#include "llvm/GlobalVariable.h"
#include "llvm/Instructions.h"
#include "llvm/IntrinsicInst.h"
+#include "llvm/LLVMContext.h"
#include "llvm/Pass.h"
#include "llvm/Analysis/Dominators.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Transforms/Utils/PromoteMemToReg.h"
+#include "llvm/Transforms/Utils/Local.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/Compiler.h"
#include "llvm/ADT/SmallVector.h"
/// isUnsafe - This is set to true if the alloca cannot be SROA'd.
bool isUnsafe : 1;
- /// needsCanon - This is set to true if there is some use of the alloca
- /// that requires canonicalization.
- bool needsCanon : 1;
+ /// needsCleanup - This is set to true if there is some use of the alloca
+ /// that requires cleanup.
+ bool needsCleanup : 1;
/// isMemCpySrc - This is true if this aggregate is memcpy'd from.
bool isMemCpySrc : 1;
bool isMemCpyDst : 1;
AllocaInfo()
- : isUnsafe(false), needsCanon(false),
+ : isUnsafe(false), needsCleanup(false),
isMemCpySrc(false), isMemCpyDst(false) {}
};
void DoScalarReplacement(AllocationInst *AI,
std::vector<AllocationInst*> &WorkList);
- void CanonicalizeAllocaUsers(AllocationInst *AI);
+ void CleanupGEP(GetElementPtrInst *GEP);
+ void CleanupAllocaUsers(AllocationInst *AI);
AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base);
void RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI,
void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI,
SmallVector<AllocaInst*, 32> &NewElts);
- bool CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&ResTy,
- uint64_t Offset);
+ bool CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy,
+ bool &SawVec, uint64_t Offset, unsigned AllocaSize);
void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset);
- Value *ConvertUsesOfLoadToScalar(LoadInst *LI, AllocaInst *NewAI,
- uint64_t Offset);
- Value *ConvertUsesOfStoreToScalar(StoreInst *SI, AllocaInst *NewAI,
- uint64_t Offset);
+ Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType,
+ uint64_t Offset, IRBuilder<> &Builder);
+ Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal,
+ uint64_t Offset, IRBuilder<> &Builder);
static Instruction *isOnlyCopiedFromConstantGlobal(AllocationInst *AI);
};
}
if (Allocas.empty()) break;
- PromoteMemToReg(Allocas, DT, DF);
+ PromoteMemToReg(Allocas, DT, DF, Context);
NumPromoted += Allocas.size();
Changed = true;
}
AI->eraseFromParent();
continue;
}
+
+ // If this alloca is impossible for us to promote, reject it early.
+ 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.
+ if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) {
+ DOUT << "Found alloca equal to global: " << *AI;
+ DOUT << " memcpy = " << *TheCopy;
+ Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2));
+ AI->replaceAllUsesWith(
+ Context->getConstantExprBitCast(TheSrc, AI->getType()));
+ TheCopy->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.
- if (!AI->isArrayAllocation() &&
- (isa<StructType>(AI->getAllocatedType()) ||
+ uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType());
+
+ // Do not promote any struct whose size is too big.
+ if (AllocaSize > SRThreshold) continue;
+
+ if ((isa<StructType>(AI->getAllocatedType()) ||
isa<ArrayType>(AI->getAllocatedType())) &&
- AI->getAllocatedType()->isSized() &&
- // Do not promote any struct whose size is larger than "128" bytes.
- TD->getTypePaddedSize(AI->getAllocatedType()) < SRThreshold &&
// Do not promote any struct into more than "32" separate vars.
- getNumSAElements(AI->getAllocatedType()) < SRThreshold/4) {
+ getNumSAElements(AI->getAllocatedType()) <= SRThreshold/4) {
// Check that all of the users of the allocation are capable of being
// transformed.
switch (isSafeAllocaToScalarRepl(AI)) {
- default: assert(0 && "Unexpected value!");
+ default: llvm_unreachable("Unexpected value!");
case 0: // Not safe to scalar replace.
break;
case 1: // Safe, but requires cleanup/canonicalizations first
- CanonicalizeAllocaUsers(AI);
+ CleanupAllocaUsers(AI);
// FALL THROUGH.
case 3: // Safe to scalar replace.
DoScalarReplacement(AI, WorkList);
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.
- if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) {
- DOUT << "Found alloca equal to global: " << *AI;
- DOUT << " memcpy = " << *TheCopy;
- Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2));
- AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType()));
- TheCopy->eraseFromParent(); // Don't mutate the global.
- AI->eraseFromParent();
- ++NumGlobals;
- Changed = true;
- continue;
- }
-
-
// If we can turn this aggregate value (potentially with casts) into a
// simple scalar value that can be mem2reg'd into a register value.
// 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.
bool IsNotTrivial = false;
- const Type *ActualTy = 0;
- if (CanConvertToScalar(AI, IsNotTrivial, ActualTy, 0))
- if (IsNotTrivial && ActualTy &&
- TD->getTypeSizeInBits(ActualTy) < SRThreshold*8) {
- DOUT << "CONVERT TO SCALAR: " << *AI << " TYPE = " << *ActualTy <<"\n";
- ++NumConverted;
+ const Type *VectorTy = 0;
+ bool HadAVector = false;
+ if (CanConvertToScalar(AI, IsNotTrivial, VectorTy, HadAVector,
+ 0, unsigned(AllocaSize)) && IsNotTrivial) {
+ AllocaInst *NewAI;
+ // 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.
+ if (VectorTy && isa<VectorType>(VectorTy) && HadAVector) {
+ DOUT << "CONVERT TO VECTOR: " << *AI << " TYPE = " << *VectorTy <<"\n";
- // Create and insert the alloca.
- AllocaInst *NewAI = new AllocaInst(ActualTy, 0, AI->getName(),
- AI->getParent()->begin());
+ // Create and insert the vector alloca.
+ NewAI = new AllocaInst(VectorTy, 0, "", AI->getParent()->begin());
+ ConvertUsesToScalar(AI, NewAI, 0);
+ } else {
+ DOUT << "CONVERT TO SCALAR INTEGER: " << *AI << "\n";
+
+ // Create and insert the integer alloca.
+ const Type *NewTy = Context->getIntegerType(AllocaSize*8);
+ NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin());
ConvertUsesToScalar(AI, NewAI, 0);
- AI->eraseFromParent();
- Changed = true;
- continue;
}
+ NewAI->takeName(AI);
+ AI->eraseFromParent();
+ ++NumConverted;
+ Changed = true;
+ continue;
+ }
- // Otherwise, couldn't process this.
+ // Otherwise, couldn't process this alloca.
}
return Changed;
// %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1
// (Also works for arrays instead of structs)
if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
- Value *Insert = UndefValue::get(LI->getType());
+ Value *Insert = Context->getUndef(LI->getType());
for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) {
Value *Load = new LoadInst(ElementAllocas[i], "load", LI);
Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI);
// expanded itself once the worklist is rerun.
//
SmallVector<Value*, 8> NewArgs;
- NewArgs.push_back(Constant::getNullValue(Type::Int32Ty));
+ NewArgs.push_back(Context->getNullValue(Type::Int32Ty));
NewArgs.append(GEPI->op_begin()+3, GEPI->op_end());
RepValue = GetElementPtrInst::Create(AllocaToUse, NewArgs.begin(),
NewArgs.end(), "", GEPI);
// The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>".
if (I == E ||
- I.getOperand() != Constant::getNullValue(I.getOperand()->getType())) {
+ I.getOperand() != Context->getNullValue(I.getOperand()->getType())) {
return MarkUnsafe(Info);
}
// out if this is the only problem.
if ((NumElements == 1 || NumElements == 2) &&
AllUsersAreLoads(GEPI)) {
- Info.needsCanon = true;
+ Info.needsCleanup = true;
return; // Canonicalization required!
}
return MarkUnsafe(Info);
// If not the whole aggregate, give up.
if (Length->getZExtValue() !=
- TD->getTypePaddedSize(AI->getType()->getElementType()))
+ TD->getTypeAllocSize(AI->getType()->getElementType()))
return MarkUnsafe(Info);
// We only know about memcpy/memset/memmove.
- if (!isa<MemCpyInst>(MI) && !isa<MemSetInst>(MI) && !isa<MemMoveInst>(MI))
+ if (!isa<MemIntrinsic>(MI))
return MarkUnsafe(Info);
// Otherwise, we can transform it. Determine whether this is a memcpy/set
// cast a {i32,i32}* to i64* and store through it. This is similar to the
// memcpy case and occurs in various "byval" cases and emulated memcpys.
if (isa<IntegerType>(SI->getOperand(0)->getType()) &&
- TD->getTypePaddedSize(SI->getOperand(0)->getType()) ==
- TD->getTypePaddedSize(AI->getType()->getElementType())) {
+ TD->getTypeAllocSize(SI->getOperand(0)->getType()) ==
+ TD->getTypeAllocSize(AI->getType()->getElementType())) {
Info.isMemCpyDst = true;
continue;
}
// cast a {i32,i32}* to i64* and load through it. This is similar to the
// memcpy case and occurs in various "byval" cases and emulated memcpys.
if (isa<IntegerType>(LI->getType()) &&
- TD->getTypePaddedSize(LI->getType()) ==
- TD->getTypePaddedSize(AI->getType()->getElementType())) {
+ TD->getTypeAllocSize(LI->getType()) ==
+ TD->getTypeAllocSize(AI->getType()->getElementType())) {
Info.isMemCpySrc = true;
continue;
}
return MarkUnsafe(Info);
- } else {
+ } else if (isa<DbgInfoIntrinsic>(UI)) {
+ // If one user is DbgInfoIntrinsic then check if all users are
+ // DbgInfoIntrinsics.
+ if (OnlyUsedByDbgInfoIntrinsics(BC)) {
+ Info.needsCleanup = true;
+ return;
+ }
+ else
+ MarkUnsafe(Info);
+ }
+ else {
return MarkUnsafe(Info);
}
if (Info.isUnsafe) return;
SmallVector<AllocaInst*, 32> &NewElts) {
// If this is a memcpy/memmove, construct the other pointer as the
- // appropriate type.
+ // 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
+ // memset, this Value* stays null.
Value *OtherPtr = 0;
- if (MemCpyInst *MCI = dyn_cast<MemCpyInst>(MI)) {
- if (BCInst == MCI->getRawDest())
- OtherPtr = MCI->getRawSource();
- else {
- assert(BCInst == MCI->getRawSource());
- OtherPtr = MCI->getRawDest();
- }
- } else if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
- if (BCInst == MMI->getRawDest())
- OtherPtr = MMI->getRawSource();
+ unsigned MemAlignment = MI->getAlignment();
+ if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy
+ if (BCInst == MTI->getRawDest())
+ OtherPtr = MTI->getRawSource();
else {
- assert(BCInst == MMI->getRawSource());
- OtherPtr = MMI->getRawDest();
+ assert(BCInst == MTI->getRawSource());
+ OtherPtr = MTI->getRawDest();
}
}
const Type *BytePtrTy = MI->getRawDest()->getType();
bool SROADest = MI->getRawDest() == BCInst;
- Constant *Zero = Constant::getNullValue(Type::Int32Ty);
+ Constant *Zero = Context->getNullValue(Type::Int32Ty);
for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
// If this is a memcpy/memmove, emit a GEP of the other element address.
Value *OtherElt = 0;
+ unsigned OtherEltAlign = MemAlignment;
+
if (OtherPtr) {
- Value *Idx[2] = { Zero, ConstantInt::get(Type::Int32Ty, i) };
+ Value *Idx[2] = { Zero, Context->getConstantInt(Type::Int32Ty, i) };
OtherElt = GetElementPtrInst::Create(OtherPtr, Idx, Idx + 2,
OtherPtr->getNameStr()+"."+utostr(i),
MI);
+ uint64_t EltOffset;
+ const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType());
+ if (const StructType *ST =
+ dyn_cast<StructType>(OtherPtrTy->getElementType())) {
+ EltOffset = TD->getStructLayout(ST)->getElementOffset(i);
+ } else {
+ const Type *EltTy =
+ cast<SequentialType>(OtherPtr->getType())->getElementType();
+ EltOffset = TD->getTypeAllocSize(EltTy)*i;
+ }
+
+ // The alignment of the other pointer is the guaranteed alignment of the
+ // element, which is affected by both the known alignment of the whole
+ // mem intrinsic and the alignment of the element. If the alignment of
+ // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the
+ // known alignment is just 4 bytes.
+ OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset);
}
Value *EltPtr = NewElts[i];
- const Type *EltTy =cast<PointerType>(EltPtr->getType())->getElementType();
+ const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType();
// If we got down to a scalar, insert a load or store as appropriate.
if (EltTy->isSingleValueType()) {
- if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
- Value *Elt = new LoadInst(SROADest ? OtherElt : EltPtr, "tmp",
- MI);
- new StoreInst(Elt, SROADest ? EltPtr : OtherElt, MI);
+ if (isa<MemTransferInst>(MI)) {
+ if (SROADest) {
+ // From Other to Alloca.
+ Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI);
+ new StoreInst(Elt, EltPtr, MI);
+ } else {
+ // From Alloca to Other.
+ Value *Elt = new LoadInst(EltPtr, "tmp", MI);
+ new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI);
+ }
continue;
}
assert(isa<MemSetInst>(MI));
Constant *StoreVal;
if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) {
if (CI->isZero()) {
- StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0>
+ StoreVal = Context->getNullValue(EltTy); // 0.0, null, 0, <0,0>
} else {
// If EltTy is a vector type, get the element type.
- const Type *ValTy = EltTy;
- if (const VectorType *VTy = dyn_cast<VectorType>(ValTy))
- ValTy = VTy->getElementType();
-
+ const Type *ValTy = EltTy->getScalarType();
+
// Construct an integer with the right value.
unsigned EltSize = TD->getTypeSizeInBits(ValTy);
APInt OneVal(EltSize, CI->getZExtValue());
}
// Convert the integer value to the appropriate type.
- StoreVal = ConstantInt::get(TotalVal);
+ StoreVal = Context->getConstantInt(TotalVal);
if (isa<PointerType>(ValTy))
- StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy);
+ StoreVal = Context->getConstantExprIntToPtr(StoreVal, ValTy);
else if (ValTy->isFloatingPoint())
- StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy);
+ StoreVal = Context->getConstantExprBitCast(StoreVal, 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();
SmallVector<Constant*, 16> Elts(NumElts, StoreVal);
- StoreVal = ConstantVector::get(&Elts[0], NumElts);
+ StoreVal = Context->getConstantVector(&Elts[0], NumElts);
}
}
new StoreInst(StoreVal, EltPtr, MI);
OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(),
MI);
- unsigned EltSize = TD->getTypePaddedSize(EltTy);
+ unsigned EltSize = TD->getTypeAllocSize(EltTy);
// Finally, insert the meminst for this element.
- if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
+ if (isa<MemTransferInst>(MI)) {
Value *Ops[] = {
SROADest ? EltPtr : OtherElt, // Dest ptr
SROADest ? OtherElt : EltPtr, // Src ptr
- ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
- Zero // Align
+ Context->getConstantInt(MI->getOperand(3)->getType(), EltSize), // Size
+ Context->getConstantInt(Type::Int32Ty, OtherEltAlign) // Align
};
CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
} else {
assert(isa<MemSetInst>(MI));
Value *Ops[] = {
EltPtr, MI->getOperand(2), // Dest, Value,
- ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
+ Context->getConstantInt(MI->getOperand(3)->getType(), EltSize), // Size
Zero // Align
};
CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
// and store the element value to the individual alloca.
Value *SrcVal = SI->getOperand(0);
const Type *AllocaEltTy = AI->getType()->getElementType();
- uint64_t AllocaSizeBits = TD->getTypePaddedSizeInBits(AllocaEltTy);
+ uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
// If this isn't a store of an integer to the whole alloca, it may be a store
// to the first element. Just ignore the store in this case and normal SROA
// will handle it.
if (!isa<IntegerType>(SrcVal->getType()) ||
- TD->getTypePaddedSizeInBits(SrcVal->getType()) != AllocaSizeBits)
+ TD->getTypeAllocSizeInBits(SrcVal->getType()) != AllocaSizeBits)
return;
+ // Handle tail padding by extending the operand
+ if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits)
+ SrcVal = new ZExtInst(SrcVal,
+ Context->getIntegerType(AllocaSizeBits), "", SI);
DOUT << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << *SI;
uint64_t Shift = Layout->getElementOffsetInBits(i);
if (TD->isBigEndian())
- Shift = AllocaSizeBits-Shift-TD->getTypePaddedSizeInBits(FieldTy);
+ Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy);
Value *EltVal = SrcVal;
if (Shift) {
- Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
+ Value *ShiftVal = Context->getConstantInt(EltVal->getType(), Shift);
EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
"sroa.store.elt", SI);
}
if (FieldSizeBits == 0) continue;
if (FieldSizeBits != AllocaSizeBits)
- EltVal = new TruncInst(EltVal, IntegerType::get(FieldSizeBits), "", SI);
+ EltVal = new TruncInst(EltVal,
+ Context->getIntegerType(FieldSizeBits), "", SI);
Value *DestField = NewElts[i];
if (EltVal->getType() == FieldTy) {
// Storing to an integer field of this size, just do it.
} else {
// Otherwise, bitcast the dest pointer (for aggregates).
DestField = new BitCastInst(DestField,
- PointerType::getUnqual(EltVal->getType()),
+ Context->getPointerTypeUnqual(EltVal->getType()),
"", SI);
}
new StoreInst(EltVal, DestField, SI);
} else {
const ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
const Type *ArrayEltTy = ATy->getElementType();
- uint64_t ElementOffset = TD->getTypePaddedSizeInBits(ArrayEltTy);
+ uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy);
uint64_t Shift;
Value *EltVal = SrcVal;
if (Shift) {
- Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
+ Value *ShiftVal = Context->getConstantInt(EltVal->getType(), Shift);
EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
"sroa.store.elt", SI);
}
// Truncate down to an integer of the right size.
if (ElementSizeBits != AllocaSizeBits)
- EltVal = new TruncInst(EltVal, IntegerType::get(ElementSizeBits),"",SI);
+ EltVal = new TruncInst(EltVal,
+ Context->getIntegerType(ElementSizeBits),"",SI);
Value *DestField = NewElts[i];
if (EltVal->getType() == ArrayEltTy) {
// Storing to an integer field of this size, just do it.
} else {
// Otherwise, bitcast the dest pointer (for aggregates).
DestField = new BitCastInst(DestField,
- PointerType::getUnqual(EltVal->getType()),
+ Context->getPointerTypeUnqual(EltVal->getType()),
"", SI);
}
new StoreInst(EltVal, DestField, SI);
// Extract each element out of the NewElts according to its structure offset
// and form the result value.
const Type *AllocaEltTy = AI->getType()->getElementType();
- uint64_t AllocaSizeBits = TD->getTypePaddedSizeInBits(AllocaEltTy);
+ uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
// If this isn't a load of the whole alloca to an integer, it may be a load
// of the first element. Just ignore the load in this case and normal SROA
// will handle it.
if (!isa<IntegerType>(LI->getType()) ||
- TD->getTypePaddedSizeInBits(LI->getType()) != AllocaSizeBits)
+ TD->getTypeAllocSizeInBits(LI->getType()) != AllocaSizeBits)
return;
DOUT << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << *LI;
Layout = TD->getStructLayout(EltSTy);
} else {
const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
- ArrayEltBitOffset = TD->getTypePaddedSizeInBits(ArrayEltTy);
+ ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
}
- Value *ResultVal = Constant::getNullValue(LI->getType());
+ Value *ResultVal =
+ Context->getNullValue(Context->getIntegerType(AllocaSizeBits));
for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
// Load the value from the alloca. If the NewElt is an aggregate, cast
// Ignore zero sized fields like {}, they obviously contain no data.
if (FieldSizeBits == 0) continue;
- const IntegerType *FieldIntTy = IntegerType::get(FieldSizeBits);
+ const IntegerType *FieldIntTy = Context->getIntegerType(FieldSizeBits);
if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() &&
!isa<VectorType>(FieldTy))
- SrcField = new BitCastInst(SrcField, PointerType::getUnqual(FieldIntTy),
+ SrcField = new BitCastInst(SrcField,
+ Context->getPointerTypeUnqual(FieldIntTy),
"", LI);
SrcField = new LoadInst(SrcField, "sroa.load.elt", LI);
Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth();
if (Shift) {
- Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift);
+ Value *ShiftVal = Context->getConstantInt(SrcField->getType(), Shift);
SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI);
}
ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI);
}
-
+
+ // Handle tail padding by truncating the result
+ if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits)
+ ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI);
+
LI->replaceAllUsesWith(ResultVal);
LI->eraseFromParent();
}
} else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
return HasPadding(VTy->getElementType(), TD);
}
- return TD.getTypeSizeInBits(Ty) != TD.getTypePaddedSizeInBits(Ty);
+ return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty);
}
/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
return 0;
// If we require cleanup, return 1, otherwise return 3.
- return Info.needsCanon ? 1 : 3;
+ return Info.needsCleanup ? 1 : 3;
}
-/// CanonicalizeAllocaUsers - If SROA reported that it can promote the specified
+/// CleanupGEP - GEP is used by an Alloca, which can be prompted after the GEP
+/// is canonicalized here.
+void SROA::CleanupGEP(GetElementPtrInst *GEPI) {
+ gep_type_iterator I = gep_type_begin(GEPI);
+ ++I;
+
+ const ArrayType *AT = dyn_cast<ArrayType>(*I);
+ if (!AT)
+ return;
+
+ uint64_t NumElements = AT->getNumElements();
+
+ if (isa<ConstantInt>(I.getOperand()))
+ return;
+
+ if (NumElements == 1) {
+ GEPI->setOperand(2, Context->getNullValue(Type::Int32Ty));
+ return;
+ }
+
+ assert(NumElements == 2 && "Unhandled case!");
+ // All users of the GEP must be loads. At each use of the GEP, insert
+ // two loads of the appropriate indexed GEP and select between them.
+ Value *IsOne = new ICmpInst(GEPI, ICmpInst::ICMP_NE, I.getOperand(),
+ Context->getNullValue(I.getOperand()->getType()),
+ "isone");
+ // Insert the new GEP instructions, which are properly indexed.
+ SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end());
+ Indices[1] = Context->getNullValue(Type::Int32Ty);
+ Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0),
+ Indices.begin(),
+ Indices.end(),
+ GEPI->getName()+".0", GEPI);
+ Indices[1] = Context->getConstantInt(Type::Int32Ty, 1);
+ Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0),
+ Indices.begin(),
+ Indices.end(),
+ GEPI->getName()+".1", GEPI);
+ // Replace all loads of the variable index GEP with loads from both
+ // indexes and a select.
+ while (!GEPI->use_empty()) {
+ LoadInst *LI = cast<LoadInst>(GEPI->use_back());
+ Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI);
+ Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI);
+ Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI);
+ LI->replaceAllUsesWith(R);
+ LI->eraseFromParent();
+ }
+ GEPI->eraseFromParent();
+}
+
+
+/// CleanupAllocaUsers - If SROA reported that it can promote the specified
/// allocation, but only if cleaned up, perform the cleanups required.
-void SROA::CanonicalizeAllocaUsers(AllocationInst *AI) {
+void SROA::CleanupAllocaUsers(AllocationInst *AI) {
// At this point, we know that the end result will be SROA'd and promoted, so
// we can insert ugly code if required so long as sroa+mem2reg will clean it
// up.
for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
UI != E; ) {
- GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI++);
- if (!GEPI) continue;
- gep_type_iterator I = gep_type_begin(GEPI);
- ++I;
-
- if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
- uint64_t NumElements = AT->getNumElements();
-
- if (!isa<ConstantInt>(I.getOperand())) {
- if (NumElements == 1) {
- GEPI->setOperand(2, Constant::getNullValue(Type::Int32Ty));
- } else {
- assert(NumElements == 2 && "Unhandled case!");
- // All users of the GEP must be loads. At each use of the GEP, insert
- // two loads of the appropriate indexed GEP and select between them.
- Value *IsOne = new ICmpInst(ICmpInst::ICMP_NE, I.getOperand(),
- Constant::getNullValue(I.getOperand()->getType()),
- "isone", GEPI);
- // Insert the new GEP instructions, which are properly indexed.
- SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end());
- Indices[1] = Constant::getNullValue(Type::Int32Ty);
- Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0),
- Indices.begin(),
- Indices.end(),
- GEPI->getName()+".0", GEPI);
- Indices[1] = ConstantInt::get(Type::Int32Ty, 1);
- Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0),
- Indices.begin(),
- Indices.end(),
- GEPI->getName()+".1", GEPI);
- // Replace all loads of the variable index GEP with loads from both
- // indexes and a select.
- while (!GEPI->use_empty()) {
- LoadInst *LI = cast<LoadInst>(GEPI->use_back());
- Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI);
- Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI);
- Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI);
- LI->replaceAllUsesWith(R);
- LI->eraseFromParent();
- }
- GEPI->eraseFromParent();
+ User *U = *UI++;
+ if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U))
+ CleanupGEP(GEPI);
+ else {
+ Instruction *I = cast<Instruction>(U);
+ SmallVector<DbgInfoIntrinsic *, 2> DbgInUses;
+ if (!isa<StoreInst>(I) && OnlyUsedByDbgInfoIntrinsics(I, &DbgInUses)) {
+ // Safe to remove debug info uses.
+ while (!DbgInUses.empty()) {
+ DbgInfoIntrinsic *DI = DbgInUses.back(); DbgInUses.pop_back();
+ DI->eraseFromParent();
}
+ I->eraseFromParent();
}
}
}
/// 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.
-static void MergeInType(const Type *In, uint64_t Offset, const Type *&Accum,
- const TargetData &TD) {
- // If this is our first type, just use it.
- if ((Accum == 0 && Offset == 0) || In == Type::VoidTy ||
- // Or if this is a same type, keep it.
- (In == Accum && Offset == 0)) {
- Accum = In;
- return;
- }
-
- // Merging something like i32 into offset 8 means that a "field" is merged in
- // before the basic type is. Make sure to consider the offset below.
- if (Accum == 0)
- Accum = Type::Int8Ty;
-
- if (const VectorType *VATy = dyn_cast<VectorType>(Accum)) {
- if (VATy->getElementType() == In &&
- Offset % TD.getTypePaddedSize(In) == 0 &&
- Offset < TD.getTypePaddedSize(VATy))
- return; // Accum is a vector, and we are accessing an element: ok.
- if (const VectorType *VInTy = dyn_cast<VectorType>(In))
- if (VInTy->getBitWidth() == VATy->getBitWidth() && Offset == 0)
- return; // Two vectors of the same size: keep either one of them.
- }
-
- if (const VectorType *VInTy = dyn_cast<VectorType>(In)) {
- // In is a vector, and we are accessing an element: keep V.
- if (VInTy->getElementType() == Accum &&
- Offset % TD.getTypePaddedSize(Accum) == 0 &&
- Offset < TD.getTypePaddedSize(VInTy)) {
- Accum = VInTy;
- return;
+static void MergeInType(const Type *In, uint64_t Offset, const Type *&VecTy,
+ unsigned AllocaSize, const TargetData &TD,
+ LLVMContext *Context) {
+ // If this could be contributing to a vector, analyze it.
+ if (VecTy != Type::VoidTy) { // either null or a vector type.
+
+ // 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)) {
+ 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 (VecTy == 0)
+ VecTy = VInTy;
+ return;
+ }
+ } else if (In == Type::FloatTy || In == Type::DoubleTy ||
+ (isa<IntegerType>(In) && In->getPrimitiveSizeInBits() >= 8 &&
+ isPowerOf2_32(In->getPrimitiveSizeInBits()))) {
+ // 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 &&
+ (VecTy == 0 ||
+ cast<VectorType>(VecTy)->getElementType()
+ ->getPrimitiveSizeInBits()/8 == EltSize)) {
+ if (VecTy == 0)
+ VecTy = Context->getVectorType(In, AllocaSize/EltSize);
+ return;
+ }
}
}
- // Otherwise, we have a case that we can't handle with an optimized form.
- // Convert the alloca to an integer that is as large as the largest store size
- // of the value values.
- uint64_t InSize = TD.getTypeStoreSizeInBits(In)+8*Offset;
- uint64_t ASize = TD.getTypeStoreSizeInBits(Accum);
- if (InSize > ASize) ASize = InSize;
- Accum = IntegerType::get(ASize);
+ // Otherwise, we have a case that we can't handle with an optimized vector
+ // form. We can still turn this into a large integer.
+ VecTy = Type::VoidTy;
}
/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all
-/// its accesses to use a to single scalar type, return true, and set ResTy to
-/// the new type. Further, if the use is not a completely trivial use that
-/// mem2reg could promote, set IsNotTrivial. Offset is the current offset from
-/// the base of the alloca being analyzed.
+/// its accesses to use a to single vector type, return true, and set VecTy to
+/// the new type. If we could convert the alloca into a single promotable
+/// integer, return true but set VecTy to VoidTy. Further, if the use is not a
+/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset
+/// is the current offset from the base of the alloca being analyzed.
+///
+/// If we see at least one access to the value that is as a vector type, set the
+/// SawVec flag.
///
-bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial,
- const Type *&ResTy, uint64_t Offset) {
+bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy,
+ bool &SawVec, uint64_t Offset,
+ unsigned AllocaSize) {
for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
Instruction *User = cast<Instruction>(*UI);
// Don't break volatile loads.
if (LI->isVolatile())
return false;
- MergeInType(LI->getType(), Offset, ResTy, *TD);
+ MergeInType(LI->getType(), Offset, VecTy, AllocaSize, *TD, Context);
+ SawVec |= isa<VectorType>(LI->getType());
continue;
}
if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
// Storing the pointer, not into the value?
if (SI->getOperand(0) == V || SI->isVolatile()) return 0;
- MergeInType(SI->getOperand(0)->getType(), Offset, ResTy, *TD);
+ MergeInType(SI->getOperand(0)->getType(), Offset,
+ VecTy, AllocaSize, *TD, Context);
+ SawVec |= isa<VectorType>(SI->getOperand(0)->getType());
continue;
}
if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
- if (!CanConvertToScalar(BCI, IsNotTrivial, ResTy, Offset))
+ if (!CanConvertToScalar(BCI, IsNotTrivial, VecTy, SawVec, Offset,
+ AllocaSize))
return false;
IsNotTrivial = true;
continue;
uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(),
&Indices[0], Indices.size());
// See if all uses can be converted.
- if (!CanConvertToScalar(GEP, IsNotTrivial, ResTy, Offset+GEPOffset))
+ if (!CanConvertToScalar(GEP, IsNotTrivial, VecTy, SawVec,Offset+GEPOffset,
+ AllocaSize))
return false;
IsNotTrivial = 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())) {
+ IsNotTrivial = true;
+ continue;
+ }
+ }
+
+ // 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)) {
+ if (ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()))
+ if (Len->getZExtValue() == AllocaSize && Offset == 0) {
+ IsNotTrivial = true;
+ continue;
+ }
+ }
+ // Ignore dbg intrinsic.
+ if (isa<DbgInfoIntrinsic>(User))
+ continue;
+
// Otherwise, we cannot handle this!
return false;
}
while (!Ptr->use_empty()) {
Instruction *User = cast<Instruction>(Ptr->use_back());
- if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
- LI->replaceAllUsesWith(ConvertUsesOfLoadToScalar(LI, NewAI, Offset));
- LI->eraseFromParent();
- continue;
- }
-
- if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
- assert(SI->getOperand(0) != Ptr && "Consistency error!");
- new StoreInst(ConvertUsesOfStoreToScalar(SI, NewAI, Offset), NewAI, SI);
- SI->eraseFromParent();
- continue;
- }
-
if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
ConvertUsesToScalar(CI, NewAI, Offset);
CI->eraseFromParent();
GEP->eraseFromParent();
continue;
}
- assert(0 && "Unsupported operation!");
- abort();
+
+ IRBuilder<> Builder(User->getParent(), User);
+
+ 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 *NewLoadVal
+ = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder);
+ LI->replaceAllUsesWith(NewLoadVal);
+ LI->eraseFromParent();
+ continue;
+ }
+
+ if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
+ assert(SI->getOperand(0) != Ptr && "Consistency error!");
+ Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").c_str());
+ Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset,
+ Builder);
+ Builder.CreateStore(New, NewAI);
+ SI->eraseFromParent();
+ continue;
+ }
+
+ // If this is a constant sized memset of a constant value (e.g. 0) we can
+ // 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) {
+ unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue();
+
+ // Compute the value replicated the right number of times.
+ APInt APVal(NumBytes*8, Val);
+
+ // Splat the value if non-zero.
+ if (Val)
+ for (unsigned i = 1; i != NumBytes; ++i)
+ APVal |= APVal << 8;
+
+ Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").c_str());
+ Value *New = ConvertScalar_InsertValue(Context->getConstantInt(APVal),
+ Old, Offset, Builder);
+ Builder.CreateStore(New, NewAI);
+ }
+ MSI->eraseFromParent();
+ continue;
+ }
+
+ // 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)) {
+ assert(Offset == 0 && "must be store to start of alloca");
+
+ // 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>(Ptr->getUnderlyingObject());
+
+ if (MTI->getSource()->getUnderlyingObject() != 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();
+ SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType());
+
+ LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval");
+ SrcVal->setAlignment(MTI->getAlignment());
+ Builder.CreateStore(SrcVal, NewAI);
+ } else if (MTI->getDest()->getUnderlyingObject() != 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");
+
+ Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType());
+ StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr);
+ NewStore->setAlignment(MTI->getAlignment());
+ } else {
+ // Noop transfer. Src == Dst
+ }
+
+
+ MTI->eraseFromParent();
+ continue;
+ }
+
+ // If user is a dbg info intrinsic then it is safe to remove it.
+ if (isa<DbgInfoIntrinsic>(User)) {
+ User->eraseFromParent();
+ continue;
+ }
+
+ llvm_unreachable("Unsupported operation!");
}
}
-/// ConvertUsesOfLoadToScalar - Convert all of the users of the specified load
-/// to use the new alloca directly, returning the value that should replace the
-/// load. This happens when we are converting an "integer union" to a single
+/// 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.
+///
+/// This happens when we are converting an "integer union" to a single
/// integer scalar, or when we are converting a "vector union" to a vector with
/// insert/extractelement instructions.
///
/// 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.
-Value *SROA::ConvertUsesOfLoadToScalar(LoadInst *LI, AllocaInst *NewAI,
- uint64_t Offset) {
- // The load is a bit extract from NewAI shifted right by Offset bits.
- Value *NV = new LoadInst(NewAI, LI->getName(), LI);
-
+/// shifted to the right.
+Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType,
+ uint64_t Offset, IRBuilder<> &Builder) {
// If the load is of the whole new alloca, no conversion is needed.
- if (NV->getType() == LI->getType() && Offset == 0)
- return NV;
+ if (FromVal->getType() == 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>(NV->getType())) {
- if (isa<VectorType>(LI->getType()))
- return new BitCastInst(NV, LI->getType(), LI->getName(), LI);
+ if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) {
+ if (isa<VectorType>(ToType))
+ return Builder.CreateBitCast(FromVal, ToType, "tmp");
// Otherwise it must be an element access.
unsigned Elt = 0;
if (Offset) {
- unsigned EltSize = TD->getTypePaddedSizeInBits(VTy->getElementType());
+ unsigned EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType());
Elt = Offset/EltSize;
assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
}
// Return the element extracted out of it.
- return new ExtractElementInst(NV, ConstantInt::get(Type::Int32Ty, Elt),
- "tmp", LI);
+ Value *V = Builder.CreateExtractElement(FromVal,
+ Context->getConstantInt(Type::Int32Ty,Elt),
+ "tmp");
+ if (V->getType() != ToType)
+ V = Builder.CreateBitCast(V, ToType, "tmp");
+ 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)) {
+ const StructLayout &Layout = *TD->getStructLayout(ST);
+ Value *Res = Context->getUndef(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");
+ }
+ return Res;
+ }
+
+ if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) {
+ uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType());
+ Value *Res = Context->getUndef(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");
+ }
+ return Res;
}
// Otherwise, this must be a union that was converted to an integer value.
- const IntegerType *NTy = cast<IntegerType>(NV->getType());
+ const 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.
// 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(LI->getType()) - Offset;
+ TD->getTypeStoreSizeInBits(ToType) - Offset;
} else {
ShAmt = Offset;
}
// We do this to support (f.e.) loads off the end of a structure where
// only some bits are used.
if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth())
- NV = BinaryOperator::CreateLShr(NV,
- ConstantInt::get(NV->getType(), ShAmt),
- LI->getName(), LI);
+ FromVal = Builder.CreateLShr(FromVal,
+ Context->getConstantInt(FromVal->getType(),
+ ShAmt), "tmp");
else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
- NV = BinaryOperator::CreateShl(NV,
- ConstantInt::get(NV->getType(), -ShAmt),
- LI->getName(), LI);
+ FromVal = Builder.CreateShl(FromVal,
+ Context->getConstantInt(FromVal->getType(),
+ -ShAmt), "tmp");
// Finally, unconditionally truncate the integer to the right width.
- unsigned LIBitWidth = TD->getTypeSizeInBits(LI->getType());
+ unsigned LIBitWidth = TD->getTypeSizeInBits(ToType);
if (LIBitWidth < NTy->getBitWidth())
- NV = new TruncInst(NV, IntegerType::get(LIBitWidth),
- LI->getName(), LI);
+ FromVal =
+ Builder.CreateTrunc(FromVal, Context->getIntegerType(LIBitWidth), "tmp");
+ else if (LIBitWidth > NTy->getBitWidth())
+ FromVal =
+ Builder.CreateZExt(FromVal, Context->getIntegerType(LIBitWidth), "tmp");
// If the result is an integer, this is a trunc or bitcast.
- if (isa<IntegerType>(LI->getType())) {
+ if (isa<IntegerType>(ToType)) {
// Should be done.
- } else if (LI->getType()->isFloatingPoint() ||
- isa<VectorType>(LI->getType())) {
+ } else if (ToType->isFloatingPoint() || isa<VectorType>(ToType)) {
// Just do a bitcast, we know the sizes match up.
- NV = new BitCastInst(NV, LI->getType(), LI->getName(), LI);
+ FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp");
} else {
// Otherwise must be a pointer.
- NV = new IntToPtrInst(NV, LI->getType(), LI->getName(), LI);
+ FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp");
}
- assert(NV->getType() == LI->getType() && "Didn't convert right?");
- return NV;
+ assert(FromVal->getType() == ToType && "Didn't convert right?");
+ return FromVal;
}
-/// ConvertUsesOfStoreToScalar - Convert the specified store to a load+store
-/// pair of the new alloca directly, returning the value that should be stored
-/// to the alloca. This happens when we are converting an "integer union" to a
+/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer
+/// or vector value "Old" at the offset specified by Offset.
+///
+/// This happens when we are converting an "integer union" to a
/// single integer scalar, or when we are converting a "vector union" to a
/// vector with insert/extractelement instructions.
///
/// 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.
-Value *SROA::ConvertUsesOfStoreToScalar(StoreInst *SI, AllocaInst *NewAI,
- uint64_t Offset) {
+/// shifted to the right.
+Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old,
+ uint64_t Offset, IRBuilder<> &Builder) {
// Convert the stored type to the actual type, shift it left to insert
// then 'or' into place.
- Value *SV = SI->getOperand(0);
- const Type *AllocaType = NewAI->getType()->getElementType();
- if (SV->getType() == AllocaType && Offset == 0) {
- return SV;
- }
+ const Type *AllocaType = Old->getType();
if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
- Value *Old = new LoadInst(NewAI, NewAI->getName()+".in", SI);
+ 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");
- // If the result alloca is a vector type, this is either an element
- // access or a bitcast to another vector type.
- if (isa<VectorType>(SV->getType())) {
- SV = new BitCastInst(SV, AllocaType, SV->getName(), SI);
- } else {
- // Must be an element insertion.
- unsigned Elt = Offset/TD->getTypePaddedSizeInBits(VTy->getElementType());
- SV = InsertElementInst::Create(Old, SV,
- ConstantInt::get(Type::Int32Ty, Elt),
- "tmp", SI);
- }
+ uint64_t EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType());
+
+ // Must be an element insertion.
+ unsigned Elt = Offset/EltSize;
+
+ if (SV->getType() != VTy->getElementType())
+ SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp");
+
+ SV = Builder.CreateInsertElement(Old, SV,
+ Context->getConstantInt(Type::Int32Ty, Elt),
+ "tmp");
return SV;
}
-
-
- Value *Old = new LoadInst(NewAI, NewAI->getName()+".in", SI);
+
+ // If SV is a first-class aggregate value, insert each value recursively.
+ if (const 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");
+ Old = ConvertScalar_InsertValue(Elt, Old,
+ Offset+Layout.getElementOffsetInBits(i),
+ Builder);
+ }
+ return Old;
+ }
+
+ if (const 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");
+ Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder);
+ }
+ return Old;
+ }
// If SV is a float, convert it to the appropriate integer type.
- // If it is a pointer, do the same, and also handle ptr->ptr casts
- // here.
+ // 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);
if (SV->getType()->isFloatingPoint() || isa<VectorType>(SV->getType()))
- SV = new BitCastInst(SV, IntegerType::get(SrcWidth), SV->getName(), SI);
+ SV = Builder.CreateBitCast(SV, Context->getIntegerType(SrcWidth), "tmp");
else if (isa<PointerType>(SV->getType()))
- SV = new PtrToIntInst(SV, TD->getIntPtrType(), SV->getName(), SI);
+ SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(), "tmp");
- // Always zero extend the value if needed.
- if (SV->getType() != AllocaType)
- SV = new ZExtInst(SV, AllocaType, SV->getName(), SI);
+ // Zero extend or truncate the value if needed.
+ if (SV->getType() != AllocaType) {
+ if (SV->getType()->getPrimitiveSizeInBits() <
+ AllocaType->getPrimitiveSizeInBits())
+ SV = Builder.CreateZExt(SV, AllocaType, "tmp");
+ else {
+ // Truncation may be needed if storing more than the alloca can hold
+ // (undefined behavior).
+ SV = Builder.CreateTrunc(SV, AllocaType, "tmp");
+ SrcWidth = DestWidth;
+ SrcStoreWidth = DestStoreWidth;
+ }
+ }
// 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.
// only some bits in the structure are set.
APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
- SV = BinaryOperator::CreateShl(SV,
- ConstantInt::get(SV->getType(), ShAmt),
- SV->getName(), SI);
+ SV = Builder.CreateShl(SV, Context->getConstantInt(SV->getType(),
+ ShAmt), "tmp");
Mask <<= ShAmt;
} else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
- SV = BinaryOperator::CreateLShr(SV,
- ConstantInt::get(SV->getType(),-ShAmt),
- SV->getName(), SI);
+ SV = Builder.CreateLShr(SV, Context->getConstantInt(SV->getType(),
+ -ShAmt), "tmp");
Mask = Mask.lshr(-ShAmt);
}
// in the new bits.
if (SrcWidth != DestWidth) {
assert(DestWidth > SrcWidth);
- Old = BinaryOperator::CreateAnd(Old, ConstantInt::get(~Mask),
- Old->getName()+".mask", SI);
- SV = BinaryOperator::CreateOr(Old, SV, SV->getName()+".ins", SI);
+ Old = Builder.CreateAnd(Old, Context->getConstantInt(~Mask), "mask");
+ SV = Builder.CreateOr(Old, SV, "ins");
}
return SV;
}
// If this is isn't our memcpy/memmove, reject it as something we can't
// handle.
- if (!isa<MemCpyInst>(*UI) && !isa<MemMoveInst>(*UI))
+ if (!isa<MemTransferInst>(*UI))
return false;
// If we already have seen a copy, reject the second one.