#include "llvm/LLVMContext.h"
#include "llvm/Module.h"
#include "llvm/Pass.h"
-#include "llvm/Analysis/DominanceFrontier.h"
+#include "llvm/Analysis/DebugInfo.h"
+#include "llvm/Analysis/DIBuilder.h"
+#include "llvm/Analysis/Dominators.h"
+#include "llvm/Analysis/Loads.h"
#include "llvm/Analysis/ValueTracking.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Transforms/Utils/PromoteMemToReg.h"
#include "llvm/Support/IRBuilder.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/raw_ostream.h"
+#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
using namespace llvm;
STATISTIC(NumReplaced, "Number of allocas broken up");
STATISTIC(NumPromoted, "Number of allocas promoted");
+STATISTIC(NumAdjusted, "Number of scalar allocas adjusted to allow promotion");
STATISTIC(NumConverted, "Number of aggregates converted to scalar");
STATISTIC(NumGlobals, "Number of allocas copied from constant global");
namespace {
struct SROA : public FunctionPass {
- SROA(int T, bool hasDF, char &ID)
- : FunctionPass(ID), HasDomFrontiers(hasDF) {
+ SROA(int T, bool hasDT, char &ID)
+ : FunctionPass(ID), HasDomTree(hasDT) {
if (T == -1)
SRThreshold = 128;
else
bool performPromotion(Function &F);
private:
- bool HasDomFrontiers;
+ bool HasDomTree;
TargetData *TD;
/// DeadInsts - Keep track of instructions we have made dead, so that
/// information about the uses. All these fields are initialized to false
/// and set to true when something is learned.
struct AllocaInfo {
+ /// The alloca to promote.
+ AllocaInst *AI;
+
+ /// CheckedPHIs - This is a set of verified PHI nodes, to prevent infinite
+ /// looping and avoid redundant work.
+ SmallPtrSet<PHINode*, 8> CheckedPHIs;
+
/// isUnsafe - This is set to true if the alloca cannot be SROA'd.
bool isUnsafe : 1;
/// not set this.
bool hasALoadOrStore : 1;
- AllocaInfo()
- : isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false),
+ explicit AllocaInfo(AllocaInst *ai)
+ : AI(ai), isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false),
hasSubelementAccess(false), hasALoadOrStore(false) {}
};
unsigned SRThreshold;
- void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; }
+ void MarkUnsafe(AllocaInfo &I, Instruction *User) {
+ I.isUnsafe = true;
+ DEBUG(dbgs() << " Transformation preventing inst: " << *User << '\n');
+ }
bool isSafeAllocaToScalarRepl(AllocaInst *AI);
- void isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
- AllocaInfo &Info);
- void isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t &Offset,
- AllocaInfo &Info);
- void isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize,
- const Type *MemOpType, bool isStore, AllocaInfo &Info);
- bool TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size);
- uint64_t FindElementAndOffset(const Type *&T, uint64_t &Offset,
- const Type *&IdxTy);
+ void isSafeForScalarRepl(Instruction *I, uint64_t Offset, AllocaInfo &Info);
+ void isSafePHISelectUseForScalarRepl(Instruction *User, uint64_t Offset,
+ AllocaInfo &Info);
+ void isSafeGEP(GetElementPtrInst *GEPI, uint64_t &Offset, AllocaInfo &Info);
+ void isSafeMemAccess(uint64_t Offset, uint64_t MemSize,
+ Type *MemOpType, bool isStore, AllocaInfo &Info,
+ Instruction *TheAccess, bool AllowWholeAccess);
+ bool TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size);
+ uint64_t FindElementAndOffset(Type *&T, uint64_t &Offset,
+ Type *&IdxTy);
void DoScalarReplacement(AllocaInst *AI,
std::vector<AllocaInst*> &WorkList);
SmallVector<AllocaInst*, 32> &NewElts);
void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
SmallVector<AllocaInst*, 32> &NewElts);
+ void RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI,
+ uint64_t Offset,
+ SmallVector<AllocaInst*, 32> &NewElts);
void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
AllocaInst *AI,
SmallVector<AllocaInst*, 32> &NewElts);
void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
SmallVector<AllocaInst*, 32> &NewElts);
- static MemTransferInst *isOnlyCopiedFromConstantGlobal(AllocaInst *AI);
+ static MemTransferInst *isOnlyCopiedFromConstantGlobal(
+ AllocaInst *AI, SmallVector<Instruction*, 4> &ToDelete);
};
- // SROA_DF - SROA that uses DominanceFrontier.
- struct SROA_DF : public SROA {
+ // SROA_DT - SROA that uses DominatorTree.
+ struct SROA_DT : public SROA {
static char ID;
public:
- SROA_DF(int T = -1) : SROA(T, true, ID) {
- initializeSROA_DFPass(*PassRegistry::getPassRegistry());
+ SROA_DT(int T = -1) : SROA(T, true, ID) {
+ initializeSROA_DTPass(*PassRegistry::getPassRegistry());
}
// getAnalysisUsage - This pass does not require any passes, but we know it
// will not alter the CFG, so say so.
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
AU.addRequired<DominatorTree>();
- AU.addRequired<DominanceFrontier>();
AU.setPreservesCFG();
}
};
}
-char SROA_DF::ID = 0;
+char SROA_DT::ID = 0;
char SROA_SSAUp::ID = 0;
-INITIALIZE_PASS_BEGIN(SROA_DF, "scalarrepl",
- "Scalar Replacement of Aggregates (DF)", false, false)
+INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl",
+ "Scalar Replacement of Aggregates (DT)", false, false)
INITIALIZE_PASS_DEPENDENCY(DominatorTree)
-INITIALIZE_PASS_DEPENDENCY(DominanceFrontier)
-INITIALIZE_PASS_END(SROA_DF, "scalarrepl",
- "Scalar Replacement of Aggregates (DF)", false, false)
+INITIALIZE_PASS_END(SROA_DT, "scalarrepl",
+ "Scalar Replacement of Aggregates (DT)", false, false)
INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa",
"Scalar Replacement of Aggregates (SSAUp)", false, false)
// Public interface to the ScalarReplAggregates pass
FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold,
- bool UseDomFrontier) {
- if (UseDomFrontier)
- return new SROA_DF(Threshold);
+ bool UseDomTree) {
+ if (UseDomTree)
+ return new SROA_DT(Threshold);
return new SROA_SSAUp(Threshold);
}
/// optimization, which scans the uses of an alloca and determines if it can
/// rewrite it in terms of a single new alloca that can be mem2reg'd.
class ConvertToScalarInfo {
- /// AllocaSize - The size of the alloca being considered.
+ /// AllocaSize - The size of the alloca being considered in bytes.
unsigned AllocaSize;
const TargetData &TD;
/// which means that mem2reg can't promote it.
bool IsNotTrivial;
+ /// ScalarKind - Tracks the kind of alloca being considered for promotion,
+ /// computed based on the uses of the alloca rather than the LLVM type system.
+ enum {
+ Unknown,
+
+ // Accesses via GEPs that are consistent with element access of a vector
+ // type. This will not be converted into a vector unless there is a later
+ // access using an actual vector type.
+ ImplicitVector,
+
+ // Accesses via vector operations and GEPs that are consistent with the
+ // layout of a vector type.
+ Vector,
+
+ // An integer bag-of-bits with bitwise operations for insertion and
+ // extraction. Any combination of types can be converted into this kind
+ // of scalar.
+ Integer
+ } ScalarKind;
+
/// VectorTy - This tracks the type that we should promote the vector to if
/// it is possible to turn it into a vector. This starts out null, and if it
/// isn't possible to turn into a vector type, it gets set to VoidTy.
- const Type *VectorTy;
+ VectorType *VectorTy;
- /// HadAVector - True if there is at least one vector access to the alloca.
- /// We don't want to turn random arrays into vectors and use vector element
- /// insert/extract, but if there are element accesses to something that is
- /// also declared as a vector, we do want to promote to a vector.
- bool HadAVector;
+ /// HadNonMemTransferAccess - True if there is at least one access to the
+ /// alloca that is not a MemTransferInst. We don't want to turn structs into
+ /// large integers unless there is some potential for optimization.
+ bool HadNonMemTransferAccess;
public:
explicit ConvertToScalarInfo(unsigned Size, const TargetData &td)
- : AllocaSize(Size), TD(td) {
- IsNotTrivial = false;
- VectorTy = 0;
- HadAVector = false;
- }
+ : AllocaSize(Size), TD(td), IsNotTrivial(false), ScalarKind(Unknown),
+ VectorTy(0), HadNonMemTransferAccess(false) { }
AllocaInst *TryConvert(AllocaInst *AI);
private:
bool CanConvertToScalar(Value *V, uint64_t Offset);
- void MergeInType(const Type *In, uint64_t Offset);
+ void MergeInTypeForLoadOrStore(Type *In, uint64_t Offset);
+ bool MergeInVectorType(VectorType *VInTy, uint64_t Offset);
void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset);
- Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType,
+ Value *ConvertScalar_ExtractValue(Value *NV, Type *ToType,
uint64_t Offset, IRBuilder<> &Builder);
Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal,
uint64_t Offset, IRBuilder<> &Builder);
} // end anonymous namespace.
-/// IsVerbotenVectorType - Return true if this is a vector type ScalarRepl isn't
-/// allowed to form. We do this to avoid MMX types, which is a complete hack,
-/// but is required until the backend is fixed.
-static bool IsVerbotenVectorType(const VectorType *VTy, const Instruction *I) {
- StringRef Triple(I->getParent()->getParent()->getParent()->getTargetTriple());
- if (!Triple.startswith("i386") &&
- !Triple.startswith("x86_64"))
- return false;
-
- // Reject all the MMX vector types.
- switch (VTy->getNumElements()) {
- default: return false;
- case 1: return VTy->getElementType()->isIntegerTy(64);
- case 2: return VTy->getElementType()->isIntegerTy(32);
- case 4: return VTy->getElementType()->isIntegerTy(16);
- case 8: return VTy->getElementType()->isIntegerTy(8);
- }
-}
-
-
/// TryConvert - Analyze the specified alloca, and if it is safe to do so,
/// rewrite it to be a new alloca which is mem2reg'able. This returns the new
/// alloca if possible or null if not.
if (!CanConvertToScalar(AI, 0) || !IsNotTrivial)
return 0;
+ // If an alloca has only memset / memcpy uses, it may still have an Unknown
+ // ScalarKind. Treat it as an Integer below.
+ if (ScalarKind == Unknown)
+ ScalarKind = Integer;
+
+ if (ScalarKind == Vector && VectorTy->getBitWidth() != AllocaSize * 8)
+ ScalarKind = Integer;
+
// If we were able to find a vector type that can handle this with
// insert/extract elements, and if there was at least one use that had
// a vector type, promote this to a vector. We don't want to promote
// random stuff that doesn't use vectors (e.g. <9 x double>) because then
// we just get a lot of insert/extracts. If at least one vector is
// involved, then we probably really do have a union of vector/array.
- const Type *NewTy;
- if (VectorTy && VectorTy->isVectorTy() && HadAVector &&
- !IsVerbotenVectorType(cast<VectorType>(VectorTy), AI)) {
+ Type *NewTy;
+ if (ScalarKind == Vector) {
+ assert(VectorTy && "Missing type for vector scalar.");
DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = "
<< *VectorTy << '\n');
NewTy = VectorTy; // Use the vector type.
} else {
+ unsigned BitWidth = AllocaSize * 8;
+ if ((ScalarKind == ImplicitVector || ScalarKind == Integer) &&
+ !HadNonMemTransferAccess && !TD.fitsInLegalInteger(BitWidth))
+ return 0;
+
DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n");
// Create and insert the integer alloca.
- NewTy = IntegerType::get(AI->getContext(), AllocaSize*8);
+ NewTy = IntegerType::get(AI->getContext(), BitWidth);
}
AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin());
ConvertUsesToScalar(AI, NewAI, 0);
return NewAI;
}
-/// MergeInType - Add the 'In' type to the accumulated vector type (VectorTy)
-/// so far at the offset specified by Offset (which is specified in bytes).
+/// MergeInTypeForLoadOrStore - Add the 'In' type to the accumulated vector type
+/// (VectorTy) so far at the offset specified by Offset (which is specified in
+/// bytes).
///
/// There are two cases we handle here:
/// 1) A union of vector types of the same size and potentially its elements.
/// large) integer type with extract and insert operations where the loads
/// and stores would mutate the memory. We mark this by setting VectorTy
/// to VoidTy.
-void ConvertToScalarInfo::MergeInType(const Type *In, uint64_t Offset) {
+void ConvertToScalarInfo::MergeInTypeForLoadOrStore(Type *In,
+ uint64_t Offset) {
// If we already decided to turn this into a blob of integer memory, there is
// nothing to be done.
- if (VectorTy && VectorTy->isVoidTy())
+ if (ScalarKind == Integer)
return;
// If this could be contributing to a vector, analyze it.
// If the In type is a vector that is the same size as the alloca, see if it
// matches the existing VecTy.
- if (const VectorType *VInTy = dyn_cast<VectorType>(In)) {
- // Remember if we saw a vector type.
- HadAVector = true;
-
- if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) {
- // If we're storing/loading a vector of the right size, allow it as a
- // vector. If this the first vector we see, remember the type so that
- // we know the element size. If this is a subsequent access, ignore it
- // even if it is a differing type but the same size. Worst case we can
- // bitcast the resultant vectors.
- if (VectorTy == 0)
- VectorTy = VInTy;
+ if (VectorType *VInTy = dyn_cast<VectorType>(In)) {
+ if (MergeInVectorType(VInTy, Offset))
return;
- }
} else if (In->isFloatTy() || In->isDoubleTy() ||
(In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 &&
isPowerOf2_32(In->getPrimitiveSizeInBits()))) {
+ // Full width accesses can be ignored, because they can always be turned
+ // into bitcasts.
+ unsigned EltSize = In->getPrimitiveSizeInBits()/8;
+ if (EltSize == AllocaSize)
+ return;
+
// If we're accessing something that could be an element of a vector, see
// if the implied vector agrees with what we already have and if Offset is
// compatible with it.
- unsigned EltSize = In->getPrimitiveSizeInBits()/8;
if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 &&
- (VectorTy == 0 ||
- cast<VectorType>(VectorTy)->getElementType()
- ->getPrimitiveSizeInBits()/8 == EltSize)) {
- if (VectorTy == 0)
+ (!VectorTy || EltSize == VectorTy->getElementType()
+ ->getPrimitiveSizeInBits()/8)) {
+ if (!VectorTy) {
+ ScalarKind = ImplicitVector;
VectorTy = VectorType::get(In, AllocaSize/EltSize);
+ }
return;
}
}
// Otherwise, we have a case that we can't handle with an optimized vector
// form. We can still turn this into a large integer.
- VectorTy = Type::getVoidTy(In->getContext());
+ ScalarKind = Integer;
+}
+
+/// MergeInVectorType - Handles the vector case of MergeInTypeForLoadOrStore,
+/// returning true if the type was successfully merged and false otherwise.
+bool ConvertToScalarInfo::MergeInVectorType(VectorType *VInTy,
+ uint64_t Offset) {
+ if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) {
+ // If we're storing/loading a vector of the right size, allow it as a
+ // vector. If this the first vector we see, remember the type so that
+ // we know the element size. If this is a subsequent access, ignore it
+ // even if it is a differing type but the same size. Worst case we can
+ // bitcast the resultant vectors.
+ if (!VectorTy)
+ VectorTy = VInTy;
+ ScalarKind = Vector;
+ return true;
+ }
+
+ return false;
}
/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all
if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
// Don't break volatile loads.
- if (LI->isVolatile())
+ if (!LI->isSimple())
return false;
// Don't touch MMX operations.
if (LI->getType()->isX86_MMXTy())
return false;
- MergeInType(LI->getType(), Offset);
+ HadNonMemTransferAccess = true;
+ MergeInTypeForLoadOrStore(LI->getType(), Offset);
continue;
}
if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
// Storing the pointer, not into the value?
- if (SI->getOperand(0) == V || SI->isVolatile()) return false;
+ if (SI->getOperand(0) == V || !SI->isSimple()) return false;
// Don't touch MMX operations.
if (SI->getOperand(0)->getType()->isX86_MMXTy())
return false;
- MergeInType(SI->getOperand(0)->getType(), Offset);
+ HadNonMemTransferAccess = true;
+ MergeInTypeForLoadOrStore(SI->getOperand(0)->getType(), Offset);
continue;
}
if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
- IsNotTrivial = true; // Can't be mem2reg'd.
+ if (!onlyUsedByLifetimeMarkers(BCI))
+ IsNotTrivial = true; // Can't be mem2reg'd.
if (!CanConvertToScalar(BCI, Offset))
return false;
continue;
// Compute the offset that this GEP adds to the pointer.
SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
+ if (!GEP->getPointerOperandType()->isPointerTy())
+ return false;
uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(),
- &Indices[0], Indices.size());
+ Indices);
// See if all uses can be converted.
if (!CanConvertToScalar(GEP, Offset+GEPOffset))
return false;
IsNotTrivial = true; // Can't be mem2reg'd.
+ HadNonMemTransferAccess = true;
continue;
}
// If this is a constant sized memset of a constant value (e.g. 0) we can
// handle it.
if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
- // Store of constant value and constant size.
- if (!isa<ConstantInt>(MSI->getValue()) ||
- !isa<ConstantInt>(MSI->getLength()))
+ // Store of constant value.
+ if (!isa<ConstantInt>(MSI->getValue()))
+ return false;
+
+ // Store of constant size.
+ ConstantInt *Len = dyn_cast<ConstantInt>(MSI->getLength());
+ if (!Len)
return false;
+
+ // If the size differs from the alloca, we can only convert the alloca to
+ // an integer bag-of-bits.
+ // FIXME: This should handle all of the cases that are currently accepted
+ // as vector element insertions.
+ if (Len->getZExtValue() != AllocaSize || Offset != 0)
+ ScalarKind = Integer;
+
IsNotTrivial = true; // Can't be mem2reg'd.
+ HadNonMemTransferAccess = true;
continue;
}
continue;
}
+ // If this is a lifetime intrinsic, we can handle it.
+ if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
+ if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
+ II->getIntrinsicID() == Intrinsic::lifetime_end) {
+ continue;
+ }
+ }
+
// Otherwise, we cannot handle this!
return false;
}
// Compute the offset that this GEP adds to the pointer.
SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(),
- &Indices[0], Indices.size());
+ Indices);
ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8);
GEP->eraseFromParent();
continue;
if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
// The load is a bit extract from NewAI shifted right by Offset bits.
- Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp");
+ Value *LoadedVal = Builder.CreateLoad(NewAI);
Value *NewLoadVal
= ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder);
LI->replaceAllUsesWith(NewLoadVal);
// If the source and destination are both to the same alloca, then this is
// a noop copy-to-self, just delete it. Otherwise, emit a load and store
// as appropriate.
- AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, 0));
+ AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &TD, 0));
- if (GetUnderlyingObject(MTI->getSource(), 0) != OrigAI) {
+ if (GetUnderlyingObject(MTI->getSource(), &TD, 0) != OrigAI) {
// Dest must be OrigAI, change this to be a load from the original
// pointer (bitcasted), then a store to our new alloca.
assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?");
Value *SrcPtr = MTI->getSource();
- const PointerType* SPTy = cast<PointerType>(SrcPtr->getType());
- const PointerType* AIPTy = cast<PointerType>(NewAI->getType());
+ PointerType* SPTy = cast<PointerType>(SrcPtr->getType());
+ PointerType* AIPTy = cast<PointerType>(NewAI->getType());
if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
AIPTy = PointerType::get(AIPTy->getElementType(),
SPTy->getAddressSpace());
LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval");
SrcVal->setAlignment(MTI->getAlignment());
Builder.CreateStore(SrcVal, NewAI);
- } else if (GetUnderlyingObject(MTI->getDest(), 0) != OrigAI) {
+ } else if (GetUnderlyingObject(MTI->getDest(), &TD, 0) != OrigAI) {
// Src must be OrigAI, change this to be a load from NewAI then a store
// through the original dest pointer (bitcasted).
assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?");
LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval");
- const PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType());
- const PointerType* AIPTy = cast<PointerType>(NewAI->getType());
+ PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType());
+ PointerType* AIPTy = cast<PointerType>(NewAI->getType());
if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
AIPTy = PointerType::get(AIPTy->getElementType(),
DPTy->getAddressSpace());
continue;
}
+ if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
+ if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
+ II->getIntrinsicID() == Intrinsic::lifetime_end) {
+ // There's no need to preserve these, as the resulting alloca will be
+ // converted to a register anyways.
+ II->eraseFromParent();
+ continue;
+ }
+ }
+
llvm_unreachable("Unsupported operation!");
}
}
/// Offset is an offset from the original alloca, in bits that need to be
/// shifted to the right.
Value *ConvertToScalarInfo::
-ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType,
+ConvertScalar_ExtractValue(Value *FromVal, Type *ToType,
uint64_t Offset, IRBuilder<> &Builder) {
// If the load is of the whole new alloca, no conversion is needed.
- if (FromVal->getType() == ToType && Offset == 0)
+ Type *FromType = FromVal->getType();
+ if (FromType == ToType && Offset == 0)
return FromVal;
// If the result alloca is a vector type, this is either an element
// access or a bitcast to another vector type of the same size.
- if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) {
- if (ToType->isVectorTy())
- return Builder.CreateBitCast(FromVal, ToType, "tmp");
+ if (VectorType *VTy = dyn_cast<VectorType>(FromType)) {
+ unsigned FromTypeSize = TD.getTypeAllocSize(FromType);
+ unsigned ToTypeSize = TD.getTypeAllocSize(ToType);
+ if (FromTypeSize == ToTypeSize)
+ return Builder.CreateBitCast(FromVal, ToType);
// Otherwise it must be an element access.
unsigned Elt = 0;
assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
}
// Return the element extracted out of it.
- Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get(
- Type::getInt32Ty(FromVal->getContext()), Elt), "tmp");
+ Value *V = Builder.CreateExtractElement(FromVal, Builder.getInt32(Elt));
if (V->getType() != ToType)
- V = Builder.CreateBitCast(V, ToType, "tmp");
+ V = Builder.CreateBitCast(V, ToType);
return V;
}
// If ToType is a first class aggregate, extract out each of the pieces and
// use insertvalue's to form the FCA.
- if (const StructType *ST = dyn_cast<StructType>(ToType)) {
+ if (StructType *ST = dyn_cast<StructType>(ToType)) {
const StructLayout &Layout = *TD.getStructLayout(ST);
Value *Res = UndefValue::get(ST);
for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i),
Offset+Layout.getElementOffsetInBits(i),
Builder);
- Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
+ Res = Builder.CreateInsertValue(Res, Elt, i);
}
return Res;
}
- if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) {
+ if (ArrayType *AT = dyn_cast<ArrayType>(ToType)) {
uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType());
Value *Res = UndefValue::get(AT);
for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(),
Offset+i*EltSize, Builder);
- Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
+ Res = Builder.CreateInsertValue(Res, Elt, i);
}
return Res;
}
// Otherwise, this must be a union that was converted to an integer value.
- const IntegerType *NTy = cast<IntegerType>(FromVal->getType());
+ IntegerType *NTy = cast<IntegerType>(FromVal->getType());
// If this is a big-endian system and the load is narrower than the
// full alloca type, we need to do a shift to get the right bits.
// only some bits are used.
if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth())
FromVal = Builder.CreateLShr(FromVal,
- ConstantInt::get(FromVal->getType(),
- ShAmt), "tmp");
+ ConstantInt::get(FromVal->getType(), ShAmt));
else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
FromVal = Builder.CreateShl(FromVal,
- ConstantInt::get(FromVal->getType(),
- -ShAmt), "tmp");
+ ConstantInt::get(FromVal->getType(), -ShAmt));
// Finally, unconditionally truncate the integer to the right width.
unsigned LIBitWidth = TD.getTypeSizeInBits(ToType);
if (LIBitWidth < NTy->getBitWidth())
FromVal =
Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(),
- LIBitWidth), "tmp");
+ LIBitWidth));
else if (LIBitWidth > NTy->getBitWidth())
FromVal =
Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(),
- LIBitWidth), "tmp");
+ LIBitWidth));
// If the result is an integer, this is a trunc or bitcast.
if (ToType->isIntegerTy()) {
// Should be done.
} else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) {
// Just do a bitcast, we know the sizes match up.
- FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp");
+ FromVal = Builder.CreateBitCast(FromVal, ToType);
} else {
// Otherwise must be a pointer.
- FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp");
+ FromVal = Builder.CreateIntToPtr(FromVal, ToType);
}
assert(FromVal->getType() == ToType && "Didn't convert right?");
return FromVal;
uint64_t Offset, IRBuilder<> &Builder) {
// Convert the stored type to the actual type, shift it left to insert
// then 'or' into place.
- const Type *AllocaType = Old->getType();
+ Type *AllocaType = Old->getType();
LLVMContext &Context = Old->getContext();
- if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
+ if (VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy);
uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType());
// Changing the whole vector with memset or with an access of a different
// vector type?
if (ValSize == VecSize)
- return Builder.CreateBitCast(SV, AllocaType, "tmp");
-
- uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType());
+ return Builder.CreateBitCast(SV, AllocaType);
// Must be an element insertion.
+ Type *EltTy = VTy->getElementType();
+ if (SV->getType() != EltTy)
+ SV = Builder.CreateBitCast(SV, EltTy);
+ uint64_t EltSize = TD.getTypeAllocSizeInBits(EltTy);
unsigned Elt = Offset/EltSize;
-
- if (SV->getType() != VTy->getElementType())
- SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp");
-
- SV = Builder.CreateInsertElement(Old, SV,
- ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt),
- "tmp");
- return SV;
+ return Builder.CreateInsertElement(Old, SV, Builder.getInt32(Elt));
}
// If SV is a first-class aggregate value, insert each value recursively.
- if (const StructType *ST = dyn_cast<StructType>(SV->getType())) {
+ if (StructType *ST = dyn_cast<StructType>(SV->getType())) {
const StructLayout &Layout = *TD.getStructLayout(ST);
for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
- Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
+ Value *Elt = Builder.CreateExtractValue(SV, i);
Old = ConvertScalar_InsertValue(Elt, Old,
Offset+Layout.getElementOffsetInBits(i),
Builder);
return Old;
}
- if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) {
+ if (ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) {
uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType());
for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
- Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
+ Value *Elt = Builder.CreateExtractValue(SV, i);
Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder);
}
return Old;
unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType());
unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType);
if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy())
- SV = Builder.CreateBitCast(SV,
- IntegerType::get(SV->getContext(),SrcWidth), "tmp");
+ SV = Builder.CreateBitCast(SV, IntegerType::get(SV->getContext(),SrcWidth));
else if (SV->getType()->isPointerTy())
- SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp");
+ SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()));
// Zero extend or truncate the value if needed.
if (SV->getType() != AllocaType) {
if (SV->getType()->getPrimitiveSizeInBits() <
AllocaType->getPrimitiveSizeInBits())
- SV = Builder.CreateZExt(SV, AllocaType, "tmp");
+ SV = Builder.CreateZExt(SV, AllocaType);
else {
// Truncation may be needed if storing more than the alloca can hold
// (undefined behavior).
- SV = Builder.CreateTrunc(SV, AllocaType, "tmp");
+ SV = Builder.CreateTrunc(SV, AllocaType);
SrcWidth = DestWidth;
SrcStoreWidth = DestStoreWidth;
}
// only some bits in the structure are set.
APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
- SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(),
- ShAmt), "tmp");
+ SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), ShAmt));
Mask <<= ShAmt;
} else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
- SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(),
- -ShAmt), "tmp");
+ SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), -ShAmt));
Mask = Mask.lshr(-ShAmt);
}
namespace {
class AllocaPromoter : public LoadAndStorePromoter {
AllocaInst *AI;
+ DIBuilder *DIB;
+ SmallVector<DbgDeclareInst *, 4> DDIs;
+ SmallVector<DbgValueInst *, 4> DVIs;
public:
- AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S)
- : LoadAndStorePromoter(Insts, S), AI(0) {}
+ AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
+ DIBuilder *DB)
+ : LoadAndStorePromoter(Insts, S), AI(0), DIB(DB) {}
void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) {
// Remember which alloca we're promoting (for isInstInList).
this->AI = AI;
+ if (MDNode *DebugNode = MDNode::getIfExists(AI->getContext(), AI)) {
+ for (Value::use_iterator UI = DebugNode->use_begin(),
+ E = DebugNode->use_end(); UI != E; ++UI)
+ if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
+ DDIs.push_back(DDI);
+ else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI))
+ DVIs.push_back(DVI);
+ }
+
LoadAndStorePromoter::run(Insts);
AI->eraseFromParent();
+ for (SmallVector<DbgDeclareInst *, 4>::iterator I = DDIs.begin(),
+ E = DDIs.end(); I != E; ++I) {
+ DbgDeclareInst *DDI = *I;
+ DDI->eraseFromParent();
+ }
+ for (SmallVector<DbgValueInst *, 4>::iterator I = DVIs.begin(),
+ E = DVIs.end(); I != E; ++I) {
+ DbgValueInst *DVI = *I;
+ DVI->eraseFromParent();
+ }
}
virtual bool isInstInList(Instruction *I,
return LI->getOperand(0) == AI;
return cast<StoreInst>(I)->getPointerOperand() == AI;
}
+
+ virtual void updateDebugInfo(Instruction *Inst) const {
+ for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(),
+ E = DDIs.end(); I != E; ++I) {
+ DbgDeclareInst *DDI = *I;
+ if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
+ ConvertDebugDeclareToDebugValue(DDI, SI, *DIB);
+ else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
+ ConvertDebugDeclareToDebugValue(DDI, LI, *DIB);
+ }
+ for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(),
+ E = DVIs.end(); I != E; ++I) {
+ DbgValueInst *DVI = *I;
+ if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
+ Instruction *DbgVal = NULL;
+ // If an argument is zero extended then use argument directly. The ZExt
+ // may be zapped by an optimization pass in future.
+ Argument *ExtendedArg = NULL;
+ if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
+ ExtendedArg = dyn_cast<Argument>(ZExt->getOperand(0));
+ if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
+ ExtendedArg = dyn_cast<Argument>(SExt->getOperand(0));
+ if (ExtendedArg)
+ DbgVal = DIB->insertDbgValueIntrinsic(ExtendedArg, 0,
+ DIVariable(DVI->getVariable()),
+ SI);
+ else
+ DbgVal = DIB->insertDbgValueIntrinsic(SI->getOperand(0), 0,
+ DIVariable(DVI->getVariable()),
+ SI);
+ DbgVal->setDebugLoc(DVI->getDebugLoc());
+ } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
+ Instruction *DbgVal =
+ DIB->insertDbgValueIntrinsic(LI->getOperand(0), 0,
+ DIVariable(DVI->getVariable()), LI);
+ DbgVal->setDebugLoc(DVI->getDebugLoc());
+ }
+ }
+ }
};
} // end anon namespace
+/// isSafeSelectToSpeculate - 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.
+static bool isSafeSelectToSpeculate(SelectInst *SI, const TargetData *TD) {
+ bool TDerefable = SI->getTrueValue()->isDereferenceablePointer();
+ bool FDerefable = SI->getFalseValue()->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(SI->getTrueValue(), LI,
+ LI->getAlignment(), TD))
+ return false;
+ if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI,
+ LI->getAlignment(), TD))
+ return false;
+ }
+
+ return true;
+}
+
+/// isSafePHIToSpeculate - 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.
+static bool isSafePHIToSpeculate(PHINode *PN, const TargetData *TD) {
+ // 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());
+ }
+
+ // Okay, we know that we have one or more loads in the same block as the PHI.
+ // We can 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 i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
+ BasicBlock *Pred = PN->getIncomingBlock(i);
+ Value *InVal = PN->getIncomingValue(i);
+
+ // If the terminator of the predecessor has side-effects (an invoke),
+ // there is no safe place to put a load in the predecessor.
+ if (Pred->getTerminator()->mayHaveSideEffects())
+ return false;
+
+ // If the value is produced by the terminator of the predecessor
+ // (an invoke), there is no valid place to put a load in the predecessor.
+ if (Pred->getTerminator() == InVal)
+ return false;
+
+ // If the predecessor has a single successor, then the edge isn't critical.
+ if (Pred->getTerminator()->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, Pred->getTerminator(), MaxAlign, TD))
+ continue;
+
+ return false;
+ }
+
+ return true;
+}
+
+
+/// tryToMakeAllocaBePromotable - This returns true if the alloca only has
+/// direct (non-volatile) loads and stores to it. If the alloca is close but
+/// not quite there, this will transform the code to allow promotion. As such,
+/// it is a non-pure predicate.
+static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) {
+ SetVector<Instruction*, SmallVector<Instruction*, 4>,
+ SmallPtrSet<Instruction*, 4> > InstsToRewrite;
+
+ for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
+ UI != UE; ++UI) {
+ User *U = *UI;
+ if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
+ if (!LI->isSimple())
+ return false;
+ continue;
+ }
+
+ if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
+ if (SI->getOperand(0) == AI || !SI->isSimple())
+ return false; // Don't allow a store OF the AI, only INTO the AI.
+ continue;
+ }
+
+ if (SelectInst *SI = dyn_cast<SelectInst>(U)) {
+ // If the condition being selected on is a constant, fold the select, yes
+ // this does (rarely) happen early on.
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) {
+ Value *Result = SI->getOperand(1+CI->isZero());
+ SI->replaceAllUsesWith(Result);
+ SI->eraseFromParent();
+
+ // This is very rare and we just scrambled the use list of AI, start
+ // over completely.
+ return tryToMakeAllocaBePromotable(AI, TD);
+ }
+
+ // If it is safe to turn "load (select c, AI, ptr)" into a select of two
+ // loads, then we can transform this by rewriting the select.
+ if (!isSafeSelectToSpeculate(SI, TD))
+ return false;
+
+ InstsToRewrite.insert(SI);
+ continue;
+ }
+
+ if (PHINode *PN = dyn_cast<PHINode>(U)) {
+ if (PN->use_empty()) { // Dead PHIs can be stripped.
+ InstsToRewrite.insert(PN);
+ continue;
+ }
+
+ // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads
+ // in the pred blocks, then we can transform this by rewriting the PHI.
+ if (!isSafePHIToSpeculate(PN, TD))
+ return false;
+
+ InstsToRewrite.insert(PN);
+ continue;
+ }
+
+ if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
+ if (onlyUsedByLifetimeMarkers(BCI)) {
+ InstsToRewrite.insert(BCI);
+ continue;
+ }
+ }
+
+ return false;
+ }
+
+ // If there are no instructions to rewrite, then all uses are load/stores and
+ // we're done!
+ if (InstsToRewrite.empty())
+ return true;
+
+ // If we have instructions that need to be rewritten for this to be promotable
+ // take care of it now.
+ for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) {
+ if (BitCastInst *BCI = dyn_cast<BitCastInst>(InstsToRewrite[i])) {
+ // This could only be a bitcast used by nothing but lifetime intrinsics.
+ for (BitCastInst::use_iterator I = BCI->use_begin(), E = BCI->use_end();
+ I != E;) {
+ Use &U = I.getUse();
+ ++I;
+ cast<Instruction>(U.getUser())->eraseFromParent();
+ }
+ BCI->eraseFromParent();
+ continue;
+ }
+
+ if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) {
+ // Selects in InstsToRewrite only have load uses. Rewrite each as two
+ // loads with a new select.
+ while (!SI->use_empty()) {
+ LoadInst *LI = cast<LoadInst>(SI->use_back());
+
+ IRBuilder<> Builder(LI);
+ LoadInst *TrueLoad =
+ Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t");
+ LoadInst *FalseLoad =
+ Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".f");
+
+ // Transfer alignment and TBAA info if present.
+ TrueLoad->setAlignment(LI->getAlignment());
+ FalseLoad->setAlignment(LI->getAlignment());
+ if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
+ TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
+ FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
+ }
+
+ Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad);
+ V->takeName(LI);
+ LI->replaceAllUsesWith(V);
+ LI->eraseFromParent();
+ }
+
+ // Now that all the loads are gone, the select is gone too.
+ SI->eraseFromParent();
+ continue;
+ }
+
+ // Otherwise, we have a PHI node which allows us to push the loads into the
+ // predecessors.
+ PHINode *PN = cast<PHINode>(InstsToRewrite[i]);
+ if (PN->use_empty()) {
+ PN->eraseFromParent();
+ continue;
+ }
+
+ Type *LoadTy = cast<PointerType>(PN->getType())->getElementType();
+ PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(),
+ PN->getName()+".ld", 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>(PN->use_back());
+ MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
+ unsigned Align = SomeLoad->getAlignment();
+
+ // Rewrite all loads of the PN to use the new PHI.
+ while (!PN->use_empty()) {
+ LoadInst *LI = cast<LoadInst>(PN->use_back());
+ LI->replaceAllUsesWith(NewPN);
+ LI->eraseFromParent();
+ }
+
+ // Inject loads into all of the pred blocks. Keep track of which blocks we
+ // insert them into in case we have multiple edges from the same block.
+ DenseMap<BasicBlock*, LoadInst*> InsertedLoads;
+
+ for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
+ BasicBlock *Pred = PN->getIncomingBlock(i);
+ LoadInst *&Load = InsertedLoads[Pred];
+ if (Load == 0) {
+ Load = new LoadInst(PN->getIncomingValue(i),
+ PN->getName() + "." + Pred->getName(),
+ Pred->getTerminator());
+ Load->setAlignment(Align);
+ if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
+ }
+
+ NewPN->addIncoming(Load, Pred);
+ }
+
+ PN->eraseFromParent();
+ }
+
+ ++NumAdjusted;
+ return true;
+}
+
bool SROA::performPromotion(Function &F) {
std::vector<AllocaInst*> Allocas;
DominatorTree *DT = 0;
- DominanceFrontier *DF = 0;
- if (HasDomFrontiers) {
+ if (HasDomTree)
DT = &getAnalysis<DominatorTree>();
- DF = &getAnalysis<DominanceFrontier>();
- }
BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
-
+ DIBuilder DIB(*F.getParent());
bool Changed = false;
SmallVector<Instruction*, 64> Insts;
while (1) {
// the entry node
for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
- if (isAllocaPromotable(AI))
+ if (tryToMakeAllocaBePromotable(AI, TD))
Allocas.push_back(AI);
if (Allocas.empty()) break;
- if (HasDomFrontiers)
- PromoteMemToReg(Allocas, *DT, *DF);
+ if (HasDomTree)
+ PromoteMemToReg(Allocas, *DT);
else {
SSAUpdater SSA;
for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
UI != E; ++UI)
Insts.push_back(cast<Instruction>(*UI));
-
- AllocaPromoter(Insts, SSA).run(AI, Insts);
+ AllocaPromoter(Insts, SSA, &DIB).run(AI, Insts);
Insts.clear();
}
}
/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for
/// SROA. It must be a struct or array type with a small number of elements.
static bool ShouldAttemptScalarRepl(AllocaInst *AI) {
- const Type *T = AI->getAllocatedType();
+ Type *T = AI->getAllocatedType();
// Do not promote any struct into more than 32 separate vars.
- if (const StructType *ST = dyn_cast<StructType>(T))
+ if (StructType *ST = dyn_cast<StructType>(T))
return ST->getNumElements() <= 32;
// Arrays are much less likely to be safe for SROA; only consider
// them if they are very small.
- if (const ArrayType *AT = dyn_cast<ArrayType>(T))
+ if (ArrayType *AT = dyn_cast<ArrayType>(T))
return AT->getNumElements() <= 8;
return false;
}
// performScalarRepl - This algorithm is a simple worklist driven algorithm,
-// which runs on all of the malloc/alloca instructions in the function, removing
-// them if they are only used by getelementptr instructions.
+// which runs on all of the alloca instructions in the function, removing them
+// if they are only used by getelementptr instructions.
//
bool SROA::performScalarRepl(Function &F) {
std::vector<AllocaInst*> WorkList;
// the constant global instead. This is commonly produced by the CFE by
// constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
// is only subsequently read.
- if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) {
+ SmallVector<Instruction *, 4> ToDelete;
+ if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(AI, ToDelete)) {
DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n');
- DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n');
- Constant *TheSrc = cast<Constant>(TheCopy->getSource());
+ DEBUG(dbgs() << " memcpy = " << *Copy << '\n');
+ for (unsigned i = 0, e = ToDelete.size(); i != e; ++i)
+ ToDelete[i]->eraseFromParent();
+ Constant *TheSrc = cast<Constant>(Copy->getSource());
AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType()));
- TheCopy->eraseFromParent(); // Don't mutate the global.
+ Copy->eraseFromParent(); // Don't mutate the global.
AI->eraseFromParent();
++NumGlobals;
Changed = true;
std::vector<AllocaInst*> &WorkList) {
DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n');
SmallVector<AllocaInst*, 32> ElementAllocas;
- if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
+ if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
ElementAllocas.reserve(ST->getNumContainedTypes());
for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
WorkList.push_back(NA); // Add to worklist for recursive processing
}
} else {
- const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
+ ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
ElementAllocas.reserve(AT->getNumElements());
- const Type *ElTy = AT->getElementType();
+ Type *ElTy = AT->getElementType();
for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
AI->getName() + "." + Twine(i), AI);
/// performing scalar replacement of alloca AI. The results are flagged in
/// the Info parameter. Offset indicates the position within AI that is
/// referenced by this instruction.
-void SROA::isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
+void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset,
AllocaInfo &Info) {
for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
Instruction *User = cast<Instruction>(*UI);
if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
- isSafeForScalarRepl(BC, AI, Offset, Info);
+ isSafeForScalarRepl(BC, Offset, Info);
} else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
uint64_t GEPOffset = Offset;
- isSafeGEP(GEPI, AI, GEPOffset, Info);
+ isSafeGEP(GEPI, GEPOffset, Info);
if (!Info.isUnsafe)
- isSafeForScalarRepl(GEPI, AI, GEPOffset, Info);
+ isSafeForScalarRepl(GEPI, GEPOffset, Info);
} else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
- if (Length)
- isSafeMemAccess(AI, Offset, Length->getZExtValue(), 0,
- UI.getOperandNo() == 0, Info);
- else
- MarkUnsafe(Info);
+ if (Length == 0)
+ return MarkUnsafe(Info, User);
+ isSafeMemAccess(Offset, Length->getZExtValue(), 0,
+ UI.getOperandNo() == 0, Info, MI,
+ true /*AllowWholeAccess*/);
} else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
- if (!LI->isVolatile()) {
- const Type *LIType = LI->getType();
- isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(LIType),
- LIType, false, Info);
- Info.hasALoadOrStore = true;
- } else
- MarkUnsafe(Info);
+ if (!LI->isSimple())
+ return MarkUnsafe(Info, User);
+ Type *LIType = LI->getType();
+ isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType),
+ LIType, false, Info, LI, true /*AllowWholeAccess*/);
+ Info.hasALoadOrStore = true;
+
} else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
// Store is ok if storing INTO the pointer, not storing the pointer
- if (!SI->isVolatile() && SI->getOperand(0) != I) {
- const Type *SIType = SI->getOperand(0)->getType();
- isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(SIType),
- SIType, true, Info);
- Info.hasALoadOrStore = true;
- } else
- MarkUnsafe(Info);
+ if (!SI->isSimple() || SI->getOperand(0) == I)
+ return MarkUnsafe(Info, User);
+
+ Type *SIType = SI->getOperand(0)->getType();
+ isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType),
+ SIType, true, Info, SI, true /*AllowWholeAccess*/);
+ Info.hasALoadOrStore = true;
+ } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
+ if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
+ II->getIntrinsicID() != Intrinsic::lifetime_end)
+ return MarkUnsafe(Info, User);
+ } else if (isa<PHINode>(User) || isa<SelectInst>(User)) {
+ isSafePHISelectUseForScalarRepl(User, Offset, Info);
+ } else {
+ return MarkUnsafe(Info, User);
+ }
+ if (Info.isUnsafe) return;
+ }
+}
+
+
+/// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer
+/// derived from the alloca, we can often still split the alloca into elements.
+/// This is useful if we have a large alloca where one element is phi'd
+/// together somewhere: we can SRoA and promote all the other elements even if
+/// we end up not being able to promote this one.
+///
+/// All we require is that the uses of the PHI do not index into other parts of
+/// the alloca. The most important use case for this is single load and stores
+/// that are PHI'd together, which can happen due to code sinking.
+void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset,
+ AllocaInfo &Info) {
+ // If we've already checked this PHI, don't do it again.
+ if (PHINode *PN = dyn_cast<PHINode>(I))
+ if (!Info.CheckedPHIs.insert(PN))
+ return;
+
+ for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
+ Instruction *User = cast<Instruction>(*UI);
+
+ if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
+ isSafePHISelectUseForScalarRepl(BC, Offset, Info);
+ } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
+ // Only allow "bitcast" GEPs for simplicity. We could generalize this,
+ // but would have to prove that we're staying inside of an element being
+ // promoted.
+ if (!GEPI->hasAllZeroIndices())
+ return MarkUnsafe(Info, User);
+ isSafePHISelectUseForScalarRepl(GEPI, Offset, Info);
+ } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
+ if (!LI->isSimple())
+ return MarkUnsafe(Info, User);
+ Type *LIType = LI->getType();
+ isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType),
+ LIType, false, Info, LI, false /*AllowWholeAccess*/);
+ Info.hasALoadOrStore = true;
+
+ } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
+ // Store is ok if storing INTO the pointer, not storing the pointer
+ if (!SI->isSimple() || SI->getOperand(0) == I)
+ return MarkUnsafe(Info, User);
+
+ Type *SIType = SI->getOperand(0)->getType();
+ isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType),
+ SIType, true, Info, SI, false /*AllowWholeAccess*/);
+ Info.hasALoadOrStore = true;
+ } else if (isa<PHINode>(User) || isa<SelectInst>(User)) {
+ isSafePHISelectUseForScalarRepl(User, Offset, Info);
} else {
- DEBUG(errs() << " Transformation preventing inst: " << *User << '\n');
- MarkUnsafe(Info);
+ return MarkUnsafe(Info, User);
}
if (Info.isUnsafe) return;
}
/// references, and when the resulting offset corresponds to an element within
/// the alloca type. The results are flagged in the Info parameter. Upon
/// return, Offset is adjusted as specified by the GEP indices.
-void SROA::isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI,
+void SROA::isSafeGEP(GetElementPtrInst *GEPI,
uint64_t &Offset, AllocaInfo &Info) {
gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI);
if (GEPIt == E)
ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand());
if (!IdxVal)
- return MarkUnsafe(Info);
+ return MarkUnsafe(Info, GEPI);
}
// Compute the offset due to this GEP and check if the alloca has a
// component element at that offset.
SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
- Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(),
- &Indices[0], Indices.size());
- if (!TypeHasComponent(AI->getAllocatedType(), Offset, 0))
- MarkUnsafe(Info);
+ Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices);
+ if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0))
+ MarkUnsafe(Info, GEPI);
}
/// isHomogeneousAggregate - Check if type T is a struct or array containing
/// elements of the same type (which is always true for arrays). If so,
/// return true with NumElts and EltTy set to the number of elements and the
/// element type, respectively.
-static bool isHomogeneousAggregate(const Type *T, unsigned &NumElts,
- const Type *&EltTy) {
- if (const ArrayType *AT = dyn_cast<ArrayType>(T)) {
+static bool isHomogeneousAggregate(Type *T, unsigned &NumElts,
+ Type *&EltTy) {
+ if (ArrayType *AT = dyn_cast<ArrayType>(T)) {
NumElts = AT->getNumElements();
EltTy = (NumElts == 0 ? 0 : AT->getElementType());
return true;
}
- if (const StructType *ST = dyn_cast<StructType>(T)) {
+ if (StructType *ST = dyn_cast<StructType>(T)) {
NumElts = ST->getNumContainedTypes();
EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0));
for (unsigned n = 1; n < NumElts; ++n) {
/// isCompatibleAggregate - Check if T1 and T2 are either the same type or are
/// "homogeneous" aggregates with the same element type and number of elements.
-static bool isCompatibleAggregate(const Type *T1, const Type *T2) {
+static bool isCompatibleAggregate(Type *T1, Type *T2) {
if (T1 == T2)
return true;
unsigned NumElts1, NumElts2;
- const Type *EltTy1, *EltTy2;
+ Type *EltTy1, *EltTy2;
if (isHomogeneousAggregate(T1, NumElts1, EltTy1) &&
isHomogeneousAggregate(T2, NumElts2, EltTy2) &&
NumElts1 == NumElts2 &&
/// alloca or has an offset and size that corresponds to a component element
/// within it. The offset checked here may have been formed from a GEP with a
/// pointer bitcasted to a different type.
-void SROA::isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize,
- const Type *MemOpType, bool isStore,
- AllocaInfo &Info) {
+///
+/// If AllowWholeAccess is true, then this allows uses of the entire alloca as a
+/// unit. If false, it only allows accesses known to be in a single element.
+void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize,
+ Type *MemOpType, bool isStore,
+ AllocaInfo &Info, Instruction *TheAccess,
+ bool AllowWholeAccess) {
// Check if this is a load/store of the entire alloca.
- if (Offset == 0 && MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) {
+ if (Offset == 0 && AllowWholeAccess &&
+ MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) {
// This can be safe for MemIntrinsics (where MemOpType is 0) and integer
// loads/stores (which are essentially the same as the MemIntrinsics with
// regard to copying padding between elements). But, if an alloca is
// This is also safe for references using a type that is compatible with
// the type of the alloca, so that loads/stores can be rewritten using
// insertvalue/extractvalue.
- if (isCompatibleAggregate(MemOpType, AI->getAllocatedType())) {
+ if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) {
Info.hasSubelementAccess = true;
return;
}
}
// Check if the offset/size correspond to a component within the alloca type.
- const Type *T = AI->getAllocatedType();
+ Type *T = Info.AI->getAllocatedType();
if (TypeHasComponent(T, Offset, MemSize)) {
Info.hasSubelementAccess = true;
return;
}
- return MarkUnsafe(Info);
+ return MarkUnsafe(Info, TheAccess);
}
/// TypeHasComponent - Return true if T has a component type with the
/// specified offset and size. If Size is zero, do not check the size.
-bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) {
- const Type *EltTy;
+bool SROA::TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size) {
+ Type *EltTy;
uint64_t EltSize;
- if (const StructType *ST = dyn_cast<StructType>(T)) {
+ if (StructType *ST = dyn_cast<StructType>(T)) {
const StructLayout *Layout = TD->getStructLayout(ST);
unsigned EltIdx = Layout->getElementContainingOffset(Offset);
EltTy = ST->getContainedType(EltIdx);
EltSize = TD->getTypeAllocSize(EltTy);
Offset -= Layout->getElementOffset(EltIdx);
- } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) {
+ } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) {
EltTy = AT->getElementType();
EltSize = TD->getTypeAllocSize(EltTy);
if (Offset >= AT->getNumElements() * EltSize)
/// instruction.
void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
SmallVector<AllocaInst*, 32> &NewElts) {
- for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
- Instruction *User = cast<Instruction>(*UI);
+ for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) {
+ Use &TheUse = UI.getUse();
+ Instruction *User = cast<Instruction>(*UI++);
if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
RewriteBitCast(BC, AI, Offset, NewElts);
- } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
+ continue;
+ }
+
+ if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
RewriteGEP(GEPI, AI, Offset, NewElts);
- } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
+ continue;
+ }
+
+ if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
uint64_t MemSize = Length->getZExtValue();
if (Offset == 0 &&
RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts);
// Otherwise the intrinsic can only touch a single element and the
// address operand will be updated, so nothing else needs to be done.
- } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
- const Type *LIType = LI->getType();
+ continue;
+ }
+
+ if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
+ if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
+ II->getIntrinsicID() == Intrinsic::lifetime_end) {
+ RewriteLifetimeIntrinsic(II, AI, Offset, NewElts);
+ }
+ continue;
+ }
+
+ if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
+ Type *LIType = LI->getType();
if (isCompatibleAggregate(LIType, AI->getAllocatedType())) {
// Replace:
// %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1
// (Also works for arrays instead of structs)
Value *Insert = UndefValue::get(LIType);
+ IRBuilder<> Builder(LI);
for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
- Value *Load = new LoadInst(NewElts[i], "load", LI);
- Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI);
+ Value *Load = Builder.CreateLoad(NewElts[i], "load");
+ Insert = Builder.CreateInsertValue(Insert, Load, i, "insert");
}
LI->replaceAllUsesWith(Insert);
DeadInsts.push_back(LI);
// If this is a load of the entire alloca to an integer, rewrite it.
RewriteLoadUserOfWholeAlloca(LI, AI, NewElts);
}
- } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
+ continue;
+ }
+
+ if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
Value *Val = SI->getOperand(0);
- const Type *SIType = Val->getType();
+ Type *SIType = Val->getType();
if (isCompatibleAggregate(SIType, AI->getAllocatedType())) {
// Replace:
// store { i32, i32 } %val, { i32, i32 }* %alloc
// %val.1 = extractvalue { i32, i32 } %val, 1
// store i32 %val.1, i32* %alloc.1
// (Also works for arrays instead of structs)
+ IRBuilder<> Builder(SI);
for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
- Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI);
- new StoreInst(Extract, NewElts[i], SI);
+ Value *Extract = Builder.CreateExtractValue(Val, i, Val->getName());
+ Builder.CreateStore(Extract, NewElts[i]);
}
DeadInsts.push_back(SI);
} else if (SIType->isIntegerTy() &&
// If this is a store of the entire alloca from an integer, rewrite it.
RewriteStoreUserOfWholeAlloca(SI, AI, NewElts);
}
+ continue;
+ }
+
+ if (isa<SelectInst>(User) || isa<PHINode>(User)) {
+ // If we have a PHI user of the alloca itself (as opposed to a GEP or
+ // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to
+ // the new pointer.
+ if (!isa<AllocaInst>(I)) continue;
+
+ assert(Offset == 0 && NewElts[0] &&
+ "Direct alloca use should have a zero offset");
+
+ // If we have a use of the alloca, we know the derived uses will be
+ // utilizing just the first element of the scalarized result. Insert a
+ // bitcast of the first alloca before the user as required.
+ AllocaInst *NewAI = NewElts[0];
+ BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI);
+ NewAI->moveBefore(BCI);
+ TheUse = BCI;
+ continue;
}
}
}
return;
// The bitcast references the original alloca. Replace its uses with
- // references to the first new element alloca.
- Instruction *Val = NewElts[0];
+ // references to the alloca containing offset zero (which is normally at
+ // index zero, but might not be in cases involving structs with elements
+ // of size zero).
+ Type *T = AI->getAllocatedType();
+ uint64_t EltOffset = 0;
+ Type *IdxTy;
+ uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy);
+ Instruction *Val = NewElts[Idx];
if (Val->getType() != BC->getDestTy()) {
Val = new BitCastInst(Val, BC->getDestTy(), "", BC);
Val->takeName(BC);
/// Sets T to the type of the element and Offset to the offset within that
/// element. IdxTy is set to the type of the index result to be used in a
/// GEP instruction.
-uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset,
- const Type *&IdxTy) {
+uint64_t SROA::FindElementAndOffset(Type *&T, uint64_t &Offset,
+ Type *&IdxTy) {
uint64_t Idx = 0;
- if (const StructType *ST = dyn_cast<StructType>(T)) {
+ if (StructType *ST = dyn_cast<StructType>(T)) {
const StructLayout *Layout = TD->getStructLayout(ST);
Idx = Layout->getElementContainingOffset(Offset);
T = ST->getContainedType(Idx);
IdxTy = Type::getInt32Ty(T->getContext());
return Idx;
}
- const ArrayType *AT = cast<ArrayType>(T);
+ ArrayType *AT = cast<ArrayType>(T);
T = AT->getElementType();
uint64_t EltSize = TD->getTypeAllocSize(T);
Idx = Offset / EltSize;
SmallVector<AllocaInst*, 32> &NewElts) {
uint64_t OldOffset = Offset;
SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
- Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(),
- &Indices[0], Indices.size());
+ Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices);
RewriteForScalarRepl(GEPI, AI, Offset, NewElts);
- const Type *T = AI->getAllocatedType();
- const Type *IdxTy;
+ Type *T = AI->getAllocatedType();
+ Type *IdxTy;
uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy);
if (GEPI->getOperand(0) == AI)
OldIdx = ~0ULL; // Force the GEP to be rewritten.
if (Idx == OldIdx)
return;
- const Type *i32Ty = Type::getInt32Ty(AI->getContext());
+ Type *i32Ty = Type::getInt32Ty(AI->getContext());
SmallVector<Value*, 8> NewArgs;
NewArgs.push_back(Constant::getNullValue(i32Ty));
while (EltOffset != 0) {
}
Instruction *Val = NewElts[Idx];
if (NewArgs.size() > 1) {
- Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(),
- NewArgs.end(), "", GEPI);
+ Val = GetElementPtrInst::CreateInBounds(Val, NewArgs, "", GEPI);
Val->takeName(GEPI);
}
if (Val->getType() != GEPI->getType())
DeadInsts.push_back(GEPI);
}
+/// RewriteLifetimeIntrinsic - II is a lifetime.start/lifetime.end. Rewrite it
+/// to mark the lifetime of the scalarized memory.
+void SROA::RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI,
+ uint64_t Offset,
+ SmallVector<AllocaInst*, 32> &NewElts) {
+ ConstantInt *OldSize = cast<ConstantInt>(II->getArgOperand(0));
+ // Put matching lifetime markers on everything from Offset up to
+ // Offset+OldSize.
+ Type *AIType = AI->getAllocatedType();
+ uint64_t NewOffset = Offset;
+ Type *IdxTy;
+ uint64_t Idx = FindElementAndOffset(AIType, NewOffset, IdxTy);
+
+ IRBuilder<> Builder(II);
+ uint64_t Size = OldSize->getLimitedValue();
+
+ if (NewOffset) {
+ // Splice the first element and index 'NewOffset' bytes in. SROA will
+ // split the alloca again later.
+ Value *V = Builder.CreateBitCast(NewElts[Idx], Builder.getInt8PtrTy());
+ V = Builder.CreateGEP(V, Builder.getInt64(NewOffset));
+
+ IdxTy = NewElts[Idx]->getAllocatedType();
+ uint64_t EltSize = TD->getTypeAllocSize(IdxTy) - NewOffset;
+ if (EltSize > Size) {
+ EltSize = Size;
+ Size = 0;
+ } else {
+ Size -= EltSize;
+ }
+ if (II->getIntrinsicID() == Intrinsic::lifetime_start)
+ Builder.CreateLifetimeStart(V, Builder.getInt64(EltSize));
+ else
+ Builder.CreateLifetimeEnd(V, Builder.getInt64(EltSize));
+ ++Idx;
+ }
+
+ for (; Idx != NewElts.size() && Size; ++Idx) {
+ IdxTy = NewElts[Idx]->getAllocatedType();
+ uint64_t EltSize = TD->getTypeAllocSize(IdxTy);
+ if (EltSize > Size) {
+ EltSize = Size;
+ Size = 0;
+ } else {
+ Size -= EltSize;
+ }
+ if (II->getIntrinsicID() == Intrinsic::lifetime_start)
+ Builder.CreateLifetimeStart(NewElts[Idx],
+ Builder.getInt64(EltSize));
+ else
+ Builder.CreateLifetimeEnd(NewElts[Idx],
+ Builder.getInt64(EltSize));
+ }
+ DeadInsts.push_back(II);
+}
+
/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI.
/// Rewrite it to copy or set the elements of the scalarized memory.
void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
// If the pointer is not the right type, insert a bitcast to the right
// type.
- const Type *NewTy =
+ Type *NewTy =
PointerType::get(AI->getType()->getElementType(), AddrSpace);
if (OtherPtr->getType() != NewTy)
if (OtherPtr) {
Value *Idx[2] = { Zero,
ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) };
- OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2,
+ OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx,
OtherPtr->getName()+"."+Twine(i),
MI);
uint64_t EltOffset;
- const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType());
- const Type *OtherTy = OtherPtrTy->getElementType();
- if (const StructType *ST = dyn_cast<StructType>(OtherTy)) {
+ PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType());
+ Type *OtherTy = OtherPtrTy->getElementType();
+ if (StructType *ST = dyn_cast<StructType>(OtherTy)) {
EltOffset = TD->getStructLayout(ST)->getElementOffset(i);
} else {
- const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType();
+ Type *EltTy = cast<SequentialType>(OtherTy)->getElementType();
EltOffset = TD->getTypeAllocSize(EltTy)*i;
}
}
Value *EltPtr = NewElts[i];
- const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType();
+ Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType();
// If we got down to a scalar, insert a load or store as appropriate.
if (EltTy->isSingleValueType()) {
StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0>
} else {
// If EltTy is a vector type, get the element type.
- const Type *ValTy = EltTy->getScalarType();
+ Type *ValTy = EltTy->getScalarType();
// Construct an integer with the right value.
unsigned EltSize = TD->getTypeSizeInBits(ValTy);
assert(StoreVal->getType() == ValTy && "Type mismatch!");
// If the requested value was a vector constant, create it.
- if (EltTy != ValTy) {
- unsigned NumElts = cast<VectorType>(ValTy)->getNumElements();
+ if (EltTy->isVectorTy()) {
+ unsigned NumElts = cast<VectorType>(EltTy)->getNumElements();
SmallVector<Constant*, 16> Elts(NumElts, StoreVal);
- StoreVal = ConstantVector::get(&Elts[0], NumElts);
+ StoreVal = ConstantVector::get(Elts);
}
}
new StoreInst(StoreVal, EltPtr, MI);
}
unsigned EltSize = TD->getTypeAllocSize(EltTy);
+ if (!EltSize)
+ continue;
IRBuilder<> Builder(MI);
// Extract each element out of the integer according to its structure offset
// and store the element value to the individual alloca.
Value *SrcVal = SI->getOperand(0);
- const Type *AllocaEltTy = AI->getAllocatedType();
+ Type *AllocaEltTy = AI->getAllocatedType();
uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
IRBuilder<> Builder(SI);
// There are two forms here: AI could be an array or struct. Both cases
// have different ways to compute the element offset.
- if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
+ if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
const StructLayout *Layout = TD->getStructLayout(EltSTy);
for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
// Get the number of bits to shift SrcVal to get the value.
- const Type *FieldTy = EltSTy->getElementType(i);
+ Type *FieldTy = EltSTy->getElementType(i);
uint64_t Shift = Layout->getElementOffsetInBits(i);
if (TD->isBigEndian())
}
} else {
- const ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
- const Type *ArrayEltTy = ATy->getElementType();
+ ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
+ Type *ArrayEltTy = ATy->getElementType();
uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy);
SmallVector<AllocaInst*, 32> &NewElts) {
// Extract each element out of the NewElts according to its structure offset
// and form the result value.
- const Type *AllocaEltTy = AI->getAllocatedType();
+ Type *AllocaEltTy = AI->getAllocatedType();
uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI
// have different ways to compute the element offset.
const StructLayout *Layout = 0;
uint64_t ArrayEltBitOffset = 0;
- if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
+ if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
Layout = TD->getStructLayout(EltSTy);
} else {
- const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
+ Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
}
// Load the value from the alloca. If the NewElt is an aggregate, cast
// the pointer to an integer of the same size before doing the load.
Value *SrcField = NewElts[i];
- const Type *FieldTy =
+ Type *FieldTy =
cast<PointerType>(SrcField->getType())->getElementType();
uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
// Ignore zero sized fields like {}, they obviously contain no data.
if (FieldSizeBits == 0) continue;
- const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(),
+ IntegerType *FieldIntTy = IntegerType::get(LI->getContext(),
FieldSizeBits);
if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() &&
!FieldTy->isVectorTy())
/// HasPadding - Return true if the specified type has any structure or
/// alignment padding in between the elements that would be split apart
/// by SROA; return false otherwise.
-static bool HasPadding(const Type *Ty, const TargetData &TD) {
- if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
+static bool HasPadding(Type *Ty, const TargetData &TD) {
+ if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
Ty = ATy->getElementType();
return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty);
}
// SROA currently handles only Arrays and Structs.
- const StructType *STy = cast<StructType>(Ty);
+ StructType *STy = cast<StructType>(Ty);
const StructLayout *SL = TD.getStructLayout(STy);
unsigned PrevFieldBitOffset = 0;
for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) {
// Loop over the use list of the alloca. We can only transform it if all of
// the users are safe to transform.
- AllocaInfo Info;
+ AllocaInfo Info(AI);
- isSafeForScalarRepl(AI, AI, 0, Info);
+ isSafeForScalarRepl(AI, 0, Info);
if (Info.isUnsafe) {
DEBUG(dbgs() << "Cannot transform: " << *AI << '\n');
return false;
// and fusion code.
if (!Info.hasSubelementAccess && Info.hasALoadOrStore) {
// If the struct/array just has one element, use basic SRoA.
- if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
+ if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
if (ST->getNumElements() > 1) return false;
} else {
if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1)
return false;
}
}
+
return true;
}
/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
/// the alloca, and if the source pointer is a pointer to a constant global, we
/// can optimize this.
-static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy,
- bool isOffset) {
+static bool
+isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy,
+ bool isOffset,
+ SmallVector<Instruction *, 4> &LifetimeMarkers) {
+ // We track lifetime intrinsics as we encounter them. If we decide to go
+ // ahead and replace the value with the global, this lets the caller quickly
+ // eliminate the markers.
+
for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
User *U = cast<Instruction>(*UI);
if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
// Ignore non-volatile loads, they are always ok.
- if (LI->isVolatile()) return false;
+ if (!LI->isSimple()) return false;
continue;
}
if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
// If uses of the bitcast are ok, we are ok.
- if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset))
+ if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset,
+ LifetimeMarkers))
return false;
continue;
}
// If the GEP has all zero indices, it doesn't offset the pointer. If it
// doesn't, it does.
if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy,
- isOffset || !GEP->hasAllZeroIndices()))
+ isOffset || !GEP->hasAllZeroIndices(),
+ LifetimeMarkers))
return false;
continue;
}
if (CallSite CS = U) {
- // If this is a readonly/readnone call site, then we know it is just a
- // load and we can ignore it.
- if (CS.onlyReadsMemory())
- continue;
-
// If this is the function being called then we treat it like a load and
// ignore it.
if (CS.isCallee(UI))
continue;
+ // If this is a readonly/readnone call site, then we know it is just a
+ // load (but one that potentially returns the value itself), so we can
+ // ignore it if we know that the value isn't captured.
+ unsigned ArgNo = CS.getArgumentNo(UI);
+ if (CS.onlyReadsMemory() &&
+ (CS.getInstruction()->use_empty() || CS.doesNotCapture(ArgNo)))
+ continue;
+
// If this is being passed as a byval argument, the caller is making a
// copy, so it is only a read of the alloca.
- unsigned ArgNo = CS.getArgumentNo(UI);
- if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal))
+ if (CS.isByValArgument(ArgNo))
continue;
}
+ // Lifetime intrinsics can be handled by the caller.
+ if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
+ if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
+ II->getIntrinsicID() == Intrinsic::lifetime_end) {
+ assert(II->use_empty() && "Lifetime markers have no result to use!");
+ LifetimeMarkers.push_back(II);
+ continue;
+ }
+ }
+
// If this is isn't our memcpy/memmove, reject it as something we can't
// handle.
MemTransferInst *MI = dyn_cast<MemTransferInst>(U);
/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
/// modified by a copy from a constant global. If we can prove this, we can
/// replace any uses of the alloca with uses of the global directly.
-MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) {
+MemTransferInst *
+SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI,
+ SmallVector<Instruction*, 4> &ToDelete) {
MemTransferInst *TheCopy = 0;
- if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false))
+ if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false, ToDelete))
return TheCopy;
return 0;
}