#include "llvm/Support/InstVisitor.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/raw_ostream.h"
-#include "llvm/Target/TargetData.h"
+#include "llvm/DataLayout.h"
#include "llvm/Transforms/Utils/Local.h"
#include "llvm/Transforms/Utils/PromoteMemToReg.h"
#include "llvm/Transforms/Utils/SSAUpdater.h"
/// splittable and eagerly split them into scalar values.
bool IsSplittable;
+ /// \brief Test whether a partition has been marked as dead.
+ bool isDead() const {
+ if (BeginOffset == UINT64_MAX) {
+ assert(EndOffset == UINT64_MAX);
+ return true;
+ }
+ return false;
+ }
+
+ /// \brief Kill a partition.
+ /// This is accomplished by setting both its beginning and end offset to
+ /// the maximum possible value.
+ void kill() {
+ assert(!isDead() && "He's Dead, Jim!");
+ BeginOffset = EndOffset = UINT64_MAX;
+ }
+
Partition() : ByteRange(), IsSplittable() {}
Partition(uint64_t BeginOffset, uint64_t EndOffset, bool IsSplittable)
: ByteRange(BeginOffset, EndOffset), IsSplittable(IsSplittable) {}
///
/// Construction does most of the work for partitioning the alloca. This
/// performs the necessary walks of users and builds a partitioning from it.
- AllocaPartitioning(const TargetData &TD, AllocaInst &AI);
+ AllocaPartitioning(const DataLayout &TD, AllocaInst &AI);
/// \brief Test whether a pointer to the allocation escapes our analysis.
///
/// correctly represent. We stash extra data to help us untangle this
/// after the partitioning is complete.
struct MemTransferOffsets {
+ /// The destination begin and end offsets when the destination is within
+ /// this alloca. If the end offset is zero the destination is not within
+ /// this alloca.
uint64_t DestBegin, DestEnd;
+
+ /// The source begin and end offsets when the source is within this alloca.
+ /// If the end offset is zero, the source is not within this alloca.
uint64_t SourceBegin, SourceEnd;
+
+ /// Flag for whether an alloca is splittable.
bool IsSplittable;
};
MemTransferOffsets getMemTransferOffsets(MemTransferInst &II) const {
class AllocaPartitioning::BuilderBase
: public InstVisitor<DerivedT, RetT> {
public:
- BuilderBase(const TargetData &TD, AllocaInst &AI, AllocaPartitioning &P)
+ BuilderBase(const DataLayout &TD, AllocaInst &AI, AllocaPartitioning &P)
: TD(TD),
AllocSize(TD.getTypeAllocSize(AI.getAllocatedType())),
P(P) {
}
protected:
- const TargetData &TD;
+ const DataLayout &TD;
const uint64_t AllocSize;
AllocaPartitioning &P;
SmallDenseMap<Instruction *, unsigned> MemTransferPartitionMap;
public:
- PartitionBuilder(const TargetData &TD, AllocaInst &AI, AllocaPartitioning &P)
+ PartitionBuilder(const DataLayout &TD, AllocaInst &AI, AllocaPartitioning &P)
: BuilderBase<PartitionBuilder, bool>(TD, AI, P) {}
/// \brief Run the builder over the allocation.
EndOffset = AllocSize;
}
- // See if we can just add a user onto the last slot currently occupied.
- if (!P.Partitions.empty() &&
- P.Partitions.back().BeginOffset == BeginOffset &&
- P.Partitions.back().EndOffset == EndOffset) {
- P.Partitions.back().IsSplittable &= IsSplittable;
- return;
- }
-
Partition New(BeginOffset, EndOffset, IsSplittable);
P.Partitions.push_back(New);
}
// Only intrinsics with a constant length can be split.
Offsets.IsSplittable = Length;
- if (*U != II.getRawDest()) {
- assert(*U == II.getRawSource());
- Offsets.SourceBegin = Offset;
- Offsets.SourceEnd = Offset + Size;
- } else {
+ if (*U == II.getRawDest()) {
Offsets.DestBegin = Offset;
Offsets.DestEnd = Offset + Size;
}
+ if (*U == II.getRawSource()) {
+ Offsets.SourceBegin = Offset;
+ Offsets.SourceEnd = Offset + Size;
+ }
- insertUse(II, Offset, Size, Offsets.IsSplittable);
- unsigned NewIdx = P.Partitions.size() - 1;
-
- SmallDenseMap<Instruction *, unsigned>::const_iterator PMI;
- bool Inserted = false;
- llvm::tie(PMI, Inserted)
- = MemTransferPartitionMap.insert(std::make_pair(&II, NewIdx));
- 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. 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.
+ // If we have set up end offsets for both the source and the destination,
+ // we have found both sides of this transfer pointing at the same alloca.
+ bool SeenBothEnds = Offsets.SourceEnd && Offsets.DestEnd;
+ if (SeenBothEnds && II.getRawDest() != II.getRawSource()) {
+ unsigned PrevIdx = MemTransferPartitionMap[&II];
+
+ // Check if the begin offsets match and this is a non-volatile transfer.
+ // In that case, we can completely elide the transfer.
+ if (!II.isVolatile() && Offsets.SourceBegin == Offsets.DestBegin) {
+ P.Partitions[PrevIdx].kill();
+ return true;
+ }
+
+ // Otherwise we have an offset transfer within the same alloca. We can't
+ // split those.
+ P.Partitions[PrevIdx].IsSplittable = Offsets.IsSplittable = false;
+ } else if (SeenBothEnds) {
+ // Handle the case where this exact use provides both ends of the
+ // operation.
+ assert(II.getRawDest() == II.getRawSource());
+
+ // For non-volatile transfers this is a no-op.
+ if (!II.isVolatile())
+ return true;
+
+ // Otherwise just suppress splitting.
Offsets.IsSplittable = false;
- P.Partitions[PMI->second].IsSplittable = false;
- P.Partitions[NewIdx].IsSplittable = false;
+ }
+
+
+ // Insert the use now that we've fixed up the splittable nature.
+ insertUse(II, Offset, Size, Offsets.IsSplittable);
+
+ // Setup the mapping from intrinsic to partition of we've not seen both
+ // ends of this transfer.
+ if (!SeenBothEnds) {
+ unsigned NewIdx = P.Partitions.size() - 1;
+ bool Inserted
+ = MemTransferPartitionMap.insert(std::make_pair(&II, NewIdx)).second;
+ assert(Inserted &&
+ "Already have intrinsic in map but haven't seen both ends");
+ (void)Inserted;
}
return true;
SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
public:
- UseBuilder(const TargetData &TD, AllocaInst &AI, AllocaPartitioning &P)
+ UseBuilder(const DataLayout &TD, AllocaInst &AI, AllocaPartitioning &P)
: BuilderBase<UseBuilder>(TD, AI, P) {}
/// \brief Run the builder over the allocation.
void visitMemTransferInst(MemTransferInst &II) {
ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
+ if (!Size)
+ return markAsDead(II);
+
+ MemTransferOffsets &Offsets = P.MemTransferInstData[&II];
+ if (!II.isVolatile() && Offsets.DestEnd && Offsets.SourceEnd &&
+ Offsets.DestBegin == Offsets.SourceBegin)
+ return markAsDead(II); // Skip identity transfers without side-effects.
+
insertUse(II, Offset, Size);
}
SplitEndOffset = std::max(SplitEndOffset, Partitions[j].EndOffset);
}
- Partitions[j].BeginOffset = Partitions[j].EndOffset = UINT64_MAX;
+ Partitions[j].kill();
++NumDeadPartitions;
++j;
}
if (New.BeginOffset != New.EndOffset)
Partitions.push_back(New);
// Mark the old one for removal.
- Partitions[i].BeginOffset = Partitions[i].EndOffset = UINT64_MAX;
+ Partitions[i].kill();
++NumDeadPartitions;
}
// replaced in the process.
std::sort(Partitions.begin(), Partitions.end());
if (NumDeadPartitions) {
- assert(Partitions.back().BeginOffset == UINT64_MAX);
- assert(Partitions.back().EndOffset == UINT64_MAX);
+ assert(Partitions.back().isDead());
assert((ptrdiff_t)NumDeadPartitions ==
std::count(Partitions.begin(), Partitions.end(), Partitions.back()));
}
Partitions.erase(Partitions.end() - NumDeadPartitions, Partitions.end());
}
-AllocaPartitioning::AllocaPartitioning(const TargetData &TD, AllocaInst &AI)
+AllocaPartitioning::AllocaPartitioning(const DataLayout &TD, AllocaInst &AI)
:
#ifndef NDEBUG
AI(AI),
if (!PB())
return;
- if (Partitions.size() > 1) {
- // Sort the uses. This arranges for the offsets to be in ascending order,
- // and the sizes to be in descending order.
- std::sort(Partitions.begin(), Partitions.end());
+ // Sort the uses. This arranges for the offsets to be in ascending order,
+ // and the sizes to be in descending order.
+ std::sort(Partitions.begin(), Partitions.end());
+ // Remove any partitions from the back which are marked as dead.
+ while (!Partitions.empty() && Partitions.back().isDead())
+ Partitions.pop_back();
+
+ if (Partitions.size() > 1) {
// Intersect splittability for all partitions with equal offsets and sizes.
// Then remove all but the first so that we have a sequence of non-equal but
// potentially overlapping partitions.
const bool RequiresDomTree;
LLVMContext *C;
- const TargetData *TD;
+ const DataLayout *TD;
DominatorTree *DT;
/// \brief Worklist of alloca instructions to simplify.
/// 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;
INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
false, false)
-/// \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
-/// GEP is computed and Offset is set to it. If the GEP has any non-constant
-/// operands, the function returns false and the value of Offset is unmodified.
-static bool accumulateGEPOffsets(const TargetData &TD, GEPOperator &GEP,
- APInt &Offset) {
- APInt GEPOffset(Offset.getBitWidth(), 0);
- for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
- GTI != GTE; ++GTI) {
- ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
- if (!OpC)
- return false;
- if (OpC->isZero()) continue;
+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>;
- // Handle a struct index, which adds its field offset to the pointer.
- if (StructType *STy = dyn_cast<StructType>(*GTI)) {
- unsigned ElementIdx = OpC->getZExtValue();
- const StructLayout *SL = TD.getStructLayout(STy);
- GEPOffset += APInt(Offset.getBitWidth(),
- SL->getElementOffset(ElementIdx));
- continue;
- }
+ const DataLayout &TD;
+ AllocaPartitioning &P;
+ SROA &Pass;
- APInt TypeSize(Offset.getBitWidth(),
- TD.getTypeAllocSize(GTI.getIndexedType()));
- if (VectorType *VTy = dyn_cast<VectorType>(*GTI)) {
- assert((VTy->getScalarSizeInBits() % 8) == 0 &&
- "vector element size is not a multiple of 8, cannot GEP over it");
- TypeSize = VTy->getScalarSizeInBits() / 8;
- }
+public:
+ PHIOrSelectSpeculator(const DataLayout &TD, AllocaPartitioning &P, SROA &Pass)
+ : TD(TD), P(P), Pass(Pass) {}
- GEPOffset += OpC->getValue().sextOrTrunc(Offset.getBitWidth()) * TypeSize;
+ /// \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()));
+ }
}
- Offset = GEPOffset;
- return true;
-}
-/// \brief Build a GEP out of a base pointer and indices.
-///
-/// 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(IRBuilder<> &IRB, Value *BasePtr,
- SmallVectorImpl<Value *> &Indices,
- const Twine &Prefix) {
- if (Indices.empty())
- return BasePtr;
+private:
+ // By default, skip this instruction.
+ void visitInstruction(Instruction &I) {}
- // A single zero index is a no-op, so check for this and avoid building a GEP
- // in that case.
- if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
- return BasePtr;
+ /// 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;
- return IRB.CreateInBoundsGEP(BasePtr, Indices, Prefix + ".idx");
-}
+ // 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;
-/// \brief Get a natural GEP off of the BasePtr walking through Ty toward
-/// TargetTy without changing the offset of the pointer.
-///
-/// This routine assumes we've already established a properly offset GEP with
-/// Indices, and arrived at the Ty type. The goal is to continue to GEP with
-/// zero-indices down through type layers until we find one the same as
-/// TargetTy. If we can't find one with the same type, we at least try to use
-/// one with the same size. If none of that works, we just produce the GEP as
-/// indicated by Indices to have the correct offset.
-static Value *getNaturalGEPWithType(IRBuilder<> &IRB, const TargetData &TD,
- Value *BasePtr, Type *Ty, Type *TargetTy,
- SmallVectorImpl<Value *> &Indices,
- const Twine &Prefix) {
- if (Ty == TargetTy)
- return buildGEP(IRB, BasePtr, Indices, Prefix);
+ // 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;
- // See if we can descend into a struct and locate a field with the correct
- // type.
- unsigned NumLayers = 0;
- Type *ElementTy = Ty;
- do {
- if (ElementTy->isPointerTy())
- break;
- if (SequentialType *SeqTy = dyn_cast<SequentialType>(ElementTy)) {
- ElementTy = SeqTy->getElementType();
- Indices.push_back(IRB.getInt(APInt(TD.getPointerSizeInBits(), 0)));
- } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
- ElementTy = *STy->element_begin();
- Indices.push_back(IRB.getInt32(0));
- } else {
- break;
+ MaxAlign = std::max(MaxAlign, LI->getAlignment());
+ Loads.push_back(LI);
}
- ++NumLayers;
- } while (ElementTy != TargetTy);
- if (ElementTy != TargetTy)
- Indices.erase(Indices.end() - NumLayers, Indices.end());
- return buildGEP(IRB, BasePtr, Indices, Prefix);
-}
+ // 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);
-/// \brief Recursively compute indices for a natural GEP.
-///
-/// This is the recursive step for getNaturalGEPWithOffset that walks down the
-/// element types adding appropriate indices for the GEP.
-static Value *getNaturalGEPRecursively(IRBuilder<> &IRB, const TargetData &TD,
- Value *Ptr, Type *Ty, APInt &Offset,
- Type *TargetTy,
- SmallVectorImpl<Value *> &Indices,
- const Twine &Prefix) {
- if (Offset == 0)
- return getNaturalGEPWithType(IRB, TD, Ptr, Ty, TargetTy, Indices, Prefix);
+ // 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;
- // We can't recurse through pointer types.
- if (Ty->isPointerTy())
- return 0;
+ // If the predecessor has a single successor, then the edge isn't
+ // critical.
+ if (TI->getNumSuccessors() == 1)
+ continue;
- // We try to analyze GEPs over vectors here, but note that these GEPs are
- // extremely poorly defined currently. The long-term goal is to remove GEPing
- // over a vector from the IR completely.
- if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
- unsigned ElementSizeInBits = VecTy->getScalarSizeInBits();
- if (ElementSizeInBits % 8)
- return 0; // GEPs over non-multiple of 8 size vector elements are invalid.
- APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
- APInt NumSkippedElements = Offset.udiv(ElementSize);
- if (NumSkippedElements.ugt(VecTy->getNumElements()))
- return 0;
- Offset -= NumSkippedElements * ElementSize;
- Indices.push_back(IRB.getInt(NumSkippedElements));
- return getNaturalGEPRecursively(IRB, TD, Ptr, VecTy->getElementType(),
- Offset, TargetTy, Indices, Prefix);
- }
+ // 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;
- if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
- Type *ElementTy = ArrTy->getElementType();
- APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
- APInt NumSkippedElements = Offset.udiv(ElementSize);
- if (NumSkippedElements.ugt(ArrTy->getNumElements()))
- return 0;
+ return false;
+ }
- Offset -= NumSkippedElements * ElementSize;
- Indices.push_back(IRB.getInt(NumSkippedElements));
- return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
- Indices, Prefix);
+ return true;
}
- StructType *STy = dyn_cast<StructType>(Ty);
- if (!STy)
- return 0;
+ void visitPHINode(PHINode &PN) {
+ DEBUG(dbgs() << " original: " << PN << "\n");
- const StructLayout *SL = TD.getStructLayout(STy);
- uint64_t StructOffset = Offset.getZExtValue();
- if (StructOffset >= SL->getSizeInBytes())
- return 0;
- unsigned Index = SL->getElementContainingOffset(StructOffset);
- Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
- Type *ElementTy = STy->getElementType(Index);
- if (Offset.uge(TD.getTypeAllocSize(ElementTy)))
- return 0; // The offset points into alignment padding.
-
- Indices.push_back(IRB.getInt32(Index));
- return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
- Indices, Prefix);
-}
-
-/// \brief Get a natural GEP from a base pointer to a particular offset and
-/// resulting in a particular type.
-///
-/// The goal is to produce a "natural" looking GEP that works with the existing
-/// composite types to arrive at the appropriate offset and element type for
-/// a pointer. TargetTy is the element type the returned GEP should point-to if
-/// possible. We recurse by decreasing Offset, adding the appropriate index to
-/// Indices, and setting Ty to the result subtype.
-///
-/// If no natural GEP can be constructed, this function returns null.
-static Value *getNaturalGEPWithOffset(IRBuilder<> &IRB, const TargetData &TD,
- Value *Ptr, APInt Offset, Type *TargetTy,
- SmallVectorImpl<Value *> &Indices,
- const Twine &Prefix) {
- PointerType *Ty = cast<PointerType>(Ptr->getType());
-
- // Don't consider any GEPs through an i8* as natural unless the TargetTy is
- // an i8.
- if (Ty == IRB.getInt8PtrTy() && TargetTy->isIntegerTy(8))
- return 0;
-
- Type *ElementTy = Ty->getElementType();
- if (!ElementTy->isSized())
- return 0; // We can't GEP through an unsized element.
- APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
- if (ElementSize == 0)
- return 0; // Zero-length arrays can't help us build a natural GEP.
- APInt NumSkippedElements = Offset.udiv(ElementSize);
-
- Offset -= NumSkippedElements * ElementSize;
- Indices.push_back(IRB.getInt(NumSkippedElements));
- return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
- Indices, Prefix);
-}
-
-/// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
-/// resulting pointer has PointerTy.
-///
-/// This tries very hard to compute a "natural" GEP which arrives at the offset
-/// and produces the pointer type desired. Where it cannot, it will try to use
-/// the natural GEP to arrive at the offset and bitcast to the type. Where that
-/// fails, it will try to use an existing i8* and GEP to the byte offset and
-/// bitcast to the type.
-///
-/// The strategy for finding the more natural GEPs is to peel off layers of the
-/// pointer, walking back through bit casts and GEPs, searching for a base
-/// pointer from which we can compute a natural GEP with the desired
-/// properities. 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(IRBuilder<> &IRB, const TargetData &TD,
- Value *Ptr, APInt Offset, Type *PointerTy,
- const Twine &Prefix) {
- // 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;
- Visited.insert(Ptr);
- SmallVector<Value *, 4> Indices;
-
- // We may end up computing an offset pointer that has the wrong type. If we
- // never are able to compute one directly that has the correct type, we'll
- // fall back to it, so keep it around here.
- Value *OffsetPtr = 0;
-
- // Remember any i8 pointer we come across to re-use if we need to do a raw
- // byte offset.
- Value *Int8Ptr = 0;
- APInt Int8PtrOffset(Offset.getBitWidth(), 0);
-
- Type *TargetTy = PointerTy->getPointerElementType();
-
- do {
- // First fold any existing GEPs into the offset.
- while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
- APInt GEPOffset(Offset.getBitWidth(), 0);
- if (!accumulateGEPOffsets(TD, *GEP, GEPOffset))
- break;
- Offset += GEPOffset;
- Ptr = GEP->getPointerOperand();
- if (!Visited.insert(Ptr))
- break;
- }
-
- // See if we can perform a natural GEP here.
- Indices.clear();
- if (Value *P = getNaturalGEPWithOffset(IRB, TD, Ptr, Offset, TargetTy,
- Indices, Prefix)) {
- if (P->getType() == PointerTy) {
- // Zap any offset pointer that we ended up computing in previous rounds.
- if (OffsetPtr && OffsetPtr->use_empty())
- if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
- I->eraseFromParent();
- return P;
- }
- if (!OffsetPtr) {
- OffsetPtr = P;
- }
- }
-
- // Stash this pointer if we've found an i8*.
- if (Ptr->getType()->isIntegerTy(8)) {
- Int8Ptr = Ptr;
- Int8PtrOffset = Offset;
- }
-
- // Peel off a layer of the pointer and update the offset appropriately.
- if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
- Ptr = cast<Operator>(Ptr)->getOperand(0);
- } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
- if (GA->mayBeOverridden())
- break;
- Ptr = GA->getAliasee();
- } else {
- break;
- }
- assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
- } while (Visited.insert(Ptr));
-
- if (!OffsetPtr) {
- if (!Int8Ptr) {
- Int8Ptr = IRB.CreateBitCast(Ptr, IRB.getInt8PtrTy(),
- Prefix + ".raw_cast");
- Int8PtrOffset = Offset;
- }
-
- OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
- IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
- Prefix + ".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, Prefix + ".cast");
-
- return Ptr;
-}
-
-/// \brief Test whether the given alloca partition can be promoted to a vector.
-///
-/// This is a quick test to check whether we can rewrite a particular alloca
-/// partition (and its newly formed alloca) into a vector alloca with only
-/// whole-vector loads and stores such that it could be promoted to a vector
-/// SSA value. We only can ensure this for a limited set of operations, and we
-/// don't want to do the rewrites unless we are confident that the result will
-/// be promotable, so we have an early test here.
-static bool isVectorPromotionViable(const TargetData &TD,
- Type *AllocaTy,
- AllocaPartitioning &P,
- uint64_t PartitionBeginOffset,
- uint64_t PartitionEndOffset,
- AllocaPartitioning::const_use_iterator I,
- AllocaPartitioning::const_use_iterator E) {
- VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
- if (!Ty)
- return false;
-
- uint64_t VecSize = TD.getTypeSizeInBits(Ty);
- uint64_t ElementSize = Ty->getScalarSizeInBits();
-
- // While the definition of LLVM vectors is bitpacked, we don't support sizes
- // that aren't byte sized.
- if (ElementSize % 8)
- return false;
- assert((VecSize % 8) == 0 && "vector size not a multiple of element size?");
- VecSize /= 8;
- 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 ||
- BeginIndex >= Ty->getNumElements())
- return false;
- uint64_t EndOffset = I->EndOffset - PartitionBeginOffset;
- uint64_t EndIndex = EndOffset / ElementSize;
- if (EndIndex * ElementSize != EndOffset ||
- EndIndex > Ty->getNumElements())
- return false;
+ SmallVector<LoadInst *, 4> Loads;
+ if (!isSafePHIToSpeculate(PN, Loads))
+ return;
- // FIXME: We should build shuffle vector instructions to handle
- // non-element-sized accesses.
- if ((EndOffset - BeginOffset) != ElementSize &&
- (EndOffset - BeginOffset) != VecSize)
- return false;
+ assert(!Loads.empty());
- if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I->U->getUser())) {
- if (MI->isVolatile())
- return false;
- if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(I->U->getUser())) {
- const AllocaPartitioning::MemTransferOffsets &MTO
- = P.getMemTransferOffsets(*MTI);
- if (!MTO.IsSplittable)
- return false;
- }
- } 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->U->getUser()) &&
- !isa<StoreInst>(I->U->getUser())) {
- return false;
- }
- }
- return true;
-}
+ Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
+ IRBuilder<> PHIBuilder(&PN);
+ PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
+ PN.getName() + ".sroa.speculated");
-/// \brief Test whether the given alloca partition can be promoted to an int.
-///
-/// This is a quick test to check whether we can rewrite a particular alloca
-/// partition (and its newly formed alloca) into an integer alloca suitable for
-/// promotion to an SSA value. We only can ensure this for a limited set of
-/// operations, and we don't want to do the rewrites unless we are confident
-/// that the result will be promotable, so we have an early test here.
-static bool isIntegerPromotionViable(const TargetData &TD,
- Type *AllocaTy,
- AllocaPartitioning &P,
- AllocaPartitioning::const_use_iterator I,
- AllocaPartitioning::const_use_iterator E) {
- IntegerType *Ty = dyn_cast<IntegerType>(AllocaTy);
- if (!Ty)
- return false;
+ // 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();
- // Check the uses to ensure the uses are (likely) promoteable integer uses.
- // Also ensure that the alloca has a covering load or store. We don't want
- // promote because of some other unsplittable entry (which we may make
- // splittable later) and lose the ability to promote each element access.
- bool WholeAllocaOp = false;
- for (; I != E; ++I) {
- if (!I->U)
- continue; // Skip dead use.
- 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->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->U->getUser())) {
- if (MI->isVolatile())
- return false;
- if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(I->U->getUser())) {
- const AllocaPartitioning::MemTransferOffsets &MTO
- = P.getMemTransferOffsets(*MTI);
- if (!MTO.IsSplittable)
- return false;
- }
- } else {
- return false;
- }
- }
- return WholeAllocaOp;
-}
+ // 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());
-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>;
+ // 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);
- const TargetData &TD;
- AllocaPartitioning &P;
- SROA &Pass;
+ 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);
-public:
- PHIOrSelectSpeculator(const TargetData &TD, AllocaPartitioning &P, SROA &Pass)
- : TD(TD), P(P), Pass(Pass) {}
+ Instruction *Ptr = dyn_cast<Instruction>(InVal);
+ if (!Ptr)
+ // No uses to rewrite.
+ continue;
- /// \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.
+ // Try to lookup and rewrite any partition uses corresponding to this phi
+ // input.
+ AllocaPartitioning::iterator PI
+ = P.findPartitionForPHIOrSelectOperand(InUse);
+ if (PI == P.end())
+ continue;
- visit(cast<Instruction>(PU.U->getUser()));
+ // 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");
}
-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.
+ /// 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 = phi [i32* %Alloca, i32* %Other]
+ /// %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 = phi [i32 %V1, i32 %V2]
+ /// %V = select i1 %cond, i32 %V1, i32 %V2
///
- /// We can do this to a select if its only uses are loads and if the operands
+ /// 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();
+ 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;
- // 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;
+ // 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);
+ }
- // 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;
+ return true;
+ }
- MaxAlign = std::max(MaxAlign, LI->getAlignment());
- Loads.push_back(LI);
+ 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
+/// GEP is computed and Offset is set to it. If the GEP has any non-constant
+/// operands, the function returns false and the value of Offset is unmodified.
+static bool accumulateGEPOffsets(const DataLayout &TD, GEPOperator &GEP,
+ APInt &Offset) {
+ APInt GEPOffset(Offset.getBitWidth(), 0);
+ for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
+ GTI != GTE; ++GTI) {
+ ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
+ if (!OpC)
+ return false;
+ if (OpC->isZero()) continue;
+
+ // Handle a struct index, which adds its field offset to the pointer.
+ if (StructType *STy = dyn_cast<StructType>(*GTI)) {
+ unsigned ElementIdx = OpC->getZExtValue();
+ const StructLayout *SL = TD.getStructLayout(STy);
+ GEPOffset += APInt(Offset.getBitWidth(),
+ SL->getElementOffset(ElementIdx));
+ continue;
+ }
+
+ APInt TypeSize(Offset.getBitWidth(),
+ TD.getTypeAllocSize(GTI.getIndexedType()));
+ if (VectorType *VTy = dyn_cast<VectorType>(*GTI)) {
+ assert((VTy->getScalarSizeInBits() % 8) == 0 &&
+ "vector element size is not a multiple of 8, cannot GEP over it");
+ TypeSize = VTy->getScalarSizeInBits() / 8;
+ }
+
+ GEPOffset += OpC->getValue().sextOrTrunc(Offset.getBitWidth()) * TypeSize;
+ }
+ Offset = GEPOffset;
+ return true;
+}
+
+/// \brief Build a GEP out of a base pointer and indices.
+///
+/// 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(IRBuilder<> &IRB, Value *BasePtr,
+ SmallVectorImpl<Value *> &Indices,
+ const Twine &Prefix) {
+ if (Indices.empty())
+ return BasePtr;
+
+ // A single zero index is a no-op, so check for this and avoid building a GEP
+ // in that case.
+ if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
+ return BasePtr;
+
+ return IRB.CreateInBoundsGEP(BasePtr, Indices, Prefix + ".idx");
+}
+
+/// \brief Get a natural GEP off of the BasePtr walking through Ty toward
+/// TargetTy without changing the offset of the pointer.
+///
+/// This routine assumes we've already established a properly offset GEP with
+/// Indices, and arrived at the Ty type. The goal is to continue to GEP with
+/// zero-indices down through type layers until we find one the same as
+/// TargetTy. If we can't find one with the same type, we at least try to use
+/// one with the same size. If none of that works, we just produce the GEP as
+/// indicated by Indices to have the correct offset.
+static Value *getNaturalGEPWithType(IRBuilder<> &IRB, const DataLayout &TD,
+ Value *BasePtr, Type *Ty, Type *TargetTy,
+ SmallVectorImpl<Value *> &Indices,
+ const Twine &Prefix) {
+ if (Ty == TargetTy)
+ return buildGEP(IRB, BasePtr, Indices, Prefix);
+
+ // See if we can descend into a struct and locate a field with the correct
+ // type.
+ unsigned NumLayers = 0;
+ Type *ElementTy = Ty;
+ do {
+ if (ElementTy->isPointerTy())
+ break;
+ if (SequentialType *SeqTy = dyn_cast<SequentialType>(ElementTy)) {
+ ElementTy = SeqTy->getElementType();
+ Indices.push_back(IRB.getInt(APInt(TD.getPointerSizeInBits(), 0)));
+ } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
+ if (STy->element_begin() == STy->element_end())
+ break; // Nothing left to descend into.
+ ElementTy = *STy->element_begin();
+ Indices.push_back(IRB.getInt32(0));
+ } else {
+ break;
}
+ ++NumLayers;
+ } while (ElementTy != TargetTy);
+ if (ElementTy != TargetTy)
+ Indices.erase(Indices.end() - NumLayers, Indices.end());
- // 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);
+ return buildGEP(IRB, BasePtr, Indices, Prefix);
+}
- // 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;
+/// \brief Recursively compute indices for a natural GEP.
+///
+/// This is the recursive step for getNaturalGEPWithOffset that walks down the
+/// element types adding appropriate indices for the GEP.
+static Value *getNaturalGEPRecursively(IRBuilder<> &IRB, const DataLayout &TD,
+ Value *Ptr, Type *Ty, APInt &Offset,
+ Type *TargetTy,
+ SmallVectorImpl<Value *> &Indices,
+ const Twine &Prefix) {
+ if (Offset == 0)
+ return getNaturalGEPWithType(IRB, TD, Ptr, Ty, TargetTy, Indices, Prefix);
- // If the predecessor has a single successor, then the edge isn't
- // critical.
- if (TI->getNumSuccessors() == 1)
- continue;
+ // We can't recurse through pointer types.
+ if (Ty->isPointerTy())
+ return 0;
- // 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;
+ // We try to analyze GEPs over vectors here, but note that these GEPs are
+ // extremely poorly defined currently. The long-term goal is to remove GEPing
+ // over a vector from the IR completely.
+ if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
+ unsigned ElementSizeInBits = VecTy->getScalarSizeInBits();
+ if (ElementSizeInBits % 8)
+ return 0; // GEPs over non-multiple of 8 size vector elements are invalid.
+ APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
+ APInt NumSkippedElements = Offset.udiv(ElementSize);
+ if (NumSkippedElements.ugt(VecTy->getNumElements()))
+ return 0;
+ Offset -= NumSkippedElements * ElementSize;
+ Indices.push_back(IRB.getInt(NumSkippedElements));
+ return getNaturalGEPRecursively(IRB, TD, Ptr, VecTy->getElementType(),
+ Offset, TargetTy, Indices, Prefix);
+ }
- return false;
- }
+ if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
+ Type *ElementTy = ArrTy->getElementType();
+ APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
+ APInt NumSkippedElements = Offset.udiv(ElementSize);
+ if (NumSkippedElements.ugt(ArrTy->getNumElements()))
+ return 0;
- return true;
+ Offset -= NumSkippedElements * ElementSize;
+ Indices.push_back(IRB.getInt(NumSkippedElements));
+ return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
+ Indices, Prefix);
}
- void visitPHINode(PHINode &PN) {
- DEBUG(dbgs() << " original: " << PN << "\n");
+ StructType *STy = dyn_cast<StructType>(Ty);
+ if (!STy)
+ return 0;
- SmallVector<LoadInst *, 4> Loads;
- if (!isSafePHIToSpeculate(PN, Loads))
- return;
+ const StructLayout *SL = TD.getStructLayout(STy);
+ uint64_t StructOffset = Offset.getZExtValue();
+ if (StructOffset >= SL->getSizeInBytes())
+ return 0;
+ unsigned Index = SL->getElementContainingOffset(StructOffset);
+ Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
+ Type *ElementTy = STy->getElementType(Index);
+ if (Offset.uge(TD.getTypeAllocSize(ElementTy)))
+ return 0; // The offset points into alignment padding.
- assert(!Loads.empty());
+ Indices.push_back(IRB.getInt32(Index));
+ return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
+ Indices, Prefix);
+}
- Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
- IRBuilder<> PHIBuilder(&PN);
- PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
- PN.getName() + ".sroa.speculated");
+/// \brief Get a natural GEP from a base pointer to a particular offset and
+/// resulting in a particular type.
+///
+/// The goal is to produce a "natural" looking GEP that works with the existing
+/// composite types to arrive at the appropriate offset and element type for
+/// a pointer. TargetTy is the element type the returned GEP should point-to if
+/// possible. We recurse by decreasing Offset, adding the appropriate index to
+/// Indices, and setting Ty to the result subtype.
+///
+/// If no natural GEP can be constructed, this function returns null.
+static Value *getNaturalGEPWithOffset(IRBuilder<> &IRB, const DataLayout &TD,
+ Value *Ptr, APInt Offset, Type *TargetTy,
+ SmallVectorImpl<Value *> &Indices,
+ const Twine &Prefix) {
+ PointerType *Ty = cast<PointerType>(Ptr->getType());
- // 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();
+ // Don't consider any GEPs through an i8* as natural unless the TargetTy is
+ // an i8.
+ if (Ty == IRB.getInt8PtrTy() && TargetTy->isIntegerTy(8))
+ return 0;
- // 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());
+ Type *ElementTy = Ty->getElementType();
+ if (!ElementTy->isSized())
+ return 0; // We can't GEP through an unsized element.
+ APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
+ if (ElementSize == 0)
+ return 0; // Zero-length arrays can't help us build a natural GEP.
+ APInt NumSkippedElements = Offset.udiv(ElementSize);
- // 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);
+ Offset -= NumSkippedElements * ElementSize;
+ Indices.push_back(IRB.getInt(NumSkippedElements));
+ return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
+ Indices, Prefix);
+}
- 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);
+/// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
+/// resulting pointer has PointerTy.
+///
+/// This tries very hard to compute a "natural" GEP which arrives at the offset
+/// and produces the pointer type desired. Where it cannot, it will try to use
+/// the natural GEP to arrive at the offset and bitcast to the type. Where that
+/// fails, it will try to use an existing i8* and GEP to the byte offset and
+/// bitcast to the type.
+///
+/// The strategy for finding the more natural GEPs is to peel off layers of the
+/// pointer, walking back through bit casts and GEPs, searching for a base
+/// pointer from which we can compute a natural GEP with the desired
+/// properities. 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(IRBuilder<> &IRB, const DataLayout &TD,
+ Value *Ptr, APInt Offset, Type *PointerTy,
+ const Twine &Prefix) {
+ // 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;
+ Visited.insert(Ptr);
+ SmallVector<Value *, 4> Indices;
- Instruction *Ptr = dyn_cast<Instruction>(InVal);
- if (!Ptr)
- // No uses to rewrite.
- continue;
+ // We may end up computing an offset pointer that has the wrong type. If we
+ // never are able to compute one directly that has the correct type, we'll
+ // fall back to it, so keep it around here.
+ Value *OffsetPtr = 0;
- // Try to lookup and rewrite any partition uses corresponding to this phi
- // input.
- AllocaPartitioning::iterator PI
- = P.findPartitionForPHIOrSelectOperand(InUse);
- if (PI == P.end())
- continue;
+ // Remember any i8 pointer we come across to re-use if we need to do a raw
+ // byte offset.
+ Value *Int8Ptr = 0;
+ APInt Int8PtrOffset(Offset.getBitWidth(), 0);
- // 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());
+ Type *TargetTy = PointerTy->getPointerElementType();
+
+ do {
+ // First fold any existing GEPs into the offset.
+ while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
+ APInt GEPOffset(Offset.getBitWidth(), 0);
+ if (!accumulateGEPOffsets(TD, *GEP, GEPOffset))
+ break;
+ Offset += GEPOffset;
+ Ptr = GEP->getPointerOperand();
+ if (!Visited.insert(Ptr))
+ break;
}
- 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();
+ // See if we can perform a natural GEP here.
+ Indices.clear();
+ if (Value *P = getNaturalGEPWithOffset(IRB, TD, Ptr, Offset, TargetTy,
+ Indices, Prefix)) {
+ if (P->getType() == PointerTy) {
+ // Zap any offset pointer that we ended up computing in previous rounds.
+ if (OffsetPtr && OffsetPtr->use_empty())
+ if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
+ I->eraseFromParent();
+ return P;
+ }
+ if (!OffsetPtr) {
+ OffsetPtr = P;
+ }
+ }
- 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;
+ // Stash this pointer if we've found an i8*.
+ if (Ptr->getType()->isIntegerTy(8)) {
+ Int8Ptr = Ptr;
+ Int8PtrOffset = Offset;
+ }
+
+ // Peel off a layer of the pointer and update the offset appropriately.
+ if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
+ Ptr = cast<Operator>(Ptr)->getOperand(0);
+ } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
+ if (GA->mayBeOverridden())
+ break;
+ Ptr = GA->getAliasee();
+ } else {
+ break;
+ }
+ assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
+ } while (Visited.insert(Ptr));
- // 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);
+ if (!OffsetPtr) {
+ if (!Int8Ptr) {
+ Int8Ptr = IRB.CreateBitCast(Ptr, IRB.getInt8PtrTy(),
+ Prefix + ".raw_cast");
+ Int8PtrOffset = Offset;
}
- return true;
+ OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
+ IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
+ Prefix + ".raw_idx");
}
+ Ptr = OffsetPtr;
- void visitSelectInst(SelectInst &SI) {
- DEBUG(dbgs() << " original: " << SI << "\n");
- IRBuilder<> IRB(&SI);
+ // On the off chance we were targeting i8*, guard the bitcast here.
+ if (Ptr->getType() != PointerTy)
+ Ptr = IRB.CreateBitCast(Ptr, PointerTy, Prefix + ".cast");
- // If the select isn't safe to speculate, just use simple logic to emit it.
- SmallVector<LoadInst *, 4> Loads;
- if (!isSafeSelectToSpeculate(SI, Loads))
- return;
+ return Ptr;
+}
- 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;
- }
- }
+/// \brief Test whether the given alloca partition can be promoted to a vector.
+///
+/// This is a quick test to check whether we can rewrite a particular alloca
+/// partition (and its newly formed alloca) into a vector alloca with only
+/// whole-vector loads and stores such that it could be promoted to a vector
+/// SSA value. We only can ensure this for a limited set of operations, and we
+/// don't want to do the rewrites unless we are confident that the result will
+/// be promotable, so we have an early test here.
+static bool isVectorPromotionViable(const DataLayout &TD,
+ Type *AllocaTy,
+ AllocaPartitioning &P,
+ uint64_t PartitionBeginOffset,
+ uint64_t PartitionEndOffset,
+ AllocaPartitioning::const_use_iterator I,
+ AllocaPartitioning::const_use_iterator E) {
+ VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
+ if (!Ty)
+ return false;
- 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();
+ uint64_t VecSize = TD.getTypeSizeInBits(Ty);
+ uint64_t ElementSize = Ty->getScalarSizeInBits();
- 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;
+ // While the definition of LLVM vectors is bitpacked, we don't support sizes
+ // that aren't byte sized.
+ if (ElementSize % 8)
+ return false;
+ assert((VecSize % 8) == 0 && "vector size not a multiple of element size?");
+ VecSize /= 8;
+ ElementSize /= 8;
- // 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);
- }
+ for (; I != E; ++I) {
+ if (!I->U)
+ continue; // Skip dead use.
- Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
- LI->getName() + ".sroa.speculated");
+ uint64_t BeginOffset = I->BeginOffset - PartitionBeginOffset;
+ uint64_t BeginIndex = BeginOffset / ElementSize;
+ if (BeginIndex * ElementSize != BeginOffset ||
+ BeginIndex >= Ty->getNumElements())
+ return false;
+ uint64_t EndOffset = I->EndOffset - PartitionBeginOffset;
+ uint64_t EndIndex = EndOffset / ElementSize;
+ if (EndIndex * ElementSize != EndOffset ||
+ EndIndex > Ty->getNumElements())
+ return false;
- 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]);
- }
+ // FIXME: We should build shuffle vector instructions to handle
+ // non-element-sized accesses.
+ if ((EndOffset - BeginOffset) != ElementSize &&
+ (EndOffset - BeginOffset) != VecSize)
+ return false;
+
+ if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I->U->getUser())) {
+ if (MI->isVolatile())
+ return false;
+ if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(I->U->getUser())) {
+ const AllocaPartitioning::MemTransferOffsets &MTO
+ = P.getMemTransferOffsets(*MTI);
+ if (!MTO.IsSplittable)
+ return false;
}
+ } 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->U->getUser()) &&
+ !isa<StoreInst>(I->U->getUser())) {
+ return false;
+ }
+ }
+ return true;
+}
- DEBUG(dbgs() << " speculated to: " << *V << "\n");
- LI->replaceAllUsesWith(V);
- Pass.DeadInsts.push_back(LI);
+/// \brief Test whether the given alloca partition can be promoted to an int.
+///
+/// This is a quick test to check whether we can rewrite a particular alloca
+/// partition (and its newly formed alloca) into an integer alloca suitable for
+/// promotion to an SSA value. We only can ensure this for a limited set of
+/// operations, and we don't want to do the rewrites unless we are confident
+/// that the result will be promotable, so we have an early test here.
+static bool isIntegerPromotionViable(const DataLayout &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 || 8*TD.getTypeStoreSize(Ty) != Ty->getBitWidth())
+ return false;
+
+ // Check the uses to ensure the uses are (likely) promoteable integer uses.
+ // Also ensure that the alloca has a covering load or store. We don't want
+ // promote because of some other unsplittable entry (which we may make
+ // splittable later) and lose the ability to promote each element access.
+ bool WholeAllocaOp = false;
+ for (; I != E; ++I) {
+ 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->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->U->getUser())) {
+ if (MI->isVolatile())
+ return false;
+ if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(I->U->getUser())) {
+ const AllocaPartitioning::MemTransferOffsets &MTO
+ = P.getMemTransferOffsets(*MTI);
+ if (!MTO.IsSplittable)
+ return false;
+ }
+ } else {
+ return false;
}
}
-};
+ return WholeAllocaOp;
+}
+namespace {
/// \brief Visitor to rewrite instructions using a partition of an alloca to
/// use a new alloca.
///
// Befriend the base class so it can delegate to private visit methods.
friend class llvm::InstVisitor<AllocaPartitionRewriter, bool>;
- const TargetData &TD;
+ const DataLayout &TD;
AllocaPartitioning &P;
SROA &Pass;
AllocaInst &OldAI, &NewAI;
const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
+ Type *NewAllocaTy;
// If we are rewriting an alloca partition which can be written as pure
// vector operations, we stash extra information here. When VecTy is
std::string NamePrefix;
public:
- AllocaPartitionRewriter(const TargetData &TD, AllocaPartitioning &P,
+ AllocaPartitionRewriter(const DataLayout &TD, AllocaPartitioning &P,
AllocaPartitioning::iterator PI,
SROA &Pass, AllocaInst &OldAI, AllocaInst &NewAI,
uint64_t NewBeginOffset, uint64_t NewEndOffset)
OldAI(OldAI), NewAI(NewAI),
NewAllocaBeginOffset(NewBeginOffset),
NewAllocaEndOffset(NewEndOffset),
+ NewAllocaTy(NewAI.getAllocatedType()),
VecTy(), ElementTy(), ElementSize(), IntPromotionTy(),
BeginOffset(), EndOffset() {
}
"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;
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!");
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);
+ 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")),
Pass.DeadInsts.push_back(I);
}
- Value *getValueCast(IRBuilder<> &IRB, Value *V, Type *Ty) {
+ /// \brief Test whether we can convert a value from the old to the new type.
+ ///
+ /// This predicate should be used to guard calls to convertValue in order to
+ /// ensure that we only try to convert viable values. The strategy is that we
+ /// will peel off single element struct and array wrappings to get to an
+ /// underlying value, and convert that value.
+ bool canConvertValue(Type *OldTy, Type *NewTy) {
+ if (OldTy == NewTy)
+ return true;
+ if (TD.getTypeSizeInBits(NewTy) != TD.getTypeSizeInBits(OldTy))
+ return false;
+ if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
+ return false;
+
+ if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
+ if (NewTy->isPointerTy() && OldTy->isPointerTy())
+ return true;
+ if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
+ return true;
+ return false;
+ }
+
+ return true;
+ }
+
+ Value *convertValue(IRBuilder<> &IRB, Value *V, Type *Ty) {
+ assert(canConvertValue(V->getType(), Ty) && "Value not convertable to type");
+ if (V->getType() == Ty)
+ return V;
if (V->getType()->isIntegerTy() && Ty->isPointerTy())
return IRB.CreateIntToPtr(V, Ty);
if (V->getType()->isPointerTy() && Ty->isIntegerTy())
getName(".load"));
}
if (Result->getType() != LI.getType())
- Result = getValueCast(IRB, Result, LI.getType());
+ Result = convertValue(IRB, Result, LI.getType());
LI.replaceAllUsesWith(Result);
Pass.DeadInsts.push_back(&LI);
if (IntPromotionTy)
return rewriteIntegerLoad(IRB, LI);
+ if (BeginOffset == NewAllocaBeginOffset &&
+ canConvertValue(NewAllocaTy, LI.getType())) {
+ Value *NewLI = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
+ LI.isVolatile(), getName(".load"));
+ Value *NewV = convertValue(IRB, NewLI, LI.getType());
+ LI.replaceAllUsesWith(NewV);
+ Pass.DeadInsts.push_back(&LI);
+
+ DEBUG(dbgs() << " to: " << *NewLI << "\n");
+ return !LI.isVolatile();
+ }
+
Value *NewPtr = getAdjustedAllocaPtr(IRB,
LI.getPointerOperand()->getType());
LI.setOperand(0, NewPtr);
if (V->getType() == ElementTy ||
BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset) {
if (V->getType() != ElementTy)
- V = getValueCast(IRB, V, ElementTy);
+ V = convertValue(IRB, V, ElementTy);
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);
+ V = convertValue(IRB, V, VecTy);
}
StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
Pass.DeadInsts.push_back(&SI);
if (IntPromotionTy)
return rewriteIntegerStore(IRB, SI);
+ Type *ValueTy = SI.getValueOperand()->getType();
+
+ // 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 (ValueTy->isPointerTy())
+ if (AllocaInst *AI = dyn_cast<AllocaInst>(SI.getValueOperand()
+ ->stripInBoundsOffsets()))
+ Pass.PostPromotionWorklist.insert(AI);
+
+ if (BeginOffset == NewAllocaBeginOffset &&
+ canConvertValue(ValueTy, NewAllocaTy)) {
+ Value *NewV = convertValue(IRB, SI.getValueOperand(), NewAllocaTy);
+ StoreInst *NewSI = IRB.CreateAlignedStore(NewV, &NewAI, NewAI.getAlignment(),
+ SI.isVolatile());
+ (void)NewSI;
+ Pass.DeadInsts.push_back(&SI);
+
+ DEBUG(dbgs() << " to: " << *NewSI << "\n");
+ return !SI.isVolatile();
+ }
+
Value *NewPtr = getAdjustedAllocaPtr(IRB,
SI.getPointerOperand()->getType());
SI.setOperand(1, NewPtr);
// Befriend the base class so it can delegate to private visit methods.
friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
- const TargetData &TD;
+ const DataLayout &TD;
/// Queue of pointer uses to analyze and potentially rewrite.
SmallVector<Use *, 8> Queue;
Use *U;
public:
- AggLoadStoreRewriter(const TargetData &TD) : TD(TD) {}
+ AggLoadStoreRewriter(const DataLayout &TD) : TD(TD) {}
/// Rewrite loads and stores through a pointer and all pointers derived from
/// it.
};
}
+/// \brief Strip aggregate type wrapping.
+///
+/// This removes no-op aggregate types wrapping an underlying type. It will
+/// strip as many layers of types as it can without changing either the type
+/// size or the allocated size.
+static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
+ if (Ty->isSingleValueType())
+ return Ty;
+
+ uint64_t AllocSize = DL.getTypeAllocSize(Ty);
+ uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
+
+ Type *InnerTy;
+ if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
+ InnerTy = ArrTy->getElementType();
+ } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
+ const StructLayout *SL = DL.getStructLayout(STy);
+ unsigned Index = SL->getElementContainingOffset(0);
+ InnerTy = STy->getElementType(Index);
+ } else {
+ return Ty;
+ }
+
+ if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
+ TypeSize > DL.getTypeSizeInBits(InnerTy))
+ return Ty;
+
+ return stripAggregateTypeWrapping(DL, InnerTy);
+}
+
/// \brief Try to find a partition of the aggregate type passed in for a given
/// offset and size.
///
/// when the size or offset cause either end of type-based partition to be off.
/// Also, this is a best-effort routine. It is reasonable to give up and not
/// return a type if necessary.
-static Type *getTypePartition(const TargetData &TD, Type *Ty,
+static Type *getTypePartition(const DataLayout &TD, Type *Ty,
uint64_t Offset, uint64_t Size) {
if (Offset == 0 && TD.getTypeAllocSize(Ty) == Size)
- return Ty;
+ return stripAggregateTypeWrapping(TD, Ty);
if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
// We can't partition pointers...
assert(Offset == 0);
if (Size == ElementSize)
- return ElementTy;
+ return stripAggregateTypeWrapping(TD, ElementTy);
assert(Size > ElementSize);
uint64_t NumElements = Size / ElementSize;
if (NumElements * ElementSize != Size)
assert(Offset == 0);
if (Size == ElementSize)
- return ElementTy;
+ return stripAggregateTypeWrapping(TD, ElementTy);
StructType::element_iterator EI = STy->element_begin() + Index,
EE = STy->element_end();
<< "[" << 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 (P.isEscaped())
return Changed;
- // No partitions to split. Leave the dead alloca for a later pass to clean up.
- if (P.begin() == P.end())
- return Changed;
-
// Delete all the dead users of this alloca before splitting and rewriting it.
for (AllocaPartitioning::dead_user_iterator DI = P.dead_user_begin(),
DE = P.dead_user_end();
}
}
+ // No partitions to split. Leave the dead alloca for a later pass to clean up.
+ if (P.begin() == P.end())
+ return Changed;
+
return splitAlloca(AI, P) || Changed;
}
bool SROA::runOnFunction(Function &F) {
DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
C = &F.getContext();
- TD = getAnalysisIfAvailable<TargetData>();
+ TD = getAnalysisIfAvailable<DataLayout>();
if (!TD) {
DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
return false;
// the list of promotable allocas.
SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
- 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));
- 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