#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 ||
// 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) {
+ 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) !=
+ if (BasicBlock::iterator(CI) !=
A->getParent()->getEntryBlock().begin()) {
- // If the CastInst is the insert point, change the insert point.
- if (CI == InsertPt) ++InsertPt;
- // Splice the cast at the beginning of the entry block.
- CI->moveBefore(A->getParent()->getEntryBlock().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(Op, V, Ty, "",
+ A->getParent()->getEntryBlock().begin());
+ NewCI->takeName(CI);
+ CI->replaceAllUsesWith(NewCI);
+ return 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();
while (isa<PHINode>(It)) ++It;
if (It != BasicBlock::iterator(CI)) {
- // If the CastInst is the insert point, change the insert point.
- if (CI == InsertPt) ++InsertPt;
- // Splice the cast immediately after the operand in question.
- CI->moveBefore(It);
+ // 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(Op, V, Ty, "", It);
+ NewCI->takeName(CI);
+ CI->replaceAllUsesWith(NewCI);
+ return NewCI;
}
return CI;
}
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(SCEVHandle &S,
- SCEVHandle &Remainder,
+static bool FactorOutConstant(const SCEV *&S,
+ const SCEV *&Remainder,
const APInt &Factor,
ScalarEvolution &SE) {
// Everything is divisible by one.
// 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));
+ ConstantInt::get(SE.getContext(), 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()) {
- SCEVHandle Div = SE.getConstant(CI);
+ const SCEV *Div = SE.getConstant(CI);
S = Div;
Remainder =
SE.getAddExpr(Remainder,
if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S))
if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
if (!C->getValue()->getValue().srem(Factor)) {
- std::vector<SCEVHandle> NewMulOps(M->getOperands());
+ 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);
// In an AddRec, check if both start and step are divisible.
if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
- SCEVHandle Step = A->getStepRecurrence(SE);
- SCEVHandle StepRem = SE.getIntegerSCEV(0, Step->getType());
+ const SCEV *Step = A->getStepRecurrence(SE);
+ const SCEV *StepRem = SE.getIntegerSCEV(0, Step->getType());
if (!FactorOutConstant(Step, StepRem, Factor, SE))
return false;
if (!StepRem->isZero())
return false;
- SCEVHandle Start = A->getStart();
+ const SCEV *Start = A->getStart();
if (!FactorOutConstant(Start, Remainder, Factor, SE))
return false;
S = SE.getAddRecExpr(Start, Step, A->getLoop());
/// 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.
+/// BasicAliasAnalysis analyze the result.
+///
+/// Design note: This 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 SCEVHandle *op_begin,
- const SCEVHandle *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;
- std::vector<SCEVHandle> Ops(op_begin, op_end);
+ SmallVector<const SCEV *, 8> Ops(op_begin, op_end);
bool AnyNonZeroIndices = false;
// Decend down the pointer's type and attempt to convert the other
for (;;) {
APInt ElSize = APInt(SE.getTypeSizeInBits(Ty),
ElTy->isSized() ? SE.TD->getTypeAllocSize(ElTy) : 0);
- std::vector<SCEVHandle> NewOps;
- std::vector<SCEVHandle> ScaledOps;
+ 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()) {
- SCEVHandle Start = A->getStart();
+ const SCEV *Start = A->getStart();
Ops.push_back(SE.getAddRecExpr(SE.getIntegerSCEV(0, A->getType()),
A->getStepRecurrence(SE),
A->getLoop()));
}
// If the scale size is not 0, attempt to factor out a scale.
if (ElSize != 0) {
- SCEVHandle Op = Ops[i];
- SCEVHandle Remainder = SE.getIntegerSCEV(0, Op->getType());
+ 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);
GepIndices.push_back(ConstantInt::get(Type::Int32Ty, ElIdx));
ElTy = STy->getTypeAtIndex(ElIdx);
Ops[0] =
- SE.getConstant(ConstantInt::get(Ty,
- FullOffset -
- SL.getElementOffset(ElIdx)));
+ SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
AnyNonZeroIndices = true;
continue;
}
// 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);
+ Value *GEP = Builder.CreateGEP(V, Idx, "scevgep");
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));
// comments on expandAddToGEP for details.
if (SE.TD)
if (const PointerType *PTy = dyn_cast<PointerType>(V->getType())) {
- const std::vector<SCEVHandle> &Ops = S->getOperands();
- return expandAddToGEP(&Ops[0], &Ops[Ops.size() - 1],
- PTy, Ty, V);
+ 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(SCEVHandle &Base, SCEVHandle &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);
- std::vector<SCEVHandle> 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)) {
- SCEVHandle Start = SE.getAnyExtendExpr(S->getStart(),
+ const SCEV *Start = SE.getAnyExtendExpr(S->getStart(),
+ CanonicalIV->getType());
+ const SCEV *Step = SE.getAnyExtendExpr(S->getStepRecurrence(SE),
CanonicalIV->getType());
- SCEVHandle 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()) {
- std::vector<SCEVHandle> NewOps(S->getOperands());
+ const SmallVectorImpl<const SCEV *> &SOperands = S->getOperands();
+ SmallVector<const SCEV *, 4> NewOps(SOperands.begin(), SOperands.end());
NewOps[0] = SE.getIntegerSCEV(0, Ty);
- SCEVHandle 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) {
- SCEVHandle Base = S->getStart();
- SCEVHandle RestArray[1] = { Rest };
+ 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.
}
}
- Value *RestV = expand(Rest);
- return expand(SE.getAddExpr(S->getStart(), SE.getUnknown(RestV)));
+ // Just do a normal add. Pre-expand the operands to suppress folding.
+ return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())),
+ SE.getUnknown(expand(Rest))));
}
// {0,+,1} --> Insert a canonical induction variable into the loop!
// 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;
getOrInsertCanonicalInductionVariable(L, Ty);
// If this is a simple linear addrec, emit it now as a special case.
- if (S->isAffine()) { // {0,+,F} --> i*F
- Value *F = expandCodeFor(S->getOperand(1), Ty);
-
- // If the insert point is directly inside of the loop, emit the multiply at
- // the insert point. Otherwise, L is a loop that is a parent of the insert
- // point loop. If we can, move the multiply to the outer most loop that it
- // is safe to be in.
- BasicBlock::iterator MulInsertPt = getInsertionPoint();
- Loop *InsertPtLoop = SE.LI->getLoopFor(MulInsertPt->getParent());
- if (InsertPtLoop != L && InsertPtLoop &&
- L->contains(InsertPtLoop->getHeader())) {
- do {
- // If we cannot hoist the multiply out of this loop, don't.
- if (!InsertPtLoop->isLoopInvariant(F)) break;
-
- BasicBlock *InsertPtLoopPH = InsertPtLoop->getLoopPreheader();
-
- // If this loop hasn't got a preheader, we aren't able to hoist the
- // multiply.
- if (!InsertPtLoopPH)
- break;
-
- // Otherwise, move the insert point to the preheader.
- MulInsertPt = InsertPtLoopPH->getTerminator();
- InsertPtLoop = InsertPtLoop->getParentLoop();
- } while (InsertPtLoop != L);
- }
-
- return InsertBinop(Instruction::Mul, I, F, MulInsertPt);
- }
+ if (S->isAffine()) // {0,+,F} --> i*F
+ return
+ expand(SE.getTruncateOrNoop(
+ SE.getMulExpr(SE.getUnknown(I),
+ SE.getNoopOrAnyExtend(S->getOperand(1),
+ I->getType())),
+ Ty));
// If this is a chain of recurrences, turn it into a closed form, using the
// 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.
- SCEVHandle 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.
- SCEVHandle NewS = S;
- SCEVHandle Ext = SE.getNoopOrAnyExtend(S, I->getType());
+ const SCEV *NewS = S;
+ const SCEV *Ext = SE.getNoopOrAnyExtend(S, I->getType());
if (isa<SCEVAddRecExpr>(Ext))
NewS = Ext;
- SCEVHandle 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.
- SCEVHandle T = SE.getTruncateOrNoop(V, Ty);
- return expand(V);
+ const SCEV *T = SE.getTruncateOrNoop(V, Ty);
+ return expand(T);
}
Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
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(SCEVHandle SH, const Type *Ty) {
+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<SCEVHandle, AssertingVH<Value> >::iterator I =
- InsertedExpressions.find(S);
+ // Compute an insertion point for this SCEV object. Hoist the instructions
+ // as far out in the loop nest as possible.
+ Instruction *InsertPt = Builder.GetInsertPoint();
+ for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ;
+ L = L->getParentLoop())
+ if (S->isLoopInvariant(L)) {
+ if (!L) break;
+ if (BasicBlock *Preheader = L->getLoopPreheader())
+ InsertPt = Preheader->getTerminator();
+ } else {
+ // If the SCEV is computable at this level, insert it into the header
+ // after the PHIs (and after any other instructions that we've inserted
+ // 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 = next(BasicBlock::iterator(InsertPt));
+ break;
+ }
+
+ // 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);
- 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!");
- SCEVHandle H = SE.getAddRecExpr(SE.getIntegerSCEV(0, Ty),
- SE.getIntegerSCEV(1, Ty), L);
- return expand(H);
+ const SCEV *H = SE.getAddRecExpr(SE.getIntegerSCEV(0, Ty),
+ SE.getIntegerSCEV(1, Ty), L);
+ BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
+ BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
+ Value *V = expandCodeFor(H, 0, L->getHeader()->begin());
+ if (SaveInsertBB)
+ Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
+ return V;
}