//
// The LLVM Compiler Infrastructure
//
-// This file was developed by the LLVM research group and is distributed under
-// the University of Illinois Open Source License. See LICENSE.TXT for details.
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
#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/Pass.h"
STATISTIC(NumReplaced, "Number of allocas broken up");
STATISTIC(NumPromoted, "Number of allocas promoted");
STATISTIC(NumConverted, "Number of aggregates converted to scalar");
+STATISTIC(NumGlobals, "Number of allocas copied from constant global");
namespace {
struct VISIBILITY_HIDDEN SROA : public FunctionPass {
+ static char ID; // Pass identification, replacement for typeid
+ explicit SROA(signed T = -1) : FunctionPass((intptr_t)&ID) {
+ if (T == -1)
+ SRThreshold = 128;
+ else
+ SRThreshold = T;
+ }
+
bool runOnFunction(Function &F);
bool performScalarRepl(Function &F);
}
private:
- int isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI);
- int isSafeUseOfAllocation(Instruction *User, AllocationInst *AI);
- bool isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI);
- bool isSafeUseOfBitCastedAllocation(BitCastInst *User, AllocationInst *AI);
+ /// AllocaInfo - When analyzing uses of an alloca instruction, this captures
+ /// information about the uses. All these fields are initialized to false
+ /// and set to true when something is learned.
+ struct AllocaInfo {
+ /// 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;
+
+ /// isMemCpySrc - This is true if this aggregate is memcpy'd from.
+ bool isMemCpySrc : 1;
+
+ /// isMemCpyDst - This is true if this aggregate is memcpy'd into.
+ bool isMemCpyDst : 1;
+
+ AllocaInfo()
+ : isUnsafe(false), needsCanon(false),
+ isMemCpySrc(false), isMemCpyDst(false) {}
+ };
+
+ unsigned SRThreshold;
+
+ void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; }
+
int isSafeAllocaToScalarRepl(AllocationInst *AI);
+
+ void isSafeUseOfAllocation(Instruction *User, AllocationInst *AI,
+ AllocaInfo &Info);
+ void isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI,
+ AllocaInfo &Info);
+ void isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI,
+ unsigned OpNo, AllocaInfo &Info);
+ void isSafeUseOfBitCastedAllocation(BitCastInst *User, AllocationInst *AI,
+ AllocaInfo &Info);
+
+ void DoScalarReplacement(AllocationInst *AI,
+ std::vector<AllocationInst*> &WorkList);
void CanonicalizeAllocaUsers(AllocationInst *AI);
AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base);
const Type *CanConvertToScalar(Value *V, bool &IsNotTrivial);
void ConvertToScalar(AllocationInst *AI, const Type *Ty);
void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, unsigned Offset);
+ static Instruction *isOnlyCopiedFromConstantGlobal(AllocationInst *AI);
};
+ char SROA::ID = 0;
RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates");
}
// Public interface to the ScalarReplAggregates pass
-FunctionPass *llvm::createScalarReplAggregatesPass() { return new SROA(); }
+FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) {
+ return new SROA(Threshold);
+}
bool SROA::runOnFunction(Function &F) {
bool SROA::performPromotion(Function &F) {
std::vector<AllocaInst*> Allocas;
- const TargetData &TD = getAnalysis<TargetData>();
- DominatorTree &DT = getAnalysis<DominatorTree>();
+ DominatorTree &DT = getAnalysis<DominatorTree>();
DominanceFrontier &DF = getAnalysis<DominanceFrontier>();
BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
// the entry node
for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
- if (isAllocaPromotable(AI, TD))
+ if (isAllocaPromotable(AI))
Allocas.push_back(AI);
if (Allocas.empty()) break;
- PromoteMemToReg(Allocas, DT, DF, TD);
+ PromoteMemToReg(Allocas, DT, DF);
NumPromoted += Allocas.size();
Changed = true;
}
if (AllocationInst *A = dyn_cast<AllocationInst>(I))
WorkList.push_back(A);
+ const TargetData &TD = getAnalysis<TargetData>();
+
// Process the worklist
bool Changed = false;
while (!WorkList.empty()) {
continue;
}
- // 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()) &&
- !isa<ArrayType>(AI->getAllocatedType()))) continue;
-
- // Check that all of the users of the allocation are capable of being
- // transformed.
- switch (isSafeAllocaToScalarRepl(AI)) {
- default: assert(0 && "Unexpected value!");
- case 0: // Not safe to scalar replace.
+ // 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()) ||
+ isa<ArrayType>(AI->getAllocatedType())) &&
+ AI->getAllocatedType()->isSized() &&
+ TD.getABITypeSize(AI->getAllocatedType()) < SRThreshold) {
+ // Check that all of the users of the allocation are capable of being
+ // transformed.
+ switch (isSafeAllocaToScalarRepl(AI)) {
+ default: assert(0 && "Unexpected value!");
+ case 0: // Not safe to scalar replace.
+ break;
+ case 1: // Safe, but requires cleanup/canonicalizations first
+ CanonicalizeAllocaUsers(AI);
+ // FALL THROUGH.
+ case 3: // Safe to scalar replace.
+ DoScalarReplacement(AI, WorkList);
+ Changed = true;
+ 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;
- case 1: // Safe, but requires cleanup/canonicalizations first
- CanonicalizeAllocaUsers(AI);
- case 3: // Safe to scalar replace.
- break;
}
+
+ // Otherwise, couldn't process this.
+ }
- DOUT << "Found inst to xform: " << *AI;
- Changed = true;
+ return Changed;
+}
- SmallVector<AllocaInst*, 32> ElementAllocas;
- if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
- ElementAllocas.reserve(ST->getNumContainedTypes());
- for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
- AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
- AI->getAlignment(),
- AI->getName() + "." + utostr(i), AI);
- ElementAllocas.push_back(NA);
- WorkList.push_back(NA); // Add to worklist for recursive processing
- }
- } else {
- const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
- ElementAllocas.reserve(AT->getNumElements());
- const Type *ElTy = AT->getElementType();
- for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
- AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
- AI->getName() + "." + utostr(i), AI);
- ElementAllocas.push_back(NA);
- WorkList.push_back(NA); // Add to worklist for recursive processing
- }
+/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl
+/// predicate, do SROA now.
+void SROA::DoScalarReplacement(AllocationInst *AI,
+ std::vector<AllocationInst*> &WorkList) {
+ DOUT << "Found inst to SROA: " << *AI;
+ SmallVector<AllocaInst*, 32> ElementAllocas;
+ if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
+ ElementAllocas.reserve(ST->getNumContainedTypes());
+ for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
+ AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
+ AI->getAlignment(),
+ AI->getName() + "." + utostr(i), AI);
+ ElementAllocas.push_back(NA);
+ WorkList.push_back(NA); // Add to worklist for recursive processing
+ }
+ } else {
+ const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
+ ElementAllocas.reserve(AT->getNumElements());
+ const Type *ElTy = AT->getElementType();
+ for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
+ AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
+ AI->getName() + "." + utostr(i), AI);
+ ElementAllocas.push_back(NA);
+ WorkList.push_back(NA); // Add to worklist for recursive processing
}
+ }
- // Now that we have created the alloca instructions that we want to use,
- // expand the getelementptr instructions to use them.
- //
- while (!AI->use_empty()) {
- Instruction *User = cast<Instruction>(AI->use_back());
- if (BitCastInst *BCInst = dyn_cast<BitCastInst>(User)) {
- RewriteBitCastUserOfAlloca(BCInst, AI, ElementAllocas);
- BCInst->eraseFromParent();
- continue;
- }
-
- GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User);
- // We now know that the GEP is of the form: GEP <ptr>, 0, <cst>
- unsigned Idx =
- (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
-
- assert(Idx < ElementAllocas.size() && "Index out of range?");
- AllocaInst *AllocaToUse = ElementAllocas[Idx];
-
- Value *RepValue;
- if (GEPI->getNumOperands() == 3) {
- // Do not insert a new getelementptr instruction with zero indices, only
- // to have it optimized out later.
- RepValue = AllocaToUse;
- } else {
- // We are indexing deeply into the structure, so we still need a
- // getelement ptr instruction to finish the indexing. This may be
- // expanded itself once the worklist is rerun.
- //
- SmallVector<Value*, 8> NewArgs;
- NewArgs.push_back(Constant::getNullValue(Type::Int32Ty));
- NewArgs.append(GEPI->op_begin()+3, GEPI->op_end());
- RepValue = new GetElementPtrInst(AllocaToUse, &NewArgs[0],
- NewArgs.size(), "", GEPI);
- RepValue->takeName(GEPI);
- }
-
- // If this GEP is to the start of the aggregate, check for memcpys.
- if (Idx == 0) {
- bool IsStartOfAggregateGEP = true;
- for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i) {
- if (!isa<ConstantInt>(GEPI->getOperand(i))) {
- IsStartOfAggregateGEP = false;
- break;
- }
- if (!cast<ConstantInt>(GEPI->getOperand(i))->isZero()) {
- IsStartOfAggregateGEP = false;
- break;
- }
+ // Now that we have created the alloca instructions that we want to use,
+ // expand the getelementptr instructions to use them.
+ //
+ while (!AI->use_empty()) {
+ Instruction *User = cast<Instruction>(AI->use_back());
+ if (BitCastInst *BCInst = dyn_cast<BitCastInst>(User)) {
+ RewriteBitCastUserOfAlloca(BCInst, AI, ElementAllocas);
+ BCInst->eraseFromParent();
+ continue;
+ }
+
+ GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User);
+ // We now know that the GEP is of the form: GEP <ptr>, 0, <cst>
+ unsigned Idx =
+ (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
+
+ assert(Idx < ElementAllocas.size() && "Index out of range?");
+ AllocaInst *AllocaToUse = ElementAllocas[Idx];
+
+ Value *RepValue;
+ if (GEPI->getNumOperands() == 3) {
+ // Do not insert a new getelementptr instruction with zero indices, only
+ // to have it optimized out later.
+ RepValue = AllocaToUse;
+ } else {
+ // We are indexing deeply into the structure, so we still need a
+ // getelement ptr instruction to finish the indexing. This may be
+ // expanded itself once the worklist is rerun.
+ //
+ SmallVector<Value*, 8> NewArgs;
+ NewArgs.push_back(Constant::getNullValue(Type::Int32Ty));
+ NewArgs.append(GEPI->op_begin()+3, GEPI->op_end());
+ RepValue = new GetElementPtrInst(AllocaToUse, NewArgs.begin(),
+ NewArgs.end(), "", GEPI);
+ RepValue->takeName(GEPI);
+ }
+
+ // If this GEP is to the start of the aggregate, check for memcpys.
+ if (Idx == 0) {
+ bool IsStartOfAggregateGEP = true;
+ for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i) {
+ if (!isa<ConstantInt>(GEPI->getOperand(i))) {
+ IsStartOfAggregateGEP = false;
+ break;
+ }
+ if (!cast<ConstantInt>(GEPI->getOperand(i))->isZero()) {
+ IsStartOfAggregateGEP = false;
+ break;
}
-
- if (IsStartOfAggregateGEP)
- RewriteBitCastUserOfAlloca(GEPI, AI, ElementAllocas);
}
-
- // Move all of the users over to the new GEP.
- GEPI->replaceAllUsesWith(RepValue);
- // Delete the old GEP
- GEPI->eraseFromParent();
+ if (IsStartOfAggregateGEP)
+ RewriteBitCastUserOfAlloca(GEPI, AI, ElementAllocas);
}
+
- // Finally, delete the Alloca instruction
- AI->eraseFromParent();
- NumReplaced++;
+ // Move all of the users over to the new GEP.
+ GEPI->replaceAllUsesWith(RepValue);
+ // Delete the old GEP
+ GEPI->eraseFromParent();
}
- return Changed;
+ // Finally, delete the Alloca instruction
+ AI->eraseFromParent();
+ NumReplaced++;
}
/// getelementptr instruction of an array aggregate allocation. isFirstElt
/// indicates whether Ptr is known to the start of the aggregate.
///
-int SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI) {
+void SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI,
+ AllocaInfo &Info) {
for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
I != E; ++I) {
Instruction *User = cast<Instruction>(*I);
case Instruction::Load: break;
case Instruction::Store:
// Store is ok if storing INTO the pointer, not storing the pointer
- if (User->getOperand(0) == Ptr) return 0;
+ if (User->getOperand(0) == Ptr) return MarkUnsafe(Info);
break;
case Instruction::GetElementPtr: {
GetElementPtrInst *GEP = cast<GetElementPtrInst>(User);
if (GEP->getNumOperands() > 1) {
if (!isa<ConstantInt>(GEP->getOperand(1)) ||
!cast<ConstantInt>(GEP->getOperand(1))->isZero())
- return 0; // Using pointer arithmetic to navigate the array.
+ // Using pointer arithmetic to navigate the array.
+ return MarkUnsafe(Info);
if (AreAllZeroIndices) {
for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i) {
}
}
}
- if (!isSafeElementUse(GEP, AreAllZeroIndices, AI)) return 0;
+ isSafeElementUse(GEP, AreAllZeroIndices, AI, Info);
+ if (Info.isUnsafe) return;
break;
}
case Instruction::BitCast:
- if (isFirstElt &&
- isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI))
+ if (isFirstElt) {
+ isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info);
+ if (Info.isUnsafe) return;
break;
+ }
DOUT << " Transformation preventing inst: " << *User;
- return 0;
+ return MarkUnsafe(Info);
case Instruction::Call:
if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
- if (isFirstElt && isSafeMemIntrinsicOnAllocation(MI, AI))
+ if (isFirstElt) {
+ isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info);
+ if (Info.isUnsafe) return;
break;
+ }
}
DOUT << " Transformation preventing inst: " << *User;
- return 0;
+ return MarkUnsafe(Info);
default:
DOUT << " Transformation preventing inst: " << *User;
- return 0;
+ return MarkUnsafe(Info);
}
}
- return 3; // All users look ok :)
+ return; // All users look ok :)
}
/// AllUsersAreLoads - Return true if all users of this value are loads.
/// isSafeUseOfAllocation - Check to see if this user is an allowed use for an
/// aggregate allocation.
///
-int SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI) {
+void SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI,
+ AllocaInfo &Info) {
if (BitCastInst *C = dyn_cast<BitCastInst>(User))
- return isSafeUseOfBitCastedAllocation(C, AI) ? 3 : 0;
- if (!isa<GetElementPtrInst>(User)) return 0;
+ return isSafeUseOfBitCastedAllocation(C, AI, Info);
+
+ GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User);
+ if (GEPI == 0)
+ return MarkUnsafe(Info);
- GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User);
gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_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()))
- return 0;
+ I.getOperand() != Constant::getNullValue(I.getOperand()->getType())) {
+ return MarkUnsafe(Info);
+ }
++I;
- if (I == E) return 0; // ran out of GEP indices??
+ if (I == E) return MarkUnsafe(Info); // ran out of GEP indices??
bool IsAllZeroIndices = true;
// something funny is going on, so we won't do the optimization.
//
if (Idx->getZExtValue() >= NumElements)
- return 0;
+ return MarkUnsafe(Info);
// We cannot scalar repl this level of the array unless any array
// sub-indices are in-range constants. In particular, consider:
NumElements = cast<VectorType>(*I)->getNumElements();
ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand());
- if (!IdxVal) return 0;
+ if (!IdxVal) return MarkUnsafe(Info);
if (IdxVal->getZExtValue() >= NumElements)
- return 0;
+ return MarkUnsafe(Info);
IsAllZeroIndices &= IdxVal->isZero();
}
// it, in which case we CAN promote it, but we have to canonicalize this
// out if this is the only problem.
if ((NumElements == 1 || NumElements == 2) &&
- AllUsersAreLoads(GEPI))
- return 1; // Canonicalization required!
- return 0;
+ AllUsersAreLoads(GEPI)) {
+ Info.needsCanon = true;
+ return; // Canonicalization required!
+ }
+ return MarkUnsafe(Info);
}
}
// If there are any non-simple uses of this getelementptr, make sure to reject
// them.
- return isSafeElementUse(GEPI, IsAllZeroIndices, AI);
+ return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info);
}
/// isSafeMemIntrinsicOnAllocation - Return true if the specified memory
/// intrinsic can be promoted by SROA. At this point, we know that the operand
/// of the memintrinsic is a pointer to the beginning of the allocation.
-bool SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI){
+void SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI,
+ unsigned OpNo, AllocaInfo &Info) {
// If not constant length, give up.
ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
- if (!Length) return false;
+ if (!Length) return MarkUnsafe(Info);
// If not the whole aggregate, give up.
const TargetData &TD = getAnalysis<TargetData>();
- if (Length->getZExtValue() != TD.getTypeSize(AI->getType()->getElementType()))
- return false;
+ if (Length->getZExtValue() !=
+ TD.getABITypeSize(AI->getType()->getElementType()))
+ return MarkUnsafe(Info);
// We only know about memcpy/memset/memmove.
if (!isa<MemCpyInst>(MI) && !isa<MemSetInst>(MI) && !isa<MemMoveInst>(MI))
- return false;
- // Otherwise, we can transform it.
- return true;
+ return MarkUnsafe(Info);
+
+ // Otherwise, we can transform it. Determine whether this is a memcpy/set
+ // into or out of the aggregate.
+ if (OpNo == 1)
+ Info.isMemCpyDst = true;
+ else {
+ assert(OpNo == 2);
+ Info.isMemCpySrc = true;
+ }
}
/// isSafeUseOfBitCastedAllocation - Return true if all users of this bitcast
/// are
-bool SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocationInst *AI) {
+void SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocationInst *AI,
+ AllocaInfo &Info) {
for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end();
UI != E; ++UI) {
if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) {
- if (!isSafeUseOfBitCastedAllocation(BCU, AI))
- return false;
+ isSafeUseOfBitCastedAllocation(BCU, AI, Info);
} else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) {
- if (!isSafeMemIntrinsicOnAllocation(MI, AI))
- return false;
+ isSafeMemIntrinsicOnAllocation(MI, AI, UI.getOperandNo(), Info);
} else {
- return false;
+ return MarkUnsafe(Info);
}
+ if (Info.isUnsafe) return;
}
- return true;
}
/// RewriteBitCastUserOfAlloca - BCInst (transitively) bitcasts AI, or indexes
// If this is a memcpy/memmove, emit a GEP of the other element address.
Value *OtherElt = 0;
if (OtherPtr) {
- OtherElt = new GetElementPtrInst(OtherPtr, Zero,
- ConstantInt::get(Type::Int32Ty, i),
+ Value *Idx[2];
+ Idx[0] = Zero;
+ Idx[1] = ConstantInt::get(Type::Int32Ty, i);
+ OtherElt = new GetElementPtrInst(OtherPtr, Idx, Idx + 2,
OtherPtr->getNameStr()+"."+utostr(i),
MI);
}
if (CI->isZero()) {
StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0>
} else {
- // If EltTy is a packed type, get the element type.
+ // 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();
-
+
// Construct an integer with the right value.
- unsigned EltSize = TD.getTypeSize(ValTy);
- APInt OneVal(EltSize*8, CI->getZExtValue());
+ unsigned EltSize = TD.getTypeSizeInBits(ValTy);
+ APInt OneVal(EltSize, CI->getZExtValue());
APInt TotalVal(OneVal);
// Set each byte.
- for (unsigned i = 0; i != EltSize-1; ++i) {
+ for (unsigned i = 0; 8*i < EltSize; ++i) {
TotalVal = TotalVal.shl(8);
TotalVal |= OneVal;
}
-
+
// Convert the integer value to the appropriate type.
StoreVal = ConstantInt::get(TotalVal);
if (isa<PointerType>(ValTy))
OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(),
MI);
- unsigned EltSize = TD.getTypeSize(EltTy);
+ unsigned EltSize = TD.getABITypeSize(EltTy);
// Finally, insert the meminst for this element.
if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
Zero // Align
};
- new CallInst(TheFn, Ops, 4, "", MI);
+ new CallInst(TheFn, Ops, Ops + 4, "", MI);
} else {
assert(isa<MemSetInst>(MI));
Value *Ops[] = {
ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
Zero // Align
};
- new CallInst(TheFn, Ops, 4, "", MI);
+ new CallInst(TheFn, Ops, Ops + 4, "", MI);
}
}
}
}
+/// HasPadding - Return true if the specified type has any structure or
+/// alignment padding, false otherwise.
+static bool HasPadding(const Type *Ty, const TargetData &TD,
+ bool inPacked = false) {
+ if (const StructType *STy = dyn_cast<StructType>(Ty)) {
+ const StructLayout *SL = TD.getStructLayout(STy);
+ unsigned PrevFieldBitOffset = 0;
+ for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
+ unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
+
+ // Padding in sub-elements?
+ if (HasPadding(STy->getElementType(i), TD, STy->isPacked()))
+ return true;
+
+ // Check to see if there is any padding between this element and the
+ // previous one.
+ if (i) {
+ unsigned PrevFieldEnd =
+ PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1));
+ if (PrevFieldEnd < FieldBitOffset)
+ return true;
+ }
+
+ PrevFieldBitOffset = FieldBitOffset;
+ }
+
+ // Check for tail padding.
+ if (unsigned EltCount = STy->getNumElements()) {
+ unsigned PrevFieldEnd = PrevFieldBitOffset +
+ TD.getTypeSizeInBits(STy->getElementType(EltCount-1));
+ if (PrevFieldEnd < SL->getSizeInBits())
+ return true;
+ }
+
+ } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
+ return HasPadding(ATy->getElementType(), TD, false);
+ } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
+ return HasPadding(VTy->getElementType(), TD, false);
+ }
+ return inPacked ?
+ false : TD.getTypeSizeInBits(Ty) != TD.getABITypeSizeInBits(Ty);
+}
/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe,
int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) {
// Loop over the use list of the alloca. We can only transform it if all of
// the users are safe to transform.
- //
- int isSafe = 3;
+ AllocaInfo Info;
+
for (Value::use_iterator I = AI->use_begin(), E = AI->use_end();
I != E; ++I) {
- isSafe &= isSafeUseOfAllocation(cast<Instruction>(*I), AI);
- if (isSafe == 0) {
+ isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info);
+ if (Info.isUnsafe) {
DOUT << "Cannot transform: " << *AI << " due to user: " << **I;
return 0;
}
}
- // If we require cleanup, isSafe is now 1, otherwise it is 3.
- return isSafe;
+
+ // Okay, we know all the users are promotable. If the aggregate is a memcpy
+ // source and destination, we have to be careful. In particular, the memcpy
+ // could be moving around elements that live in structure padding of the LLVM
+ // types, but may actually be used. In these cases, we refuse to promote the
+ // struct.
+ if (Info.isMemCpySrc && Info.isMemCpyDst &&
+ HasPadding(AI->getType()->getElementType(), getAnalysis<TargetData>()))
+ return 0;
+
+ // If we require cleanup, return 1, otherwise return 3.
+ return Info.needsCanon ? 1 : 3;
}
/// CanonicalizeAllocaUsers - If SROA reported that it can promote the specified
// up.
for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
UI != E; ) {
- GetElementPtrInst *GEPI = cast<GetElementPtrInst>(*UI++);
+ GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI++);
+ if (!GEPI) continue;
gep_type_iterator I = gep_type_begin(GEPI);
++I;
SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end());
Indices[1] = Constant::getNullValue(Type::Int32Ty);
Value *ZeroIdx = new GetElementPtrInst(GEPI->getOperand(0),
- &Indices[0], Indices.size(),
+ Indices.begin(),
+ Indices.end(),
GEPI->getName()+".0", GEPI);
Indices[1] = ConstantInt::get(Type::Int32Ty, 1);
Value *OneIdx = new GetElementPtrInst(GEPI->getOperand(0),
- &Indices[0], Indices.size(),
+ Indices.begin(),
+ Indices.end(),
GEPI->getName()+".1", GEPI);
// Replace all loads of the variable index GEP with loads from both
// indexes and a select.
case Type::PointerTyID: Accum = TD.getIntPtrType(); break;
case Type::FloatTyID: Accum = Type::Int32Ty; break;
case Type::DoubleTyID: Accum = Type::Int64Ty; break;
+ case Type::X86_FP80TyID: return true;
+ case Type::FP128TyID: return true;
+ case Type::PPC_FP128TyID: return true;
default:
assert(Accum->isInteger() && "Unknown FP type!");
break;
case Type::PointerTyID: In = TD.getIntPtrType(); break;
case Type::FloatTyID: In = Type::Int32Ty; break;
case Type::DoubleTyID: In = Type::Int64Ty; break;
+ case Type::X86_FP80TyID: return true;
+ case Type::FP128TyID: return true;
+ case Type::PPC_FP128TyID: return true;
default:
assert(In->isInteger() && "Unknown FP type!");
break;
return false;
}
-/// getUIntAtLeastAsBitAs - Return an unsigned integer type that is at least
+/// getUIntAtLeastAsBigAs - Return an unsigned integer type that is at least
/// as big as the specified type. If there is no suitable type, this returns
/// null.
-const Type *getUIntAtLeastAsBitAs(unsigned NumBits) {
+const Type *getUIntAtLeastAsBigAs(unsigned NumBits) {
if (NumBits > 64) return 0;
if (NumBits > 32) return Type::Int64Ty;
if (NumBits > 16) return Type::Int32Ty;
// Check to see if this is stepping over an element: GEP Ptr, int C
if (GEP->getNumOperands() == 2 && isa<ConstantInt>(GEP->getOperand(1))) {
unsigned Idx = cast<ConstantInt>(GEP->getOperand(1))->getZExtValue();
- unsigned ElSize = TD.getTypeSize(PTy->getElementType());
+ unsigned ElSize = TD.getABITypeSize(PTy->getElementType());
unsigned BitOffset = Idx*ElSize*8;
if (BitOffset > 64 || !isPowerOf2_32(ElSize)) return 0;
if (SubElt == 0) return 0;
if (SubElt != Type::VoidTy && SubElt->isInteger()) {
const Type *NewTy =
- getUIntAtLeastAsBitAs(TD.getTypeSize(SubElt)*8+BitOffset);
+ getUIntAtLeastAsBigAs(TD.getABITypeSizeInBits(SubElt)+BitOffset);
if (NewTy == 0 || MergeInType(NewTy, UsedType, TD)) return 0;
continue;
}
} else if (GEP->getNumOperands() == 3 &&
isa<ConstantInt>(GEP->getOperand(1)) &&
isa<ConstantInt>(GEP->getOperand(2)) &&
- cast<Constant>(GEP->getOperand(1))->isNullValue()) {
+ cast<ConstantInt>(GEP->getOperand(1))->isZero()) {
// We are stepping into an element, e.g. a structure or an array:
// GEP Ptr, int 0, uint C
const Type *AggTy = PTy->getElementType();
if (const ArrayType *ATy = dyn_cast<ArrayType>(AggTy)) {
if (Idx >= ATy->getNumElements()) return 0; // Out of range.
} else if (const VectorType *VectorTy = dyn_cast<VectorType>(AggTy)) {
- // Getting an element of the packed vector.
+ // Getting an element of the vector.
if (Idx >= VectorTy->getNumElements()) return 0; // Out of range.
// Merge in the vector type.
} else {
return 0;
}
- const Type *NTy = getUIntAtLeastAsBitAs(TD.getTypeSize(AggTy)*8);
+ const Type *NTy = getUIntAtLeastAsBigAs(TD.getABITypeSizeInBits(AggTy));
if (NTy == 0 || MergeInType(NTy, UsedType, TD)) return 0;
const Type *SubTy = CanConvertToScalar(GEP, IsNotTrivial);
if (SubTy == 0) return 0;
++NumConverted;
BasicBlock *EntryBlock = AI->getParent();
- assert(EntryBlock == &EntryBlock->getParent()->front() &&
+ assert(EntryBlock == &EntryBlock->getParent()->getEntryBlock() &&
"Not in the entry block!");
EntryBlock->getInstList().remove(AI); // Take the alloca out of the program.
/// 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 SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, unsigned Offset) {
- bool isVectorInsert = isa<VectorType>(NewAI->getType()->getElementType());
const TargetData &TD = getAnalysis<TargetData>();
while (!Ptr->use_empty()) {
Instruction *User = cast<Instruction>(Ptr->use_back());
if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
// The load is a bit extract from NewAI shifted right by Offset bits.
Value *NV = new LoadInst(NewAI, LI->getName(), LI);
- if (NV->getType() != LI->getType()) {
- if (const VectorType *PTy = dyn_cast<VectorType>(NV->getType())) {
- // If the result alloca is a vector type, this is either an element
- // access or a bitcast to another vector type.
- if (isa<VectorType>(LI->getType())) {
- NV = new BitCastInst(NV, LI->getType(), LI->getName(), LI);
- } else {
- // Must be an element access.
- unsigned Elt = Offset/(TD.getTypeSize(PTy->getElementType())*8);
- NV = new ExtractElementInst(
- NV, ConstantInt::get(Type::Int32Ty, Elt), "tmp", LI);
- }
- } else if (isa<PointerType>(NV->getType())) {
- assert(isa<PointerType>(LI->getType()));
- // Must be ptr->ptr cast. Anything else would result in NV being
- // an integer.
+ if (NV->getType() == LI->getType()) {
+ // We win, no conversion needed.
+ } else if (const VectorType *PTy = dyn_cast<VectorType>(NV->getType())) {
+ // If the result alloca is a vector type, this is either an element
+ // access or a bitcast to another vector type.
+ if (isa<VectorType>(LI->getType())) {
NV = new BitCastInst(NV, LI->getType(), LI->getName(), LI);
} else {
- assert(NV->getType()->isInteger() && "Unknown promotion!");
- if (Offset && Offset < TD.getTypeSize(NV->getType())*8) {
- NV = BinaryOperator::createLShr(NV,
- ConstantInt::get(NV->getType(), Offset),
- LI->getName(), LI);
- }
-
- // If the result is an integer, this is a trunc or bitcast.
- if (LI->getType()->isInteger()) {
- NV = CastInst::createTruncOrBitCast(NV, LI->getType(),
- LI->getName(), LI);
- } else if (LI->getType()->isFloatingPoint()) {
- // If needed, truncate the integer to the appropriate size.
- if (NV->getType()->getPrimitiveSizeInBits() >
- LI->getType()->getPrimitiveSizeInBits()) {
- switch (LI->getType()->getTypeID()) {
- default: assert(0 && "Unknown FP type!");
- case Type::FloatTyID:
- NV = new TruncInst(NV, Type::Int32Ty, LI->getName(), LI);
- break;
- case Type::DoubleTyID:
- NV = new TruncInst(NV, Type::Int64Ty, LI->getName(), LI);
- break;
- }
- }
-
- // Then do a bitcast.
- NV = new BitCastInst(NV, LI->getType(), LI->getName(), LI);
- } else {
- // Otherwise must be a pointer.
- NV = new IntToPtrInst(NV, LI->getType(), LI->getName(), LI);
- }
+ // Must be an element access.
+ unsigned Elt = Offset/TD.getABITypeSizeInBits(PTy->getElementType());
+ NV = new ExtractElementInst(
+ NV, ConstantInt::get(Type::Int32Ty, Elt), "tmp", LI);
+ }
+ } else if (isa<PointerType>(NV->getType())) {
+ assert(isa<PointerType>(LI->getType()));
+ // Must be ptr->ptr cast. Anything else would result in NV being
+ // an integer.
+ NV = new BitCastInst(NV, LI->getType(), LI->getName(), LI);
+ } else {
+ const IntegerType *NTy = cast<IntegerType>(NV->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.
+ int ShAmt = 0;
+ if (TD.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(LI->getType()) - Offset;
+ } else {
+ ShAmt = Offset;
+ }
+
+ // Note: we support negative bitwidths (with shl) which are not defined.
+ // 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);
+ else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
+ NV = BinaryOperator::createShl(NV,
+ ConstantInt::get(NV->getType(),-ShAmt),
+ LI->getName(), LI);
+
+ // Finally, unconditionally truncate the integer to the right width.
+ unsigned LIBitWidth = TD.getTypeSizeInBits(LI->getType());
+ if (LIBitWidth < NTy->getBitWidth())
+ NV = new TruncInst(NV, IntegerType::get(LIBitWidth),
+ LI->getName(), LI);
+
+ // If the result is an integer, this is a trunc or bitcast.
+ if (isa<IntegerType>(LI->getType())) {
+ assert(NV->getType() == LI->getType() && "Truncate wasn't enough?");
+ } else if (LI->getType()->isFloatingPoint()) {
+ // Just do a bitcast, we know the sizes match up.
+ NV = new BitCastInst(NV, LI->getType(), LI->getName(), LI);
+ } else {
+ // Otherwise must be a pointer.
+ NV = new IntToPtrInst(NV, LI->getType(), LI->getName(), LI);
}
}
LI->replaceAllUsesWith(NV);
// then 'or' into place.
Value *SV = SI->getOperand(0);
const Type *AllocaType = NewAI->getType()->getElementType();
- if (SV->getType() != AllocaType) {
+ if (SV->getType() == AllocaType) {
+ // All is well.
+ } else if (const VectorType *PTy = dyn_cast<VectorType>(AllocaType)) {
Value *Old = new LoadInst(NewAI, NewAI->getName()+".in", SI);
-
- if (const VectorType *PTy = dyn_cast<VectorType>(AllocaType)) {
- // 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.getTypeSize(PTy->getElementType())*8);
- SV = new InsertElementInst(Old, SV,
- ConstantInt::get(Type::Int32Ty, Elt),
- "tmp", SI);
- }
+
+ // 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 {
- // 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.
- switch (SV->getType()->getTypeID()) {
- default:
- assert(!SV->getType()->isFloatingPoint() && "Unknown FP type!");
- break;
- case Type::FloatTyID:
- SV = new BitCastInst(SV, Type::Int32Ty, SV->getName(), SI);
- break;
- case Type::DoubleTyID:
- SV = new BitCastInst(SV, Type::Int64Ty, SV->getName(), SI);
- break;
- case Type::PointerTyID:
- if (isa<PointerType>(AllocaType))
- SV = new BitCastInst(SV, AllocaType, SV->getName(), SI);
- else
- SV = new PtrToIntInst(SV, TD.getIntPtrType(), SV->getName(), SI);
- break;
- }
+ // Must be an element insertion.
+ unsigned Elt = Offset/TD.getABITypeSizeInBits(PTy->getElementType());
+ SV = new InsertElementInst(Old, SV,
+ ConstantInt::get(Type::Int32Ty, Elt),
+ "tmp", SI);
+ }
+ } else if (isa<PointerType>(AllocaType)) {
+ // If the alloca type is a pointer, then all the elements must be
+ // pointers.
+ if (SV->getType() != AllocaType)
+ SV = new BitCastInst(SV, AllocaType, SV->getName(), SI);
+ } else {
+ Value *Old = new LoadInst(NewAI, NewAI->getName()+".in", SI);
- unsigned SrcSize = TD.getTypeSize(SV->getType())*8;
-
- // Always zero extend the value if needed.
- if (SV->getType() != AllocaType)
- SV = CastInst::createZExtOrBitCast(SV, AllocaType,
- SV->getName(), SI);
- if (Offset && Offset < AllocaType->getPrimitiveSizeInBits())
- SV = BinaryOperator::createShl(SV,
- ConstantInt::get(SV->getType(), Offset),
- SV->getName()+".adj", SI);
- // Mask out the bits we are about to insert from the old value.
- unsigned TotalBits = TD.getTypeSize(SV->getType())*8;
- if (TotalBits != SrcSize) {
- assert(TotalBits > SrcSize);
- uint64_t Mask = ~(((1ULL << SrcSize)-1) << Offset);
- Mask = Mask & cast<IntegerType>(SV->getType())->getBitMask();
- Old = BinaryOperator::createAnd(Old,
- ConstantInt::get(Old->getType(), Mask),
- Old->getName()+".mask", SI);
- SV = BinaryOperator::createOr(Old, SV, SV->getName()+".ins", SI);
- }
+ // 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.
+ 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())
+ SV = new BitCastInst(SV, IntegerType::get(SrcWidth),
+ SV->getName(), SI);
+ else if (isa<PointerType>(SV->getType()))
+ SV = new PtrToIntInst(SV, TD.getIntPtrType(), SV->getName(), SI);
+
+ // Always zero extend the value if needed.
+ if (SV->getType() != AllocaType)
+ SV = new ZExtInst(SV, AllocaType, SV->getName(), SI);
+
+ // 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()) {
+ // 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 = DestStoreWidth - SrcStoreWidth - Offset;
+ } else {
+ ShAmt = Offset;
+ }
+
+ // Note: we support negative bitwidths (with shr) which are not defined.
+ // We do this to support (f.e.) stores off the end of a structure where
+ // 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);
+ Mask <<= ShAmt;
+ } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
+ SV = BinaryOperator::createLShr(SV,
+ ConstantInt::get(SV->getType(),-ShAmt),
+ SV->getName(), SI);
+ Mask = Mask.lshr(ShAmt);
+ }
+
+ // Mask out the bits we are about to insert from the old value, and or
+ // 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);
}
}
new StoreInst(SV, NewAI, SI);
SI->eraseFromParent();
- } else if (CastInst *CI = dyn_cast<CastInst>(User)) {
- unsigned NewOff = Offset;
- const TargetData &TD = getAnalysis<TargetData>();
- if (TD.isBigEndian() && !isVectorInsert) {
- // Adjust the pointer. For example, storing 16-bits into a 32-bit
- // alloca with just a cast makes it modify the top 16-bits.
- const Type *SrcTy = cast<PointerType>(Ptr->getType())->getElementType();
- const Type *DstTy = cast<PointerType>(CI->getType())->getElementType();
- int PtrDiffBits = TD.getTypeSize(SrcTy)*8-TD.getTypeSize(DstTy)*8;
- NewOff += PtrDiffBits;
- }
- ConvertUsesToScalar(CI, NewAI, NewOff);
+ } else if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
+ ConvertUsesToScalar(CI, NewAI, Offset);
CI->eraseFromParent();
} else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
const PointerType *AggPtrTy =
cast<PointerType>(GEP->getOperand(0)->getType());
const TargetData &TD = getAnalysis<TargetData>();
- unsigned AggSizeInBits = TD.getTypeSize(AggPtrTy->getElementType())*8;
-
+ unsigned AggSizeInBits =
+ TD.getABITypeSizeInBits(AggPtrTy->getElementType());
+
// Check to see if this is stepping over an element: GEP Ptr, int C
unsigned NewOffset = Offset;
if (GEP->getNumOperands() == 2) {
unsigned Idx = cast<ConstantInt>(GEP->getOperand(1))->getZExtValue();
unsigned BitOffset = Idx*AggSizeInBits;
- if (TD.isLittleEndian() || isVectorInsert)
- NewOffset += BitOffset;
- else
- NewOffset -= BitOffset;
-
+ NewOffset += BitOffset;
} else if (GEP->getNumOperands() == 3) {
// We know that operand #2 is zero.
unsigned Idx = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
const Type *AggTy = AggPtrTy->getElementType();
if (const SequentialType *SeqTy = dyn_cast<SequentialType>(AggTy)) {
- unsigned ElSizeBits = TD.getTypeSize(SeqTy->getElementType())*8;
+ unsigned ElSizeBits =
+ TD.getABITypeSizeInBits(SeqTy->getElementType());
- if (TD.isLittleEndian() || isVectorInsert)
- NewOffset += ElSizeBits*Idx;
- else
- NewOffset += AggSizeInBits-ElSizeBits*(Idx+1);
+ NewOffset += ElSizeBits*Idx;
} else if (const StructType *STy = dyn_cast<StructType>(AggTy)) {
unsigned EltBitOffset =
- TD.getStructLayout(STy)->getElementOffset(Idx)*8;
-
- if (TD.isLittleEndian() || isVectorInsert)
- NewOffset += EltBitOffset;
- else {
- const PointerType *ElPtrTy = cast<PointerType>(GEP->getType());
- unsigned ElSizeBits = TD.getTypeSize(ElPtrTy->getElementType())*8;
- NewOffset += AggSizeInBits-(EltBitOffset+ElSizeBits);
- }
+ TD.getStructLayout(STy)->getElementOffsetInBits(Idx);
+ NewOffset += EltBitOffset;
} else {
assert(0 && "Unsupported operation!");
abort();
}
}
}
+
+
+/// 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, Instruction *&TheCopy,
+ bool isOffset) {
+ for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
+ if (isa<LoadInst>(*UI)) {
+ // Ignore loads, they are always ok.
+ continue;
+ }
+ if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) {
+ // If uses of the bitcast are ok, we are ok.
+ if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset))
+ return false;
+ continue;
+ }
+ if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) {
+ // 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()))
+ return false;
+ continue;
+ }
+
+ // If this is isn't our memcpy/memmove, reject it as something we can't
+ // handle.
+ if (!isa<MemCpyInst>(*UI) && !isa<MemMoveInst>(*UI))
+ return false;
+
+ // 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() != 1) return false;
+
+ MemIntrinsic *MI = cast<MemIntrinsic>(*UI);
+
+ // If the source of the memcpy/move is not a constant global, reject it.
+ if (!PointsToConstantGlobal(MI->getOperand(2)))
+ 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.
+Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocationInst *AI) {
+ Instruction *TheCopy = 0;
+ if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false))
+ return TheCopy;
+ return 0;
+}