// each member (if possible). Then, if possible, it transforms the individual
// alloca instructions into nice clean scalar SSA form.
//
-// This combines a simple SRoA algorithm with the Mem2Reg algorithm because
+// This combines a simple SRoA algorithm with the Mem2Reg algorithm because they
// often interact, especially for C++ programs. As such, iterating between
// SRoA, then Mem2Reg until we run out of things to promote works well.
//
#define DEBUG_TYPE "scalarrepl"
#include "llvm/Transforms/Scalar.h"
-#include "llvm/Constants.h"
-#include "llvm/DerivedTypes.h"
-#include "llvm/Function.h"
-#include "llvm/GlobalVariable.h"
-#include "llvm/Instructions.h"
-#include "llvm/IntrinsicInst.h"
-#include "llvm/LLVMContext.h"
-#include "llvm/Module.h"
-#include "llvm/Pass.h"
-#include "llvm/Analysis/DebugInfo.h"
-#include "llvm/Analysis/DIBuilder.h"
-#include "llvm/Analysis/Dominators.h"
+#include "llvm/ADT/SetVector.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/Loads.h"
#include "llvm/Analysis/ValueTracking.h"
-#include "llvm/Target/TargetData.h"
-#include "llvm/Transforms/Utils/PromoteMemToReg.h"
-#include "llvm/Transforms/Utils/Local.h"
-#include "llvm/Transforms/Utils/SSAUpdater.h"
+#include "llvm/DIBuilder.h"
+#include "llvm/DebugInfo.h"
+#include "llvm/IR/Constants.h"
+#include "llvm/IR/DataLayout.h"
+#include "llvm/IR/DerivedTypes.h"
+#include "llvm/IR/Dominators.h"
+#include "llvm/IR/Function.h"
+#include "llvm/IR/GlobalVariable.h"
+#include "llvm/IR/IRBuilder.h"
+#include "llvm/IR/Instructions.h"
+#include "llvm/IR/IntrinsicInst.h"
+#include "llvm/IR/LLVMContext.h"
+#include "llvm/IR/Module.h"
+#include "llvm/IR/Operator.h"
+#include "llvm/Pass.h"
#include "llvm/Support/CallSite.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/GetElementPtrTypeIterator.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"
+#include "llvm/Transforms/Utils/Local.h"
+#include "llvm/Transforms/Utils/PromoteMemToReg.h"
+#include "llvm/Transforms/Utils/SSAUpdater.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 hasDT, char &ID)
+ SROA(int T, bool hasDT, char &ID, int ST, int AT, int SLT)
: FunctionPass(ID), HasDomTree(hasDT) {
if (T == -1)
SRThreshold = 128;
else
SRThreshold = T;
+ if (ST == -1)
+ StructMemberThreshold = 32;
+ else
+ StructMemberThreshold = ST;
+ if (AT == -1)
+ ArrayElementThreshold = 8;
+ else
+ ArrayElementThreshold = AT;
+ if (SLT == -1)
+ // Do not limit the scalar integer load size if no threshold is given.
+ ScalarLoadThreshold = -1;
+ else
+ ScalarLoadThreshold = SLT;
}
bool runOnFunction(Function &F);
private:
bool HasDomTree;
- TargetData *TD;
+ const DataLayout *DL;
/// DeadInsts - Keep track of instructions we have made dead, so that
/// we can remove them after we are done working.
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;
/// ever accessed, or false if the alloca is only accessed with mem
/// intrinsics or load/store that only access the entire alloca at once.
bool hasSubelementAccess : 1;
-
+
/// hasALoadOrStore - This is true if there are any loads or stores to it.
/// The alloca may just be accessed with memcpy, for example, which would
/// not set this.
bool hasALoadOrStore : 1;
-
+
explicit AllocaInfo(AllocaInst *ai)
: AI(ai), isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false),
hasSubelementAccess(false), hasALoadOrStore(false) {}
};
+ /// SRThreshold - The maximum alloca size to considered for SROA.
unsigned SRThreshold;
+ /// StructMemberThreshold - The maximum number of members a struct can
+ /// contain to be considered for SROA.
+ unsigned StructMemberThreshold;
+
+ /// ArrayElementThreshold - The maximum number of elements an array can
+ /// have to be considered for SROA.
+ unsigned ArrayElementThreshold;
+
+ /// ScalarLoadThreshold - The maximum size in bits of scalars to load when
+ /// converting to scalar
+ unsigned ScalarLoadThreshold;
+
void MarkUnsafe(AllocaInfo &I, Instruction *User) {
I.isUnsafe = true;
DEBUG(dbgs() << " Transformation preventing inst: " << *User << '\n');
void DeleteDeadInstructions();
void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
- SmallVector<AllocaInst*, 32> &NewElts);
+ SmallVectorImpl<AllocaInst *> &NewElts);
void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
- SmallVector<AllocaInst*, 32> &NewElts);
+ SmallVectorImpl<AllocaInst *> &NewElts);
void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
- SmallVector<AllocaInst*, 32> &NewElts);
+ SmallVectorImpl<AllocaInst *> &NewElts);
void RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI,
uint64_t Offset,
- SmallVector<AllocaInst*, 32> &NewElts);
+ SmallVectorImpl<AllocaInst *> &NewElts);
void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
AllocaInst *AI,
- SmallVector<AllocaInst*, 32> &NewElts);
+ SmallVectorImpl<AllocaInst *> &NewElts);
void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
- SmallVector<AllocaInst*, 32> &NewElts);
+ SmallVectorImpl<AllocaInst *> &NewElts);
void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
- SmallVector<AllocaInst*, 32> &NewElts);
-
- static MemTransferInst *isOnlyCopiedFromConstantGlobal(
- AllocaInst *AI, SmallVector<Instruction*, 4> &ToDelete);
+ SmallVectorImpl<AllocaInst *> &NewElts);
+ bool ShouldAttemptScalarRepl(AllocaInst *AI);
};
-
+
// SROA_DT - SROA that uses DominatorTree.
struct SROA_DT : public SROA {
static char ID;
public:
- SROA_DT(int T = -1) : SROA(T, true, ID) {
+ SROA_DT(int T = -1, int ST = -1, int AT = -1, int SLT = -1) :
+ SROA(T, true, ID, ST, AT, SLT) {
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<DominatorTreeWrapperPass>();
AU.setPreservesCFG();
}
};
-
+
// SROA_SSAUp - SROA that uses SSAUpdater.
struct SROA_SSAUp : public SROA {
static char ID;
public:
- SROA_SSAUp(int T = -1) : SROA(T, false, ID) {
+ SROA_SSAUp(int T = -1, int ST = -1, int AT = -1, int SLT = -1) :
+ SROA(T, false, ID, ST, AT, SLT) {
initializeSROA_SSAUpPass(*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.setPreservesCFG();
}
};
-
+
}
char SROA_DT::ID = 0;
INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl",
"Scalar Replacement of Aggregates (DT)", false, false)
-INITIALIZE_PASS_DEPENDENCY(DominatorTree)
+INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
INITIALIZE_PASS_END(SROA_DT, "scalarrepl",
"Scalar Replacement of Aggregates (DT)", false, false)
// Public interface to the ScalarReplAggregates pass
FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold,
- bool UseDomTree) {
+ bool UseDomTree,
+ int StructMemberThreshold,
+ int ArrayElementThreshold,
+ int ScalarLoadThreshold) {
if (UseDomTree)
- return new SROA_DT(Threshold);
- return new SROA_SSAUp(Threshold);
+ return new SROA_DT(Threshold, StructMemberThreshold, ArrayElementThreshold,
+ ScalarLoadThreshold);
+ return new SROA_SSAUp(Threshold, StructMemberThreshold,
+ ArrayElementThreshold, ScalarLoadThreshold);
}
class ConvertToScalarInfo {
/// AllocaSize - The size of the alloca being considered in bytes.
unsigned AllocaSize;
- const TargetData &TD;
+ const DataLayout &DL;
+ unsigned ScalarLoadThreshold;
/// IsNotTrivial - This is set to true if there is some access to the object
/// which means that mem2reg can't promote it.
/// isn't possible to turn into a vector type, it gets set to VoidTy.
VectorType *VectorTy;
- /// HadNonMemTransferAccess - True if there is at least one access to the
+ /// 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;
+ /// HadDynamicAccess - True if some element of this alloca was dynamic.
+ /// We don't yet have support for turning a dynamic access into a large
+ /// integer.
+ bool HadDynamicAccess;
+
public:
- explicit ConvertToScalarInfo(unsigned Size, const TargetData &td)
- : AllocaSize(Size), TD(td), IsNotTrivial(false), ScalarKind(Unknown),
- VectorTy(0), HadNonMemTransferAccess(false) { }
+ explicit ConvertToScalarInfo(unsigned Size, const DataLayout &DL,
+ unsigned SLT)
+ : AllocaSize(Size), DL(DL), ScalarLoadThreshold(SLT), IsNotTrivial(false),
+ ScalarKind(Unknown), VectorTy(0), HadNonMemTransferAccess(false),
+ HadDynamicAccess(false) { }
AllocaInst *TryConvert(AllocaInst *AI);
private:
- bool CanConvertToScalar(Value *V, uint64_t Offset);
+ bool CanConvertToScalar(Value *V, uint64_t Offset, Value* NonConstantIdx);
void MergeInTypeForLoadOrStore(Type *In, uint64_t Offset);
bool MergeInVectorType(VectorType *VInTy, uint64_t Offset);
- void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset);
+ void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset,
+ Value *NonConstantIdx);
Value *ConvertScalar_ExtractValue(Value *NV, Type *ToType,
- uint64_t Offset, IRBuilder<> &Builder);
+ uint64_t Offset, Value* NonConstantIdx,
+ IRBuilder<> &Builder);
Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal,
- uint64_t Offset, IRBuilder<> &Builder);
+ uint64_t Offset, Value* NonConstantIdx,
+ IRBuilder<> &Builder);
};
} // end anonymous namespace.
AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) {
// If we can't convert this scalar, or if mem2reg can trivially do it, bail
// out.
- if (!CanConvertToScalar(AI, 0) || !IsNotTrivial)
+ if (!CanConvertToScalar(AI, 0, 0) || !IsNotTrivial)
return 0;
// If an alloca has only memset / memcpy uses, it may still have an Unknown
if (ScalarKind == Unknown)
ScalarKind = Integer;
- // FIXME: It should be possible to promote the vector type up to the alloca's
- // size.
if (ScalarKind == Vector && VectorTy->getBitWidth() != AllocaSize * 8)
ScalarKind = Integer;
NewTy = VectorTy; // Use the vector type.
} else {
unsigned BitWidth = AllocaSize * 8;
+
+ // Do not convert to scalar integer if the alloca size exceeds the
+ // scalar load threshold.
+ if (BitWidth > ScalarLoadThreshold)
+ return 0;
+
if ((ScalarKind == ImplicitVector || ScalarKind == Integer) &&
- !HadNonMemTransferAccess && !TD.fitsInLegalInteger(BitWidth))
+ !HadNonMemTransferAccess && !DL.fitsInLegalInteger(BitWidth))
+ return 0;
+ // Dynamic accesses on integers aren't yet supported. They need us to shift
+ // by a dynamic amount which could be difficult to work out as we might not
+ // know whether to use a left or right shift.
+ if (ScalarKind == Integer && HadDynamicAccess)
return 0;
DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n");
NewTy = IntegerType::get(AI->getContext(), BitWidth);
}
AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin());
- ConvertUsesToScalar(AI, NewAI, 0);
+ ConvertUsesToScalar(AI, NewAI, 0, 0);
return NewAI;
}
/// (VectorTy) so far at the offset specified by Offset (which is specified in
/// bytes).
///
-/// There are three cases we handle here:
+/// There are two cases we handle here:
/// 1) A union of vector types of the same size and potentially its elements.
/// Here we turn element accesses into insert/extract element operations.
/// This promotes a <4 x float> with a store of float to the third element
/// into a <4 x float> that uses insert element.
-/// 2) A union of vector types with power-of-2 size differences, e.g. a float,
-/// <2 x float> and <4 x float>. Here we turn element accesses into insert
-/// and extract element operations, and <2 x float> accesses into a cast to
-/// <2 x double>, an extract, and a cast back to <2 x float>.
-/// 3) A fully general blob of memory, which we turn into some (potentially
+/// 2) A fully general blob of memory, which we turn into some (potentially
/// 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.
// if the implied vector agrees with what we already have and if Offset is
// compatible with it.
if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 &&
- (!VectorTy || Offset * 8 < VectorTy->getPrimitiveSizeInBits())) {
+ (!VectorTy || EltSize == VectorTy->getElementType()
+ ->getPrimitiveSizeInBits()/8)) {
if (!VectorTy) {
ScalarKind = ImplicitVector;
VectorTy = VectorType::get(In, AllocaSize/EltSize);
- return;
}
-
- unsigned CurrentEltSize = VectorTy->getElementType()
- ->getPrimitiveSizeInBits()/8;
- if (EltSize == CurrentEltSize)
- return;
-
- if (In->isIntegerTy() && isPowerOf2_32(AllocaSize / EltSize))
- return;
+ return;
}
}
/// returning true if the type was successfully merged and false otherwise.
bool ConvertToScalarInfo::MergeInVectorType(VectorType *VInTy,
uint64_t Offset) {
- // TODO: Support nonzero offsets?
- if (Offset != 0)
- return false;
-
- // Only allow vectors that are a power-of-2 away from the size of the alloca.
- if (!isPowerOf2_64(AllocaSize / (VInTy->getBitWidth() / 8)))
- return false;
-
- // If this the first vector we see, remember the type so that we know the
- // element size.
- if (!VectorTy) {
- ScalarKind = Vector;
- VectorTy = VInTy;
- return true;
- }
-
- unsigned BitWidth = VectorTy->getBitWidth();
- unsigned InBitWidth = VInTy->getBitWidth();
-
- // Vectors of the same size can be converted using a simple bitcast.
- if (InBitWidth == BitWidth && AllocaSize == (InBitWidth / 8)) {
+ 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;
}
- Type *ElementTy = VectorTy->getElementType();
- Type *InElementTy = VInTy->getElementType();
-
- // If they're the same alloc size, we'll be attempting to convert between
- // them with a vector shuffle, which requires the element types to match.
- if (TD.getTypeAllocSize(VectorTy) == TD.getTypeAllocSize(VInTy) &&
- ElementTy != InElementTy)
- return false;
-
- // Do not allow mixed integer and floating-point accesses from vectors of
- // different sizes.
- if (ElementTy->isFloatingPointTy() != InElementTy->isFloatingPointTy())
- return false;
-
- if (ElementTy->isFloatingPointTy()) {
- // Only allow floating-point vectors of different sizes if they have the
- // same element type.
- // TODO: This could be loosened a bit, but would anything benefit?
- if (ElementTy != InElementTy)
- return false;
-
- // There are no arbitrary-precision floating-point types, which limits the
- // number of legal vector types with larger element types that we can form
- // to bitcast and extract a subvector.
- // TODO: We could support some more cases with mixed fp128 and double here.
- if (!(BitWidth == 64 || BitWidth == 128) ||
- !(InBitWidth == 64 || InBitWidth == 128))
- return false;
- } else {
- assert(ElementTy->isIntegerTy() && "Vector elements must be either integer "
- "or floating-point.");
- unsigned BitWidth = ElementTy->getPrimitiveSizeInBits();
- unsigned InBitWidth = InElementTy->getPrimitiveSizeInBits();
-
- // Do not allow integer types smaller than a byte or types whose widths are
- // not a multiple of a byte.
- if (BitWidth < 8 || InBitWidth < 8 ||
- BitWidth % 8 != 0 || InBitWidth % 8 != 0)
- return false;
- }
-
- // Pick the largest of the two vector types.
- ScalarKind = Vector;
- if (InBitWidth > BitWidth)
- VectorTy = VInTy;
-
- return true;
+ return false;
}
/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all
///
/// If we see at least one access to the value that is as a vector type, set the
/// SawVec flag.
-bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) {
+bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset,
+ Value* NonConstantIdx) {
for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
Instruction *User = cast<Instruction>(*UI);
if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
if (!onlyUsedByLifetimeMarkers(BCI))
IsNotTrivial = true; // Can't be mem2reg'd.
- if (!CanConvertToScalar(BCI, Offset))
+ if (!CanConvertToScalar(BCI, Offset, NonConstantIdx))
return false;
continue;
}
if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
// If this is a GEP with a variable indices, we can't handle it.
- if (!GEP->hasAllConstantIndices())
+ PointerType* PtrTy = dyn_cast<PointerType>(GEP->getPointerOperandType());
+ if (!PtrTy)
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(),
+ Value *GEPNonConstantIdx = 0;
+ if (!GEP->hasAllConstantIndices()) {
+ if (!isa<VectorType>(PtrTy->getElementType()))
+ return false;
+ if (NonConstantIdx)
+ return false;
+ GEPNonConstantIdx = Indices.pop_back_val();
+ if (!GEPNonConstantIdx->getType()->isIntegerTy(32))
+ return false;
+ HadDynamicAccess = true;
+ } else
+ GEPNonConstantIdx = NonConstantIdx;
+ uint64_t GEPOffset = DL.getIndexedOffset(PtrTy,
Indices);
// See if all uses can be converted.
- if (!CanConvertToScalar(GEP, Offset+GEPOffset))
+ if (!CanConvertToScalar(GEP, Offset+GEPOffset, GEPNonConstantIdx))
return false;
IsNotTrivial = true; // Can't be mem2reg'd.
HadNonMemTransferAccess = true;
// 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 to dynamic index.
+ if (NonConstantIdx)
+ return false;
// Store of constant value.
if (!isa<ConstantInt>(MSI->getValue()))
return false;
// If this is a memcpy or memmove into or out of the whole allocation, we
// can handle it like a load or store of the scalar type.
if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
+ // Store to dynamic index.
+ if (NonConstantIdx)
+ return false;
ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength());
if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0)
return false;
/// Offset is an offset from the original alloca, in bits that need to be
/// shifted to the right. By the end of this, there should be no uses of Ptr.
void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI,
- uint64_t Offset) {
+ uint64_t Offset,
+ Value* NonConstantIdx) {
while (!Ptr->use_empty()) {
Instruction *User = cast<Instruction>(Ptr->use_back());
if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
- ConvertUsesToScalar(CI, NewAI, Offset);
+ ConvertUsesToScalar(CI, NewAI, Offset, NonConstantIdx);
CI->eraseFromParent();
continue;
}
if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
// 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(),
+ Value* GEPNonConstantIdx = 0;
+ if (!GEP->hasAllConstantIndices()) {
+ assert(!NonConstantIdx &&
+ "Dynamic GEP reading from dynamic GEP unsupported");
+ GEPNonConstantIdx = Indices.pop_back_val();
+ } else
+ GEPNonConstantIdx = NonConstantIdx;
+ uint64_t GEPOffset = DL.getIndexedOffset(GEP->getPointerOperandType(),
Indices);
- ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8);
+ ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8, GEPNonConstantIdx);
GEP->eraseFromParent();
continue;
}
// The load is a bit extract from NewAI shifted right by Offset bits.
Value *LoadedVal = Builder.CreateLoad(NewAI);
Value *NewLoadVal
- = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder);
+ = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset,
+ NonConstantIdx, Builder);
LI->replaceAllUsesWith(NewLoadVal);
LI->eraseFromParent();
continue;
assert(SI->getOperand(0) != Ptr && "Consistency error!");
Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset,
- Builder);
+ NonConstantIdx, Builder);
Builder.CreateStore(New, NewAI);
SI->eraseFromParent();
// transform it into a store of the expanded constant value.
if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
assert(MSI->getRawDest() == Ptr && "Consistency error!");
- unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
- if (NumBytes != 0) {
+ assert(!NonConstantIdx && "Cannot replace dynamic memset with insert");
+ int64_t SNumBytes = cast<ConstantInt>(MSI->getLength())->getSExtValue();
+ if (SNumBytes > 0 && (SNumBytes >> 32) == 0) {
+ unsigned NumBytes = static_cast<unsigned>(SNumBytes);
unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue();
// Compute the value replicated the right number of times.
Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
Value *New = ConvertScalar_InsertValue(
ConstantInt::get(User->getContext(), APVal),
- Old, Offset, Builder);
+ Old, Offset, 0, Builder);
Builder.CreateStore(New, NewAI);
// If the load we just inserted is now dead, then the memset overwrote
// can handle it like a load or store of the scalar type.
if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
assert(Offset == 0 && "must be store to start of alloca");
+ assert(!NonConstantIdx && "Cannot replace dynamic transfer with insert");
// 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, &TD, 0));
+ AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &DL, 0));
- if (GetUnderlyingObject(MTI->getSource(), &TD, 0) != OrigAI) {
+ if (GetUnderlyingObject(MTI->getSource(), &DL, 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?");
LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval");
SrcVal->setAlignment(MTI->getAlignment());
Builder.CreateStore(SrcVal, NewAI);
- } else if (GetUnderlyingObject(MTI->getDest(), &TD, 0) != OrigAI) {
+ } else if (GetUnderlyingObject(MTI->getDest(), &DL, 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?");
}
}
-/// getScaledElementType - Gets a scaled element type for a partial vector
-/// access of an alloca. The input types must be integer or floating-point
-/// scalar or vector types, and the resulting type is an integer, float or
-/// double.
-static Type *getScaledElementType(Type *Ty1, Type *Ty2,
- unsigned NewBitWidth) {
- bool IsFP1 = Ty1->isFloatingPointTy() ||
- (Ty1->isVectorTy() &&
- cast<VectorType>(Ty1)->getElementType()->isFloatingPointTy());
- bool IsFP2 = Ty2->isFloatingPointTy() ||
- (Ty2->isVectorTy() &&
- cast<VectorType>(Ty2)->getElementType()->isFloatingPointTy());
-
- LLVMContext &Context = Ty1->getContext();
-
- // Prefer floating-point types over integer types, as integer types may have
- // been created by earlier scalar replacement.
- if (IsFP1 || IsFP2) {
- if (NewBitWidth == 32)
- return Type::getFloatTy(Context);
- if (NewBitWidth == 64)
- return Type::getDoubleTy(Context);
- }
-
- return Type::getIntNTy(Context, NewBitWidth);
-}
-
-/// CreateShuffleVectorCast - Creates a shuffle vector to convert one vector
-/// to another vector of the same element type which has the same allocation
-/// size but different primitive sizes (e.g. <3 x i32> and <4 x i32>).
-static Value *CreateShuffleVectorCast(Value *FromVal, Type *ToType,
- IRBuilder<> &Builder) {
- Type *FromType = FromVal->getType();
- VectorType *FromVTy = cast<VectorType>(FromType);
- VectorType *ToVTy = cast<VectorType>(ToType);
- assert((ToVTy->getElementType() == FromVTy->getElementType()) &&
- "Vectors must have the same element type");
- Value *UnV = UndefValue::get(FromType);
- unsigned numEltsFrom = FromVTy->getNumElements();
- unsigned numEltsTo = ToVTy->getNumElements();
-
- SmallVector<Constant*, 3> Args;
- Type* Int32Ty = Builder.getInt32Ty();
- unsigned minNumElts = std::min(numEltsFrom, numEltsTo);
- unsigned i;
- for (i=0; i != minNumElts; ++i)
- Args.push_back(ConstantInt::get(Int32Ty, i));
-
- if (i < numEltsTo) {
- Constant* UnC = UndefValue::get(Int32Ty);
- for (; i != numEltsTo; ++i)
- Args.push_back(UnC);
- }
- Constant *Mask = ConstantVector::get(Args);
- return Builder.CreateShuffleVector(FromVal, UnV, Mask, "tmpV");
-}
-
/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer
/// or vector value FromVal, extracting the bits from the offset specified by
/// Offset. This returns the value, which is of type ToType.
/// shifted to the right.
Value *ConvertToScalarInfo::
ConvertScalar_ExtractValue(Value *FromVal, Type *ToType,
- uint64_t Offset, IRBuilder<> &Builder) {
+ uint64_t Offset, Value* NonConstantIdx,
+ IRBuilder<> &Builder) {
// If the load is of the whole new alloca, no conversion is needed.
Type *FromType = FromVal->getType();
if (FromType == ToType && Offset == 0)
// 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 (VectorType *VTy = dyn_cast<VectorType>(FromType)) {
- unsigned FromTypeSize = TD.getTypeAllocSize(FromType);
- unsigned ToTypeSize = TD.getTypeAllocSize(ToType);
- if (FromTypeSize == ToTypeSize) {
- // If the two types have the same primitive size, use a bit cast.
- // Otherwise, it is two vectors with the same element type that has
- // the same allocation size but different number of elements so use
- // a shuffle vector.
- if (FromType->getPrimitiveSizeInBits() ==
- ToType->getPrimitiveSizeInBits())
+ unsigned FromTypeSize = DL.getTypeAllocSize(FromType);
+ unsigned ToTypeSize = DL.getTypeAllocSize(ToType);
+ if (FromTypeSize == ToTypeSize)
return Builder.CreateBitCast(FromVal, ToType);
- else
- return CreateShuffleVectorCast(FromVal, ToType, Builder);
- }
-
- if (isPowerOf2_64(FromTypeSize / ToTypeSize)) {
- assert(!(ToType->isVectorTy() && Offset != 0) && "Can't extract a value "
- "of a smaller vector type at a nonzero offset.");
-
- Type *CastElementTy = getScaledElementType(FromType, ToType,
- ToTypeSize * 8);
- unsigned NumCastVectorElements = FromTypeSize / ToTypeSize;
-
- LLVMContext &Context = FromVal->getContext();
- Type *CastTy = VectorType::get(CastElementTy,
- NumCastVectorElements);
- Value *Cast = Builder.CreateBitCast(FromVal, CastTy);
-
- unsigned EltSize = TD.getTypeAllocSizeInBits(CastElementTy);
- unsigned Elt = Offset/EltSize;
- assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
- Value *Extract = Builder.CreateExtractElement(Cast, ConstantInt::get(
- Type::getInt32Ty(Context), Elt));
- return Builder.CreateBitCast(Extract, ToType);
- }
// Otherwise it must be an element access.
unsigned Elt = 0;
if (Offset) {
- unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType());
+ unsigned EltSize = DL.getTypeAllocSizeInBits(VTy->getElementType());
Elt = Offset/EltSize;
assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
}
// Return the element extracted out of it.
- Value *V = Builder.CreateExtractElement(FromVal, Builder.getInt32(Elt));
+ Value *Idx;
+ if (NonConstantIdx) {
+ if (Elt)
+ Idx = Builder.CreateAdd(NonConstantIdx,
+ Builder.getInt32(Elt),
+ "dyn.offset");
+ else
+ Idx = NonConstantIdx;
+ } else
+ Idx = Builder.getInt32(Elt);
+ Value *V = Builder.CreateExtractElement(FromVal, Idx);
if (V->getType() != ToType)
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 (StructType *ST = dyn_cast<StructType>(ToType)) {
- const StructLayout &Layout = *TD.getStructLayout(ST);
+ assert(!NonConstantIdx &&
+ "Dynamic indexing into struct types not supported");
+ const StructLayout &Layout = *DL.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);
+ 0, Builder);
Res = Builder.CreateInsertValue(Res, Elt, i);
}
return Res;
}
if (ArrayType *AT = dyn_cast<ArrayType>(ToType)) {
- uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType());
+ assert(!NonConstantIdx &&
+ "Dynamic indexing into array types not supported");
+ uint64_t EltSize = DL.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);
+ Offset+i*EltSize, 0, Builder);
Res = Builder.CreateInsertValue(Res, Elt, i);
}
return Res;
// 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.
int ShAmt = 0;
- if (TD.isBigEndian()) {
+ if (DL.isBigEndian()) {
// On big-endian machines, the lowest bit is stored at the bit offset
// from the pointer given by getTypeStoreSizeInBits. This matters for
// integers with a bitwidth that is not a multiple of 8.
- ShAmt = TD.getTypeStoreSizeInBits(NTy) -
- TD.getTypeStoreSizeInBits(ToType) - Offset;
+ ShAmt = DL.getTypeStoreSizeInBits(NTy) -
+ DL.getTypeStoreSizeInBits(ToType) - Offset;
} else {
ShAmt = Offset;
}
ConstantInt::get(FromVal->getType(), -ShAmt));
// Finally, unconditionally truncate the integer to the right width.
- unsigned LIBitWidth = TD.getTypeSizeInBits(ToType);
+ unsigned LIBitWidth = DL.getTypeSizeInBits(ToType);
if (LIBitWidth < NTy->getBitWidth())
FromVal =
Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(),
///
/// Offset is an offset from the original alloca, in bits that need to be
/// shifted to the right.
+///
+/// NonConstantIdx is an index value if there was a GEP with a non-constant
+/// index value. If this is 0 then all GEPs used to find this insert address
+/// are constant.
Value *ConvertToScalarInfo::
ConvertScalar_InsertValue(Value *SV, Value *Old,
- uint64_t Offset, IRBuilder<> &Builder) {
+ uint64_t Offset, Value* NonConstantIdx,
+ IRBuilder<> &Builder) {
// Convert the stored type to the actual type, shift it left to insert
// then 'or' into place.
Type *AllocaType = Old->getType();
LLVMContext &Context = Old->getContext();
if (VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
- uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy);
- uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType());
+ uint64_t VecSize = DL.getTypeAllocSizeInBits(VTy);
+ uint64_t ValSize = DL.getTypeAllocSizeInBits(SV->getType());
// Changing the whole vector with memset or with an access of a different
// vector type?
- if (ValSize == VecSize) {
- // If the two types have the same primitive size, use a bit cast.
- // Otherwise, it is two vectors with the same element type that has
- // the same allocation size but different number of elements so use
- // a shuffle vector.
- if (VTy->getPrimitiveSizeInBits() ==
- SV->getType()->getPrimitiveSizeInBits())
+ if (ValSize == VecSize)
return Builder.CreateBitCast(SV, AllocaType);
- else
- return CreateShuffleVectorCast(SV, VTy, Builder);
- }
-
- if (isPowerOf2_64(VecSize / ValSize)) {
- assert(!(SV->getType()->isVectorTy() && Offset != 0) && "Can't insert a "
- "value of a smaller vector type at a nonzero offset.");
-
- Type *CastElementTy = getScaledElementType(VTy, SV->getType(),
- ValSize);
- unsigned NumCastVectorElements = VecSize / ValSize;
-
- Type *OldCastTy = VectorType::get(CastElementTy, NumCastVectorElements);
- Value *OldCast = Builder.CreateBitCast(Old, OldCastTy);
-
- Value *SVCast = Builder.CreateBitCast(SV, CastElementTy);
-
- unsigned EltSize = TD.getTypeAllocSizeInBits(CastElementTy);
- unsigned Elt = Offset/EltSize;
- assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
- Value *Insert =
- Builder.CreateInsertElement(OldCast, SVCast, Builder.getInt32(Elt));
- return Builder.CreateBitCast(Insert, AllocaType);
- }
// Must be an element insertion.
- assert(SV->getType() == VTy->getElementType());
- uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType());
+ Type *EltTy = VTy->getElementType();
+ if (SV->getType() != EltTy)
+ SV = Builder.CreateBitCast(SV, EltTy);
+ uint64_t EltSize = DL.getTypeAllocSizeInBits(EltTy);
unsigned Elt = Offset/EltSize;
- return Builder.CreateInsertElement(Old, SV, Builder.getInt32(Elt));
+ Value *Idx;
+ if (NonConstantIdx) {
+ if (Elt)
+ Idx = Builder.CreateAdd(NonConstantIdx,
+ Builder.getInt32(Elt),
+ "dyn.offset");
+ else
+ Idx = NonConstantIdx;
+ } else
+ Idx = Builder.getInt32(Elt);
+ return Builder.CreateInsertElement(Old, SV, Idx);
}
// If SV is a first-class aggregate value, insert each value recursively.
if (StructType *ST = dyn_cast<StructType>(SV->getType())) {
- const StructLayout &Layout = *TD.getStructLayout(ST);
+ assert(!NonConstantIdx &&
+ "Dynamic indexing into struct types not supported");
+ const StructLayout &Layout = *DL.getStructLayout(ST);
for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
Value *Elt = Builder.CreateExtractValue(SV, i);
Old = ConvertScalar_InsertValue(Elt, Old,
Offset+Layout.getElementOffsetInBits(i),
- Builder);
+ 0, Builder);
}
return Old;
}
if (ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) {
- uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType());
+ assert(!NonConstantIdx &&
+ "Dynamic indexing into array types not supported");
+ uint64_t EltSize = DL.getTypeAllocSizeInBits(AT->getElementType());
for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
Value *Elt = Builder.CreateExtractValue(SV, i);
- Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder);
+ Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, 0, Builder);
}
return Old;
}
// If SV is a float, convert it to the appropriate integer type.
// If it is a pointer, do the same.
- unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType());
- unsigned DestWidth = TD.getTypeSizeInBits(AllocaType);
- unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType());
- unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType);
+ unsigned SrcWidth = DL.getTypeSizeInBits(SV->getType());
+ unsigned DestWidth = DL.getTypeSizeInBits(AllocaType);
+ unsigned SrcStoreWidth = DL.getTypeStoreSizeInBits(SV->getType());
+ unsigned DestStoreWidth = DL.getTypeStoreSizeInBits(AllocaType);
if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy())
SV = Builder.CreateBitCast(SV, IntegerType::get(SV->getContext(),SrcWidth));
else if (SV->getType()->isPointerTy())
- SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()));
+ SV = Builder.CreatePtrToInt(SV, DL.getIntPtrType(SV->getType()));
// Zero extend or truncate the value if needed.
if (SV->getType() != AllocaType) {
// If this is a big-endian system and the store is narrower than the
// full alloca type, we need to do a shift to get the right bits.
int ShAmt = 0;
- if (TD.isBigEndian()) {
+ if (DL.isBigEndian()) {
// On big-endian machines, the lowest bit is stored at the bit offset
// from the pointer given by getTypeStoreSizeInBits. This matters for
// integers with a bitwidth that is not a multiple of 8.
bool SROA::runOnFunction(Function &F) {
- TD = getAnalysisIfAvailable<TargetData>();
+ if (skipOptnoneFunction(F))
+ return false;
+
+ DL = getAnalysisIfAvailable<DataLayout>();
bool Changed = performPromotion(F);
- // FIXME: ScalarRepl currently depends on TargetData more than it
+ // FIXME: ScalarRepl currently depends on DataLayout more than it
// theoretically needs to. It should be refactored in order to support
// target-independent IR. Until this is done, just skip the actual
// scalar-replacement portion of this pass.
- if (!TD) return Changed;
+ if (!DL) return Changed;
while (1) {
bool LocalChange = performScalarRepl(F);
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))
+ 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(),
+ for (SmallVectorImpl<DbgDeclareInst *>::iterator I = DDIs.begin(),
E = DDIs.end(); I != E; ++I) {
DbgDeclareInst *DDI = *I;
DDI->eraseFromParent();
}
- for (SmallVector<DbgValueInst *, 4>::iterator I = DVIs.begin(),
+ for (SmallVectorImpl<DbgValueInst *>::iterator I = DVIs.begin(),
E = DVIs.end(); I != E; ++I) {
DbgValueInst *DVI = *I;
DVI->eraseFromParent();
}
}
-
+
virtual bool isInstInList(Instruction *I,
const SmallVectorImpl<Instruction*> &Insts) const {
if (LoadInst *LI = dyn_cast<LoadInst>(I))
}
virtual void updateDebugInfo(Instruction *Inst) const {
- for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(),
+ for (SmallVectorImpl<DbgDeclareInst *>::const_iterator I = DDIs.begin(),
E = DDIs.end(); I != E; ++I) {
DbgDeclareInst *DDI = *I;
if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
ConvertDebugDeclareToDebugValue(DDI, LI, *DIB);
}
- for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(),
+ for (SmallVectorImpl<DbgValueInst *>::const_iterator I = DVIs.begin(),
E = DVIs.end(); I != E; ++I) {
DbgValueInst *DVI = *I;
+ Value *Arg = NULL;
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));
+ Arg = 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());
+ Arg = dyn_cast<Argument>(SExt->getOperand(0));
+ if (!Arg)
+ Arg = SI->getOperand(0);
} else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
- Instruction *DbgVal =
- DIB->insertDbgValueIntrinsic(LI->getOperand(0), 0,
- DIVariable(DVI->getVariable()), LI);
- DbgVal->setDebugLoc(DVI->getDebugLoc());
+ Arg = LI->getOperand(0);
+ } else {
+ continue;
}
+ Instruction *DbgVal =
+ DIB->insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
+ Inst);
+ DbgVal->setDebugLoc(DVI->getDebugLoc());
}
}
};
///
/// 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) {
+static bool isSafeSelectToSpeculate(SelectInst *SI, const DataLayout *DL) {
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))
+ LI->getAlignment(), DL))
return false;
if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI,
- LI->getAlignment(), TD))
+ LI->getAlignment(), DL))
return false;
}
-
+
return true;
}
///
/// 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) {
+static bool isSafePHIToSpeculate(PHINode *PN, const DataLayout *DL) {
// 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.
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
// 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))
+ isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, DL))
continue;
-
+
return false;
}
-
+
return true;
}
/// 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) {
+static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const DataLayout *DL) {
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;
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.
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);
+ return tryToMakeAllocaBePromotable(AI, DL);
}
// 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))
+ if (!isSafeSelectToSpeculate(SI, DL))
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))
+ if (!isSafePHIToSpeculate(PN, DL))
return false;
-
+
InstsToRewrite.insert(PN);
continue;
}
-
+
if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
if (onlyUsedByLifetimeMarkers(BCI)) {
InstsToRewrite.insert(BCI);
continue;
}
}
-
+
return false;
}
// 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) {
// loads with a new select.
while (!SI->use_empty()) {
LoadInst *LI = cast<LoadInst>(SI->use_back());
-
+
IRBuilder<> Builder(LI);
- LoadInst *TrueLoad =
+ LoadInst *TrueLoad =
Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t");
- LoadInst *FalseLoad =
+ LoadInst *FalseLoad =
Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".f");
-
+
// Transfer alignment and TBAA info if present.
TrueLoad->setAlignment(LI->getAlignment());
FalseLoad->setAlignment(LI->getAlignment());
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]);
PN->eraseFromParent();
continue;
}
-
+
Type *LoadTy = cast<PointerType>(PN->getType())->getElementType();
PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(),
PN->getName()+".ld", PN);
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];
Load->setAlignment(Align);
if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
}
-
+
NewPN->addIncoming(Load, Pred);
}
-
+
PN->eraseFromParent();
}
-
+
++NumAdjusted;
return true;
}
std::vector<AllocaInst*> Allocas;
DominatorTree *DT = 0;
if (HasDomTree)
- DT = &getAnalysis<DominatorTree>();
+ DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
DIBuilder DIB(*F.getParent());
// 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 (tryToMakeAllocaBePromotable(AI, TD))
+ if (tryToMakeAllocaBePromotable(AI, DL))
Allocas.push_back(AI);
if (Allocas.empty()) break;
SSAUpdater SSA;
for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
AllocaInst *AI = Allocas[i];
-
+
// Build list of instructions to promote.
for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
UI != E; ++UI)
/// 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) {
+bool SROA::ShouldAttemptScalarRepl(AllocaInst *AI) {
Type *T = AI->getAllocatedType();
- // Do not promote any struct into more than 32 separate vars.
+ // Do not promote any struct that has too many members.
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.
+ return ST->getNumElements() <= StructMemberThreshold;
+ // Do not promote any array that has too many elements.
if (ArrayType *AT = dyn_cast<ArrayType>(T))
- return AT->getNumElements() <= 8;
+ return AT->getNumElements() <= ArrayElementThreshold;
return false;
}
-
// performScalarRepl - This algorithm is a simple worklist driven algorithm,
-// which runs on all of the 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 entry block, removing
+// them if they are only used by getelementptr instructions.
//
bool SROA::performScalarRepl(Function &F) {
std::vector<AllocaInst*> WorkList;
if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized())
continue;
- // Check to see if this allocation is only modified by a memcpy/memmove from
- // a constant global. If this is the case, we can change all users to use
- // 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.
- SmallVector<Instruction *, 4> ToDelete;
- if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(AI, ToDelete)) {
- DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n');
- 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()));
- Copy->eraseFromParent(); // Don't mutate the global.
- AI->eraseFromParent();
- ++NumGlobals;
- Changed = true;
- continue;
- }
-
// Check to see if we can perform the core SROA transformation. We cannot
// transform the allocation instruction if it is an array allocation
// (allocations OF arrays are ok though), and an allocation of a scalar
// value cannot be decomposed at all.
- uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType());
+ uint64_t AllocaSize = DL->getTypeAllocSize(AI->getAllocatedType());
// Do not promote [0 x %struct].
if (AllocaSize == 0) continue;
// promoted itself. If so, we don't want to transform it needlessly. Note
// that we can't just check based on the type: the alloca may be of an i32
// but that has pointer arithmetic to set byte 3 of it or something.
- if (AllocaInst *NewAI =
- ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) {
+ if (AllocaInst *NewAI = ConvertToScalarInfo(
+ (unsigned)AllocaSize, *DL, ScalarLoadThreshold).TryConvert(AI)) {
NewAI->takeName(AI);
AI->eraseFromParent();
++NumConverted;
ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
if (Length == 0)
return MarkUnsafe(Info, User);
+ if (Length->isNegative())
+ return MarkUnsafe(Info, User);
+
isSafeMemAccess(Offset, Length->getZExtValue(), 0,
UI.getOperandNo() == 0, Info, MI,
true /*AllowWholeAccess*/);
if (!LI->isSimple())
return MarkUnsafe(Info, User);
Type *LIType = LI->getType();
- isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType),
+ isSafeMemAccess(Offset, DL->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->isSimple() || SI->getOperand(0) == I)
return MarkUnsafe(Info, User);
-
+
Type *SIType = SI->getOperand(0)->getType();
- isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType),
+ isSafeMemAccess(Offset, DL->getTypeAllocSize(SIType),
SIType, true, Info, SI, true /*AllowWholeAccess*/);
Info.hasALoadOrStore = true;
} else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(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.
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)) {
if (!LI->isSimple())
return MarkUnsafe(Info, User);
Type *LIType = LI->getType();
- isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType),
+ isSafeMemAccess(Offset, DL->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),
+ isSafeMemAccess(Offset, DL->getTypeAllocSize(SIType),
SIType, true, Info, SI, false /*AllowWholeAccess*/);
Info.hasALoadOrStore = true;
} else if (isa<PHINode>(User) || isa<SelectInst>(User)) {
gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI);
if (GEPIt == E)
return;
+ bool NonConstant = false;
+ unsigned NonConstantIdxSize = 0;
// Walk through the GEP type indices, checking the types that this indexes
// into.
// 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);
- if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0))
+ // If this GEP is non-constant then the last operand must have been a
+ // dynamic index into a vector. Pop this now as it has no impact on the
+ // constant part of the offset.
+ if (NonConstant)
+ Indices.pop_back();
+ Offset += DL->getIndexedOffset(GEPI->getPointerOperandType(), Indices);
+ if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset,
+ NonConstantIdxSize))
MarkUnsafe(Info, GEPI);
}
bool AllowWholeAccess) {
// Check if this is a load/store of the entire alloca.
if (Offset == 0 && AllowWholeAccess &&
- MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) {
+ MemSize == DL->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
Type *EltTy;
uint64_t EltSize;
if (StructType *ST = dyn_cast<StructType>(T)) {
- const StructLayout *Layout = TD->getStructLayout(ST);
+ const StructLayout *Layout = DL->getStructLayout(ST);
unsigned EltIdx = Layout->getElementContainingOffset(Offset);
EltTy = ST->getContainedType(EltIdx);
- EltSize = TD->getTypeAllocSize(EltTy);
+ EltSize = DL->getTypeAllocSize(EltTy);
Offset -= Layout->getElementOffset(EltIdx);
} else if (ArrayType *AT = dyn_cast<ArrayType>(T)) {
EltTy = AT->getElementType();
- EltSize = TD->getTypeAllocSize(EltTy);
+ EltSize = DL->getTypeAllocSize(EltTy);
if (Offset >= AT->getNumElements() * EltSize)
return false;
Offset %= EltSize;
+ } else if (VectorType *VT = dyn_cast<VectorType>(T)) {
+ EltTy = VT->getElementType();
+ EltSize = DL->getTypeAllocSize(EltTy);
+ if (Offset >= VT->getNumElements() * EltSize)
+ return false;
+ Offset %= EltSize;
} else {
return false;
}
/// Offset indicates the position within AI that is referenced by this
/// instruction.
void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
- SmallVector<AllocaInst*, 32> &NewElts) {
+ SmallVectorImpl<AllocaInst *> &NewElts) {
for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) {
Use &TheUse = UI.getUse();
Instruction *User = cast<Instruction>(*UI++);
RewriteBitCast(BC, AI, Offset, NewElts);
continue;
}
-
+
if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
RewriteGEP(GEPI, AI, Offset, NewElts);
continue;
}
-
+
if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
uint64_t MemSize = Length->getZExtValue();
if (Offset == 0 &&
- MemSize == TD->getTypeAllocSize(AI->getAllocatedType()))
+ MemSize == DL->getTypeAllocSize(AI->getAllocatedType()))
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.
}
continue;
}
-
+
if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
Type *LIType = LI->getType();
-
+
if (isCompatibleAggregate(LIType, AI->getAllocatedType())) {
// Replace:
// %res = load { i32, i32 }* %alloc
LI->replaceAllUsesWith(Insert);
DeadInsts.push_back(LI);
} else if (LIType->isIntegerTy() &&
- TD->getTypeAllocSize(LIType) ==
- TD->getTypeAllocSize(AI->getAllocatedType())) {
+ DL->getTypeAllocSize(LIType) ==
+ DL->getTypeAllocSize(AI->getAllocatedType())) {
// If this is a load of the entire alloca to an integer, rewrite it.
RewriteLoadUserOfWholeAlloca(LI, AI, NewElts);
}
continue;
}
-
+
if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
Value *Val = SI->getOperand(0);
Type *SIType = Val->getType();
}
DeadInsts.push_back(SI);
} else if (SIType->isIntegerTy() &&
- TD->getTypeAllocSize(SIType) ==
- TD->getTypeAllocSize(AI->getAllocatedType())) {
+ DL->getTypeAllocSize(SIType) ==
+ DL->getTypeAllocSize(AI->getAllocatedType())) {
// 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
+ // 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.
/// RewriteBitCast - Update a bitcast reference to the alloca being replaced
/// and recursively continue updating all of its uses.
void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
- SmallVector<AllocaInst*, 32> &NewElts) {
+ SmallVectorImpl<AllocaInst *> &NewElts) {
RewriteForScalarRepl(BC, AI, Offset, NewElts);
if (BC->getOperand(0) != AI)
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);
Type *&IdxTy) {
uint64_t Idx = 0;
if (StructType *ST = dyn_cast<StructType>(T)) {
- const StructLayout *Layout = TD->getStructLayout(ST);
+ const StructLayout *Layout = DL->getStructLayout(ST);
Idx = Layout->getElementContainingOffset(Offset);
T = ST->getContainedType(Idx);
Offset -= Layout->getElementOffset(Idx);
IdxTy = Type::getInt32Ty(T->getContext());
return Idx;
+ } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) {
+ T = AT->getElementType();
+ uint64_t EltSize = DL->getTypeAllocSize(T);
+ Idx = Offset / EltSize;
+ Offset -= Idx * EltSize;
+ IdxTy = Type::getInt64Ty(T->getContext());
+ return Idx;
}
- ArrayType *AT = cast<ArrayType>(T);
- T = AT->getElementType();
- uint64_t EltSize = TD->getTypeAllocSize(T);
+ VectorType *VT = cast<VectorType>(T);
+ T = VT->getElementType();
+ uint64_t EltSize = DL->getTypeAllocSize(T);
Idx = Offset / EltSize;
Offset -= Idx * EltSize;
IdxTy = Type::getInt64Ty(T->getContext());
/// elements of the alloca that are being split apart, and if so, rewrite
/// the GEP to be relative to the new element.
void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
- SmallVector<AllocaInst*, 32> &NewElts) {
+ SmallVectorImpl<AllocaInst *> &NewElts) {
uint64_t OldOffset = Offset;
SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
- Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices);
+ // If the GEP was dynamic then it must have been a dynamic vector lookup.
+ // In this case, it must be the last GEP operand which is dynamic so keep that
+ // aside until we've found the constant GEP offset then add it back in at the
+ // end.
+ Value* NonConstantIdx = 0;
+ if (!GEPI->hasAllConstantIndices())
+ NonConstantIdx = Indices.pop_back_val();
+ Offset += DL->getIndexedOffset(GEPI->getPointerOperandType(), Indices);
RewriteForScalarRepl(GEPI, AI, Offset, NewElts);
uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy);
NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx));
}
+ if (NonConstantIdx) {
+ Type* GepTy = T;
+ // This GEP has a dynamic index. We need to add "i32 0" to index through
+ // any structs or arrays in the original type until we get to the vector
+ // to index.
+ while (!isa<VectorType>(GepTy)) {
+ NewArgs.push_back(Constant::getNullValue(i32Ty));
+ GepTy = cast<CompositeType>(GepTy)->getTypeAtIndex(0U);
+ }
+ NewArgs.push_back(NonConstantIdx);
+ }
Instruction *Val = NewElts[Idx];
if (NewArgs.size() > 1) {
Val = GetElementPtrInst::CreateInBounds(Val, NewArgs, "", GEPI);
/// to mark the lifetime of the scalarized memory.
void SROA::RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI,
uint64_t Offset,
- SmallVector<AllocaInst*, 32> &NewElts) {
+ SmallVectorImpl<AllocaInst *> &NewElts) {
ConstantInt *OldSize = cast<ConstantInt>(II->getArgOperand(0));
// Put matching lifetime markers on everything from Offset up to
// Offset+OldSize.
V = Builder.CreateGEP(V, Builder.getInt64(NewOffset));
IdxTy = NewElts[Idx]->getAllocatedType();
- uint64_t EltSize = TD->getTypeAllocSize(IdxTy) - NewOffset;
+ uint64_t EltSize = DL->getTypeAllocSize(IdxTy) - NewOffset;
if (EltSize > Size) {
EltSize = Size;
Size = 0;
for (; Idx != NewElts.size() && Size; ++Idx) {
IdxTy = NewElts[Idx]->getAllocatedType();
- uint64_t EltSize = TD->getTypeAllocSize(IdxTy);
+ uint64_t EltSize = DL->getTypeAllocSize(IdxTy);
if (EltSize > Size) {
EltSize = Size;
Size = 0;
/// 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,
- AllocaInst *AI,
- SmallVector<AllocaInst*, 32> &NewElts) {
+void
+SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
+ AllocaInst *AI,
+ SmallVectorImpl<AllocaInst *> &NewElts) {
// If this is a memcpy/memmove, construct the other pointer as the
// appropriate type. The "Other" pointer is the pointer that goes to memory
// that doesn't have anything to do with the alloca that we are promoting. For
if (OtherPtr == AI || OtherPtr == NewElts[0]) {
// This code will run twice for a no-op memcpy -- once for each operand.
// Put only one reference to MI on the DeadInsts list.
- for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(),
+ for (SmallVectorImpl<Value *>::const_iterator I = DeadInsts.begin(),
E = DeadInsts.end(); I != E; ++I)
if (*I == MI) return;
DeadInsts.push_back(MI);
PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType());
Type *OtherTy = OtherPtrTy->getElementType();
if (StructType *ST = dyn_cast<StructType>(OtherTy)) {
- EltOffset = TD->getStructLayout(ST)->getElementOffset(i);
+ EltOffset = DL->getStructLayout(ST)->getElementOffset(i);
} else {
Type *EltTy = cast<SequentialType>(OtherTy)->getElementType();
- EltOffset = TD->getTypeAllocSize(EltTy)*i;
+ EltOffset = DL->getTypeAllocSize(EltTy)*i;
}
// The alignment of the other pointer is the guaranteed alignment of the
Type *ValTy = EltTy->getScalarType();
// Construct an integer with the right value.
- unsigned EltSize = TD->getTypeSizeInBits(ValTy);
+ unsigned EltSize = DL->getTypeSizeInBits(ValTy);
APInt OneVal(EltSize, CI->getZExtValue());
APInt TotalVal(OneVal);
// Set each byte.
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();
- SmallVector<Constant*, 16> Elts(NumElts, StoreVal);
- StoreVal = ConstantVector::get(Elts);
+ if (EltTy->isVectorTy()) {
+ unsigned NumElts = cast<VectorType>(EltTy)->getNumElements();
+ StoreVal = ConstantVector::getSplat(NumElts, StoreVal);
}
}
new StoreInst(StoreVal, EltPtr, MI);
// this element.
}
- unsigned EltSize = TD->getTypeAllocSize(EltTy);
+ unsigned EltSize = DL->getTypeAllocSize(EltTy);
+ if (!EltSize)
+ continue;
IRBuilder<> Builder(MI);
/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that
/// overwrites the entire allocation. Extract out the pieces of the stored
/// integer and store them individually.
-void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
- SmallVector<AllocaInst*, 32> &NewElts){
+void
+SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
+ SmallVectorImpl<AllocaInst *> &NewElts) {
// 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);
Type *AllocaEltTy = AI->getAllocatedType();
- uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
+ uint64_t AllocaSizeBits = DL->getTypeAllocSizeInBits(AllocaEltTy);
IRBuilder<> Builder(SI);
-
+
// Handle tail padding by extending the operand
- if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits)
+ if (DL->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits)
SrcVal = Builder.CreateZExt(SrcVal,
IntegerType::get(SI->getContext(), AllocaSizeBits));
// There are two forms here: AI could be an array or struct. Both cases
// have different ways to compute the element offset.
if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
- const StructLayout *Layout = TD->getStructLayout(EltSTy);
+ const StructLayout *Layout = DL->getStructLayout(EltSTy);
for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
// Get the number of bits to shift SrcVal to get the value.
Type *FieldTy = EltSTy->getElementType(i);
uint64_t Shift = Layout->getElementOffsetInBits(i);
- if (TD->isBigEndian())
- Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy);
+ if (DL->isBigEndian())
+ Shift = AllocaSizeBits-Shift-DL->getTypeAllocSizeInBits(FieldTy);
Value *EltVal = SrcVal;
if (Shift) {
}
// Truncate down to an integer of the right size.
- uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
+ uint64_t FieldSizeBits = DL->getTypeSizeInBits(FieldTy);
// Ignore zero sized fields like {}, they obviously contain no data.
if (FieldSizeBits == 0) continue;
} else {
ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
Type *ArrayEltTy = ATy->getElementType();
- uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
- uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy);
+ uint64_t ElementOffset = DL->getTypeAllocSizeInBits(ArrayEltTy);
+ uint64_t ElementSizeBits = DL->getTypeSizeInBits(ArrayEltTy);
uint64_t Shift;
- if (TD->isBigEndian())
+ if (DL->isBigEndian())
Shift = AllocaSizeBits-ElementOffset;
else
Shift = 0;
}
new StoreInst(EltVal, DestField, SI);
- if (TD->isBigEndian())
+ if (DL->isBigEndian())
Shift -= ElementOffset;
else
Shift += ElementOffset;
/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to
/// an integer. Load the individual pieces to form the aggregate value.
-void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
- SmallVector<AllocaInst*, 32> &NewElts) {
+void
+SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
+ SmallVectorImpl<AllocaInst *> &NewElts) {
// Extract each element out of the NewElts according to its structure offset
// and form the result value.
Type *AllocaEltTy = AI->getAllocatedType();
- uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
+ uint64_t AllocaSizeBits = DL->getTypeAllocSizeInBits(AllocaEltTy);
DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI
<< '\n');
const StructLayout *Layout = 0;
uint64_t ArrayEltBitOffset = 0;
if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
- Layout = TD->getStructLayout(EltSTy);
+ Layout = DL->getStructLayout(EltSTy);
} else {
Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
- ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
+ ArrayEltBitOffset = DL->getTypeAllocSizeInBits(ArrayEltTy);
}
Value *ResultVal =
Value *SrcField = NewElts[i];
Type *FieldTy =
cast<PointerType>(SrcField->getType())->getElementType();
- uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
+ uint64_t FieldSizeBits = DL->getTypeSizeInBits(FieldTy);
// Ignore zero sized fields like {}, they obviously contain no data.
if (FieldSizeBits == 0) continue;
else // Array case.
Shift = i*ArrayEltBitOffset;
- if (TD->isBigEndian())
+ if (DL->isBigEndian())
Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth();
if (Shift) {
}
// Handle tail padding by truncating the result
- if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits)
+ if (DL->getTypeSizeInBits(LI->getType()) != AllocaSizeBits)
ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI);
LI->replaceAllUsesWith(ResultVal);
/// 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(Type *Ty, const TargetData &TD) {
+static bool HasPadding(Type *Ty, const DataLayout &DL) {
if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
Ty = ATy->getElementType();
- return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty);
+ return DL.getTypeSizeInBits(Ty) != DL.getTypeAllocSizeInBits(Ty);
}
// SROA currently handles only Arrays and Structs.
StructType *STy = cast<StructType>(Ty);
- const StructLayout *SL = TD.getStructLayout(STy);
+ const StructLayout *SL = DL.getStructLayout(STy);
unsigned PrevFieldBitOffset = 0;
for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
// previous one.
if (i) {
unsigned PrevFieldEnd =
- PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1));
+ PrevFieldBitOffset+DL.getTypeSizeInBits(STy->getElementType(i-1));
if (PrevFieldEnd < FieldBitOffset)
return true;
}
// Check for tail padding.
if (unsigned EltCount = STy->getNumElements()) {
unsigned PrevFieldEnd = PrevFieldBitOffset +
- TD.getTypeSizeInBits(STy->getElementType(EltCount-1));
+ DL.getTypeSizeInBits(STy->getElementType(EltCount-1));
if (PrevFieldEnd < SL->getSizeInBits())
return true;
}
// types, but may actually be used. In these cases, we refuse to promote the
// struct.
if (Info.isMemCpySrc && Info.isMemCpyDst &&
- HasPadding(AI->getAllocatedType(), *TD))
+ HasPadding(AI->getAllocatedType(), *DL))
return false;
// If the alloca never has an access to just *part* of it, but is accessed
return false;
}
}
-
- return true;
-}
-
-
-
-/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to
-/// some part of a constant global variable. This intentionally only accepts
-/// constant expressions because we don't can't rewrite arbitrary instructions.
-static bool PointsToConstantGlobal(Value *V) {
- if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
- return GV->isConstant();
- if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
- if (CE->getOpcode() == Instruction::BitCast ||
- CE->getOpcode() == Instruction::GetElementPtr)
- return PointsToConstantGlobal(CE->getOperand(0));
- return false;
-}
-
-/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
-/// pointer to an alloca. Ignore any reads of the pointer, return false if we
-/// see any stores or other unknown uses. If we see pointer arithmetic, keep
-/// track of whether it moves the pointer (with isOffset) but otherwise traverse
-/// 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,
- 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->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,
- LifetimeMarkers))
- return false;
- continue;
- }
- if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
- // 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(),
- LifetimeMarkers))
- return false;
- continue;
- }
- if (CallSite CS = U) {
- // 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.paramHasAttr(ArgNo+1, Attribute::NoCapture)))
- 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.
- if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal))
- 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);
- if (MI == 0)
- return false;
-
- // If the transfer is using the alloca as a source of the transfer, then
- // ignore it since it is a load (unless the transfer is volatile).
- if (UI.getOperandNo() == 1) {
- if (MI->isVolatile()) return false;
- continue;
- }
-
- // If we already have seen a copy, reject the second one.
- if (TheCopy) return false;
-
- // If the pointer has been offset from the start of the alloca, we can't
- // safely handle this.
- if (isOffset) return false;
-
- // If the memintrinsic isn't using the alloca as the dest, reject it.
- if (UI.getOperandNo() != 0) return false;
-
- // If the source of the memcpy/move is not a constant global, reject it.
- if (!PointsToConstantGlobal(MI->getSource()))
- return false;
-
- // Otherwise, the transform is safe. Remember the copy instruction.
- TheCopy = MI;
- }
return true;
}
-
-/// 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,
- SmallVector<Instruction*, 4> &ToDelete) {
- MemTransferInst *TheCopy = 0;
- if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false, ToDelete))
- return TheCopy;
- return 0;
-}