#include "llvm/Analysis/ScalarEvolutionExpander.h"
#include "llvm/Analysis/LoopInfo.h"
+#include "llvm/LLVMContext.h"
#include "llvm/Target/TargetData.h"
#include "llvm/ADT/STLExtras.h"
using namespace llvm;
-/// InsertCastOfTo - Insert a cast of V to the specified type, doing what
-/// we can to share the casts.
-Value *SCEVExpander::InsertCastOfTo(Instruction::CastOps opcode, Value *V,
- const Type *Ty) {
+/// InsertNoopCastOfTo - Insert a cast of V to the specified type,
+/// which must be possible with a noop cast, doing what we can to share
+/// the casts.
+Value *SCEVExpander::InsertNoopCastOfTo(Value *V, const Type *Ty) {
+ Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false);
+ assert((Op == Instruction::BitCast ||
+ Op == Instruction::PtrToInt ||
+ Op == Instruction::IntToPtr) &&
+ "InsertNoopCastOfTo cannot perform non-noop casts!");
+ assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) &&
+ "InsertNoopCastOfTo cannot change sizes!");
+
// Short-circuit unnecessary bitcasts.
- if (opcode == Instruction::BitCast && V->getType() == Ty)
+ if (Op == Instruction::BitCast && V->getType() == Ty)
return V;
// Short-circuit unnecessary inttoptr<->ptrtoint casts.
- if ((opcode == Instruction::PtrToInt || opcode == Instruction::IntToPtr) &&
+ if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) &&
SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) {
if (CastInst *CI = dyn_cast<CastInst>(V))
if ((CI->getOpcode() == Instruction::PtrToInt ||
return CE->getOperand(0);
}
- // FIXME: keep track of the cast instruction.
if (Constant *C = dyn_cast<Constant>(V))
- return ConstantExpr::getCast(opcode, C, Ty);
-
+ return ConstantExpr::getCast(Op, C, Ty);
+
if (Argument *A = dyn_cast<Argument>(V)) {
// Check to see if there is already a cast!
for (Value::use_iterator UI = A->use_begin(), E = A->use_end();
UI != E; ++UI)
if ((*UI)->getType() == Ty)
if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI)))
- if (CI->getOpcode() == opcode) {
+ if (CI->getOpcode() == Op) {
// If the cast isn't the first instruction of the function, move it.
if (BasicBlock::iterator(CI) !=
A->getParent()->getEntryBlock().begin()) {
// The old cast is left in place in case it is being used
// as an insert point.
Instruction *NewCI =
- CastInst::Create(opcode, V, Ty, "",
+ CastInst::Create(Op, V, Ty, "",
A->getParent()->getEntryBlock().begin());
NewCI->takeName(CI);
CI->replaceAllUsesWith(NewCI);
return CI;
}
- Instruction *I = CastInst::Create(opcode, V, Ty, V->getName(),
+ Instruction *I = CastInst::Create(Op, V, Ty, V->getName(),
A->getParent()->getEntryBlock().begin());
InsertedValues.insert(I);
return I;
UI != E; ++UI) {
if ((*UI)->getType() == Ty)
if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI)))
- if (CI->getOpcode() == opcode) {
+ if (CI->getOpcode() == Op) {
BasicBlock::iterator It = I; ++It;
if (isa<InvokeInst>(I))
It = cast<InvokeInst>(I)->getNormalDest()->begin();
// Recreate the cast at the beginning of the entry block.
// The old cast is left in place in case it is being used
// as an insert point.
- Instruction *NewCI = CastInst::Create(opcode, V, Ty, "", It);
+ Instruction *NewCI = CastInst::Create(Op, V, Ty, "", It);
NewCI->takeName(CI);
CI->replaceAllUsesWith(NewCI);
return NewCI;
if (InvokeInst *II = dyn_cast<InvokeInst>(I))
IP = II->getNormalDest()->begin();
while (isa<PHINode>(IP)) ++IP;
- Instruction *CI = CastInst::Create(opcode, V, Ty, V->getName(), IP);
+ Instruction *CI = CastInst::Create(Op, V, Ty, V->getName(), IP);
InsertedValues.insert(CI);
return CI;
}
-/// InsertNoopCastOfTo - Insert a cast of V to the specified type,
-/// which must be possible with a noop cast.
-Value *SCEVExpander::InsertNoopCastOfTo(Value *V, const Type *Ty) {
- Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false);
- assert((Op == Instruction::BitCast ||
- Op == Instruction::PtrToInt ||
- Op == Instruction::IntToPtr) &&
- "InsertNoopCastOfTo cannot perform non-noop casts!");
- assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) &&
- "InsertNoopCastOfTo cannot change sizes!");
- return InsertCastOfTo(Op, V, Ty);
-}
-
/// InsertBinop - Insert the specified binary operator, doing a small amount
/// of work to avoid inserting an obviously redundant operation.
-Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode, Value *LHS,
- Value *RHS, BasicBlock::iterator InsertPt) {
+Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode,
+ Value *LHS, Value *RHS) {
// Fold a binop with constant operands.
if (Constant *CLHS = dyn_cast<Constant>(LHS))
if (Constant *CRHS = dyn_cast<Constant>(RHS))
// Do a quick scan to see if we have this binop nearby. If so, reuse it.
unsigned ScanLimit = 6;
- BasicBlock::iterator BlockBegin = InsertPt->getParent()->begin();
- if (InsertPt != BlockBegin) {
- // Scanning starts from the last instruction before InsertPt.
- BasicBlock::iterator IP = InsertPt;
+ BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
+ // Scanning starts from the last instruction before the insertion point.
+ BasicBlock::iterator IP = Builder.GetInsertPoint();
+ if (IP != BlockBegin) {
--IP;
for (; ScanLimit; --IP, --ScanLimit) {
if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
if (IP == BlockBegin) break;
}
}
-
+
// If we haven't found this binop, insert it.
- Instruction *BO = BinaryOperator::Create(Opcode, LHS, RHS, "tmp", InsertPt);
+ Value *BO = Builder.CreateBinOp(Opcode, LHS, RHS, "tmp");
InsertedValues.insert(BO);
return BO;
}
/// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
/// unnecessary; in its place, just signed-divide Ops[i] by the scale and
/// check to see if the divide was folded.
-static bool FactorOutConstant(const SCEV* &S,
- const SCEV* &Remainder,
- const APInt &Factor,
- ScalarEvolution &SE) {
+static bool FactorOutConstant(const SCEV *&S,
+ const SCEV *&Remainder,
+ const SCEV *Factor,
+ ScalarEvolution &SE,
+ const TargetData *TD) {
// Everything is divisible by one.
- if (Factor == 1)
+ if (Factor->isOne())
+ return true;
+
+ // x/x == 1.
+ if (S == Factor) {
+ S = SE.getIntegerSCEV(1, S->getType());
return true;
+ }
// For a Constant, check for a multiple of the given factor.
if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
- ConstantInt *CI =
- ConstantInt::get(C->getValue()->getValue().sdiv(Factor));
- // If the quotient is zero and the remainder is non-zero, reject
- // the value at this scale. It will be considered for subsequent
- // smaller scales.
- if (C->isZero() || !CI->isZero()) {
- const SCEV* Div = SE.getConstant(CI);
- S = Div;
- Remainder =
- SE.getAddExpr(Remainder,
- SE.getConstant(C->getValue()->getValue().srem(Factor)));
+ // 0/x == 0.
+ if (C->isZero())
return true;
+ // Check for divisibility.
+ if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) {
+ ConstantInt *CI =
+ ConstantInt::get(SE.getContext(),
+ C->getValue()->getValue().sdiv(
+ FC->getValue()->getValue()));
+ // If the quotient is zero and the remainder is non-zero, reject
+ // the value at this scale. It will be considered for subsequent
+ // smaller scales.
+ if (!CI->isZero()) {
+ const SCEV *Div = SE.getConstant(CI);
+ S = Div;
+ Remainder =
+ SE.getAddExpr(Remainder,
+ SE.getConstant(C->getValue()->getValue().srem(
+ FC->getValue()->getValue())));
+ return true;
+ }
}
}
// In a Mul, check if there is a constant operand which is a multiple
// of the given factor.
- if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S))
- if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
- if (!C->getValue()->getValue().srem(Factor)) {
- const SmallVectorImpl<const SCEV*> &MOperands = M->getOperands();
- SmallVector<const SCEV*, 4> NewMulOps(MOperands.begin(), MOperands.end());
- NewMulOps[0] =
- SE.getConstant(C->getValue()->getValue().sdiv(Factor));
- S = SE.getMulExpr(NewMulOps);
- return true;
+ if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
+ if (TD) {
+ // With TargetData, the size is known. Check if there is a constant
+ // operand which is a multiple of the given factor. If so, we can
+ // factor it.
+ const SCEVConstant *FC = cast<SCEVConstant>(Factor);
+ if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
+ if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) {
+ const SmallVectorImpl<const SCEV *> &MOperands = M->getOperands();
+ SmallVector<const SCEV *, 4> NewMulOps(MOperands.begin(),
+ MOperands.end());
+ NewMulOps[0] =
+ SE.getConstant(C->getValue()->getValue().sdiv(
+ FC->getValue()->getValue()));
+ S = SE.getMulExpr(NewMulOps);
+ return true;
+ }
+ } else {
+ // Without TargetData, check if Factor can be factored out of any of the
+ // Mul's operands. If so, we can just remove it.
+ for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
+ const SCEV *SOp = M->getOperand(i);
+ const SCEV *Remainder = SE.getIntegerSCEV(0, SOp->getType());
+ if (FactorOutConstant(SOp, Remainder, Factor, SE, TD) &&
+ Remainder->isZero()) {
+ const SmallVectorImpl<const SCEV *> &MOperands = M->getOperands();
+ SmallVector<const SCEV *, 4> NewMulOps(MOperands.begin(),
+ MOperands.end());
+ NewMulOps[i] = SOp;
+ S = SE.getMulExpr(NewMulOps);
+ return true;
+ }
}
+ }
+ }
// In an AddRec, check if both start and step are divisible.
if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
- const SCEV* Step = A->getStepRecurrence(SE);
- const SCEV* StepRem = SE.getIntegerSCEV(0, Step->getType());
- if (!FactorOutConstant(Step, StepRem, Factor, SE))
+ const SCEV *Step = A->getStepRecurrence(SE);
+ const SCEV *StepRem = SE.getIntegerSCEV(0, Step->getType());
+ if (!FactorOutConstant(Step, StepRem, Factor, SE, TD))
return false;
if (!StepRem->isZero())
return false;
- const SCEV* Start = A->getStart();
- if (!FactorOutConstant(Start, Remainder, Factor, SE))
+ const SCEV *Start = A->getStart();
+ if (!FactorOutConstant(Start, Remainder, Factor, SE, TD))
return false;
S = SE.getAddRecExpr(Start, Step, A->getLoop());
return true;
return false;
}
-/// expandAddToGEP - Expand a SCEVAddExpr with a pointer type into a GEP
-/// instead of using ptrtoint+arithmetic+inttoptr. This helps
-/// BasicAliasAnalysis analyze the result. However, it suffers from the
-/// underlying bug described in PR2831. Addition in LLVM currently always
-/// has two's complement wrapping guaranteed. However, the semantics for
-/// getelementptr overflow are ambiguous. In the common case though, this
-/// expansion gets used when a GEP in the original code has been converted
-/// into integer arithmetic, in which case the resulting code will be no
-/// more undefined than it was originally.
+/// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs
+/// is the number of SCEVAddRecExprs present, which are kept at the end of
+/// the list.
+///
+static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops,
+ const Type *Ty,
+ ScalarEvolution &SE) {
+ unsigned NumAddRecs = 0;
+ for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i)
+ ++NumAddRecs;
+ // Group Ops into non-addrecs and addrecs.
+ SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs);
+ SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end());
+ // Let ScalarEvolution sort and simplify the non-addrecs list.
+ const SCEV *Sum = NoAddRecs.empty() ?
+ SE.getIntegerSCEV(0, Ty) :
+ SE.getAddExpr(NoAddRecs);
+ // If it returned an add, use the operands. Otherwise it simplified
+ // the sum into a single value, so just use that.
+ if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum))
+ Ops = Add->getOperands();
+ else {
+ Ops.clear();
+ if (!Sum->isZero())
+ Ops.push_back(Sum);
+ }
+ // Then append the addrecs.
+ Ops.insert(Ops.end(), AddRecs.begin(), AddRecs.end());
+}
+
+/// SplitAddRecs - Flatten a list of add operands, moving addrec start values
+/// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}.
+/// This helps expose more opportunities for folding parts of the expressions
+/// into GEP indices.
+///
+static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops,
+ const Type *Ty,
+ ScalarEvolution &SE) {
+ // Find the addrecs.
+ SmallVector<const SCEV *, 8> AddRecs;
+ for (unsigned i = 0, e = Ops.size(); i != e; ++i)
+ while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) {
+ const SCEV *Start = A->getStart();
+ if (Start->isZero()) break;
+ const SCEV *Zero = SE.getIntegerSCEV(0, Ty);
+ AddRecs.push_back(SE.getAddRecExpr(Zero,
+ A->getStepRecurrence(SE),
+ A->getLoop()));
+ if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) {
+ Ops[i] = Zero;
+ Ops.insert(Ops.end(), Add->op_begin(), Add->op_end());
+ e += Add->getNumOperands();
+ } else {
+ Ops[i] = Start;
+ }
+ }
+ if (!AddRecs.empty()) {
+ // Add the addrecs onto the end of the list.
+ Ops.insert(Ops.end(), AddRecs.begin(), AddRecs.end());
+ // Resort the operand list, moving any constants to the front.
+ SimplifyAddOperands(Ops, Ty, SE);
+ }
+}
+
+/// expandAddToGEP - Expand an addition expression with a pointer type into
+/// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps
+/// BasicAliasAnalysis and other passes analyze the result. See the rules
+/// for getelementptr vs. inttoptr in
+/// http://llvm.org/docs/LangRef.html#pointeraliasing
+/// for details.
+///
+/// Design note: The correctness of using getelmeentptr here depends on
+/// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as
+/// they may introduce pointer arithmetic which may not be safely converted
+/// into getelementptr.
///
/// Design note: It might seem desirable for this function to be more
/// loop-aware. If some of the indices are loop-invariant while others
/// loop-invariant portions of expressions, after considering what
/// can be folded using target addressing modes.
///
-Value *SCEVExpander::expandAddToGEP(const SCEV* const *op_begin,
- const SCEV* const *op_end,
+Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin,
+ const SCEV *const *op_end,
const PointerType *PTy,
const Type *Ty,
Value *V) {
const Type *ElTy = PTy->getElementType();
SmallVector<Value *, 4> GepIndices;
- SmallVector<const SCEV*, 8> Ops(op_begin, op_end);
+ SmallVector<const SCEV *, 8> Ops(op_begin, op_end);
bool AnyNonZeroIndices = false;
+ // Split AddRecs up into parts as either of the parts may be usable
+ // without the other.
+ SplitAddRecs(Ops, Ty, SE);
+
// Decend down the pointer's type and attempt to convert the other
// operands into GEP indices, at each level. The first index in a GEP
// indexes into the array implied by the pointer operand; the rest of
// the indices index into the element or field type selected by the
// preceding index.
for (;;) {
- APInt ElSize = APInt(SE.getTypeSizeInBits(Ty),
- ElTy->isSized() ? SE.TD->getTypeAllocSize(ElTy) : 0);
- SmallVector<const SCEV*, 8> NewOps;
- SmallVector<const SCEV*, 8> ScaledOps;
- for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
- // Split AddRecs up into parts as either of the parts may be usable
- // without the other.
- if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i]))
- if (!A->getStart()->isZero()) {
- const SCEV* Start = A->getStart();
- Ops.push_back(SE.getAddRecExpr(SE.getIntegerSCEV(0, A->getType()),
- A->getStepRecurrence(SE),
- A->getLoop()));
- Ops[i] = Start;
- ++e;
- }
- // If the scale size is not 0, attempt to factor out a scale.
- if (ElSize != 0) {
- const SCEV* Op = Ops[i];
- const SCEV* Remainder = SE.getIntegerSCEV(0, Op->getType());
- if (FactorOutConstant(Op, Remainder, ElSize, SE)) {
- ScaledOps.push_back(Op); // Op now has ElSize factored out.
- NewOps.push_back(Remainder);
- continue;
+ const SCEV *ElSize = SE.getAllocSizeExpr(ElTy);
+ // If the scale size is not 0, attempt to factor out a scale for
+ // array indexing.
+ SmallVector<const SCEV *, 8> ScaledOps;
+ if (ElTy->isSized() && !ElSize->isZero()) {
+ SmallVector<const SCEV *, 8> NewOps;
+ for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
+ const SCEV *Op = Ops[i];
+ const SCEV *Remainder = SE.getIntegerSCEV(0, Ty);
+ if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.TD)) {
+ // Op now has ElSize factored out.
+ ScaledOps.push_back(Op);
+ if (!Remainder->isZero())
+ NewOps.push_back(Remainder);
+ AnyNonZeroIndices = true;
+ } else {
+ // The operand was not divisible, so add it to the list of operands
+ // we'll scan next iteration.
+ NewOps.push_back(Ops[i]);
}
}
- // If the operand was not divisible, add it to the list of operands
- // we'll scan next iteration.
- NewOps.push_back(Ops[i]);
+ // If we made any changes, update Ops.
+ if (!ScaledOps.empty()) {
+ Ops = NewOps;
+ SimplifyAddOperands(Ops, Ty, SE);
+ }
}
- Ops = NewOps;
- AnyNonZeroIndices |= !ScaledOps.empty();
+
+ // Record the scaled array index for this level of the type. If
+ // we didn't find any operands that could be factored, tentatively
+ // assume that element zero was selected (since the zero offset
+ // would obviously be folded away).
Value *Scaled = ScaledOps.empty() ?
Constant::getNullValue(Ty) :
expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
GepIndices.push_back(Scaled);
// Collect struct field index operands.
- if (!Ops.empty())
- while (const StructType *STy = dyn_cast<StructType>(ElTy)) {
+ while (const StructType *STy = dyn_cast<StructType>(ElTy)) {
+ bool FoundFieldNo = false;
+ // An empty struct has no fields.
+ if (STy->getNumElements() == 0) break;
+ if (SE.TD) {
+ // With TargetData, field offsets are known. See if a constant offset
+ // falls within any of the struct fields.
+ if (Ops.empty()) break;
if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
if (SE.getTypeSizeInBits(C->getType()) <= 64) {
const StructLayout &SL = *SE.TD->getStructLayout(STy);
uint64_t FullOffset = C->getValue()->getZExtValue();
if (FullOffset < SL.getSizeInBytes()) {
unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
- GepIndices.push_back(ConstantInt::get(Type::Int32Ty, ElIdx));
+ GepIndices.push_back(
+ ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx));
ElTy = STy->getTypeAtIndex(ElIdx);
Ops[0] =
SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
AnyNonZeroIndices = true;
- continue;
+ FoundFieldNo = true;
}
}
- break;
+ } else {
+ // Without TargetData, just check for a SCEVFieldOffsetExpr of the
+ // appropriate struct type.
+ for (unsigned i = 0, e = Ops.size(); i != e; ++i)
+ if (const SCEVFieldOffsetExpr *FO =
+ dyn_cast<SCEVFieldOffsetExpr>(Ops[i]))
+ if (FO->getStructType() == STy) {
+ unsigned FieldNo = FO->getFieldNo();
+ GepIndices.push_back(
+ ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
+ FieldNo));
+ ElTy = STy->getTypeAtIndex(FieldNo);
+ Ops[i] = SE.getConstant(Ty, 0);
+ AnyNonZeroIndices = true;
+ FoundFieldNo = true;
+ break;
+ }
+ }
+ // If no struct field offsets were found, tentatively assume that
+ // field zero was selected (since the zero offset would obviously
+ // be folded away).
+ if (!FoundFieldNo) {
+ ElTy = STy->getTypeAtIndex(0u);
+ GepIndices.push_back(
+ Constant::getNullValue(Type::getInt32Ty(Ty->getContext())));
}
+ }
- if (const ArrayType *ATy = dyn_cast<ArrayType>(ElTy)) {
+ if (const ArrayType *ATy = dyn_cast<ArrayType>(ElTy))
ElTy = ATy->getElementType();
- continue;
- }
- break;
+ else
+ break;
}
// If none of the operands were convertable to proper GEP indices, cast
// the base to i8* and do an ugly getelementptr with that. It's still
// better than ptrtoint+arithmetic+inttoptr at least.
if (!AnyNonZeroIndices) {
+ // Cast the base to i8*.
V = InsertNoopCastOfTo(V,
- Type::Int8Ty->getPointerTo(PTy->getAddressSpace()));
+ Type::getInt8Ty(Ty->getContext())->getPointerTo(PTy->getAddressSpace()));
+
+ // Expand the operands for a plain byte offset.
Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
// Fold a GEP with constant operands.
// Do a quick scan to see if we have this GEP nearby. If so, reuse it.
unsigned ScanLimit = 6;
- BasicBlock::iterator BlockBegin = InsertPt->getParent()->begin();
- if (InsertPt != BlockBegin) {
- // Scanning starts from the last instruction before InsertPt.
- BasicBlock::iterator IP = InsertPt;
+ BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
+ // Scanning starts from the last instruction before the insertion point.
+ BasicBlock::iterator IP = Builder.GetInsertPoint();
+ if (IP != BlockBegin) {
--IP;
for (; ScanLimit; --IP, --ScanLimit) {
if (IP->getOpcode() == Instruction::GetElementPtr &&
}
}
- Value *GEP = GetElementPtrInst::Create(V, Idx, "scevgep", InsertPt);
+ // Emit a GEP.
+ Value *GEP = Builder.CreateGEP(V, Idx, "uglygep");
InsertedValues.insert(GEP);
return GEP;
}
- // Insert a pretty getelementptr.
- Value *GEP = GetElementPtrInst::Create(V,
- GepIndices.begin(),
- GepIndices.end(),
- "scevgep", InsertPt);
+ // Insert a pretty getelementptr. Note that this GEP is not marked inbounds,
+ // because ScalarEvolution may have changed the address arithmetic to
+ // compute a value which is beyond the end of the allocated object.
+ Value *GEP = Builder.CreateGEP(V,
+ GepIndices.begin(),
+ GepIndices.end(),
+ "scevgep");
Ops.push_back(SE.getUnknown(GEP));
InsertedValues.insert(GEP);
return expand(SE.getAddExpr(Ops));
// Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
// comments on expandAddToGEP for details.
- if (SE.TD)
- if (const PointerType *PTy = dyn_cast<PointerType>(V->getType())) {
- const SmallVectorImpl<const SCEV*> &Ops = S->getOperands();
- return expandAddToGEP(&Ops[0], &Ops[Ops.size() - 1],
- PTy, Ty, V);
- }
+ if (const PointerType *PTy = dyn_cast<PointerType>(V->getType())) {
+ const SmallVectorImpl<const SCEV *> &Ops = S->getOperands();
+ return expandAddToGEP(&Ops[0], &Ops[Ops.size() - 1], PTy, Ty, V);
+ }
V = InsertNoopCastOfTo(V, Ty);
// Emit a bunch of add instructions
for (int i = S->getNumOperands()-2; i >= 0; --i) {
Value *W = expandCodeFor(S->getOperand(i), Ty);
- V = InsertBinop(Instruction::Add, V, W, InsertPt);
+ V = InsertBinop(Instruction::Add, V, W);
}
return V;
}
// Emit a bunch of multiply instructions
for (; i >= FirstOp; --i) {
Value *W = expandCodeFor(S->getOperand(i), Ty);
- V = InsertBinop(Instruction::Mul, V, W, InsertPt);
+ V = InsertBinop(Instruction::Mul, V, W);
}
// -1 * ... ---> 0 - ...
if (FirstOp == 1)
- V = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), V, InsertPt);
+ V = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), V);
return V;
}
const APInt &RHS = SC->getValue()->getValue();
if (RHS.isPowerOf2())
return InsertBinop(Instruction::LShr, LHS,
- ConstantInt::get(Ty, RHS.logBase2()),
- InsertPt);
+ ConstantInt::get(Ty, RHS.logBase2()));
}
Value *RHS = expandCodeFor(S->getRHS(), Ty);
- return InsertBinop(Instruction::UDiv, LHS, RHS, InsertPt);
+ return InsertBinop(Instruction::UDiv, LHS, RHS);
}
/// Move parts of Base into Rest to leave Base with the minimal
/// expression that provides a pointer operand suitable for a
/// GEP expansion.
-static void ExposePointerBase(const SCEV* &Base, const SCEV* &Rest,
+static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest,
ScalarEvolution &SE) {
while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
Base = A->getStart();
}
if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
Base = A->getOperand(A->getNumOperands()-1);
- SmallVector<const SCEV*, 8> NewAddOps(A->op_begin(), A->op_end());
+ SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end());
NewAddOps.back() = Rest;
Rest = SE.getAddExpr(NewAddOps);
ExposePointerBase(Base, Rest, SE);
if (CanonicalIV &&
SE.getTypeSizeInBits(CanonicalIV->getType()) >
SE.getTypeSizeInBits(Ty)) {
- const SCEV* Start = SE.getAnyExtendExpr(S->getStart(),
+ const SCEV *Start = SE.getAnyExtendExpr(S->getStart(),
+ CanonicalIV->getType());
+ const SCEV *Step = SE.getAnyExtendExpr(S->getStepRecurrence(SE),
CanonicalIV->getType());
- const SCEV* Step = SE.getAnyExtendExpr(S->getStepRecurrence(SE),
- CanonicalIV->getType());
Value *V = expand(SE.getAddRecExpr(Start, Step, S->getLoop()));
- BasicBlock::iterator SaveInsertPt = getInsertionPoint();
+ BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
+ BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
BasicBlock::iterator NewInsertPt =
next(BasicBlock::iterator(cast<Instruction>(V)));
while (isa<PHINode>(NewInsertPt)) ++NewInsertPt;
V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0,
NewInsertPt);
- setInsertionPoint(SaveInsertPt);
+ Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
return V;
}
// {X,+,F} --> X + {0,+,F}
if (!S->getStart()->isZero()) {
- const SmallVectorImpl<const SCEV*> &SOperands = S->getOperands();
- SmallVector<const SCEV*, 4> NewOps(SOperands.begin(), SOperands.end());
+ const SmallVectorImpl<const SCEV *> &SOperands = S->getOperands();
+ SmallVector<const SCEV *, 4> NewOps(SOperands.begin(), SOperands.end());
NewOps[0] = SE.getIntegerSCEV(0, Ty);
- const SCEV* Rest = SE.getAddRecExpr(NewOps, L);
+ const SCEV *Rest = SE.getAddRecExpr(NewOps, L);
// Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
// comments on expandAddToGEP for details.
- if (SE.TD) {
- const SCEV* Base = S->getStart();
- const SCEV* RestArray[1] = { Rest };
- // Dig into the expression to find the pointer base for a GEP.
- ExposePointerBase(Base, RestArray[0], SE);
- // If we found a pointer, expand the AddRec with a GEP.
- if (const PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
- // Make sure the Base isn't something exotic, such as a multiplied
- // or divided pointer value. In those cases, the result type isn't
- // actually a pointer type.
- if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
- Value *StartV = expand(Base);
- assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
- return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
- }
+ const SCEV *Base = S->getStart();
+ const SCEV *RestArray[1] = { Rest };
+ // Dig into the expression to find the pointer base for a GEP.
+ ExposePointerBase(Base, RestArray[0], SE);
+ // If we found a pointer, expand the AddRec with a GEP.
+ if (const PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
+ // Make sure the Base isn't something exotic, such as a multiplied
+ // or divided pointer value. In those cases, the result type isn't
+ // actually a pointer type.
+ if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
+ Value *StartV = expand(Base);
+ assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
+ return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
}
}
// Create and insert the PHI node for the induction variable in the
// specified loop.
BasicBlock *Header = L->getHeader();
+ BasicBlock *Preheader = L->getLoopPreheader();
PHINode *PN = PHINode::Create(Ty, "indvar", Header->begin());
InsertedValues.insert(PN);
- PN->addIncoming(Constant::getNullValue(Ty), L->getLoopPreheader());
+ PN->addIncoming(Constant::getNullValue(Ty), Preheader);
pred_iterator HPI = pred_begin(Header);
assert(HPI != pred_end(Header) && "Loop with zero preds???");
InsertedValues.insert(Add);
pred_iterator PI = pred_begin(Header);
- if (*PI == L->getLoopPreheader())
+ if (*PI == Preheader)
++PI;
PN->addIncoming(Add, *PI);
return PN;
// folders, then expandCodeFor the closed form. This allows the folders to
// simplify the expression without having to build a bunch of special code
// into this folder.
- const SCEV* IH = SE.getUnknown(I); // Get I as a "symbolic" SCEV.
+ const SCEV *IH = SE.getUnknown(I); // Get I as a "symbolic" SCEV.
// Promote S up to the canonical IV type, if the cast is foldable.
- const SCEV* NewS = S;
- const SCEV* Ext = SE.getNoopOrAnyExtend(S, I->getType());
+ const SCEV *NewS = S;
+ const SCEV *Ext = SE.getNoopOrAnyExtend(S, I->getType());
if (isa<SCEVAddRecExpr>(Ext))
NewS = Ext;
- const SCEV* V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
+ const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
//cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
// Truncate the result down to the original type, if needed.
- const SCEV* T = SE.getTruncateOrNoop(V, Ty);
+ const SCEV *T = SE.getTruncateOrNoop(V, Ty);
return expand(T);
}
const Type *Ty = SE.getEffectiveSCEVType(S->getType());
Value *V = expandCodeFor(S->getOperand(),
SE.getEffectiveSCEVType(S->getOperand()->getType()));
- Instruction *I = new TruncInst(V, Ty, "tmp.", InsertPt);
+ Value *I = Builder.CreateTrunc(V, Ty, "tmp");
InsertedValues.insert(I);
return I;
}
const Type *Ty = SE.getEffectiveSCEVType(S->getType());
Value *V = expandCodeFor(S->getOperand(),
SE.getEffectiveSCEVType(S->getOperand()->getType()));
- Instruction *I = new ZExtInst(V, Ty, "tmp.", InsertPt);
+ Value *I = Builder.CreateZExt(V, Ty, "tmp");
InsertedValues.insert(I);
return I;
}
const Type *Ty = SE.getEffectiveSCEVType(S->getType());
Value *V = expandCodeFor(S->getOperand(),
SE.getEffectiveSCEVType(S->getOperand()->getType()));
- Instruction *I = new SExtInst(V, Ty, "tmp.", InsertPt);
+ Value *I = Builder.CreateSExt(V, Ty, "tmp");
InsertedValues.insert(I);
return I;
}
Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
- const Type *Ty = SE.getEffectiveSCEVType(S->getType());
- Value *LHS = expandCodeFor(S->getOperand(0), Ty);
- for (unsigned i = 1; i < S->getNumOperands(); ++i) {
+ Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
+ const Type *Ty = LHS->getType();
+ for (int i = S->getNumOperands()-2; i >= 0; --i) {
+ // In the case of mixed integer and pointer types, do the
+ // rest of the comparisons as integer.
+ if (S->getOperand(i)->getType() != Ty) {
+ Ty = SE.getEffectiveSCEVType(Ty);
+ LHS = InsertNoopCastOfTo(LHS, Ty);
+ }
Value *RHS = expandCodeFor(S->getOperand(i), Ty);
- Instruction *ICmp =
- new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS, "tmp", InsertPt);
+ Value *ICmp = Builder.CreateICmpSGT(LHS, RHS, "tmp");
InsertedValues.insert(ICmp);
- Instruction *Sel = SelectInst::Create(ICmp, LHS, RHS, "smax", InsertPt);
+ Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
InsertedValues.insert(Sel);
LHS = Sel;
}
+ // In the case of mixed integer and pointer types, cast the
+ // final result back to the pointer type.
+ if (LHS->getType() != S->getType())
+ LHS = InsertNoopCastOfTo(LHS, S->getType());
return LHS;
}
Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
- const Type *Ty = SE.getEffectiveSCEVType(S->getType());
- Value *LHS = expandCodeFor(S->getOperand(0), Ty);
- for (unsigned i = 1; i < S->getNumOperands(); ++i) {
+ Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
+ const Type *Ty = LHS->getType();
+ for (int i = S->getNumOperands()-2; i >= 0; --i) {
+ // In the case of mixed integer and pointer types, do the
+ // rest of the comparisons as integer.
+ if (S->getOperand(i)->getType() != Ty) {
+ Ty = SE.getEffectiveSCEVType(Ty);
+ LHS = InsertNoopCastOfTo(LHS, Ty);
+ }
Value *RHS = expandCodeFor(S->getOperand(i), Ty);
- Instruction *ICmp =
- new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS, "tmp", InsertPt);
+ Value *ICmp = Builder.CreateICmpUGT(LHS, RHS, "tmp");
InsertedValues.insert(ICmp);
- Instruction *Sel = SelectInst::Create(ICmp, LHS, RHS, "umax", InsertPt);
+ Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
InsertedValues.insert(Sel);
LHS = Sel;
}
+ // In the case of mixed integer and pointer types, cast the
+ // final result back to the pointer type.
+ if (LHS->getType() != S->getType())
+ LHS = InsertNoopCastOfTo(LHS, S->getType());
return LHS;
}
-Value *SCEVExpander::expandCodeFor(const SCEV* SH, const Type *Ty) {
+Value *SCEVExpander::visitFieldOffsetExpr(const SCEVFieldOffsetExpr *S) {
+ return ConstantExpr::getOffsetOf(S->getStructType(), S->getFieldNo());
+}
+
+Value *SCEVExpander::visitAllocSizeExpr(const SCEVAllocSizeExpr *S) {
+ return ConstantExpr::getSizeOf(S->getAllocType());
+}
+
+Value *SCEVExpander::expandCodeFor(const SCEV *SH, const Type *Ty) {
// Expand the code for this SCEV.
Value *V = expand(SH);
if (Ty) {
}
Value *SCEVExpander::expand(const SCEV *S) {
- // Check to see if we already expanded this.
- std::map<const SCEV*, AssertingVH<Value> >::iterator I =
- InsertedExpressions.find(S);
- if (I != InsertedExpressions.end())
- return I->second;
-
// Compute an insertion point for this SCEV object. Hoist the instructions
// as far out in the loop nest as possible.
- BasicBlock::iterator InsertPt = getInsertionPoint();
- BasicBlock::iterator SaveInsertPt = InsertPt;
- for (Loop *L = SE.LI->getLoopFor(InsertPt->getParent()); ;
+ Instruction *InsertPt = Builder.GetInsertPoint();
+ for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ;
L = L->getParentLoop())
if (S->isLoopInvariant(L)) {
if (!L) break;
// there) so that it is guaranteed to dominate any user inside the loop.
if (L && S->hasComputableLoopEvolution(L))
InsertPt = L->getHeader()->getFirstNonPHI();
- while (isInsertedInstruction(InsertPt)) ++InsertPt;
+ while (isInsertedInstruction(InsertPt))
+ InsertPt = next(BasicBlock::iterator(InsertPt));
break;
}
- setInsertionPoint(InsertPt);
+ // Check to see if we already expanded this here.
+ std::map<std::pair<const SCEV *, Instruction *>,
+ AssertingVH<Value> >::iterator I =
+ InsertedExpressions.find(std::make_pair(S, InsertPt));
+ if (I != InsertedExpressions.end())
+ return I->second;
+
+ BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
+ BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
+ Builder.SetInsertPoint(InsertPt->getParent(), InsertPt);
+
+ // Expand the expression into instructions.
Value *V = visit(S);
- setInsertionPoint(SaveInsertPt);
- InsertedExpressions[S] = V;
+ // Remember the expanded value for this SCEV at this location.
+ InsertedExpressions[std::make_pair(S, InsertPt)] = V;
+
+ Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
return V;
}
SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
const Type *Ty) {
assert(Ty->isInteger() && "Can only insert integer induction variables!");
- const SCEV* H = SE.getAddRecExpr(SE.getIntegerSCEV(0, Ty),
+ const SCEV *H = SE.getAddRecExpr(SE.getIntegerSCEV(0, Ty),
SE.getIntegerSCEV(1, Ty), L);
- BasicBlock::iterator SaveInsertPt = getInsertionPoint();
+ BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
+ BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
Value *V = expandCodeFor(H, 0, L->getHeader()->begin());
- setInsertionPoint(SaveInsertPt);
+ if (SaveInsertBB)
+ Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
return V;
}