#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
-#include "llvm/Analysis/Dominators.h"
#include "llvm/Analysis/Loads.h"
#include "llvm/Analysis/PtrUseVisitor.h"
#include "llvm/Analysis/ValueTracking.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/DerivedTypes.h"
+#include "llvm/IR/Dominators.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/Instructions.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/MathExtras.h"
+#include "llvm/Support/TimeValue.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Transforms/Utils/Local.h"
#include "llvm/Transforms/Utils/PromoteMemToReg.h"
#include "llvm/Transforms/Utils/SSAUpdater.h"
+
+#if __cplusplus >= 201103L && !defined(NDEBUG)
+// We only use this for a debug check in C++11
+#include <random>
+#endif
+
using namespace llvm;
STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
static cl::opt<bool>
ForceSSAUpdater("force-ssa-updater", cl::init(false), cl::Hidden);
+/// Hidden option to enable randomly shuffling the slices to help uncover
+/// instability in their order.
+static cl::opt<bool> SROARandomShuffleSlices("sroa-random-shuffle-slices",
+ cl::init(false), cl::Hidden);
+
namespace {
/// \brief A custom IRBuilder inserter which prefixes all names if they are
/// preserved.
void visitMemTransferInst(MemTransferInst &II) {
ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
- if ((Length && Length->getValue() == 0) ||
- (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize)))
+ if (Length && Length->getValue() == 0)
// Zero-length mem transfer intrinsics can be ignored entirely.
return markAsDead(II);
+ // Because we can visit these intrinsics twice, also check to see if the
+ // first time marked this instruction as dead. If so, skip it.
+ if (VisitedDeadInsts.count(&II))
+ return;
+
if (!IsOffsetKnown)
return PI.setAborted(&II);
+ // This side of the transfer is completely out-of-bounds, and so we can
+ // nuke the entire transfer. However, we also need to nuke the other side
+ // if already added to our partitions.
+ // FIXME: Yet another place we really should bypass this when
+ // instrumenting for ASan.
+ if (!Offset.isNegative() && Offset.uge(AllocSize)) {
+ SmallDenseMap<Instruction *, unsigned>::iterator MTPI = MemTransferSliceMap.find(&II);
+ if (MTPI != MemTransferSliceMap.end())
+ S.Slices[MTPI->second].kill();
+ return markAsDead(II);
+ }
+
uint64_t RawOffset = Offset.getLimitedValue();
uint64_t Size = Length ? Length->getLimitedValue()
: AllocSize - RawOffset;
std::mem_fun_ref(&Slice::isDead)),
Slices.end());
+#if __cplusplus >= 201103L && !defined(NDEBUG)
+ if (SROARandomShuffleSlices) {
+ std::mt19937 MT(static_cast<unsigned>(sys::TimeValue::now().msec()));
+ std::shuffle(Slices.begin(), Slices.end(), MT);
+ }
+#endif
+
// Sort the uses. This arranges for the offsets to be in ascending order,
// and the sizes to be in descending order.
std::sort(Slices.begin(), Slices.end());
ArrayRef<AllocaSlices::iterator> SplitUses);
bool splitAlloca(AllocaInst &AI, AllocaSlices &S);
bool runOnAlloca(AllocaInst &AI);
+ void clobberUse(Use &U);
void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas);
bool promoteAllocas(Function &F);
};
INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates",
false, false)
-INITIALIZE_PASS_DEPENDENCY(DominatorTree)
+INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
false, false)
AllocaSlices::const_iterator E,
uint64_t EndOffset) {
Type *Ty = 0;
- bool IgnoreNonIntegralTypes = false;
+ bool TyIsCommon = true;
+ IntegerType *ITy = 0;
+
+ // Note that we need to look at *every* alloca slice's Use to ensure we
+ // always get consistent results regardless of the order of slices.
for (AllocaSlices::const_iterator I = B; I != E; ++I) {
Use *U = I->getUse();
if (isa<IntrinsicInst>(*U->getUser()))
UserTy = LI->getType();
} else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
UserTy = SI->getValueOperand()->getType();
- } else {
- IgnoreNonIntegralTypes = true; // Give up on anything but an iN type.
- continue;
}
- if (IntegerType *ITy = dyn_cast<IntegerType>(UserTy)) {
+ if (!UserTy || (Ty && Ty != UserTy))
+ TyIsCommon = false; // Give up on anything but an iN type.
+ else
+ Ty = UserTy;
+
+ if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
// If the type is larger than the partition, skip it. We only encounter
// this for split integer operations where we want to use the type of the
// entity causing the split. Also skip if the type is not a byte width
// multiple.
- if (ITy->getBitWidth() % 8 != 0 ||
- ITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
+ if (UserITy->getBitWidth() % 8 != 0 ||
+ UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
continue;
- // If we have found an integer type use covering the alloca, use that
- // regardless of the other types, as integers are often used for
- // a "bucket of bits" type.
- //
- // NB: This *must* be the only return from inside the loop so that the
- // order of slices doesn't impact the computed type.
- return ITy;
- } else if (IgnoreNonIntegralTypes) {
- continue;
+ // Track the largest bitwidth integer type used in this way in case there
+ // is no common type.
+ if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
+ ITy = UserITy;
}
-
- if (Ty && Ty != UserTy)
- IgnoreNonIntegralTypes = true; // Give up on anything but an iN type.
-
- Ty = UserTy;
}
- return Ty;
+
+ return TyIsCommon ? Ty : ITy;
}
/// PHI instructions that use an alloca and are subsequently loaded can be
/// This will return the BasePtr if that is valid, or build a new GEP
/// instruction using the IRBuilder if GEP-ing is needed.
static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
- SmallVectorImpl<Value *> &Indices) {
+ SmallVectorImpl<Value *> &Indices, Twine NamePrefix) {
if (Indices.empty())
return BasePtr;
if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
return BasePtr;
- return IRB.CreateInBoundsGEP(BasePtr, Indices, "idx");
+ return IRB.CreateInBoundsGEP(BasePtr, Indices, NamePrefix + "sroa_idx");
}
/// \brief Get a natural GEP off of the BasePtr walking through Ty toward
/// indicated by Indices to have the correct offset.
static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
Value *BasePtr, Type *Ty, Type *TargetTy,
- SmallVectorImpl<Value *> &Indices) {
+ SmallVectorImpl<Value *> &Indices,
+ Twine NamePrefix) {
if (Ty == TargetTy)
- return buildGEP(IRB, BasePtr, Indices);
+ return buildGEP(IRB, BasePtr, Indices, NamePrefix);
// See if we can descend into a struct and locate a field with the correct
// type.
if (ElementTy != TargetTy)
Indices.erase(Indices.end() - NumLayers, Indices.end());
- return buildGEP(IRB, BasePtr, Indices);
+ return buildGEP(IRB, BasePtr, Indices, NamePrefix);
}
/// \brief Recursively compute indices for a natural GEP.
static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
Value *Ptr, Type *Ty, APInt &Offset,
Type *TargetTy,
- SmallVectorImpl<Value *> &Indices) {
+ SmallVectorImpl<Value *> &Indices,
+ Twine NamePrefix) {
if (Offset == 0)
- return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices);
+ return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices, NamePrefix);
// We can't recurse through pointer types.
if (Ty->isPointerTy())
Offset -= NumSkippedElements * ElementSize;
Indices.push_back(IRB.getInt(NumSkippedElements));
return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
- Offset, TargetTy, Indices);
+ Offset, TargetTy, Indices, NamePrefix);
}
if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
Offset -= NumSkippedElements * ElementSize;
Indices.push_back(IRB.getInt(NumSkippedElements));
return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
- Indices);
+ Indices, NamePrefix);
}
StructType *STy = dyn_cast<StructType>(Ty);
Indices.push_back(IRB.getInt32(Index));
return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
- Indices);
+ Indices, NamePrefix);
}
/// \brief Get a natural GEP from a base pointer to a particular offset and
/// If no natural GEP can be constructed, this function returns null.
static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
Value *Ptr, APInt Offset, Type *TargetTy,
- SmallVectorImpl<Value *> &Indices) {
+ SmallVectorImpl<Value *> &Indices,
+ Twine NamePrefix) {
PointerType *Ty = cast<PointerType>(Ptr->getType());
// Don't consider any GEPs through an i8* as natural unless the TargetTy is
Offset -= NumSkippedElements * ElementSize;
Indices.push_back(IRB.getInt(NumSkippedElements));
return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
- Indices);
+ Indices, NamePrefix);
}
/// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
/// properties. The algorithm tries to fold as many constant indices into
/// a single GEP as possible, thus making each GEP more independent of the
/// surrounding code.
-static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL,
- Value *Ptr, APInt Offset, Type *PointerTy) {
+static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr,
+ APInt Offset, Type *PointerTy,
+ Twine NamePrefix) {
// Even though we don't look through PHI nodes, we could be called on an
// instruction in an unreachable block, which may be on a cycle.
SmallPtrSet<Value *, 4> Visited;
// See if we can perform a natural GEP here.
Indices.clear();
if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
- Indices)) {
+ Indices, NamePrefix)) {
if (P->getType() == PointerTy) {
// Zap any offset pointer that we ended up computing in previous rounds.
if (OffsetPtr && OffsetPtr->use_empty())
if (!OffsetPtr) {
if (!Int8Ptr) {
Int8Ptr = IRB.CreateBitCast(Ptr, IRB.getInt8PtrTy(),
- "raw_cast");
+ NamePrefix + "sroa_raw_cast");
Int8PtrOffset = Offset;
}
OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
- "raw_idx");
+ NamePrefix + "sroa_raw_idx");
}
Ptr = OffsetPtr;
// On the off chance we were targeting i8*, guard the bitcast here.
if (Ptr->getType() != PointerTy)
- Ptr = IRB.CreateBitCast(Ptr, PointerTy, "cast");
+ Ptr = IRB.CreateBitCast(Ptr, PointerTy, NamePrefix + "sroa_cast");
return Ptr;
}
Use *OldUse;
Instruction *OldPtr;
- // Output members carrying state about the result of visiting and rewriting
- // the slice of the alloca.
- bool IsUsedByRewrittenSpeculatableInstructions;
+ // Track post-rewrite users which are PHI nodes and Selects.
+ SmallPtrSetImpl<PHINode *> &PHIUsers;
+ SmallPtrSetImpl<SelectInst *> &SelectUsers;
// Utility IR builder, whose name prefix is setup for each visited use, and
// the insertion point is set to point to the user.
AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &S, SROA &Pass,
AllocaInst &OldAI, AllocaInst &NewAI,
uint64_t NewBeginOffset, uint64_t NewEndOffset,
- bool IsVectorPromotable = false,
- bool IsIntegerPromotable = false)
+ bool IsVectorPromotable, bool IsIntegerPromotable,
+ SmallPtrSetImpl<PHINode *> &PHIUsers,
+ SmallPtrSetImpl<SelectInst *> &SelectUsers)
: DL(DL), S(S), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
NewAllocaBeginOffset(NewBeginOffset), NewAllocaEndOffset(NewEndOffset),
NewAllocaTy(NewAI.getAllocatedType()),
DL.getTypeSizeInBits(NewAI.getAllocatedType()))
: 0),
BeginOffset(), EndOffset(), IsSplittable(), IsSplit(), OldUse(),
- OldPtr(), IsUsedByRewrittenSpeculatableInstructions(false),
+ OldPtr(), PHIUsers(PHIUsers), SelectUsers(SelectUsers),
IRB(NewAI.getContext(), ConstantFolder()) {
if (VecTy) {
assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
return CanSROA;
}
- /// \brief Query whether this slice is used by speculatable instructions after
- /// rewriting.
- ///
- /// These instructions (PHIs and Selects currently) require the alloca slice
- /// to run back through the rewriter. Thus, they are promotable, but not on
- /// this iteration. This is distinct from a slice which is unpromotable for
- /// some other reason, in which case we don't even want to perform the
- /// speculation. This can be querried at any time and reflects whether (at
- /// that point) a visit call has rewritten a speculatable instruction on the
- /// current slice.
- bool isUsedByRewrittenSpeculatableInstructions() const {
- return IsUsedByRewrittenSpeculatableInstructions;
- }
-
private:
// Make sure the other visit overloads are visible.
using Base::visit;
Value *getAdjustedAllocaPtr(IRBuilderTy &IRB, uint64_t Offset,
Type *PointerTy) {
assert(Offset >= NewAllocaBeginOffset);
+#ifndef NDEBUG
+ StringRef OldName = OldPtr->getName();
+ // Skip through the last '.sroa.' component of the name.
+ size_t LastSROAPrefix = OldName.rfind(".sroa.");
+ if (LastSROAPrefix != StringRef::npos) {
+ OldName = OldName.substr(LastSROAPrefix + strlen(".sroa."));
+ // Look for an SROA slice index.
+ size_t IndexEnd = OldName.find_first_not_of("0123456789");
+ if (IndexEnd != StringRef::npos && OldName[IndexEnd] == '.') {
+ // Strip the index and look for the offset.
+ OldName = OldName.substr(IndexEnd + 1);
+ size_t OffsetEnd = OldName.find_first_not_of("0123456789");
+ if (OffsetEnd != StringRef::npos && OldName[OffsetEnd] == '.')
+ // Strip the offset.
+ OldName = OldName.substr(OffsetEnd + 1);
+ }
+ }
+ // Strip any SROA suffixes as well.
+ OldName = OldName.substr(0, OldName.find(".sroa_"));
+#endif
return getAdjustedPtr(IRB, DL, &NewAI, APInt(DL.getPointerSizeInBits(),
Offset - NewAllocaBeginOffset),
- PointerTy);
+ PointerTy,
+#ifndef NDEBUG
+ Twine(OldName) + "."
+#else
+ Twine()
+#endif
+ );
}
/// \brief Compute suitable alignment to access an offset into the new alloca.
Value *NewPtr = getAdjustedAllocaPtr(IRB, NewBeginOffset,
V->getType()->getPointerTo());
NewSI = IRB.CreateAlignedStore(
- V, NewPtr, getOffsetTypeAlign(
- V->getType(), NewBeginOffset - NewAllocaBeginOffset),
+ V, NewPtr, getOffsetTypeAlign(V->getType(),
+ NewBeginOffset - NewAllocaBeginOffset),
SI.isVolatile());
}
(void)NewSI;
if (!isa<Constant>(II.getLength())) {
assert(!IsSplit);
assert(BeginOffset >= NewAllocaBeginOffset);
- II.setDest(
- getAdjustedAllocaPtr(IRB, BeginOffset, II.getRawDest()->getType()));
+ II.setDest(getAdjustedAllocaPtr(IRB, BeginOffset, OldPtr->getType()));
Type *CstTy = II.getAlignmentCst()->getType();
II.setAlignment(ConstantInt::get(CstTy, getOffsetAlign(BeginOffset)));
Type *SizeTy = II.getLength()->getType();
Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
CallInst *New = IRB.CreateMemSet(
- getAdjustedAllocaPtr(IRB, NewBeginOffset, II.getRawDest()->getType()),
+ getAdjustedAllocaPtr(IRB, NewBeginOffset, OldPtr->getType()),
II.getValue(), Size, getOffsetAlign(SliceOffset), II.isVolatile());
(void)New;
DEBUG(dbgs() << " to: " << *New << "\n");
uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
- assert(II.getRawSource() == OldPtr || II.getRawDest() == OldPtr);
- bool IsDest = II.getRawDest() == OldPtr;
+ bool IsDest = &II.getRawDestUse() == OldUse;
+ assert((IsDest && II.getRawDest() == OldPtr) ||
+ (!IsDest && II.getRawSource() == OldPtr));
// Compute the relative offset within the transfer.
unsigned IntPtrWidth = DL.getPointerSizeInBits();
// memcpy, and so simply updating the pointers is the necessary for us to
// update both source and dest of a single call.
if (!IsSplittable) {
- Value *OldOp = IsDest ? II.getRawDest() : II.getRawSource();
+ Value *AdjustedPtr =
+ getAdjustedAllocaPtr(IRB, BeginOffset, OldPtr->getType());
if (IsDest)
- II.setDest(
- getAdjustedAllocaPtr(IRB, BeginOffset, II.getRawDest()->getType()));
+ II.setDest(AdjustedPtr);
else
- II.setSource(getAdjustedAllocaPtr(IRB, BeginOffset,
- II.getRawSource()->getType()));
+ II.setSource(AdjustedPtr);
Type *CstTy = II.getAlignmentCst()->getType();
II.setAlignment(ConstantInt::get(CstTy, Align));
DEBUG(dbgs() << " to: " << II << "\n");
- deleteIfTriviallyDead(OldOp);
+ deleteIfTriviallyDead(OldPtr);
return false;
}
// For split transfer intrinsics we have an incredibly useful assurance:
// alloca that should be re-examined after rewriting this instruction.
Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
if (AllocaInst *AI
- = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets()))
+ = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
+ assert(AI != &OldAI && AI != &NewAI &&
+ "Splittable transfers cannot reach the same alloca on both ends.");
Pass.Worklist.insert(AI);
+ }
if (EmitMemCpy) {
- Type *OtherPtrTy = IsDest ? II.getRawSource()->getType()
- : II.getRawDest()->getType();
+ Type *OtherPtrTy = OtherPtr->getType();
// Compute the other pointer, folding as much as possible to produce
// a single, simple GEP in most cases.
- OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, RelOffset, OtherPtrTy);
+ OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, RelOffset, OtherPtrTy,
+ OtherPtr->getName() + ".");
- Value *OurPtr = getAdjustedAllocaPtr(
- IRB, NewBeginOffset,
- IsDest ? II.getRawDest()->getType() : II.getRawSource()->getType());
+ Value *OurPtr =
+ getAdjustedAllocaPtr(IRB, NewBeginOffset, OldPtr->getType());
Type *SizeTy = II.getLength()->getType();
Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
OtherPtrTy = SubIntTy->getPointerTo();
}
- Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, RelOffset, OtherPtrTy);
+ Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, RelOffset, OtherPtrTy,
+ OtherPtr->getName() + ".");
Value *DstPtr = &NewAI;
if (!IsDest)
std::swap(SrcPtr, DstPtr);
ConstantInt *Size
= ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
NewEndOffset - NewBeginOffset);
- Value *Ptr =
- getAdjustedAllocaPtr(IRB, NewBeginOffset, II.getArgOperand(1)->getType());
+ Value *Ptr = getAdjustedAllocaPtr(IRB, NewBeginOffset, OldPtr->getType());
Value *New;
if (II.getIntrinsicID() == Intrinsic::lifetime_start)
New = IRB.CreateLifetimeStart(Ptr, Size);
// as local as possible to the PHI. To do that, we re-use the location of
// the old pointer, which necessarily must be in the right position to
// dominate the PHI.
- IRBuilderTy PtrBuilder(OldPtr);
- PtrBuilder.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) +
- ".");
+ IRBuilderTy PtrBuilder(IRB);
+ PtrBuilder.SetInsertPoint(OldPtr);
+ PtrBuilder.SetCurrentDebugLocation(OldPtr->getDebugLoc());
Value *NewPtr =
getAdjustedAllocaPtr(PtrBuilder, BeginOffset, OldPtr->getType());
DEBUG(dbgs() << " to: " << PN << "\n");
deleteIfTriviallyDead(OldPtr);
- // Check whether we can speculate this PHI node, and if so remember that
- // fact and queue it up for another iteration after the speculation
- // occurs.
- if (isSafePHIToSpeculate(PN, &DL)) {
- Pass.SpeculatablePHIs.insert(&PN);
- IsUsedByRewrittenSpeculatableInstructions = true;
- return true;
- }
-
- return false; // PHIs can't be promoted on their own.
+ // PHIs can't be promoted on their own, but often can be speculated. We
+ // check the speculation outside of the rewriter so that we see the
+ // fully-rewritten alloca.
+ PHIUsers.insert(&PN);
+ return true;
}
bool visitSelectInst(SelectInst &SI) {
DEBUG(dbgs() << " to: " << SI << "\n");
deleteIfTriviallyDead(OldPtr);
- // Check whether we can speculate this select instruction, and if so
- // remember that fact and queue it up for another iteration after the
- // speculation occurs.
- if (isSafeSelectToSpeculate(SI, &DL)) {
- Pass.SpeculatableSelects.insert(&SI);
- IsUsedByRewrittenSpeculatableInstructions = true;
- return true;
- }
-
- return false; // Selects can't be promoted on their own.
+ // Selects can't be promoted on their own, but often can be speculated. We
+ // check the speculation outside of the rewriter so that we see the
+ // fully-rewritten alloca.
+ SelectUsers.insert(&SI);
+ return true;
}
};
<< "[" << BeginOffset << "," << EndOffset << ") to: " << *NewAI
<< "\n");
- // Track the high watermark on several worklists that are only relevant for
+ // Track the high watermark on the worklist as it is only relevant for
// promoted allocas. We will reset it to this point if the alloca is not in
// fact scheduled for promotion.
unsigned PPWOldSize = PostPromotionWorklist.size();
- unsigned SPOldSize = SpeculatablePHIs.size();
- unsigned SSOldSize = SpeculatableSelects.size();
unsigned NumUses = 0;
+ SmallPtrSet<PHINode *, 8> PHIUsers;
+ SmallPtrSet<SelectInst *, 8> SelectUsers;
AllocaSliceRewriter Rewriter(*DL, S, *this, AI, *NewAI, BeginOffset,
EndOffset, IsVectorPromotable,
- IsIntegerPromotable);
+ IsIntegerPromotable, PHIUsers, SelectUsers);
bool Promotable = true;
for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
SUE = SplitUses.end();
MaxUsesPerAllocaPartition =
std::max<unsigned>(NumUses, MaxUsesPerAllocaPartition);
- if (Promotable && !Rewriter.isUsedByRewrittenSpeculatableInstructions()) {
- DEBUG(dbgs() << " and queuing for promotion\n");
- PromotableAllocas.push_back(NewAI);
- } else if (NewAI != &AI ||
- (Promotable &&
- Rewriter.isUsedByRewrittenSpeculatableInstructions())) {
+ // Now that we've processed all the slices in the new partition, check if any
+ // PHIs or Selects would block promotion.
+ for (SmallPtrSetImpl<PHINode *>::iterator I = PHIUsers.begin(),
+ E = PHIUsers.end();
+ I != E; ++I)
+ if (!isSafePHIToSpeculate(**I, DL)) {
+ Promotable = false;
+ PHIUsers.clear();
+ SelectUsers.clear();
+ break;
+ }
+ for (SmallPtrSetImpl<SelectInst *>::iterator I = SelectUsers.begin(),
+ E = SelectUsers.end();
+ I != E; ++I)
+ if (!isSafeSelectToSpeculate(**I, DL)) {
+ Promotable = false;
+ PHIUsers.clear();
+ SelectUsers.clear();
+ break;
+ }
+
+ if (Promotable) {
+ if (PHIUsers.empty() && SelectUsers.empty()) {
+ // Promote the alloca.
+ PromotableAllocas.push_back(NewAI);
+ } else {
+ // If we have either PHIs or Selects to speculate, add them to those
+ // worklists and re-queue the new alloca so that we promote in on the
+ // next iteration.
+ for (SmallPtrSetImpl<PHINode *>::iterator I = PHIUsers.begin(),
+ E = PHIUsers.end();
+ I != E; ++I)
+ SpeculatablePHIs.insert(*I);
+ for (SmallPtrSetImpl<SelectInst *>::iterator I = SelectUsers.begin(),
+ E = SelectUsers.end();
+ I != E; ++I)
+ SpeculatableSelects.insert(*I);
+ Worklist.insert(NewAI);
+ }
+ } else {
// If we can't promote the alloca, iterate on it to check for new
// refinements exposed by splitting the current alloca. Don't iterate on an
// alloca which didn't actually change and didn't get promoted.
- //
- // Alternatively, if we could promote the alloca but have speculatable
- // instructions then we will speculate them after finishing our processing
- // of the original alloca. Mark the new one for re-visiting in the next
- // iteration so the speculated operations can be rewritten.
- //
- // FIXME: We should actually track whether the rewriter changed anything.
- Worklist.insert(NewAI);
- }
-
- // Drop any post-promotion work items if promotion didn't happen.
- if (!Promotable) {
+ if (NewAI != &AI)
+ Worklist.insert(NewAI);
+
+ // Drop any post-promotion work items if promotion didn't happen.
while (PostPromotionWorklist.size() > PPWOldSize)
PostPromotionWorklist.pop_back();
- while (SpeculatablePHIs.size() > SPOldSize)
- SpeculatablePHIs.pop_back();
- while (SpeculatableSelects.size() > SSOldSize)
- SpeculatableSelects.pop_back();
}
return true;
return Changed;
}
+/// \brief Clobber a use with undef, deleting the used value if it becomes dead.
+void SROA::clobberUse(Use &U) {
+ Value *OldV = U;
+ // Replace the use with an undef value.
+ U = UndefValue::get(OldV->getType());
+
+ // Check for this making an instruction dead. We have to garbage collect
+ // all the dead instructions to ensure the uses of any alloca end up being
+ // minimal.
+ if (Instruction *OldI = dyn_cast<Instruction>(OldV))
+ if (isInstructionTriviallyDead(OldI)) {
+ DeadInsts.insert(OldI);
+ }
+}
+
/// \brief Analyze an alloca for SROA.
///
/// This analyzes the alloca to ensure we can reason about it, builds
for (AllocaSlices::dead_user_iterator DI = S.dead_user_begin(),
DE = S.dead_user_end();
DI != DE; ++DI) {
- Changed = true;
+ // Free up everything used by this instruction.
+ for (User::op_iterator DOI = (*DI)->op_begin(), DOE = (*DI)->op_end();
+ DOI != DOE; ++DOI)
+ clobberUse(*DOI);
+
+ // Now replace the uses of this instruction.
(*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType()));
+
+ // And mark it for deletion.
DeadInsts.insert(*DI);
+ Changed = true;
}
for (AllocaSlices::dead_op_iterator DO = S.dead_op_begin(),
DE = S.dead_op_end();
DO != DE; ++DO) {
- Value *OldV = **DO;
- // Clobber the use with an undef value.
- **DO = UndefValue::get(OldV->getType());
- if (Instruction *OldI = dyn_cast<Instruction>(OldV))
- if (isInstructionTriviallyDead(OldI)) {
- Changed = true;
- DeadInsts.insert(OldI);
- }
+ clobberUse(**DO);
+ Changed = true;
}
// No slices to split. Leave the dead alloca for a later pass to clean up.
}
bool SROA::runOnFunction(Function &F) {
+ if (skipOptnoneFunction(F))
+ return false;
+
DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
C = &F.getContext();
DL = getAnalysisIfAvailable<DataLayout>();
DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
return false;
}
- DT = getAnalysisIfAvailable<DominatorTree>();
+ DominatorTreeWrapperPass *DTWP =
+ getAnalysisIfAvailable<DominatorTreeWrapperPass>();
+ DT = DTWP ? &DTWP->getDomTree() : 0;
BasicBlock &EntryBB = F.getEntryBlock();
for (BasicBlock::iterator I = EntryBB.begin(), E = llvm::prior(EntryBB.end());
void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
if (RequiresDomTree)
- AU.addRequired<DominatorTree>();
+ AU.addRequired<DominatorTreeWrapperPass>();
AU.setPreservesCFG();
}