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/Loads.h"
32 #include "llvm/Analysis/PtrUseVisitor.h"
33 #include "llvm/Analysis/ValueTracking.h"
34 #include "llvm/IR/Constants.h"
35 #include "llvm/IR/DIBuilder.h"
36 #include "llvm/IR/DataLayout.h"
37 #include "llvm/IR/DebugInfo.h"
38 #include "llvm/IR/DerivedTypes.h"
39 #include "llvm/IR/Dominators.h"
40 #include "llvm/IR/Function.h"
41 #include "llvm/IR/IRBuilder.h"
42 #include "llvm/IR/InstVisitor.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/IntrinsicInst.h"
45 #include "llvm/IR/LLVMContext.h"
46 #include "llvm/IR/Operator.h"
47 #include "llvm/Pass.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/Compiler.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Support/TimeValue.h"
54 #include "llvm/Support/raw_ostream.h"
55 #include "llvm/Transforms/Utils/Local.h"
56 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
57 #include "llvm/Transforms/Utils/SSAUpdater.h"
59 #if __cplusplus >= 201103L && !defined(NDEBUG)
60 // We only use this for a debug check in C++11
66 #define DEBUG_TYPE "sroa"
68 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
69 STATISTIC(NumAllocaPartitions, "Number of alloca partitions formed");
70 STATISTIC(MaxPartitionsPerAlloca, "Maximum number of partitions per alloca");
71 STATISTIC(NumAllocaPartitionUses, "Number of alloca partition uses rewritten");
72 STATISTIC(MaxUsesPerAllocaPartition, "Maximum number of uses of a partition");
73 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
74 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
75 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
76 STATISTIC(NumDeleted, "Number of instructions deleted");
77 STATISTIC(NumVectorized, "Number of vectorized aggregates");
79 /// Hidden option to force the pass to not use DomTree and mem2reg, instead
80 /// forming SSA values through the SSAUpdater infrastructure.
82 ForceSSAUpdater("force-ssa-updater", cl::init(false), cl::Hidden);
84 /// Hidden option to enable randomly shuffling the slices to help uncover
85 /// instability in their order.
86 static cl::opt<bool> SROARandomShuffleSlices("sroa-random-shuffle-slices",
87 cl::init(false), cl::Hidden);
89 /// Hidden option to experiment with completely strict handling of inbounds
91 static cl::opt<bool> SROAStrictInbounds("sroa-strict-inbounds",
92 cl::init(false), cl::Hidden);
95 /// \brief A custom IRBuilder inserter which prefixes all names if they are
97 template <bool preserveNames = true>
98 class IRBuilderPrefixedInserter :
99 public IRBuilderDefaultInserter<preserveNames> {
103 void SetNamePrefix(const Twine &P) { Prefix = P.str(); }
106 void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
107 BasicBlock::iterator InsertPt) const {
108 IRBuilderDefaultInserter<preserveNames>::InsertHelper(
109 I, Name.isTriviallyEmpty() ? Name : Prefix + Name, BB, InsertPt);
113 // Specialization for not preserving the name is trivial.
115 class IRBuilderPrefixedInserter<false> :
116 public IRBuilderDefaultInserter<false> {
118 void SetNamePrefix(const Twine &P) {}
121 /// \brief Provide a typedef for IRBuilder that drops names in release builds.
123 typedef llvm::IRBuilder<true, ConstantFolder,
124 IRBuilderPrefixedInserter<true> > IRBuilderTy;
126 typedef llvm::IRBuilder<false, ConstantFolder,
127 IRBuilderPrefixedInserter<false> > IRBuilderTy;
132 /// \brief A used slice of an alloca.
134 /// This structure represents a slice of an alloca used by some instruction. It
135 /// stores both the begin and end offsets of this use, a pointer to the use
136 /// itself, and a flag indicating whether we can classify the use as splittable
137 /// or not when forming partitions of the alloca.
139 /// \brief The beginning offset of the range.
140 uint64_t BeginOffset;
142 /// \brief The ending offset, not included in the range.
145 /// \brief Storage for both the use of this slice and whether it can be
147 PointerIntPair<Use *, 1, bool> UseAndIsSplittable;
150 Slice() : BeginOffset(), EndOffset() {}
151 Slice(uint64_t BeginOffset, uint64_t EndOffset, Use *U, bool IsSplittable)
152 : BeginOffset(BeginOffset), EndOffset(EndOffset),
153 UseAndIsSplittable(U, IsSplittable) {}
155 uint64_t beginOffset() const { return BeginOffset; }
156 uint64_t endOffset() const { return EndOffset; }
158 bool isSplittable() const { return UseAndIsSplittable.getInt(); }
159 void makeUnsplittable() { UseAndIsSplittable.setInt(false); }
161 Use *getUse() const { return UseAndIsSplittable.getPointer(); }
163 bool isDead() const { return getUse() == nullptr; }
164 void kill() { UseAndIsSplittable.setPointer(nullptr); }
166 /// \brief Support for ordering ranges.
168 /// This provides an ordering over ranges such that start offsets are
169 /// always increasing, and within equal start offsets, the end offsets are
170 /// decreasing. Thus the spanning range comes first in a cluster with the
171 /// same start position.
172 bool operator<(const Slice &RHS) const {
173 if (beginOffset() < RHS.beginOffset()) return true;
174 if (beginOffset() > RHS.beginOffset()) return false;
175 if (isSplittable() != RHS.isSplittable()) return !isSplittable();
176 if (endOffset() > RHS.endOffset()) return true;
180 /// \brief Support comparison with a single offset to allow binary searches.
181 friend LLVM_ATTRIBUTE_UNUSED bool operator<(const Slice &LHS,
182 uint64_t RHSOffset) {
183 return LHS.beginOffset() < RHSOffset;
185 friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
187 return LHSOffset < RHS.beginOffset();
190 bool operator==(const Slice &RHS) const {
191 return isSplittable() == RHS.isSplittable() &&
192 beginOffset() == RHS.beginOffset() && endOffset() == RHS.endOffset();
194 bool operator!=(const Slice &RHS) const { return !operator==(RHS); }
196 } // end anonymous namespace
199 template <typename T> struct isPodLike;
200 template <> struct isPodLike<Slice> {
201 static const bool value = true;
206 /// \brief Representation of the alloca slices.
208 /// This class represents the slices of an alloca which are formed by its
209 /// various uses. If a pointer escapes, we can't fully build a representation
210 /// for the slices used and we reflect that in this structure. The uses are
211 /// stored, sorted by increasing beginning offset and with unsplittable slices
212 /// starting at a particular offset before splittable slices.
215 /// \brief Construct the slices of a particular alloca.
216 AllocaSlices(const DataLayout &DL, AllocaInst &AI);
218 /// \brief Test whether a pointer to the allocation escapes our analysis.
220 /// If this is true, the slices are never fully built and should be
222 bool isEscaped() const { return PointerEscapingInstr; }
224 /// \brief Support for iterating over the slices.
226 typedef SmallVectorImpl<Slice>::iterator iterator;
227 iterator begin() { return Slices.begin(); }
228 iterator end() { return Slices.end(); }
230 typedef SmallVectorImpl<Slice>::const_iterator const_iterator;
231 const_iterator begin() const { return Slices.begin(); }
232 const_iterator end() const { return Slices.end(); }
235 /// \brief Allow iterating the dead users for this alloca.
237 /// These are instructions which will never actually use the alloca as they
238 /// are outside the allocated range. They are safe to replace with undef and
241 typedef SmallVectorImpl<Instruction *>::const_iterator dead_user_iterator;
242 dead_user_iterator dead_user_begin() const { return DeadUsers.begin(); }
243 dead_user_iterator dead_user_end() const { return DeadUsers.end(); }
246 /// \brief Allow iterating the dead expressions referring to this alloca.
248 /// These are operands which have cannot actually be used to refer to the
249 /// alloca as they are outside its range and the user doesn't correct for
250 /// that. These mostly consist of PHI node inputs and the like which we just
251 /// need to replace with undef.
253 typedef SmallVectorImpl<Use *>::const_iterator dead_op_iterator;
254 dead_op_iterator dead_op_begin() const { return DeadOperands.begin(); }
255 dead_op_iterator dead_op_end() const { return DeadOperands.end(); }
258 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
259 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
260 void printSlice(raw_ostream &OS, const_iterator I,
261 StringRef Indent = " ") const;
262 void printUse(raw_ostream &OS, const_iterator I,
263 StringRef Indent = " ") const;
264 void print(raw_ostream &OS) const;
265 void dump(const_iterator I) const;
270 template <typename DerivedT, typename RetT = void> class BuilderBase;
272 friend class AllocaSlices::SliceBuilder;
274 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
275 /// \brief Handle to alloca instruction to simplify method interfaces.
279 /// \brief The instruction responsible for this alloca not having a known set
282 /// When an instruction (potentially) escapes the pointer to the alloca, we
283 /// store a pointer to that here and abort trying to form slices of the
284 /// alloca. This will be null if the alloca slices are analyzed successfully.
285 Instruction *PointerEscapingInstr;
287 /// \brief The slices of the alloca.
289 /// We store a vector of the slices formed by uses of the alloca here. This
290 /// vector is sorted by increasing begin offset, and then the unsplittable
291 /// slices before the splittable ones. See the Slice inner class for more
293 SmallVector<Slice, 8> Slices;
295 /// \brief Instructions which will become dead if we rewrite the alloca.
297 /// Note that these are not separated by slice. This is because we expect an
298 /// alloca to be completely rewritten or not rewritten at all. If rewritten,
299 /// all these instructions can simply be removed and replaced with undef as
300 /// they come from outside of the allocated space.
301 SmallVector<Instruction *, 8> DeadUsers;
303 /// \brief Operands which will become dead if we rewrite the alloca.
305 /// These are operands that in their particular use can be replaced with
306 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
307 /// to PHI nodes and the like. They aren't entirely dead (there might be
308 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
309 /// want to swap this particular input for undef to simplify the use lists of
311 SmallVector<Use *, 8> DeadOperands;
315 static Value *foldSelectInst(SelectInst &SI) {
316 // If the condition being selected on is a constant or the same value is
317 // being selected between, fold the select. Yes this does (rarely) happen
319 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
320 return SI.getOperand(1+CI->isZero());
321 if (SI.getOperand(1) == SI.getOperand(2))
322 return SI.getOperand(1);
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 visitPHINode(PHINode &PN) {
651 return markAsDead(PN);
653 return PI.setAborted(&PN);
655 // See if we already have computed info on this node.
656 uint64_t &PHISize = PHIOrSelectSizes[&PN];
658 // This is a new PHI node, check for an unsafe use of the PHI node.
659 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&PN, PHISize))
660 return PI.setAborted(UnsafeI);
663 // For PHI and select operands outside the alloca, we can't nuke the entire
664 // phi or select -- the other side might still be relevant, so we special
665 // case them here and use a separate structure to track the operands
666 // themselves which should be replaced with undef.
667 // FIXME: This should instead be escaped in the event we're instrumenting
668 // for address sanitization.
669 if (Offset.uge(AllocSize)) {
670 S.DeadOperands.push_back(U);
674 insertUse(PN, Offset, PHISize);
677 void visitSelectInst(SelectInst &SI) {
679 return markAsDead(SI);
680 if (Value *Result = foldSelectInst(SI)) {
682 // If the result of the constant fold will be the pointer, recurse
683 // through the select as if we had RAUW'ed it.
686 // Otherwise the operand to the select is dead, and we can replace it
688 S.DeadOperands.push_back(U);
693 return PI.setAborted(&SI);
695 // See if we already have computed info on this node.
696 uint64_t &SelectSize = PHIOrSelectSizes[&SI];
698 // This is a new Select, check for an unsafe use of it.
699 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&SI, SelectSize))
700 return PI.setAborted(UnsafeI);
703 // For PHI and select operands outside the alloca, we can't nuke the entire
704 // phi or select -- the other side might still be relevant, so we special
705 // case them here and use a separate structure to track the operands
706 // themselves which should be replaced with undef.
707 // FIXME: This should instead be escaped in the event we're instrumenting
708 // for address sanitization.
709 if (Offset.uge(AllocSize)) {
710 S.DeadOperands.push_back(U);
714 insertUse(SI, Offset, SelectSize);
717 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
718 void visitInstruction(Instruction &I) {
723 AllocaSlices::AllocaSlices(const DataLayout &DL, AllocaInst &AI)
725 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
728 PointerEscapingInstr(nullptr) {
729 SliceBuilder PB(DL, AI, *this);
730 SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
731 if (PtrI.isEscaped() || PtrI.isAborted()) {
732 // FIXME: We should sink the escape vs. abort info into the caller nicely,
733 // possibly by just storing the PtrInfo in the AllocaSlices.
734 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
735 : PtrI.getAbortingInst();
736 assert(PointerEscapingInstr && "Did not track a bad instruction");
740 Slices.erase(std::remove_if(Slices.begin(), Slices.end(),
741 std::mem_fun_ref(&Slice::isDead)),
744 #if __cplusplus >= 201103L && !defined(NDEBUG)
745 if (SROARandomShuffleSlices) {
746 std::mt19937 MT(static_cast<unsigned>(sys::TimeValue::now().msec()));
747 std::shuffle(Slices.begin(), Slices.end(), MT);
751 // Sort the uses. This arranges for the offsets to be in ascending order,
752 // and the sizes to be in descending order.
753 std::sort(Slices.begin(), Slices.end());
756 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
758 void AllocaSlices::print(raw_ostream &OS, const_iterator I,
759 StringRef Indent) const {
760 printSlice(OS, I, Indent);
761 printUse(OS, I, Indent);
764 void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
765 StringRef Indent) const {
766 OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
767 << " slice #" << (I - begin())
768 << (I->isSplittable() ? " (splittable)" : "") << "\n";
771 void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
772 StringRef Indent) const {
773 OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
776 void AllocaSlices::print(raw_ostream &OS) const {
777 if (PointerEscapingInstr) {
778 OS << "Can't analyze slices for alloca: " << AI << "\n"
779 << " A pointer to this alloca escaped by:\n"
780 << " " << *PointerEscapingInstr << "\n";
784 OS << "Slices of alloca: " << AI << "\n";
785 for (const_iterator I = begin(), E = end(); I != E; ++I)
789 LLVM_DUMP_METHOD void AllocaSlices::dump(const_iterator I) const {
792 LLVM_DUMP_METHOD void AllocaSlices::dump() const { print(dbgs()); }
794 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
797 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
799 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
800 /// the loads and stores of an alloca instruction, as well as updating its
801 /// debug information. This is used when a domtree is unavailable and thus
802 /// mem2reg in its full form can't be used to handle promotion of allocas to
804 class AllocaPromoter : public LoadAndStorePromoter {
808 SmallVector<DbgDeclareInst *, 4> DDIs;
809 SmallVector<DbgValueInst *, 4> DVIs;
812 AllocaPromoter(const SmallVectorImpl<Instruction *> &Insts, SSAUpdater &S,
813 AllocaInst &AI, DIBuilder &DIB)
814 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
816 void run(const SmallVectorImpl<Instruction*> &Insts) {
817 // Retain the debug information attached to the alloca for use when
818 // rewriting loads and stores.
819 if (MDNode *DebugNode = MDNode::getIfExists(AI.getContext(), &AI)) {
820 for (User *U : DebugNode->users())
821 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U))
823 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(U))
827 LoadAndStorePromoter::run(Insts);
829 // While we have the debug information, clear it off of the alloca. The
830 // caller takes care of deleting the alloca.
831 while (!DDIs.empty())
832 DDIs.pop_back_val()->eraseFromParent();
833 while (!DVIs.empty())
834 DVIs.pop_back_val()->eraseFromParent();
837 bool isInstInList(Instruction *I,
838 const SmallVectorImpl<Instruction*> &Insts) const override {
840 if (LoadInst *LI = dyn_cast<LoadInst>(I))
841 Ptr = LI->getOperand(0);
843 Ptr = cast<StoreInst>(I)->getPointerOperand();
845 // Only used to detect cycles, which will be rare and quickly found as
846 // we're walking up a chain of defs rather than down through uses.
847 SmallPtrSet<Value *, 4> Visited;
853 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr))
854 Ptr = BCI->getOperand(0);
855 else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Ptr))
856 Ptr = GEPI->getPointerOperand();
860 } while (Visited.insert(Ptr));
865 void updateDebugInfo(Instruction *Inst) const override {
866 for (SmallVectorImpl<DbgDeclareInst *>::const_iterator I = DDIs.begin(),
867 E = DDIs.end(); I != E; ++I) {
868 DbgDeclareInst *DDI = *I;
869 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
870 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
871 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
872 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
874 for (SmallVectorImpl<DbgValueInst *>::const_iterator I = DVIs.begin(),
875 E = DVIs.end(); I != E; ++I) {
876 DbgValueInst *DVI = *I;
877 Value *Arg = nullptr;
878 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
879 // If an argument is zero extended then use argument directly. The ZExt
880 // may be zapped by an optimization pass in future.
881 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
882 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
883 else if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
884 Arg = dyn_cast<Argument>(SExt->getOperand(0));
886 Arg = SI->getValueOperand();
887 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
888 Arg = LI->getPointerOperand();
892 Instruction *DbgVal =
893 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
895 DbgVal->setDebugLoc(DVI->getDebugLoc());
899 } // end anon namespace
903 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
905 /// This pass takes allocations which can be completely analyzed (that is, they
906 /// don't escape) and tries to turn them into scalar SSA values. There are
907 /// a few steps to this process.
909 /// 1) It takes allocations of aggregates and analyzes the ways in which they
910 /// are used to try to split them into smaller allocations, ideally of
911 /// a single scalar data type. It will split up memcpy and memset accesses
912 /// as necessary and try to isolate individual scalar accesses.
913 /// 2) It will transform accesses into forms which are suitable for SSA value
914 /// promotion. This can be replacing a memset with a scalar store of an
915 /// integer value, or it can involve speculating operations on a PHI or
916 /// select to be a PHI or select of the results.
917 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
918 /// onto insert and extract operations on a vector value, and convert them to
919 /// this form. By doing so, it will enable promotion of vector aggregates to
920 /// SSA vector values.
921 class SROA : public FunctionPass {
922 const bool RequiresDomTree;
925 const DataLayout *DL;
928 /// \brief Worklist of alloca instructions to simplify.
930 /// Each alloca in the function is added to this. Each new alloca formed gets
931 /// added to it as well to recursively simplify unless that alloca can be
932 /// directly promoted. Finally, each time we rewrite a use of an alloca other
933 /// the one being actively rewritten, we add it back onto the list if not
934 /// already present to ensure it is re-visited.
935 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > Worklist;
937 /// \brief A collection of instructions to delete.
938 /// We try to batch deletions to simplify code and make things a bit more
940 SetVector<Instruction *, SmallVector<Instruction *, 8> > DeadInsts;
942 /// \brief Post-promotion worklist.
944 /// Sometimes we discover an alloca which has a high probability of becoming
945 /// viable for SROA after a round of promotion takes place. In those cases,
946 /// the alloca is enqueued here for re-processing.
948 /// Note that we have to be very careful to clear allocas out of this list in
949 /// the event they are deleted.
950 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > PostPromotionWorklist;
952 /// \brief A collection of alloca instructions we can directly promote.
953 std::vector<AllocaInst *> PromotableAllocas;
955 /// \brief A worklist of PHIs to speculate prior to promoting allocas.
957 /// All of these PHIs have been checked for the safety of speculation and by
958 /// being speculated will allow promoting allocas currently in the promotable
960 SetVector<PHINode *, SmallVector<PHINode *, 2> > SpeculatablePHIs;
962 /// \brief A worklist of select instructions to speculate prior to promoting
965 /// All of these select instructions have been checked for the safety of
966 /// speculation and by being speculated will allow promoting allocas
967 /// currently in the promotable queue.
968 SetVector<SelectInst *, SmallVector<SelectInst *, 2> > SpeculatableSelects;
971 SROA(bool RequiresDomTree = true)
972 : FunctionPass(ID), RequiresDomTree(RequiresDomTree),
973 C(nullptr), DL(nullptr), DT(nullptr) {
974 initializeSROAPass(*PassRegistry::getPassRegistry());
976 bool runOnFunction(Function &F) override;
977 void getAnalysisUsage(AnalysisUsage &AU) const override;
979 const char *getPassName() const override { return "SROA"; }
983 friend class PHIOrSelectSpeculator;
984 friend class AllocaSliceRewriter;
986 bool rewritePartition(AllocaInst &AI, AllocaSlices &S,
987 AllocaSlices::iterator B, AllocaSlices::iterator E,
988 int64_t BeginOffset, int64_t EndOffset,
989 ArrayRef<AllocaSlices::iterator> SplitUses);
990 bool splitAlloca(AllocaInst &AI, AllocaSlices &S);
991 bool runOnAlloca(AllocaInst &AI);
992 void clobberUse(Use &U);
993 void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas);
994 bool promoteAllocas(Function &F);
1000 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
1001 return new SROA(RequiresDomTree);
1004 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates",
1006 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1007 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
1010 /// Walk the range of a partitioning looking for a common type to cover this
1011 /// sequence of slices.
1012 static Type *findCommonType(AllocaSlices::const_iterator B,
1013 AllocaSlices::const_iterator E,
1014 uint64_t EndOffset) {
1016 bool TyIsCommon = true;
1017 IntegerType *ITy = nullptr;
1019 // Note that we need to look at *every* alloca slice's Use to ensure we
1020 // always get consistent results regardless of the order of slices.
1021 for (AllocaSlices::const_iterator I = B; I != E; ++I) {
1022 Use *U = I->getUse();
1023 if (isa<IntrinsicInst>(*U->getUser()))
1025 if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
1028 Type *UserTy = nullptr;
1029 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1030 UserTy = LI->getType();
1031 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1032 UserTy = SI->getValueOperand()->getType();
1035 if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
1036 // If the type is larger than the partition, skip it. We only encounter
1037 // this for split integer operations where we want to use the type of the
1038 // entity causing the split. Also skip if the type is not a byte width
1040 if (UserITy->getBitWidth() % 8 != 0 ||
1041 UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
1044 // Track the largest bitwidth integer type used in this way in case there
1045 // is no common type.
1046 if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
1050 // To avoid depending on the order of slices, Ty and TyIsCommon must not
1051 // depend on types skipped above.
1052 if (!UserTy || (Ty && Ty != UserTy))
1053 TyIsCommon = false; // Give up on anything but an iN type.
1058 return TyIsCommon ? Ty : ITy;
1061 /// PHI instructions that use an alloca and are subsequently loaded can be
1062 /// rewritten to load both input pointers in the pred blocks and then PHI the
1063 /// results, allowing the load of the alloca to be promoted.
1065 /// %P2 = phi [i32* %Alloca, i32* %Other]
1066 /// %V = load i32* %P2
1068 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1070 /// %V2 = load i32* %Other
1072 /// %V = phi [i32 %V1, i32 %V2]
1074 /// We can do this to a select if its only uses are loads and if the operands
1075 /// to the select can be loaded unconditionally.
1077 /// FIXME: This should be hoisted into a generic utility, likely in
1078 /// Transforms/Util/Local.h
1079 static bool isSafePHIToSpeculate(PHINode &PN,
1080 const DataLayout *DL = nullptr) {
1081 // For now, we can only do this promotion if the load is in the same block
1082 // as the PHI, and if there are no stores between the phi and load.
1083 // TODO: Allow recursive phi users.
1084 // TODO: Allow stores.
1085 BasicBlock *BB = PN.getParent();
1086 unsigned MaxAlign = 0;
1087 bool HaveLoad = false;
1088 for (User *U : PN.users()) {
1089 LoadInst *LI = dyn_cast<LoadInst>(U);
1090 if (!LI || !LI->isSimple())
1093 // For now we only allow loads in the same block as the PHI. This is
1094 // a common case that happens when instcombine merges two loads through
1096 if (LI->getParent() != BB)
1099 // Ensure that there are no instructions between the PHI and the load that
1101 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1102 if (BBI->mayWriteToMemory())
1105 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1112 // We can only transform this if it is safe to push the loads into the
1113 // predecessor blocks. The only thing to watch out for is that we can't put
1114 // a possibly trapping load in the predecessor if it is a critical edge.
1115 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1116 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1117 Value *InVal = PN.getIncomingValue(Idx);
1119 // If the value is produced by the terminator of the predecessor (an
1120 // invoke) or it has side-effects, there is no valid place to put a load
1121 // in the predecessor.
1122 if (TI == InVal || TI->mayHaveSideEffects())
1125 // If the predecessor has a single successor, then the edge isn't
1127 if (TI->getNumSuccessors() == 1)
1130 // If this pointer is always safe to load, or if we can prove that there
1131 // is already a load in the block, then we can move the load to the pred
1133 if (InVal->isDereferenceablePointer(DL) ||
1134 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, DL))
1143 static void speculatePHINodeLoads(PHINode &PN) {
1144 DEBUG(dbgs() << " original: " << PN << "\n");
1146 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1147 IRBuilderTy PHIBuilder(&PN);
1148 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1149 PN.getName() + ".sroa.speculated");
1151 // Get the AA tags and alignment to use from one of the loads. It doesn't
1152 // matter which one we get and if any differ.
1153 LoadInst *SomeLoad = cast<LoadInst>(PN.user_back());
1156 SomeLoad->getAAMetadata(AATags);
1157 unsigned Align = SomeLoad->getAlignment();
1159 // Rewrite all loads of the PN to use the new PHI.
1160 while (!PN.use_empty()) {
1161 LoadInst *LI = cast<LoadInst>(PN.user_back());
1162 LI->replaceAllUsesWith(NewPN);
1163 LI->eraseFromParent();
1166 // Inject loads into all of the pred blocks.
1167 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1168 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1169 TerminatorInst *TI = Pred->getTerminator();
1170 Value *InVal = PN.getIncomingValue(Idx);
1171 IRBuilderTy PredBuilder(TI);
1173 LoadInst *Load = PredBuilder.CreateLoad(
1174 InVal, (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1175 ++NumLoadsSpeculated;
1176 Load->setAlignment(Align);
1178 Load->setAAMetadata(AATags);
1179 NewPN->addIncoming(Load, Pred);
1182 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1183 PN.eraseFromParent();
1186 /// Select instructions that use an alloca and are subsequently loaded can be
1187 /// rewritten to load both input pointers and then select between the result,
1188 /// allowing the load of the alloca to be promoted.
1190 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1191 /// %V = load i32* %P2
1193 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1194 /// %V2 = load i32* %Other
1195 /// %V = select i1 %cond, i32 %V1, i32 %V2
1197 /// We can do this to a select if its only uses are loads and if the operand
1198 /// to the select can be loaded unconditionally.
1199 static bool isSafeSelectToSpeculate(SelectInst &SI,
1200 const DataLayout *DL = nullptr) {
1201 Value *TValue = SI.getTrueValue();
1202 Value *FValue = SI.getFalseValue();
1203 bool TDerefable = TValue->isDereferenceablePointer(DL);
1204 bool FDerefable = FValue->isDereferenceablePointer(DL);
1206 for (User *U : SI.users()) {
1207 LoadInst *LI = dyn_cast<LoadInst>(U);
1208 if (!LI || !LI->isSimple())
1211 // Both operands to the select need to be dereferencable, either
1212 // absolutely (e.g. allocas) or at this point because we can see other
1215 !isSafeToLoadUnconditionally(TValue, LI, LI->getAlignment(), DL))
1218 !isSafeToLoadUnconditionally(FValue, LI, LI->getAlignment(), DL))
1225 static void speculateSelectInstLoads(SelectInst &SI) {
1226 DEBUG(dbgs() << " original: " << SI << "\n");
1228 IRBuilderTy IRB(&SI);
1229 Value *TV = SI.getTrueValue();
1230 Value *FV = SI.getFalseValue();
1231 // Replace the loads of the select with a select of two loads.
1232 while (!SI.use_empty()) {
1233 LoadInst *LI = cast<LoadInst>(SI.user_back());
1234 assert(LI->isSimple() && "We only speculate simple loads");
1236 IRB.SetInsertPoint(LI);
1238 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1240 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1241 NumLoadsSpeculated += 2;
1243 // Transfer alignment and AA info if present.
1244 TL->setAlignment(LI->getAlignment());
1245 FL->setAlignment(LI->getAlignment());
1248 LI->getAAMetadata(Tags);
1250 TL->setAAMetadata(Tags);
1251 FL->setAAMetadata(Tags);
1254 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1255 LI->getName() + ".sroa.speculated");
1257 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1258 LI->replaceAllUsesWith(V);
1259 LI->eraseFromParent();
1261 SI.eraseFromParent();
1264 /// \brief Build a GEP out of a base pointer and indices.
1266 /// This will return the BasePtr if that is valid, or build a new GEP
1267 /// instruction using the IRBuilder if GEP-ing is needed.
1268 static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1269 SmallVectorImpl<Value *> &Indices, Twine NamePrefix) {
1270 if (Indices.empty())
1273 // A single zero index is a no-op, so check for this and avoid building a GEP
1275 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1278 return IRB.CreateInBoundsGEP(BasePtr, Indices, NamePrefix + "sroa_idx");
1281 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1282 /// TargetTy without changing the offset of the pointer.
1284 /// This routine assumes we've already established a properly offset GEP with
1285 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1286 /// zero-indices down through type layers until we find one the same as
1287 /// TargetTy. If we can't find one with the same type, we at least try to use
1288 /// one with the same size. If none of that works, we just produce the GEP as
1289 /// indicated by Indices to have the correct offset.
1290 static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1291 Value *BasePtr, Type *Ty, Type *TargetTy,
1292 SmallVectorImpl<Value *> &Indices,
1295 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1297 // Pointer size to use for the indices.
1298 unsigned PtrSize = DL.getPointerTypeSizeInBits(BasePtr->getType());
1300 // See if we can descend into a struct and locate a field with the correct
1302 unsigned NumLayers = 0;
1303 Type *ElementTy = Ty;
1305 if (ElementTy->isPointerTy())
1308 if (ArrayType *ArrayTy = dyn_cast<ArrayType>(ElementTy)) {
1309 ElementTy = ArrayTy->getElementType();
1310 Indices.push_back(IRB.getIntN(PtrSize, 0));
1311 } else if (VectorType *VectorTy = dyn_cast<VectorType>(ElementTy)) {
1312 ElementTy = VectorTy->getElementType();
1313 Indices.push_back(IRB.getInt32(0));
1314 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1315 if (STy->element_begin() == STy->element_end())
1316 break; // Nothing left to descend into.
1317 ElementTy = *STy->element_begin();
1318 Indices.push_back(IRB.getInt32(0));
1323 } while (ElementTy != TargetTy);
1324 if (ElementTy != TargetTy)
1325 Indices.erase(Indices.end() - NumLayers, Indices.end());
1327 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1330 /// \brief Recursively compute indices for a natural GEP.
1332 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1333 /// element types adding appropriate indices for the GEP.
1334 static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1335 Value *Ptr, Type *Ty, APInt &Offset,
1337 SmallVectorImpl<Value *> &Indices,
1340 return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices, NamePrefix);
1342 // We can't recurse through pointer types.
1343 if (Ty->isPointerTy())
1346 // We try to analyze GEPs over vectors here, but note that these GEPs are
1347 // extremely poorly defined currently. The long-term goal is to remove GEPing
1348 // over a vector from the IR completely.
1349 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1350 unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
1351 if (ElementSizeInBits % 8 != 0) {
1352 // GEPs over non-multiple of 8 size vector elements are invalid.
1355 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1356 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1357 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1359 Offset -= NumSkippedElements * ElementSize;
1360 Indices.push_back(IRB.getInt(NumSkippedElements));
1361 return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1362 Offset, TargetTy, Indices, NamePrefix);
1365 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1366 Type *ElementTy = ArrTy->getElementType();
1367 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1368 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1369 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1372 Offset -= NumSkippedElements * ElementSize;
1373 Indices.push_back(IRB.getInt(NumSkippedElements));
1374 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1375 Indices, NamePrefix);
1378 StructType *STy = dyn_cast<StructType>(Ty);
1382 const StructLayout *SL = DL.getStructLayout(STy);
1383 uint64_t StructOffset = Offset.getZExtValue();
1384 if (StructOffset >= SL->getSizeInBytes())
1386 unsigned Index = SL->getElementContainingOffset(StructOffset);
1387 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1388 Type *ElementTy = STy->getElementType(Index);
1389 if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
1390 return nullptr; // The offset points into alignment padding.
1392 Indices.push_back(IRB.getInt32(Index));
1393 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1394 Indices, NamePrefix);
1397 /// \brief Get a natural GEP from a base pointer to a particular offset and
1398 /// resulting in a particular type.
1400 /// The goal is to produce a "natural" looking GEP that works with the existing
1401 /// composite types to arrive at the appropriate offset and element type for
1402 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1403 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1404 /// Indices, and setting Ty to the result subtype.
1406 /// If no natural GEP can be constructed, this function returns null.
1407 static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1408 Value *Ptr, APInt Offset, Type *TargetTy,
1409 SmallVectorImpl<Value *> &Indices,
1411 PointerType *Ty = cast<PointerType>(Ptr->getType());
1413 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1415 if (Ty == IRB.getInt8PtrTy(Ty->getAddressSpace()) && TargetTy->isIntegerTy(8))
1418 Type *ElementTy = Ty->getElementType();
1419 if (!ElementTy->isSized())
1420 return nullptr; // We can't GEP through an unsized element.
1421 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1422 if (ElementSize == 0)
1423 return nullptr; // Zero-length arrays can't help us build a natural GEP.
1424 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1426 Offset -= NumSkippedElements * ElementSize;
1427 Indices.push_back(IRB.getInt(NumSkippedElements));
1428 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1429 Indices, NamePrefix);
1432 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1433 /// resulting pointer has PointerTy.
1435 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1436 /// and produces the pointer type desired. Where it cannot, it will try to use
1437 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1438 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1439 /// bitcast to the type.
1441 /// The strategy for finding the more natural GEPs is to peel off layers of the
1442 /// pointer, walking back through bit casts and GEPs, searching for a base
1443 /// pointer from which we can compute a natural GEP with the desired
1444 /// properties. The algorithm tries to fold as many constant indices into
1445 /// a single GEP as possible, thus making each GEP more independent of the
1446 /// surrounding code.
1447 static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr,
1448 APInt Offset, Type *PointerTy,
1450 // Even though we don't look through PHI nodes, we could be called on an
1451 // instruction in an unreachable block, which may be on a cycle.
1452 SmallPtrSet<Value *, 4> Visited;
1453 Visited.insert(Ptr);
1454 SmallVector<Value *, 4> Indices;
1456 // We may end up computing an offset pointer that has the wrong type. If we
1457 // never are able to compute one directly that has the correct type, we'll
1458 // fall back to it, so keep it around here.
1459 Value *OffsetPtr = nullptr;
1461 // Remember any i8 pointer we come across to re-use if we need to do a raw
1463 Value *Int8Ptr = nullptr;
1464 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1466 Type *TargetTy = PointerTy->getPointerElementType();
1469 // First fold any existing GEPs into the offset.
1470 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1471 APInt GEPOffset(Offset.getBitWidth(), 0);
1472 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1474 Offset += GEPOffset;
1475 Ptr = GEP->getPointerOperand();
1476 if (!Visited.insert(Ptr))
1480 // See if we can perform a natural GEP here.
1482 if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1483 Indices, NamePrefix)) {
1484 if (P->getType() == PointerTy) {
1485 // Zap any offset pointer that we ended up computing in previous rounds.
1486 if (OffsetPtr && OffsetPtr->use_empty())
1487 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1488 I->eraseFromParent();
1496 // Stash this pointer if we've found an i8*.
1497 if (Ptr->getType()->isIntegerTy(8)) {
1499 Int8PtrOffset = Offset;
1502 // Peel off a layer of the pointer and update the offset appropriately.
1503 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1504 Ptr = cast<Operator>(Ptr)->getOperand(0);
1505 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1506 if (GA->mayBeOverridden())
1508 Ptr = GA->getAliasee();
1512 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1513 } while (Visited.insert(Ptr));
1517 Int8Ptr = IRB.CreateBitCast(
1518 Ptr, IRB.getInt8PtrTy(PointerTy->getPointerAddressSpace()),
1519 NamePrefix + "sroa_raw_cast");
1520 Int8PtrOffset = Offset;
1523 OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
1524 IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1525 NamePrefix + "sroa_raw_idx");
1529 // On the off chance we were targeting i8*, guard the bitcast here.
1530 if (Ptr->getType() != PointerTy)
1531 Ptr = IRB.CreateBitCast(Ptr, PointerTy, NamePrefix + "sroa_cast");
1536 /// \brief Test whether we can convert a value from the old to the new type.
1538 /// This predicate should be used to guard calls to convertValue in order to
1539 /// ensure that we only try to convert viable values. The strategy is that we
1540 /// will peel off single element struct and array wrappings to get to an
1541 /// underlying value, and convert that value.
1542 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1545 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1546 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1547 if (NewITy->getBitWidth() >= OldITy->getBitWidth())
1549 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1551 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1554 // We can convert pointers to integers and vice-versa. Same for vectors
1555 // of pointers and integers.
1556 OldTy = OldTy->getScalarType();
1557 NewTy = NewTy->getScalarType();
1558 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1559 if (NewTy->isPointerTy() && OldTy->isPointerTy())
1561 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
1569 /// \brief Generic routine to convert an SSA value to a value of a different
1572 /// This will try various different casting techniques, such as bitcasts,
1573 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1574 /// two types for viability with this routine.
1575 static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1577 Type *OldTy = V->getType();
1578 assert(canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type");
1583 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1584 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1585 if (NewITy->getBitWidth() > OldITy->getBitWidth())
1586 return IRB.CreateZExt(V, NewITy);
1588 // See if we need inttoptr for this type pair. A cast involving both scalars
1589 // and vectors requires and additional bitcast.
1590 if (OldTy->getScalarType()->isIntegerTy() &&
1591 NewTy->getScalarType()->isPointerTy()) {
1592 // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1593 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1594 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1597 // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1598 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1599 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1602 return IRB.CreateIntToPtr(V, NewTy);
1605 // See if we need ptrtoint for this type pair. A cast involving both scalars
1606 // and vectors requires and additional bitcast.
1607 if (OldTy->getScalarType()->isPointerTy() &&
1608 NewTy->getScalarType()->isIntegerTy()) {
1609 // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1610 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1611 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1614 // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1615 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1616 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1619 return IRB.CreatePtrToInt(V, NewTy);
1622 return IRB.CreateBitCast(V, NewTy);
1625 /// \brief Test whether the given slice use can be promoted to a vector.
1627 /// This function is called to test each entry in a partioning which is slated
1628 /// for a single slice.
1629 static bool isVectorPromotionViableForSlice(
1630 const DataLayout &DL, AllocaSlices &S, uint64_t SliceBeginOffset,
1631 uint64_t SliceEndOffset, VectorType *Ty, uint64_t ElementSize,
1632 AllocaSlices::const_iterator I) {
1633 // First validate the slice offsets.
1634 uint64_t BeginOffset =
1635 std::max(I->beginOffset(), SliceBeginOffset) - SliceBeginOffset;
1636 uint64_t BeginIndex = BeginOffset / ElementSize;
1637 if (BeginIndex * ElementSize != BeginOffset ||
1638 BeginIndex >= Ty->getNumElements())
1640 uint64_t EndOffset =
1641 std::min(I->endOffset(), SliceEndOffset) - SliceBeginOffset;
1642 uint64_t EndIndex = EndOffset / ElementSize;
1643 if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
1646 assert(EndIndex > BeginIndex && "Empty vector!");
1647 uint64_t NumElements = EndIndex - BeginIndex;
1649 (NumElements == 1) ? Ty->getElementType()
1650 : VectorType::get(Ty->getElementType(), NumElements);
1653 Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1655 Use *U = I->getUse();
1657 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1658 if (MI->isVolatile())
1660 if (!I->isSplittable())
1661 return false; // Skip any unsplittable intrinsics.
1662 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1663 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1664 II->getIntrinsicID() != Intrinsic::lifetime_end)
1666 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1667 // Disable vector promotion when there are loads or stores of an FCA.
1669 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1670 if (LI->isVolatile())
1672 Type *LTy = LI->getType();
1673 if (SliceBeginOffset > I->beginOffset() ||
1674 SliceEndOffset < I->endOffset()) {
1675 assert(LTy->isIntegerTy());
1678 if (!canConvertValue(DL, SliceTy, LTy))
1680 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1681 if (SI->isVolatile())
1683 Type *STy = SI->getValueOperand()->getType();
1684 if (SliceBeginOffset > I->beginOffset() ||
1685 SliceEndOffset < I->endOffset()) {
1686 assert(STy->isIntegerTy());
1689 if (!canConvertValue(DL, STy, SliceTy))
1698 /// \brief Test whether the given alloca partitioning and range of slices can be
1699 /// promoted to a vector.
1701 /// This is a quick test to check whether we can rewrite a particular alloca
1702 /// partition (and its newly formed alloca) into a vector alloca with only
1703 /// whole-vector loads and stores such that it could be promoted to a vector
1704 /// SSA value. We only can ensure this for a limited set of operations, and we
1705 /// don't want to do the rewrites unless we are confident that the result will
1706 /// be promotable, so we have an early test here.
1708 isVectorPromotionViable(const DataLayout &DL, Type *AllocaTy, AllocaSlices &S,
1709 uint64_t SliceBeginOffset, uint64_t SliceEndOffset,
1710 AllocaSlices::const_iterator I,
1711 AllocaSlices::const_iterator E,
1712 ArrayRef<AllocaSlices::iterator> SplitUses) {
1713 VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
1717 uint64_t ElementSize = DL.getTypeSizeInBits(Ty->getScalarType());
1719 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1720 // that aren't byte sized.
1721 if (ElementSize % 8)
1723 assert((DL.getTypeSizeInBits(Ty) % 8) == 0 &&
1724 "vector size not a multiple of element size?");
1728 if (!isVectorPromotionViableForSlice(DL, S, SliceBeginOffset,
1729 SliceEndOffset, Ty, ElementSize, I))
1732 for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
1733 SUE = SplitUses.end();
1735 if (!isVectorPromotionViableForSlice(DL, S, SliceBeginOffset,
1736 SliceEndOffset, Ty, ElementSize, *SUI))
1742 /// \brief Test whether a slice of an alloca is valid for integer widening.
1744 /// This implements the necessary checking for the \c isIntegerWideningViable
1745 /// test below on a single slice of the alloca.
1746 static bool isIntegerWideningViableForSlice(const DataLayout &DL,
1748 uint64_t AllocBeginOffset,
1749 uint64_t Size, AllocaSlices &S,
1750 AllocaSlices::const_iterator I,
1751 bool &WholeAllocaOp) {
1752 uint64_t RelBegin = I->beginOffset() - AllocBeginOffset;
1753 uint64_t RelEnd = I->endOffset() - AllocBeginOffset;
1755 // We can't reasonably handle cases where the load or store extends past
1756 // the end of the aloca's type and into its padding.
1760 Use *U = I->getUse();
1762 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1763 if (LI->isVolatile())
1765 if (RelBegin == 0 && RelEnd == Size)
1766 WholeAllocaOp = true;
1767 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
1768 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1770 } else if (RelBegin != 0 || RelEnd != Size ||
1771 !canConvertValue(DL, AllocaTy, LI->getType())) {
1772 // Non-integer loads need to be convertible from the alloca type so that
1773 // they are promotable.
1776 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1777 Type *ValueTy = SI->getValueOperand()->getType();
1778 if (SI->isVolatile())
1780 if (RelBegin == 0 && RelEnd == Size)
1781 WholeAllocaOp = true;
1782 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
1783 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1785 } else if (RelBegin != 0 || RelEnd != Size ||
1786 !canConvertValue(DL, ValueTy, AllocaTy)) {
1787 // Non-integer stores need to be convertible to the alloca type so that
1788 // they are promotable.
1791 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1792 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
1794 if (!I->isSplittable())
1795 return false; // Skip any unsplittable intrinsics.
1796 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1797 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1798 II->getIntrinsicID() != Intrinsic::lifetime_end)
1807 /// \brief Test whether the given alloca partition's integer operations can be
1808 /// widened to promotable ones.
1810 /// This is a quick test to check whether we can rewrite the integer loads and
1811 /// stores to a particular alloca into wider loads and stores and be able to
1812 /// promote the resulting alloca.
1814 isIntegerWideningViable(const DataLayout &DL, Type *AllocaTy,
1815 uint64_t AllocBeginOffset, AllocaSlices &S,
1816 AllocaSlices::const_iterator I,
1817 AllocaSlices::const_iterator E,
1818 ArrayRef<AllocaSlices::iterator> SplitUses) {
1819 uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
1820 // Don't create integer types larger than the maximum bitwidth.
1821 if (SizeInBits > IntegerType::MAX_INT_BITS)
1824 // Don't try to handle allocas with bit-padding.
1825 if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
1828 // We need to ensure that an integer type with the appropriate bitwidth can
1829 // be converted to the alloca type, whatever that is. We don't want to force
1830 // the alloca itself to have an integer type if there is a more suitable one.
1831 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
1832 if (!canConvertValue(DL, AllocaTy, IntTy) ||
1833 !canConvertValue(DL, IntTy, AllocaTy))
1836 uint64_t Size = DL.getTypeStoreSize(AllocaTy);
1838 // While examining uses, we ensure that the alloca has a covering load or
1839 // store. We don't want to widen the integer operations only to fail to
1840 // promote due to some other unsplittable entry (which we may make splittable
1841 // later). However, if there are only splittable uses, go ahead and assume
1842 // that we cover the alloca.
1843 bool WholeAllocaOp = (I != E) ? false : DL.isLegalInteger(SizeInBits);
1846 if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
1847 S, I, WholeAllocaOp))
1850 for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
1851 SUE = SplitUses.end();
1853 if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
1854 S, *SUI, WholeAllocaOp))
1857 return WholeAllocaOp;
1860 static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1861 IntegerType *Ty, uint64_t Offset,
1862 const Twine &Name) {
1863 DEBUG(dbgs() << " start: " << *V << "\n");
1864 IntegerType *IntTy = cast<IntegerType>(V->getType());
1865 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
1866 "Element extends past full value");
1867 uint64_t ShAmt = 8*Offset;
1868 if (DL.isBigEndian())
1869 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
1871 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
1872 DEBUG(dbgs() << " shifted: " << *V << "\n");
1874 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
1875 "Cannot extract to a larger integer!");
1877 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
1878 DEBUG(dbgs() << " trunced: " << *V << "\n");
1883 static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
1884 Value *V, uint64_t Offset, const Twine &Name) {
1885 IntegerType *IntTy = cast<IntegerType>(Old->getType());
1886 IntegerType *Ty = cast<IntegerType>(V->getType());
1887 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
1888 "Cannot insert a larger integer!");
1889 DEBUG(dbgs() << " start: " << *V << "\n");
1891 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
1892 DEBUG(dbgs() << " extended: " << *V << "\n");
1894 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
1895 "Element store outside of alloca store");
1896 uint64_t ShAmt = 8*Offset;
1897 if (DL.isBigEndian())
1898 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
1900 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
1901 DEBUG(dbgs() << " shifted: " << *V << "\n");
1904 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
1905 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
1906 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
1907 DEBUG(dbgs() << " masked: " << *Old << "\n");
1908 V = IRB.CreateOr(Old, V, Name + ".insert");
1909 DEBUG(dbgs() << " inserted: " << *V << "\n");
1914 static Value *extractVector(IRBuilderTy &IRB, Value *V,
1915 unsigned BeginIndex, unsigned EndIndex,
1916 const Twine &Name) {
1917 VectorType *VecTy = cast<VectorType>(V->getType());
1918 unsigned NumElements = EndIndex - BeginIndex;
1919 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
1921 if (NumElements == VecTy->getNumElements())
1924 if (NumElements == 1) {
1925 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
1927 DEBUG(dbgs() << " extract: " << *V << "\n");
1931 SmallVector<Constant*, 8> Mask;
1932 Mask.reserve(NumElements);
1933 for (unsigned i = BeginIndex; i != EndIndex; ++i)
1934 Mask.push_back(IRB.getInt32(i));
1935 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
1936 ConstantVector::get(Mask),
1938 DEBUG(dbgs() << " shuffle: " << *V << "\n");
1942 static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
1943 unsigned BeginIndex, const Twine &Name) {
1944 VectorType *VecTy = cast<VectorType>(Old->getType());
1945 assert(VecTy && "Can only insert a vector into a vector");
1947 VectorType *Ty = dyn_cast<VectorType>(V->getType());
1949 // Single element to insert.
1950 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
1952 DEBUG(dbgs() << " insert: " << *V << "\n");
1956 assert(Ty->getNumElements() <= VecTy->getNumElements() &&
1957 "Too many elements!");
1958 if (Ty->getNumElements() == VecTy->getNumElements()) {
1959 assert(V->getType() == VecTy && "Vector type mismatch");
1962 unsigned EndIndex = BeginIndex + Ty->getNumElements();
1964 // When inserting a smaller vector into the larger to store, we first
1965 // use a shuffle vector to widen it with undef elements, and then
1966 // a second shuffle vector to select between the loaded vector and the
1968 SmallVector<Constant*, 8> Mask;
1969 Mask.reserve(VecTy->getNumElements());
1970 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
1971 if (i >= BeginIndex && i < EndIndex)
1972 Mask.push_back(IRB.getInt32(i - BeginIndex));
1974 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
1975 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
1976 ConstantVector::get(Mask),
1978 DEBUG(dbgs() << " shuffle: " << *V << "\n");
1981 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
1982 Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
1984 V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
1986 DEBUG(dbgs() << " blend: " << *V << "\n");
1991 /// \brief Visitor to rewrite instructions using p particular slice of an alloca
1992 /// to use a new alloca.
1994 /// Also implements the rewriting to vector-based accesses when the partition
1995 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
1997 class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
1998 // Befriend the base class so it can delegate to private visit methods.
1999 friend class llvm::InstVisitor<AllocaSliceRewriter, bool>;
2000 typedef llvm::InstVisitor<AllocaSliceRewriter, bool> Base;
2002 const DataLayout &DL;
2005 AllocaInst &OldAI, &NewAI;
2006 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2009 // If we are rewriting an alloca partition which can be written as pure
2010 // vector operations, we stash extra information here. When VecTy is
2011 // non-null, we have some strict guarantees about the rewritten alloca:
2012 // - The new alloca is exactly the size of the vector type here.
2013 // - The accesses all either map to the entire vector or to a single
2015 // - The set of accessing instructions is only one of those handled above
2016 // in isVectorPromotionViable. Generally these are the same access kinds
2017 // which are promotable via mem2reg.
2020 uint64_t ElementSize;
2022 // This is a convenience and flag variable that will be null unless the new
2023 // alloca's integer operations should be widened to this integer type due to
2024 // passing isIntegerWideningViable above. If it is non-null, the desired
2025 // integer type will be stored here for easy access during rewriting.
2028 // The original offset of the slice currently being rewritten relative to
2029 // the original alloca.
2030 uint64_t BeginOffset, EndOffset;
2031 // The new offsets of the slice currently being rewritten relative to the
2033 uint64_t NewBeginOffset, NewEndOffset;
2039 Instruction *OldPtr;
2041 // Track post-rewrite users which are PHI nodes and Selects.
2042 SmallPtrSetImpl<PHINode *> &PHIUsers;
2043 SmallPtrSetImpl<SelectInst *> &SelectUsers;
2045 // Utility IR builder, whose name prefix is setup for each visited use, and
2046 // the insertion point is set to point to the user.
2050 AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &S, SROA &Pass,
2051 AllocaInst &OldAI, AllocaInst &NewAI,
2052 uint64_t NewAllocaBeginOffset,
2053 uint64_t NewAllocaEndOffset, bool IsVectorPromotable,
2054 bool IsIntegerPromotable,
2055 SmallPtrSetImpl<PHINode *> &PHIUsers,
2056 SmallPtrSetImpl<SelectInst *> &SelectUsers)
2057 : DL(DL), S(S), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
2058 NewAllocaBeginOffset(NewAllocaBeginOffset),
2059 NewAllocaEndOffset(NewAllocaEndOffset),
2060 NewAllocaTy(NewAI.getAllocatedType()),
2061 VecTy(IsVectorPromotable ? cast<VectorType>(NewAllocaTy) : nullptr),
2062 ElementTy(VecTy ? VecTy->getElementType() : nullptr),
2063 ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
2064 IntTy(IsIntegerPromotable
2067 DL.getTypeSizeInBits(NewAI.getAllocatedType()))
2069 BeginOffset(), EndOffset(), IsSplittable(), IsSplit(), OldUse(),
2070 OldPtr(), PHIUsers(PHIUsers), SelectUsers(SelectUsers),
2071 IRB(NewAI.getContext(), ConstantFolder()) {
2073 assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
2074 "Only multiple-of-8 sized vector elements are viable");
2077 assert((!IsVectorPromotable && !IsIntegerPromotable) ||
2078 IsVectorPromotable != IsIntegerPromotable);
2081 bool visit(AllocaSlices::const_iterator I) {
2082 bool CanSROA = true;
2083 BeginOffset = I->beginOffset();
2084 EndOffset = I->endOffset();
2085 IsSplittable = I->isSplittable();
2087 BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
2089 // Compute the intersecting offset range.
2090 assert(BeginOffset < NewAllocaEndOffset);
2091 assert(EndOffset > NewAllocaBeginOffset);
2092 NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2093 NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2095 SliceSize = NewEndOffset - NewBeginOffset;
2097 OldUse = I->getUse();
2098 OldPtr = cast<Instruction>(OldUse->get());
2100 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2101 IRB.SetInsertPoint(OldUserI);
2102 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2103 IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
2105 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2112 // Make sure the other visit overloads are visible.
2115 // Every instruction which can end up as a user must have a rewrite rule.
2116 bool visitInstruction(Instruction &I) {
2117 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2118 llvm_unreachable("No rewrite rule for this instruction!");
2121 Value *getNewAllocaSlicePtr(IRBuilderTy &IRB, Type *PointerTy) {
2122 // Note that the offset computation can use BeginOffset or NewBeginOffset
2123 // interchangeably for unsplit slices.
2124 assert(IsSplit || BeginOffset == NewBeginOffset);
2125 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2128 StringRef OldName = OldPtr->getName();
2129 // Skip through the last '.sroa.' component of the name.
2130 size_t LastSROAPrefix = OldName.rfind(".sroa.");
2131 if (LastSROAPrefix != StringRef::npos) {
2132 OldName = OldName.substr(LastSROAPrefix + strlen(".sroa."));
2133 // Look for an SROA slice index.
2134 size_t IndexEnd = OldName.find_first_not_of("0123456789");
2135 if (IndexEnd != StringRef::npos && OldName[IndexEnd] == '.') {
2136 // Strip the index and look for the offset.
2137 OldName = OldName.substr(IndexEnd + 1);
2138 size_t OffsetEnd = OldName.find_first_not_of("0123456789");
2139 if (OffsetEnd != StringRef::npos && OldName[OffsetEnd] == '.')
2140 // Strip the offset.
2141 OldName = OldName.substr(OffsetEnd + 1);
2144 // Strip any SROA suffixes as well.
2145 OldName = OldName.substr(0, OldName.find(".sroa_"));
2148 return getAdjustedPtr(IRB, DL, &NewAI,
2149 APInt(DL.getPointerSizeInBits(), Offset), PointerTy,
2151 Twine(OldName) + "."
2158 /// \brief Compute suitable alignment to access this slice of the *new* alloca.
2160 /// You can optionally pass a type to this routine and if that type's ABI
2161 /// alignment is itself suitable, this will return zero.
2162 unsigned getSliceAlign(Type *Ty = nullptr) {
2163 unsigned NewAIAlign = NewAI.getAlignment();
2165 NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
2166 unsigned Align = MinAlign(NewAIAlign, NewBeginOffset - NewAllocaBeginOffset);
2167 return (Ty && Align == DL.getABITypeAlignment(Ty)) ? 0 : Align;
2170 unsigned getIndex(uint64_t Offset) {
2171 assert(VecTy && "Can only call getIndex when rewriting a vector");
2172 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2173 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2174 uint32_t Index = RelOffset / ElementSize;
2175 assert(Index * ElementSize == RelOffset);
2179 void deleteIfTriviallyDead(Value *V) {
2180 Instruction *I = cast<Instruction>(V);
2181 if (isInstructionTriviallyDead(I))
2182 Pass.DeadInsts.insert(I);
2185 Value *rewriteVectorizedLoadInst() {
2186 unsigned BeginIndex = getIndex(NewBeginOffset);
2187 unsigned EndIndex = getIndex(NewEndOffset);
2188 assert(EndIndex > BeginIndex && "Empty vector!");
2190 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2192 return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2195 Value *rewriteIntegerLoad(LoadInst &LI) {
2196 assert(IntTy && "We cannot insert an integer to the alloca");
2197 assert(!LI.isVolatile());
2198 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2200 V = convertValue(DL, IRB, V, IntTy);
2201 assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2202 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2203 if (Offset > 0 || NewEndOffset < NewAllocaEndOffset)
2204 V = extractInteger(DL, IRB, V, cast<IntegerType>(LI.getType()), Offset,
2209 bool visitLoadInst(LoadInst &LI) {
2210 DEBUG(dbgs() << " original: " << LI << "\n");
2211 Value *OldOp = LI.getOperand(0);
2212 assert(OldOp == OldPtr);
2214 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), SliceSize * 8)
2216 bool IsPtrAdjusted = false;
2219 V = rewriteVectorizedLoadInst();
2220 } else if (IntTy && LI.getType()->isIntegerTy()) {
2221 V = rewriteIntegerLoad(LI);
2222 } else if (NewBeginOffset == NewAllocaBeginOffset &&
2223 canConvertValue(DL, NewAllocaTy, LI.getType())) {
2224 V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2225 LI.isVolatile(), LI.getName());
2227 Type *LTy = TargetTy->getPointerTo();
2228 V = IRB.CreateAlignedLoad(getNewAllocaSlicePtr(IRB, LTy),
2229 getSliceAlign(TargetTy), LI.isVolatile(),
2231 IsPtrAdjusted = true;
2233 V = convertValue(DL, IRB, V, TargetTy);
2236 assert(!LI.isVolatile());
2237 assert(LI.getType()->isIntegerTy() &&
2238 "Only integer type loads and stores are split");
2239 assert(SliceSize < DL.getTypeStoreSize(LI.getType()) &&
2240 "Split load isn't smaller than original load");
2241 assert(LI.getType()->getIntegerBitWidth() ==
2242 DL.getTypeStoreSizeInBits(LI.getType()) &&
2243 "Non-byte-multiple bit width");
2244 // Move the insertion point just past the load so that we can refer to it.
2245 IRB.SetInsertPoint(std::next(BasicBlock::iterator(&LI)));
2246 // Create a placeholder value with the same type as LI to use as the
2247 // basis for the new value. This allows us to replace the uses of LI with
2248 // the computed value, and then replace the placeholder with LI, leaving
2249 // LI only used for this computation.
2251 = new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
2252 V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset,
2254 LI.replaceAllUsesWith(V);
2255 Placeholder->replaceAllUsesWith(&LI);
2258 LI.replaceAllUsesWith(V);
2261 Pass.DeadInsts.insert(&LI);
2262 deleteIfTriviallyDead(OldOp);
2263 DEBUG(dbgs() << " to: " << *V << "\n");
2264 return !LI.isVolatile() && !IsPtrAdjusted;
2267 bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp) {
2268 if (V->getType() != VecTy) {
2269 unsigned BeginIndex = getIndex(NewBeginOffset);
2270 unsigned EndIndex = getIndex(NewEndOffset);
2271 assert(EndIndex > BeginIndex && "Empty vector!");
2272 unsigned NumElements = EndIndex - BeginIndex;
2273 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2275 (NumElements == 1) ? ElementTy
2276 : VectorType::get(ElementTy, NumElements);
2277 if (V->getType() != SliceTy)
2278 V = convertValue(DL, IRB, V, SliceTy);
2280 // Mix in the existing elements.
2281 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2283 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2285 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2286 Pass.DeadInsts.insert(&SI);
2289 DEBUG(dbgs() << " to: " << *Store << "\n");
2293 bool rewriteIntegerStore(Value *V, StoreInst &SI) {
2294 assert(IntTy && "We cannot extract an integer from the alloca");
2295 assert(!SI.isVolatile());
2296 if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2297 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2299 Old = convertValue(DL, IRB, Old, IntTy);
2300 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2301 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2302 V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset,
2305 V = convertValue(DL, IRB, V, NewAllocaTy);
2306 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2307 Pass.DeadInsts.insert(&SI);
2309 DEBUG(dbgs() << " to: " << *Store << "\n");
2313 bool visitStoreInst(StoreInst &SI) {
2314 DEBUG(dbgs() << " original: " << SI << "\n");
2315 Value *OldOp = SI.getOperand(1);
2316 assert(OldOp == OldPtr);
2318 Value *V = SI.getValueOperand();
2320 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2321 // alloca that should be re-examined after promoting this alloca.
2322 if (V->getType()->isPointerTy())
2323 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2324 Pass.PostPromotionWorklist.insert(AI);
2326 if (SliceSize < DL.getTypeStoreSize(V->getType())) {
2327 assert(!SI.isVolatile());
2328 assert(V->getType()->isIntegerTy() &&
2329 "Only integer type loads and stores are split");
2330 assert(V->getType()->getIntegerBitWidth() ==
2331 DL.getTypeStoreSizeInBits(V->getType()) &&
2332 "Non-byte-multiple bit width");
2333 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), SliceSize * 8);
2334 V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset,
2339 return rewriteVectorizedStoreInst(V, SI, OldOp);
2340 if (IntTy && V->getType()->isIntegerTy())
2341 return rewriteIntegerStore(V, SI);
2344 if (NewBeginOffset == NewAllocaBeginOffset &&
2345 NewEndOffset == NewAllocaEndOffset &&
2346 canConvertValue(DL, V->getType(), NewAllocaTy)) {
2347 V = convertValue(DL, IRB, V, NewAllocaTy);
2348 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2351 Value *NewPtr = getNewAllocaSlicePtr(IRB, V->getType()->getPointerTo());
2352 NewSI = IRB.CreateAlignedStore(V, NewPtr, getSliceAlign(V->getType()),
2356 Pass.DeadInsts.insert(&SI);
2357 deleteIfTriviallyDead(OldOp);
2359 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2360 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2363 /// \brief Compute an integer value from splatting an i8 across the given
2364 /// number of bytes.
2366 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2367 /// call this routine.
2368 /// FIXME: Heed the advice above.
2370 /// \param V The i8 value to splat.
2371 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2372 Value *getIntegerSplat(Value *V, unsigned Size) {
2373 assert(Size > 0 && "Expected a positive number of bytes.");
2374 IntegerType *VTy = cast<IntegerType>(V->getType());
2375 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2379 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size*8);
2380 V = IRB.CreateMul(IRB.CreateZExt(V, SplatIntTy, "zext"),
2381 ConstantExpr::getUDiv(
2382 Constant::getAllOnesValue(SplatIntTy),
2383 ConstantExpr::getZExt(
2384 Constant::getAllOnesValue(V->getType()),
2390 /// \brief Compute a vector splat for a given element value.
2391 Value *getVectorSplat(Value *V, unsigned NumElements) {
2392 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2393 DEBUG(dbgs() << " splat: " << *V << "\n");
2397 bool visitMemSetInst(MemSetInst &II) {
2398 DEBUG(dbgs() << " original: " << II << "\n");
2399 assert(II.getRawDest() == OldPtr);
2401 // If the memset has a variable size, it cannot be split, just adjust the
2402 // pointer to the new alloca.
2403 if (!isa<Constant>(II.getLength())) {
2405 assert(NewBeginOffset == BeginOffset);
2406 II.setDest(getNewAllocaSlicePtr(IRB, OldPtr->getType()));
2407 Type *CstTy = II.getAlignmentCst()->getType();
2408 II.setAlignment(ConstantInt::get(CstTy, getSliceAlign()));
2410 deleteIfTriviallyDead(OldPtr);
2414 // Record this instruction for deletion.
2415 Pass.DeadInsts.insert(&II);
2417 Type *AllocaTy = NewAI.getAllocatedType();
2418 Type *ScalarTy = AllocaTy->getScalarType();
2420 // If this doesn't map cleanly onto the alloca type, and that type isn't
2421 // a single value type, just emit a memset.
2422 if (!VecTy && !IntTy &&
2423 (BeginOffset > NewAllocaBeginOffset ||
2424 EndOffset < NewAllocaEndOffset ||
2425 !AllocaTy->isSingleValueType() ||
2426 !DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy)) ||
2427 DL.getTypeSizeInBits(ScalarTy)%8 != 0)) {
2428 Type *SizeTy = II.getLength()->getType();
2429 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2430 CallInst *New = IRB.CreateMemSet(
2431 getNewAllocaSlicePtr(IRB, OldPtr->getType()), II.getValue(), Size,
2432 getSliceAlign(), II.isVolatile());
2434 DEBUG(dbgs() << " to: " << *New << "\n");
2438 // If we can represent this as a simple value, we have to build the actual
2439 // value to store, which requires expanding the byte present in memset to
2440 // a sensible representation for the alloca type. This is essentially
2441 // splatting the byte to a sufficiently wide integer, splatting it across
2442 // any desired vector width, and bitcasting to the final type.
2446 // If this is a memset of a vectorized alloca, insert it.
2447 assert(ElementTy == ScalarTy);
2449 unsigned BeginIndex = getIndex(NewBeginOffset);
2450 unsigned EndIndex = getIndex(NewEndOffset);
2451 assert(EndIndex > BeginIndex && "Empty vector!");
2452 unsigned NumElements = EndIndex - BeginIndex;
2453 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2456 getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
2457 Splat = convertValue(DL, IRB, Splat, ElementTy);
2458 if (NumElements > 1)
2459 Splat = getVectorSplat(Splat, NumElements);
2461 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2463 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2465 // If this is a memset on an alloca where we can widen stores, insert the
2467 assert(!II.isVolatile());
2469 uint64_t Size = NewEndOffset - NewBeginOffset;
2470 V = getIntegerSplat(II.getValue(), Size);
2472 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2473 EndOffset != NewAllocaBeginOffset)) {
2474 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2476 Old = convertValue(DL, IRB, Old, IntTy);
2477 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2478 V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2480 assert(V->getType() == IntTy &&
2481 "Wrong type for an alloca wide integer!");
2483 V = convertValue(DL, IRB, V, AllocaTy);
2485 // Established these invariants above.
2486 assert(NewBeginOffset == NewAllocaBeginOffset);
2487 assert(NewEndOffset == NewAllocaEndOffset);
2489 V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
2490 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2491 V = getVectorSplat(V, AllocaVecTy->getNumElements());
2493 V = convertValue(DL, IRB, V, AllocaTy);
2496 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2499 DEBUG(dbgs() << " to: " << *New << "\n");
2500 return !II.isVolatile();
2503 bool visitMemTransferInst(MemTransferInst &II) {
2504 // Rewriting of memory transfer instructions can be a bit tricky. We break
2505 // them into two categories: split intrinsics and unsplit intrinsics.
2507 DEBUG(dbgs() << " original: " << II << "\n");
2509 bool IsDest = &II.getRawDestUse() == OldUse;
2510 assert((IsDest && II.getRawDest() == OldPtr) ||
2511 (!IsDest && II.getRawSource() == OldPtr));
2513 unsigned SliceAlign = getSliceAlign();
2515 // For unsplit intrinsics, we simply modify the source and destination
2516 // pointers in place. This isn't just an optimization, it is a matter of
2517 // correctness. With unsplit intrinsics we may be dealing with transfers
2518 // within a single alloca before SROA ran, or with transfers that have
2519 // a variable length. We may also be dealing with memmove instead of
2520 // memcpy, and so simply updating the pointers is the necessary for us to
2521 // update both source and dest of a single call.
2522 if (!IsSplittable) {
2523 Value *AdjustedPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2525 II.setDest(AdjustedPtr);
2527 II.setSource(AdjustedPtr);
2529 if (II.getAlignment() > SliceAlign) {
2530 Type *CstTy = II.getAlignmentCst()->getType();
2532 ConstantInt::get(CstTy, MinAlign(II.getAlignment(), SliceAlign)));
2535 DEBUG(dbgs() << " to: " << II << "\n");
2536 deleteIfTriviallyDead(OldPtr);
2539 // For split transfer intrinsics we have an incredibly useful assurance:
2540 // the source and destination do not reside within the same alloca, and at
2541 // least one of them does not escape. This means that we can replace
2542 // memmove with memcpy, and we don't need to worry about all manner of
2543 // downsides to splitting and transforming the operations.
2545 // If this doesn't map cleanly onto the alloca type, and that type isn't
2546 // a single value type, just emit a memcpy.
2548 = !VecTy && !IntTy && (BeginOffset > NewAllocaBeginOffset ||
2549 EndOffset < NewAllocaEndOffset ||
2550 !NewAI.getAllocatedType()->isSingleValueType());
2552 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2553 // size hasn't been shrunk based on analysis of the viable range, this is
2555 if (EmitMemCpy && &OldAI == &NewAI) {
2556 // Ensure the start lines up.
2557 assert(NewBeginOffset == BeginOffset);
2559 // Rewrite the size as needed.
2560 if (NewEndOffset != EndOffset)
2561 II.setLength(ConstantInt::get(II.getLength()->getType(),
2562 NewEndOffset - NewBeginOffset));
2565 // Record this instruction for deletion.
2566 Pass.DeadInsts.insert(&II);
2568 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2569 // alloca that should be re-examined after rewriting this instruction.
2570 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2572 = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
2573 assert(AI != &OldAI && AI != &NewAI &&
2574 "Splittable transfers cannot reach the same alloca on both ends.");
2575 Pass.Worklist.insert(AI);
2578 Type *OtherPtrTy = OtherPtr->getType();
2579 unsigned OtherAS = OtherPtrTy->getPointerAddressSpace();
2581 // Compute the relative offset for the other pointer within the transfer.
2582 unsigned IntPtrWidth = DL.getPointerSizeInBits(OtherAS);
2583 APInt OtherOffset(IntPtrWidth, NewBeginOffset - BeginOffset);
2584 unsigned OtherAlign = MinAlign(II.getAlignment() ? II.getAlignment() : 1,
2585 OtherOffset.zextOrTrunc(64).getZExtValue());
2588 // Compute the other pointer, folding as much as possible to produce
2589 // a single, simple GEP in most cases.
2590 OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2591 OtherPtr->getName() + ".");
2593 Value *OurPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2594 Type *SizeTy = II.getLength()->getType();
2595 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2597 CallInst *New = IRB.CreateMemCpy(
2598 IsDest ? OurPtr : OtherPtr, IsDest ? OtherPtr : OurPtr, Size,
2599 MinAlign(SliceAlign, OtherAlign), II.isVolatile());
2601 DEBUG(dbgs() << " to: " << *New << "\n");
2605 bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
2606 NewEndOffset == NewAllocaEndOffset;
2607 uint64_t Size = NewEndOffset - NewBeginOffset;
2608 unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
2609 unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
2610 unsigned NumElements = EndIndex - BeginIndex;
2611 IntegerType *SubIntTy
2612 = IntTy ? Type::getIntNTy(IntTy->getContext(), Size*8) : nullptr;
2614 // Reset the other pointer type to match the register type we're going to
2615 // use, but using the address space of the original other pointer.
2616 if (VecTy && !IsWholeAlloca) {
2617 if (NumElements == 1)
2618 OtherPtrTy = VecTy->getElementType();
2620 OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2622 OtherPtrTy = OtherPtrTy->getPointerTo(OtherAS);
2623 } else if (IntTy && !IsWholeAlloca) {
2624 OtherPtrTy = SubIntTy->getPointerTo(OtherAS);
2626 OtherPtrTy = NewAllocaTy->getPointerTo(OtherAS);
2629 Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2630 OtherPtr->getName() + ".");
2631 unsigned SrcAlign = OtherAlign;
2632 Value *DstPtr = &NewAI;
2633 unsigned DstAlign = SliceAlign;
2635 std::swap(SrcPtr, DstPtr);
2636 std::swap(SrcAlign, DstAlign);
2640 if (VecTy && !IsWholeAlloca && !IsDest) {
2641 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2643 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
2644 } else if (IntTy && !IsWholeAlloca && !IsDest) {
2645 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2647 Src = convertValue(DL, IRB, Src, IntTy);
2648 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2649 Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
2651 Src = IRB.CreateAlignedLoad(SrcPtr, SrcAlign, II.isVolatile(),
2655 if (VecTy && !IsWholeAlloca && IsDest) {
2656 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2658 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
2659 } else if (IntTy && !IsWholeAlloca && IsDest) {
2660 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2662 Old = convertValue(DL, IRB, Old, IntTy);
2663 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2664 Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
2665 Src = convertValue(DL, IRB, Src, NewAllocaTy);
2668 StoreInst *Store = cast<StoreInst>(
2669 IRB.CreateAlignedStore(Src, DstPtr, DstAlign, II.isVolatile()));
2671 DEBUG(dbgs() << " to: " << *Store << "\n");
2672 return !II.isVolatile();
2675 bool visitIntrinsicInst(IntrinsicInst &II) {
2676 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2677 II.getIntrinsicID() == Intrinsic::lifetime_end);
2678 DEBUG(dbgs() << " original: " << II << "\n");
2679 assert(II.getArgOperand(1) == OldPtr);
2681 // Record this instruction for deletion.
2682 Pass.DeadInsts.insert(&II);
2685 = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2686 NewEndOffset - NewBeginOffset);
2687 Value *Ptr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2689 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2690 New = IRB.CreateLifetimeStart(Ptr, Size);
2692 New = IRB.CreateLifetimeEnd(Ptr, Size);
2695 DEBUG(dbgs() << " to: " << *New << "\n");
2699 bool visitPHINode(PHINode &PN) {
2700 DEBUG(dbgs() << " original: " << PN << "\n");
2701 assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
2702 assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
2704 // We would like to compute a new pointer in only one place, but have it be
2705 // as local as possible to the PHI. To do that, we re-use the location of
2706 // the old pointer, which necessarily must be in the right position to
2707 // dominate the PHI.
2708 IRBuilderTy PtrBuilder(IRB);
2709 PtrBuilder.SetInsertPoint(OldPtr);
2710 PtrBuilder.SetCurrentDebugLocation(OldPtr->getDebugLoc());
2712 Value *NewPtr = getNewAllocaSlicePtr(PtrBuilder, OldPtr->getType());
2713 // Replace the operands which were using the old pointer.
2714 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
2716 DEBUG(dbgs() << " to: " << PN << "\n");
2717 deleteIfTriviallyDead(OldPtr);
2719 // PHIs can't be promoted on their own, but often can be speculated. We
2720 // check the speculation outside of the rewriter so that we see the
2721 // fully-rewritten alloca.
2722 PHIUsers.insert(&PN);
2726 bool visitSelectInst(SelectInst &SI) {
2727 DEBUG(dbgs() << " original: " << SI << "\n");
2728 assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
2729 "Pointer isn't an operand!");
2730 assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
2731 assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
2733 Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2734 // Replace the operands which were using the old pointer.
2735 if (SI.getOperand(1) == OldPtr)
2736 SI.setOperand(1, NewPtr);
2737 if (SI.getOperand(2) == OldPtr)
2738 SI.setOperand(2, NewPtr);
2740 DEBUG(dbgs() << " to: " << SI << "\n");
2741 deleteIfTriviallyDead(OldPtr);
2743 // Selects can't be promoted on their own, but often can be speculated. We
2744 // check the speculation outside of the rewriter so that we see the
2745 // fully-rewritten alloca.
2746 SelectUsers.insert(&SI);
2754 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
2756 /// This pass aggressively rewrites all aggregate loads and stores on
2757 /// a particular pointer (or any pointer derived from it which we can identify)
2758 /// with scalar loads and stores.
2759 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
2760 // Befriend the base class so it can delegate to private visit methods.
2761 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
2763 const DataLayout &DL;
2765 /// Queue of pointer uses to analyze and potentially rewrite.
2766 SmallVector<Use *, 8> Queue;
2768 /// Set to prevent us from cycling with phi nodes and loops.
2769 SmallPtrSet<User *, 8> Visited;
2771 /// The current pointer use being rewritten. This is used to dig up the used
2772 /// value (as opposed to the user).
2776 AggLoadStoreRewriter(const DataLayout &DL) : DL(DL) {}
2778 /// Rewrite loads and stores through a pointer and all pointers derived from
2780 bool rewrite(Instruction &I) {
2781 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
2783 bool Changed = false;
2784 while (!Queue.empty()) {
2785 U = Queue.pop_back_val();
2786 Changed |= visit(cast<Instruction>(U->getUser()));
2792 /// Enqueue all the users of the given instruction for further processing.
2793 /// This uses a set to de-duplicate users.
2794 void enqueueUsers(Instruction &I) {
2795 for (Use &U : I.uses())
2796 if (Visited.insert(U.getUser()))
2797 Queue.push_back(&U);
2800 // Conservative default is to not rewrite anything.
2801 bool visitInstruction(Instruction &I) { return false; }
2803 /// \brief Generic recursive split emission class.
2804 template <typename Derived>
2807 /// The builder used to form new instructions.
2809 /// The indices which to be used with insert- or extractvalue to select the
2810 /// appropriate value within the aggregate.
2811 SmallVector<unsigned, 4> Indices;
2812 /// The indices to a GEP instruction which will move Ptr to the correct slot
2813 /// within the aggregate.
2814 SmallVector<Value *, 4> GEPIndices;
2815 /// The base pointer of the original op, used as a base for GEPing the
2816 /// split operations.
2819 /// Initialize the splitter with an insertion point, Ptr and start with a
2820 /// single zero GEP index.
2821 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
2822 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
2825 /// \brief Generic recursive split emission routine.
2827 /// This method recursively splits an aggregate op (load or store) into
2828 /// scalar or vector ops. It splits recursively until it hits a single value
2829 /// and emits that single value operation via the template argument.
2831 /// The logic of this routine relies on GEPs and insertvalue and
2832 /// extractvalue all operating with the same fundamental index list, merely
2833 /// formatted differently (GEPs need actual values).
2835 /// \param Ty The type being split recursively into smaller ops.
2836 /// \param Agg The aggregate value being built up or stored, depending on
2837 /// whether this is splitting a load or a store respectively.
2838 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
2839 if (Ty->isSingleValueType())
2840 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
2842 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2843 unsigned OldSize = Indices.size();
2845 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
2847 assert(Indices.size() == OldSize && "Did not return to the old size");
2848 Indices.push_back(Idx);
2849 GEPIndices.push_back(IRB.getInt32(Idx));
2850 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
2851 GEPIndices.pop_back();
2857 if (StructType *STy = dyn_cast<StructType>(Ty)) {
2858 unsigned OldSize = Indices.size();
2860 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
2862 assert(Indices.size() == OldSize && "Did not return to the old size");
2863 Indices.push_back(Idx);
2864 GEPIndices.push_back(IRB.getInt32(Idx));
2865 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
2866 GEPIndices.pop_back();
2872 llvm_unreachable("Only arrays and structs are aggregate loadable types");
2876 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
2877 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2878 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
2880 /// Emit a leaf load of a single value. This is called at the leaves of the
2881 /// recursive emission to actually load values.
2882 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2883 assert(Ty->isSingleValueType());
2884 // Load the single value and insert it using the indices.
2885 Value *GEP = IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep");
2886 Value *Load = IRB.CreateLoad(GEP, Name + ".load");
2887 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
2888 DEBUG(dbgs() << " to: " << *Load << "\n");
2892 bool visitLoadInst(LoadInst &LI) {
2893 assert(LI.getPointerOperand() == *U);
2894 if (!LI.isSimple() || LI.getType()->isSingleValueType())
2897 // We have an aggregate being loaded, split it apart.
2898 DEBUG(dbgs() << " original: " << LI << "\n");
2899 LoadOpSplitter Splitter(&LI, *U);
2900 Value *V = UndefValue::get(LI.getType());
2901 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
2902 LI.replaceAllUsesWith(V);
2903 LI.eraseFromParent();
2907 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
2908 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2909 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
2911 /// Emit a leaf store of a single value. This is called at the leaves of the
2912 /// recursive emission to actually produce stores.
2913 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2914 assert(Ty->isSingleValueType());
2915 // Extract the single value and store it using the indices.
2916 Value *Store = IRB.CreateStore(
2917 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
2918 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
2920 DEBUG(dbgs() << " to: " << *Store << "\n");
2924 bool visitStoreInst(StoreInst &SI) {
2925 if (!SI.isSimple() || SI.getPointerOperand() != *U)
2927 Value *V = SI.getValueOperand();
2928 if (V->getType()->isSingleValueType())
2931 // We have an aggregate being stored, split it apart.
2932 DEBUG(dbgs() << " original: " << SI << "\n");
2933 StoreOpSplitter Splitter(&SI, *U);
2934 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
2935 SI.eraseFromParent();
2939 bool visitBitCastInst(BitCastInst &BC) {
2944 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
2949 bool visitPHINode(PHINode &PN) {
2954 bool visitSelectInst(SelectInst &SI) {
2961 /// \brief Strip aggregate type wrapping.
2963 /// This removes no-op aggregate types wrapping an underlying type. It will
2964 /// strip as many layers of types as it can without changing either the type
2965 /// size or the allocated size.
2966 static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
2967 if (Ty->isSingleValueType())
2970 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
2971 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
2974 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
2975 InnerTy = ArrTy->getElementType();
2976 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
2977 const StructLayout *SL = DL.getStructLayout(STy);
2978 unsigned Index = SL->getElementContainingOffset(0);
2979 InnerTy = STy->getElementType(Index);
2984 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
2985 TypeSize > DL.getTypeSizeInBits(InnerTy))
2988 return stripAggregateTypeWrapping(DL, InnerTy);
2991 /// \brief Try to find a partition of the aggregate type passed in for a given
2992 /// offset and size.
2994 /// This recurses through the aggregate type and tries to compute a subtype
2995 /// based on the offset and size. When the offset and size span a sub-section
2996 /// of an array, it will even compute a new array type for that sub-section,
2997 /// and the same for structs.
2999 /// Note that this routine is very strict and tries to find a partition of the
3000 /// type which produces the *exact* right offset and size. It is not forgiving
3001 /// when the size or offset cause either end of type-based partition to be off.
3002 /// Also, this is a best-effort routine. It is reasonable to give up and not
3003 /// return a type if necessary.
3004 static Type *getTypePartition(const DataLayout &DL, Type *Ty,
3005 uint64_t Offset, uint64_t Size) {
3006 if (Offset == 0 && DL.getTypeAllocSize(Ty) == Size)
3007 return stripAggregateTypeWrapping(DL, Ty);
3008 if (Offset > DL.getTypeAllocSize(Ty) ||
3009 (DL.getTypeAllocSize(Ty) - Offset) < Size)
3012 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3013 // We can't partition pointers...
3014 if (SeqTy->isPointerTy())
3017 Type *ElementTy = SeqTy->getElementType();
3018 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3019 uint64_t NumSkippedElements = Offset / ElementSize;
3020 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy)) {
3021 if (NumSkippedElements >= ArrTy->getNumElements())
3023 } else if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy)) {
3024 if (NumSkippedElements >= VecTy->getNumElements())
3027 Offset -= NumSkippedElements * ElementSize;
3029 // First check if we need to recurse.
3030 if (Offset > 0 || Size < ElementSize) {
3031 // Bail if the partition ends in a different array element.
3032 if ((Offset + Size) > ElementSize)
3034 // Recurse through the element type trying to peel off offset bytes.
3035 return getTypePartition(DL, ElementTy, Offset, Size);
3037 assert(Offset == 0);
3039 if (Size == ElementSize)
3040 return stripAggregateTypeWrapping(DL, ElementTy);
3041 assert(Size > ElementSize);
3042 uint64_t NumElements = Size / ElementSize;
3043 if (NumElements * ElementSize != Size)
3045 return ArrayType::get(ElementTy, NumElements);
3048 StructType *STy = dyn_cast<StructType>(Ty);
3052 const StructLayout *SL = DL.getStructLayout(STy);
3053 if (Offset >= SL->getSizeInBytes())
3055 uint64_t EndOffset = Offset + Size;
3056 if (EndOffset > SL->getSizeInBytes())
3059 unsigned Index = SL->getElementContainingOffset(Offset);
3060 Offset -= SL->getElementOffset(Index);
3062 Type *ElementTy = STy->getElementType(Index);
3063 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3064 if (Offset >= ElementSize)
3065 return nullptr; // The offset points into alignment padding.
3067 // See if any partition must be contained by the element.
3068 if (Offset > 0 || Size < ElementSize) {
3069 if ((Offset + Size) > ElementSize)
3071 return getTypePartition(DL, ElementTy, Offset, Size);
3073 assert(Offset == 0);
3075 if (Size == ElementSize)
3076 return stripAggregateTypeWrapping(DL, ElementTy);
3078 StructType::element_iterator EI = STy->element_begin() + Index,
3079 EE = STy->element_end();
3080 if (EndOffset < SL->getSizeInBytes()) {
3081 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3082 if (Index == EndIndex)
3083 return nullptr; // Within a single element and its padding.
3085 // Don't try to form "natural" types if the elements don't line up with the
3087 // FIXME: We could potentially recurse down through the last element in the
3088 // sub-struct to find a natural end point.
3089 if (SL->getElementOffset(EndIndex) != EndOffset)
3092 assert(Index < EndIndex);
3093 EE = STy->element_begin() + EndIndex;
3096 // Try to build up a sub-structure.
3097 StructType *SubTy = StructType::get(STy->getContext(), makeArrayRef(EI, EE),
3099 const StructLayout *SubSL = DL.getStructLayout(SubTy);
3100 if (Size != SubSL->getSizeInBytes())
3101 return nullptr; // The sub-struct doesn't have quite the size needed.
3106 /// \brief Rewrite an alloca partition's users.
3108 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3109 /// to rewrite uses of an alloca partition to be conducive for SSA value
3110 /// promotion. If the partition needs a new, more refined alloca, this will
3111 /// build that new alloca, preserving as much type information as possible, and
3112 /// rewrite the uses of the old alloca to point at the new one and have the
3113 /// appropriate new offsets. It also evaluates how successful the rewrite was
3114 /// at enabling promotion and if it was successful queues the alloca to be
3116 bool SROA::rewritePartition(AllocaInst &AI, AllocaSlices &S,
3117 AllocaSlices::iterator B, AllocaSlices::iterator E,
3118 int64_t BeginOffset, int64_t EndOffset,
3119 ArrayRef<AllocaSlices::iterator> SplitUses) {
3120 assert(BeginOffset < EndOffset);
3121 uint64_t SliceSize = EndOffset - BeginOffset;
3123 // Try to compute a friendly type for this partition of the alloca. This
3124 // won't always succeed, in which case we fall back to a legal integer type
3125 // or an i8 array of an appropriate size.
3126 Type *SliceTy = nullptr;
3127 if (Type *CommonUseTy = findCommonType(B, E, EndOffset))
3128 if (DL->getTypeAllocSize(CommonUseTy) >= SliceSize)
3129 SliceTy = CommonUseTy;
3131 if (Type *TypePartitionTy = getTypePartition(*DL, AI.getAllocatedType(),
3132 BeginOffset, SliceSize))
3133 SliceTy = TypePartitionTy;
3134 if ((!SliceTy || (SliceTy->isArrayTy() &&
3135 SliceTy->getArrayElementType()->isIntegerTy())) &&
3136 DL->isLegalInteger(SliceSize * 8))
3137 SliceTy = Type::getIntNTy(*C, SliceSize * 8);
3139 SliceTy = ArrayType::get(Type::getInt8Ty(*C), SliceSize);
3140 assert(DL->getTypeAllocSize(SliceTy) >= SliceSize);
3142 bool IsVectorPromotable = isVectorPromotionViable(
3143 *DL, SliceTy, S, BeginOffset, EndOffset, B, E, SplitUses);
3145 bool IsIntegerPromotable =
3146 !IsVectorPromotable &&
3147 isIntegerWideningViable(*DL, SliceTy, BeginOffset, S, B, E, SplitUses);
3149 // Check for the case where we're going to rewrite to a new alloca of the
3150 // exact same type as the original, and with the same access offsets. In that
3151 // case, re-use the existing alloca, but still run through the rewriter to
3152 // perform phi and select speculation.
3154 if (SliceTy == AI.getAllocatedType()) {
3155 assert(BeginOffset == 0 &&
3156 "Non-zero begin offset but same alloca type");
3158 // FIXME: We should be able to bail at this point with "nothing changed".
3159 // FIXME: We might want to defer PHI speculation until after here.
3161 unsigned Alignment = AI.getAlignment();
3163 // The minimum alignment which users can rely on when the explicit
3164 // alignment is omitted or zero is that required by the ABI for this
3166 Alignment = DL->getABITypeAlignment(AI.getAllocatedType());
3168 Alignment = MinAlign(Alignment, BeginOffset);
3169 // If we will get at least this much alignment from the type alone, leave
3170 // the alloca's alignment unconstrained.
3171 if (Alignment <= DL->getABITypeAlignment(SliceTy))
3173 NewAI = new AllocaInst(SliceTy, nullptr, Alignment,
3174 AI.getName() + ".sroa." + Twine(B - S.begin()), &AI);
3178 DEBUG(dbgs() << "Rewriting alloca partition "
3179 << "[" << BeginOffset << "," << EndOffset << ") to: " << *NewAI
3182 // Track the high watermark on the worklist as it is only relevant for
3183 // promoted allocas. We will reset it to this point if the alloca is not in
3184 // fact scheduled for promotion.
3185 unsigned PPWOldSize = PostPromotionWorklist.size();
3186 unsigned NumUses = 0;
3187 SmallPtrSet<PHINode *, 8> PHIUsers;
3188 SmallPtrSet<SelectInst *, 8> SelectUsers;
3190 AllocaSliceRewriter Rewriter(*DL, S, *this, AI, *NewAI, BeginOffset,
3191 EndOffset, IsVectorPromotable,
3192 IsIntegerPromotable, PHIUsers, SelectUsers);
3193 bool Promotable = true;
3194 for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
3195 SUE = SplitUses.end();
3196 SUI != SUE; ++SUI) {
3197 DEBUG(dbgs() << " rewriting split ");
3198 DEBUG(S.printSlice(dbgs(), *SUI, ""));
3199 Promotable &= Rewriter.visit(*SUI);
3202 for (AllocaSlices::iterator I = B; I != E; ++I) {
3203 DEBUG(dbgs() << " rewriting ");
3204 DEBUG(S.printSlice(dbgs(), I, ""));
3205 Promotable &= Rewriter.visit(I);
3209 NumAllocaPartitionUses += NumUses;
3210 MaxUsesPerAllocaPartition =
3211 std::max<unsigned>(NumUses, MaxUsesPerAllocaPartition);
3213 // Now that we've processed all the slices in the new partition, check if any
3214 // PHIs or Selects would block promotion.
3215 for (SmallPtrSetImpl<PHINode *>::iterator I = PHIUsers.begin(),
3218 if (!isSafePHIToSpeculate(**I, DL)) {
3221 SelectUsers.clear();
3224 for (SmallPtrSetImpl<SelectInst *>::iterator I = SelectUsers.begin(),
3225 E = SelectUsers.end();
3227 if (!isSafeSelectToSpeculate(**I, DL)) {
3230 SelectUsers.clear();
3235 if (PHIUsers.empty() && SelectUsers.empty()) {
3236 // Promote the alloca.
3237 PromotableAllocas.push_back(NewAI);
3239 // If we have either PHIs or Selects to speculate, add them to those
3240 // worklists and re-queue the new alloca so that we promote in on the
3242 for (SmallPtrSetImpl<PHINode *>::iterator I = PHIUsers.begin(),
3245 SpeculatablePHIs.insert(*I);
3246 for (SmallPtrSetImpl<SelectInst *>::iterator I = SelectUsers.begin(),
3247 E = SelectUsers.end();
3249 SpeculatableSelects.insert(*I);
3250 Worklist.insert(NewAI);
3253 // If we can't promote the alloca, iterate on it to check for new
3254 // refinements exposed by splitting the current alloca. Don't iterate on an
3255 // alloca which didn't actually change and didn't get promoted.
3257 Worklist.insert(NewAI);
3259 // Drop any post-promotion work items if promotion didn't happen.
3260 while (PostPromotionWorklist.size() > PPWOldSize)
3261 PostPromotionWorklist.pop_back();
3268 removeFinishedSplitUses(SmallVectorImpl<AllocaSlices::iterator> &SplitUses,
3269 uint64_t &MaxSplitUseEndOffset, uint64_t Offset) {
3270 if (Offset >= MaxSplitUseEndOffset) {
3272 MaxSplitUseEndOffset = 0;
3276 size_t SplitUsesOldSize = SplitUses.size();
3277 SplitUses.erase(std::remove_if(SplitUses.begin(), SplitUses.end(),
3278 [Offset](const AllocaSlices::iterator &I) {
3279 return I->endOffset() <= Offset;
3282 if (SplitUsesOldSize == SplitUses.size())
3285 // Recompute the max. While this is linear, so is remove_if.
3286 MaxSplitUseEndOffset = 0;
3287 for (SmallVectorImpl<AllocaSlices::iterator>::iterator
3288 SUI = SplitUses.begin(),
3289 SUE = SplitUses.end();
3291 MaxSplitUseEndOffset = std::max((*SUI)->endOffset(), MaxSplitUseEndOffset);
3294 /// \brief Walks the slices of an alloca and form partitions based on them,
3295 /// rewriting each of their uses.
3296 bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &S) {
3297 if (S.begin() == S.end())
3300 unsigned NumPartitions = 0;
3301 bool Changed = false;
3302 SmallVector<AllocaSlices::iterator, 4> SplitUses;
3303 uint64_t MaxSplitUseEndOffset = 0;
3305 uint64_t BeginOffset = S.begin()->beginOffset();
3307 for (AllocaSlices::iterator SI = S.begin(), SJ = std::next(SI), SE = S.end();
3308 SI != SE; SI = SJ) {
3309 uint64_t MaxEndOffset = SI->endOffset();
3311 if (!SI->isSplittable()) {
3312 // When we're forming an unsplittable region, it must always start at the
3313 // first slice and will extend through its end.
3314 assert(BeginOffset == SI->beginOffset());
3316 // Form a partition including all of the overlapping slices with this
3317 // unsplittable slice.
3318 while (SJ != SE && SJ->beginOffset() < MaxEndOffset) {
3319 if (!SJ->isSplittable())
3320 MaxEndOffset = std::max(MaxEndOffset, SJ->endOffset());
3324 assert(SI->isSplittable()); // Established above.
3326 // Collect all of the overlapping splittable slices.
3327 while (SJ != SE && SJ->beginOffset() < MaxEndOffset &&
3328 SJ->isSplittable()) {
3329 MaxEndOffset = std::max(MaxEndOffset, SJ->endOffset());
3333 // Back up MaxEndOffset and SJ if we ended the span early when
3334 // encountering an unsplittable slice.
3335 if (SJ != SE && SJ->beginOffset() < MaxEndOffset) {
3336 assert(!SJ->isSplittable());
3337 MaxEndOffset = SJ->beginOffset();
3341 // Check if we have managed to move the end offset forward yet. If so,
3342 // we'll have to rewrite uses and erase old split uses.
3343 if (BeginOffset < MaxEndOffset) {
3344 // Rewrite a sequence of overlapping slices.
3346 rewritePartition(AI, S, SI, SJ, BeginOffset, MaxEndOffset, SplitUses);
3349 removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset, MaxEndOffset);
3352 // Accumulate all the splittable slices from the [SI,SJ) region which
3353 // overlap going forward.
3354 for (AllocaSlices::iterator SK = SI; SK != SJ; ++SK)
3355 if (SK->isSplittable() && SK->endOffset() > MaxEndOffset) {
3356 SplitUses.push_back(SK);
3357 MaxSplitUseEndOffset = std::max(SK->endOffset(), MaxSplitUseEndOffset);
3360 // If we're already at the end and we have no split uses, we're done.
3361 if (SJ == SE && SplitUses.empty())
3364 // If we have no split uses or no gap in offsets, we're ready to move to
3366 if (SplitUses.empty() || (SJ != SE && MaxEndOffset == SJ->beginOffset())) {
3367 BeginOffset = SJ->beginOffset();
3371 // Even if we have split slices, if the next slice is splittable and the
3372 // split slices reach it, we can simply set up the beginning offset of the
3373 // next iteration to bridge between them.
3374 if (SJ != SE && SJ->isSplittable() &&
3375 MaxSplitUseEndOffset > SJ->beginOffset()) {
3376 BeginOffset = MaxEndOffset;
3380 // Otherwise, we have a tail of split slices. Rewrite them with an empty
3382 uint64_t PostSplitEndOffset =
3383 SJ == SE ? MaxSplitUseEndOffset : SJ->beginOffset();
3385 Changed |= rewritePartition(AI, S, SJ, SJ, MaxEndOffset, PostSplitEndOffset,
3390 break; // Skip the rest, we don't need to do any cleanup.
3392 removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset,
3393 PostSplitEndOffset);
3395 // Now just reset the begin offset for the next iteration.
3396 BeginOffset = SJ->beginOffset();
3399 NumAllocaPartitions += NumPartitions;
3400 MaxPartitionsPerAlloca =
3401 std::max<unsigned>(NumPartitions, MaxPartitionsPerAlloca);
3406 /// \brief Clobber a use with undef, deleting the used value if it becomes dead.
3407 void SROA::clobberUse(Use &U) {
3409 // Replace the use with an undef value.
3410 U = UndefValue::get(OldV->getType());
3412 // Check for this making an instruction dead. We have to garbage collect
3413 // all the dead instructions to ensure the uses of any alloca end up being
3415 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3416 if (isInstructionTriviallyDead(OldI)) {
3417 DeadInsts.insert(OldI);
3421 /// \brief Analyze an alloca for SROA.
3423 /// This analyzes the alloca to ensure we can reason about it, builds
3424 /// the slices of the alloca, and then hands it off to be split and
3425 /// rewritten as needed.
3426 bool SROA::runOnAlloca(AllocaInst &AI) {
3427 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3428 ++NumAllocasAnalyzed;
3430 // Special case dead allocas, as they're trivial.
3431 if (AI.use_empty()) {
3432 AI.eraseFromParent();
3436 // Skip alloca forms that this analysis can't handle.
3437 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3438 DL->getTypeAllocSize(AI.getAllocatedType()) == 0)
3441 bool Changed = false;
3443 // First, split any FCA loads and stores touching this alloca to promote
3444 // better splitting and promotion opportunities.
3445 AggLoadStoreRewriter AggRewriter(*DL);
3446 Changed |= AggRewriter.rewrite(AI);
3448 // Build the slices using a recursive instruction-visiting builder.
3449 AllocaSlices S(*DL, AI);
3450 DEBUG(S.print(dbgs()));
3454 // Delete all the dead users of this alloca before splitting and rewriting it.
3455 for (AllocaSlices::dead_user_iterator DI = S.dead_user_begin(),
3456 DE = S.dead_user_end();
3458 // Free up everything used by this instruction.
3459 for (Use &DeadOp : (*DI)->operands())
3462 // Now replace the uses of this instruction.
3463 (*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType()));
3465 // And mark it for deletion.
3466 DeadInsts.insert(*DI);
3469 for (AllocaSlices::dead_op_iterator DO = S.dead_op_begin(),
3470 DE = S.dead_op_end();
3476 // No slices to split. Leave the dead alloca for a later pass to clean up.
3477 if (S.begin() == S.end())
3480 Changed |= splitAlloca(AI, S);
3482 DEBUG(dbgs() << " Speculating PHIs\n");
3483 while (!SpeculatablePHIs.empty())
3484 speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
3486 DEBUG(dbgs() << " Speculating Selects\n");
3487 while (!SpeculatableSelects.empty())
3488 speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
3493 /// \brief Delete the dead instructions accumulated in this run.
3495 /// Recursively deletes the dead instructions we've accumulated. This is done
3496 /// at the very end to maximize locality of the recursive delete and to
3497 /// minimize the problems of invalidated instruction pointers as such pointers
3498 /// are used heavily in the intermediate stages of the algorithm.
3500 /// We also record the alloca instructions deleted here so that they aren't
3501 /// subsequently handed to mem2reg to promote.
3502 void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
3503 while (!DeadInsts.empty()) {
3504 Instruction *I = DeadInsts.pop_back_val();
3505 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3507 I->replaceAllUsesWith(UndefValue::get(I->getType()));
3509 for (Use &Operand : I->operands())
3510 if (Instruction *U = dyn_cast<Instruction>(Operand)) {
3511 // Zero out the operand and see if it becomes trivially dead.
3513 if (isInstructionTriviallyDead(U))
3514 DeadInsts.insert(U);
3517 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3518 DeletedAllocas.insert(AI);
3521 I->eraseFromParent();
3525 static void enqueueUsersInWorklist(Instruction &I,
3526 SmallVectorImpl<Instruction *> &Worklist,
3527 SmallPtrSet<Instruction *, 8> &Visited) {
3528 for (User *U : I.users())
3529 if (Visited.insert(cast<Instruction>(U)))
3530 Worklist.push_back(cast<Instruction>(U));
3533 /// \brief Promote the allocas, using the best available technique.
3535 /// This attempts to promote whatever allocas have been identified as viable in
3536 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
3537 /// If there is a domtree available, we attempt to promote using the full power
3538 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3539 /// based on the SSAUpdater utilities. This function returns whether any
3540 /// promotion occurred.
3541 bool SROA::promoteAllocas(Function &F) {
3542 if (PromotableAllocas.empty())
3545 NumPromoted += PromotableAllocas.size();
3547 if (DT && !ForceSSAUpdater) {
3548 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3549 PromoteMemToReg(PromotableAllocas, *DT);
3550 PromotableAllocas.clear();
3554 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3556 DIBuilder DIB(*F.getParent());
3557 SmallVector<Instruction *, 64> Insts;
3559 // We need a worklist to walk the uses of each alloca.
3560 SmallVector<Instruction *, 8> Worklist;
3561 SmallPtrSet<Instruction *, 8> Visited;
3562 SmallVector<Instruction *, 32> DeadInsts;
3564 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3565 AllocaInst *AI = PromotableAllocas[Idx];
3570 enqueueUsersInWorklist(*AI, Worklist, Visited);
3572 while (!Worklist.empty()) {
3573 Instruction *I = Worklist.pop_back_val();
3575 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3576 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3577 // leading to them) here. Eventually it should use them to optimize the
3578 // scalar values produced.
3579 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3580 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3581 II->getIntrinsicID() == Intrinsic::lifetime_end);
3582 II->eraseFromParent();
3586 // Push the loads and stores we find onto the list. SROA will already
3587 // have validated that all loads and stores are viable candidates for
3589 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
3590 assert(LI->getType() == AI->getAllocatedType());
3591 Insts.push_back(LI);
3594 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
3595 assert(SI->getValueOperand()->getType() == AI->getAllocatedType());
3596 Insts.push_back(SI);
3600 // For everything else, we know that only no-op bitcasts and GEPs will
3601 // make it this far, just recurse through them and recall them for later
3603 DeadInsts.push_back(I);
3604 enqueueUsersInWorklist(*I, Worklist, Visited);
3606 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3607 while (!DeadInsts.empty())
3608 DeadInsts.pop_back_val()->eraseFromParent();
3609 AI->eraseFromParent();
3612 PromotableAllocas.clear();
3616 bool SROA::runOnFunction(Function &F) {
3617 if (skipOptnoneFunction(F))
3620 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3621 C = &F.getContext();
3622 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
3624 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3627 DL = &DLP->getDataLayout();
3628 DominatorTreeWrapperPass *DTWP =
3629 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
3630 DT = DTWP ? &DTWP->getDomTree() : nullptr;
3632 BasicBlock &EntryBB = F.getEntryBlock();
3633 for (BasicBlock::iterator I = EntryBB.begin(), E = std::prev(EntryBB.end());
3635 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3636 Worklist.insert(AI);
3638 bool Changed = false;
3639 // A set of deleted alloca instruction pointers which should be removed from
3640 // the list of promotable allocas.
3641 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3644 while (!Worklist.empty()) {
3645 Changed |= runOnAlloca(*Worklist.pop_back_val());
3646 deleteDeadInstructions(DeletedAllocas);
3648 // Remove the deleted allocas from various lists so that we don't try to
3649 // continue processing them.
3650 if (!DeletedAllocas.empty()) {
3651 auto IsInSet = [&](AllocaInst *AI) {
3652 return DeletedAllocas.count(AI);
3654 Worklist.remove_if(IsInSet);
3655 PostPromotionWorklist.remove_if(IsInSet);
3656 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3657 PromotableAllocas.end(),
3659 PromotableAllocas.end());
3660 DeletedAllocas.clear();
3664 Changed |= promoteAllocas(F);
3666 Worklist = PostPromotionWorklist;
3667 PostPromotionWorklist.clear();
3668 } while (!Worklist.empty());
3673 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3674 if (RequiresDomTree)
3675 AU.addRequired<DominatorTreeWrapperPass>();
3676 AU.setPreservesCFG();