment(), DL))
1206 !isSafeToLoadUnconditionally(FValue, LI, LI->getAlignment(), DL))
1213 static void speculateSelectInstLoads(SelectInst &SI) {
1214 DEBUG(dbgs() << " original: " << SI << "\n");
1216 IRBuilderTy IRB(&SI);
1217 Value *TV = SI.getTrueValue();
1218 Value *FV = SI.getFalseValue();
1219 // Replace the loads of the select with a select of two loads.
1220 while (!SI.use_empty()) {
1221 LoadInst *LI = cast<LoadInst>(SI.user_back());
1222 assert(LI->isSimple() && "We only speculate simple loads");
1224 IRB.SetInsertPoint(LI);
1226 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1228 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1229 NumLoadsSpeculated += 2;
1231 // Transfer alignment and AA info if present.
1232 TL->setAlignment(LI->getAlignment());
1233 FL->setAlignment(LI->getAlignment());
1236 LI->getAAMetadata(Tags);
1238 TL->setAAMetadata(Tags);
1239 FL->setAAMetadata(Tags);
1242 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1243 LI->getName() + ".sroa.speculated");
1245 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1246 LI->replaceAllUsesWith(V);
1247 LI->eraseFromParent();
1249 SI.eraseFromParent();
1252 /// \brief Build a GEP out of a base pointer and indices.
1254 /// This will return the BasePtr if that is valid, or build a new GEP
1255 /// instruction using the IRBuilder if GEP-ing is needed.
1256 static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1257 SmallVectorImpl<Value *> &Indices, Twine NamePrefix) {
1258 if (Indices.empty())
1261 // A single zero index is a no-op, so check for this and avoid building a GEP
1263 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1266 return IRB.CreateInBoundsGEP(BasePtr, Indices, NamePrefix + "sroa_idx");
1269 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1270 /// TargetTy without changing the offset of the pointer.
1272 /// This routine assumes we've already established a properly offset GEP with
1273 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1274 /// zero-indices down through type layers until we find one the same as
1275 /// TargetTy. If we can't find one with the same type, we at least try to use
1276 /// one with the same size. If none of that works, we just produce the GEP as
1277 /// indicated by Indices to have the correct offset.
1278 static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1279 Value *BasePtr, Type *Ty, Type *TargetTy,
1280 SmallVectorImpl<Value *> &Indices,
1283 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1285 // Pointer size to use for the indices.
1286 unsigned PtrSize = DL.getPointerTypeSizeInBits(BasePtr->getType());
1288 // See if we can descend into a struct and locate a field with the correct
1290 unsigned NumLayers = 0;
1291 Type *ElementTy = Ty;
1293 if (ElementTy->isPointerTy())
1296 if (ArrayType *ArrayTy = dyn_cast<ArrayType>(ElementTy)) {
1297 ElementTy = ArrayTy->getElementType();
1298 Indices.push_back(IRB.getIntN(PtrSize, 0));
1299 } else if (VectorType *VectorTy = dyn_cast<VectorType>(ElementTy)) {
1300 ElementTy = VectorTy->getElementType();
1301 Indices.push_back(IRB.getInt32(0));
1302 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1303 if (STy->element_begin() == STy->element_end())
1304 break; // Nothing left to descend into.
1305 ElementTy = *STy->element_begin();
1306 Indices.push_back(IRB.getInt32(0));
1311 } while (ElementTy != TargetTy);
1312 if (ElementTy != TargetTy)
1313 Indices.erase(Indices.end() - NumLayers, Indices.end());
1315 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1318 /// \brief Recursively compute indices for a natural GEP.
1320 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1321 /// element types adding appropriate indices for the GEP.
1322 static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1323 Value *Ptr, Type *Ty, APInt &Offset,
1325 SmallVectorImpl<Value *> &Indices,
1328 return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices, NamePrefix);
1330 // We can't recurse through pointer types.
1331 if (Ty->isPointerTy())
1334 // We try to analyze GEPs over vectors here, but note that these GEPs are
1335 // extremely poorly defined currently. The long-term goal is to remove GEPing
1336 // over a vector from the IR completely.
1337 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1338 unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
1339 if (ElementSizeInBits % 8 != 0) {
1340 // GEPs over non-multiple of 8 size vector elements are invalid.
1343 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1344 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1345 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1347 Offset -= NumSkippedElements * ElementSize;
1348 Indices.push_back(IRB.getInt(NumSkippedElements));
1349 return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1350 Offset, TargetTy, Indices, NamePrefix);
1353 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1354 Type *ElementTy = ArrTy->getElementType();
1355 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1356 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1357 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1360 Offset -= NumSkippedElements * ElementSize;
1361 Indices.push_back(IRB.getInt(NumSkippedElements));
1362 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1363 Indices, NamePrefix);
1366 StructType *STy = dyn_cast<StructType>(Ty);
1370 const StructLayout *SL = DL.getStructLayout(STy);
1371 uint64_t StructOffset = Offset.getZExtValue();
1372 if (StructOffset >= SL->getSizeInBytes())
1374 unsigned Index = SL->getElementContainingOffset(StructOffset);
1375 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1376 Type *ElementTy = STy->getElementType(Index);
1377 if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
1378 return nullptr; // The offset points into alignment padding.
1380 Indices.push_back(IRB.getInt32(Index));
1381 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1382 Indices, NamePrefix);
1385 /// \brief Get a natural GEP from a base pointer to a particular offset and
1386 /// resulting in a particular type.
1388 /// The goal is to produce a "natural" looking GEP that works with the existing
1389 /// composite types to arrive at the appropriate offset and element type for
1390 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1391 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1392 /// Indices, and setting Ty to the result subtype.
1394 /// If no natural GEP can be constructed, this function returns null.
1395 static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1396 Value *Ptr, APInt Offset, Type *TargetTy,
1397 SmallVectorImpl<Value *> &Indices,
1399 PointerType *Ty = cast<PointerType>(Ptr->getType());
1401 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1403 if (Ty == IRB.getInt8PtrTy(Ty->getAddressSpace()) && TargetTy->isIntegerTy(8))
1406 Type *ElementTy = Ty->getElementType();
1407 if (!ElementTy->isSized())
1408 return nullptr; // We can't GEP through an unsized element.
1409 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1410 if (ElementSize == 0)
1411 return nullptr; // Zero-length arrays can't help us build a natural GEP.
1412 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1414 Offset -= NumSkippedElements * ElementSize;
1415 Indices.push_back(IRB.getInt(NumSkippedElements));
1416 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1417 Indices, NamePrefix);
1420 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1421 /// resulting pointer has PointerTy.
1423 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1424 /// and produces the pointer type desired. Where it cannot, it will try to use
1425 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1426 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1427 /// bitcast to the type.
1429 /// The strategy for finding the more natural GEPs is to peel off layers of the
1430 /// pointer, walking back through bit casts and GEPs, searching for a base
1431 /// pointer from which we can compute a natural GEP with the desired
1432 /// properties. The algorithm tries to fold as many constant indices into
1433 /// a single GEP as possible, thus making each GEP more independent of the
1434 /// surrounding code.
1435 static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr,
1436 APInt Offset, Type *PointerTy,
1438 // Even though we don't look through PHI nodes, we could be called on an
1439 // instruction in an unreachable block, which may be on a cycle.
1440 SmallPtrSet<Value *, 4> Visited;
1441 Visited.insert(Ptr);
1442 SmallVector<Value *, 4> Indices;
1444 // We may end up computing an offset pointer that has the wrong type. If we
1445 // never are able to compute one directly that has the correct type, we'll
1446 // fall back to it, so keep it around here.
1447 Value *OffsetPtr = nullptr;
1449 // Remember any i8 pointer we come across to re-use if we need to do a raw
1451 Value *Int8Ptr = nullptr;
1452 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1454 Type *TargetTy = PointerTy->getPointerElementType();
1457 // First fold any existing GEPs into the offset.
1458 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1459 APInt GEPOffset(Offset.getBitWidth(), 0);
1460 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1462 Offset += GEPOffset;
1463 Ptr = GEP->getPointerOperand();
1464 if (!Visited.insert(Ptr))
1468 // See if we can perform a natural GEP here.
1470 if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1471 Indices, NamePrefix)) {
1472 if (P->getType() == PointerTy) {
1473 // Zap any offset pointer that we ended up computing in previous rounds.
1474 if (OffsetPtr && OffsetPtr->use_empty())
1475 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1476 I->eraseFromParent();
1484 // Stash this pointer if we've found an i8*.
1485 if (Ptr->getType()->isIntegerTy(8)) {
1487 Int8PtrOffset = Offset;
1490 // Peel off a layer of the pointer and update the offset appropriately.
1491 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1492 Ptr = cast<Operator>(Ptr)->getOperand(0);
1493 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1494 if (GA->mayBeOverridden())
1496 Ptr = GA->getAliasee();
1500 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1501 } while (Visited.insert(Ptr));
1505 Int8Ptr = IRB.CreateBitCast(
1506 Ptr, IRB.getInt8PtrTy(PointerTy->getPointerAddressSpace()),
1507 NamePrefix + "sroa_raw_cast");
1508 Int8PtrOffset = Offset;
1511 OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
1512 IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1513 NamePrefix + "sroa_raw_idx");
1517 // On the off chance we were targeting i8*, guard the bitcast here.
1518 if (Ptr->getType() != PointerTy)
1519 Ptr = IRB.CreateBitCast(Ptr, PointerTy, NamePrefix + "sroa_cast");
1524 /// \brief Test whether we can convert a value from the old to the new type.
1526 /// This predicate should be used to guard calls to convertValue in order to
1527 /// ensure that we only try to convert viable values. The strategy is that we
1528 /// will peel off single element struct and array wrappings to get to an
1529 /// underlying value, and convert that value.
1530 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1533 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1534 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1535 if (NewITy->getBitWidth() >= OldITy->getBitWidth())
1537 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1539 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1542 // We can convert pointers to integers and vice-versa. Same for vectors
1543 // of pointers and integers.
1544 OldTy = OldTy->getScalarType();
1545 NewTy = NewTy->getScalarType();
1546 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1547 if (NewTy->isPointerTy() && OldTy->isPointerTy())
1549 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
1557 /// \brief Generic routine to convert an SSA value to a value of a different
1560 /// This will try various different casting techniques, such as bitcasts,
1561 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1562 /// two types for viability with this routine.
1563 static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1565 Type *OldTy = V->getType();
1566 assert(canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type");
1571 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1572 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1573 if (NewITy->getBitWidth() > OldITy->getBitWidth())
1574 return IRB.CreateZExt(V, NewITy);
1576 // See if we need inttoptr for this type pair. A cast involving both scalars
1577 // and vectors requires and additional bitcast.
1578 if (OldTy->getScalarType()->isIntegerTy() &&
1579 NewTy->getScalarType()->isPointerTy()) {
1580 // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1581 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1582 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1585 // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1586 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1587 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1590 return IRB.CreateIntToPtr(V, NewTy);
1593 // See if we need ptrtoint for this type pair. A cast involving both scalars
1594 // and vectors requires and additional bitcast.
1595 if (OldTy->getScalarType()->isPointerTy() &&
1596 NewTy->getScalarType()->isIntegerTy()) {
1597 // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1598 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1599 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1602 // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1603 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1604 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1607 return IRB.CreatePtrToInt(V, NewTy);
1610 return IRB.CreateBitCast(V, NewTy);
1613 /// \brief Test whether the given slice use can be promoted to a vector.
1615 /// This function is called to test each entry in a partioning which is slated
1616 /// for a single slice.
1618 isVectorPromotionViableForSlice(const DataLayout &DL, uint64_t SliceBeginOffset,
1619 uint64_t SliceEndOffset, VectorType *Ty,
1620 uint64_t ElementSize, const Slice &S) {
1621 // First validate the slice offsets.
1622 uint64_t BeginOffset =
1623 std::max(S.beginOffset(), SliceBeginOffset) - SliceBeginOffset;
1624 uint64_t BeginIndex = BeginOffset / ElementSize;
1625 if (BeginIndex * ElementSize != BeginOffset ||
1626 BeginIndex >= Ty->getNumElements())
1628 uint64_t EndOffset =
1629 std::min(S.endOffset(), SliceEndOffset) - SliceBeginOffset;
1630 uint64_t EndIndex = EndOffset / ElementSize;
1631 if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
1634 assert(EndIndex > BeginIndex && "Empty vector!");
1635 uint64_t NumElements = EndIndex - BeginIndex;
1636 Type *SliceTy = (NumElements == 1)
1637 ? Ty->getElementType()
1638 : VectorType::get(Ty->getElementType(), NumElements);
1641 Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1643 Use *U = S.getUse();
1645 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1646 if (MI->isVolatile())
1648 if (!S.isSplittable())
1649 return false; // Skip any unsplittable intrinsics.
1650 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1651 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1652 II->getIntrinsicID() != Intrinsic::lifetime_end)
1654 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1655 // Disable vector promotion when there are loads or stores of an FCA.
1657 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1658 if (LI->isVolatile())
1660 Type *LTy = LI->getType();
1661 if (SliceBeginOffset > S.beginOffset() || SliceEndOffset < S.endOffset()) {
1662 assert(LTy->isIntegerTy());
1665 if (!canConvertValue(DL, SliceTy, LTy))
1667 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1668 if (SI->isVolatile())
1670 Type *STy = SI->getValueOperand()->getType();
1671 if (SliceBeginOffset > S.beginOffset() || SliceEndOffset < S.endOffset()) {
1672 assert(STy->isIntegerTy());
1675 if (!canConvertValue(DL, STy, SliceTy))
1684 /// \brief Test whether the given alloca partitioning and range of slices can be
1685 /// promoted to a vector.
1687 /// This is a quick test to check whether we can rewrite a particular alloca
1688 /// partition (and its newly formed alloca) into a vector alloca with only
1689 /// whole-vector loads and stores such that it could be promoted to a vector
1690 /// SSA value. We only can ensure this for a limited set of operations, and we
1691 /// don't want to do the rewrites unless we are confident that the result will
1692 /// be promotable, so we have an early test here.
1694 isVectorPromotionViable(const DataLayout &DL, Type *AllocaTy,
1695 uint64_t SliceBeginOffset, uint64_t SliceEndOffset,
1696 AllocaSlices::const_range Slices,
1697 ArrayRef<AllocaSlices::iterator> SplitUses) {
1698 VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
1702 uint64_t ElementSize = DL.getTypeSizeInBits(Ty->getScalarType());
1704 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1705 // that aren't byte sized.
1706 if (ElementSize % 8)
1708 assert((DL.getTypeSizeInBits(Ty) % 8) == 0 &&
1709 "vector size not a multiple of element size?");
1712 for (const auto &S : Slices)
1713 if (!isVectorPromotionViableForSlice(DL, SliceBeginOffset, SliceEndOffset,
1714 Ty, ElementSize, S))
1717 for (const auto &SI : SplitUses)
1718 if (!isVectorPromotionViableForSlice(DL, SliceBeginOffset, SliceEndOffset,
1719 Ty, ElementSize, *SI))
1725 /// \brief Test whether a slice of an alloca is valid for integer widening.
1727 /// This implements the necessary checking for the \c isIntegerWideningViable
1728 /// test below on a single slice of the alloca.
1729 static bool isIntegerWideningViableForSlice(const DataLayout &DL,
1731 uint64_t AllocBeginOffset,
1734 bool &WholeAllocaOp) {
1735 uint64_t RelBegin = S.beginOffset() - AllocBeginOffset;
1736 uint64_t RelEnd = S.endOffset() - AllocBeginOffset;
1738 // We can't reasonably handle cases where the load or store extends past
1739 // the end of the aloca's type and into its padding.
1743 Use *U = S.getUse();
1745 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1746 if (LI->isVolatile())
1748 if (RelBegin == 0 && RelEnd == Size)
1749 WholeAllocaOp = true;
1750 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
1751 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1753 } else if (RelBegin != 0 || RelEnd != Size ||
1754 !canConvertValue(DL, AllocaTy, LI->getType())) {
1755 // Non-integer loads need to be convertible from the alloca type so that
1756 // they are promotable.
1759 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1760 Type *ValueTy = SI->getValueOperand()->getType();
1761 if (SI->isVolatile())
1763 if (RelBegin == 0 && RelEnd == Size)
1764 WholeAllocaOp = true;
1765 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
1766 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1768 } else if (RelBegin != 0 || RelEnd != Size ||
1769 !canConvertValue(DL, ValueTy, AllocaTy)) {
1770 // Non-integer stores need to be convertible to the alloca type so that
1771 // they are promotable.
1774 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1775 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
1777 if (!S.isSplittable())
1778 return false; // Skip any unsplittable intrinsics.
1779 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1780 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1781 II->getIntrinsicID() != Intrinsic::lifetime_end)
1790 /// \brief Test whether the given alloca partition's integer operations can be
1791 /// widened to promotable ones.
1793 /// This is a quick test to check whether we can rewrite the integer loads and
1794 /// stores to a particular alloca into wider loads and stores and be able to
1795 /// promote the resulting alloca.
1797 isIntegerWideningViable(const DataLayout &DL, Type *AllocaTy,
1798 uint64_t AllocBeginOffset,
1799 AllocaSlices::const_range Slices,
1800 ArrayRef<AllocaSlices::iterator> SplitUses) {
1801 uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
1802 // Don't create integer types larger than the maximum bitwidth.
1803 if (SizeInBits > IntegerType::MAX_INT_BITS)
1806 // Don't try to handle allocas with bit-padding.
1807 if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
1810 // We need to ensure that an integer type with the appropriate bitwidth can
1811 // be converted to the alloca type, whatever that is. We don't want to force
1812 // the alloca itself to have an integer type if there is a more suitable one.
1813 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
1814 if (!canConvertValue(DL, AllocaTy, IntTy) ||
1815 !canConvertValue(DL, IntTy, AllocaTy))
1818 uint64_t Size = DL.getTypeStoreSize(AllocaTy);
1820 // While examining uses, we ensure that the alloca has a covering load or
1821 // store. We don't want to widen the integer operations only to fail to
1822 // promote due to some other unsplittable entry (which we may make splittable
1823 // later). However, if there are only splittable uses, go ahead and assume
1824 // that we cover the alloca.
1825 bool WholeAllocaOp =
1826 Slices.begin() != Slices.end() ? false : DL.isLegalInteger(SizeInBits);
1828 for (const auto &S : Slices)
1829 if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
1833 for (const auto &SI : SplitUses)
1834 if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
1835 *SI, WholeAllocaOp))
1838 return WholeAllocaOp;
1841 static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1842 IntegerType *Ty, uint64_t Offset,
1843 const Twine &Name) {
1844 DEBUG(dbgs() << " start: " << *V << "\n");
1845 IntegerType *IntTy = cast<IntegerType>(V->getType());
1846 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
1847 "Element extends past full value");
1848 uint64_t ShAmt = 8*Offset;
1849 if (DL.isBigEndian())
1850 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
1852 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
1853 DEBUG(dbgs() << " shifted: " << *V << "\n");
1855 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
1856 "Cannot extract to a larger integer!");
1858 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
1859 DEBUG(dbgs() << " trunced: " << *V << "\n");
1864 static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
1865 Value *V, uint64_t Offset, const Twine &Name) {
1866 IntegerType *IntTy = cast<IntegerType>(Old->getType());
1867 IntegerType *Ty = cast<IntegerType>(V->getType());
1868 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
1869 "Cannot insert a larger integer!");
1870 DEBUG(dbgs() << " start: " << *V << "\n");
1872 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
1873 DEBUG(dbgs() << " extended: " << *V << "\n");
1875 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
1876 "Element store outside of alloca store");
1877 uint64_t ShAmt = 8*Offset;
1878 if (DL.isBigEndian())
1879 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
1881 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
1882 DEBUG(dbgs() << " shifted: " << *V << "\n");
1885 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
1886 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
1887 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
1888 DEBUG(dbgs() << " masked: " << *Old << "\n");
1889 V = IRB.CreateOr(Old, V, Name + ".insert");
1890 DEBUG(dbgs() << " inserted: " << *V << "\n");
1895 static Value *extractVector(IRBuilderTy &IRB, Value *V,
1896 unsigned BeginIndex, unsigned EndIndex,
1897 const Twine &Name) {
1898 VectorType *VecTy = cast<VectorType>(V->getType());
1899 unsigned NumElements = EndIndex - BeginIndex;
1900 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
1902 if (NumElements == VecTy->getNumElements())
1905 if (NumElements == 1) {
1906 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
1908 DEBUG(dbgs() << " extract: " << *V << "\n");
1912 SmallVector<Constant*, 8> Mask;
1913 Mask.reserve(NumElements);
1914 for (unsigned i = BeginIndex; i != EndIndex; ++i)
1915 Mask.push_back(IRB.getInt32(i));
1916 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
1917 ConstantVector::get(Mask),
1919 DEBUG(dbgs() << " shuffle: " << *V << "\n");
1923 static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
1924 unsigned BeginIndex, const Twine &Name) {
1925 VectorType *VecTy = cast<VectorType>(Old->getType());
1926 assert(VecTy && "Can only insert a vector into a vector");
1928 VectorType *Ty = dyn_cast<VectorType>(V->getType());
1930 // Single element to insert.
1931 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
1933 DEBUG(dbgs() << " insert: " << *V << "\n");
1937 assert(Ty->getNumElements() <= VecTy->getNumElements() &&
1938 "Too many elements!");
1939 if (Ty->getNumElements() == VecTy->getNumElements()) {
1940 assert(V->getType() == VecTy && "Vector type mismatch");
1943 unsigned EndIndex = BeginIndex + Ty->getNumElements();
1945 // When inserting a smaller vector into the larger to store, we first
1946 // use a shuffle vector to widen it with undef elements, and then
1947 // a second shuffle vector to select between the loaded vector and the
1949 SmallVector<Constant*, 8> Mask;
1950 Mask.reserve(VecTy->getNumElements());
1951 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
1952 if (i >= BeginIndex && i < EndIndex)
1953 Mask.push_back(IRB.getInt32(i - BeginIndex));
1955 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
1956 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
1957 ConstantVector::get(Mask),
1959 DEBUG(dbgs() << " shuffle: " << *V << "\n");
1962 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
1963 Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
1965 V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
1967 DEBUG(dbgs() << " blend: " << *V << "\n");
1972 /// \brief Visitor to rewrite instructions using p particular slice of an alloca
1973 /// to use a new alloca.
1975 /// Also implements the rewriting to vector-based accesses when the partition
1976 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
1978 class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
1979 // Befriend the base class so it can delegate to private visit methods.
1980 friend class llvm::InstVisitor<AllocaSliceRewriter, bool>;
1981 typedef llvm::InstVisitor<AllocaSliceRewriter, bool> Base;
1983 const DataLayout &DL;
1986 AllocaInst &OldAI, &NewAI;
1987 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
1990 // If we are rewriting an alloca partition which can be written as pure
1991 // vector operations, we stash extra information here. When VecTy is
1992 // non-null, we have some strict guarantees about the rewritten alloca:
1993 // - The new alloca is exactly the size of the vector type here.
1994 // - The accesses all either map to the entire vector or to a single
1996 // - The set of accessing instructions is only one of those handled above
1997 // in isVectorPromotionViable. Generally these are the same access kinds
1998 // which are promotable via mem2reg.
2001 uint64_t ElementSize;
2003 // This is a convenience and flag variable that will be null unless the new
2004 // alloca's integer operations should be widened to this integer type due to
2005 // passing isIntegerWideningViable above. If it is non-null, the desired
2006 // integer type will be stored here for easy access during rewriting.
2009 // The original offset of the slice currently being rewritten relative to
2010 // the original alloca.
2011 uint64_t BeginOffset, EndOffset;
2012 // The new offsets of the slice currently being rewritten relative to the
2014 uint64_t NewBeginOffset, NewEndOffset;
2020 Instruction *OldPtr;
2022 // Track post-rewrite users which are PHI nodes and Selects.
2023 SmallPtrSetImpl<PHINode *> &PHIUsers;
2024 SmallPtrSetImpl<SelectInst *> &SelectUsers;
2026 // Utility IR builder, whose name prefix is setup for each visited use, and
2027 // the insertion point is set to point to the user.
2031 AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &S, SROA &Pass,
2032 AllocaInst &OldAI, AllocaInst &NewAI,
2033 uint64_t NewAllocaBeginOffset,
2034 uint64_t NewAllocaEndOffset, bool IsVectorPromotable,
2035 bool IsIntegerPromotable,
2036 SmallPtrSetImpl<PHINode *> &PHIUsers,
2037 SmallPtrSetImpl<SelectInst *> &SelectUsers)
2038 : DL(DL), S(S), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
2039 NewAllocaBeginOffset(NewAllocaBeginOffset),
2040 NewAllocaEndOffset(NewAllocaEndOffset),
2041 NewAllocaTy(NewAI.getAllocatedType()),
2042 VecTy(IsVectorPromotable ? cast<VectorType>(NewAllocaTy) : nullptr),
2043 ElementTy(VecTy ? VecTy->getElementType() : nullptr),
2044 ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
2045 IntTy(IsIntegerPromotable
2048 DL.getTypeSizeInBits(NewAI.getAllocatedType()))
2050 BeginOffset(), EndOffset(), IsSplittable(), IsSplit(), OldUse(),
2051 OldPtr(), PHIUsers(PHIUsers), SelectUsers(SelectUsers),
2052 IRB(NewAI.getContext(), ConstantFolder()) {
2054 assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
2055 "Only multiple-of-8 sized vector elements are viable");
2058 assert((!IsVectorPromotable && !IsIntegerPromotable) ||
2059 IsVectorPromotable != IsIntegerPromotable);
2062 bool visit(AllocaSlices::const_iterator I) {
2063 bool CanSROA = true;
2064 BeginOffset = I->beginOffset();
2065 EndOffset = I->endOffset();
2066 IsSplittable = I->isSplittable();
2068 BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
2070 // Compute the intersecting offset range.
2071 assert(BeginOffset < NewAllocaEndOffset);
2072 assert(EndOffset > NewAllocaBeginOffset);
2073 NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2074 NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2076 SliceSize = NewEndOffset - NewBeginOffset;
2078 OldUse = I->getUse();
2079 OldPtr = cast<Instruction>(OldUse->get());
2081 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2082 IRB.SetInsertPoint(OldUserI);
2083 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2084 IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
2086 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2093 // Make sure the other visit overloads are visible.
2096 // Every instruction which can end up as a user must have a rewrite rule.
2097 bool visitInstruction(Instruction &I) {
2098 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2099 llvm_unreachable("No rewrite rule for this instruction!");
2102 Value *getNewAllocaSlicePtr(IRBuilderTy &IRB, Type *PointerTy) {
2103 // Note that the offset computation can use BeginOffset or NewBeginOffset
2104 // interchangeably for unsplit slices.
2105 assert(IsSplit || BeginOffset == NewBeginOffset);
2106 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2109 StringRef OldName = OldPtr->getName();
2110 // Skip through the last '.sroa.' component of the name.
2111 size_t LastSROAPrefix = OldName.rfind(".sroa.");
2112 if (LastSROAPrefix != StringRef::npos) {
2113 OldName = OldName.substr(LastSROAPrefix + strlen(".sroa."));
2114 // Look for an SROA slice index.
2115 size_t IndexEnd = OldName.find_first_not_of("0123456789");
2116 if (IndexEnd != StringRef::npos && OldName[IndexEnd] == '.') {
2117 // Strip the index and look for the offset.
2118 OldName = OldName.substr(IndexEnd + 1);
2119 size_t OffsetEnd = OldName.find_first_not_of("0123456789");
2120 if (OffsetEnd != StringRef::npos && OldName[OffsetEnd] == '.')
2121 // Strip the offset.
2122 OldName = OldName.substr(OffsetEnd + 1);
2125 // Strip any SROA suffixes as well.
2126 OldName = OldName.substr(0, OldName.find(".sroa_"));
2129 return getAdjustedPtr(IRB, DL, &NewAI,
2130 APInt(DL.getPointerSizeInBits(), Offset), PointerTy,
2132 Twine(OldName) + "."
2139 /// \brief Compute suitable alignment to access this slice of the *new* alloca.
2141 /// You can optionally pass a type to this routine and if that type's ABI
2142 /// alignment is itself suitable, this will return zero.
2143 unsigned getSliceAlign(Type *Ty = nullptr) {
2144 unsigned NewAIAlign = NewAI.getAlignment();
2146 NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
2147 unsigned Align = MinAlign(NewAIAlign, NewBeginOffset - NewAllocaBeginOffset);
2148 return (Ty && Align == DL.getABITypeAlignment(Ty)) ? 0 : Align;
2151 unsigned getIndex(uint64_t Offset) {
2152 assert(VecTy && "Can only call getIndex when rewriting a vector");
2153 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2154 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2155 uint32_t Index = RelOffset / ElementSize;
2156 assert(Index * ElementSize == RelOffset);
2160 void deleteIfTriviallyDead(Value *V) {
2161 Instruction *I = cast<Instruction>(V);
2162 if (isInstructionTriviallyDead(I))
2163 Pass.DeadInsts.insert(I);
2166 Value *rewriteVectorizedLoadInst() {
2167 unsigned BeginIndex = getIndex(NewBeginOffset);
2168 unsigned EndIndex = getIndex(NewEndOffset);
2169 assert(EndIndex > BeginIndex && "Empty vector!");
2171 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2173 return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2176 Value *rewriteIntegerLoad(LoadInst &LI) {
2177 assert(IntTy && "We cannot insert an integer to the alloca");
2178 assert(!LI.isVolatile());
2179 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2181 V = convertValue(DL, IRB, V, IntTy);
2182 assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2183 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2184 if (Offset > 0 || NewEndOffset < NewAllocaEndOffset)
2185 V = extractInteger(DL, IRB, V, cast<IntegerType>(LI.getType()), Offset,
2190 bool visitLoadInst(LoadInst &LI) {
2191 DEBUG(dbgs() << " original: " << LI << "\n");
2192 Value *OldOp = LI.getOperand(0);
2193 assert(OldOp == OldPtr);
2195 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), SliceSize * 8)
2197 bool IsPtrAdjusted = false;
2200 V = rewriteVectorizedLoadInst();
2201 } else if (IntTy && LI.getType()->isIntegerTy()) {
2202 V = rewriteIntegerLoad(LI);
2203 } else if (NewBeginOffset == NewAllocaBeginOffset &&
2204 canConvertValue(DL, NewAllocaTy, LI.getType())) {
2205 V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2206 LI.isVolatile(), LI.getName());
2208 Type *LTy = TargetTy->getPointerTo();
2209 V = IRB.CreateAlignedLoad(getNewAllocaSlicePtr(IRB, LTy),
2210 getSliceAlign(TargetTy), LI.isVolatile(),
2212 IsPtrAdjusted = true;
2214 V = convertValue(DL, IRB, V, TargetTy);
2217 assert(!LI.isVolatile());
2218 assert(LI.getType()->isIntegerTy() &&
2219 "Only integer type loads and stores are split");
2220 assert(SliceSize < DL.getTypeStoreSize(LI.getType()) &&
2221 "Split load isn't smaller than original load");
2222 assert(LI.getType()->getIntegerBitWidth() ==
2223 DL.getTypeStoreSizeInBits(LI.getType()) &&
2224 "Non-byte-multiple bit width");
2225 // Move the insertion point just past the load so that we can refer to it.
2226 IRB.SetInsertPoint(std::next(BasicBlock::iterator(&LI)));
2227 // Create a placeholder value with the same type as LI to use as the
2228 // basis for the new value. This allows us to replace the uses of LI with
2229 // the computed value, and then replace the placeholder with LI, leaving
2230 // LI only used for this computation.
2232 = new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
2233 V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset,
2235 LI.replaceAllUsesWith(V);
2236 Placeholder->replaceAllUsesWith(&LI);
2239 LI.replaceAllUsesWith(V);
2242 Pass.DeadInsts.insert(&LI);
2243 deleteIfTriviallyDead(OldOp);
2244 DEBUG(dbgs() << " to: " << *V << "\n");
2245 return !LI.isVolatile() && !IsPtrAdjusted;
2248 bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp) {
2249 if (V->getType() != VecTy) {
2250 unsigned BeginIndex = getIndex(NewBeginOffset);
2251 unsigned EndIndex = getIndex(NewEndOffset);
2252 assert(EndIndex > BeginIndex && "Empty vector!");
2253 unsigned NumElements = EndIndex - BeginIndex;
2254 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2256 (NumElements == 1) ? ElementTy
2257 : VectorType::get(ElementTy, NumElements);
2258 if (V->getType() != SliceTy)
2259 V = convertValue(DL, IRB, V, SliceTy);
2261 // Mix in the existing elements.
2262 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2264 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2266 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2267 Pass.DeadInsts.insert(&SI);
2270 DEBUG(dbgs() << " to: " << *Store << "\n");
2274 bool rewriteIntegerStore(Value *V, StoreInst &SI) {
2275 assert(IntTy && "We cannot extract an integer from the alloca");
2276 assert(!SI.isVolatile());
2277 if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2278 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2280 Old = convertValue(DL, IRB, Old, IntTy);
2281 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2282 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2283 V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset,
2286 V = convertValue(DL, IRB, V, NewAllocaTy);
2287 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2288 Pass.DeadInsts.insert(&SI);
2290 DEBUG(dbgs() << " to: " << *Store << "\n");
2294 bool visitStoreInst(StoreInst &SI) {
2295 DEBUG(dbgs() << " original: " << SI << "\n");
2296 Value *OldOp = SI.getOperand(1);
2297 assert(OldOp == OldPtr);
2299 Value *V = SI.getValueOperand();
2301 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2302 // alloca that should be re-examined after promoting this alloca.
2303 if (V->getType()->isPointerTy())
2304 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2305 Pass.PostPromotionWorklist.insert(AI);
2307 if (SliceSize < DL.getTypeStoreSize(V->getType())) {
2308 assert(!SI.isVolatile());
2309 assert(V->getType()->isIntegerTy() &&
2310 "Only integer type loads and stores are split");
2311 assert(V->getType()->getIntegerBitWidth() ==
2312 DL.getTypeStoreSizeInBits(V->getType()) &&
2313 "Non-byte-multiple bit width");
2314 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), SliceSize * 8);
2315 V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset,
2320 return rewriteVectorizedStoreInst(V, SI, OldOp);
2321 if (IntTy && V->getType()->isIntegerTy())
2322 return rewriteIntegerStore(V, SI);
2325 if (NewBeginOffset == NewAllocaBeginOffset &&
2326 NewEndOffset == NewAllocaEndOffset &&
2327 canConvertValue(DL, V->getType(), NewAllocaTy)) {
2328 V = convertValue(DL, IRB, V, NewAllocaTy);
2329 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2332 Value *NewPtr = getNewAllocaSlicePtr(IRB, V->getType()->getPointerTo());
2333 NewSI = IRB.CreateAlignedStore(V, NewPtr, getSliceAlign(V->getType()),
2337 Pass.DeadInsts.insert(&SI);
2338 deleteIfTriviallyDead(OldOp);
2340 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2341 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2344 /// \brief Compute an integer value from splatting an i8 across the given
2345 /// number of bytes.
2347 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2348 /// call this routine.
2349 /// FIXME: Heed the advice above.
2351 /// \param V The i8 value to splat.
2352 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2353 Value *getIntegerSplat(Value *V, unsigned Size) {
2354 assert(Size > 0 && "Expected a positive number of bytes.");
2355 IntegerType *VTy = cast<IntegerType>(V->getType());
2356 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2360 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size*8);
2361 V = IRB.CreateMul(IRB.CreateZExt(V, SplatIntTy, "zext"),
2362 ConstantExpr::getUDiv(
2363 Constant::getAllOnesValue(SplatIntTy),
2364 ConstantExpr::getZExt(
2365 Constant::getAllOnesValue(V->getType()),
2371 /// \brief Compute a vector splat for a given element value.
2372 Value *getVectorSplat(Value *V, unsigned NumElements) {
2373 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2374 DEBUG(dbgs() << " splat: " << *V << "\n");
2378 bool visitMemSetInst(MemSetInst &II) {
2379 DEBUG(dbgs() << " original: " << II << "\n");
2380 assert(II.getRawDest() == OldPtr);
2382 // If the memset has a variable size, it cannot be split, just adjust the
2383 // pointer to the new alloca.
2384 if (!isa<Constant>(II.getLength())) {
2386 assert(NewBeginOffset == BeginOffset);
2387 II.setDest(getNewAllocaSlicePtr(IRB, OldPtr->getType()));
2388 Type *CstTy = II.getAlignmentCst()->getType();
2389 II.setAlignment(ConstantInt::get(CstTy, getSliceAlign()));
2391 deleteIfTriviallyDead(OldPtr);
2395 // Record this instruction for deletion.
2396 Pass.DeadInsts.insert(&II);
2398 Type *AllocaTy = NewAI.getAllocatedType();
2399 Type *ScalarTy = AllocaTy->getScalarType();
2401 // If this doesn't map cleanly onto the alloca type, and that type isn't
2402 // a single value type, just emit a memset.
2403 if (!VecTy && !IntTy &&
2404 (BeginOffset > NewAllocaBeginOffset ||
2405 EndOffset < NewAllocaEndOffset ||
2406 SliceSize != DL.getTypeStoreSize(AllocaTy) ||
2407 !AllocaTy->isSingleValueType() ||
2408 !DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy)) ||
2409 DL.getTypeSizeInBits(ScalarTy)%8 != 0)) {
2410 Type *SizeTy = II.getLength()->getType();
2411 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2412 CallInst *New = IRB.CreateMemSet(
2413 getNewAllocaSlicePtr(IRB, OldPtr->getType()), II.getValue(), Size,
2414 getSliceAlign(), II.isVolatile());
2416 DEBUG(dbgs() << " to: " << *New << "\n");
2420 // If we can represent this as a simple value, we have to build the actual
2421 // value to store, which requires expanding the byte present in memset to
2422 // a sensible representation for the alloca type. This is essentially
2423 // splatting the byte to a sufficiently wide integer, splatting it across
2424 // any desired vector width, and bitcasting to the final type.
2428 // If this is a memset of a vectorized alloca, insert it.
2429 assert(ElementTy == ScalarTy);
2431 unsigned BeginIndex = getIndex(NewBeginOffset);
2432 unsigned EndIndex = getIndex(NewEndOffset);
2433 assert(EndIndex > BeginIndex && "Empty vector!");
2434 unsigned NumElements = EndIndex - BeginIndex;
2435 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2438 getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
2439 Splat = convertValue(DL, IRB, Splat, ElementTy);
2440 if (NumElements > 1)
2441 Splat = getVectorSplat(Splat, NumElements);
2443 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2445 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2447 // If this is a memset on an alloca where we can widen stores, insert the
2449 assert(!II.isVolatile());
2451 uint64_t Size = NewEndOffset - NewBeginOffset;
2452 V = getIntegerSplat(II.getValue(), Size);
2454 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2455 EndOffset != NewAllocaBeginOffset)) {
2456 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2458 Old = convertValue(DL, IRB, Old, IntTy);
2459 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2460 V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2462 assert(V->getType() == IntTy &&
2463 "Wrong type for an alloca wide integer!");
2465 V = convertValue(DL, IRB, V, AllocaTy);
2467 // Established these invariants above.
2468 assert(NewBeginOffset == NewAllocaBeginOffset);
2469 assert(NewEndOffset == NewAllocaEndOffset);
2471 V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
2472 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2473 V = getVectorSplat(V, AllocaVecTy->getNumElements());
2475 V = convertValue(DL, IRB, V, AllocaTy);
2478 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2481 DEBUG(dbgs() << " to: " << *New << "\n");
2482 return !II.isVolatile();
2485 bool visitMemTransferInst(MemTransferInst &II) {
2486 // Rewriting of memory transfer instructions can be a bit tricky. We break
2487 // them into two categories: split intrinsics and unsplit intrinsics.
2489 DEBUG(dbgs() << " original: " << II << "\n");
2491 bool IsDest = &II.getRawDestUse() == OldUse;
2492 assert((IsDest && II.getRawDest() == OldPtr) ||
2493 (!IsDest && II.getRawSource() == OldPtr));
2495 unsigned SliceAlign = getSliceAlign();
2497 // For unsplit intrinsics, we simply modify the source and destination
2498 // pointers in place. This isn't just an optimization, it is a matter of
2499 // correctness. With unsplit intrinsics we may be dealing with transfers
2500 // within a single alloca before SROA ran, or with transfers that have
2501 // a variable length. We may also be dealing with memmove instead of
2502 // memcpy, and so simply updating the pointers is the necessary for us to
2503 // update both source and dest of a single call.
2504 if (!IsSplittable) {
2505 Value *AdjustedPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2507 II.setDest(AdjustedPtr);
2509 II.setSource(AdjustedPtr);
2511 if (II.getAlignment() > SliceAlign) {
2512 Type *CstTy = II.getAlignmentCst()->getType();
2514 ConstantInt::get(CstTy, MinAlign(II.getAlignment(), SliceAlign)));
2517 DEBUG(dbgs() << " to: " << II << "\n");
2518 deleteIfTriviallyDead(OldPtr);
2521 // For split transfer intrinsics we have an incredibly useful assurance:
2522 // the source and destination do not reside within the same alloca, and at
2523 // least one of them does not escape. This means that we can replace
2524 // memmove with memcpy, and we don't need to worry about all manner of
2525 // downsides to splitting and transforming the operations.
2527 // If this doesn't map cleanly onto the alloca type, and that type isn't
2528 // a single value type, just emit a memcpy.
2531 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2532 SliceSize != DL.getTypeStoreSize(NewAI.getAllocatedType()) ||
2533 !NewAI.getAllocatedType()->isSingleValueType());
2535 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2536 // size hasn't been shrunk based on analysis of the viable range, this is
2538 if (EmitMemCpy && &OldAI == &NewAI) {
2539 // Ensure the start lines up.
2540 assert(NewBeginOffset == BeginOffset);
2542 // Rewrite the size as needed.
2543 if (NewEndOffset != EndOffset)
2544 II.setLength(ConstantInt::get(II.getLength()->getType(),
2545 NewEndOffset - NewBeginOffset));
2548 // Record this instruction for deletion.
2549 Pass.DeadInsts.insert(&II);
2551 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2552 // alloca that should be re-examined after rewriting this instruction.
2553 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2555 = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
2556 assert(AI != &OldAI && AI != &NewAI &&
2557 "Splittable transfers cannot reach the same alloca on both ends.");
2558 Pass.Worklist.insert(AI);
2561 Type *OtherPtrTy = OtherPtr->getType();
2562 unsigned OtherAS = OtherPtrTy->getPointerAddressSpace();
2564 // Compute the relative offset for the other pointer within the transfer.
2565 unsigned IntPtrWidth = DL.getPointerSizeInBits(OtherAS);
2566 APInt OtherOffset(IntPtrWidth, NewBeginOffset - BeginOffset);
2567 unsigned OtherAlign = MinAlign(II.getAlignment() ? II.getAlignment() : 1,
2568 OtherOffset.zextOrTrunc(64).getZExtValue());
2571 // Compute the other pointer, folding as much as possible to produce
2572 // a single, simple GEP in most cases.
2573 OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2574 OtherPtr->getName() + ".");
2576 Value *OurPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2577 Type *SizeTy = II.getLength()->getType();
2578 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2580 CallInst *New = IRB.CreateMemCpy(
2581 IsDest ? OurPtr : OtherPtr, IsDest ? OtherPtr : OurPtr, Size,
2582 MinAlign(SliceAlign, OtherAlign), II.isVolatile());
2584 DEBUG(dbgs() << " to: " << *New << "\n");
2588 bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
2589 NewEndOffset == NewAllocaEndOffset;
2590 uint64_t Size = NewEndOffset - NewBeginOffset;
2591 unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
2592 unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
2593 unsigned NumElements = EndIndex - BeginIndex;
2594 IntegerType *SubIntTy
2595 = IntTy ? Type::getIntNTy(IntTy->getContext(), Size*8) : nullptr;
2597 // Reset the other pointer type to match the register type we're going to
2598 // use, but using the address space of the original other pointer.
2599 if (VecTy && !IsWholeAlloca) {
2600 if (NumElements == 1)
2601 OtherPtrTy = VecTy->getElementType();
2603 OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2605 OtherPtrTy = OtherPtrTy->getPointerTo(OtherAS);
2606 } else if (IntTy && !IsWholeAlloca) {
2607 OtherPtrTy = SubIntTy->getPointerTo(OtherAS);
2609 OtherPtrTy = NewAllocaTy->getPointerTo(OtherAS);
2612 Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2613 OtherPtr->getName() + ".");
2614 unsigned SrcAlign = OtherAlign;
2615 Value *DstPtr = &NewAI;
2616 unsigned DstAlign = SliceAlign;
2618 std::swap(SrcPtr, DstPtr);
2619 std::swap(SrcAlign, DstAlign);
2623 if (VecTy && !IsWholeAlloca && !IsDest) {
2624 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2626 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
2627 } else if (IntTy && !IsWholeAlloca && !IsDest) {
2628 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2630 Src = convertValue(DL, IRB, Src, IntTy);
2631 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2632 Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
2634 Src = IRB.CreateAlignedLoad(SrcPtr, SrcAlign, II.isVolatile(),
2638 if (VecTy && !IsWholeAlloca && IsDest) {
2639 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2641 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
2642 } else if (IntTy && !IsWholeAlloca && IsDest) {
2643 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2645 Old = convertValue(DL, IRB, Old, IntTy);
2646 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2647 Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
2648 Src = convertValue(DL, IRB, Src, NewAllocaTy);
2651 StoreInst *Store = cast<StoreInst>(
2652 IRB.CreateAlignedStore(Src, DstPtr, DstAlign, II.isVolatile()));
2654 DEBUG(dbgs() << " to: " << *Store << "\n");
2655 return !II.isVolatile();
2658 bool visitIntrinsicInst(IntrinsicInst &II) {
2659 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2660 II.getIntrinsicID() == Intrinsic::lifetime_end);
2661 DEBUG(dbgs() << " original: " << II << "\n");
2662 assert(II.getArgOperand(1) == OldPtr);
2664 // Record this instruction for deletion.
2665 Pass.DeadInsts.insert(&II);
2668 = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2669 NewEndOffset - NewBeginOffset);
2670 Value *Ptr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2672 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2673 New = IRB.CreateLifetimeStart(Ptr, Size);
2675 New = IRB.CreateLifetimeEnd(Ptr, Size);
2678 DEBUG(dbgs() << " to: " << *New << "\n");
2682 bool visitPHINode(PHINode &PN) {
2683 DEBUG(dbgs() << " original: " << PN << "\n");
2684 assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
2685 assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
2687 // We would like to compute a new pointer in only one place, but have it be
2688 // as local as possible to the PHI. To do that, we re-use the location of
2689 // the old pointer, which necessarily must be in the right position to
2690 // dominate the PHI.
2691 IRBuilderTy PtrBuilder(IRB);
2692 if (isa<PHINode>(OldPtr))
2693 PtrBuilder.SetInsertPoint(OldPtr->getParent()->getFirstInsertionPt());
2695 PtrBuilder.SetInsertPoint(OldPtr);
2696 PtrBuilder.SetCurrentDebugLocation(OldPtr->getDebugLoc());
2698 Value *NewPtr = getNewAllocaSlicePtr(PtrBuilder, OldPtr->getType());
2699 // Replace the operands which were using the old pointer.
2700 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
2702 DEBUG(dbgs() << " to: " << PN << "\n");
2703 deleteIfTriviallyDead(OldPtr);
2705 // PHIs can't be promoted on their own, but often can be speculated. We
2706 // check the speculation outside of the rewriter so that we see the
2707 // fully-rewritten alloca.
2708 PHIUsers.insert(&PN);
2712 bool visitSelectInst(SelectInst &SI) {
2713 DEBUG(dbgs() << " original: " << SI << "\n");
2714 assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
2715 "Pointer isn't an operand!");
2716 assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
2717 assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
2719 Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2720 // Replace the operands which were using the old pointer.
2721 if (SI.getOperand(1) == OldPtr)
2722 SI.setOperand(1, NewPtr);
2723 if (SI.getOperand(2) == OldPtr)
2724 SI.setOperand(2, NewPtr);
2726 DEBUG(dbgs() << " to: " << SI << "\n");
2727 deleteIfTriviallyDead(OldPtr);
2729 // Selects can't be promoted on their own, but often can be speculated. We
2730 // check the speculation outside of the rewriter so that we see the
2731 // fully-rewritten alloca.
2732 SelectUsers.insert(&SI);
2740 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
2742 /// This pass aggressively rewrites all aggregate loads and stores on
2743 /// a particular pointer (or any pointer derived from it which we can identify)
2744 /// with scalar loads and stores.
2745 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
2746 // Befriend the base class so it can delegate to private visit methods.
2747 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
2749 const DataLayout &DL;
2751 /// Queue of pointer uses to analyze and potentially rewrite.
2752 SmallVector<Use *, 8> Queue;
2754 /// Set to prevent us from cycling with phi nodes and loops.
2755 SmallPtrSet<User *, 8> Visited;
2757 /// The current pointer use being rewritten. This is used to dig up the used
2758 /// value (as opposed to the user).
2762 AggLoadStoreRewriter(const DataLayout &DL) : DL(DL) {}
2764 /// Rewrite loads and stores through a pointer and all pointers derived from
2766 bool rewrite(Instruction &I) {
2767 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
2769 bool Changed = false;
2770 while (!Queue.empty()) {
2771 U = Queue.pop_back_val();
2772 Changed |= visit(cast<Instruction>(U->getUser()));
2778 /// Enqueue all the users of the given instruction for further processing.
2779 /// This uses a set to de-duplicate users.
2780 void enqueueUsers(Instruction &I) {
2781 for (Use &U : I.uses())
2782 if (Visited.insert(U.getUser()))
2783 Queue.push_back(&U);
2786 // Conservative default is to not rewrite anything.
2787 bool visitInstruction(Instruction &I) { return false; }
2789 /// \brief Generic recursive split emission class.
2790 template <typename Derived>
2793 /// The builder used to form new instructions.
2795 /// The indices which to be used with insert- or extractvalue to select the
2796 /// appropriate value within the aggregate.
2797 SmallVector<unsigned, 4> Indices;
2798 /// The indices to a GEP instruction which will move Ptr to the correct slot
2799 /// within the aggregate.
2800 SmallVector<Value *, 4> GEPIndices;
2801 /// The base pointer of the original op, used as a base for GEPing the
2802 /// split operations.
2805 /// Initialize the splitter with an insertion point, Ptr and start with a
2806 /// single zero GEP index.
2807 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
2808 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
2811 /// \brief Generic recursive split emission routine.
2813 /// This method recursively splits an aggregate op (load or store) into
2814 /// scalar or vector ops. It splits recursively until it hits a single value
2815 /// and emits that single value operation via the template argument.
2817 /// The logic of this routine relies on GEPs and insertvalue and
2818 /// extractvalue all operating with the same fundamental index list, merely
2819 /// formatted differently (GEPs need actual values).
2821 /// \param Ty The type being split recursively into smaller ops.
2822 /// \param Agg The aggregate value being built up or stored, depending on
2823 /// whether this is splitting a load or a store respectively.
2824 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
2825 if (Ty->isSingleValueType())
2826 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
2828 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2829 unsigned OldSize = Indices.size();
2831 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
2833 assert(Indices.size() == OldSize && "Did not return to the old size");
2834 Indices.push_back(Idx);
2835 GEPIndices.push_back(IRB.getInt32(Idx));
2836 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
2837 GEPIndices.pop_back();
2843 if (StructType *STy = dyn_cast<StructType>(Ty)) {
2844 unsigned OldSize = Indices.size();
2846 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
2848 assert(Indices.size() == OldSize && "Did not return to the old size");
2849 Indices.push_back(Idx);
2850 GEPIndices.push_back(IRB.getInt32(Idx));
2851 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
2852 GEPIndices.pop_back();
2858 llvm_unreachable("Only arrays and structs are aggregate loadable types");
2862 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
2863 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2864 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
2866 /// Emit a leaf load of a single value. This is called at the leaves of the
2867 /// recursive emission to actually load values.
2868 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2869 assert(Ty->isSingleValueType());
2870 // Load the single value and insert it using the indices.
2871 Value *GEP = IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep");
2872 Value *Load = IRB.CreateLoad(GEP, Name + ".load");
2873 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
2874 DEBUG(dbgs() << " to: " << *Load << "\n");
2878 bool visitLoadInst(LoadInst &LI) {
2879 assert(LI.getPointerOperand() == *U);
2880 if (!LI.isSimple() || LI.getType()->isSingleValueType())
2883 // We have an aggregate being loaded, split it apart.
2884 DEBUG(dbgs() << " original: " << LI << "\n");
2885 LoadOpSplitter Splitter(&LI, *U);
2886 Value *V = UndefValue::get(LI.getType());
2887 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
2888 LI.replaceAllUsesWith(V);
2889 LI.eraseFromParent();
2893 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
2894 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2895 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
2897 /// Emit a leaf store of a single value. This is called at the leaves of the
2898 /// recursive emission to actually produce stores.
2899 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2900 assert(Ty->isSingleValueType());
2901 // Extract the single value and store it using the indices.
2902 Value *Store = IRB.CreateStore(
2903 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
2904 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
2906 DEBUG(dbgs() << " to: " << *Store << "\n");
2910 bool visitStoreInst(StoreInst &SI) {
2911 if (!SI.isSimple() || SI.getPointerOperand() != *U)
2913 Value *V = SI.getValueOperand();
2914 if (V->getType()->isSingleValueType())
2917 // We have an aggregate being stored, split it apart.
2918 DEBUG(dbgs() << " original: " << SI << "\n");
2919 StoreOpSplitter Splitter(&SI, *U);
2920 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
2921 SI.eraseFromParent();
2925 bool visitBitCastInst(BitCastInst &BC) {
2930 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
2935 bool visitPHINode(PHINode &PN) {
2940 bool visitSelectInst(SelectInst &SI) {
2947 /// \brief Strip aggregate type wrapping.
2949 /// This removes no-op aggregate types wrapping an underlying type. It will
2950 /// strip as many layers of types as it can without changing either the type
2951 /// size or the allocated size.
2952 static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
2953 if (Ty->isSingleValueType())
2956 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
2957 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
2960 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
2961 InnerTy = ArrTy->getElementType();
2962 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
2963 const StructLayout *SL = DL.getStructLayout(STy);
2964 unsigned Index = SL->getElementContainingOffset(0);
2965 InnerTy = STy->getElementType(Index);
2970 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
2971 TypeSize > DL.getTypeSizeInBits(InnerTy))
2974 return stripAggregateTypeWrapping(DL, InnerTy);
2977 /// \brief Try to find a partition of the aggregate type passed in for a given
2978 /// offset and size.
2980 /// This recurses through the aggregate type and tries to compute a subtype
2981 /// based on the offset and size. When the offset and size span a sub-section
2982 /// of an array, it will even compute a new array type for that sub-section,
2983 /// and the same for structs.
2985 /// Note that this routine is very strict and tries to find a partition of the
2986 /// type which produces the *exact* right offset and size. It is not forgiving
2987 /// when the size or offset cause either end of type-based partition to be off.
2988 /// Also, this is a best-effort routine. It is reasonable to give up and not
2989 /// return a type if necessary.
2990 static Type *getTypePartition(const DataLayout &DL, Type *Ty,
2991 uint64_t Offset, uint64_t Size) {
2992 if (Offset == 0 && DL.getTypeAllocSize(Ty) == Size)
2993 return stripAggregateTypeWrapping(DL, Ty);
2994 if (Offset > DL.getTypeAllocSize(Ty) ||
2995 (DL.getTypeAllocSize(Ty) - Offset) < Size)
2998 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
2999 // We can't partition pointers...
3000 if (SeqTy->isPointerTy())
3003 Type *ElementTy = SeqTy->getElementType();
3004 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3005 uint64_t NumSkippedElements = Offset / ElementSize;
3006 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy)) {
3007 if (NumSkippedElements >= ArrTy->getNumElements())
3009 } else if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy)) {
3010 if (NumSkippedElements >= VecTy->getNumElements())
3013 Offset -= NumSkippedElements * ElementSize;
3015 // First check if we need to recurse.
3016 if (Offset > 0 || Size < ElementSize) {
3017 // Bail if the partition ends in a different array element.
3018 if ((Offset + Size) > ElementSize)
3020 // Recurse through the element type trying to peel off offset bytes.
3021 return getTypePartition(DL, ElementTy, Offset, Size);
3023 assert(Offset == 0);
3025 if (Size == ElementSize)
3026 return stripAggregateTypeWrapping(DL, ElementTy);
3027 assert(Size > ElementSize);
3028 uint64_t NumElements = Size / ElementSize;
3029 if (NumElements * ElementSize != Size)
3031 return ArrayType::get(ElementTy, NumElements);
3034 StructType *STy = dyn_cast<StructType>(Ty);
3038 const StructLayout *SL = DL.getStructLayout(STy);
3039 if (Offset >= SL->getSizeInBytes())
3041 uint64_t EndOffset = Offset + Size;
3042 if (EndOffset > SL->getSizeInBytes())
3045 unsigned Index = SL->getElementContainingOffset(Offset);
3046 Offset -= SL->getElementOffset(Index);
3048 Type *ElementTy = STy->getElementType(Index);
3049 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3050 if (Offset >= ElementSize)
3051 return nullptr; // The offset points into alignment padding.
3053 // See if any partition must be contained by the element.
3054 if (Offset > 0 || Size < ElementSize) {
3055 if ((Offset + Size) > ElementSize)
3057 return getTypePartition(DL, ElementTy, Offset, Size);
3059 assert(Offset == 0);
3061 if (Size == ElementSize)
3062 return stripAggregateTypeWrapping(DL, ElementTy);
3064 StructType::element_iterator EI = STy->element_begin() + Index,
3065 EE = STy->element_end();
3066 if (EndOffset < SL->getSizeInBytes()) {
3067 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3068 if (Index == EndIndex)
3069 return nullptr; // Within a single element and its padding.
3071 // Don't try to form "natural" types if the elements don't line up with the
3073 // FIXME: We could potentially recurse down through the last element in the
3074 // sub-struct to find a natural end point.
3075 if (SL->getElementOffset(EndIndex) != EndOffset)
3078 assert(Index < EndIndex);
3079 EE = STy->element_begin() + EndIndex;
3082 // Try to build up a sub-structure.
3083 StructType *SubTy = StructType::get(STy->getContext(), makeArrayRef(EI, EE),
3085 const StructLayout *SubSL = DL.getStructLayout(SubTy);
3086 if (Size != SubSL->getSizeInBytes())
3087 return nullptr; // The sub-struct doesn't have quite the size needed.
3092 /// \brief Rewrite an alloca partition's users.
3094 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3095 /// to rewrite uses of an alloca partition to be conducive for SSA value
3096 /// promotion. If the partition needs a new, more refined alloca, this will
3097 /// build that new alloca, preserving as much type information as possible, and
3098 /// rewrite the uses of the old alloca to point at the new one and have the
3099 /// appropriate new offsets. It also evaluates how successful the rewrite was
3100 /// at enabling promotion and if it was successful queues the alloca to be
3102 bool SROA::rewritePartition(AllocaInst &AI, AllocaSlices &S,
3103 AllocaSlices::iterator B, AllocaSlices::iterator E,
3104 int64_t BeginOffset, int64_t EndOffset,
3105 ArrayRef<AllocaSlices::iterator> SplitUses) {
3106 assert(BeginOffset < EndOffset);
3107 uint64_t SliceSize = EndOffset - BeginOffset;
3109 // Try to compute a friendly type for this partition of the alloca. This
3110 // won't always succeed, in which case we fall back to a legal integer type
3111 // or an i8 array of an appropriate size.
3112 Type *SliceTy = nullptr;
3113 if (Type *CommonUseTy = findCommonType(B, E, EndOffset))
3114 if (DL->getTypeAllocSize(CommonUseTy) >= SliceSize)
3115 SliceTy = CommonUseTy;
3117 if (Type *TypePartitionTy = getTypePartition(*DL, AI.getAllocatedType(),
3118 BeginOffset, SliceSize))
3119 SliceTy = TypePartitionTy;
3120 if ((!SliceTy || (SliceTy->isArrayTy() &&
3121 SliceTy->getArrayElementType()->isIntegerTy())) &&
3122 DL->isLegalInteger(SliceSize * 8))
3123 SliceTy = Type::getIntNTy(*C, SliceSize * 8);
3125 SliceTy = ArrayType::get(Type::getInt8Ty(*C), SliceSize);
3126 assert(DL->getTypeAllocSize(SliceTy) >= SliceSize);
3128 bool IsVectorPromotable =
3129 isVectorPromotionViable(*DL, SliceTy, BeginOffset, EndOffset,
3130 AllocaSlices::const_range(B, E), SplitUses);
3132 bool IsIntegerPromotable =
3133 !IsVectorPromotable &&
3134 isIntegerWideningViable(*DL, SliceTy, BeginOffset,
3135 AllocaSlices::const_range(B, E), SplitUses);
3137 // Check for the case where we're going to rewrite to a new alloca of the
3138 // exact same type as the original, and with the same access offsets. In that
3139 // case, re-use the existing alloca, but still run through the rewriter to
3140 // perform phi and select speculation.
3142 if (SliceTy == AI.getAllocatedType()) {
3143 assert(BeginOffset == 0 &&
3144 "Non-zero begin offset but same alloca type");
3146 // FIXME: We should be able to bail at this point with "nothing changed".
3147 // FIXME: We might want to defer PHI speculation until after here.
3149 unsigned Alignment = AI.getAlignment();
3151 // The minimum alignment which users can rely on when the explicit
3152 // alignment is omitted or zero is that required by the ABI for this
3154 Alignment = DL->getABITypeAlignment(AI.getAllocatedType());
3156 Alignment = MinAlign(Alignment, BeginOffset);
3157 // If we will get at least this much alignment from the type alone, leave
3158 // the alloca's alignment unconstrained.
3159 if (Alignment <= DL->getABITypeAlignment(SliceTy))
3161 NewAI = new AllocaInst(SliceTy, nullptr, Alignment,
3162 AI.getName() + ".sroa." + Twine(B - S.begin()), &AI);
3166 DEBUG(dbgs() << "Rewriting alloca partition "
3167 << "[" << BeginOffset << "," << EndOffset << ") to: " << *NewAI
3170 // Track the high watermark on the worklist as it is only relevant for
3171 // promoted allocas. We will reset it to this point if the alloca is not in
3172 // fact scheduled for promotion.
3173 unsigned PPWOldSize = PostPromotionWorklist.size();
3174 unsigned NumUses = 0;
3175 SmallPtrSet<PHINode *, 8> PHIUsers;
3176 SmallPtrSet<SelectInst *, 8> SelectUsers;
3178 AllocaSliceRewriter Rewriter(*DL, S, *this, AI, *NewAI, BeginOffset,
3179 EndOffset, IsVectorPromotable,
3180 IsIntegerPromotable, PHIUsers, SelectUsers);
3181 bool Promotable = true;
3182 for (auto & SplitUse : SplitUses) {
3183 DEBUG(dbgs() << " rewriting split ");
3184 DEBUG(S.printSlice(dbgs(), SplitUse, ""));
3185 Promotable &= Rewriter.visit(SplitUse);
3188 for (AllocaSlices::iterator I = B; I != E; ++I) {
3189 DEBUG(dbgs() << " rewriting ");
3190 DEBUG(S.printSlice(dbgs(), I, ""));
3191 Promotable &= Rewriter.visit(I);
3195 NumAllocaPartitionUses += NumUses;
3196 MaxUsesPerAllocaPartition =
3197 std::max<unsigned>(NumUses, MaxUsesPerAllocaPartition);
3199 // Now that we've processed all the slices in the new partition, check if any
3200 // PHIs or Selects would block promotion.
3201 for (SmallPtrSetImpl<PHINode *>::iterator I = PHIUsers.begin(),
3204 if (!isSafePHIToSpeculate(**I, DL)) {
3207 SelectUsers.clear();
3210 for (SmallPtrSetImpl<SelectInst *>::iterator I = SelectUsers.begin(),
3211 E = SelectUsers.end();
3213 if (!isSafeSelectToSpeculate(**I, DL)) {
3216 SelectUsers.clear();
3221 if (PHIUsers.empty() && SelectUsers.empty()) {
3222 // Promote the alloca.
3223 PromotableAllocas.push_back(NewAI);
3225 // If we have either PHIs or Selects to speculate, add them to those
3226 // worklists and re-queue the new alloca so that we promote in on the
3228 for (PHINode *PHIUser : PHIUsers)
3229 SpeculatablePHIs.insert(PHIUser);
3230 for (SelectInst *SelectUser : SelectUsers)
3231 SpeculatableSelects.insert(SelectUser);
3232 Worklist.insert(NewAI);
3235 // If we can't promote the alloca, iterate on it to check for new
3236 // refinements exposed by splitting the current alloca. Don't iterate on an
3237 // alloca which didn't actually change and didn't get promoted.
3239 Worklist.insert(NewAI);
3241 // Drop any post-promotion work items if promotion didn't happen.
3242 while (PostPromotionWorklist.size() > PPWOldSize)
3243 PostPromotionWorklist.pop_back();
3250 removeFinishedSplitUses(SmallVectorImpl<AllocaSlices::iterator> &SplitUses,
3251 uint64_t &MaxSplitUseEndOffset, uint64_t Offset) {
3252 if (Offset >= MaxSplitUseEndOffset) {
3254 MaxSplitUseEndOffset = 0;
3258 size_t SplitUsesOldSize = SplitUses.size();
3259 SplitUses.erase(std::remove_if(SplitUses.begin(), SplitUses.end(),
3260 [Offset](const AllocaSlices::iterator &I) {
3261 return I->endOffset() <= Offset;
3264 if (SplitUsesOldSize == SplitUses.size())
3267 // Recompute the max. While this is linear, so is remove_if.
3268 MaxSplitUseEndOffset = 0;
3269 for (AllocaSlices::iterator SplitUse : SplitUses)
3270 MaxSplitUseEndOffset =
3271 std::max(SplitUse->endOffset(), MaxSplitUseEndOffset);
3274 /// \brief Walks the slices of an alloca and form partitions based on them,
3275 /// rewriting each of their uses.
3276 bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &S) {
3277 if (S.begin() == S.end())
3280 unsigned NumPartitions = 0;
3281 bool Changed = false;
3282 SmallVector<AllocaSlices::iterator, 4> SplitUses;
3283 uint64_t MaxSplitUseEndOffset = 0;
3285 uint64_t BeginOffset = S.begin()->beginOffset();
3287 for (AllocaSlices::iterator SI = S.begin(), SJ = std::next(SI), SE = S.end();
3288 SI != SE; SI = SJ) {
3289 uint64_t MaxEndOffset = SI->endOffset();
3291 if (!SI->isSplittable()) {
3292 // When we're forming an unsplittable region, it must always start at the
3293 // first slice and will extend through its end.
3294 assert(BeginOffset == SI->beginOffset());
3296 // Form a partition including all of the overlapping slices with this
3297 // unsplittable slice.
3298 while (SJ != SE && SJ->beginOffset() < MaxEndOffset) {
3299 if (!SJ->isSplittable())
3300 MaxEndOffset = std::max(MaxEndOffset, SJ->endOffset());
3304 assert(SI->isSplittable()); // Established above.
3306 // Collect all of the overlapping splittable slices.
3307 while (SJ != SE && SJ->beginOffset() < MaxEndOffset &&
3308 SJ->isSplittable()) {
3309 MaxEndOffset = std::max(MaxEndOffset, SJ->endOffset());
3313 // Back up MaxEndOffset and SJ if we ended the span early when
3314 // encountering an unsplittable slice.
3315 if (SJ != SE && SJ->beginOffset() < MaxEndOffset) {
3316 assert(!SJ->isSplittable());
3317 MaxEndOffset = SJ->beginOffset();
3321 // Check if we have managed to move the end offset forward yet. If so,
3322 // we'll have to rewrite uses and erase old split uses.
3323 if (BeginOffset < MaxEndOffset) {
3324 // Rewrite a sequence of overlapping slices.
3326 rewritePartition(AI, S, SI, SJ, BeginOffset, MaxEndOffset, SplitUses);
3329 removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset, MaxEndOffset);
3332 // Accumulate all the splittable slices from the [SI,SJ) region which
3333 // overlap going forward.
3334 for (AllocaSlices::iterator SK = SI; SK != SJ; ++SK)
3335 if (SK->isSplittable() && SK->endOffset() > MaxEndOffset) {
3336 SplitUses.push_back(SK);
3337 MaxSplitUseEndOffset = std::max(SK->endOffset(), MaxSplitUseEndOffset);
3340 // If we're already at the end and we have no split uses, we're done.
3341 if (SJ == SE && SplitUses.empty())
3344 // If we have no split uses or no gap in offsets, we're ready to move to
3346 if (SplitUses.empty() || (SJ != SE && MaxEndOffset == SJ->beginOffset())) {
3347 BeginOffset = SJ->beginOffset();
3351 // Even if we have split slices, if the next slice is splittable and the
3352 // split slices reach it, we can simply set up the beginning offset of the
3353 // next iteration to bridge between them.
3354 if (SJ != SE && SJ->isSplittable() &&
3355 MaxSplitUseEndOffset > SJ->beginOffset()) {
3356 BeginOffset = MaxEndOffset;
3360 // Otherwise, we have a tail of split slices. Rewrite them with an empty
3362 uint64_t PostSplitEndOffset =
3363 SJ == SE ? MaxSplitUseEndOffset : SJ->beginOffset();
3365 Changed |= rewritePartition(AI, S, SJ, SJ, MaxEndOffset, PostSplitEndOffset,
3370 break; // Skip the rest, we don't need to do any cleanup.
3372 removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset,
3373 PostSplitEndOffset);
3375 // Now just reset the begin offset for the next iteration.
3376 BeginOffset = SJ->beginOffset();
3379 NumAllocaPartitions += NumPartitions;
3380 MaxPartitionsPerAlloca =
3381 std::max<unsigned>(NumPartitions, MaxPartitionsPerAlloca);
3386 /// \brief Clobber a use with undef, deleting the used value if it becomes dead.
3387 void SROA::clobberUse(Use &U) {
3389 // Replace the use with an undef value.
3390 U = UndefValue::get(OldV->getType());
3392 // Check for this making an instruction dead. We have to garbage collect
3393 // all the dead instructions to ensure the uses of any alloca end up being
3395 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3396 if (isInstructionTriviallyDead(OldI)) {
3397 DeadInsts.insert(OldI);
3401 /// \brief Analyze an alloca for SROA.
3403 /// This analyzes the alloca to ensure we can reason about it, builds
3404 /// the slices of the alloca, and then hands it off to be split and
3405 /// rewritten as needed.
3406 bool SROA::runOnAlloca(AllocaInst &AI) {
3407 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3408 ++NumAllocasAnalyzed;
3410 // Special case dead allocas, as they're trivial.
3411 if (AI.use_empty()) {
3412 AI.eraseFromParent();
3416 // Skip alloca forms that this analysis can't handle.
3417 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3418 DL->getTypeAllocSize(AI.getAllocatedType()) == 0)
3421 bool Changed = false;
3423 // First, split any FCA loads and stores touching this alloca to promote
3424 // better splitting and promotion opportunities.
3425 AggLoadStoreRewriter AggRewriter(*DL);
3426 Changed |= AggRewriter.rewrite(AI);
3428 // Build the slices using a recursive instruction-visiting builder.
3429 AllocaSlices S(*DL, AI);
3430 DEBUG(S.print(dbgs()));
3434 // Delete all the dead users of this alloca before splitting and rewriting it.
3435 for (Instruction *DeadUser : S.getDeadUsers()) {
3436 // Free up everything used by this instruction.
3437 for (Use &DeadOp : DeadUser->operands())
3440 // Now replace the uses of this instruction.
3441 DeadUser->replaceAllUsesWith(UndefValue::get(DeadUser->getType()));
3443 // And mark it for deletion.
3444 DeadInsts.insert(DeadUser);
3447 for (Use *DeadOp : S.getDeadOperands()) {
3448 clobberUse(*DeadOp);
3452 // No slices to split. Leave the dead alloca for a later pass to clean up.
3453 if (S.begin() == S.end())
3456 Changed |= splitAlloca(AI, S);
3458 DEBUG(dbgs() << " Speculating PHIs\n");
3459 while (!SpeculatablePHIs.empty())
3460 speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
3462 DEBUG(dbgs() << " Speculating Selects\n");
3463 while (!SpeculatableSelects.empty())
3464 speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
3469 /// \brief Delete the dead instructions accumulated in this run.
3471 /// Recursively deletes the dead instructions we've accumulated. This is done
3472 /// at the very end to maximize locality of the recursive delete and to
3473 /// minimize the problems of invalidated instruction pointers as such pointers
3474 /// are used heavily in the intermediate stages of the algorithm.
3476 /// We also record the alloca instructions deleted here so that they aren't
3477 /// subsequently handed to mem2reg to promote.
3478 void SROA::deleteDeadInstructions(SmallPtrSetImpl<AllocaInst*> &DeletedAllocas) {
3479 while (!DeadInsts.empty()) {
3480 Instruction *I = DeadInsts.pop_back_val();
3481 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3483 I->replaceAllUsesWith(UndefValue::get(I->getType()));
3485 for (Use &Operand : I->operands())
3486 if (Instruction *U = dyn_cast<Instruction>(Operand)) {
3487 // Zero out the operand and see if it becomes trivially dead.
3489 if (isInstructionTriviallyDead(U))
3490 DeadInsts.insert(U);
3493 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3494 DeletedAllocas.insert(AI);
3497 I->eraseFromParent();
3501 static void enqueueUsersInWorklist(Instruction &I,
3502 SmallVectorImpl<Instruction *> &Worklist,
3503 SmallPtrSetImpl<Instruction *> &Visited) {
3504 for (User *U : I.users())
3505 if (Visited.insert(cast<Instruction>(U)))
3506 Worklist.push_back(cast<Instruction>(U));
3509 /// \brief Promote the allocas, using the best available technique.
3511 /// This attempts to promote whatever allocas have been identified as viable in
3512 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
3513 /// If there is a domtree available, we attempt to promote using the full power
3514 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3515 /// based on the SSAUpdater utilities. This function returns whether any
3516 /// promotion occurred.
3517 bool SROA::promoteAllocas(Function &F) {
3518 if (PromotableAllocas.empty())
3521 NumPromoted += PromotableAllocas.size();
3523 if (DT && !ForceSSAUpdater) {
3524 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3525 PromoteMemToReg(PromotableAllocas, *DT, nullptr, AT);
3526 PromotableAllocas.clear();
3530 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3532 DIBuilder DIB(*F.getParent());
3533 SmallVector<Instruction *, 64> Insts;
3535 // We need a worklist to walk the uses of each alloca.
3536 SmallVector<Instruction *, 8> Worklist;
3537 SmallPtrSet<Instruction *, 8> Visited;
3538 SmallVector<Instruction *, 32> DeadInsts;
3540 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3541 AllocaInst *AI = PromotableAllocas[Idx];
3546 enqueueUsersInWorklist(*AI, Worklist, Visited);
3548 while (!Worklist.empty()) {
3549 Instruction *I = Worklist.pop_back_val();
3551 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3552 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3553 // leading to them) here. Eventually it should use them to optimize the
3554 // scalar values produced.
3555 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3556 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3557 II->getIntrinsicID() == Intrinsic::lifetime_end);
3558 II->eraseFromParent();
3562 // Push the loads and stores we find onto the list. SROA will already
3563 // have validated that all loads and stores are viable candidates for
3565 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
3566 assert(LI->getType() == AI->getAllocatedType());
3567 Insts.push_back(LI);
3570 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
3571 assert(SI->getValueOperand()->getType() == AI->getAllocatedType());
3572 Insts.push_back(SI);
3576 // For everything else, we know that only no-op bitcasts and GEPs will
3577 // make it this far, just recurse through them and recall them for later
3579 DeadInsts.push_back(I);
3580 enqueueUsersInWorklist(*I, Worklist, Visited);
3582 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3583 while (!DeadInsts.empty())
3584 DeadInsts.pop_back_val()->eraseFromParent();
3585 AI->eraseFromParent();
3588 PromotableAllocas.clear();
3592 bool SROA::runOnFunction(Function &F) {
3593 if (skipOptnoneFunction(F))
3596 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3597 C = &F.getContext();
3598 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
3600 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3603 DL = &DLP->getDataLayout();
3604 DominatorTreeWrapperPass *DTWP =
3605 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
3606 DT = DTWP ? &DTWP->getDomTree() : nullptr;
3607 AT = &getAnalysis<AssumptionTracker>();
3609 BasicBlock &EntryBB = F.getEntryBlock();
3610 for (BasicBlock::iterator I = EntryBB.begin(), E = std::prev(EntryBB.end());
3612 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3613 Worklist.insert(AI);
3615 bool Changed = false;
3616 // A set of deleted alloca instruction pointers which should be removed from
3617 // the list of promotable allocas.
3618 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3621 while (!Worklist.empty()) {
3622 Changed |= runOnAlloca(*Worklist.pop_back_val());
3623 deleteDeadInstructions(DeletedAllocas);
3625 // Remove the deleted allocas from various lists so that we don't try to
3626 // continue processing them.
3627 if (!DeletedAllocas.empty()) {
3628 auto IsInSet = [&](AllocaInst *AI) {
3629 return DeletedAllocas.count(AI);
3631 Worklist.remove_if(IsInSet);
3632 PostPromotionWorklist.remove_if(IsInSet);
3633 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3634 PromotableAllocas.end(),
3636 PromotableAllocas.end());
3637 DeletedAllocas.clear();
3641 Changed |= promoteAllocas(F);
3643 Worklist = PostPromotionWorklist;
3644 PostPromotionWorklist.clear();
3645 } while (!Worklist.empty());
3650 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3651 AU.addRequired<AssumptionTracker>();
3652 if (RequiresDomTree)
3653 AU.addRequired<DominatorTreeWrapperPass>();
3654 AU.setPreservesCFG();