#include "llvm/DebugInfo.h"
#include "llvm/DerivedTypes.h"
#include "llvm/Function.h"
-#include "llvm/GlobalVariable.h"
#include "llvm/IRBuilder.h"
#include "llvm/Instructions.h"
#include "llvm/IntrinsicInst.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/STLExtras.h"
-#include "llvm/ADT/TinyPtrVector.h"
#include "llvm/Analysis/Dominators.h"
#include "llvm/Analysis/Loads.h"
#include "llvm/Analysis/ValueTracking.h"
-#include "llvm/Support/CallSite.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/GetElementPtrTypeIterator.h"
#include "llvm/Support/InstVisitor.h"
#include "llvm/Support/MathExtras.h"
-#include "llvm/Support/ValueHandle.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Transforms/Utils/Local.h"
/// memory use itself with a particular partition. As a consequence there is
/// intentionally overlap between various uses of the same partition.
struct PartitionUse : public ByteRange {
- /// \brief The user of this range of the alloca.
- AssertingVH<Instruction> User;
-
- /// \brief The particular pointer value derived from this alloca in use.
- AssertingVH<Instruction> Ptr;
+ /// \brief The use in question. Provides access to both user and used value.
+ ///
+ /// Note that this may be null if the partition use is *dead*, that is, it
+ /// should be ignored.
+ Use *U;
- PartitionUse() : ByteRange(), User(), Ptr() {}
- PartitionUse(uint64_t BeginOffset, uint64_t EndOffset,
- Instruction *User, Instruction *Ptr)
- : ByteRange(BeginOffset, EndOffset), User(User), Ptr(Ptr) {}
+ PartitionUse() : ByteRange(), U() {}
+ PartitionUse(uint64_t BeginOffset, uint64_t EndOffset, Use *U)
+ : ByteRange(BeginOffset, EndOffset), U(U) {}
};
/// \brief Construct a partitioning of a particular alloca.
use_iterator use_begin(const_iterator I) { return Uses[I - begin()].begin(); }
use_iterator use_end(unsigned Idx) { return Uses[Idx].end(); }
use_iterator use_end(const_iterator I) { return Uses[I - begin()].end(); }
- void use_push_back(unsigned Idx, const PartitionUse &U) {
- Uses[Idx].push_back(U);
- }
- void use_push_back(const_iterator I, const PartitionUse &U) {
- Uses[I - begin()].push_back(U);
- }
- void use_erase(unsigned Idx, use_iterator UI) { Uses[Idx].erase(UI); }
- void use_erase(const_iterator I, use_iterator UI) {
- Uses[I - begin()].erase(UI);
- }
typedef SmallVectorImpl<PartitionUse>::const_iterator const_use_iterator;
const_use_iterator use_begin(unsigned Idx) const { return Uses[Idx].begin(); }
const_use_iterator use_end(const_iterator I) const {
return Uses[I - begin()].end();
}
+
+ unsigned use_size(unsigned Idx) const { return Uses[Idx].size(); }
+ unsigned use_size(const_iterator I) const { return Uses[I - begin()].size(); }
+ const PartitionUse &getUse(unsigned PIdx, unsigned UIdx) const {
+ return Uses[PIdx][UIdx];
+ }
+ const PartitionUse &getUse(const_iterator I, unsigned UIdx) const {
+ return Uses[I - begin()][UIdx];
+ }
+
+ void use_push_back(unsigned Idx, const PartitionUse &PU) {
+ Uses[Idx].push_back(PU);
+ }
+ void use_push_back(const_iterator I, const PartitionUse &PU) {
+ Uses[I - begin()].push_back(PU);
+ }
/// @}
/// \brief Allow iterating the dead users for this alloca.
/// When manipulating PHI nodes or selects, they can use more than one
/// partition of an alloca. We store a special mapping to allow finding the
/// partition referenced by each of these operands, if any.
- iterator findPartitionForPHIOrSelectOperand(Instruction &I, Value *Op) {
- SmallDenseMap<std::pair<Instruction *, Value *>,
- std::pair<unsigned, unsigned> >::const_iterator MapIt
- = PHIOrSelectOpMap.find(std::make_pair(&I, Op));
+ iterator findPartitionForPHIOrSelectOperand(Use *U) {
+ SmallDenseMap<Use *, std::pair<unsigned, unsigned> >::const_iterator MapIt
+ = PHIOrSelectOpMap.find(U);
if (MapIt == PHIOrSelectOpMap.end())
return end();
///
/// Similar to mapping these operands back to the partitions, this maps
/// directly to the use structure of that partition.
- use_iterator findPartitionUseForPHIOrSelectOperand(Instruction &I,
- Value *Op) {
- SmallDenseMap<std::pair<Instruction *, Value *>,
- std::pair<unsigned, unsigned> >::const_iterator MapIt
- = PHIOrSelectOpMap.find(std::make_pair(&I, Op));
+ use_iterator findPartitionUseForPHIOrSelectOperand(Use *U) {
+ SmallDenseMap<Use *, std::pair<unsigned, unsigned> >::const_iterator MapIt
+ = PHIOrSelectOpMap.find(U);
assert(MapIt != PHIOrSelectOpMap.end());
return Uses[MapIt->second.first].begin() + MapIt->second.second;
}
SmallDenseMap<Instruction *, std::pair<uint64_t, bool> > PHIOrSelectSizes;
/// \brief Auxiliary information for particular PHI or select operands.
- SmallDenseMap<std::pair<Instruction *, Value *>,
- std::pair<unsigned, unsigned>, 4> PHIOrSelectOpMap;
+ SmallDenseMap<Use *, std::pair<unsigned, unsigned>, 4> PHIOrSelectOpMap;
/// \brief A utility routine called from the constructor.
///
const uint64_t AllocSize;
AllocaPartitioning &P;
+ SmallPtrSet<Use *, 8> VisitedUses;
+
struct OffsetUse {
Use *U;
int64_t Offset;
int64_t Offset;
void enqueueUsers(Instruction &I, int64_t UserOffset) {
- SmallPtrSet<User *, 8> UserSet;
for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
UI != UE; ++UI) {
- if (!UserSet.insert(*UI))
- continue;
-
- OffsetUse OU = { &UI.getUse(), UserOffset };
- Queue.push_back(OU);
+ if (VisitedUses.insert(&UI.getUse())) {
+ OffsetUse OU = { &UI.getUse(), UserOffset };
+ Queue.push_back(OU);
+ }
}
}
bool Inserted = false;
llvm::tie(PMI, Inserted)
= MemTransferPartitionMap.insert(std::make_pair(&II, NewIdx));
- if (!Inserted && Offsets.IsSplittable) {
+ if (Offsets.IsSplittable &&
+ (!Inserted || II.getRawSource() == II.getRawDest())) {
// We've found a memory transfer intrinsic which refers to the alloca as
- // both a source and dest. We refuse to split these to simplify splitting
- // logic. If possible, SROA will still split them into separate allocas
- // and then re-analyze.
+ // both a source and dest. This is detected either by direct equality of
+ // the operand values, or when we visit the intrinsic twice due to two
+ // different chains of values leading to it. We refuse to split these to
+ // simplify splitting logic. If possible, SROA will still split them into
+ // separate allocas and then re-analyze.
Offsets.IsSplittable = false;
P.Partitions[PMI->second].IsSplittable = false;
P.Partitions[NewIdx].IsSplittable = false;
B = llvm::prior(B);
for (iterator I = B, E = P.end(); I != E && I->BeginOffset < EndOffset;
++I) {
- PartitionUse NewUse(std::max(I->BeginOffset, BeginOffset),
- std::min(I->EndOffset, EndOffset),
- &User, cast<Instruction>(*U));
- P.use_push_back(I, NewUse);
+ PartitionUse NewPU(std::max(I->BeginOffset, BeginOffset),
+ std::min(I->EndOffset, EndOffset), U);
+ P.use_push_back(I, NewPU);
if (isa<PHINode>(U->getUser()) || isa<SelectInst>(U->getUser()))
- P.PHIOrSelectOpMap[std::make_pair(&User, U->get())]
+ P.PHIOrSelectOpMap[U]
= std::make_pair(I - P.begin(), P.Uses[I - P.begin()].size() - 1);
}
}
Type *AllocaPartitioning::getCommonType(iterator I) const {
Type *Ty = 0;
for (const_use_iterator UI = use_begin(I), UE = use_end(I); UI != UE; ++UI) {
- if (isa<IntrinsicInst>(*UI->User))
+ if (!UI->U)
+ continue; // Skip dead uses.
+ if (isa<IntrinsicInst>(*UI->U->getUser()))
continue;
if (UI->BeginOffset != I->BeginOffset || UI->EndOffset != I->EndOffset)
continue;
Type *UserTy = 0;
- if (LoadInst *LI = dyn_cast<LoadInst>(&*UI->User)) {
+ if (LoadInst *LI = dyn_cast<LoadInst>(UI->U->getUser())) {
UserTy = LI->getType();
- } else if (StoreInst *SI = dyn_cast<StoreInst>(&*UI->User)) {
+ } else if (StoreInst *SI = dyn_cast<StoreInst>(UI->U->getUser())) {
UserTy = SI->getValueOperand()->getType();
- } else if (SelectInst *SI = dyn_cast<SelectInst>(&*UI->User)) {
- if (PointerType *PtrTy = dyn_cast<PointerType>(SI->getType()))
- UserTy = PtrTy->getElementType();
- } else if (PHINode *PN = dyn_cast<PHINode>(&*UI->User)) {
- if (PointerType *PtrTy = dyn_cast<PointerType>(PN->getType()))
- UserTy = PtrTy->getElementType();
}
if (Ty && Ty != UserTy)
StringRef Indent) const {
for (const_use_iterator UI = use_begin(I), UE = use_end(I);
UI != UE; ++UI) {
+ if (!UI->U)
+ continue; // Skip dead uses.
OS << Indent << " [" << UI->BeginOffset << "," << UI->EndOffset << ") "
- << "used by: " << *UI->User << "\n";
- if (MemTransferInst *II = dyn_cast<MemTransferInst>(&*UI->User)) {
+ << "used by: " << *UI->U->getUser() << "\n";
+ if (MemTransferInst *II = dyn_cast<MemTransferInst>(UI->U->getUser())) {
const MemTransferOffsets &MTO = MemTransferInstData.lookup(II);
bool IsDest;
if (!MTO.IsSplittable)
/// uses as dead. Only used to guard insertion into DeadInsts.
SmallPtrSet<Instruction *, 4> DeadSplitInsts;
+ /// \brief Post-promotion worklist.
+ ///
+ /// Sometimes we discover an alloca which has a high probability of becoming
+ /// viable for SROA after a round of promotion takes place. In those cases,
+ /// the alloca is enqueued here for re-processing.
+ ///
+ /// Note that we have to be very careful to clear allocas out of this list in
+ /// the event they are deleted.
+ SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > PostPromotionWorklist;
+
/// \brief A collection of alloca instructions we can directly promote.
std::vector<AllocaInst *> PromotableAllocas;
static char ID;
private:
+ friend class PHIOrSelectSpeculator;
friend class AllocaPartitionRewriter;
friend class AllocaPartitionVectorRewriter;
INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
false, false)
+namespace {
+/// \brief Visitor to speculate PHIs and Selects where possible.
+class PHIOrSelectSpeculator : public InstVisitor<PHIOrSelectSpeculator> {
+ // Befriend the base class so it can delegate to private visit methods.
+ friend class llvm::InstVisitor<PHIOrSelectSpeculator>;
+
+ const TargetData &TD;
+ AllocaPartitioning &P;
+ SROA &Pass;
+
+public:
+ PHIOrSelectSpeculator(const TargetData &TD, AllocaPartitioning &P, SROA &Pass)
+ : TD(TD), P(P), Pass(Pass) {}
+
+ /// \brief Visit the users of an alloca partition and rewrite them.
+ void visitUsers(AllocaPartitioning::const_iterator PI) {
+ // Note that we need to use an index here as the underlying vector of uses
+ // may be grown during speculation. However, we never need to re-visit the
+ // new uses, and so we can use the initial size bound.
+ for (unsigned Idx = 0, Size = P.use_size(PI); Idx != Size; ++Idx) {
+ const AllocaPartitioning::PartitionUse &PU = P.getUse(PI, Idx);
+ if (!PU.U)
+ continue; // Skip dead use.
+
+ visit(cast<Instruction>(PU.U->getUser()));
+ }
+ }
+
+private:
+ // By default, skip this instruction.
+ void visitInstruction(Instruction &I) {}
+
+ /// PHI instructions that use an alloca and are subsequently loaded can be
+ /// rewritten to load both input pointers in the pred blocks and then PHI the
+ /// results, allowing the load of the alloca to be promoted.
+ /// From this:
+ /// %P2 = phi [i32* %Alloca, i32* %Other]
+ /// %V = load i32* %P2
+ /// to:
+ /// %V1 = load i32* %Alloca -> will be mem2reg'd
+ /// ...
+ /// %V2 = load i32* %Other
+ /// ...
+ /// %V = phi [i32 %V1, i32 %V2]
+ ///
+ /// We can do this to a select if its only uses are loads and if the operands
+ /// to the select can be loaded unconditionally.
+ ///
+ /// FIXME: This should be hoisted into a generic utility, likely in
+ /// Transforms/Util/Local.h
+ bool isSafePHIToSpeculate(PHINode &PN, SmallVectorImpl<LoadInst *> &Loads) {
+ // For now, we can only do this promotion if the load is in the same block
+ // as the PHI, and if there are no stores between the phi and load.
+ // TODO: Allow recursive phi users.
+ // TODO: Allow stores.
+ BasicBlock *BB = PN.getParent();
+ unsigned MaxAlign = 0;
+ for (Value::use_iterator UI = PN.use_begin(), UE = PN.use_end();
+ UI != UE; ++UI) {
+ LoadInst *LI = dyn_cast<LoadInst>(*UI);
+ if (LI == 0 || !LI->isSimple()) return false;
+
+ // For now we only allow loads in the same block as the PHI. This is
+ // a common case that happens when instcombine merges two loads through
+ // a PHI.
+ if (LI->getParent() != BB) return false;
+
+ // Ensure that there are no instructions between the PHI and the load that
+ // could store.
+ for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
+ if (BBI->mayWriteToMemory())
+ return false;
+
+ MaxAlign = std::max(MaxAlign, LI->getAlignment());
+ Loads.push_back(LI);
+ }
+
+ // We can only transform this if it is safe to push the loads into the
+ // predecessor blocks. The only thing to watch out for is that we can't put
+ // a possibly trapping load in the predecessor if it is a critical edge.
+ for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num;
+ ++Idx) {
+ TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
+ Value *InVal = PN.getIncomingValue(Idx);
+
+ // If the value is produced by the terminator of the predecessor (an
+ // invoke) or it has side-effects, there is no valid place to put a load
+ // in the predecessor.
+ if (TI == InVal || TI->mayHaveSideEffects())
+ return false;
+
+ // If the predecessor has a single successor, then the edge isn't
+ // critical.
+ if (TI->getNumSuccessors() == 1)
+ continue;
+
+ // If this pointer is always safe to load, or if we can prove that there
+ // is already a load in the block, then we can move the load to the pred
+ // block.
+ if (InVal->isDereferenceablePointer() ||
+ isSafeToLoadUnconditionally(InVal, TI, MaxAlign, &TD))
+ continue;
+
+ return false;
+ }
+
+ return true;
+ }
+
+ void visitPHINode(PHINode &PN) {
+ DEBUG(dbgs() << " original: " << PN << "\n");
+
+ SmallVector<LoadInst *, 4> Loads;
+ if (!isSafePHIToSpeculate(PN, Loads))
+ return;
+
+ assert(!Loads.empty());
+
+ Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
+ IRBuilder<> PHIBuilder(&PN);
+ PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
+ PN.getName() + ".sroa.speculated");
+
+ // Get the TBAA tag and alignment to use from one of the loads. It doesn't
+ // matter which one we get and if any differ, it doesn't matter.
+ LoadInst *SomeLoad = cast<LoadInst>(Loads.back());
+ MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
+ unsigned Align = SomeLoad->getAlignment();
+
+ // Rewrite all loads of the PN to use the new PHI.
+ do {
+ LoadInst *LI = Loads.pop_back_val();
+ LI->replaceAllUsesWith(NewPN);
+ Pass.DeadInsts.push_back(LI);
+ } while (!Loads.empty());
+
+ // Inject loads into all of the pred blocks.
+ for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
+ BasicBlock *Pred = PN.getIncomingBlock(Idx);
+ TerminatorInst *TI = Pred->getTerminator();
+ Use *InUse = &PN.getOperandUse(PN.getOperandNumForIncomingValue(Idx));
+ Value *InVal = PN.getIncomingValue(Idx);
+ IRBuilder<> PredBuilder(TI);
+
+ LoadInst *Load
+ = PredBuilder.CreateLoad(InVal, (PN.getName() + ".sroa.speculate.load." +
+ Pred->getName()));
+ ++NumLoadsSpeculated;
+ Load->setAlignment(Align);
+ if (TBAATag)
+ Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
+ NewPN->addIncoming(Load, Pred);
+
+ Instruction *Ptr = dyn_cast<Instruction>(InVal);
+ if (!Ptr)
+ // No uses to rewrite.
+ continue;
+
+ // Try to lookup and rewrite any partition uses corresponding to this phi
+ // input.
+ AllocaPartitioning::iterator PI
+ = P.findPartitionForPHIOrSelectOperand(InUse);
+ if (PI == P.end())
+ continue;
+
+ // Replace the Use in the PartitionUse for this operand with the Use
+ // inside the load.
+ AllocaPartitioning::use_iterator UI
+ = P.findPartitionUseForPHIOrSelectOperand(InUse);
+ assert(isa<PHINode>(*UI->U->getUser()));
+ UI->U = &Load->getOperandUse(Load->getPointerOperandIndex());
+ }
+ DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
+ }
+
+ /// Select instructions that use an alloca and are subsequently loaded can be
+ /// rewritten to load both input pointers and then select between the result,
+ /// allowing the load of the alloca to be promoted.
+ /// From this:
+ /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
+ /// %V = load i32* %P2
+ /// to:
+ /// %V1 = load i32* %Alloca -> will be mem2reg'd
+ /// %V2 = load i32* %Other
+ /// %V = select i1 %cond, i32 %V1, i32 %V2
+ ///
+ /// We can do this to a select if its only uses are loads and if the operand
+ /// to the select can be loaded unconditionally.
+ bool isSafeSelectToSpeculate(SelectInst &SI,
+ SmallVectorImpl<LoadInst *> &Loads) {
+ Value *TValue = SI.getTrueValue();
+ Value *FValue = SI.getFalseValue();
+ bool TDerefable = TValue->isDereferenceablePointer();
+ bool FDerefable = FValue->isDereferenceablePointer();
+
+ for (Value::use_iterator UI = SI.use_begin(), UE = SI.use_end();
+ UI != UE; ++UI) {
+ LoadInst *LI = dyn_cast<LoadInst>(*UI);
+ if (LI == 0 || !LI->isSimple()) return false;
+
+ // Both operands to the select need to be dereferencable, either
+ // absolutely (e.g. allocas) or at this point because we can see other
+ // accesses to it.
+ if (!TDerefable && !isSafeToLoadUnconditionally(TValue, LI,
+ LI->getAlignment(), &TD))
+ return false;
+ if (!FDerefable && !isSafeToLoadUnconditionally(FValue, LI,
+ LI->getAlignment(), &TD))
+ return false;
+ Loads.push_back(LI);
+ }
+
+ return true;
+ }
+
+ void visitSelectInst(SelectInst &SI) {
+ DEBUG(dbgs() << " original: " << SI << "\n");
+ IRBuilder<> IRB(&SI);
+
+ // If the select isn't safe to speculate, just use simple logic to emit it.
+ SmallVector<LoadInst *, 4> Loads;
+ if (!isSafeSelectToSpeculate(SI, Loads))
+ return;
+
+ Use *Ops[2] = { &SI.getOperandUse(1), &SI.getOperandUse(2) };
+ AllocaPartitioning::iterator PIs[2];
+ AllocaPartitioning::PartitionUse PUs[2];
+ for (unsigned i = 0, e = 2; i != e; ++i) {
+ PIs[i] = P.findPartitionForPHIOrSelectOperand(Ops[i]);
+ if (PIs[i] != P.end()) {
+ // If the pointer is within the partitioning, remove the select from
+ // its uses. We'll add in the new loads below.
+ AllocaPartitioning::use_iterator UI
+ = P.findPartitionUseForPHIOrSelectOperand(Ops[i]);
+ PUs[i] = *UI;
+ // Clear out the use here so that the offsets into the use list remain
+ // stable but this use is ignored when rewriting.
+ UI->U = 0;
+ }
+ }
+
+ Value *TV = SI.getTrueValue();
+ Value *FV = SI.getFalseValue();
+ // Replace the loads of the select with a select of two loads.
+ while (!Loads.empty()) {
+ LoadInst *LI = Loads.pop_back_val();
+
+ IRB.SetInsertPoint(LI);
+ LoadInst *TL =
+ IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
+ LoadInst *FL =
+ IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
+ NumLoadsSpeculated += 2;
+
+ // Transfer alignment and TBAA info if present.
+ TL->setAlignment(LI->getAlignment());
+ FL->setAlignment(LI->getAlignment());
+ if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
+ TL->setMetadata(LLVMContext::MD_tbaa, Tag);
+ FL->setMetadata(LLVMContext::MD_tbaa, Tag);
+ }
+
+ Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
+ LI->getName() + ".sroa.speculated");
+
+ LoadInst *Loads[2] = { TL, FL };
+ for (unsigned i = 0, e = 2; i != e; ++i) {
+ if (PIs[i] != P.end()) {
+ Use *LoadUse = &Loads[i]->getOperandUse(0);
+ assert(PUs[i].U->get() == LoadUse->get());
+ PUs[i].U = LoadUse;
+ P.use_push_back(PIs[i], PUs[i]);
+ }
+ }
+
+ DEBUG(dbgs() << " speculated to: " << *V << "\n");
+ LI->replaceAllUsesWith(V);
+ Pass.DeadInsts.push_back(LI);
+ }
+ }
+};
+}
+
/// \brief Accumulate the constant offsets in a GEP into a single APInt offset.
///
/// If the provided GEP is all-constant, the total byte offset formed by the
ElementSize /= 8;
for (; I != E; ++I) {
+ if (!I->U)
+ continue; // Skip dead use.
+
uint64_t BeginOffset = I->BeginOffset - PartitionBeginOffset;
uint64_t BeginIndex = BeginOffset / ElementSize;
if (BeginIndex * ElementSize != BeginOffset ||
(EndOffset - BeginOffset) != VecSize)
return false;
- if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(&*I->User)) {
+ if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I->U->getUser())) {
if (MI->isVolatile())
return false;
- if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(&*I->User)) {
+ if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(I->U->getUser())) {
const AllocaPartitioning::MemTransferOffsets &MTO
= P.getMemTransferOffsets(*MTI);
if (!MTO.IsSplittable)
return false;
}
- } else if (I->Ptr->getType()->getPointerElementType()->isStructTy()) {
+ } else if (I->U->get()->getType()->getPointerElementType()->isStructTy()) {
// Disable vector promotion when there are loads or stores of an FCA.
return false;
- } else if (!isa<LoadInst>(*I->User) && !isa<StoreInst>(*I->User)) {
+ } else if (!isa<LoadInst>(I->U->getUser()) &&
+ !isa<StoreInst>(I->U->getUser())) {
return false;
}
}
/// that the result will be promotable, so we have an early test here.
static bool isIntegerPromotionViable(const TargetData &TD,
Type *AllocaTy,
+ uint64_t AllocBeginOffset,
AllocaPartitioning &P,
AllocaPartitioning::const_use_iterator I,
AllocaPartitioning::const_use_iterator E) {
IntegerType *Ty = dyn_cast<IntegerType>(AllocaTy);
- if (!Ty)
+ if (!Ty || 8*TD.getTypeStoreSize(Ty) != Ty->getBitWidth())
return false;
// Check the uses to ensure the uses are (likely) promoteable integer uses.
// splittable later) and lose the ability to promote each element access.
bool WholeAllocaOp = false;
for (; I != E; ++I) {
- if (LoadInst *LI = dyn_cast<LoadInst>(&*I->User)) {
+ if (!I->U)
+ continue; // Skip dead use.
+
+ // We can't reasonably handle cases where the load or store extends past
+ // the end of the aloca's type and into its padding.
+ if ((I->EndOffset - AllocBeginOffset) > TD.getTypeStoreSize(Ty))
+ return false;
+
+ if (LoadInst *LI = dyn_cast<LoadInst>(I->U->getUser())) {
if (LI->isVolatile() || !LI->getType()->isIntegerTy())
return false;
if (LI->getType() == Ty)
WholeAllocaOp = true;
- } else if (StoreInst *SI = dyn_cast<StoreInst>(&*I->User)) {
+ } else if (StoreInst *SI = dyn_cast<StoreInst>(I->U->getUser())) {
if (SI->isVolatile() || !SI->getValueOperand()->getType()->isIntegerTy())
return false;
if (SI->getValueOperand()->getType() == Ty)
WholeAllocaOp = true;
- } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(&*I->User)) {
+ } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I->U->getUser())) {
if (MI->isVolatile())
return false;
- if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(&*I->User)) {
+ if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(I->U->getUser())) {
const AllocaPartitioning::MemTransferOffsets &MTO
= P.getMemTransferOffsets(*MTI);
if (!MTO.IsSplittable)
// The offset of the partition user currently being rewritten.
uint64_t BeginOffset, EndOffset;
+ Use *OldUse;
Instruction *OldPtr;
// The name prefix to use when rewriting instructions for this alloca.
"Only multiple-of-8 sized vector elements are viable");
ElementSize = VecTy->getScalarSizeInBits() / 8;
} else if (isIntegerPromotionViable(TD, NewAI.getAllocatedType(),
- P, I, E)) {
+ NewAllocaBeginOffset, P, I, E)) {
IntPromotionTy = cast<IntegerType>(NewAI.getAllocatedType());
}
bool CanSROA = true;
for (; I != E; ++I) {
+ if (!I->U)
+ continue; // Skip dead uses.
BeginOffset = I->BeginOffset;
EndOffset = I->EndOffset;
- OldPtr = I->Ptr;
+ OldUse = I->U;
+ OldPtr = cast<Instruction>(I->U->get());
NamePrefix = (Twine(NewAI.getName()) + "." + Twine(BeginOffset)).str();
- CanSROA &= visit(I->User);
+ CanSROA &= visit(cast<Instruction>(I->U->getUser()));
}
if (VecTy) {
assert(CanSROA);
return getAdjustedPtr(IRB, TD, &NewAI, Offset, PointerTy, getName(""));
}
+ /// \brief Compute suitable alignment to access an offset into the new alloca.
+ unsigned getOffsetAlign(uint64_t Offset) {
+ unsigned NewAIAlign = NewAI.getAlignment();
+ if (!NewAIAlign)
+ NewAIAlign = TD.getABITypeAlignment(NewAI.getAllocatedType());
+ return MinAlign(NewAIAlign, Offset);
+ }
+
+ /// \brief Compute suitable alignment to access this partition of the new
+ /// alloca.
+ unsigned getPartitionAlign() {
+ return getOffsetAlign(BeginOffset - NewAllocaBeginOffset);
+ }
+
+ /// \brief Compute suitable alignment to access a type at an offset of the
+ /// new alloca.
+ ///
+ /// \returns zero if the type's ABI alignment is a suitable alignment,
+ /// otherwise returns the maximal suitable alignment.
+ unsigned getOffsetTypeAlign(Type *Ty, uint64_t Offset) {
+ unsigned Align = getOffsetAlign(Offset);
+ return Align == TD.getABITypeAlignment(Ty) ? 0 : Align;
+ }
+
+ /// \brief Compute suitable alignment to access a type at the beginning of
+ /// this partition of the new alloca.
+ ///
+ /// See \c getOffsetTypeAlign for details; this routine delegates to it.
+ unsigned getPartitionTypeAlign(Type *Ty) {
+ return getOffsetTypeAlign(Ty, BeginOffset - NewAllocaBeginOffset);
+ }
+
ConstantInt *getIndex(IRBuilder<> &IRB, uint64_t Offset) {
assert(VecTy && "Can only call getIndex when rewriting a vector");
uint64_t RelOffset = Offset - NewAllocaBeginOffset;
Value *extractInteger(IRBuilder<> &IRB, IntegerType *TargetTy,
uint64_t Offset) {
assert(IntPromotionTy && "Alloca is not an integer we can extract from");
- Value *V = IRB.CreateLoad(&NewAI, getName(".load"));
+ Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
+ getName(".load"));
assert(Offset >= NewAllocaBeginOffset && "Out of bounds offset");
uint64_t RelOffset = Offset - NewAllocaBeginOffset;
- if (RelOffset)
- V = IRB.CreateLShr(V, RelOffset*8, getName(".shift"));
+ assert(TD.getTypeStoreSize(TargetTy) + RelOffset <=
+ TD.getTypeStoreSize(IntPromotionTy) &&
+ "Element load outside of alloca store");
+ uint64_t ShAmt = 8*RelOffset;
+ if (TD.isBigEndian())
+ ShAmt = 8*(TD.getTypeStoreSize(IntPromotionTy) -
+ TD.getTypeStoreSize(TargetTy) - RelOffset);
+ if (ShAmt)
+ V = IRB.CreateLShr(V, ShAmt, getName(".shift"));
if (TargetTy != IntPromotionTy) {
assert(TargetTy->getBitWidth() < IntPromotionTy->getBitWidth() &&
"Cannot extract to a larger integer!");
StoreInst *insertInteger(IRBuilder<> &IRB, Value *V, uint64_t Offset) {
IntegerType *Ty = cast<IntegerType>(V->getType());
if (Ty == IntPromotionTy)
- return IRB.CreateStore(V, &NewAI);
+ return IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
assert(Ty->getBitWidth() < IntPromotionTy->getBitWidth() &&
"Cannot insert a larger integer!");
V = IRB.CreateZExt(V, IntPromotionTy, getName(".ext"));
assert(Offset >= NewAllocaBeginOffset && "Out of bounds offset");
uint64_t RelOffset = Offset - NewAllocaBeginOffset;
- if (RelOffset)
- V = IRB.CreateShl(V, RelOffset*8, getName(".shift"));
-
- APInt Mask = ~Ty->getMask().zext(IntPromotionTy->getBitWidth())
- .shl(RelOffset*8);
- Value *Old = IRB.CreateAnd(IRB.CreateLoad(&NewAI, getName(".oldload")),
+ assert(TD.getTypeStoreSize(Ty) + RelOffset <=
+ TD.getTypeStoreSize(IntPromotionTy) &&
+ "Element store outside of alloca store");
+ uint64_t ShAmt = 8*RelOffset;
+ if (TD.isBigEndian())
+ ShAmt = 8*(TD.getTypeStoreSize(IntPromotionTy) - TD.getTypeStoreSize(Ty)
+ - RelOffset);
+ if (ShAmt)
+ V = IRB.CreateShl(V, ShAmt, getName(".shift"));
+
+ APInt Mask = ~Ty->getMask().zext(IntPromotionTy->getBitWidth()).shl(ShAmt);
+ Value *Old = IRB.CreateAnd(IRB.CreateAlignedLoad(&NewAI,
+ NewAI.getAlignment(),
+ getName(".oldload")),
Mask, getName(".mask"));
- return IRB.CreateStore(IRB.CreateOr(Old, V, getName(".insert")),
- &NewAI);
+ return IRB.CreateAlignedStore(IRB.CreateOr(Old, V, getName(".insert")),
+ &NewAI, NewAI.getAlignment());
}
void deleteIfTriviallyDead(Value *V) {
Value *Result;
if (LI.getType() == VecTy->getElementType() ||
BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset) {
- Result
- = IRB.CreateExtractElement(IRB.CreateLoad(&NewAI, getName(".load")),
- getIndex(IRB, BeginOffset),
- getName(".extract"));
+ Result = IRB.CreateExtractElement(
+ IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), getName(".load")),
+ getIndex(IRB, BeginOffset), getName(".extract"));
} else {
- Result = IRB.CreateLoad(&NewAI, getName(".load"));
+ Result = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
+ getName(".load"));
}
if (Result->getType() != LI.getType())
Result = getValueCast(IRB, Result, LI.getType());
Value *NewPtr = getAdjustedAllocaPtr(IRB,
LI.getPointerOperand()->getType());
LI.setOperand(0, NewPtr);
+ LI.setAlignment(getPartitionTypeAlign(LI.getType()));
DEBUG(dbgs() << " to: " << LI << "\n");
deleteIfTriviallyDead(OldOp);
BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset) {
if (V->getType() != ElementTy)
V = getValueCast(IRB, V, ElementTy);
- V = IRB.CreateInsertElement(IRB.CreateLoad(&NewAI, getName(".load")), V,
- getIndex(IRB, BeginOffset),
+ LoadInst *LI = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
+ getName(".load"));
+ V = IRB.CreateInsertElement(LI, V, getIndex(IRB, BeginOffset),
getName(".insert"));
} else if (V->getType() != VecTy) {
V = getValueCast(IRB, V, VecTy);
}
- StoreInst *Store = IRB.CreateStore(V, &NewAI);
+ StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
Pass.DeadInsts.push_back(&SI);
(void)Store;
if (IntPromotionTy)
return rewriteIntegerStore(IRB, SI);
+ // Strip all inbounds GEPs and pointer casts to try to dig out any root
+ // alloca that should be re-examined after promoting this alloca.
+ if (SI.getValueOperand()->getType()->isPointerTy())
+ if (AllocaInst *AI = dyn_cast<AllocaInst>(SI.getValueOperand()
+ ->stripInBoundsOffsets()))
+ Pass.PostPromotionWorklist.insert(AI);
+
Value *NewPtr = getAdjustedAllocaPtr(IRB,
SI.getPointerOperand()->getType());
SI.setOperand(1, NewPtr);
+ SI.setAlignment(getPartitionTypeAlign(SI.getValueOperand()->getType()));
DEBUG(dbgs() << " to: " << SI << "\n");
deleteIfTriviallyDead(OldOp);
// pointer to the new alloca.
if (!isa<Constant>(II.getLength())) {
II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
+ Type *CstTy = II.getAlignmentCst()->getType();
+ II.setAlignment(ConstantInt::get(CstTy, getPartitionAlign()));
+
deleteIfTriviallyDead(OldPtr);
return false;
}
!TD.isLegalInteger(TD.getTypeSizeInBits(ScalarTy)))) {
Type *SizeTy = II.getLength()->getType();
Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
-
CallInst *New
= IRB.CreateMemSet(getAdjustedAllocaPtr(IRB,
II.getRawDest()->getType()),
- II.getValue(), Size, II.getAlignment(),
+ II.getValue(), Size, getPartitionAlign(),
II.isVolatile());
(void)New;
DEBUG(dbgs() << " to: " << *New << "\n");
// If this is an element-wide memset of a vectorizable alloca, insert it.
if (VecTy && (BeginOffset > NewAllocaBeginOffset ||
EndOffset < NewAllocaEndOffset)) {
- StoreInst *Store = IRB.CreateStore(
- IRB.CreateInsertElement(IRB.CreateLoad(&NewAI, getName(".load")), V,
- getIndex(IRB, BeginOffset),
+ StoreInst *Store = IRB.CreateAlignedStore(
+ IRB.CreateInsertElement(IRB.CreateAlignedLoad(&NewAI,
+ NewAI.getAlignment(),
+ getName(".load")),
+ V, getIndex(IRB, BeginOffset),
getName(".insert")),
- &NewAI);
+ &NewAI, NewAI.getAlignment());
(void)Store;
DEBUG(dbgs() << " to: " << *Store << "\n");
return true;
assert(V->getType() == VecTy);
}
- Value *New = IRB.CreateStore(V, &NewAI, II.isVolatile());
+ Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
+ II.isVolatile());
(void)New;
DEBUG(dbgs() << " to: " << *New << "\n");
return !II.isVolatile();
const AllocaPartitioning::MemTransferOffsets &MTO
= P.getMemTransferOffsets(II);
+ // Compute the relative offset within the transfer.
+ unsigned IntPtrWidth = TD.getPointerSizeInBits();
+ APInt RelOffset(IntPtrWidth, BeginOffset - (IsDest ? MTO.DestBegin
+ : MTO.SourceBegin));
+
+ unsigned Align = II.getAlignment();
+ if (Align > 1)
+ Align = MinAlign(RelOffset.zextOrTrunc(64).getZExtValue(),
+ MinAlign(II.getAlignment(), getPartitionAlign()));
+
// For unsplit intrinsics, we simply modify the source and destination
// pointers in place. This isn't just an optimization, it is a matter of
// correctness. With unsplit intrinsics we may be dealing with transfers
else
II.setSource(getAdjustedAllocaPtr(IRB, II.getRawSource()->getType()));
+ Type *CstTy = II.getAlignmentCst()->getType();
+ II.setAlignment(ConstantInt::get(CstTy, Align));
+
DEBUG(dbgs() << " to: " << II << "\n");
deleteIfTriviallyDead(OldOp);
return false;
// memmove with memcpy, and we don't need to worry about all manner of
// downsides to splitting and transforming the operations.
- // Compute the relative offset within the transfer.
- unsigned IntPtrWidth = TD.getPointerSizeInBits();
- APInt RelOffset(IntPtrWidth, BeginOffset - (IsDest ? MTO.DestBegin
- : MTO.SourceBegin));
-
// If this doesn't map cleanly onto the alloca type, and that type isn't
// a single value type, just emit a memcpy.
bool EmitMemCpy
CallInst *New = IRB.CreateMemCpy(IsDest ? OurPtr : OtherPtr,
IsDest ? OtherPtr : OurPtr,
- Size, II.getAlignment(),
- II.isVolatile());
+ Size, Align, II.isVolatile());
(void)New;
DEBUG(dbgs() << " to: " << *New << "\n");
return false;
}
+ // Note that we clamp the alignment to 1 here as a 0 alignment for a memcpy
+ // is equivalent to 1, but that isn't true if we end up rewriting this as
+ // a load or store.
+ if (!Align)
+ Align = 1;
+
Value *SrcPtr = OtherPtr;
Value *DstPtr = &NewAI;
if (!IsDest)
Value *Src;
if (IsVectorElement && !IsDest) {
// We have to extract rather than load.
- Src = IRB.CreateExtractElement(IRB.CreateLoad(SrcPtr,
- getName(".copyload")),
- getIndex(IRB, BeginOffset),
- getName(".copyextract"));
+ Src = IRB.CreateExtractElement(
+ IRB.CreateAlignedLoad(SrcPtr, Align, getName(".copyload")),
+ getIndex(IRB, BeginOffset),
+ getName(".copyextract"));
} else {
- Src = IRB.CreateLoad(SrcPtr, II.isVolatile(), getName(".copyload"));
+ Src = IRB.CreateAlignedLoad(SrcPtr, Align, II.isVolatile(),
+ getName(".copyload"));
}
if (IsVectorElement && IsDest) {
// We have to insert into a loaded copy before storing.
- Src = IRB.CreateInsertElement(IRB.CreateLoad(&NewAI, getName(".load")),
- Src, getIndex(IRB, BeginOffset),
- getName(".insert"));
+ Src = IRB.CreateInsertElement(
+ IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), getName(".load")),
+ Src, getIndex(IRB, BeginOffset),
+ getName(".insert"));
}
- Value *Store = IRB.CreateStore(Src, DstPtr, II.isVolatile());
+ StoreInst *Store = cast<StoreInst>(
+ IRB.CreateAlignedStore(Src, DstPtr, Align, II.isVolatile()));
(void)Store;
DEBUG(dbgs() << " to: " << *Store << "\n");
return !II.isVolatile();
return true;
}
- /// PHI instructions that use an alloca and are subsequently loaded can be
- /// rewritten to load both input pointers in the pred blocks and then PHI the
- /// results, allowing the load of the alloca to be promoted.
- /// From this:
- /// %P2 = phi [i32* %Alloca, i32* %Other]
- /// %V = load i32* %P2
- /// to:
- /// %V1 = load i32* %Alloca -> will be mem2reg'd
- /// ...
- /// %V2 = load i32* %Other
- /// ...
- /// %V = phi [i32 %V1, i32 %V2]
- ///
- /// We can do this to a select if its only uses are loads and if the operand
- /// to the select can be loaded unconditionally.
- ///
- /// FIXME: This should be hoisted into a generic utility, likely in
- /// Transforms/Util/Local.h
- bool isSafePHIToSpeculate(PHINode &PN, SmallVectorImpl<LoadInst *> &Loads) {
- // For now, we can only do this promotion if the load is in the same block
- // as the PHI, and if there are no stores between the phi and load.
- // TODO: Allow recursive phi users.
- // TODO: Allow stores.
- BasicBlock *BB = PN.getParent();
- unsigned MaxAlign = 0;
- for (Value::use_iterator UI = PN.use_begin(), UE = PN.use_end();
- UI != UE; ++UI) {
- LoadInst *LI = dyn_cast<LoadInst>(*UI);
- if (LI == 0 || !LI->isSimple()) return false;
-
- // For now we only allow loads in the same block as the PHI. This is
- // a common case that happens when instcombine merges two loads through
- // a PHI.
- if (LI->getParent() != BB) return false;
-
- // Ensure that there are no instructions between the PHI and the load that
- // could store.
- for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
- if (BBI->mayWriteToMemory())
- return false;
-
- MaxAlign = std::max(MaxAlign, LI->getAlignment());
- Loads.push_back(LI);
- }
-
- // We can only transform this if it is safe to push the loads into the
- // predecessor blocks. The only thing to watch out for is that we can't put
- // a possibly trapping load in the predecessor if it is a critical edge.
- for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num;
- ++Idx) {
- TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
- Value *InVal = PN.getIncomingValue(Idx);
-
- // If the value is produced by the terminator of the predecessor (an
- // invoke) or it has side-effects, there is no valid place to put a load
- // in the predecessor.
- if (TI == InVal || TI->mayHaveSideEffects())
- return false;
-
- // If the predecessor has a single successor, then the edge isn't
- // critical.
- if (TI->getNumSuccessors() == 1)
- continue;
-
- // If this pointer is always safe to load, or if we can prove that there
- // is already a load in the block, then we can move the load to the pred
- // block.
- if (InVal->isDereferenceablePointer() ||
- isSafeToLoadUnconditionally(InVal, TI, MaxAlign, &TD))
- continue;
-
- return false;
- }
-
- return true;
- }
-
bool visitPHINode(PHINode &PN) {
DEBUG(dbgs() << " original: " << PN << "\n");
+
// We would like to compute a new pointer in only one place, but have it be
// 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.
IRBuilder<> PtrBuilder(cast<Instruction>(OldPtr));
- SmallVector<LoadInst *, 4> Loads;
- if (!isSafePHIToSpeculate(PN, Loads)) {
- Value *NewPtr = getAdjustedAllocaPtr(PtrBuilder, OldPtr->getType());
- // Replace the operands which were using the old pointer.
- User::op_iterator OI = PN.op_begin(), OE = PN.op_end();
- for (; OI != OE; ++OI)
- if (*OI == OldPtr)
- *OI = NewPtr;
-
- DEBUG(dbgs() << " to: " << PN << "\n");
- deleteIfTriviallyDead(OldPtr);
- return false;
- }
- assert(!Loads.empty());
-
- Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
- IRBuilder<> PHIBuilder(&PN);
- PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues());
- NewPN->takeName(&PN);
-
- // Get the TBAA tag and alignment to use from one of the loads. It doesn't
- // matter which one we get and if any differ, it doesn't matter.
- LoadInst *SomeLoad = cast<LoadInst>(Loads.back());
- MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
- unsigned Align = SomeLoad->getAlignment();
Value *NewPtr = getAdjustedAllocaPtr(PtrBuilder, OldPtr->getType());
+ // Replace the operands which were using the old pointer.
+ User::op_iterator OI = PN.op_begin(), OE = PN.op_end();
+ for (; OI != OE; ++OI)
+ if (*OI == OldPtr)
+ *OI = NewPtr;
- // Rewrite all loads of the PN to use the new PHI.
- do {
- LoadInst *LI = Loads.pop_back_val();
- LI->replaceAllUsesWith(NewPN);
- Pass.DeadInsts.push_back(LI);
- } while (!Loads.empty());
-
- // Inject loads into all of the pred blocks.
- for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
- BasicBlock *Pred = PN.getIncomingBlock(Idx);
- TerminatorInst *TI = Pred->getTerminator();
- Value *InVal = PN.getIncomingValue(Idx);
- IRBuilder<> PredBuilder(TI);
-
- // Map the value to the new alloca pointer if this was the old alloca
- // pointer.
- bool ThisOperand = InVal == OldPtr;
- if (ThisOperand)
- InVal = NewPtr;
-
- LoadInst *Load
- = PredBuilder.CreateLoad(InVal, getName(".sroa.speculate." +
- Pred->getName()));
- ++NumLoadsSpeculated;
- Load->setAlignment(Align);
- if (TBAATag)
- Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
- NewPN->addIncoming(Load, Pred);
-
- if (ThisOperand)
- continue;
- Instruction *OtherPtr = dyn_cast<Instruction>(InVal);
- if (!OtherPtr)
- // No uses to rewrite.
- continue;
-
- // Try to lookup and rewrite any partition uses corresponding to this phi
- // input.
- AllocaPartitioning::iterator PI
- = P.findPartitionForPHIOrSelectOperand(PN, OtherPtr);
- if (PI != P.end()) {
- // If the other pointer is within the partitioning, replace the PHI in
- // its uses with the load we just speculated, or add another load for
- // it to rewrite if we've already replaced the PHI.
- AllocaPartitioning::use_iterator UI
- = P.findPartitionUseForPHIOrSelectOperand(PN, OtherPtr);
- if (isa<PHINode>(*UI->User))
- UI->User = Load;
- else {
- AllocaPartitioning::PartitionUse OtherUse = *UI;
- OtherUse.User = Load;
- P.use_push_back(PI, OtherUse);
- }
- }
- }
- DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
- return NewPtr == &NewAI;
- }
-
- /// Select instructions that use an alloca and are subsequently loaded can be
- /// rewritten to load both input pointers and then select between the result,
- /// allowing the load of the alloca to be promoted.
- /// From this:
- /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
- /// %V = load i32* %P2
- /// to:
- /// %V1 = load i32* %Alloca -> will be mem2reg'd
- /// %V2 = load i32* %Other
- /// %V = select i1 %cond, i32 %V1, i32 %V2
- ///
- /// We can do this to a select if its only uses are loads and if the operand
- /// to the select can be loaded unconditionally.
- bool isSafeSelectToSpeculate(SelectInst &SI,
- SmallVectorImpl<LoadInst *> &Loads) {
- Value *TValue = SI.getTrueValue();
- Value *FValue = SI.getFalseValue();
- bool TDerefable = TValue->isDereferenceablePointer();
- bool FDerefable = FValue->isDereferenceablePointer();
-
- for (Value::use_iterator UI = SI.use_begin(), UE = SI.use_end();
- UI != UE; ++UI) {
- LoadInst *LI = dyn_cast<LoadInst>(*UI);
- if (LI == 0 || !LI->isSimple()) return false;
-
- // Both operands to the select need to be dereferencable, either
- // absolutely (e.g. allocas) or at this point because we can see other
- // accesses to it.
- if (!TDerefable && !isSafeToLoadUnconditionally(TValue, LI,
- LI->getAlignment(), &TD))
- return false;
- if (!FDerefable && !isSafeToLoadUnconditionally(FValue, LI,
- LI->getAlignment(), &TD))
- return false;
- Loads.push_back(LI);
- }
-
- return true;
+ DEBUG(dbgs() << " to: " << PN << "\n");
+ deleteIfTriviallyDead(OldPtr);
+ return false;
}
bool visitSelectInst(SelectInst &SI) {
assert(SI.getFalseValue() != OldPtr && "Pointer is both operands!");
else
assert(SI.getFalseValue() == OldPtr && "Pointer isn't an operand!");
- Value *NewPtr = getAdjustedAllocaPtr(IRB, OldPtr->getType());
-
- // If the select isn't safe to speculate, just use simple logic to emit it.
- SmallVector<LoadInst *, 4> Loads;
- if (!isSafeSelectToSpeculate(SI, Loads)) {
- SI.setOperand(IsTrueVal ? 1 : 2, NewPtr);
- DEBUG(dbgs() << " to: " << SI << "\n");
- deleteIfTriviallyDead(OldPtr);
- return false;
- }
-
- Value *OtherPtr = IsTrueVal ? SI.getFalseValue() : SI.getTrueValue();
- AllocaPartitioning::iterator PI
- = P.findPartitionForPHIOrSelectOperand(SI, OtherPtr);
- AllocaPartitioning::PartitionUse OtherUse;
- if (PI != P.end()) {
- // If the other pointer is within the partitioning, remove the select
- // from its uses. We'll add in the new loads below.
- AllocaPartitioning::use_iterator UI
- = P.findPartitionUseForPHIOrSelectOperand(SI, OtherPtr);
- OtherUse = *UI;
- P.use_erase(PI, UI);
- }
-
- Value *TV = IsTrueVal ? NewPtr : SI.getTrueValue();
- Value *FV = IsTrueVal ? SI.getFalseValue() : NewPtr;
- // Replace the loads of the select with a select of two loads.
- while (!Loads.empty()) {
- LoadInst *LI = Loads.pop_back_val();
-
- IRB.SetInsertPoint(LI);
- LoadInst *TL =
- IRB.CreateLoad(TV, getName("." + LI->getName() + ".true"));
- LoadInst *FL =
- IRB.CreateLoad(FV, getName("." + LI->getName() + ".false"));
- NumLoadsSpeculated += 2;
- if (PI != P.end()) {
- LoadInst *OtherLoad = IsTrueVal ? FL : TL;
- assert(OtherUse.Ptr == OtherLoad->getOperand(0));
- OtherUse.User = OtherLoad;
- P.use_push_back(PI, OtherUse);
- }
-
- // Transfer alignment and TBAA info if present.
- TL->setAlignment(LI->getAlignment());
- FL->setAlignment(LI->getAlignment());
- if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
- TL->setMetadata(LLVMContext::MD_tbaa, Tag);
- FL->setMetadata(LLVMContext::MD_tbaa, Tag);
- }
-
- Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL);
- V->takeName(LI);
- DEBUG(dbgs() << " speculated to: " << *V << "\n");
- LI->replaceAllUsesWith(V);
- Pass.DeadInsts.push_back(LI);
- }
+ Value *NewPtr = getAdjustedAllocaPtr(IRB, OldPtr->getType());
+ SI.setOperand(IsTrueVal ? 1 : 2, NewPtr);
+ DEBUG(dbgs() << " to: " << SI << "\n");
deleteIfTriviallyDead(OldPtr);
- return NewPtr == &NewAI;
+ return false;
}
};
AllocaPartitioning &P,
AllocaPartitioning::iterator PI) {
uint64_t AllocaSize = PI->EndOffset - PI->BeginOffset;
- if (P.use_begin(PI) == P.use_end(PI))
+ bool IsLive = false;
+ for (AllocaPartitioning::use_iterator UI = P.use_begin(PI),
+ UE = P.use_end(PI);
+ UI != UE && !IsLive; ++UI)
+ if (UI->U)
+ IsLive = true;
+ if (!IsLive)
return false; // No live uses left of this partition.
+ DEBUG(dbgs() << "Speculating PHIs and selects in partition "
+ << "[" << PI->BeginOffset << "," << PI->EndOffset << ")\n");
+
+ PHIOrSelectSpeculator Speculator(*TD, P, *this);
+ DEBUG(dbgs() << " speculating ");
+ DEBUG(P.print(dbgs(), PI, ""));
+ Speculator.visitUsers(PI);
+
// Try to compute a friendly type for this partition of the alloca. This
// won't always succeed, in which case we fall back to a legal integer type
// or an i8 array of an appropriate size.
assert(PI == P.begin() && "Begin offset is zero on later partition");
NewAI = &AI;
} else {
- // FIXME: The alignment here is overly conservative -- we could in many
- // cases get away with much weaker alignment constraints.
- NewAI = new AllocaInst(AllocaTy, 0, AI.getAlignment(),
+ unsigned Alignment = AI.getAlignment();
+ if (!Alignment) {
+ // The minimum alignment which users can rely on when the explicit
+ // alignment is omitted or zero is that required by the ABI for this
+ // type.
+ Alignment = TD->getABITypeAlignment(AI.getAllocatedType());
+ }
+ Alignment = MinAlign(Alignment, PI->BeginOffset);
+ // If we will get at least this much alignment from the type alone, leave
+ // the alloca's alignment unconstrained.
+ if (Alignment <= TD->getABITypeAlignment(AllocaTy))
+ Alignment = 0;
+ NewAI = new AllocaInst(AllocaTy, 0, Alignment,
AI.getName() + ".sroa." + Twine(PI - P.begin()),
&AI);
++NumNewAllocas;
<< "[" << PI->BeginOffset << "," << PI->EndOffset << ") to: "
<< *NewAI << "\n");
+ // Track the high watermark of the post-promotion worklist. We will reset it
+ // to this point if the alloca is not in fact scheduled for promotion.
+ unsigned PPWOldSize = PostPromotionWorklist.size();
+
AllocaPartitionRewriter Rewriter(*TD, P, PI, *this, AI, *NewAI,
PI->BeginOffset, PI->EndOffset);
DEBUG(dbgs() << " rewriting ");
DEBUG(P.print(dbgs(), PI, ""));
- if (Rewriter.visitUsers(P.use_begin(PI), P.use_end(PI))) {
+ bool Promotable = Rewriter.visitUsers(P.use_begin(PI), P.use_end(PI));
+ if (Promotable) {
DEBUG(dbgs() << " and queuing for promotion\n");
PromotableAllocas.push_back(NewAI);
} else if (NewAI != &AI) {
// alloca which didn't actually change and didn't get promoted.
Worklist.insert(NewAI);
}
+
+ // Drop any post-promotion work items if promotion didn't happen.
+ if (!Promotable)
+ while (PostPromotionWorklist.size() > PPWOldSize)
+ PostPromotionWorklist.pop_back();
+
return true;
}
TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
return false;
- // First check if this is a non-aggregate type that we should simply promote.
- if (!AI.getAllocatedType()->isAggregateType() && isAllocaPromotable(&AI)) {
- DEBUG(dbgs() << " Trivially scalar type, queuing for promotion...\n");
- PromotableAllocas.push_back(&AI);
- return false;
- }
-
bool Changed = false;
// First, split any FCA loads and stores touching this alloca to promote
if (PromotableAllocas.empty())
return false;
- // Ensure that the list is unique.
- std::sort(PromotableAllocas.begin(), PromotableAllocas.end());
- PromotableAllocas.erase(std::unique(PromotableAllocas.begin(),
- PromotableAllocas.end()),
- PromotableAllocas.end());
-
NumPromoted += PromotableAllocas.size();
if (DT && !ForceSSAUpdater) {
const SetType &Set;
public:
+ typedef AllocaInst *argument_type;
+
IsAllocaInSet(const SetType &Set) : Set(Set) {}
- bool operator()(AllocaInst *AI) { return Set.count(AI); }
+ bool operator()(AllocaInst *AI) const { return Set.count(AI); }
};
}
// the list of promotable allocas.
SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
- while (!Worklist.empty()) {
- Changed |= runOnAlloca(*Worklist.pop_back_val());
- deleteDeadInstructions(DeletedAllocas);
- if (!DeletedAllocas.empty()) {
- PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
- PromotableAllocas.end(),
- IsAllocaInSet(DeletedAllocas)),
- PromotableAllocas.end());
- DeletedAllocas.clear();
+ do {
+ while (!Worklist.empty()) {
+ Changed |= runOnAlloca(*Worklist.pop_back_val());
+ deleteDeadInstructions(DeletedAllocas);
+
+ // Remove the deleted allocas from various lists so that we don't try to
+ // continue processing them.
+ if (!DeletedAllocas.empty()) {
+ Worklist.remove_if(IsAllocaInSet(DeletedAllocas));
+ PostPromotionWorklist.remove_if(IsAllocaInSet(DeletedAllocas));
+ PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
+ PromotableAllocas.end(),
+ IsAllocaInSet(DeletedAllocas)),
+ PromotableAllocas.end());
+ DeletedAllocas.clear();
+ }
}
- }
- Changed |= promoteAllocas(F);
+ Changed |= promoteAllocas(F);
+
+ Worklist = PostPromotionWorklist;
+ PostPromotionWorklist.clear();
+ } while (!Worklist.empty());
return Changed;
}
projects / oota-llvm.git / blobdiff