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 #define DEBUG_TYPE "sroa"
27 #include "llvm/Transforms/Scalar.h"
28 #include "llvm/ADT/STLExtras.h"
29 #include "llvm/ADT/SetVector.h"
30 #include "llvm/ADT/SmallVector.h"
31 #include "llvm/ADT/Statistic.h"
32 #include "llvm/Analysis/Dominators.h"
33 #include "llvm/Analysis/Loads.h"
34 #include "llvm/Analysis/PtrUseVisitor.h"
35 #include "llvm/Analysis/ValueTracking.h"
36 #include "llvm/DIBuilder.h"
37 #include "llvm/DebugInfo.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/IR/DerivedTypes.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/IRBuilder.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/InstVisitor.h"
48 #include "llvm/Pass.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Compiler.h"
51 #include "llvm/Support/Debug.h"
52 #include "llvm/Support/ErrorHandling.h"
53 #include "llvm/Support/MathExtras.h"
54 #include "llvm/Support/raw_ostream.h"
55 #include "llvm/Transforms/Utils/Local.h"
56 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
57 #include "llvm/Transforms/Utils/SSAUpdater.h"
60 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
61 STATISTIC(NumAllocaPartitions, "Number of alloca partitions formed");
62 STATISTIC(MaxPartitionsPerAlloca, "Maximum number of partitions");
63 STATISTIC(NumAllocaPartitionUses, "Number of alloca partition uses found");
64 STATISTIC(MaxPartitionUsesPerAlloca, "Maximum number of partition uses");
65 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
66 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
67 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
68 STATISTIC(NumDeleted, "Number of instructions deleted");
69 STATISTIC(NumVectorized, "Number of vectorized aggregates");
71 /// Hidden option to force the pass to not use DomTree and mem2reg, instead
72 /// forming SSA values through the SSAUpdater infrastructure.
74 ForceSSAUpdater("force-ssa-updater", cl::init(false), cl::Hidden);
77 /// \brief A custom IRBuilder inserter which prefixes all names if they are
79 template <bool preserveNames = true>
80 class IRBuilderPrefixedInserter :
81 public IRBuilderDefaultInserter<preserveNames> {
85 void SetNamePrefix(const Twine &P) { Prefix = P.str(); }
88 void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
89 BasicBlock::iterator InsertPt) const {
90 IRBuilderDefaultInserter<preserveNames>::InsertHelper(
91 I, Name.isTriviallyEmpty() ? Name : Prefix + Name, BB, InsertPt);
95 // Specialization for not preserving the name is trivial.
97 class IRBuilderPrefixedInserter<false> :
98 public IRBuilderDefaultInserter<false> {
100 void SetNamePrefix(const Twine &P) {}
103 /// \brief Provide a typedef for IRBuilder that drops names in release builds.
105 typedef llvm::IRBuilder<true, ConstantFolder,
106 IRBuilderPrefixedInserter<true> > IRBuilderTy;
108 typedef llvm::IRBuilder<false, ConstantFolder,
109 IRBuilderPrefixedInserter<false> > IRBuilderTy;
114 /// \brief A partition of an alloca.
116 /// This structure represents a contiguous partition of the alloca. These are
117 /// formed by examining the uses of the alloca. During formation, they may
118 /// overlap but once an AllocaPartitioning is built, the Partitions within it
119 /// are all disjoint. The partition also contains a chain of uses of that
122 /// \brief The beginning offset of the range.
123 uint64_t BeginOffset;
125 /// \brief The ending offset, not included in the range.
128 /// \brief Storage for both the use of this partition and whether it can be
130 PointerIntPair<Use *, 1, bool> PartitionUseAndIsSplittable;
133 Partition() : BeginOffset(), EndOffset() {}
134 Partition(uint64_t BeginOffset, uint64_t EndOffset, Use *U, bool IsSplittable)
135 : BeginOffset(BeginOffset), EndOffset(EndOffset),
136 PartitionUseAndIsSplittable(U, IsSplittable) {}
138 uint64_t beginOffset() const { return BeginOffset; }
139 uint64_t endOffset() const { return EndOffset; }
141 bool isSplittable() const { return PartitionUseAndIsSplittable.getInt(); }
142 void makeUnsplittable() { PartitionUseAndIsSplittable.setInt(false); }
144 Use *getUse() const { return PartitionUseAndIsSplittable.getPointer(); }
146 bool isDead() const { return getUse() == 0; }
147 void kill() { PartitionUseAndIsSplittable.setPointer(0); }
149 /// \brief Support for ordering ranges.
151 /// This provides an ordering over ranges such that start offsets are
152 /// always increasing, and within equal start offsets, the end offsets are
153 /// decreasing. Thus the spanning range comes first in a cluster with the
154 /// same start position.
155 bool operator<(const Partition &RHS) const {
156 if (beginOffset() < RHS.beginOffset()) return true;
157 if (beginOffset() > RHS.beginOffset()) return false;
158 if (isSplittable() != RHS.isSplittable()) return !isSplittable();
159 if (endOffset() > RHS.endOffset()) return true;
163 /// \brief Support comparison with a single offset to allow binary searches.
164 friend LLVM_ATTRIBUTE_UNUSED bool operator<(const Partition &LHS,
165 uint64_t RHSOffset) {
166 return LHS.beginOffset() < RHSOffset;
168 friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
169 const Partition &RHS) {
170 return LHSOffset < RHS.beginOffset();
173 bool operator==(const Partition &RHS) const {
174 return isSplittable() == RHS.isSplittable() &&
175 beginOffset() == RHS.beginOffset() && endOffset() == RHS.endOffset();
177 bool operator!=(const Partition &RHS) const { return !operator==(RHS); }
179 } // end anonymous namespace
182 template <typename T> struct isPodLike;
183 template <> struct isPodLike<Partition> {
184 static const bool value = true;
189 /// \brief Alloca partitioning representation.
191 /// This class represents a partitioning of an alloca into slices, and
192 /// information about the nature of uses of each slice of the alloca. The goal
193 /// is that this information is sufficient to decide if and how to split the
194 /// alloca apart and replace slices with scalars. It is also intended that this
195 /// structure can capture the relevant information needed both to decide about
196 /// and to enact these transformations.
197 class AllocaPartitioning {
199 /// \brief Construct a partitioning of a particular alloca.
201 /// Construction does most of the work for partitioning the alloca. This
202 /// performs the necessary walks of users and builds a partitioning from it.
203 AllocaPartitioning(const DataLayout &DL, AllocaInst &AI);
205 /// \brief Test whether a pointer to the allocation escapes our analysis.
207 /// If this is true, the partitioning is never fully built and should be
209 bool isEscaped() const { return PointerEscapingInstr; }
211 /// \brief Support for iterating over the partitions.
213 typedef SmallVectorImpl<Partition>::iterator iterator;
214 iterator begin() { return Partitions.begin(); }
215 iterator end() { return Partitions.end(); }
217 typedef SmallVectorImpl<Partition>::const_iterator const_iterator;
218 const_iterator begin() const { return Partitions.begin(); }
219 const_iterator end() const { return Partitions.end(); }
222 /// \brief Allow iterating the dead users for this alloca.
224 /// These are instructions which will never actually use the alloca as they
225 /// are outside the allocated range. They are safe to replace with undef and
228 typedef SmallVectorImpl<Instruction *>::const_iterator dead_user_iterator;
229 dead_user_iterator dead_user_begin() const { return DeadUsers.begin(); }
230 dead_user_iterator dead_user_end() const { return DeadUsers.end(); }
233 /// \brief Allow iterating the dead expressions referring to this alloca.
235 /// These are operands which have cannot actually be used to refer to the
236 /// alloca as they are outside its range and the user doesn't correct for
237 /// that. These mostly consist of PHI node inputs and the like which we just
238 /// need to replace with undef.
240 typedef SmallVectorImpl<Use *>::const_iterator dead_op_iterator;
241 dead_op_iterator dead_op_begin() const { return DeadOperands.begin(); }
242 dead_op_iterator dead_op_end() const { return DeadOperands.end(); }
245 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
246 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
247 void printPartition(raw_ostream &OS, const_iterator I,
248 StringRef Indent = " ") const;
249 void printUse(raw_ostream &OS, const_iterator I,
250 StringRef Indent = " ") const;
251 void print(raw_ostream &OS) const;
252 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump(const_iterator I) const;
253 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump() const;
257 template <typename DerivedT, typename RetT = void> class BuilderBase;
258 class PartitionBuilder;
259 friend class AllocaPartitioning::PartitionBuilder;
261 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
262 /// \brief Handle to alloca instruction to simplify method interfaces.
266 /// \brief The instruction responsible for this alloca having no partitioning.
268 /// When an instruction (potentially) escapes the pointer to the alloca, we
269 /// store a pointer to that here and abort trying to partition the alloca.
270 /// This will be null if the alloca is partitioned successfully.
271 Instruction *PointerEscapingInstr;
273 /// \brief The partitions of the alloca.
275 /// We store a vector of the partitions over the alloca here. This vector is
276 /// sorted by increasing begin offset, and then by decreasing end offset. See
277 /// the Partition inner class for more details. Initially (during
278 /// construction) there are overlaps, but we form a disjoint sequence of
279 /// partitions while finishing construction and a fully constructed object is
280 /// expected to always have this as a disjoint space.
281 SmallVector<Partition, 8> Partitions;
283 /// \brief Instructions which will become dead if we rewrite the alloca.
285 /// Note that these are not separated by partition. This is because we expect
286 /// a partitioned alloca to be completely rewritten or not rewritten at all.
287 /// If rewritten, all these instructions can simply be removed and replaced
288 /// with undef as they come from outside of the allocated space.
289 SmallVector<Instruction *, 8> DeadUsers;
291 /// \brief Operands which will become dead if we rewrite the alloca.
293 /// These are operands that in their particular use can be replaced with
294 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
295 /// to PHI nodes and the like. They aren't entirely dead (there might be
296 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
297 /// want to swap this particular input for undef to simplify the use lists of
299 SmallVector<Use *, 8> DeadOperands;
303 static Value *foldSelectInst(SelectInst &SI) {
304 // If the condition being selected on is a constant or the same value is
305 // being selected between, fold the select. Yes this does (rarely) happen
307 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
308 return SI.getOperand(1+CI->isZero());
309 if (SI.getOperand(1) == SI.getOperand(2))
310 return SI.getOperand(1);
315 /// \brief Builder for the alloca partitioning.
317 /// This class builds an alloca partitioning by recursively visiting the uses
318 /// of an alloca and splitting the partitions for each load and store at each
320 class AllocaPartitioning::PartitionBuilder
321 : public PtrUseVisitor<PartitionBuilder> {
322 friend class PtrUseVisitor<PartitionBuilder>;
323 friend class InstVisitor<PartitionBuilder>;
324 typedef PtrUseVisitor<PartitionBuilder> Base;
326 const uint64_t AllocSize;
327 AllocaPartitioning &P;
329 SmallDenseMap<Instruction *, unsigned> MemTransferPartitionMap;
330 SmallDenseMap<Instruction *, uint64_t> PHIOrSelectSizes;
332 /// \brief Set to de-duplicate dead instructions found in the use walk.
333 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
336 PartitionBuilder(const DataLayout &DL, AllocaInst &AI, AllocaPartitioning &P)
337 : PtrUseVisitor<PartitionBuilder>(DL),
338 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())),
342 void markAsDead(Instruction &I) {
343 if (VisitedDeadInsts.insert(&I))
344 P.DeadUsers.push_back(&I);
347 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
348 bool IsSplittable = false) {
349 // Completely skip uses which have a zero size or start either before or
350 // past the end of the allocation.
351 if (Size == 0 || Offset.isNegative() || Offset.uge(AllocSize)) {
352 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
353 << " which has zero size or starts outside of the "
354 << AllocSize << " byte alloca:\n"
355 << " alloca: " << P.AI << "\n"
356 << " use: " << I << "\n");
357 return markAsDead(I);
360 uint64_t BeginOffset = Offset.getZExtValue();
361 uint64_t EndOffset = BeginOffset + Size;
363 // Clamp the end offset to the end of the allocation. Note that this is
364 // formulated to handle even the case where "BeginOffset + Size" overflows.
365 // This may appear superficially to be something we could ignore entirely,
366 // but that is not so! There may be widened loads or PHI-node uses where
367 // some instructions are dead but not others. We can't completely ignore
368 // them, and so have to record at least the information here.
369 assert(AllocSize >= BeginOffset); // Established above.
370 if (Size > AllocSize - BeginOffset) {
371 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
372 << " to remain within the " << AllocSize << " byte alloca:\n"
373 << " alloca: " << P.AI << "\n"
374 << " use: " << I << "\n");
375 EndOffset = AllocSize;
378 P.Partitions.push_back(Partition(BeginOffset, EndOffset, U, IsSplittable));
381 void visitBitCastInst(BitCastInst &BC) {
383 return markAsDead(BC);
385 return Base::visitBitCastInst(BC);
388 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
389 if (GEPI.use_empty())
390 return markAsDead(GEPI);
392 return Base::visitGetElementPtrInst(GEPI);
395 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
396 uint64_t Size, bool IsVolatile) {
397 // We allow splitting of loads and stores where the type is an integer type
398 // and cover the entire alloca. This prevents us from splitting over
400 // FIXME: In the great blue eventually, we should eagerly split all integer
401 // loads and stores, and then have a separate step that merges adjacent
402 // alloca partitions into a single partition suitable for integer widening.
403 // Or we should skip the merge step and rely on GVN and other passes to
404 // merge adjacent loads and stores that survive mem2reg.
406 Ty->isIntegerTy() && !IsVolatile && Offset == 0 && Size >= AllocSize;
408 insertUse(I, Offset, Size, IsSplittable);
411 void visitLoadInst(LoadInst &LI) {
412 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
413 "All simple FCA loads should have been pre-split");
416 return PI.setAborted(&LI);
418 uint64_t Size = DL.getTypeStoreSize(LI.getType());
419 return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
422 void visitStoreInst(StoreInst &SI) {
423 Value *ValOp = SI.getValueOperand();
425 return PI.setEscapedAndAborted(&SI);
427 return PI.setAborted(&SI);
429 uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
431 // If this memory access can be shown to *statically* extend outside the
432 // bounds of of the allocation, it's behavior is undefined, so simply
433 // ignore it. Note that this is more strict than the generic clamping
434 // behavior of insertUse. We also try to handle cases which might run the
436 // FIXME: We should instead consider the pointer to have escaped if this
437 // function is being instrumented for addressing bugs or race conditions.
438 if (Offset.isNegative() || Size > AllocSize ||
439 Offset.ugt(AllocSize - Size)) {
440 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @" << Offset
441 << " which extends past the end of the " << AllocSize
443 << " alloca: " << P.AI << "\n"
444 << " use: " << SI << "\n");
445 return markAsDead(SI);
448 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
449 "All simple FCA stores should have been pre-split");
450 handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
454 void visitMemSetInst(MemSetInst &II) {
455 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
456 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
457 if ((Length && Length->getValue() == 0) ||
458 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize)))
459 // Zero-length mem transfer intrinsics can be ignored entirely.
460 return markAsDead(II);
463 return PI.setAborted(&II);
465 insertUse(II, Offset,
466 Length ? Length->getLimitedValue()
467 : AllocSize - Offset.getLimitedValue(),
471 void visitMemTransferInst(MemTransferInst &II) {
472 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
473 if ((Length && Length->getValue() == 0) ||
474 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize)))
475 // Zero-length mem transfer intrinsics can be ignored entirely.
476 return markAsDead(II);
479 return PI.setAborted(&II);
481 uint64_t RawOffset = Offset.getLimitedValue();
482 uint64_t Size = Length ? Length->getLimitedValue()
483 : AllocSize - RawOffset;
485 // Check for the special case where the same exact value is used for both
487 if (*U == II.getRawDest() && *U == II.getRawSource()) {
488 // For non-volatile transfers this is a no-op.
489 if (!II.isVolatile())
490 return markAsDead(II);
492 return insertUse(II, Offset, Size, /*IsSplittable=*/false);;
495 // If we have seen both source and destination for a mem transfer, then
496 // they both point to the same alloca.
498 SmallDenseMap<Instruction *, unsigned>::iterator MTPI;
499 llvm::tie(MTPI, Inserted) =
500 MemTransferPartitionMap.insert(std::make_pair(&II, P.Partitions.size()));
501 unsigned PrevIdx = MTPI->second;
503 Partition &PrevP = P.Partitions[PrevIdx];
505 // Check if the begin offsets match and this is a non-volatile transfer.
506 // In that case, we can completely elide the transfer.
507 if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
509 return markAsDead(II);
512 // Otherwise we have an offset transfer within the same alloca. We can't
514 PrevP.makeUnsplittable();
517 // Insert the use now that we've fixed up the splittable nature.
518 insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
520 // Check that we ended up with a valid index in the map.
521 assert(P.Partitions[PrevIdx].getUse()->getUser() == &II &&
522 "Map index doesn't point back to a partition with this user.");
525 // Disable SRoA for any intrinsics except for lifetime invariants.
526 // FIXME: What about debug intrinsics? This matches old behavior, but
527 // doesn't make sense.
528 void visitIntrinsicInst(IntrinsicInst &II) {
530 return PI.setAborted(&II);
532 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
533 II.getIntrinsicID() == Intrinsic::lifetime_end) {
534 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
535 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
536 Length->getLimitedValue());
537 insertUse(II, Offset, Size, true);
541 Base::visitIntrinsicInst(II);
544 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
545 // We consider any PHI or select that results in a direct load or store of
546 // the same offset to be a viable use for partitioning purposes. These uses
547 // are considered unsplittable and the size is the maximum loaded or stored
549 SmallPtrSet<Instruction *, 4> Visited;
550 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
551 Visited.insert(Root);
552 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
553 // If there are no loads or stores, the access is dead. We mark that as
554 // a size zero access.
557 Instruction *I, *UsedI;
558 llvm::tie(UsedI, I) = Uses.pop_back_val();
560 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
561 Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
564 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
565 Value *Op = SI->getOperand(0);
568 Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
572 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
573 if (!GEP->hasAllZeroIndices())
575 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
576 !isa<SelectInst>(I)) {
580 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); UI != UE;
582 if (Visited.insert(cast<Instruction>(*UI)))
583 Uses.push_back(std::make_pair(I, cast<Instruction>(*UI)));
584 } while (!Uses.empty());
589 void visitPHINode(PHINode &PN) {
591 return markAsDead(PN);
593 return PI.setAborted(&PN);
595 // See if we already have computed info on this node.
596 uint64_t &PHISize = PHIOrSelectSizes[&PN];
598 // This is a new PHI node, check for an unsafe use of the PHI node.
599 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&PN, PHISize))
600 return PI.setAborted(UnsafeI);
603 // For PHI and select operands outside the alloca, we can't nuke the entire
604 // phi or select -- the other side might still be relevant, so we special
605 // case them here and use a separate structure to track the operands
606 // themselves which should be replaced with undef.
607 // FIXME: This should instead be escaped in the event we're instrumenting
608 // for address sanitization.
609 if ((Offset.isNegative() && (-Offset).uge(PHISize)) ||
610 (!Offset.isNegative() && Offset.uge(AllocSize))) {
611 P.DeadOperands.push_back(U);
615 insertUse(PN, Offset, PHISize);
618 void visitSelectInst(SelectInst &SI) {
620 return markAsDead(SI);
621 if (Value *Result = foldSelectInst(SI)) {
623 // If the result of the constant fold will be the pointer, recurse
624 // through the select as if we had RAUW'ed it.
627 // Otherwise the operand to the select is dead, and we can replace it
629 P.DeadOperands.push_back(U);
634 return PI.setAborted(&SI);
636 // See if we already have computed info on this node.
637 uint64_t &SelectSize = PHIOrSelectSizes[&SI];
639 // This is a new Select, check for an unsafe use of it.
640 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&SI, SelectSize))
641 return PI.setAborted(UnsafeI);
644 // For PHI and select operands outside the alloca, we can't nuke the entire
645 // phi or select -- the other side might still be relevant, so we special
646 // case them here and use a separate structure to track the operands
647 // themselves which should be replaced with undef.
648 // FIXME: This should instead be escaped in the event we're instrumenting
649 // for address sanitization.
650 if ((Offset.isNegative() && Offset.uge(SelectSize)) ||
651 (!Offset.isNegative() && Offset.uge(AllocSize))) {
652 P.DeadOperands.push_back(U);
656 insertUse(SI, Offset, SelectSize);
659 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
660 void visitInstruction(Instruction &I) {
666 struct IsPartitionDead {
667 bool operator()(const Partition &P) { return P.isDead(); }
671 AllocaPartitioning::AllocaPartitioning(const DataLayout &DL, AllocaInst &AI)
673 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
676 PointerEscapingInstr(0) {
677 PartitionBuilder PB(DL, AI, *this);
678 PartitionBuilder::PtrInfo PtrI = PB.visitPtr(AI);
679 if (PtrI.isEscaped() || PtrI.isAborted()) {
680 // FIXME: We should sink the escape vs. abort info into the caller nicely,
681 // possibly by just storing the PtrInfo in the AllocaPartitioning.
682 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
683 : PtrI.getAbortingInst();
684 assert(PointerEscapingInstr && "Did not track a bad instruction");
688 // Sort the uses. This arranges for the offsets to be in ascending order,
689 // and the sizes to be in descending order.
690 std::sort(Partitions.begin(), Partitions.end());
693 std::remove_if(Partitions.begin(), Partitions.end(), IsPartitionDead()),
696 // Record how many partitions we end up with.
697 NumAllocaPartitions += Partitions.size();
698 MaxPartitionsPerAlloca = std::max<unsigned>(Partitions.size(), MaxPartitionsPerAlloca);
700 NumAllocaPartitionUses += Partitions.size();
701 MaxPartitionUsesPerAlloca =
702 std::max<unsigned>(Partitions.size(), MaxPartitionUsesPerAlloca);
705 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
707 void AllocaPartitioning::print(raw_ostream &OS, const_iterator I,
708 StringRef Indent) const {
709 printPartition(OS, I, Indent);
710 printUse(OS, I, Indent);
713 void AllocaPartitioning::printPartition(raw_ostream &OS, const_iterator I,
714 StringRef Indent) const {
715 OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
716 << " partition #" << (I - begin())
717 << (I->isSplittable() ? " (splittable)" : "") << "\n";
720 void AllocaPartitioning::printUse(raw_ostream &OS, const_iterator I,
721 StringRef Indent) const {
722 OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
725 void AllocaPartitioning::print(raw_ostream &OS) const {
726 if (PointerEscapingInstr) {
727 OS << "No partitioning for alloca: " << AI << "\n"
728 << " A pointer to this alloca escaped by:\n"
729 << " " << *PointerEscapingInstr << "\n";
733 OS << "Partitioning of alloca: " << AI << "\n";
734 for (const_iterator I = begin(), E = end(); I != E; ++I)
738 void AllocaPartitioning::dump(const_iterator I) const { print(dbgs(), I); }
739 void AllocaPartitioning::dump() const { print(dbgs()); }
741 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
744 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
746 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
747 /// the loads and stores of an alloca instruction, as well as updating its
748 /// debug information. This is used when a domtree is unavailable and thus
749 /// mem2reg in its full form can't be used to handle promotion of allocas to
751 class AllocaPromoter : public LoadAndStorePromoter {
755 SmallVector<DbgDeclareInst *, 4> DDIs;
756 SmallVector<DbgValueInst *, 4> DVIs;
759 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
760 AllocaInst &AI, DIBuilder &DIB)
761 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
763 void run(const SmallVectorImpl<Instruction*> &Insts) {
764 // Remember which alloca we're promoting (for isInstInList).
765 if (MDNode *DebugNode = MDNode::getIfExists(AI.getContext(), &AI)) {
766 for (Value::use_iterator UI = DebugNode->use_begin(),
767 UE = DebugNode->use_end();
769 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
771 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI))
775 LoadAndStorePromoter::run(Insts);
776 AI.eraseFromParent();
777 while (!DDIs.empty())
778 DDIs.pop_back_val()->eraseFromParent();
779 while (!DVIs.empty())
780 DVIs.pop_back_val()->eraseFromParent();
783 virtual bool isInstInList(Instruction *I,
784 const SmallVectorImpl<Instruction*> &Insts) const {
785 if (LoadInst *LI = dyn_cast<LoadInst>(I))
786 return LI->getOperand(0) == &AI;
787 return cast<StoreInst>(I)->getPointerOperand() == &AI;
790 virtual void updateDebugInfo(Instruction *Inst) const {
791 for (SmallVectorImpl<DbgDeclareInst *>::const_iterator I = DDIs.begin(),
792 E = DDIs.end(); I != E; ++I) {
793 DbgDeclareInst *DDI = *I;
794 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
795 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
796 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
797 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
799 for (SmallVectorImpl<DbgValueInst *>::const_iterator I = DVIs.begin(),
800 E = DVIs.end(); I != E; ++I) {
801 DbgValueInst *DVI = *I;
803 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
804 // If an argument is zero extended then use argument directly. The ZExt
805 // may be zapped by an optimization pass in future.
806 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
807 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
808 else if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
809 Arg = dyn_cast<Argument>(SExt->getOperand(0));
811 Arg = SI->getValueOperand();
812 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
813 Arg = LI->getPointerOperand();
817 Instruction *DbgVal =
818 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
820 DbgVal->setDebugLoc(DVI->getDebugLoc());
824 } // end anon namespace
828 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
830 /// This pass takes allocations which can be completely analyzed (that is, they
831 /// don't escape) and tries to turn them into scalar SSA values. There are
832 /// a few steps to this process.
834 /// 1) It takes allocations of aggregates and analyzes the ways in which they
835 /// are used to try to split them into smaller allocations, ideally of
836 /// a single scalar data type. It will split up memcpy and memset accesses
837 /// as necessary and try to isolate individual scalar accesses.
838 /// 2) It will transform accesses into forms which are suitable for SSA value
839 /// promotion. This can be replacing a memset with a scalar store of an
840 /// integer value, or it can involve speculating operations on a PHI or
841 /// select to be a PHI or select of the results.
842 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
843 /// onto insert and extract operations on a vector value, and convert them to
844 /// this form. By doing so, it will enable promotion of vector aggregates to
845 /// SSA vector values.
846 class SROA : public FunctionPass {
847 const bool RequiresDomTree;
850 const DataLayout *DL;
853 /// \brief Worklist of alloca instructions to simplify.
855 /// Each alloca in the function is added to this. Each new alloca formed gets
856 /// added to it as well to recursively simplify unless that alloca can be
857 /// directly promoted. Finally, each time we rewrite a use of an alloca other
858 /// the one being actively rewritten, we add it back onto the list if not
859 /// already present to ensure it is re-visited.
860 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > Worklist;
862 /// \brief A collection of instructions to delete.
863 /// We try to batch deletions to simplify code and make things a bit more
865 SetVector<Instruction *, SmallVector<Instruction *, 8> > DeadInsts;
867 /// \brief Post-promotion worklist.
869 /// Sometimes we discover an alloca which has a high probability of becoming
870 /// viable for SROA after a round of promotion takes place. In those cases,
871 /// the alloca is enqueued here for re-processing.
873 /// Note that we have to be very careful to clear allocas out of this list in
874 /// the event they are deleted.
875 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > PostPromotionWorklist;
877 /// \brief A collection of alloca instructions we can directly promote.
878 std::vector<AllocaInst *> PromotableAllocas;
880 /// \brief A worklist of PHIs to speculate prior to promoting allocas.
882 /// All of these PHIs have been checked for the safety of speculation and by
883 /// being speculated will allow promoting allocas currently in the promotable
885 SetVector<PHINode *, SmallVector<PHINode *, 2> > SpeculatablePHIs;
887 /// \brief A worklist of select instructions to speculate prior to promoting
890 /// All of these select instructions have been checked for the safety of
891 /// speculation and by being speculated will allow promoting allocas
892 /// currently in the promotable queue.
893 SetVector<SelectInst *, SmallVector<SelectInst *, 2> > SpeculatableSelects;
896 SROA(bool RequiresDomTree = true)
897 : FunctionPass(ID), RequiresDomTree(RequiresDomTree),
899 initializeSROAPass(*PassRegistry::getPassRegistry());
901 bool runOnFunction(Function &F);
902 void getAnalysisUsage(AnalysisUsage &AU) const;
904 const char *getPassName() const { return "SROA"; }
908 friend class PHIOrSelectSpeculator;
909 friend class AllocaPartitionRewriter;
910 friend class AllocaPartitionVectorRewriter;
912 bool rewritePartitions(AllocaInst &AI, AllocaPartitioning &P,
913 AllocaPartitioning::iterator B,
914 AllocaPartitioning::iterator E,
915 int64_t BeginOffset, int64_t EndOffset,
916 ArrayRef<AllocaPartitioning::iterator> SplitUses);
917 bool splitAlloca(AllocaInst &AI, AllocaPartitioning &P);
918 bool runOnAlloca(AllocaInst &AI);
919 void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas);
920 bool promoteAllocas(Function &F);
926 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
927 return new SROA(RequiresDomTree);
930 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates",
932 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
933 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
936 /// Walk a range of a partitioning looking for a common type to cover this
937 /// sequence of partition uses.
938 static Type *findCommonType(AllocaPartitioning::const_iterator B,
939 AllocaPartitioning::const_iterator E,
940 uint64_t EndOffset) {
942 for (AllocaPartitioning::const_iterator I = B; I != E; ++I) {
943 Use *U = I->getUse();
944 if (isa<IntrinsicInst>(*U->getUser()))
946 if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
950 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser()))
951 UserTy = LI->getType();
952 else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser()))
953 UserTy = SI->getValueOperand()->getType();
955 return 0; // Bail if we have weird uses.
957 if (IntegerType *ITy = dyn_cast<IntegerType>(UserTy)) {
958 // If the type is larger than the partition, skip it. We only encounter
959 // this for split integer operations where we want to use the type of
961 // entity causing the split.
962 if (ITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
965 // If we have found an integer type use covering the alloca, use that
966 // regardless of the other types, as integers are often used for a
972 if (Ty && Ty != UserTy)
980 /// PHI instructions that use an alloca and are subsequently loaded can be
981 /// rewritten to load both input pointers in the pred blocks and then PHI the
982 /// results, allowing the load of the alloca to be promoted.
984 /// %P2 = phi [i32* %Alloca, i32* %Other]
985 /// %V = load i32* %P2
987 /// %V1 = load i32* %Alloca -> will be mem2reg'd
989 /// %V2 = load i32* %Other
991 /// %V = phi [i32 %V1, i32 %V2]
993 /// We can do this to a select if its only uses are loads and if the operands
994 /// to the select can be loaded unconditionally.
996 /// FIXME: This should be hoisted into a generic utility, likely in
997 /// Transforms/Util/Local.h
998 static bool isSafePHIToSpeculate(PHINode &PN,
999 const DataLayout *DL = 0) {
1000 // For now, we can only do this promotion if the load is in the same block
1001 // as the PHI, and if there are no stores between the phi and load.
1002 // TODO: Allow recursive phi users.
1003 // TODO: Allow stores.
1004 BasicBlock *BB = PN.getParent();
1005 unsigned MaxAlign = 0;
1006 bool HaveLoad = false;
1007 for (Value::use_iterator UI = PN.use_begin(), UE = PN.use_end(); UI != UE;
1009 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1010 if (LI == 0 || !LI->isSimple())
1013 // For now we only allow loads in the same block as the PHI. This is
1014 // a common case that happens when instcombine merges two loads through
1016 if (LI->getParent() != BB)
1019 // Ensure that there are no instructions between the PHI and the load that
1021 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1022 if (BBI->mayWriteToMemory())
1025 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1032 // We can only transform this if it is safe to push the loads into the
1033 // predecessor blocks. The only thing to watch out for is that we can't put
1034 // a possibly trapping load in the predecessor if it is a critical edge.
1035 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1036 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1037 Value *InVal = PN.getIncomingValue(Idx);
1039 // If the value is produced by the terminator of the predecessor (an
1040 // invoke) or it has side-effects, there is no valid place to put a load
1041 // in the predecessor.
1042 if (TI == InVal || TI->mayHaveSideEffects())
1045 // If the predecessor has a single successor, then the edge isn't
1047 if (TI->getNumSuccessors() == 1)
1050 // If this pointer is always safe to load, or if we can prove that there
1051 // is already a load in the block, then we can move the load to the pred
1053 if (InVal->isDereferenceablePointer() ||
1054 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, DL))
1063 static void speculatePHINodeLoads(PHINode &PN) {
1064 DEBUG(dbgs() << " original: " << PN << "\n");
1066 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1067 IRBuilderTy PHIBuilder(&PN);
1068 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1069 PN.getName() + ".sroa.speculated");
1071 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1072 // matter which one we get and if any differ.
1073 LoadInst *SomeLoad = cast<LoadInst>(*PN.use_begin());
1074 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1075 unsigned Align = SomeLoad->getAlignment();
1077 // Rewrite all loads of the PN to use the new PHI.
1078 while (!PN.use_empty()) {
1079 LoadInst *LI = cast<LoadInst>(*PN.use_begin());
1080 LI->replaceAllUsesWith(NewPN);
1081 LI->eraseFromParent();
1084 // Inject loads into all of the pred blocks.
1085 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1086 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1087 TerminatorInst *TI = Pred->getTerminator();
1088 Value *InVal = PN.getIncomingValue(Idx);
1089 IRBuilderTy PredBuilder(TI);
1091 LoadInst *Load = PredBuilder.CreateLoad(
1092 InVal, (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1093 ++NumLoadsSpeculated;
1094 Load->setAlignment(Align);
1096 Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1097 NewPN->addIncoming(Load, Pred);
1100 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1101 PN.eraseFromParent();
1104 /// Select instructions that use an alloca and are subsequently loaded can be
1105 /// rewritten to load both input pointers and then select between the result,
1106 /// allowing the load of the alloca to be promoted.
1108 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1109 /// %V = load i32* %P2
1111 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1112 /// %V2 = load i32* %Other
1113 /// %V = select i1 %cond, i32 %V1, i32 %V2
1115 /// We can do this to a select if its only uses are loads and if the operand
1116 /// to the select can be loaded unconditionally.
1117 static bool isSafeSelectToSpeculate(SelectInst &SI, const DataLayout *DL = 0) {
1118 Value *TValue = SI.getTrueValue();
1119 Value *FValue = SI.getFalseValue();
1120 bool TDerefable = TValue->isDereferenceablePointer();
1121 bool FDerefable = FValue->isDereferenceablePointer();
1123 for (Value::use_iterator UI = SI.use_begin(), UE = SI.use_end(); UI != UE;
1125 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1126 if (LI == 0 || !LI->isSimple())
1129 // Both operands to the select need to be dereferencable, either
1130 // absolutely (e.g. allocas) or at this point because we can see other
1133 !isSafeToLoadUnconditionally(TValue, LI, LI->getAlignment(), DL))
1136 !isSafeToLoadUnconditionally(FValue, LI, LI->getAlignment(), DL))
1143 static void speculateSelectInstLoads(SelectInst &SI) {
1144 DEBUG(dbgs() << " original: " << SI << "\n");
1146 IRBuilderTy IRB(&SI);
1147 Value *TV = SI.getTrueValue();
1148 Value *FV = SI.getFalseValue();
1149 // Replace the loads of the select with a select of two loads.
1150 while (!SI.use_empty()) {
1151 LoadInst *LI = cast<LoadInst>(*SI.use_begin());
1152 assert(LI->isSimple() && "We only speculate simple loads");
1154 IRB.SetInsertPoint(LI);
1156 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1158 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1159 NumLoadsSpeculated += 2;
1161 // Transfer alignment and TBAA info if present.
1162 TL->setAlignment(LI->getAlignment());
1163 FL->setAlignment(LI->getAlignment());
1164 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
1165 TL->setMetadata(LLVMContext::MD_tbaa, Tag);
1166 FL->setMetadata(LLVMContext::MD_tbaa, Tag);
1169 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1170 LI->getName() + ".sroa.speculated");
1172 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1173 LI->replaceAllUsesWith(V);
1174 LI->eraseFromParent();
1176 SI.eraseFromParent();
1179 /// \brief Build a GEP out of a base pointer and indices.
1181 /// This will return the BasePtr if that is valid, or build a new GEP
1182 /// instruction using the IRBuilder if GEP-ing is needed.
1183 static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1184 SmallVectorImpl<Value *> &Indices) {
1185 if (Indices.empty())
1188 // A single zero index is a no-op, so check for this and avoid building a GEP
1190 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1193 return IRB.CreateInBoundsGEP(BasePtr, Indices, "idx");
1196 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1197 /// TargetTy without changing the offset of the pointer.
1199 /// This routine assumes we've already established a properly offset GEP with
1200 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1201 /// zero-indices down through type layers until we find one the same as
1202 /// TargetTy. If we can't find one with the same type, we at least try to use
1203 /// one with the same size. If none of that works, we just produce the GEP as
1204 /// indicated by Indices to have the correct offset.
1205 static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1206 Value *BasePtr, Type *Ty, Type *TargetTy,
1207 SmallVectorImpl<Value *> &Indices) {
1209 return buildGEP(IRB, BasePtr, Indices);
1211 // See if we can descend into a struct and locate a field with the correct
1213 unsigned NumLayers = 0;
1214 Type *ElementTy = Ty;
1216 if (ElementTy->isPointerTy())
1218 if (SequentialType *SeqTy = dyn_cast<SequentialType>(ElementTy)) {
1219 ElementTy = SeqTy->getElementType();
1220 // Note that we use the default address space as this index is over an
1221 // array or a vector, not a pointer.
1222 Indices.push_back(IRB.getInt(APInt(DL.getPointerSizeInBits(0), 0)));
1223 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1224 if (STy->element_begin() == STy->element_end())
1225 break; // Nothing left to descend into.
1226 ElementTy = *STy->element_begin();
1227 Indices.push_back(IRB.getInt32(0));
1232 } while (ElementTy != TargetTy);
1233 if (ElementTy != TargetTy)
1234 Indices.erase(Indices.end() - NumLayers, Indices.end());
1236 return buildGEP(IRB, BasePtr, Indices);
1239 /// \brief Recursively compute indices for a natural GEP.
1241 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1242 /// element types adding appropriate indices for the GEP.
1243 static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1244 Value *Ptr, Type *Ty, APInt &Offset,
1246 SmallVectorImpl<Value *> &Indices) {
1248 return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices);
1250 // We can't recurse through pointer types.
1251 if (Ty->isPointerTy())
1254 // We try to analyze GEPs over vectors here, but note that these GEPs are
1255 // extremely poorly defined currently. The long-term goal is to remove GEPing
1256 // over a vector from the IR completely.
1257 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1258 unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
1259 if (ElementSizeInBits % 8)
1260 return 0; // GEPs over non-multiple of 8 size vector elements are invalid.
1261 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1262 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1263 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1265 Offset -= NumSkippedElements * ElementSize;
1266 Indices.push_back(IRB.getInt(NumSkippedElements));
1267 return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1268 Offset, TargetTy, Indices);
1271 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1272 Type *ElementTy = ArrTy->getElementType();
1273 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1274 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1275 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1278 Offset -= NumSkippedElements * ElementSize;
1279 Indices.push_back(IRB.getInt(NumSkippedElements));
1280 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1284 StructType *STy = dyn_cast<StructType>(Ty);
1288 const StructLayout *SL = DL.getStructLayout(STy);
1289 uint64_t StructOffset = Offset.getZExtValue();
1290 if (StructOffset >= SL->getSizeInBytes())
1292 unsigned Index = SL->getElementContainingOffset(StructOffset);
1293 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1294 Type *ElementTy = STy->getElementType(Index);
1295 if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
1296 return 0; // The offset points into alignment padding.
1298 Indices.push_back(IRB.getInt32(Index));
1299 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1303 /// \brief Get a natural GEP from a base pointer to a particular offset and
1304 /// resulting in a particular type.
1306 /// The goal is to produce a "natural" looking GEP that works with the existing
1307 /// composite types to arrive at the appropriate offset and element type for
1308 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1309 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1310 /// Indices, and setting Ty to the result subtype.
1312 /// If no natural GEP can be constructed, this function returns null.
1313 static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1314 Value *Ptr, APInt Offset, Type *TargetTy,
1315 SmallVectorImpl<Value *> &Indices) {
1316 PointerType *Ty = cast<PointerType>(Ptr->getType());
1318 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1320 if (Ty == IRB.getInt8PtrTy() && TargetTy->isIntegerTy(8))
1323 Type *ElementTy = Ty->getElementType();
1324 if (!ElementTy->isSized())
1325 return 0; // We can't GEP through an unsized element.
1326 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1327 if (ElementSize == 0)
1328 return 0; // Zero-length arrays can't help us build a natural GEP.
1329 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1331 Offset -= NumSkippedElements * ElementSize;
1332 Indices.push_back(IRB.getInt(NumSkippedElements));
1333 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1337 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1338 /// resulting pointer has PointerTy.
1340 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1341 /// and produces the pointer type desired. Where it cannot, it will try to use
1342 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1343 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1344 /// bitcast to the type.
1346 /// The strategy for finding the more natural GEPs is to peel off layers of the
1347 /// pointer, walking back through bit casts and GEPs, searching for a base
1348 /// pointer from which we can compute a natural GEP with the desired
1349 /// properties. The algorithm tries to fold as many constant indices into
1350 /// a single GEP as possible, thus making each GEP more independent of the
1351 /// surrounding code.
1352 static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL,
1353 Value *Ptr, APInt Offset, Type *PointerTy) {
1354 // Even though we don't look through PHI nodes, we could be called on an
1355 // instruction in an unreachable block, which may be on a cycle.
1356 SmallPtrSet<Value *, 4> Visited;
1357 Visited.insert(Ptr);
1358 SmallVector<Value *, 4> Indices;
1360 // We may end up computing an offset pointer that has the wrong type. If we
1361 // never are able to compute one directly that has the correct type, we'll
1362 // fall back to it, so keep it around here.
1363 Value *OffsetPtr = 0;
1365 // Remember any i8 pointer we come across to re-use if we need to do a raw
1368 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1370 Type *TargetTy = PointerTy->getPointerElementType();
1373 // First fold any existing GEPs into the offset.
1374 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1375 APInt GEPOffset(Offset.getBitWidth(), 0);
1376 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1378 Offset += GEPOffset;
1379 Ptr = GEP->getPointerOperand();
1380 if (!Visited.insert(Ptr))
1384 // See if we can perform a natural GEP here.
1386 if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1388 if (P->getType() == PointerTy) {
1389 // Zap any offset pointer that we ended up computing in previous rounds.
1390 if (OffsetPtr && OffsetPtr->use_empty())
1391 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1392 I->eraseFromParent();
1400 // Stash this pointer if we've found an i8*.
1401 if (Ptr->getType()->isIntegerTy(8)) {
1403 Int8PtrOffset = Offset;
1406 // Peel off a layer of the pointer and update the offset appropriately.
1407 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1408 Ptr = cast<Operator>(Ptr)->getOperand(0);
1409 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1410 if (GA->mayBeOverridden())
1412 Ptr = GA->getAliasee();
1416 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1417 } while (Visited.insert(Ptr));
1421 Int8Ptr = IRB.CreateBitCast(Ptr, IRB.getInt8PtrTy(),
1423 Int8PtrOffset = Offset;
1426 OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
1427 IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1432 // On the off chance we were targeting i8*, guard the bitcast here.
1433 if (Ptr->getType() != PointerTy)
1434 Ptr = IRB.CreateBitCast(Ptr, PointerTy, "cast");
1439 /// \brief Test whether we can convert a value from the old to the new type.
1441 /// This predicate should be used to guard calls to convertValue in order to
1442 /// ensure that we only try to convert viable values. The strategy is that we
1443 /// will peel off single element struct and array wrappings to get to an
1444 /// underlying value, and convert that value.
1445 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1448 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1449 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1450 if (NewITy->getBitWidth() >= OldITy->getBitWidth())
1452 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1454 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1457 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1458 if (NewTy->isPointerTy() && OldTy->isPointerTy())
1460 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
1468 /// \brief Generic routine to convert an SSA value to a value of a different
1471 /// This will try various different casting techniques, such as bitcasts,
1472 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1473 /// two types for viability with this routine.
1474 static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1476 assert(canConvertValue(DL, V->getType(), Ty) &&
1477 "Value not convertable to type");
1478 if (V->getType() == Ty)
1480 if (IntegerType *OldITy = dyn_cast<IntegerType>(V->getType()))
1481 if (IntegerType *NewITy = dyn_cast<IntegerType>(Ty))
1482 if (NewITy->getBitWidth() > OldITy->getBitWidth())
1483 return IRB.CreateZExt(V, NewITy);
1484 if (V->getType()->isIntegerTy() && Ty->isPointerTy())
1485 return IRB.CreateIntToPtr(V, Ty);
1486 if (V->getType()->isPointerTy() && Ty->isIntegerTy())
1487 return IRB.CreatePtrToInt(V, Ty);
1489 return IRB.CreateBitCast(V, Ty);
1492 /// \brief Test whether the given partition use can be promoted to a vector.
1494 /// This function is called to test each entry in a partioning which is slated
1495 /// for a single partition.
1496 static bool isVectorPromotionViableForPartitioning(
1497 const DataLayout &DL, AllocaPartitioning &P,
1498 uint64_t PartitionBeginOffset, uint64_t PartitionEndOffset, VectorType *Ty,
1499 uint64_t ElementSize, AllocaPartitioning::const_iterator I) {
1500 // First validate the partitioning offsets.
1501 uint64_t BeginOffset =
1502 std::max(I->beginOffset(), PartitionBeginOffset) - PartitionBeginOffset;
1503 uint64_t BeginIndex = BeginOffset / ElementSize;
1504 if (BeginIndex * ElementSize != BeginOffset ||
1505 BeginIndex >= Ty->getNumElements())
1507 uint64_t EndOffset =
1508 std::min(I->endOffset(), PartitionEndOffset) - PartitionBeginOffset;
1509 uint64_t EndIndex = EndOffset / ElementSize;
1510 if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
1513 assert(EndIndex > BeginIndex && "Empty vector!");
1514 uint64_t NumElements = EndIndex - BeginIndex;
1516 (NumElements == 1) ? Ty->getElementType()
1517 : VectorType::get(Ty->getElementType(), NumElements);
1520 Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1522 Use *U = I->getUse();
1524 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1525 if (MI->isVolatile())
1527 if (!I->isSplittable())
1528 return false; // Skip any unsplittable intrinsics.
1529 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1530 // Disable vector promotion when there are loads or stores of an FCA.
1532 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1533 if (LI->isVolatile())
1535 Type *LTy = LI->getType();
1536 if (PartitionBeginOffset > I->beginOffset() ||
1537 PartitionEndOffset < I->endOffset()) {
1538 assert(LTy->isIntegerTy());
1541 if (!canConvertValue(DL, PartitionTy, LTy))
1543 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1544 if (SI->isVolatile())
1546 Type *STy = SI->getValueOperand()->getType();
1547 if (PartitionBeginOffset > I->beginOffset() ||
1548 PartitionEndOffset < I->endOffset()) {
1549 assert(STy->isIntegerTy());
1552 if (!canConvertValue(DL, STy, PartitionTy))
1559 /// \brief Test whether the given alloca partition can be promoted to a vector.
1561 /// This is a quick test to check whether we can rewrite a particular alloca
1562 /// partition (and its newly formed alloca) into a vector alloca with only
1563 /// whole-vector loads and stores such that it could be promoted to a vector
1564 /// SSA value. We only can ensure this for a limited set of operations, and we
1565 /// don't want to do the rewrites unless we are confident that the result will
1566 /// be promotable, so we have an early test here.
1567 static bool isVectorPromotionViable(
1568 const DataLayout &DL, Type *AllocaTy, AllocaPartitioning &P,
1569 uint64_t PartitionBeginOffset, uint64_t PartitionEndOffset,
1570 AllocaPartitioning::const_iterator I, AllocaPartitioning::const_iterator E,
1571 ArrayRef<AllocaPartitioning::iterator> SplitUses) {
1572 VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
1576 uint64_t ElementSize = DL.getTypeSizeInBits(Ty->getScalarType());
1578 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1579 // that aren't byte sized.
1580 if (ElementSize % 8)
1582 assert((DL.getTypeSizeInBits(Ty) % 8) == 0 &&
1583 "vector size not a multiple of element size?");
1587 if (!isVectorPromotionViableForPartitioning(
1588 DL, P, PartitionBeginOffset, PartitionEndOffset, Ty, ElementSize,
1592 for (ArrayRef<AllocaPartitioning::iterator>::const_iterator
1593 SUI = SplitUses.begin(),
1594 SUE = SplitUses.end();
1596 if (!isVectorPromotionViableForPartitioning(
1597 DL, P, PartitionBeginOffset, PartitionEndOffset, Ty, ElementSize,
1604 /// \brief Test whether a partitioning slice of an alloca is a valid set of
1605 /// operations for integer widening.
1607 /// This implements the necessary checking for the \c isIntegerWideningViable
1608 /// test below on a single partitioning slice of the alloca.
1609 static bool isIntegerWideningViableForPartitioning(
1610 const DataLayout &DL, Type *AllocaTy, uint64_t AllocBeginOffset,
1611 uint64_t Size, AllocaPartitioning &P, AllocaPartitioning::const_iterator I,
1612 bool &WholeAllocaOp) {
1613 uint64_t RelBegin = I->beginOffset() - AllocBeginOffset;
1614 uint64_t RelEnd = I->endOffset() - AllocBeginOffset;
1616 // We can't reasonably handle cases where the load or store extends past
1617 // the end of the aloca's type and into its padding.
1621 Use *U = I->getUse();
1623 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1624 if (LI->isVolatile())
1626 if (RelBegin == 0 && RelEnd == Size)
1627 WholeAllocaOp = true;
1628 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
1629 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1631 } else if (RelBegin != 0 || RelEnd != Size ||
1632 !canConvertValue(DL, AllocaTy, LI->getType())) {
1633 // Non-integer loads need to be convertible from the alloca type so that
1634 // they are promotable.
1637 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1638 Type *ValueTy = SI->getValueOperand()->getType();
1639 if (SI->isVolatile())
1641 if (RelBegin == 0 && RelEnd == Size)
1642 WholeAllocaOp = true;
1643 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
1644 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1646 } else if (RelBegin != 0 || RelEnd != Size ||
1647 !canConvertValue(DL, ValueTy, AllocaTy)) {
1648 // Non-integer stores need to be convertible to the alloca type so that
1649 // they are promotable.
1652 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1653 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
1655 if (!I->isSplittable())
1656 return false; // Skip any unsplittable intrinsics.
1657 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1658 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1659 II->getIntrinsicID() != Intrinsic::lifetime_end)
1668 /// \brief Test whether the given alloca partition's integer operations can be
1669 /// widened to promotable ones.
1671 /// This is a quick test to check whether we can rewrite the integer loads and
1672 /// stores to a particular alloca into wider loads and stores and be able to
1673 /// promote the resulting alloca.
1675 isIntegerWideningViable(const DataLayout &DL, Type *AllocaTy,
1676 uint64_t AllocBeginOffset, AllocaPartitioning &P,
1677 AllocaPartitioning::const_iterator I,
1678 AllocaPartitioning::const_iterator E,
1679 ArrayRef<AllocaPartitioning::iterator> SplitUses) {
1680 uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
1681 // Don't create integer types larger than the maximum bitwidth.
1682 if (SizeInBits > IntegerType::MAX_INT_BITS)
1685 // Don't try to handle allocas with bit-padding.
1686 if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
1689 // We need to ensure that an integer type with the appropriate bitwidth can
1690 // be converted to the alloca type, whatever that is. We don't want to force
1691 // the alloca itself to have an integer type if there is a more suitable one.
1692 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
1693 if (!canConvertValue(DL, AllocaTy, IntTy) ||
1694 !canConvertValue(DL, IntTy, AllocaTy))
1697 uint64_t Size = DL.getTypeStoreSize(AllocaTy);
1699 // While examining uses, we ensure that the alloca has a covering load or
1700 // store. We don't want to widen the integer operations only to fail to
1701 // promote due to some other unsplittable entry (which we may make splittable
1702 // later). However, if there are only splittable uses, go ahead and assume
1703 // that we cover the alloca.
1704 bool WholeAllocaOp = (I != E) ? false : DL.isLegalInteger(SizeInBits);
1707 if (!isIntegerWideningViableForPartitioning(DL, AllocaTy, AllocBeginOffset,
1708 Size, P, I, WholeAllocaOp))
1711 for (ArrayRef<AllocaPartitioning::iterator>::const_iterator
1712 SUI = SplitUses.begin(),
1713 SUE = SplitUses.end();
1715 if (!isIntegerWideningViableForPartitioning(DL, AllocaTy, AllocBeginOffset,
1716 Size, P, *SUI, WholeAllocaOp))
1719 return WholeAllocaOp;
1722 static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1723 IntegerType *Ty, uint64_t Offset,
1724 const Twine &Name) {
1725 DEBUG(dbgs() << " start: " << *V << "\n");
1726 IntegerType *IntTy = cast<IntegerType>(V->getType());
1727 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
1728 "Element extends past full value");
1729 uint64_t ShAmt = 8*Offset;
1730 if (DL.isBigEndian())
1731 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
1733 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
1734 DEBUG(dbgs() << " shifted: " << *V << "\n");
1736 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
1737 "Cannot extract to a larger integer!");
1739 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
1740 DEBUG(dbgs() << " trunced: " << *V << "\n");
1745 static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
1746 Value *V, uint64_t Offset, const Twine &Name) {
1747 IntegerType *IntTy = cast<IntegerType>(Old->getType());
1748 IntegerType *Ty = cast<IntegerType>(V->getType());
1749 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
1750 "Cannot insert a larger integer!");
1751 DEBUG(dbgs() << " start: " << *V << "\n");
1753 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
1754 DEBUG(dbgs() << " extended: " << *V << "\n");
1756 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
1757 "Element store outside of alloca store");
1758 uint64_t ShAmt = 8*Offset;
1759 if (DL.isBigEndian())
1760 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
1762 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
1763 DEBUG(dbgs() << " shifted: " << *V << "\n");
1766 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
1767 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
1768 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
1769 DEBUG(dbgs() << " masked: " << *Old << "\n");
1770 V = IRB.CreateOr(Old, V, Name + ".insert");
1771 DEBUG(dbgs() << " inserted: " << *V << "\n");
1776 static Value *extractVector(IRBuilderTy &IRB, Value *V,
1777 unsigned BeginIndex, unsigned EndIndex,
1778 const Twine &Name) {
1779 VectorType *VecTy = cast<VectorType>(V->getType());
1780 unsigned NumElements = EndIndex - BeginIndex;
1781 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
1783 if (NumElements == VecTy->getNumElements())
1786 if (NumElements == 1) {
1787 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
1789 DEBUG(dbgs() << " extract: " << *V << "\n");
1793 SmallVector<Constant*, 8> Mask;
1794 Mask.reserve(NumElements);
1795 for (unsigned i = BeginIndex; i != EndIndex; ++i)
1796 Mask.push_back(IRB.getInt32(i));
1797 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
1798 ConstantVector::get(Mask),
1800 DEBUG(dbgs() << " shuffle: " << *V << "\n");
1804 static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
1805 unsigned BeginIndex, const Twine &Name) {
1806 VectorType *VecTy = cast<VectorType>(Old->getType());
1807 assert(VecTy && "Can only insert a vector into a vector");
1809 VectorType *Ty = dyn_cast<VectorType>(V->getType());
1811 // Single element to insert.
1812 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
1814 DEBUG(dbgs() << " insert: " << *V << "\n");
1818 assert(Ty->getNumElements() <= VecTy->getNumElements() &&
1819 "Too many elements!");
1820 if (Ty->getNumElements() == VecTy->getNumElements()) {
1821 assert(V->getType() == VecTy && "Vector type mismatch");
1824 unsigned EndIndex = BeginIndex + Ty->getNumElements();
1826 // When inserting a smaller vector into the larger to store, we first
1827 // use a shuffle vector to widen it with undef elements, and then
1828 // a second shuffle vector to select between the loaded vector and the
1830 SmallVector<Constant*, 8> Mask;
1831 Mask.reserve(VecTy->getNumElements());
1832 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
1833 if (i >= BeginIndex && i < EndIndex)
1834 Mask.push_back(IRB.getInt32(i - BeginIndex));
1836 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
1837 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
1838 ConstantVector::get(Mask),
1840 DEBUG(dbgs() << " shuffle: " << *V << "\n");
1843 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
1844 Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
1846 V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
1848 DEBUG(dbgs() << " blend: " << *V << "\n");
1853 /// \brief Visitor to rewrite instructions using a partition of an alloca to
1854 /// use a new alloca.
1856 /// Also implements the rewriting to vector-based accesses when the partition
1857 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
1859 class AllocaPartitionRewriter : public InstVisitor<AllocaPartitionRewriter,
1861 // Befriend the base class so it can delegate to private visit methods.
1862 friend class llvm::InstVisitor<AllocaPartitionRewriter, bool>;
1863 typedef llvm::InstVisitor<AllocaPartitionRewriter, bool> Base;
1865 const DataLayout &DL;
1866 AllocaPartitioning &P;
1868 AllocaInst &OldAI, &NewAI;
1869 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
1872 // If we are rewriting an alloca partition which can be written as pure
1873 // vector operations, we stash extra information here. When VecTy is
1874 // non-null, we have some strict guarantees about the rewritten alloca:
1875 // - The new alloca is exactly the size of the vector type here.
1876 // - The accesses all either map to the entire vector or to a single
1878 // - The set of accessing instructions is only one of those handled above
1879 // in isVectorPromotionViable. Generally these are the same access kinds
1880 // which are promotable via mem2reg.
1883 uint64_t ElementSize;
1885 // This is a convenience and flag variable that will be null unless the new
1886 // alloca's integer operations should be widened to this integer type due to
1887 // passing isIntegerWideningViable above. If it is non-null, the desired
1888 // integer type will be stored here for easy access during rewriting.
1891 // The offset of the partition user currently being rewritten.
1892 uint64_t BeginOffset, EndOffset;
1896 Instruction *OldPtr;
1898 // Utility IR builder, whose name prefix is setup for each visited use, and
1899 // the insertion point is set to point to the user.
1903 AllocaPartitionRewriter(const DataLayout &DL, AllocaPartitioning &P,
1904 SROA &Pass, AllocaInst &OldAI, AllocaInst &NewAI,
1905 uint64_t NewBeginOffset, uint64_t NewEndOffset,
1906 bool IsVectorPromotable = false,
1907 bool IsIntegerPromotable = false)
1908 : DL(DL), P(P), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
1909 NewAllocaBeginOffset(NewBeginOffset), NewAllocaEndOffset(NewEndOffset),
1910 NewAllocaTy(NewAI.getAllocatedType()),
1911 VecTy(IsVectorPromotable ? cast<VectorType>(NewAllocaTy) : 0),
1912 ElementTy(VecTy ? VecTy->getElementType() : 0),
1913 ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
1914 IntTy(IsIntegerPromotable
1917 DL.getTypeSizeInBits(NewAI.getAllocatedType()))
1919 BeginOffset(), EndOffset(), IsSplittable(), IsSplit(), OldUse(),
1920 OldPtr(), IRB(NewAI.getContext(), ConstantFolder()) {
1922 assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
1923 "Only multiple-of-8 sized vector elements are viable");
1926 assert((!IsVectorPromotable && !IsIntegerPromotable) ||
1927 IsVectorPromotable != IsIntegerPromotable);
1930 bool visit(AllocaPartitioning::const_iterator I) {
1931 bool CanSROA = true;
1932 BeginOffset = I->beginOffset();
1933 EndOffset = I->endOffset();
1934 IsSplittable = I->isSplittable();
1936 BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
1938 OldUse = I->getUse();
1939 OldPtr = cast<Instruction>(OldUse->get());
1941 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
1942 IRB.SetInsertPoint(OldUserI);
1943 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
1944 IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
1946 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
1953 // Make sure the other visit overloads are visible.
1956 // Every instruction which can end up as a user must have a rewrite rule.
1957 bool visitInstruction(Instruction &I) {
1958 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
1959 llvm_unreachable("No rewrite rule for this instruction!");
1962 Value *getAdjustedAllocaPtr(IRBuilderTy &IRB, uint64_t Offset,
1964 assert(Offset >= NewAllocaBeginOffset);
1965 return getAdjustedPtr(IRB, DL, &NewAI, APInt(DL.getPointerSizeInBits(),
1966 Offset - NewAllocaBeginOffset),
1970 /// \brief Compute suitable alignment to access an offset into the new alloca.
1971 unsigned getOffsetAlign(uint64_t Offset) {
1972 unsigned NewAIAlign = NewAI.getAlignment();
1974 NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
1975 return MinAlign(NewAIAlign, Offset);
1978 /// \brief Compute suitable alignment to access a type at an offset of the
1981 /// \returns zero if the type's ABI alignment is a suitable alignment,
1982 /// otherwise returns the maximal suitable alignment.
1983 unsigned getOffsetTypeAlign(Type *Ty, uint64_t Offset) {
1984 unsigned Align = getOffsetAlign(Offset);
1985 return Align == DL.getABITypeAlignment(Ty) ? 0 : Align;
1988 unsigned getIndex(uint64_t Offset) {
1989 assert(VecTy && "Can only call getIndex when rewriting a vector");
1990 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
1991 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
1992 uint32_t Index = RelOffset / ElementSize;
1993 assert(Index * ElementSize == RelOffset);
1997 void deleteIfTriviallyDead(Value *V) {
1998 Instruction *I = cast<Instruction>(V);
1999 if (isInstructionTriviallyDead(I))
2000 Pass.DeadInsts.insert(I);
2003 Value *rewriteVectorizedLoadInst(uint64_t NewBeginOffset,
2004 uint64_t NewEndOffset) {
2005 unsigned BeginIndex = getIndex(NewBeginOffset);
2006 unsigned EndIndex = getIndex(NewEndOffset);
2007 assert(EndIndex > BeginIndex && "Empty vector!");
2009 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2011 return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2014 Value *rewriteIntegerLoad(LoadInst &LI, uint64_t NewBeginOffset,
2015 uint64_t NewEndOffset) {
2016 assert(IntTy && "We cannot insert an integer to the alloca");
2017 assert(!LI.isVolatile());
2018 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2020 V = convertValue(DL, IRB, V, IntTy);
2021 assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2022 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2023 if (Offset > 0 || NewEndOffset < NewAllocaEndOffset)
2024 V = extractInteger(DL, IRB, V, cast<IntegerType>(LI.getType()), Offset,
2029 bool visitLoadInst(LoadInst &LI) {
2030 DEBUG(dbgs() << " original: " << LI << "\n");
2031 Value *OldOp = LI.getOperand(0);
2032 assert(OldOp == OldPtr);
2034 // Compute the intersecting offset range.
2035 assert(BeginOffset < NewAllocaEndOffset);
2036 assert(EndOffset > NewAllocaBeginOffset);
2037 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2038 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2040 uint64_t Size = NewEndOffset - NewBeginOffset;
2042 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), Size * 8)
2044 bool IsPtrAdjusted = false;
2047 V = rewriteVectorizedLoadInst(NewBeginOffset, NewEndOffset);
2048 } else if (IntTy && LI.getType()->isIntegerTy()) {
2049 V = rewriteIntegerLoad(LI, NewBeginOffset, NewEndOffset);
2050 } else if (NewBeginOffset == NewAllocaBeginOffset &&
2051 canConvertValue(DL, NewAllocaTy, LI.getType())) {
2052 V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2053 LI.isVolatile(), "load");
2055 Type *LTy = TargetTy->getPointerTo();
2056 V = IRB.CreateAlignedLoad(
2057 getAdjustedAllocaPtr(IRB, NewBeginOffset, LTy),
2058 getOffsetTypeAlign(TargetTy, NewBeginOffset - NewAllocaBeginOffset),
2059 LI.isVolatile(), "load");
2060 IsPtrAdjusted = true;
2062 V = convertValue(DL, IRB, V, TargetTy);
2065 assert(!LI.isVolatile());
2066 assert(LI.getType()->isIntegerTy() &&
2067 "Only integer type loads and stores are split");
2068 assert(Size < DL.getTypeStoreSize(LI.getType()) &&
2069 "Split load isn't smaller than original load");
2070 assert(LI.getType()->getIntegerBitWidth() ==
2071 DL.getTypeStoreSizeInBits(LI.getType()) &&
2072 "Non-byte-multiple bit width");
2073 // Move the insertion point just past the load so that we can refer to it.
2074 IRB.SetInsertPoint(llvm::next(BasicBlock::iterator(&LI)));
2075 // Create a placeholder value with the same type as LI to use as the
2076 // basis for the new value. This allows us to replace the uses of LI with
2077 // the computed value, and then replace the placeholder with LI, leaving
2078 // LI only used for this computation.
2080 = new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
2081 V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset,
2083 LI.replaceAllUsesWith(V);
2084 Placeholder->replaceAllUsesWith(&LI);
2087 LI.replaceAllUsesWith(V);
2090 Pass.DeadInsts.insert(&LI);
2091 deleteIfTriviallyDead(OldOp);
2092 DEBUG(dbgs() << " to: " << *V << "\n");
2093 return !LI.isVolatile() && !IsPtrAdjusted;
2096 bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp,
2097 uint64_t NewBeginOffset,
2098 uint64_t NewEndOffset) {
2099 if (V->getType() != VecTy) {
2100 unsigned BeginIndex = getIndex(NewBeginOffset);
2101 unsigned EndIndex = getIndex(NewEndOffset);
2102 assert(EndIndex > BeginIndex && "Empty vector!");
2103 unsigned NumElements = EndIndex - BeginIndex;
2104 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2106 = (NumElements == 1) ? ElementTy
2107 : VectorType::get(ElementTy, NumElements);
2108 if (V->getType() != PartitionTy)
2109 V = convertValue(DL, IRB, V, PartitionTy);
2111 // Mix in the existing elements.
2112 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2114 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2116 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2117 Pass.DeadInsts.insert(&SI);
2120 DEBUG(dbgs() << " to: " << *Store << "\n");
2124 bool rewriteIntegerStore(Value *V, StoreInst &SI,
2125 uint64_t NewBeginOffset, uint64_t NewEndOffset) {
2126 assert(IntTy && "We cannot extract an integer from the alloca");
2127 assert(!SI.isVolatile());
2128 if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2129 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2131 Old = convertValue(DL, IRB, Old, IntTy);
2132 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2133 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2134 V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset,
2137 V = convertValue(DL, IRB, V, NewAllocaTy);
2138 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2139 Pass.DeadInsts.insert(&SI);
2141 DEBUG(dbgs() << " to: " << *Store << "\n");
2145 bool visitStoreInst(StoreInst &SI) {
2146 DEBUG(dbgs() << " original: " << SI << "\n");
2147 Value *OldOp = SI.getOperand(1);
2148 assert(OldOp == OldPtr);
2150 Value *V = SI.getValueOperand();
2152 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2153 // alloca that should be re-examined after promoting this alloca.
2154 if (V->getType()->isPointerTy())
2155 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2156 Pass.PostPromotionWorklist.insert(AI);
2158 // Compute the intersecting offset range.
2159 assert(BeginOffset < NewAllocaEndOffset);
2160 assert(EndOffset > NewAllocaBeginOffset);
2161 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2162 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2164 uint64_t Size = NewEndOffset - NewBeginOffset;
2165 if (Size < DL.getTypeStoreSize(V->getType())) {
2166 assert(!SI.isVolatile());
2167 assert(V->getType()->isIntegerTy() &&
2168 "Only integer type loads and stores are split");
2169 assert(V->getType()->getIntegerBitWidth() ==
2170 DL.getTypeStoreSizeInBits(V->getType()) &&
2171 "Non-byte-multiple bit width");
2172 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), Size * 8);
2173 V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset,
2178 return rewriteVectorizedStoreInst(V, SI, OldOp, NewBeginOffset,
2180 if (IntTy && V->getType()->isIntegerTy())
2181 return rewriteIntegerStore(V, SI, NewBeginOffset, NewEndOffset);
2184 if (NewBeginOffset == NewAllocaBeginOffset &&
2185 NewEndOffset == NewAllocaEndOffset &&
2186 canConvertValue(DL, V->getType(), NewAllocaTy)) {
2187 V = convertValue(DL, IRB, V, NewAllocaTy);
2188 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2191 Value *NewPtr = getAdjustedAllocaPtr(IRB, NewBeginOffset,
2192 V->getType()->getPointerTo());
2193 NewSI = IRB.CreateAlignedStore(
2194 V, NewPtr, getOffsetTypeAlign(
2195 V->getType(), NewBeginOffset - NewAllocaBeginOffset),
2199 Pass.DeadInsts.insert(&SI);
2200 deleteIfTriviallyDead(OldOp);
2202 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2203 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2206 /// \brief Compute an integer value from splatting an i8 across the given
2207 /// number of bytes.
2209 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2210 /// call this routine.
2211 /// FIXME: Heed the advice above.
2213 /// \param V The i8 value to splat.
2214 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2215 Value *getIntegerSplat(Value *V, unsigned Size) {
2216 assert(Size > 0 && "Expected a positive number of bytes.");
2217 IntegerType *VTy = cast<IntegerType>(V->getType());
2218 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2222 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size*8);
2223 V = IRB.CreateMul(IRB.CreateZExt(V, SplatIntTy, "zext"),
2224 ConstantExpr::getUDiv(
2225 Constant::getAllOnesValue(SplatIntTy),
2226 ConstantExpr::getZExt(
2227 Constant::getAllOnesValue(V->getType()),
2233 /// \brief Compute a vector splat for a given element value.
2234 Value *getVectorSplat(Value *V, unsigned NumElements) {
2235 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2236 DEBUG(dbgs() << " splat: " << *V << "\n");
2240 bool visitMemSetInst(MemSetInst &II) {
2241 DEBUG(dbgs() << " original: " << II << "\n");
2242 assert(II.getRawDest() == OldPtr);
2244 // If the memset has a variable size, it cannot be split, just adjust the
2245 // pointer to the new alloca.
2246 if (!isa<Constant>(II.getLength())) {
2248 assert(BeginOffset >= NewAllocaBeginOffset);
2250 getAdjustedAllocaPtr(IRB, BeginOffset, II.getRawDest()->getType()));
2251 Type *CstTy = II.getAlignmentCst()->getType();
2252 II.setAlignment(ConstantInt::get(CstTy, getOffsetAlign(BeginOffset)));
2254 deleteIfTriviallyDead(OldPtr);
2258 // Record this instruction for deletion.
2259 Pass.DeadInsts.insert(&II);
2261 Type *AllocaTy = NewAI.getAllocatedType();
2262 Type *ScalarTy = AllocaTy->getScalarType();
2264 // Compute the intersecting offset range.
2265 assert(BeginOffset < NewAllocaEndOffset);
2266 assert(EndOffset > NewAllocaBeginOffset);
2267 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2268 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2269 uint64_t PartitionOffset = NewBeginOffset - NewAllocaBeginOffset;
2271 // If this doesn't map cleanly onto the alloca type, and that type isn't
2272 // a single value type, just emit a memset.
2273 if (!VecTy && !IntTy &&
2274 (BeginOffset > NewAllocaBeginOffset ||
2275 EndOffset < NewAllocaEndOffset ||
2276 !AllocaTy->isSingleValueType() ||
2277 !DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy)) ||
2278 DL.getTypeSizeInBits(ScalarTy)%8 != 0)) {
2279 Type *SizeTy = II.getLength()->getType();
2280 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2281 CallInst *New = IRB.CreateMemSet(
2282 getAdjustedAllocaPtr(IRB, NewBeginOffset, II.getRawDest()->getType()),
2283 II.getValue(), Size, getOffsetAlign(PartitionOffset),
2286 DEBUG(dbgs() << " to: " << *New << "\n");
2290 // If we can represent this as a simple value, we have to build the actual
2291 // value to store, which requires expanding the byte present in memset to
2292 // a sensible representation for the alloca type. This is essentially
2293 // splatting the byte to a sufficiently wide integer, splatting it across
2294 // any desired vector width, and bitcasting to the final type.
2298 // If this is a memset of a vectorized alloca, insert it.
2299 assert(ElementTy == ScalarTy);
2301 unsigned BeginIndex = getIndex(NewBeginOffset);
2302 unsigned EndIndex = getIndex(NewEndOffset);
2303 assert(EndIndex > BeginIndex && "Empty vector!");
2304 unsigned NumElements = EndIndex - BeginIndex;
2305 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2308 getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
2309 Splat = convertValue(DL, IRB, Splat, ElementTy);
2310 if (NumElements > 1)
2311 Splat = getVectorSplat(Splat, NumElements);
2313 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2315 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2317 // If this is a memset on an alloca where we can widen stores, insert the
2319 assert(!II.isVolatile());
2321 uint64_t Size = NewEndOffset - NewBeginOffset;
2322 V = getIntegerSplat(II.getValue(), Size);
2324 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2325 EndOffset != NewAllocaBeginOffset)) {
2326 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2328 Old = convertValue(DL, IRB, Old, IntTy);
2329 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2330 V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2332 assert(V->getType() == IntTy &&
2333 "Wrong type for an alloca wide integer!");
2335 V = convertValue(DL, IRB, V, AllocaTy);
2337 // Established these invariants above.
2338 assert(NewBeginOffset == NewAllocaBeginOffset);
2339 assert(NewEndOffset == NewAllocaEndOffset);
2341 V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
2342 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2343 V = getVectorSplat(V, AllocaVecTy->getNumElements());
2345 V = convertValue(DL, IRB, V, AllocaTy);
2348 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2351 DEBUG(dbgs() << " to: " << *New << "\n");
2352 return !II.isVolatile();
2355 bool visitMemTransferInst(MemTransferInst &II) {
2356 // Rewriting of memory transfer instructions can be a bit tricky. We break
2357 // them into two categories: split intrinsics and unsplit intrinsics.
2359 DEBUG(dbgs() << " original: " << II << "\n");
2361 // Compute the intersecting offset range.
2362 assert(BeginOffset < NewAllocaEndOffset);
2363 assert(EndOffset > NewAllocaBeginOffset);
2364 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2365 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2367 assert(II.getRawSource() == OldPtr || II.getRawDest() == OldPtr);
2368 bool IsDest = II.getRawDest() == OldPtr;
2370 // Compute the relative offset within the transfer.
2371 unsigned IntPtrWidth = DL.getPointerSizeInBits();
2372 APInt RelOffset(IntPtrWidth, NewBeginOffset - BeginOffset);
2374 unsigned Align = II.getAlignment();
2375 uint64_t PartitionOffset = NewBeginOffset - NewAllocaBeginOffset;
2378 RelOffset.zextOrTrunc(64).getZExtValue(),
2379 MinAlign(II.getAlignment(), getOffsetAlign(PartitionOffset)));
2381 // For unsplit intrinsics, we simply modify the source and destination
2382 // pointers in place. This isn't just an optimization, it is a matter of
2383 // correctness. With unsplit intrinsics we may be dealing with transfers
2384 // within a single alloca before SROA ran, or with transfers that have
2385 // a variable length. We may also be dealing with memmove instead of
2386 // memcpy, and so simply updating the pointers is the necessary for us to
2387 // update both source and dest of a single call.
2388 if (!IsSplittable) {
2389 Value *OldOp = IsDest ? II.getRawDest() : II.getRawSource();
2392 getAdjustedAllocaPtr(IRB, BeginOffset, II.getRawDest()->getType()));
2394 II.setSource(getAdjustedAllocaPtr(IRB, BeginOffset,
2395 II.getRawSource()->getType()));
2397 Type *CstTy = II.getAlignmentCst()->getType();
2398 II.setAlignment(ConstantInt::get(CstTy, Align));
2400 DEBUG(dbgs() << " to: " << II << "\n");
2401 deleteIfTriviallyDead(OldOp);
2404 // For split transfer intrinsics we have an incredibly useful assurance:
2405 // the source and destination do not reside within the same alloca, and at
2406 // least one of them does not escape. This means that we can replace
2407 // memmove with memcpy, and we don't need to worry about all manner of
2408 // downsides to splitting and transforming the operations.
2410 // If this doesn't map cleanly onto the alloca type, and that type isn't
2411 // a single value type, just emit a memcpy.
2413 = !VecTy && !IntTy && (BeginOffset > NewAllocaBeginOffset ||
2414 EndOffset < NewAllocaEndOffset ||
2415 !NewAI.getAllocatedType()->isSingleValueType());
2417 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2418 // size hasn't been shrunk based on analysis of the viable range, this is
2420 if (EmitMemCpy && &OldAI == &NewAI) {
2421 // Ensure the start lines up.
2422 assert(NewBeginOffset == BeginOffset);
2424 // Rewrite the size as needed.
2425 if (NewEndOffset != EndOffset)
2426 II.setLength(ConstantInt::get(II.getLength()->getType(),
2427 NewEndOffset - NewBeginOffset));
2430 // Record this instruction for deletion.
2431 Pass.DeadInsts.insert(&II);
2433 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2434 // alloca that should be re-examined after rewriting this instruction.
2435 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2437 = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets()))
2438 Pass.Worklist.insert(AI);
2441 Type *OtherPtrTy = IsDest ? II.getRawSource()->getType()
2442 : II.getRawDest()->getType();
2444 // Compute the other pointer, folding as much as possible to produce
2445 // a single, simple GEP in most cases.
2446 OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, RelOffset, OtherPtrTy);
2448 Value *OurPtr = getAdjustedAllocaPtr(
2449 IRB, NewBeginOffset,
2450 IsDest ? II.getRawDest()->getType() : II.getRawSource()->getType());
2451 Type *SizeTy = II.getLength()->getType();
2452 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2454 CallInst *New = IRB.CreateMemCpy(IsDest ? OurPtr : OtherPtr,
2455 IsDest ? OtherPtr : OurPtr,
2456 Size, Align, II.isVolatile());
2458 DEBUG(dbgs() << " to: " << *New << "\n");
2462 // Note that we clamp the alignment to 1 here as a 0 alignment for a memcpy
2463 // is equivalent to 1, but that isn't true if we end up rewriting this as
2468 bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
2469 NewEndOffset == NewAllocaEndOffset;
2470 uint64_t Size = NewEndOffset - NewBeginOffset;
2471 unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
2472 unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
2473 unsigned NumElements = EndIndex - BeginIndex;
2474 IntegerType *SubIntTy
2475 = IntTy ? Type::getIntNTy(IntTy->getContext(), Size*8) : 0;
2477 Type *OtherPtrTy = NewAI.getType();
2478 if (VecTy && !IsWholeAlloca) {
2479 if (NumElements == 1)
2480 OtherPtrTy = VecTy->getElementType();
2482 OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2484 OtherPtrTy = OtherPtrTy->getPointerTo();
2485 } else if (IntTy && !IsWholeAlloca) {
2486 OtherPtrTy = SubIntTy->getPointerTo();
2489 Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, RelOffset, OtherPtrTy);
2490 Value *DstPtr = &NewAI;
2492 std::swap(SrcPtr, DstPtr);
2495 if (VecTy && !IsWholeAlloca && !IsDest) {
2496 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2498 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
2499 } else if (IntTy && !IsWholeAlloca && !IsDest) {
2500 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2502 Src = convertValue(DL, IRB, Src, IntTy);
2503 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2504 Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
2506 Src = IRB.CreateAlignedLoad(SrcPtr, Align, II.isVolatile(),
2510 if (VecTy && !IsWholeAlloca && IsDest) {
2511 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2513 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
2514 } else if (IntTy && !IsWholeAlloca && IsDest) {
2515 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2517 Old = convertValue(DL, IRB, Old, IntTy);
2518 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2519 Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
2520 Src = convertValue(DL, IRB, Src, NewAllocaTy);
2523 StoreInst *Store = cast<StoreInst>(
2524 IRB.CreateAlignedStore(Src, DstPtr, Align, II.isVolatile()));
2526 DEBUG(dbgs() << " to: " << *Store << "\n");
2527 return !II.isVolatile();
2530 bool visitIntrinsicInst(IntrinsicInst &II) {
2531 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2532 II.getIntrinsicID() == Intrinsic::lifetime_end);
2533 DEBUG(dbgs() << " original: " << II << "\n");
2534 assert(II.getArgOperand(1) == OldPtr);
2536 // Compute the intersecting offset range.
2537 assert(BeginOffset < NewAllocaEndOffset);
2538 assert(EndOffset > NewAllocaBeginOffset);
2539 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2540 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2542 // Record this instruction for deletion.
2543 Pass.DeadInsts.insert(&II);
2546 = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2547 NewEndOffset - NewBeginOffset);
2549 getAdjustedAllocaPtr(IRB, NewBeginOffset, II.getArgOperand(1)->getType());
2551 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2552 New = IRB.CreateLifetimeStart(Ptr, Size);
2554 New = IRB.CreateLifetimeEnd(Ptr, Size);
2557 DEBUG(dbgs() << " to: " << *New << "\n");
2561 bool visitPHINode(PHINode &PN) {
2562 DEBUG(dbgs() << " original: " << PN << "\n");
2563 assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
2564 assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
2566 // We would like to compute a new pointer in only one place, but have it be
2567 // as local as possible to the PHI. To do that, we re-use the location of
2568 // the old pointer, which necessarily must be in the right position to
2569 // dominate the PHI.
2570 IRBuilderTy PtrBuilder(cast<Instruction>(OldPtr));
2571 PtrBuilder.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) +
2575 getAdjustedAllocaPtr(PtrBuilder, BeginOffset, OldPtr->getType());
2576 // Replace the operands which were using the old pointer.
2577 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
2579 DEBUG(dbgs() << " to: " << PN << "\n");
2580 deleteIfTriviallyDead(OldPtr);
2582 // Check whether we can speculate this PHI node, and if so remember that
2583 // fact and return that this alloca remains viable for promotion to an SSA
2585 if (isSafePHIToSpeculate(PN, &DL)) {
2586 Pass.SpeculatablePHIs.insert(&PN);
2590 return false; // PHIs can't be promoted on their own.
2593 bool visitSelectInst(SelectInst &SI) {
2594 DEBUG(dbgs() << " original: " << SI << "\n");
2595 assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
2596 "Pointer isn't an operand!");
2597 assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
2598 assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
2600 Value *NewPtr = getAdjustedAllocaPtr(IRB, BeginOffset, OldPtr->getType());
2601 // Replace the operands which were using the old pointer.
2602 if (SI.getOperand(1) == OldPtr)
2603 SI.setOperand(1, NewPtr);
2604 if (SI.getOperand(2) == OldPtr)
2605 SI.setOperand(2, NewPtr);
2607 DEBUG(dbgs() << " to: " << SI << "\n");
2608 deleteIfTriviallyDead(OldPtr);
2610 // Check whether we can speculate this select instruction, and if so
2611 // remember that fact and return that this alloca remains viable for
2612 // promotion to an SSA value.
2613 if (isSafeSelectToSpeculate(SI, &DL)) {
2614 Pass.SpeculatableSelects.insert(&SI);
2618 return false; // Selects can't be promoted on their own.
2625 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
2627 /// This pass aggressively rewrites all aggregate loads and stores on
2628 /// a particular pointer (or any pointer derived from it which we can identify)
2629 /// with scalar loads and stores.
2630 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
2631 // Befriend the base class so it can delegate to private visit methods.
2632 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
2634 const DataLayout &DL;
2636 /// Queue of pointer uses to analyze and potentially rewrite.
2637 SmallVector<Use *, 8> Queue;
2639 /// Set to prevent us from cycling with phi nodes and loops.
2640 SmallPtrSet<User *, 8> Visited;
2642 /// The current pointer use being rewritten. This is used to dig up the used
2643 /// value (as opposed to the user).
2647 AggLoadStoreRewriter(const DataLayout &DL) : DL(DL) {}
2649 /// Rewrite loads and stores through a pointer and all pointers derived from
2651 bool rewrite(Instruction &I) {
2652 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
2654 bool Changed = false;
2655 while (!Queue.empty()) {
2656 U = Queue.pop_back_val();
2657 Changed |= visit(cast<Instruction>(U->getUser()));
2663 /// Enqueue all the users of the given instruction for further processing.
2664 /// This uses a set to de-duplicate users.
2665 void enqueueUsers(Instruction &I) {
2666 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE;
2668 if (Visited.insert(*UI))
2669 Queue.push_back(&UI.getUse());
2672 // Conservative default is to not rewrite anything.
2673 bool visitInstruction(Instruction &I) { return false; }
2675 /// \brief Generic recursive split emission class.
2676 template <typename Derived>
2679 /// The builder used to form new instructions.
2681 /// The indices which to be used with insert- or extractvalue to select the
2682 /// appropriate value within the aggregate.
2683 SmallVector<unsigned, 4> Indices;
2684 /// The indices to a GEP instruction which will move Ptr to the correct slot
2685 /// within the aggregate.
2686 SmallVector<Value *, 4> GEPIndices;
2687 /// The base pointer of the original op, used as a base for GEPing the
2688 /// split operations.
2691 /// Initialize the splitter with an insertion point, Ptr and start with a
2692 /// single zero GEP index.
2693 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
2694 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
2697 /// \brief Generic recursive split emission routine.
2699 /// This method recursively splits an aggregate op (load or store) into
2700 /// scalar or vector ops. It splits recursively until it hits a single value
2701 /// and emits that single value operation via the template argument.
2703 /// The logic of this routine relies on GEPs and insertvalue and
2704 /// extractvalue all operating with the same fundamental index list, merely
2705 /// formatted differently (GEPs need actual values).
2707 /// \param Ty The type being split recursively into smaller ops.
2708 /// \param Agg The aggregate value being built up or stored, depending on
2709 /// whether this is splitting a load or a store respectively.
2710 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
2711 if (Ty->isSingleValueType())
2712 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
2714 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2715 unsigned OldSize = Indices.size();
2717 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
2719 assert(Indices.size() == OldSize && "Did not return to the old size");
2720 Indices.push_back(Idx);
2721 GEPIndices.push_back(IRB.getInt32(Idx));
2722 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
2723 GEPIndices.pop_back();
2729 if (StructType *STy = dyn_cast<StructType>(Ty)) {
2730 unsigned OldSize = Indices.size();
2732 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
2734 assert(Indices.size() == OldSize && "Did not return to the old size");
2735 Indices.push_back(Idx);
2736 GEPIndices.push_back(IRB.getInt32(Idx));
2737 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
2738 GEPIndices.pop_back();
2744 llvm_unreachable("Only arrays and structs are aggregate loadable types");
2748 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
2749 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2750 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
2752 /// Emit a leaf load of a single value. This is called at the leaves of the
2753 /// recursive emission to actually load values.
2754 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2755 assert(Ty->isSingleValueType());
2756 // Load the single value and insert it using the indices.
2757 Value *GEP = IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep");
2758 Value *Load = IRB.CreateLoad(GEP, Name + ".load");
2759 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
2760 DEBUG(dbgs() << " to: " << *Load << "\n");
2764 bool visitLoadInst(LoadInst &LI) {
2765 assert(LI.getPointerOperand() == *U);
2766 if (!LI.isSimple() || LI.getType()->isSingleValueType())
2769 // We have an aggregate being loaded, split it apart.
2770 DEBUG(dbgs() << " original: " << LI << "\n");
2771 LoadOpSplitter Splitter(&LI, *U);
2772 Value *V = UndefValue::get(LI.getType());
2773 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
2774 LI.replaceAllUsesWith(V);
2775 LI.eraseFromParent();
2779 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
2780 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2781 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
2783 /// Emit a leaf store of a single value. This is called at the leaves of the
2784 /// recursive emission to actually produce stores.
2785 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2786 assert(Ty->isSingleValueType());
2787 // Extract the single value and store it using the indices.
2788 Value *Store = IRB.CreateStore(
2789 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
2790 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
2792 DEBUG(dbgs() << " to: " << *Store << "\n");
2796 bool visitStoreInst(StoreInst &SI) {
2797 if (!SI.isSimple() || SI.getPointerOperand() != *U)
2799 Value *V = SI.getValueOperand();
2800 if (V->getType()->isSingleValueType())
2803 // We have an aggregate being stored, split it apart.
2804 DEBUG(dbgs() << " original: " << SI << "\n");
2805 StoreOpSplitter Splitter(&SI, *U);
2806 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
2807 SI.eraseFromParent();
2811 bool visitBitCastInst(BitCastInst &BC) {
2816 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
2821 bool visitPHINode(PHINode &PN) {
2826 bool visitSelectInst(SelectInst &SI) {
2833 /// \brief Strip aggregate type wrapping.
2835 /// This removes no-op aggregate types wrapping an underlying type. It will
2836 /// strip as many layers of types as it can without changing either the type
2837 /// size or the allocated size.
2838 static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
2839 if (Ty->isSingleValueType())
2842 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
2843 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
2846 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
2847 InnerTy = ArrTy->getElementType();
2848 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
2849 const StructLayout *SL = DL.getStructLayout(STy);
2850 unsigned Index = SL->getElementContainingOffset(0);
2851 InnerTy = STy->getElementType(Index);
2856 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
2857 TypeSize > DL.getTypeSizeInBits(InnerTy))
2860 return stripAggregateTypeWrapping(DL, InnerTy);
2863 /// \brief Try to find a partition of the aggregate type passed in for a given
2864 /// offset and size.
2866 /// This recurses through the aggregate type and tries to compute a subtype
2867 /// based on the offset and size. When the offset and size span a sub-section
2868 /// of an array, it will even compute a new array type for that sub-section,
2869 /// and the same for structs.
2871 /// Note that this routine is very strict and tries to find a partition of the
2872 /// type which produces the *exact* right offset and size. It is not forgiving
2873 /// when the size or offset cause either end of type-based partition to be off.
2874 /// Also, this is a best-effort routine. It is reasonable to give up and not
2875 /// return a type if necessary.
2876 static Type *getTypePartition(const DataLayout &DL, Type *Ty,
2877 uint64_t Offset, uint64_t Size) {
2878 if (Offset == 0 && DL.getTypeAllocSize(Ty) == Size)
2879 return stripAggregateTypeWrapping(DL, Ty);
2880 if (Offset > DL.getTypeAllocSize(Ty) ||
2881 (DL.getTypeAllocSize(Ty) - Offset) < Size)
2884 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
2885 // We can't partition pointers...
2886 if (SeqTy->isPointerTy())
2889 Type *ElementTy = SeqTy->getElementType();
2890 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
2891 uint64_t NumSkippedElements = Offset / ElementSize;
2892 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy)) {
2893 if (NumSkippedElements >= ArrTy->getNumElements())
2895 } else if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy)) {
2896 if (NumSkippedElements >= VecTy->getNumElements())
2899 Offset -= NumSkippedElements * ElementSize;
2901 // First check if we need to recurse.
2902 if (Offset > 0 || Size < ElementSize) {
2903 // Bail if the partition ends in a different array element.
2904 if ((Offset + Size) > ElementSize)
2906 // Recurse through the element type trying to peel off offset bytes.
2907 return getTypePartition(DL, ElementTy, Offset, Size);
2909 assert(Offset == 0);
2911 if (Size == ElementSize)
2912 return stripAggregateTypeWrapping(DL, ElementTy);
2913 assert(Size > ElementSize);
2914 uint64_t NumElements = Size / ElementSize;
2915 if (NumElements * ElementSize != Size)
2917 return ArrayType::get(ElementTy, NumElements);
2920 StructType *STy = dyn_cast<StructType>(Ty);
2924 const StructLayout *SL = DL.getStructLayout(STy);
2925 if (Offset >= SL->getSizeInBytes())
2927 uint64_t EndOffset = Offset + Size;
2928 if (EndOffset > SL->getSizeInBytes())
2931 unsigned Index = SL->getElementContainingOffset(Offset);
2932 Offset -= SL->getElementOffset(Index);
2934 Type *ElementTy = STy->getElementType(Index);
2935 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
2936 if (Offset >= ElementSize)
2937 return 0; // The offset points into alignment padding.
2939 // See if any partition must be contained by the element.
2940 if (Offset > 0 || Size < ElementSize) {
2941 if ((Offset + Size) > ElementSize)
2943 return getTypePartition(DL, ElementTy, Offset, Size);
2945 assert(Offset == 0);
2947 if (Size == ElementSize)
2948 return stripAggregateTypeWrapping(DL, ElementTy);
2950 StructType::element_iterator EI = STy->element_begin() + Index,
2951 EE = STy->element_end();
2952 if (EndOffset < SL->getSizeInBytes()) {
2953 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
2954 if (Index == EndIndex)
2955 return 0; // Within a single element and its padding.
2957 // Don't try to form "natural" types if the elements don't line up with the
2959 // FIXME: We could potentially recurse down through the last element in the
2960 // sub-struct to find a natural end point.
2961 if (SL->getElementOffset(EndIndex) != EndOffset)
2964 assert(Index < EndIndex);
2965 EE = STy->element_begin() + EndIndex;
2968 // Try to build up a sub-structure.
2969 StructType *SubTy = StructType::get(STy->getContext(), makeArrayRef(EI, EE),
2971 const StructLayout *SubSL = DL.getStructLayout(SubTy);
2972 if (Size != SubSL->getSizeInBytes())
2973 return 0; // The sub-struct doesn't have quite the size needed.
2978 /// \brief Rewrite an alloca partition's users.
2980 /// This routine drives both of the rewriting goals of the SROA pass. It tries
2981 /// to rewrite uses of an alloca partition to be conducive for SSA value
2982 /// promotion. If the partition needs a new, more refined alloca, this will
2983 /// build that new alloca, preserving as much type information as possible, and
2984 /// rewrite the uses of the old alloca to point at the new one and have the
2985 /// appropriate new offsets. It also evaluates how successful the rewrite was
2986 /// at enabling promotion and if it was successful queues the alloca to be
2988 bool SROA::rewritePartitions(AllocaInst &AI, AllocaPartitioning &P,
2989 AllocaPartitioning::iterator B,
2990 AllocaPartitioning::iterator E,
2991 int64_t BeginOffset, int64_t EndOffset,
2992 ArrayRef<AllocaPartitioning::iterator> SplitUses) {
2993 assert(BeginOffset < EndOffset);
2994 uint64_t PartitionSize = EndOffset - BeginOffset;
2996 // Try to compute a friendly type for this partition of the alloca. This
2997 // won't always succeed, in which case we fall back to a legal integer type
2998 // or an i8 array of an appropriate size.
2999 Type *PartitionTy = 0;
3000 if (Type *CommonUseTy = findCommonType(B, E, EndOffset))
3001 if (DL->getTypeAllocSize(CommonUseTy) >= PartitionSize)
3002 PartitionTy = CommonUseTy;
3004 if (Type *TypePartitionTy = getTypePartition(*DL, AI.getAllocatedType(),
3005 BeginOffset, PartitionSize))
3006 PartitionTy = TypePartitionTy;
3007 if ((!PartitionTy || (PartitionTy->isArrayTy() &&
3008 PartitionTy->getArrayElementType()->isIntegerTy())) &&
3009 DL->isLegalInteger(PartitionSize * 8))
3010 PartitionTy = Type::getIntNTy(*C, PartitionSize * 8);
3012 PartitionTy = ArrayType::get(Type::getInt8Ty(*C), PartitionSize);
3013 assert(DL->getTypeAllocSize(PartitionTy) >= PartitionSize);
3015 bool IsVectorPromotable = isVectorPromotionViable(
3016 *DL, PartitionTy, P, BeginOffset, EndOffset, B, E, SplitUses);
3018 bool IsIntegerPromotable =
3019 !IsVectorPromotable &&
3020 isIntegerWideningViable(*DL, PartitionTy, BeginOffset, P, B, E,
3023 // Check for the case where we're going to rewrite to a new alloca of the
3024 // exact same type as the original, and with the same access offsets. In that
3025 // case, re-use the existing alloca, but still run through the rewriter to
3026 // perform phi and select speculation.
3028 if (PartitionTy == AI.getAllocatedType()) {
3029 assert(BeginOffset == 0 &&
3030 "Non-zero begin offset but same alloca type");
3032 // FIXME: We should be able to bail at this point with "nothing changed".
3033 // FIXME: We might want to defer PHI speculation until after here.
3035 unsigned Alignment = AI.getAlignment();
3037 // The minimum alignment which users can rely on when the explicit
3038 // alignment is omitted or zero is that required by the ABI for this
3040 Alignment = DL->getABITypeAlignment(AI.getAllocatedType());
3042 Alignment = MinAlign(Alignment, BeginOffset);
3043 // If we will get at least this much alignment from the type alone, leave
3044 // the alloca's alignment unconstrained.
3045 if (Alignment <= DL->getABITypeAlignment(PartitionTy))
3047 NewAI = new AllocaInst(PartitionTy, 0, Alignment,
3048 AI.getName() + ".sroa." + Twine(B - P.begin()), &AI);
3052 DEBUG(dbgs() << "Rewriting alloca partition "
3053 << "[" << BeginOffset << "," << EndOffset << ") to: " << *NewAI
3056 // Track the high watermark on several worklists that are only relevant for
3057 // promoted allocas. We will reset it to this point if the alloca is not in
3058 // fact scheduled for promotion.
3059 unsigned PPWOldSize = PostPromotionWorklist.size();
3060 unsigned SPOldSize = SpeculatablePHIs.size();
3061 unsigned SSOldSize = SpeculatableSelects.size();
3063 AllocaPartitionRewriter Rewriter(*DL, P, *this, AI, *NewAI, BeginOffset,
3064 EndOffset, IsVectorPromotable,
3065 IsIntegerPromotable);
3066 bool Promotable = true;
3067 for (ArrayRef<AllocaPartitioning::iterator>::const_iterator
3068 SUI = SplitUses.begin(),
3069 SUE = SplitUses.end();
3070 SUI != SUE; ++SUI) {
3071 DEBUG(dbgs() << " rewriting split ");
3072 DEBUG(P.printPartition(dbgs(), *SUI, ""));
3073 Promotable &= Rewriter.visit(*SUI);
3075 for (AllocaPartitioning::iterator I = B; I != E; ++I) {
3076 DEBUG(dbgs() << " rewriting ");
3077 DEBUG(P.printPartition(dbgs(), I, ""));
3078 Promotable &= Rewriter.visit(I);
3081 if (Promotable && (SpeculatablePHIs.size() > SPOldSize ||
3082 SpeculatableSelects.size() > SSOldSize)) {
3083 // If we have a promotable alloca except for some unspeculated loads below
3084 // PHIs or Selects, iterate once. We will speculate the loads and on the
3085 // next iteration rewrite them into a promotable form.
3086 Worklist.insert(NewAI);
3087 } else if (Promotable) {
3088 DEBUG(dbgs() << " and queuing for promotion\n");
3089 PromotableAllocas.push_back(NewAI);
3090 } else if (NewAI != &AI) {
3091 // If we can't promote the alloca, iterate on it to check for new
3092 // refinements exposed by splitting the current alloca. Don't iterate on an
3093 // alloca which didn't actually change and didn't get promoted.
3094 // FIXME: We should actually track whether the rewriter changed anything.
3095 Worklist.insert(NewAI);
3098 // Drop any post-promotion work items if promotion didn't happen.
3100 while (PostPromotionWorklist.size() > PPWOldSize)
3101 PostPromotionWorklist.pop_back();
3102 while (SpeculatablePHIs.size() > SPOldSize)
3103 SpeculatablePHIs.pop_back();
3104 while (SpeculatableSelects.size() > SSOldSize)
3105 SpeculatableSelects.pop_back();
3112 struct IsPartitionEndLessOrEqualTo {
3113 uint64_t UpperBound;
3115 IsPartitionEndLessOrEqualTo(uint64_t UpperBound) : UpperBound(UpperBound) {}
3117 bool operator()(const AllocaPartitioning::iterator &I) {
3118 return I->endOffset() <= UpperBound;
3123 static void removeFinishedSplitUses(
3124 SmallVectorImpl<AllocaPartitioning::iterator> &SplitUses,
3125 uint64_t &MaxSplitUseEndOffset, uint64_t Offset) {
3126 if (Offset >= MaxSplitUseEndOffset) {
3128 MaxSplitUseEndOffset = 0;
3132 size_t SplitUsesOldSize = SplitUses.size();
3133 SplitUses.erase(std::remove_if(SplitUses.begin(), SplitUses.end(),
3134 IsPartitionEndLessOrEqualTo(Offset)),
3136 if (SplitUsesOldSize == SplitUses.size())
3139 // Recompute the max. While this is linear, so is remove_if.
3140 MaxSplitUseEndOffset = 0;
3141 for (SmallVectorImpl<AllocaPartitioning::iterator>::iterator
3142 SUI = SplitUses.begin(),
3143 SUE = SplitUses.end();
3145 MaxSplitUseEndOffset = std::max((*SUI)->endOffset(), MaxSplitUseEndOffset);
3148 /// \brief Walks the partitioning of an alloca rewriting uses of each partition.
3149 bool SROA::splitAlloca(AllocaInst &AI, AllocaPartitioning &P) {
3150 if (P.begin() == P.end())
3153 bool Changed = false;
3154 SmallVector<AllocaPartitioning::iterator, 4> SplitUses;
3155 uint64_t MaxSplitUseEndOffset = 0;
3157 uint64_t BeginOffset = P.begin()->beginOffset();
3159 for (AllocaPartitioning::iterator PI = P.begin(), PJ = llvm::next(PI),
3161 PI != PE; PI = PJ) {
3162 uint64_t MaxEndOffset = PI->endOffset();
3164 if (!PI->isSplittable()) {
3165 // When we're forming an unsplittable region, it must always start at he
3166 // first partitioning use and will extend through its end.
3167 assert(BeginOffset == PI->beginOffset());
3169 // Rewrite a partition including all of the overlapping uses with this
3170 // unsplittable partition.
3171 while (PJ != PE && PJ->beginOffset() < MaxEndOffset) {
3172 if (!PJ->isSplittable())
3173 MaxEndOffset = std::max(MaxEndOffset, PJ->endOffset());
3177 assert(PI->isSplittable()); // Established above.
3179 // Collect all of the overlapping splittable partitions.
3180 while (PJ != PE && PJ->beginOffset() < MaxEndOffset &&
3181 PJ->isSplittable()) {
3182 MaxEndOffset = std::max(MaxEndOffset, PJ->endOffset());
3186 // Back up MaxEndOffset and PJ if we ended the span early when
3187 // encountering an unsplittable partition.
3188 if (PJ != PE && PJ->beginOffset() < MaxEndOffset) {
3189 assert(!PJ->isSplittable());
3190 MaxEndOffset = PJ->beginOffset();
3194 // Check if we have managed to move the end offset forward yet. If so,
3195 // we'll have to rewrite uses and erase old split uses.
3196 if (BeginOffset < MaxEndOffset) {
3197 // Rewrite a sequence of overlapping partition uses.
3198 Changed |= rewritePartitions(AI, P, PI, PJ, BeginOffset,
3199 MaxEndOffset, SplitUses);
3201 removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset, MaxEndOffset);
3204 // Accumulate all the splittable partitions from the [PI,PJ) region which
3205 // overlap going forward.
3206 for (AllocaPartitioning::iterator PII = PI, PIE = PJ; PII != PIE; ++PII)
3207 if (PII->isSplittable() && PII->endOffset() > MaxEndOffset) {
3208 SplitUses.push_back(PII);
3209 MaxSplitUseEndOffset = std::max(PII->endOffset(), MaxSplitUseEndOffset);
3212 // If we're already at the end and we have no split uses, we're done.
3213 if (PJ == PE && SplitUses.empty())
3216 // If we have no split uses or no gap in offsets, we're ready to move to
3217 // the next partitioning use.
3218 if (SplitUses.empty() || (PJ != PE && MaxEndOffset == PJ->beginOffset())) {
3219 BeginOffset = PJ->beginOffset();
3223 // Even if we have split uses, if the next partitioning use is splittable
3224 // and the split uses reach it, we can simply set up the beginning offset
3225 // to bridge between them.
3226 if (PJ != PE && PJ->isSplittable() && MaxSplitUseEndOffset > PJ->beginOffset()) {
3227 BeginOffset = MaxEndOffset;
3231 // Otherwise, we have a tail of split uses. Rewrite them with an empty
3232 // range of partitioning uses.
3233 uint64_t PostSplitEndOffset =
3234 PJ == PE ? MaxSplitUseEndOffset : PJ->beginOffset();
3236 Changed |= rewritePartitions(AI, P, PJ, PJ, MaxEndOffset,
3237 PostSplitEndOffset, SplitUses);
3239 break; // Skip the rest, we don't need to do any cleanup.
3241 removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset,
3242 PostSplitEndOffset);
3244 // Now just reset the begin offset for the next iteration.
3245 BeginOffset = PJ->beginOffset();
3251 /// \brief Analyze an alloca for SROA.
3253 /// This analyzes the alloca to ensure we can reason about it, builds
3254 /// a partitioning of the alloca, and then hands it off to be split and
3255 /// rewritten as needed.
3256 bool SROA::runOnAlloca(AllocaInst &AI) {
3257 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3258 ++NumAllocasAnalyzed;
3260 // Special case dead allocas, as they're trivial.
3261 if (AI.use_empty()) {
3262 AI.eraseFromParent();
3266 // Skip alloca forms that this analysis can't handle.
3267 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3268 DL->getTypeAllocSize(AI.getAllocatedType()) == 0)
3271 bool Changed = false;
3273 // First, split any FCA loads and stores touching this alloca to promote
3274 // better splitting and promotion opportunities.
3275 AggLoadStoreRewriter AggRewriter(*DL);
3276 Changed |= AggRewriter.rewrite(AI);
3278 // Build the partition set using a recursive instruction-visiting builder.
3279 AllocaPartitioning P(*DL, AI);
3280 DEBUG(P.print(dbgs()));
3284 // Delete all the dead users of this alloca before splitting and rewriting it.
3285 for (AllocaPartitioning::dead_user_iterator DI = P.dead_user_begin(),
3286 DE = P.dead_user_end();
3289 (*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType()));
3290 DeadInsts.insert(*DI);
3292 for (AllocaPartitioning::dead_op_iterator DO = P.dead_op_begin(),
3293 DE = P.dead_op_end();
3296 // Clobber the use with an undef value.
3297 **DO = UndefValue::get(OldV->getType());
3298 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3299 if (isInstructionTriviallyDead(OldI)) {
3301 DeadInsts.insert(OldI);
3305 // No partitions to split. Leave the dead alloca for a later pass to clean up.
3306 if (P.begin() == P.end())
3309 Changed |= splitAlloca(AI, P);
3311 DEBUG(dbgs() << " Speculating PHIs\n");
3312 while (!SpeculatablePHIs.empty())
3313 speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
3315 DEBUG(dbgs() << " Speculating Selects\n");
3316 while (!SpeculatableSelects.empty())
3317 speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
3322 /// \brief Delete the dead instructions accumulated in this run.
3324 /// Recursively deletes the dead instructions we've accumulated. This is done
3325 /// at the very end to maximize locality of the recursive delete and to
3326 /// minimize the problems of invalidated instruction pointers as such pointers
3327 /// are used heavily in the intermediate stages of the algorithm.
3329 /// We also record the alloca instructions deleted here so that they aren't
3330 /// subsequently handed to mem2reg to promote.
3331 void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
3332 while (!DeadInsts.empty()) {
3333 Instruction *I = DeadInsts.pop_back_val();
3334 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3336 I->replaceAllUsesWith(UndefValue::get(I->getType()));
3338 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
3339 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
3340 // Zero out the operand and see if it becomes trivially dead.
3342 if (isInstructionTriviallyDead(U))
3343 DeadInsts.insert(U);
3346 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3347 DeletedAllocas.insert(AI);
3350 I->eraseFromParent();
3354 /// \brief Promote the allocas, using the best available technique.
3356 /// This attempts to promote whatever allocas have been identified as viable in
3357 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
3358 /// If there is a domtree available, we attempt to promote using the full power
3359 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3360 /// based on the SSAUpdater utilities. This function returns whether any
3361 /// promotion occurred.
3362 bool SROA::promoteAllocas(Function &F) {
3363 if (PromotableAllocas.empty())
3366 NumPromoted += PromotableAllocas.size();
3368 if (DT && !ForceSSAUpdater) {
3369 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3370 PromoteMemToReg(PromotableAllocas, *DT);
3371 PromotableAllocas.clear();
3375 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3377 DIBuilder DIB(*F.getParent());
3378 SmallVector<Instruction*, 64> Insts;
3380 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3381 AllocaInst *AI = PromotableAllocas[Idx];
3382 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
3384 Instruction *I = cast<Instruction>(*UI++);
3385 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3386 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3387 // leading to them) here. Eventually it should use them to optimize the
3388 // scalar values produced.
3389 if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
3390 assert(onlyUsedByLifetimeMarkers(I) &&
3391 "Found a bitcast used outside of a lifetime marker.");
3392 while (!I->use_empty())
3393 cast<Instruction>(*I->use_begin())->eraseFromParent();
3394 I->eraseFromParent();
3397 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3398 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3399 II->getIntrinsicID() == Intrinsic::lifetime_end);
3400 II->eraseFromParent();
3406 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3410 PromotableAllocas.clear();
3415 /// \brief A predicate to test whether an alloca belongs to a set.
3416 class IsAllocaInSet {
3417 typedef SmallPtrSet<AllocaInst *, 4> SetType;
3421 typedef AllocaInst *argument_type;
3423 IsAllocaInSet(const SetType &Set) : Set(Set) {}
3424 bool operator()(AllocaInst *AI) const { return Set.count(AI); }
3428 bool SROA::runOnFunction(Function &F) {
3429 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3430 C = &F.getContext();
3431 DL = getAnalysisIfAvailable<DataLayout>();
3433 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3436 DT = getAnalysisIfAvailable<DominatorTree>();
3438 BasicBlock &EntryBB = F.getEntryBlock();
3439 for (BasicBlock::iterator I = EntryBB.begin(), E = llvm::prior(EntryBB.end());
3441 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3442 Worklist.insert(AI);
3444 bool Changed = false;
3445 // A set of deleted alloca instruction pointers which should be removed from
3446 // the list of promotable allocas.
3447 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3450 while (!Worklist.empty()) {
3451 Changed |= runOnAlloca(*Worklist.pop_back_val());
3452 deleteDeadInstructions(DeletedAllocas);
3454 // Remove the deleted allocas from various lists so that we don't try to
3455 // continue processing them.
3456 if (!DeletedAllocas.empty()) {
3457 Worklist.remove_if(IsAllocaInSet(DeletedAllocas));
3458 PostPromotionWorklist.remove_if(IsAllocaInSet(DeletedAllocas));
3459 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3460 PromotableAllocas.end(),
3461 IsAllocaInSet(DeletedAllocas)),
3462 PromotableAllocas.end());
3463 DeletedAllocas.clear();
3467 Changed |= promoteAllocas(F);
3469 Worklist = PostPromotionWorklist;
3470 PostPromotionWorklist.clear();
3471 } while (!Worklist.empty());
3476 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3477 if (RequiresDomTree)
3478 AU.addRequired<DominatorTree>();
3479 AU.setPreservesCFG();