#include "llvm/Analysis/ScalarEvolutionExpander.h"
#include "llvm/Analysis/LoopInfo.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
if (opcode == Instruction::BitCast && V->getType() == Ty)
return V;
+ // Short-circuit unnecessary inttoptr<->ptrtoint casts.
+ if ((opcode == Instruction::PtrToInt || opcode == Instruction::IntToPtr) &&
+ SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) {
+ if (CastInst *CI = dyn_cast<CastInst>(V))
+ if ((CI->getOpcode() == Instruction::PtrToInt ||
+ CI->getOpcode() == Instruction::IntToPtr) &&
+ SE.getTypeSizeInBits(CI->getType()) ==
+ SE.getTypeSizeInBits(CI->getOperand(0)->getType()))
+ return CI->getOperand(0);
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
+ if ((CE->getOpcode() == Instruction::PtrToInt ||
+ CE->getOpcode() == Instruction::IntToPtr) &&
+ SE.getTypeSizeInBits(CE->getType()) ==
+ SE.getTypeSizeInBits(CE->getOperand(0)->getType()))
+ return CE->getOperand(0);
+ }
+
// FIXME: keep track of the cast instruction.
if (Constant *C = dyn_cast<Constant>(V))
return ConstantExpr::getCast(opcode, C, Ty);
// If the cast isn't the first instruction of the function, move it.
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());
}
return CI;
}
}
- return CastInst::Create(opcode, V, Ty, V->getName(),
- A->getParent()->getEntryBlock().begin());
+ Instruction *I = CastInst::Create(opcode, V, Ty, V->getName(),
+ A->getParent()->getEntryBlock().begin());
+ InsertedValues.insert(I);
+ return I;
}
Instruction *I = cast<Instruction>(V);
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);
}
if (InvokeInst *II = dyn_cast<InvokeInst>(I))
IP = II->getNormalDest()->begin();
while (isa<PHINode>(IP)) ++IP;
- return CastInst::Create(opcode, V, Ty, V->getName(), IP);
+ Instruction *CI = CastInst::Create(opcode, 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, Instruction *InsertPt) {
+ Value *RHS, BasicBlock::iterator InsertPt) {
// Fold a binop with constant operands.
if (Constant *CLHS = dyn_cast<Constant>(LHS))
if (Constant *CRHS = dyn_cast<Constant>(RHS))
BasicBlock::iterator IP = InsertPt;
--IP;
for (; ScanLimit; --IP, --ScanLimit) {
- if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(IP))
- if (BinOp->getOpcode() == Opcode && BinOp->getOperand(0) == LHS &&
- BinOp->getOperand(1) == RHS)
- return BinOp;
+ if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
+ IP->getOperand(1) == RHS)
+ return IP;
if (IP == BlockBegin) break;
}
}
// If we haven't found this binop, insert it.
- return BinaryOperator::Create(Opcode, LHS, RHS, "tmp", InsertPt);
+ Instruction *BO = BinaryOperator::Create(Opcode, LHS, RHS, "tmp", InsertPt);
+ InsertedValues.insert(BO);
+ return BO;
+}
+
+/// FactorOutConstant - Test if S is divisible by Factor, using signed
+/// division. If so, update S with Factor divided out and return true.
+/// S need not be evenly divisble if a reasonable remainder can be
+/// computed.
+/// 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,
+ const APInt &Factor,
+ ScalarEvolution &SE) {
+ // Everything is divisible by one.
+ if (Factor == 1)
+ 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()) {
+ SCEVHandle Div = SE.getConstant(CI);
+ S = Div;
+ Remainder =
+ SE.getAddExpr(Remainder,
+ SE.getConstant(C->getValue()->getValue().srem(Factor)));
+ 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<SCEVHandle> &MOperands = M->getOperands();
+ SmallVector<SCEVHandle, 4> NewMulOps(MOperands.begin(), MOperands.end());
+ NewMulOps[0] =
+ SE.getConstant(C->getValue()->getValue().sdiv(Factor));
+ 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)) {
+ SCEVHandle Step = A->getStepRecurrence(SE);
+ SCEVHandle StepRem = SE.getIntegerSCEV(0, Step->getType());
+ if (!FactorOutConstant(Step, StepRem, Factor, SE))
+ return false;
+ if (!StepRem->isZero())
+ return false;
+ SCEVHandle Start = A->getStart();
+ if (!FactorOutConstant(Start, Remainder, Factor, SE))
+ 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.
+///
+/// Design note: It might seem desirable for this function to be more
+/// loop-aware. If some of the indices are loop-invariant while others
+/// aren't, it might seem desirable to emit multiple GEPs, keeping the
+/// loop-invariant portions of the overall computation outside the loop.
+/// However, there are a few reasons this is not done here. Hoisting simple
+/// arithmetic is a low-level optimization that often isn't very
+/// important until late in the optimization process. In fact, passes
+/// like InstructionCombining will combine GEPs, even if it means
+/// pushing loop-invariant computation down into loops, so even if the
+/// GEPs were split here, the work would quickly be undone. The
+/// LoopStrengthReduction pass, which is usually run quite late (and
+/// after the last InstructionCombining pass), takes care of hoisting
+/// 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,
+ const PointerType *PTy,
+ const Type *Ty,
+ Value *V) {
+ const Type *ElTy = PTy->getElementType();
+ SmallVector<Value *, 4> GepIndices;
+ SmallVector<SCEVHandle, 8> Ops(op_begin, op_end);
+ bool AnyNonZeroIndices = false;
+
+ // 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<SCEVHandle, 8> NewOps;
+ SmallVector<SCEVHandle, 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();
+ 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) {
+ SCEVHandle Op = Ops[i];
+ SCEVHandle 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;
+ }
+ }
+ // If the operand was not divisible, add it to the list of operands
+ // we'll scan next iteration.
+ NewOps.push_back(Ops[i]);
+ }
+ Ops = NewOps;
+ AnyNonZeroIndices |= !ScaledOps.empty();
+ 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)) {
+ 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));
+ ElTy = STy->getTypeAtIndex(ElIdx);
+ Ops[0] =
+ SE.getConstant(ConstantInt::get(Ty,
+ FullOffset -
+ SL.getElementOffset(ElIdx)));
+ AnyNonZeroIndices = true;
+ continue;
+ }
+ }
+ break;
+ }
+
+ if (const ArrayType *ATy = dyn_cast<ArrayType>(ElTy)) {
+ ElTy = ATy->getElementType();
+ continue;
+ }
+ 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) {
+ V = InsertNoopCastOfTo(V,
+ Type::Int8Ty->getPointerTo(PTy->getAddressSpace()));
+ Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
+
+ // Fold a GEP with constant operands.
+ if (Constant *CLHS = dyn_cast<Constant>(V))
+ if (Constant *CRHS = dyn_cast<Constant>(Idx))
+ return ConstantExpr::getGetElementPtr(CLHS, &CRHS, 1);
+
+ // 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;
+ --IP;
+ for (; ScanLimit; --IP, --ScanLimit) {
+ if (IP->getOpcode() == Instruction::GetElementPtr &&
+ IP->getOperand(0) == V && IP->getOperand(1) == Idx)
+ return IP;
+ if (IP == BlockBegin) break;
+ }
+ }
+
+ Value *GEP = GetElementPtrInst::Create(V, Idx, "scevgep", InsertPt);
+ InsertedValues.insert(GEP);
+ return GEP;
+ }
+
+ // Insert a pretty getelementptr.
+ Value *GEP = GetElementPtrInst::Create(V,
+ GepIndices.begin(),
+ GepIndices.end(),
+ "scevgep", InsertPt);
+ Ops.push_back(SE.getUnknown(GEP));
+ InsertedValues.insert(GEP);
+ return expand(SE.getAddExpr(Ops));
}
-Value *SCEVExpander::visitAddExpr(SCEVAddExpr *S) {
- const Type *Ty = S->getType();
- if (isa<PointerType>(Ty)) Ty = TD.getIntPtrType();
+Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
+ const Type *Ty = SE.getEffectiveSCEVType(S->getType());
Value *V = expand(S->getOperand(S->getNumOperands()-1));
- V = InsertCastOfTo(CastInst::getCastOpcode(V, false, Ty, false), V, Ty);
+
+ // 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<SCEVHandle> &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 = expand(S->getOperand(i));
- W = InsertCastOfTo(CastInst::getCastOpcode(W, false, Ty, false), W, Ty);
+ Value *W = expandCodeFor(S->getOperand(i), Ty);
V = InsertBinop(Instruction::Add, V, W, InsertPt);
}
return V;
}
-
-Value *SCEVExpander::visitMulExpr(SCEVMulExpr *S) {
- const Type *Ty = S->getType();
- if (isa<PointerType>(Ty)) Ty = TD.getIntPtrType();
+
+Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
+ const Type *Ty = SE.getEffectiveSCEVType(S->getType());
int FirstOp = 0; // Set if we should emit a subtract.
- if (SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getOperand(0)))
+ if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getOperand(0)))
if (SC->getValue()->isAllOnesValue())
FirstOp = 1;
int i = S->getNumOperands()-2;
- Value *V = expand(S->getOperand(i+1));
- V = InsertCastOfTo(CastInst::getCastOpcode(V, false, Ty, false), V, Ty);
+ Value *V = expandCodeFor(S->getOperand(i+1), Ty);
// Emit a bunch of multiply instructions
for (; i >= FirstOp; --i) {
- Value *W = expand(S->getOperand(i));
- W = InsertCastOfTo(CastInst::getCastOpcode(W, false, Ty, false), W, Ty);
+ Value *W = expandCodeFor(S->getOperand(i), Ty);
V = InsertBinop(Instruction::Mul, V, W, InsertPt);
}
return V;
}
-Value *SCEVExpander::visitUDivExpr(SCEVUDivExpr *S) {
- const Type *Ty = S->getType();
- if (isa<PointerType>(Ty)) Ty = TD.getIntPtrType();
+Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
+ const Type *Ty = SE.getEffectiveSCEVType(S->getType());
- Value *LHS = expand(S->getLHS());
- LHS = InsertCastOfTo(CastInst::getCastOpcode(LHS, false, Ty, false), LHS, Ty);
- if (SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
+ Value *LHS = expandCodeFor(S->getLHS(), Ty);
+ if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
const APInt &RHS = SC->getValue()->getValue();
if (RHS.isPowerOf2())
return InsertBinop(Instruction::LShr, LHS,
InsertPt);
}
- Value *RHS = expand(S->getRHS());
- RHS = InsertCastOfTo(CastInst::getCastOpcode(RHS, false, Ty, false), RHS, Ty);
+ Value *RHS = expandCodeFor(S->getRHS(), Ty);
return InsertBinop(Instruction::UDiv, LHS, RHS, InsertPt);
}
-Value *SCEVExpander::visitAddRecExpr(SCEVAddRecExpr *S) {
- const Type *Ty = S->getType();
+/// 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,
+ ScalarEvolution &SE) {
+ while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
+ Base = A->getStart();
+ Rest = SE.getAddExpr(Rest,
+ SE.getAddRecExpr(SE.getIntegerSCEV(0, A->getType()),
+ A->getStepRecurrence(SE),
+ A->getLoop()));
+ }
+ if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
+ Base = A->getOperand(A->getNumOperands()-1);
+ SmallVector<SCEVHandle, 8> NewAddOps(A->op_begin(), A->op_end());
+ NewAddOps.back() = Rest;
+ Rest = SE.getAddExpr(NewAddOps);
+ ExposePointerBase(Base, Rest, SE);
+ }
+}
+
+Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
+ const Type *Ty = SE.getEffectiveSCEVType(S->getType());
const Loop *L = S->getLoop();
- // We cannot yet do fp recurrences, e.g. the xform of {X,+,F} --> X+{0,+,F}
- assert((Ty->isInteger() || isa<PointerType>(Ty)) &&
- "Cannot expand fp recurrences yet!");
+
+ // First check for an existing canonical IV in a suitable type.
+ PHINode *CanonicalIV = 0;
+ if (PHINode *PN = L->getCanonicalInductionVariable())
+ if (SE.isSCEVable(PN->getType()) &&
+ isa<IntegerType>(SE.getEffectiveSCEVType(PN->getType())) &&
+ SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
+ CanonicalIV = PN;
+
+ // Rewrite an AddRec in terms of the canonical induction variable, if
+ // its type is more narrow.
+ if (CanonicalIV &&
+ SE.getTypeSizeInBits(CanonicalIV->getType()) >
+ SE.getTypeSizeInBits(Ty)) {
+ SCEVHandle Start = SE.getAnyExtendExpr(S->getStart(),
+ 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::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);
+ return V;
+ }
// {X,+,F} --> X + {0,+,F}
if (!S->getStart()->isZero()) {
- Value *Start = expand(S->getStart());
- if (isa<PointerType>(Start->getType()))
- Start = InsertCastOfTo(Instruction::PtrToInt, Start, TD.getIntPtrType());
- std::vector<SCEVHandle> NewOps(S->op_begin(), S->op_end());
+ const SmallVectorImpl<SCEVHandle> &SOperands = S->getOperands();
+ SmallVector<SCEVHandle, 4> NewOps(SOperands.begin(), SOperands.end());
NewOps[0] = SE.getIntegerSCEV(0, Ty);
- Value *Rest = expand(SE.getAddRecExpr(NewOps, L));
- if (isa<PointerType>(Rest->getType()))
- Rest = InsertCastOfTo(Instruction::PtrToInt, Rest, TD.getIntPtrType());
+ SCEVHandle 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 };
+ // 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);
+ }
+ }
+ }
- // FIXME: look for an existing add to use.
- return InsertBinop(Instruction::Add, Rest, Start, InsertPt);
+ Value *RestV = expand(Rest);
+ return expand(SE.getAddExpr(S->getStart(), SE.getUnknown(RestV)));
}
// {0,+,1} --> Insert a canonical induction variable into the loop!
if (S->isAffine() &&
S->getOperand(1) == SE.getIntegerSCEV(1, Ty)) {
+ // If there's a canonical IV, just use it.
+ if (CanonicalIV) {
+ assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
+ "IVs with types different from the canonical IV should "
+ "already have been handled!");
+ return CanonicalIV;
+ }
+
// Create and insert the PHI node for the induction variable in the
// specified loop.
BasicBlock *Header = L->getHeader();
PHINode *PN = PHINode::Create(Ty, "indvar", Header->begin());
+ InsertedValues.insert(PN);
PN->addIncoming(Constant::getNullValue(Ty), L->getLoopPreheader());
pred_iterator HPI = pred_begin(Header);
Constant *One = ConstantInt::get(Ty, 1);
Instruction *Add = BinaryOperator::CreateAdd(PN, One, "indvar.next",
(*HPI)->getTerminator());
+ InsertedValues.insert(Add);
pred_iterator PI = pred_begin(Header);
if (*PI == L->getLoopPreheader())
return PN;
}
+ // {0,+,F} --> {0,+,1} * F
// Get the canonical induction variable I for this loop.
- Value *I = getOrInsertCanonicalInductionVariable(L, Ty);
+ Value *I = CanonicalIV ?
+ CanonicalIV :
+ 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 = expand(S->getOperand(1));
-
- // IF the step is by one, just return the inserted IV.
- if (ConstantInt *CI = dyn_cast<ConstantInt>(F))
- if (CI->getValue() == 1)
- return I;
-
+ 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.
- Instruction *MulInsertPt = InsertPt;
- Loop *InsertPtLoop = LI.getLoopFor(MulInsertPt->getParent());
+ BasicBlock::iterator MulInsertPt = getInsertionPoint();
+ Loop *InsertPtLoop = SE.LI->getLoopFor(MulInsertPt->getParent());
if (InsertPtLoop != L && InsertPtLoop &&
L->contains(InsertPtLoop->getHeader())) {
do {
// into this folder.
SCEVHandle IH = SE.getUnknown(I); // Get I as a "symbolic" SCEV.
- SCEVHandle V = S->evaluateAtIteration(IH, SE);
+ // Promote S up to the canonical IV type, if the cast is foldable.
+ SCEVHandle NewS = S;
+ SCEVHandle Ext = SE.getNoopOrAnyExtend(S, I->getType());
+ if (isa<SCEVAddRecExpr>(Ext))
+ NewS = Ext;
+
+ SCEVHandle 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);
}
-Value *SCEVExpander::visitTruncateExpr(SCEVTruncateExpr *S) {
- Value *V = expand(S->getOperand());
- if (isa<PointerType>(V->getType()))
- V = InsertCastOfTo(Instruction::PtrToInt, V, TD.getIntPtrType());
- return CastInst::CreateTruncOrBitCast(V, S->getType(), "tmp.", InsertPt);
+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);
+ InsertedValues.insert(I);
+ return I;
}
-Value *SCEVExpander::visitZeroExtendExpr(SCEVZeroExtendExpr *S) {
- Value *V = expand(S->getOperand());
- if (isa<PointerType>(V->getType()))
- V = InsertCastOfTo(Instruction::PtrToInt, V, TD.getIntPtrType());
- return CastInst::CreateZExtOrBitCast(V, S->getType(), "tmp.", InsertPt);
+Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
+ 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);
+ InsertedValues.insert(I);
+ return I;
}
-Value *SCEVExpander::visitSignExtendExpr(SCEVSignExtendExpr *S) {
- Value *V = expand(S->getOperand());
- if (isa<PointerType>(V->getType()))
- V = InsertCastOfTo(Instruction::PtrToInt, V, TD.getIntPtrType());
- return CastInst::CreateSExtOrBitCast(V, S->getType(), "tmp.", InsertPt);
+Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
+ 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);
+ InsertedValues.insert(I);
+ return I;
}
-Value *SCEVExpander::visitSMaxExpr(SCEVSMaxExpr *S) {
- Value *LHS = expand(S->getOperand(0));
+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 *RHS = expand(S->getOperand(i));
- Value *ICmp = new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS, "tmp", InsertPt);
- LHS = SelectInst::Create(ICmp, LHS, RHS, "smax", InsertPt);
+ Value *RHS = expandCodeFor(S->getOperand(i), Ty);
+ Instruction *ICmp =
+ new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS, "tmp", InsertPt);
+ InsertedValues.insert(ICmp);
+ Instruction *Sel = SelectInst::Create(ICmp, LHS, RHS, "smax", InsertPt);
+ InsertedValues.insert(Sel);
+ LHS = Sel;
}
return LHS;
}
-Value *SCEVExpander::visitUMaxExpr(SCEVUMaxExpr *S) {
- Value *LHS = expand(S->getOperand(0));
+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 *RHS = expand(S->getOperand(i));
- Value *ICmp = new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS, "tmp", InsertPt);
- LHS = SelectInst::Create(ICmp, LHS, RHS, "umax", InsertPt);
+ Value *RHS = expandCodeFor(S->getOperand(i), Ty);
+ Instruction *ICmp =
+ new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS, "tmp", InsertPt);
+ InsertedValues.insert(ICmp);
+ Instruction *Sel = SelectInst::Create(ICmp, LHS, RHS, "umax", InsertPt);
+ InsertedValues.insert(Sel);
+ LHS = Sel;
}
return LHS;
}
-Value *SCEVExpander::expandCodeFor(SCEVHandle SH, const Type *Ty,
- Instruction *IP) {
+Value *SCEVExpander::expandCodeFor(SCEVHandle SH, const Type *Ty) {
// Expand the code for this SCEV.
- assert(TD.getTypeSizeInBits(Ty) == TD.getTypeSizeInBits(SH->getType()) &&
- "non-trivial casts should be done with the SCEVs directly!");
- this->InsertPt = IP;
Value *V = expand(SH);
- return InsertCastOfTo(CastInst::getCastOpcode(V, false, Ty, false), V, Ty);
+ if (Ty) {
+ assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
+ "non-trivial casts should be done with the SCEVs directly!");
+ V = InsertNoopCastOfTo(V, Ty);
+ }
+ return V;
}
-Value *SCEVExpander::expand(SCEV *S) {
+Value *SCEVExpander::expand(const SCEV *S) {
// Check to see if we already expanded this.
- std::map<SCEVHandle, Value*>::iterator I = InsertedExpressions.find(S);
+ std::map<SCEVHandle, AssertingVH<Value> >::iterator I =
+ InsertedExpressions.find(S);
if (I != InsertedExpressions.end())
return I->second;
InsertedExpressions[S] = V;
return V;
}
+
+/// getOrInsertCanonicalInductionVariable - This method returns the
+/// canonical induction variable of the specified type for the specified
+/// loop (inserting one if there is none). A canonical induction variable
+/// starts at zero and steps by one on each iteration.
+Value *
+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);
+}