#include "llvm/Analysis/ValueTracking.h"
#include "llvm/Assembly/Writer.h"
#include "llvm/Target/TargetData.h"
+#include "llvm/Target/TargetLibraryInfo.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/ConstantRange.h"
#include "llvm/Support/Debug.h"
"Scalar Evolution Analysis", false, true)
INITIALIZE_PASS_DEPENDENCY(LoopInfo)
INITIALIZE_PASS_DEPENDENCY(DominatorTree)
+INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
INITIALIZE_PASS_END(ScalarEvolution, "scalar-evolution",
"Scalar Evolution Analysis", false, true)
char ScalarEvolution::ID = 0;
for (unsigned i = 1, e = AR->getNumOperands(); i != e; ++i)
OS << ",+," << *AR->getOperand(i);
OS << "}<";
- if (AR->hasNoUnsignedWrap())
+ if (AR->getNoWrapFlags(FlagNUW))
OS << "nuw><";
- if (AR->hasNoSignedWrap())
+ if (AR->getNoWrapFlags(FlagNSW))
OS << "nsw><";
+ if (AR->getNoWrapFlags(FlagNW) &&
+ !AR->getNoWrapFlags((NoWrapFlags)(FlagNUW | FlagNSW)))
+ OS << "nw><";
WriteAsOperand(OS, AR->getLoop()->getHeader(), /*PrintType=*/false);
OS << ">";
return;
OS << OpStr;
}
OS << ")";
+ switch (NAry->getSCEVType()) {
+ case scAddExpr:
+ case scMulExpr:
+ if (NAry->getNoWrapFlags(FlagNUW))
+ OS << "<nuw>";
+ if (NAry->getNoWrapFlags(FlagNSW))
+ OS << "<nsw>";
+ }
return;
}
case scUDivExpr: {
}
case scUnknown: {
const SCEVUnknown *U = cast<SCEVUnknown>(this);
- const Type *AllocTy;
+ Type *AllocTy;
if (U->isSizeOf(AllocTy)) {
OS << "sizeof(" << *AllocTy << ")";
return;
OS << "alignof(" << *AllocTy << ")";
return;
}
-
- const Type *CTy;
+
+ Type *CTy;
Constant *FieldNo;
if (U->isOffsetOf(CTy, FieldNo)) {
OS << "offsetof(" << *CTy << ", ";
OS << ")";
return;
}
-
+
// Otherwise just print it normally.
WriteAsOperand(OS, U->getValue(), false);
return;
llvm_unreachable("Unknown SCEV kind!");
}
-const Type *SCEV::getType() const {
+Type *SCEV::getType() const {
switch (getSCEVType()) {
case scConstant:
return cast<SCEVConstant>(this)->getType();
return cast<SCEVUnknown>(this)->getType();
case scCouldNotCompute:
llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
- return 0;
- default: break;
+ default:
+ llvm_unreachable("Unknown SCEV kind!");
}
- llvm_unreachable("Unknown SCEV kind!");
- return 0;
}
bool SCEV::isZero() const {
return false;
}
+/// isNonConstantNegative - Return true if the specified scev is negated, but
+/// not a constant.
+bool SCEV::isNonConstantNegative() const {
+ const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(this);
+ if (!Mul) return false;
+
+ // If there is a constant factor, it will be first.
+ const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
+ if (!SC) return false;
+
+ // Return true if the value is negative, this matches things like (-42 * V).
+ return SC->getValue()->getValue().isNegative();
+}
+
SCEVCouldNotCompute::SCEVCouldNotCompute() :
SCEV(FoldingSetNodeIDRef(), scCouldNotCompute) {}
}
const SCEV *
-ScalarEvolution::getConstant(const Type *Ty, uint64_t V, bool isSigned) {
- const IntegerType *ITy = cast<IntegerType>(getEffectiveSCEVType(Ty));
+ScalarEvolution::getConstant(Type *Ty, uint64_t V, bool isSigned) {
+ IntegerType *ITy = cast<IntegerType>(getEffectiveSCEVType(Ty));
return getConstant(ConstantInt::get(ITy, V, isSigned));
}
SCEVCastExpr::SCEVCastExpr(const FoldingSetNodeIDRef ID,
- unsigned SCEVTy, const SCEV *op, const Type *ty)
+ unsigned SCEVTy, const SCEV *op, Type *ty)
: SCEV(ID, SCEVTy), Op(op), Ty(ty) {}
SCEVTruncateExpr::SCEVTruncateExpr(const FoldingSetNodeIDRef ID,
- const SCEV *op, const Type *ty)
+ const SCEV *op, Type *ty)
: SCEVCastExpr(ID, scTruncate, op, ty) {
assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) &&
(Ty->isIntegerTy() || Ty->isPointerTy()) &&
}
SCEVZeroExtendExpr::SCEVZeroExtendExpr(const FoldingSetNodeIDRef ID,
- const SCEV *op, const Type *ty)
+ const SCEV *op, Type *ty)
: SCEVCastExpr(ID, scZeroExtend, op, ty) {
assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) &&
(Ty->isIntegerTy() || Ty->isPointerTy()) &&
}
SCEVSignExtendExpr::SCEVSignExtendExpr(const FoldingSetNodeIDRef ID,
- const SCEV *op, const Type *ty)
+ const SCEV *op, Type *ty)
: SCEVCastExpr(ID, scSignExtend, op, ty) {
assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) &&
(Ty->isIntegerTy() || Ty->isPointerTy()) &&
setValPtr(New);
}
-bool SCEVUnknown::isSizeOf(const Type *&AllocTy) const {
+bool SCEVUnknown::isSizeOf(Type *&AllocTy) const {
if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
if (VCE->getOpcode() == Instruction::PtrToInt)
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
return false;
}
-bool SCEVUnknown::isAlignOf(const Type *&AllocTy) const {
+bool SCEVUnknown::isAlignOf(Type *&AllocTy) const {
if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
if (VCE->getOpcode() == Instruction::PtrToInt)
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
if (CE->getOpcode() == Instruction::GetElementPtr &&
CE->getOperand(0)->isNullValue()) {
- const Type *Ty =
+ Type *Ty =
cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
- if (const StructType *STy = dyn_cast<StructType>(Ty))
+ if (StructType *STy = dyn_cast<StructType>(Ty))
if (!STy->isPacked() &&
CE->getNumOperands() == 3 &&
CE->getOperand(1)->isNullValue()) {
return false;
}
-bool SCEVUnknown::isOffsetOf(const Type *&CTy, Constant *&FieldNo) const {
+bool SCEVUnknown::isOffsetOf(Type *&CTy, Constant *&FieldNo) const {
if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
if (VCE->getOpcode() == Instruction::PtrToInt)
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
CE->getNumOperands() == 3 &&
CE->getOperand(0)->isNullValue() &&
CE->getOperand(1)->isNullValue()) {
- const Type *Ty =
+ Type *Ty =
cast<PointerType>(CE->getOperand(0)->getType())->getElementType();
// Ignore vector types here so that ScalarEvolutionExpander doesn't
// emit getelementptrs that index into vectors.
}
default:
- break;
+ llvm_unreachable("Unknown SCEV kind!");
}
-
- llvm_unreachable("Unknown SCEV kind!");
- return 0;
}
};
}
/// Assume, K > 0.
static const SCEV *BinomialCoefficient(const SCEV *It, unsigned K,
ScalarEvolution &SE,
- const Type* ResultTy) {
+ Type *ResultTy) {
// Handle the simplest case efficiently.
if (K == 1)
return SE.getTruncateOrZeroExtend(It, ResultTy);
MultiplyFactor = MultiplyFactor.trunc(W);
// Calculate the product, at width T+W
- const IntegerType *CalculationTy = IntegerType::get(SE.getContext(),
+ IntegerType *CalculationTy = IntegerType::get(SE.getContext(),
CalculationBits);
const SCEV *Dividend = SE.getTruncateOrZeroExtend(It, CalculationTy);
for (unsigned i = 1; i != K; ++i) {
//===----------------------------------------------------------------------===//
const SCEV *ScalarEvolution::getTruncateExpr(const SCEV *Op,
- const Type *Ty) {
+ Type *Ty) {
assert(getTypeSizeInBits(Op->getType()) > getTypeSizeInBits(Ty) &&
"This is not a truncating conversion!");
assert(isSCEVable(Ty) &&
Operands.push_back(S);
}
if (!hasTrunc)
- return getAddExpr(Operands, false, false);
+ return getAddExpr(Operands);
UniqueSCEVs.FindNodeOrInsertPos(ID, IP); // Mutates IP, returns NULL.
}
Operands.push_back(S);
}
if (!hasTrunc)
- return getMulExpr(Operands, false, false);
+ return getMulExpr(Operands);
UniqueSCEVs.FindNodeOrInsertPos(ID, IP); // Mutates IP, returns NULL.
}
SmallVector<const SCEV *, 4> Operands;
for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
Operands.push_back(getTruncateExpr(AddRec->getOperand(i), Ty));
- return getAddRecExpr(Operands, AddRec->getLoop());
+ return getAddRecExpr(Operands, AddRec->getLoop(), SCEV::FlagAnyWrap);
}
// As a special case, fold trunc(undef) to undef. We don't want to
}
const SCEV *ScalarEvolution::getZeroExtendExpr(const SCEV *Op,
- const Type *Ty) {
+ Type *Ty) {
assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
"This is not an extending conversion!");
assert(isSCEVable(Ty) &&
// If we have special knowledge that this addrec won't overflow,
// we don't need to do any further analysis.
- if (AR->hasNoUnsignedWrap())
+ if (AR->getNoWrapFlags(SCEV::FlagNUW))
return getAddRecExpr(getZeroExtendExpr(Start, Ty),
getZeroExtendExpr(Step, Ty),
- L);
+ L, AR->getNoWrapFlags());
// Check whether the backedge-taken count is SCEVCouldNotCompute.
// Note that this serves two purposes: It filters out loops that are
const SCEV *RecastedMaxBECount =
getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
if (MaxBECount == RecastedMaxBECount) {
- const Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
+ Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
// Check whether Start+Step*MaxBECount has no unsigned overflow.
const SCEV *ZMul = getMulExpr(CastedMaxBECount, Step);
const SCEV *Add = getAddExpr(Start, ZMul);
getAddExpr(getZeroExtendExpr(Start, WideTy),
getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
getZeroExtendExpr(Step, WideTy)));
- if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd)
+ if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd) {
+ // Cache knowledge of AR NUW, which is propagated to this AddRec.
+ const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNUW);
// Return the expression with the addrec on the outside.
return getAddRecExpr(getZeroExtendExpr(Start, Ty),
getZeroExtendExpr(Step, Ty),
- L);
-
+ L, AR->getNoWrapFlags());
+ }
// Similar to above, only this time treat the step value as signed.
// This covers loops that count down.
const SCEV *SMul = getMulExpr(CastedMaxBECount, Step);
getAddExpr(getZeroExtendExpr(Start, WideTy),
getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
getSignExtendExpr(Step, WideTy)));
- if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd)
+ if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd) {
+ // Cache knowledge of AR NW, which is propagated to this AddRec.
+ // Negative step causes unsigned wrap, but it still can't self-wrap.
+ const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNW);
// Return the expression with the addrec on the outside.
return getAddRecExpr(getZeroExtendExpr(Start, Ty),
getSignExtendExpr(Step, Ty),
- L);
+ L, AR->getNoWrapFlags());
+ }
}
// If the backedge is guarded by a comparison with the pre-inc value
if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT, AR, N) ||
(isLoopEntryGuardedByCond(L, ICmpInst::ICMP_ULT, Start, N) &&
isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT,
- AR->getPostIncExpr(*this), N)))
+ AR->getPostIncExpr(*this), N))) {
+ // Cache knowledge of AR NUW, which is propagated to this AddRec.
+ const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNUW);
// Return the expression with the addrec on the outside.
return getAddRecExpr(getZeroExtendExpr(Start, Ty),
getZeroExtendExpr(Step, Ty),
- L);
+ L, AR->getNoWrapFlags());
+ }
} else if (isKnownNegative(Step)) {
const SCEV *N = getConstant(APInt::getMaxValue(BitWidth) -
getSignedRange(Step).getSignedMin());
if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT, AR, N) ||
(isLoopEntryGuardedByCond(L, ICmpInst::ICMP_UGT, Start, N) &&
isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT,
- AR->getPostIncExpr(*this), N)))
+ AR->getPostIncExpr(*this), N))) {
+ // Cache knowledge of AR NW, which is propagated to this AddRec.
+ // Negative step causes unsigned wrap, but it still can't self-wrap.
+ const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNW);
// Return the expression with the addrec on the outside.
return getAddRecExpr(getZeroExtendExpr(Start, Ty),
getSignExtendExpr(Step, Ty),
- L);
+ L, AR->getNoWrapFlags());
+ }
}
}
}
return S;
}
+// Get the limit of a recurrence such that incrementing by Step cannot cause
+// signed overflow as long as the value of the recurrence within the loop does
+// not exceed this limit before incrementing.
+static const SCEV *getOverflowLimitForStep(const SCEV *Step,
+ ICmpInst::Predicate *Pred,
+ ScalarEvolution *SE) {
+ unsigned BitWidth = SE->getTypeSizeInBits(Step->getType());
+ if (SE->isKnownPositive(Step)) {
+ *Pred = ICmpInst::ICMP_SLT;
+ return SE->getConstant(APInt::getSignedMinValue(BitWidth) -
+ SE->getSignedRange(Step).getSignedMax());
+ }
+ if (SE->isKnownNegative(Step)) {
+ *Pred = ICmpInst::ICMP_SGT;
+ return SE->getConstant(APInt::getSignedMaxValue(BitWidth) -
+ SE->getSignedRange(Step).getSignedMin());
+ }
+ return 0;
+}
+
+// The recurrence AR has been shown to have no signed wrap. Typically, if we can
+// prove NSW for AR, then we can just as easily prove NSW for its preincrement
+// or postincrement sibling. This allows normalizing a sign extended AddRec as
+// such: {sext(Step + Start),+,Step} => {(Step + sext(Start),+,Step} As a
+// result, the expression "Step + sext(PreIncAR)" is congruent with
+// "sext(PostIncAR)"
+static const SCEV *getPreStartForSignExtend(const SCEVAddRecExpr *AR,
+ Type *Ty,
+ ScalarEvolution *SE) {
+ const Loop *L = AR->getLoop();
+ const SCEV *Start = AR->getStart();
+ const SCEV *Step = AR->getStepRecurrence(*SE);
+
+ // Check for a simple looking step prior to loop entry.
+ const SCEVAddExpr *SA = dyn_cast<SCEVAddExpr>(Start);
+ if (!SA)
+ return 0;
+
+ // Create an AddExpr for "PreStart" after subtracting Step. Full SCEV
+ // subtraction is expensive. For this purpose, perform a quick and dirty
+ // difference, by checking for Step in the operand list.
+ SmallVector<const SCEV *, 4> DiffOps;
+ for (SCEVAddExpr::op_iterator I = SA->op_begin(), E = SA->op_end();
+ I != E; ++I) {
+ if (*I != Step)
+ DiffOps.push_back(*I);
+ }
+ if (DiffOps.size() == SA->getNumOperands())
+ return 0;
+
+ // This is a postinc AR. Check for overflow on the preinc recurrence using the
+ // same three conditions that getSignExtendedExpr checks.
+
+ // 1. NSW flags on the step increment.
+ const SCEV *PreStart = SE->getAddExpr(DiffOps, SA->getNoWrapFlags());
+ const SCEVAddRecExpr *PreAR = dyn_cast<SCEVAddRecExpr>(
+ SE->getAddRecExpr(PreStart, Step, L, SCEV::FlagAnyWrap));
+
+ if (PreAR && PreAR->getNoWrapFlags(SCEV::FlagNSW))
+ return PreStart;
+
+ // 2. Direct overflow check on the step operation's expression.
+ unsigned BitWidth = SE->getTypeSizeInBits(AR->getType());
+ Type *WideTy = IntegerType::get(SE->getContext(), BitWidth * 2);
+ const SCEV *OperandExtendedStart =
+ SE->getAddExpr(SE->getSignExtendExpr(PreStart, WideTy),
+ SE->getSignExtendExpr(Step, WideTy));
+ if (SE->getSignExtendExpr(Start, WideTy) == OperandExtendedStart) {
+ // Cache knowledge of PreAR NSW.
+ if (PreAR)
+ const_cast<SCEVAddRecExpr *>(PreAR)->setNoWrapFlags(SCEV::FlagNSW);
+ // FIXME: this optimization needs a unit test
+ DEBUG(dbgs() << "SCEV: untested prestart overflow check\n");
+ return PreStart;
+ }
+
+ // 3. Loop precondition.
+ ICmpInst::Predicate Pred;
+ const SCEV *OverflowLimit = getOverflowLimitForStep(Step, &Pred, SE);
+
+ if (OverflowLimit &&
+ SE->isLoopEntryGuardedByCond(L, Pred, PreStart, OverflowLimit)) {
+ return PreStart;
+ }
+ return 0;
+}
+
+// Get the normalized sign-extended expression for this AddRec's Start.
+static const SCEV *getSignExtendAddRecStart(const SCEVAddRecExpr *AR,
+ Type *Ty,
+ ScalarEvolution *SE) {
+ const SCEV *PreStart = getPreStartForSignExtend(AR, Ty, SE);
+ if (!PreStart)
+ return SE->getSignExtendExpr(AR->getStart(), Ty);
+
+ return SE->getAddExpr(SE->getSignExtendExpr(AR->getStepRecurrence(*SE), Ty),
+ SE->getSignExtendExpr(PreStart, Ty));
+}
+
const SCEV *ScalarEvolution::getSignExtendExpr(const SCEV *Op,
- const Type *Ty) {
+ Type *Ty) {
assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
"This is not an extending conversion!");
assert(isSCEVable(Ty) &&
// If we have special knowledge that this addrec won't overflow,
// we don't need to do any further analysis.
- if (AR->hasNoSignedWrap())
- return getAddRecExpr(getSignExtendExpr(Start, Ty),
+ if (AR->getNoWrapFlags(SCEV::FlagNSW))
+ return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
getSignExtendExpr(Step, Ty),
- L);
+ L, SCEV::FlagNSW);
// Check whether the backedge-taken count is SCEVCouldNotCompute.
// Note that this serves two purposes: It filters out loops that are
const SCEV *RecastedMaxBECount =
getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType());
if (MaxBECount == RecastedMaxBECount) {
- const Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
+ Type *WideTy = IntegerType::get(getContext(), BitWidth * 2);
// Check whether Start+Step*MaxBECount has no signed overflow.
const SCEV *SMul = getMulExpr(CastedMaxBECount, Step);
const SCEV *Add = getAddExpr(Start, SMul);
getAddExpr(getSignExtendExpr(Start, WideTy),
getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
getSignExtendExpr(Step, WideTy)));
- if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd)
+ if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd) {
+ // Cache knowledge of AR NSW, which is propagated to this AddRec.
+ const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNSW);
// Return the expression with the addrec on the outside.
- return getAddRecExpr(getSignExtendExpr(Start, Ty),
+ return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
getSignExtendExpr(Step, Ty),
- L);
-
+ L, AR->getNoWrapFlags());
+ }
// Similar to above, only this time treat the step value as unsigned.
// This covers loops that count up with an unsigned step.
const SCEV *UMul = getMulExpr(CastedMaxBECount, Step);
getAddExpr(getSignExtendExpr(Start, WideTy),
getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
getZeroExtendExpr(Step, WideTy)));
- if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd)
+ if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd) {
+ // Cache knowledge of AR NSW, which is propagated to this AddRec.
+ const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNSW);
// Return the expression with the addrec on the outside.
- return getAddRecExpr(getSignExtendExpr(Start, Ty),
+ return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
getZeroExtendExpr(Step, Ty),
- L);
+ L, AR->getNoWrapFlags());
+ }
}
// If the backedge is guarded by a comparison with the pre-inc value
// the addrec is safe. Also, if the entry is guarded by a comparison
// with the start value and the backedge is guarded by a comparison
// with the post-inc value, the addrec is safe.
- if (isKnownPositive(Step)) {
- const SCEV *N = getConstant(APInt::getSignedMinValue(BitWidth) -
- getSignedRange(Step).getSignedMax());
- if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SLT, AR, N) ||
- (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_SLT, Start, N) &&
- isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SLT,
- AR->getPostIncExpr(*this), N)))
- // Return the expression with the addrec on the outside.
- return getAddRecExpr(getSignExtendExpr(Start, Ty),
- getSignExtendExpr(Step, Ty),
- L);
- } else if (isKnownNegative(Step)) {
- const SCEV *N = getConstant(APInt::getSignedMaxValue(BitWidth) -
- getSignedRange(Step).getSignedMin());
- if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SGT, AR, N) ||
- (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_SGT, Start, N) &&
- isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_SGT,
- AR->getPostIncExpr(*this), N)))
- // Return the expression with the addrec on the outside.
- return getAddRecExpr(getSignExtendExpr(Start, Ty),
- getSignExtendExpr(Step, Ty),
- L);
+ ICmpInst::Predicate Pred;
+ const SCEV *OverflowLimit = getOverflowLimitForStep(Step, &Pred, this);
+ if (OverflowLimit &&
+ (isLoopBackedgeGuardedByCond(L, Pred, AR, OverflowLimit) ||
+ (isLoopEntryGuardedByCond(L, Pred, Start, OverflowLimit) &&
+ isLoopBackedgeGuardedByCond(L, Pred, AR->getPostIncExpr(*this),
+ OverflowLimit)))) {
+ // Cache knowledge of AR NSW, then propagate NSW to the wide AddRec.
+ const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNSW);
+ return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this),
+ getSignExtendExpr(Step, Ty),
+ L, AR->getNoWrapFlags());
}
}
}
/// unspecified bits out to the given type.
///
const SCEV *ScalarEvolution::getAnyExtendExpr(const SCEV *Op,
- const Type *Ty) {
+ Type *Ty) {
assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) &&
"This is not an extending conversion!");
assert(isSCEVable(Ty) &&
for (SCEVAddRecExpr::op_iterator I = AR->op_begin(), E = AR->op_end();
I != E; ++I)
Ops.push_back(getAnyExtendExpr(*I, Ty));
- return getAddRecExpr(Ops, AR->getLoop());
+ return getAddRecExpr(Ops, AR->getLoop(), SCEV::FlagNW);
}
// As a special case, fold anyext(undef) to undef. We don't want to
/// getAddExpr - Get a canonical add expression, or something simpler if
/// possible.
const SCEV *ScalarEvolution::getAddExpr(SmallVectorImpl<const SCEV *> &Ops,
- bool HasNUW, bool HasNSW) {
+ SCEV::NoWrapFlags Flags) {
+ assert(!(Flags & ~(SCEV::FlagNUW | SCEV::FlagNSW)) &&
+ "only nuw or nsw allowed");
assert(!Ops.empty() && "Cannot get empty add!");
if (Ops.size() == 1) return Ops[0];
#ifndef NDEBUG
- const Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
+ Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
for (unsigned i = 1, e = Ops.size(); i != e; ++i)
assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
"SCEVAddExpr operand types don't match!");
#endif
- // If HasNSW is true and all the operands are non-negative, infer HasNUW.
- if (!HasNUW && HasNSW) {
+ // If FlagNSW is true and all the operands are non-negative, infer FlagNUW.
+ // And vice-versa.
+ int SignOrUnsignMask = SCEV::FlagNUW | SCEV::FlagNSW;
+ SCEV::NoWrapFlags SignOrUnsignWrap = maskFlags(Flags, SignOrUnsignMask);
+ if (SignOrUnsignWrap && (SignOrUnsignWrap != SignOrUnsignMask)) {
bool All = true;
for (SmallVectorImpl<const SCEV *>::const_iterator I = Ops.begin(),
E = Ops.end(); I != E; ++I)
All = false;
break;
}
- if (All) HasNUW = true;
+ if (All) Flags = setFlags(Flags, (SCEV::NoWrapFlags)SignOrUnsignMask);
}
// Sort by complexity, this groups all similar expression types together.
// Okay, check to see if the same value occurs in the operand list more than
// once. If so, merge them together into an multiply expression. Since we
// sorted the list, these values are required to be adjacent.
- const Type *Ty = Ops[0]->getType();
+ Type *Ty = Ops[0]->getType();
bool FoundMatch = false;
for (unsigned i = 0, e = Ops.size(); i != e-1; ++i)
if (Ops[i] == Ops[i+1]) { // X + Y + Y --> X + Y*2
FoundMatch = true;
}
if (FoundMatch)
- return getAddExpr(Ops, HasNUW, HasNSW);
+ return getAddExpr(Ops, Flags);
// Check for truncates. If all the operands are truncated from the same
// type, see if factoring out the truncate would permit the result to be
// if the contents of the resulting outer trunc fold to something simple.
for (; Idx < Ops.size() && isa<SCEVTruncateExpr>(Ops[Idx]); ++Idx) {
const SCEVTruncateExpr *Trunc = cast<SCEVTruncateExpr>(Ops[Idx]);
- const Type *DstType = Trunc->getType();
- const Type *SrcType = Trunc->getOperand()->getType();
+ Type *DstType = Trunc->getType();
+ Type *SrcType = Trunc->getOperand()->getType();
SmallVector<const SCEV *, 8> LargeOps;
bool Ok = true;
// Check all the operands to see if they can be represented in the
}
if (Ok) {
// Evaluate the expression in the larger type.
- const SCEV *Fold = getAddExpr(LargeOps, HasNUW, HasNSW);
+ const SCEV *Fold = getAddExpr(LargeOps, Flags);
// If it folds to something simple, use it. Otherwise, don't.
if (isa<SCEVConstant>(Fold) || isa<SCEVUnknown>(Fold))
return getTruncateExpr(Fold, DstType);
// Build the new addrec. Propagate the NUW and NSW flags if both the
// outer add and the inner addrec are guaranteed to have no overflow.
- const SCEV *NewRec = getAddRecExpr(AddRecOps, AddRecLoop,
- HasNUW && AddRec->hasNoUnsignedWrap(),
- HasNSW && AddRec->hasNoSignedWrap());
+ // Always propagate NW.
+ Flags = AddRec->getNoWrapFlags(setFlags(Flags, SCEV::FlagNW));
+ const SCEV *NewRec = getAddRecExpr(AddRecOps, AddRecLoop, Flags);
// If all of the other operands were loop invariant, we are done.
if (Ops.size() == 1) return NewRec;
- // Otherwise, add the folded AddRec by the non-liv parts.
+ // Otherwise, add the folded AddRec by the non-invariant parts.
for (unsigned i = 0;; ++i)
if (Ops[i] == AddRec) {
Ops[i] = NewRec;
}
Ops.erase(Ops.begin() + OtherIdx); --OtherIdx;
}
- Ops[Idx] = getAddRecExpr(AddRecOps, AddRecLoop);
+ // Step size has changed, so we cannot guarantee no self-wraparound.
+ Ops[Idx] = getAddRecExpr(AddRecOps, AddRecLoop, SCEV::FlagAnyWrap);
return getAddExpr(Ops);
}
O, Ops.size());
UniqueSCEVs.InsertNode(S, IP);
}
- if (HasNUW) S->setHasNoUnsignedWrap(true);
- if (HasNSW) S->setHasNoSignedWrap(true);
+ S->setNoWrapFlags(Flags);
return S;
}
+static uint64_t umul_ov(uint64_t i, uint64_t j, bool &Overflow) {
+ uint64_t k = i*j;
+ if (j > 1 && k / j != i) Overflow = true;
+ return k;
+}
+
+/// Compute the result of "n choose k", the binomial coefficient. If an
+/// intermediate computation overflows, Overflow will be set and the return will
+/// be garbage. Overflow is not cleared on absense of overflow.
+static uint64_t Choose(uint64_t n, uint64_t k, bool &Overflow) {
+ // We use the multiplicative formula:
+ // n(n-1)(n-2)...(n-(k-1)) / k(k-1)(k-2)...1 .
+ // At each iteration, we take the n-th term of the numeral and divide by the
+ // (k-n)th term of the denominator. This division will always produce an
+ // integral result, and helps reduce the chance of overflow in the
+ // intermediate computations. However, we can still overflow even when the
+ // final result would fit.
+
+ if (n == 0 || n == k) return 1;
+ if (k > n) return 0;
+
+ if (k > n/2)
+ k = n-k;
+
+ uint64_t r = 1;
+ for (uint64_t i = 1; i <= k; ++i) {
+ r = umul_ov(r, n-(i-1), Overflow);
+ r /= i;
+ }
+ return r;
+}
+
/// getMulExpr - Get a canonical multiply expression, or something simpler if
/// possible.
const SCEV *ScalarEvolution::getMulExpr(SmallVectorImpl<const SCEV *> &Ops,
- bool HasNUW, bool HasNSW) {
+ SCEV::NoWrapFlags Flags) {
+ assert(Flags == maskFlags(Flags, SCEV::FlagNUW | SCEV::FlagNSW) &&
+ "only nuw or nsw allowed");
assert(!Ops.empty() && "Cannot get empty mul!");
if (Ops.size() == 1) return Ops[0];
#ifndef NDEBUG
- const Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
+ Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
for (unsigned i = 1, e = Ops.size(); i != e; ++i)
assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
"SCEVMulExpr operand types don't match!");
#endif
- // If HasNSW is true and all the operands are non-negative, infer HasNUW.
- if (!HasNUW && HasNSW) {
+ // If FlagNSW is true and all the operands are non-negative, infer FlagNUW.
+ // And vice-versa.
+ int SignOrUnsignMask = SCEV::FlagNUW | SCEV::FlagNSW;
+ SCEV::NoWrapFlags SignOrUnsignWrap = maskFlags(Flags, SignOrUnsignMask);
+ if (SignOrUnsignWrap && (SignOrUnsignWrap != SignOrUnsignMask)) {
bool All = true;
for (SmallVectorImpl<const SCEV *>::const_iterator I = Ops.begin(),
E = Ops.end(); I != E; ++I)
All = false;
break;
}
- if (All) HasNUW = true;
+ if (All) Flags = setFlags(Flags, (SCEV::NoWrapFlags)SignOrUnsignMask);
}
// Sort by complexity, this groups all similar expression types together.
} else if (Ops[0]->isAllOnesValue()) {
// If we have a mul by -1 of an add, try distributing the -1 among the
// add operands.
- if (Ops.size() == 2)
+ if (Ops.size() == 2) {
if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1])) {
SmallVector<const SCEV *, 4> NewOps;
bool AnyFolded = false;
- for (SCEVAddRecExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
- I != E; ++I) {
+ for (SCEVAddRecExpr::op_iterator I = Add->op_begin(),
+ E = Add->op_end(); I != E; ++I) {
const SCEV *Mul = getMulExpr(Ops[0], *I);
if (!isa<SCEVMulExpr>(Mul)) AnyFolded = true;
NewOps.push_back(Mul);
if (AnyFolded)
return getAddExpr(NewOps);
}
+ else if (const SCEVAddRecExpr *
+ AddRec = dyn_cast<SCEVAddRecExpr>(Ops[1])) {
+ // Negation preserves a recurrence's no self-wrap property.
+ SmallVector<const SCEV *, 4> Operands;
+ for (SCEVAddRecExpr::op_iterator I = AddRec->op_begin(),
+ E = AddRec->op_end(); I != E; ++I) {
+ Operands.push_back(getMulExpr(Ops[0], *I));
+ }
+ return getAddRecExpr(Operands, AddRec->getLoop(),
+ AddRec->getNoWrapFlags(SCEV::FlagNW));
+ }
+ }
}
if (Ops.size() == 1)
// Build the new addrec. Propagate the NUW and NSW flags if both the
// outer mul and the inner addrec are guaranteed to have no overflow.
- const SCEV *NewRec = getAddRecExpr(NewOps, AddRecLoop,
- HasNUW && AddRec->hasNoUnsignedWrap(),
- HasNSW && AddRec->hasNoSignedWrap());
+ //
+ // No self-wrap cannot be guaranteed after changing the step size, but
+ // will be inferred if either NUW or NSW is true.
+ Flags = AddRec->getNoWrapFlags(clearFlags(Flags, SCEV::FlagNW));
+ const SCEV *NewRec = getAddRecExpr(NewOps, AddRecLoop, Flags);
// If all of the other operands were loop invariant, we are done.
if (Ops.size() == 1) return NewRec;
- // Otherwise, multiply the folded AddRec by the non-liv parts.
+ // Otherwise, multiply the folded AddRec by the non-invariant parts.
for (unsigned i = 0;; ++i)
if (Ops[i] == AddRec) {
Ops[i] = NewRec;
// multiplied together. If so, we can fold them.
for (unsigned OtherIdx = Idx+1;
OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
- ++OtherIdx)
+ ++OtherIdx) {
if (AddRecLoop == cast<SCEVAddRecExpr>(Ops[OtherIdx])->getLoop()) {
- // F * G, where F = {A,+,B}<L> and G = {C,+,D}<L> -->
- // {A*C,+,F*D + G*B + B*D}<L>
+ // {A1,+,A2,+,...,+,An}<L> * {B1,+,B2,+,...,+,Bn}<L>
+ // = {x=1 in [ sum y=x..2x [ sum z=max(y-x, y-n)..min(x,n) [
+ // choose(x, 2x)*choose(2x-y, x-z)*A_{y-z}*B_z
+ // ]]],+,...up to x=2n}.
+ // Note that the arguments to choose() are always integers with values
+ // known at compile time, never SCEV objects.
+ //
+ // The implementation avoids pointless extra computations when the two
+ // addrec's are of different length (mathematically, it's equivalent to
+ // an infinite stream of zeros on the right).
+ bool OpsModified = false;
for (; OtherIdx != Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
++OtherIdx)
if (const SCEVAddRecExpr *OtherAddRec =
dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx]))
if (OtherAddRec->getLoop() == AddRecLoop) {
- const SCEVAddRecExpr *F = AddRec, *G = OtherAddRec;
- const SCEV *NewStart = getMulExpr(F->getStart(), G->getStart());
- const SCEV *B = F->getStepRecurrence(*this);
- const SCEV *D = G->getStepRecurrence(*this);
- const SCEV *NewStep = getAddExpr(getMulExpr(F, D),
- getMulExpr(G, B),
- getMulExpr(B, D));
- const SCEV *NewAddRec = getAddRecExpr(NewStart, NewStep,
- F->getLoop());
- if (Ops.size() == 2) return NewAddRec;
- Ops[Idx] = AddRec = cast<SCEVAddRecExpr>(NewAddRec);
- Ops.erase(Ops.begin() + OtherIdx); --OtherIdx;
+ bool Overflow = false;
+ Type *Ty = AddRec->getType();
+ bool LargerThan64Bits = getTypeSizeInBits(Ty) > 64;
+ SmallVector<const SCEV*, 7> AddRecOps;
+ for (int x = 0, xe = AddRec->getNumOperands() +
+ OtherAddRec->getNumOperands() - 1;
+ x != xe && !Overflow; ++x) {
+ const SCEV *Term = getConstant(Ty, 0);
+ for (int y = x, ye = 2*x+1; y != ye && !Overflow; ++y) {
+ uint64_t Coeff1 = Choose(x, 2*x - y, Overflow);
+ for (int z = std::max(y-x, y-(int)AddRec->getNumOperands()+1),
+ ze = std::min(x+1, (int)OtherAddRec->getNumOperands());
+ z < ze && !Overflow; ++z) {
+ uint64_t Coeff2 = Choose(2*x - y, x-z, Overflow);
+ uint64_t Coeff;
+ if (LargerThan64Bits)
+ Coeff = umul_ov(Coeff1, Coeff2, Overflow);
+ else
+ Coeff = Coeff1*Coeff2;
+ const SCEV *CoeffTerm = getConstant(Ty, Coeff);
+ const SCEV *Term1 = AddRec->getOperand(y-z);
+ const SCEV *Term2 = OtherAddRec->getOperand(z);
+ Term = getAddExpr(Term, getMulExpr(CoeffTerm, Term1,Term2));
+ }
+ }
+ AddRecOps.push_back(Term);
+ }
+ if (!Overflow) {
+ const SCEV *NewAddRec = getAddRecExpr(AddRecOps,
+ AddRec->getLoop(),
+ SCEV::FlagAnyWrap);
+ if (Ops.size() == 2) return NewAddRec;
+ Ops[Idx] = AddRec = cast<SCEVAddRecExpr>(NewAddRec);
+ Ops.erase(Ops.begin() + OtherIdx); --OtherIdx;
+ OpsModified = true;
+ }
}
- return getMulExpr(Ops);
+ if (OpsModified)
+ return getMulExpr(Ops);
}
+ }
// Otherwise couldn't fold anything into this recurrence. Move onto the
// next one.
O, Ops.size());
UniqueSCEVs.InsertNode(S, IP);
}
- if (HasNUW) S->setHasNoUnsignedWrap(true);
- if (HasNSW) S->setHasNoSignedWrap(true);
+ S->setNoWrapFlags(Flags);
return S;
}
// Determine if the division can be folded into the operands of
// its operands.
// TODO: Generalize this to non-constants by using known-bits information.
- const Type *Ty = LHS->getType();
+ Type *Ty = LHS->getType();
unsigned LZ = RHSC->getValue()->getValue().countLeadingZeros();
unsigned MaxShiftAmt = getTypeSizeInBits(Ty) - LZ - 1;
// For non-power-of-two values, effectively round the value up to the
// nearest power of two.
if (!RHSC->getValue()->getValue().isPowerOf2())
++MaxShiftAmt;
- const IntegerType *ExtTy =
+ IntegerType *ExtTy =
IntegerType::get(getContext(), getTypeSizeInBits(Ty) + MaxShiftAmt);
- // {X,+,N}/C --> {X/C,+,N/C} if safe and N/C can be folded.
if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS))
if (const SCEVConstant *Step =
- dyn_cast<SCEVConstant>(AR->getStepRecurrence(*this)))
- if (!Step->getValue()->getValue()
- .urem(RHSC->getValue()->getValue()) &&
+ dyn_cast<SCEVConstant>(AR->getStepRecurrence(*this))) {
+ // {X,+,N}/C --> {X/C,+,N/C} if safe and N/C can be folded.
+ const APInt &StepInt = Step->getValue()->getValue();
+ const APInt &DivInt = RHSC->getValue()->getValue();
+ if (!StepInt.urem(DivInt) &&
getZeroExtendExpr(AR, ExtTy) ==
getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy),
getZeroExtendExpr(Step, ExtTy),
- AR->getLoop())) {
+ AR->getLoop(), SCEV::FlagAnyWrap)) {
SmallVector<const SCEV *, 4> Operands;
for (unsigned i = 0, e = AR->getNumOperands(); i != e; ++i)
Operands.push_back(getUDivExpr(AR->getOperand(i), RHS));
- return getAddRecExpr(Operands, AR->getLoop());
+ return getAddRecExpr(Operands, AR->getLoop(),
+ SCEV::FlagNW);
}
+ /// Get a canonical UDivExpr for a recurrence.
+ /// {X,+,N}/C => {Y,+,N}/C where Y=X-(X%N). Safe when C%N=0.
+ // We can currently only fold X%N if X is constant.
+ const SCEVConstant *StartC = dyn_cast<SCEVConstant>(AR->getStart());
+ if (StartC && !DivInt.urem(StepInt) &&
+ getZeroExtendExpr(AR, ExtTy) ==
+ getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy),
+ getZeroExtendExpr(Step, ExtTy),
+ AR->getLoop(), SCEV::FlagAnyWrap)) {
+ const APInt &StartInt = StartC->getValue()->getValue();
+ const APInt &StartRem = StartInt.urem(StepInt);
+ if (StartRem != 0)
+ LHS = getAddRecExpr(getConstant(StartInt - StartRem), Step,
+ AR->getLoop(), SCEV::FlagNW);
+ }
+ }
// (A*B)/C --> A*(B/C) if safe and B/C can be folded.
if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(LHS)) {
SmallVector<const SCEV *, 4> Operands;
}
}
// (A+B)/C --> (A/C + B/C) if safe and A/C and B/C can be folded.
- if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(LHS)) {
+ if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(LHS)) {
SmallVector<const SCEV *, 4> Operands;
for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i)
Operands.push_back(getZeroExtendExpr(A->getOperand(i), ExtTy));
/// getAddRecExpr - Get an add recurrence expression for the specified loop.
/// Simplify the expression as much as possible.
-const SCEV *ScalarEvolution::getAddRecExpr(const SCEV *Start,
- const SCEV *Step, const Loop *L,
- bool HasNUW, bool HasNSW) {
+const SCEV *ScalarEvolution::getAddRecExpr(const SCEV *Start, const SCEV *Step,
+ const Loop *L,
+ SCEV::NoWrapFlags Flags) {
SmallVector<const SCEV *, 4> Operands;
Operands.push_back(Start);
if (const SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step))
if (StepChrec->getLoop() == L) {
Operands.append(StepChrec->op_begin(), StepChrec->op_end());
- return getAddRecExpr(Operands, L);
+ return getAddRecExpr(Operands, L, maskFlags(Flags, SCEV::FlagNW));
}
Operands.push_back(Step);
- return getAddRecExpr(Operands, L, HasNUW, HasNSW);
+ return getAddRecExpr(Operands, L, Flags);
}
/// getAddRecExpr - Get an add recurrence expression for the specified loop.
/// Simplify the expression as much as possible.
const SCEV *
ScalarEvolution::getAddRecExpr(SmallVectorImpl<const SCEV *> &Operands,
- const Loop *L,
- bool HasNUW, bool HasNSW) {
+ const Loop *L, SCEV::NoWrapFlags Flags) {
if (Operands.size() == 1) return Operands[0];
#ifndef NDEBUG
- const Type *ETy = getEffectiveSCEVType(Operands[0]->getType());
+ Type *ETy = getEffectiveSCEVType(Operands[0]->getType());
for (unsigned i = 1, e = Operands.size(); i != e; ++i)
assert(getEffectiveSCEVType(Operands[i]->getType()) == ETy &&
"SCEVAddRecExpr operand types don't match!");
if (Operands.back()->isZero()) {
Operands.pop_back();
- return getAddRecExpr(Operands, L, HasNUW, HasNSW); // {X,+,0} --> X
+ return getAddRecExpr(Operands, L, SCEV::FlagAnyWrap); // {X,+,0} --> X
}
// It's tempting to want to call getMaxBackedgeTakenCount count here and
// meaningful BE count at this point (and if we don't, we'd be stuck
// with a SCEVCouldNotCompute as the cached BE count).
- // If HasNSW is true and all the operands are non-negative, infer HasNUW.
- if (!HasNUW && HasNSW) {
+ // If FlagNSW is true and all the operands are non-negative, infer FlagNUW.
+ // And vice-versa.
+ int SignOrUnsignMask = SCEV::FlagNUW | SCEV::FlagNSW;
+ SCEV::NoWrapFlags SignOrUnsignWrap = maskFlags(Flags, SignOrUnsignMask);
+ if (SignOrUnsignWrap && (SignOrUnsignWrap != SignOrUnsignMask)) {
bool All = true;
for (SmallVectorImpl<const SCEV *>::const_iterator I = Operands.begin(),
E = Operands.end(); I != E; ++I)
All = false;
break;
}
- if (All) HasNUW = true;
+ if (All) Flags = setFlags(Flags, (SCEV::NoWrapFlags)SignOrUnsignMask);
}
// Canonicalize nested AddRecs in by nesting them in order of loop depth.
break;
}
if (AllInvariant) {
- NestedOperands[0] = getAddRecExpr(Operands, L);
+ // Create a recurrence for the outer loop with the same step size.
+ //
+ // The outer recurrence keeps its NW flag but only keeps NUW/NSW if the
+ // inner recurrence has the same property.
+ SCEV::NoWrapFlags OuterFlags =
+ maskFlags(Flags, SCEV::FlagNW | NestedAR->getNoWrapFlags());
+
+ NestedOperands[0] = getAddRecExpr(Operands, L, OuterFlags);
AllInvariant = true;
for (unsigned i = 0, e = NestedOperands.size(); i != e; ++i)
if (!isLoopInvariant(NestedOperands[i], NestedLoop)) {
AllInvariant = false;
break;
}
- if (AllInvariant)
+ if (AllInvariant) {
// Ok, both add recurrences are valid after the transformation.
- return getAddRecExpr(NestedOperands, NestedLoop, HasNUW, HasNSW);
+ //
+ // The inner recurrence keeps its NW flag but only keeps NUW/NSW if
+ // the outer recurrence has the same property.
+ SCEV::NoWrapFlags InnerFlags =
+ maskFlags(NestedAR->getNoWrapFlags(), SCEV::FlagNW | Flags);
+ return getAddRecExpr(NestedOperands, NestedLoop, InnerFlags);
+ }
}
// Reset Operands to its original state.
Operands[0] = NestedAR;
O, Operands.size(), L);
UniqueSCEVs.InsertNode(S, IP);
}
- if (HasNUW) S->setHasNoUnsignedWrap(true);
- if (HasNSW) S->setHasNoSignedWrap(true);
+ S->setNoWrapFlags(Flags);
return S;
}
assert(!Ops.empty() && "Cannot get empty smax!");
if (Ops.size() == 1) return Ops[0];
#ifndef NDEBUG
- const Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
+ Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
for (unsigned i = 1, e = Ops.size(); i != e; ++i)
assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
"SCEVSMaxExpr operand types don't match!");
assert(!Ops.empty() && "Cannot get empty umax!");
if (Ops.size() == 1) return Ops[0];
#ifndef NDEBUG
- const Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
+ Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
for (unsigned i = 1, e = Ops.size(); i != e; ++i)
assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
"SCEVUMaxExpr operand types don't match!");
return getNotSCEV(getUMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
}
-const SCEV *ScalarEvolution::getSizeOfExpr(const Type *AllocTy) {
+const SCEV *ScalarEvolution::getSizeOfExpr(Type *AllocTy) {
// If we have TargetData, we can bypass creating a target-independent
// constant expression and then folding it back into a ConstantInt.
// This is just a compile-time optimization.
Constant *C = ConstantExpr::getSizeOf(AllocTy);
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
- if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
+ if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
C = Folded;
- const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
+ Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
return getTruncateOrZeroExtend(getSCEV(C), Ty);
}
-const SCEV *ScalarEvolution::getAlignOfExpr(const Type *AllocTy) {
+const SCEV *ScalarEvolution::getAlignOfExpr(Type *AllocTy) {
Constant *C = ConstantExpr::getAlignOf(AllocTy);
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
- if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
+ if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
C = Folded;
- const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
+ Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
return getTruncateOrZeroExtend(getSCEV(C), Ty);
}
-const SCEV *ScalarEvolution::getOffsetOfExpr(const StructType *STy,
+const SCEV *ScalarEvolution::getOffsetOfExpr(StructType *STy,
unsigned FieldNo) {
// If we have TargetData, we can bypass creating a target-independent
// constant expression and then folding it back into a ConstantInt.
Constant *C = ConstantExpr::getOffsetOf(STy, FieldNo);
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
- if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
+ if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
C = Folded;
- const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(STy));
+ Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(STy));
return getTruncateOrZeroExtend(getSCEV(C), Ty);
}
-const SCEV *ScalarEvolution::getOffsetOfExpr(const Type *CTy,
+const SCEV *ScalarEvolution::getOffsetOfExpr(Type *CTy,
Constant *FieldNo) {
Constant *C = ConstantExpr::getOffsetOf(CTy, FieldNo);
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
- if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
+ if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI))
C = Folded;
- const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(CTy));
+ Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(CTy));
return getTruncateOrZeroExtend(getSCEV(C), Ty);
}
/// the SCEV framework. This primarily includes integer types, and it
/// can optionally include pointer types if the ScalarEvolution class
/// has access to target-specific information.
-bool ScalarEvolution::isSCEVable(const Type *Ty) const {
+bool ScalarEvolution::isSCEVable(Type *Ty) const {
// Integers and pointers are always SCEVable.
return Ty->isIntegerTy() || Ty->isPointerTy();
}
/// getTypeSizeInBits - Return the size in bits of the specified type,
/// for which isSCEVable must return true.
-uint64_t ScalarEvolution::getTypeSizeInBits(const Type *Ty) const {
+uint64_t ScalarEvolution::getTypeSizeInBits(Type *Ty) const {
assert(isSCEVable(Ty) && "Type is not SCEVable!");
// If we have a TargetData, use it!
/// the given type and which represents how SCEV will treat the given
/// type, for which isSCEVable must return true. For pointer types,
/// this is the pointer-sized integer type.
-const Type *ScalarEvolution::getEffectiveSCEVType(const Type *Ty) const {
+Type *ScalarEvolution::getEffectiveSCEVType(Type *Ty) const {
assert(isSCEVable(Ty) && "Type is not SCEVable!");
if (Ty->isIntegerTy())
return getConstant(
cast<ConstantInt>(ConstantExpr::getNeg(VC->getValue())));
- const Type *Ty = V->getType();
+ Type *Ty = V->getType();
Ty = getEffectiveSCEVType(Ty);
return getMulExpr(V,
getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty))));
return getConstant(
cast<ConstantInt>(ConstantExpr::getNot(VC->getValue())));
- const Type *Ty = V->getType();
+ Type *Ty = V->getType();
Ty = getEffectiveSCEVType(Ty);
const SCEV *AllOnes =
getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty)));
return getMinusSCEV(AllOnes, V);
}
-/// getMinusSCEV - Return LHS-RHS. Minus is represented in SCEV as A+B*-1,
-/// and thus the HasNUW and HasNSW bits apply to the resultant add, not
-/// whether the sub would have overflowed.
+/// getMinusSCEV - Return LHS-RHS. Minus is represented in SCEV as A+B*-1.
const SCEV *ScalarEvolution::getMinusSCEV(const SCEV *LHS, const SCEV *RHS,
- bool HasNUW, bool HasNSW) {
+ SCEV::NoWrapFlags Flags) {
+ assert(!maskFlags(Flags, SCEV::FlagNUW) && "subtraction does not have NUW");
+
// Fast path: X - X --> 0.
if (LHS == RHS)
return getConstant(LHS->getType(), 0);
// X - Y --> X + -Y
- return getAddExpr(LHS, getNegativeSCEV(RHS), HasNUW, HasNSW);
+ return getAddExpr(LHS, getNegativeSCEV(RHS), Flags);
}
/// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion of the
/// input value to the specified type. If the type must be extended, it is zero
/// extended.
const SCEV *
-ScalarEvolution::getTruncateOrZeroExtend(const SCEV *V, const Type *Ty) {
- const Type *SrcTy = V->getType();
+ScalarEvolution::getTruncateOrZeroExtend(const SCEV *V, Type *Ty) {
+ Type *SrcTy = V->getType();
assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
(Ty->isIntegerTy() || Ty->isPointerTy()) &&
"Cannot truncate or zero extend with non-integer arguments!");
/// extended.
const SCEV *
ScalarEvolution::getTruncateOrSignExtend(const SCEV *V,
- const Type *Ty) {
- const Type *SrcTy = V->getType();
+ Type *Ty) {
+ Type *SrcTy = V->getType();
assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
(Ty->isIntegerTy() || Ty->isPointerTy()) &&
"Cannot truncate or zero extend with non-integer arguments!");
/// input value to the specified type. If the type must be extended, it is zero
/// extended. The conversion must not be narrowing.
const SCEV *
-ScalarEvolution::getNoopOrZeroExtend(const SCEV *V, const Type *Ty) {
- const Type *SrcTy = V->getType();
+ScalarEvolution::getNoopOrZeroExtend(const SCEV *V, Type *Ty) {
+ Type *SrcTy = V->getType();
assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
(Ty->isIntegerTy() || Ty->isPointerTy()) &&
"Cannot noop or zero extend with non-integer arguments!");
/// input value to the specified type. If the type must be extended, it is sign
/// extended. The conversion must not be narrowing.
const SCEV *
-ScalarEvolution::getNoopOrSignExtend(const SCEV *V, const Type *Ty) {
- const Type *SrcTy = V->getType();
+ScalarEvolution::getNoopOrSignExtend(const SCEV *V, Type *Ty) {
+ Type *SrcTy = V->getType();
assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
(Ty->isIntegerTy() || Ty->isPointerTy()) &&
"Cannot noop or sign extend with non-integer arguments!");
/// it is extended with unspecified bits. The conversion must not be
/// narrowing.
const SCEV *
-ScalarEvolution::getNoopOrAnyExtend(const SCEV *V, const Type *Ty) {
- const Type *SrcTy = V->getType();
+ScalarEvolution::getNoopOrAnyExtend(const SCEV *V, Type *Ty) {
+ Type *SrcTy = V->getType();
assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
(Ty->isIntegerTy() || Ty->isPointerTy()) &&
"Cannot noop or any extend with non-integer arguments!");
/// getTruncateOrNoop - Return a SCEV corresponding to a conversion of the
/// input value to the specified type. The conversion must not be widening.
const SCEV *
-ScalarEvolution::getTruncateOrNoop(const SCEV *V, const Type *Ty) {
- const Type *SrcTy = V->getType();
+ScalarEvolution::getTruncateOrNoop(const SCEV *V, Type *Ty) {
+ Type *SrcTy = V->getType();
assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
(Ty->isIntegerTy() || Ty->isPointerTy()) &&
"Cannot truncate or noop with non-integer arguments!");
return getUMinExpr(PromotedLHS, PromotedRHS);
}
+/// getPointerBase - Transitively follow the chain of pointer-type operands
+/// until reaching a SCEV that does not have a single pointer operand. This
+/// returns a SCEVUnknown pointer for well-formed pointer-type expressions,
+/// but corner cases do exist.
+const SCEV *ScalarEvolution::getPointerBase(const SCEV *V) {
+ // A pointer operand may evaluate to a nonpointer expression, such as null.
+ if (!V->getType()->isPointerTy())
+ return V;
+
+ if (const SCEVCastExpr *Cast = dyn_cast<SCEVCastExpr>(V)) {
+ return getPointerBase(Cast->getOperand());
+ }
+ else if (const SCEVNAryExpr *NAry = dyn_cast<SCEVNAryExpr>(V)) {
+ const SCEV *PtrOp = 0;
+ for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
+ I != E; ++I) {
+ if ((*I)->getType()->isPointerTy()) {
+ // Cannot find the base of an expression with multiple pointer operands.
+ if (PtrOp)
+ return V;
+ PtrOp = *I;
+ }
+ }
+ if (!PtrOp)
+ return V;
+ return getPointerBase(PtrOp);
+ }
+ return V;
+}
+
/// PushDefUseChildren - Push users of the given Instruction
/// onto the given Worklist.
static void
if (isLoopInvariant(Accum, L) ||
(isa<SCEVAddRecExpr>(Accum) &&
cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) {
- bool HasNUW = false;
- bool HasNSW = false;
+ SCEV::NoWrapFlags Flags = SCEV::FlagAnyWrap;
// If the increment doesn't overflow, then neither the addrec nor
// the post-increment will overflow.
if (const AddOperator *OBO = dyn_cast<AddOperator>(BEValueV)) {
if (OBO->hasNoUnsignedWrap())
- HasNUW = true;
+ Flags = setFlags(Flags, SCEV::FlagNUW);
if (OBO->hasNoSignedWrap())
- HasNSW = true;
- } else if (const GEPOperator *GEP =
- dyn_cast<GEPOperator>(BEValueV)) {
- // If the increment is a GEP, then we know it won't perform an
- // unsigned overflow, because the address space cannot be
- // wrapped around.
- HasNUW |= GEP->isInBounds();
+ Flags = setFlags(Flags, SCEV::FlagNSW);
+ } else if (const GEPOperator *GEP =
+ dyn_cast<GEPOperator>(BEValueV)) {
+ // If the increment is an inbounds GEP, then we know the address
+ // space cannot be wrapped around. We cannot make any guarantee
+ // about signed or unsigned overflow because pointers are
+ // unsigned but we may have a negative index from the base
+ // pointer.
+ if (GEP->isInBounds())
+ Flags = setFlags(Flags, SCEV::FlagNW);
}
const SCEV *StartVal = getSCEV(StartValueV);
- const SCEV *PHISCEV =
- getAddRecExpr(StartVal, Accum, L, HasNUW, HasNSW);
+ const SCEV *PHISCEV = getAddRecExpr(StartVal, Accum, L, Flags);
// Since the no-wrap flags are on the increment, they apply to the
// post-incremented value as well.
if (isLoopInvariant(Accum, L))
(void)getAddRecExpr(getAddExpr(StartVal, Accum),
- Accum, L, HasNUW, HasNSW);
+ Accum, L, Flags);
// Okay, for the entire analysis of this edge we assumed the PHI
// to be symbolic. We now need to go back and purge all of the
// initial step of the addrec evolution.
if (StartVal == getMinusSCEV(AddRec->getOperand(0),
AddRec->getOperand(1))) {
+ // FIXME: For constant StartVal, we should be able to infer
+ // no-wrap flags.
const SCEV *PHISCEV =
- getAddRecExpr(StartVal, AddRec->getOperand(1), L);
+ getAddRecExpr(StartVal, AddRec->getOperand(1), L,
+ SCEV::FlagAnyWrap);
// Okay, for the entire analysis of this edge we assumed the PHI
// to be symbolic. We now need to go back and purge all of the
// PHI's incoming blocks are in a different loop, in which case doing so
// risks breaking LCSSA form. Instcombine would normally zap these, but
// it doesn't have DominatorTree information, so it may miss cases.
- if (Value *V = SimplifyInstruction(PN, TD, DT))
+ if (Value *V = SimplifyInstruction(PN, TD, TLI, DT))
if (LI->replacementPreservesLCSSAForm(PN, V))
return getSCEV(V);
// Add expression, because the Instruction may be guarded by control flow
// and the no-overflow bits may not be valid for the expression in any
// context.
+ bool isInBounds = GEP->isInBounds();
- const Type *IntPtrTy = getEffectiveSCEVType(GEP->getType());
+ Type *IntPtrTy = getEffectiveSCEVType(GEP->getType());
Value *Base = GEP->getOperand(0);
// Don't attempt to analyze GEPs over unsized objects.
if (!cast<PointerType>(Base->getType())->getElementType()->isSized())
I != E; ++I) {
Value *Index = *I;
// Compute the (potentially symbolic) offset in bytes for this index.
- if (const StructType *STy = dyn_cast<StructType>(*GTI++)) {
+ if (StructType *STy = dyn_cast<StructType>(*GTI++)) {
// For a struct, add the member offset.
unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
const SCEV *FieldOffset = getOffsetOfExpr(STy, FieldNo);
IndexS = getTruncateOrSignExtend(IndexS, IntPtrTy);
// Multiply the index by the element size to compute the element offset.
- const SCEV *LocalOffset = getMulExpr(IndexS, ElementSize);
+ const SCEV *LocalOffset = getMulExpr(IndexS, ElementSize,
+ isInBounds ? SCEV::FlagNSW :
+ SCEV::FlagAnyWrap);
// Add the element offset to the running total offset.
TotalOffset = getAddExpr(TotalOffset, LocalOffset);
const SCEV *BaseS = getSCEV(Base);
// Add the total offset from all the GEP indices to the base.
- return getAddExpr(BaseS, TotalOffset);
+ return getAddExpr(BaseS, TotalOffset,
+ isInBounds ? SCEV::FlagNSW : SCEV::FlagAnyWrap);
}
/// GetMinTrailingZeros - Determine the minimum number of zero bits that S is
if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
// For a SCEVUnknown, ask ValueTracking.
unsigned BitWidth = getTypeSizeInBits(U->getType());
- APInt Mask = APInt::getAllOnesValue(BitWidth);
APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
- ComputeMaskedBits(U->getValue(), Mask, Zeros, Ones);
+ ComputeMaskedBits(U->getValue(), Zeros, Ones);
return Zeros.countTrailingOnes();
}
if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
// If there's no unsigned wrap, the value will never be less than its
// initial value.
- if (AddRec->hasNoUnsignedWrap())
+ if (AddRec->getNoWrapFlags(SCEV::FlagNUW))
if (const SCEVConstant *C = dyn_cast<SCEVConstant>(AddRec->getStart()))
if (!C->getValue()->isZero())
ConservativeResult =
// TODO: non-affine addrec
if (AddRec->isAffine()) {
- const Type *Ty = AddRec->getType();
+ Type *Ty = AddRec->getType();
const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
// For a SCEVUnknown, ask ValueTracking.
- APInt Mask = APInt::getAllOnesValue(BitWidth);
APInt Zeros(BitWidth, 0), Ones(BitWidth, 0);
- ComputeMaskedBits(U->getValue(), Mask, Zeros, Ones, TD);
+ ComputeMaskedBits(U->getValue(), Zeros, Ones, TD);
if (Ones == ~Zeros + 1)
return setUnsignedRange(U, ConservativeResult);
return setUnsignedRange(U,
if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
// If there's no signed wrap, and all the operands have the same sign or
// zero, the value won't ever change sign.
- if (AddRec->hasNoSignedWrap()) {
+ if (AddRec->getNoWrapFlags(SCEV::FlagNSW)) {
bool AllNonNeg = true;
bool AllNonPos = true;
for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
// TODO: non-affine addrec
if (AddRec->isAffine()) {
- const Type *Ty = AddRec->getType();
+ Type *Ty = AddRec->getType();
const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop());
if (!isa<SCEVCouldNotCompute>(MaxBECount) &&
getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) {
// because it leads to N-1 getAddExpr calls for N ultimate operands.
// Instead, gather up all the operands and make a single getAddExpr call.
// LLVM IR canonical form means we need only traverse the left operands.
+ //
+ // Don't apply this instruction's NSW or NUW flags to the new
+ // expression. The instruction may be guarded by control flow that the
+ // no-wrap behavior depends on. Non-control-equivalent instructions can be
+ // mapped to the same SCEV expression, and it would be incorrect to transfer
+ // NSW/NUW semantics to those operations.
SmallVector<const SCEV *, 4> AddOps;
AddOps.push_back(getSCEV(U->getOperand(1)));
for (Value *Op = U->getOperand(0); ; Op = U->getOperand(0)) {
return getAddExpr(AddOps);
}
case Instruction::Mul: {
- // See the Add code above.
+ // Don't transfer NSW/NUW for the same reason as AddExpr.
SmallVector<const SCEV *, 4> MulOps;
MulOps.push_back(getSCEV(U->getOperand(1)));
for (Value *Op = U->getOperand(0);
- Op->getValueID() == Instruction::Mul + Value::InstructionVal;
+ Op->getValueID() == Instruction::Mul + Value::InstructionVal;
Op = U->getOperand(0)) {
U = cast<Operator>(Op);
MulOps.push_back(getSCEV(U->getOperand(1)));
// knew about to reconstruct a low-bits mask value.
unsigned LZ = A.countLeadingZeros();
unsigned BitWidth = A.getBitWidth();
- APInt AllOnes = APInt::getAllOnesValue(BitWidth);
APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
- ComputeMaskedBits(U->getOperand(0), AllOnes, KnownZero, KnownOne, TD);
+ ComputeMaskedBits(U->getOperand(0), KnownZero, KnownOne, TD);
APInt EffectiveMask = APInt::getLowBitsSet(BitWidth, BitWidth - LZ);
// transfer the no-wrap flags, since an or won't introduce a wrap.
if (const SCEVAddRecExpr *NewAR = dyn_cast<SCEVAddRecExpr>(S)) {
const SCEVAddRecExpr *OldAR = cast<SCEVAddRecExpr>(LHS);
- if (OldAR->hasNoUnsignedWrap())
- const_cast<SCEVAddRecExpr *>(NewAR)->setHasNoUnsignedWrap(true);
- if (OldAR->hasNoSignedWrap())
- const_cast<SCEVAddRecExpr *>(NewAR)->setHasNoSignedWrap(true);
+ const_cast<SCEVAddRecExpr *>(NewAR)->setNoWrapFlags(
+ OldAR->getNoWrapFlags());
}
return S;
}
LCI->getValue() == CI->getValue())
if (const SCEVZeroExtendExpr *Z =
dyn_cast<SCEVZeroExtendExpr>(getSCEV(U->getOperand(0)))) {
- const Type *UTy = U->getType();
+ Type *UTy = U->getType();
const SCEV *Z0 = Z->getOperand();
- const Type *Z0Ty = Z0->getType();
+ Type *Z0Ty = Z0->getType();
unsigned Z0TySize = getTypeSizeInBits(Z0Ty);
// If C is a low-bits mask, the zero extend is serving to
// Iteration Count Computation Code
//
+/// getSmallConstantTripCount - Returns the maximum trip count of this loop as a
+/// normal unsigned value. Returns 0 if the trip count is unknown or not
+/// constant. Will also return 0 if the maximum trip count is very large (>=
+/// 2^32).
+///
+/// This "trip count" assumes that control exits via ExitingBlock. More
+/// precisely, it is the number of times that control may reach ExitingBlock
+/// before taking the branch. For loops with multiple exits, it may not be the
+/// number times that the loop header executes because the loop may exit
+/// prematurely via another branch.
+unsigned ScalarEvolution::
+getSmallConstantTripCount(Loop *L, BasicBlock *ExitingBlock) {
+ const SCEVConstant *ExitCount =
+ dyn_cast<SCEVConstant>(getExitCount(L, ExitingBlock));
+ if (!ExitCount)
+ return 0;
+
+ ConstantInt *ExitConst = ExitCount->getValue();
+
+ // Guard against huge trip counts.
+ if (ExitConst->getValue().getActiveBits() > 32)
+ return 0;
+
+ // In case of integer overflow, this returns 0, which is correct.
+ return ((unsigned)ExitConst->getZExtValue()) + 1;
+}
+
+/// getSmallConstantTripMultiple - Returns the largest constant divisor of the
+/// trip count of this loop as a normal unsigned value, if possible. This
+/// means that the actual trip count is always a multiple of the returned
+/// value (don't forget the trip count could very well be zero as well!).
+///
+/// Returns 1 if the trip count is unknown or not guaranteed to be the
+/// multiple of a constant (which is also the case if the trip count is simply
+/// constant, use getSmallConstantTripCount for that case), Will also return 1
+/// if the trip count is very large (>= 2^32).
+///
+/// As explained in the comments for getSmallConstantTripCount, this assumes
+/// that control exits the loop via ExitingBlock.
+unsigned ScalarEvolution::
+getSmallConstantTripMultiple(Loop *L, BasicBlock *ExitingBlock) {
+ const SCEV *ExitCount = getExitCount(L, ExitingBlock);
+ if (ExitCount == getCouldNotCompute())
+ return 1;
+
+ // Get the trip count from the BE count by adding 1.
+ const SCEV *TCMul = getAddExpr(ExitCount,
+ getConstant(ExitCount->getType(), 1));
+ // FIXME: SCEV distributes multiplication as V1*C1 + V2*C1. We could attempt
+ // to factor simple cases.
+ if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(TCMul))
+ TCMul = Mul->getOperand(0);
+
+ const SCEVConstant *MulC = dyn_cast<SCEVConstant>(TCMul);
+ if (!MulC)
+ return 1;
+
+ ConstantInt *Result = MulC->getValue();
+
+ // Guard against huge trip counts.
+ if (!Result || Result->getValue().getActiveBits() > 32)
+ return 1;
+
+ return (unsigned)Result->getZExtValue();
+}
+
+// getExitCount - Get the expression for the number of loop iterations for which
+// this loop is guaranteed not to exit via ExitintBlock. Otherwise return
+// SCEVCouldNotCompute.
+const SCEV *ScalarEvolution::getExitCount(Loop *L, BasicBlock *ExitingBlock) {
+ return getBackedgeTakenInfo(L).getExact(ExitingBlock, this);
+}
+
/// getBackedgeTakenCount - If the specified loop has a predictable
/// backedge-taken count, return it, otherwise return a SCEVCouldNotCompute
/// object. The backedge-taken count is the number of times the loop header
/// hasLoopInvariantBackedgeTakenCount).
///
const SCEV *ScalarEvolution::getBackedgeTakenCount(const Loop *L) {
- return getBackedgeTakenInfo(L).Exact;
+ return getBackedgeTakenInfo(L).getExact(this);
}
/// getMaxBackedgeTakenCount - Similar to getBackedgeTakenCount, except
/// return the least SCEV value that is known never to be less than the
/// actual backedge taken count.
const SCEV *ScalarEvolution::getMaxBackedgeTakenCount(const Loop *L) {
- return getBackedgeTakenInfo(L).Max;
+ return getBackedgeTakenInfo(L).getMax(this);
}
/// PushLoopPHIs - Push PHI nodes in the header of the given loop
const ScalarEvolution::BackedgeTakenInfo &
ScalarEvolution::getBackedgeTakenInfo(const Loop *L) {
- // Initially insert a CouldNotCompute for this loop. If the insertion
+ // Initially insert an invalid entry for this loop. If the insertion
// succeeds, proceed to actually compute a backedge-taken count and
// update the value. The temporary CouldNotCompute value tells SCEV
// code elsewhere that it shouldn't attempt to request a new
// backedge-taken count, which could result in infinite recursion.
- std::pair<std::map<const Loop *, BackedgeTakenInfo>::iterator, bool> Pair =
- BackedgeTakenCounts.insert(std::make_pair(L, getCouldNotCompute()));
+ std::pair<DenseMap<const Loop *, BackedgeTakenInfo>::iterator, bool> Pair =
+ BackedgeTakenCounts.insert(std::make_pair(L, BackedgeTakenInfo()));
if (!Pair.second)
return Pair.first->second;
- BackedgeTakenInfo BECount = ComputeBackedgeTakenCount(L);
- if (BECount.Exact != getCouldNotCompute()) {
- assert(isLoopInvariant(BECount.Exact, L) &&
- isLoopInvariant(BECount.Max, L) &&
+ // ComputeBackedgeTakenCount may allocate memory for its result. Inserting it
+ // into the BackedgeTakenCounts map transfers ownership. Otherwise, the result
+ // must be cleared in this scope.
+ BackedgeTakenInfo Result = ComputeBackedgeTakenCount(L);
+
+ if (Result.getExact(this) != getCouldNotCompute()) {
+ assert(isLoopInvariant(Result.getExact(this), L) &&
+ isLoopInvariant(Result.getMax(this), L) &&
"Computed backedge-taken count isn't loop invariant for loop!");
++NumTripCountsComputed;
-
- // Update the value in the map.
- Pair.first->second = BECount;
- } else {
- if (BECount.Max != getCouldNotCompute())
- // Update the value in the map.
- Pair.first->second = BECount;
- if (isa<PHINode>(L->getHeader()->begin()))
- // Only count loops that have phi nodes as not being computable.
- ++NumTripCountsNotComputed;
+ }
+ else if (Result.getMax(this) == getCouldNotCompute() &&
+ isa<PHINode>(L->getHeader()->begin())) {
+ // Only count loops that have phi nodes as not being computable.
+ ++NumTripCountsNotComputed;
}
// Now that we know more about the trip count for this loop, forget any
// conservative estimates made without the benefit of trip count
// information. This is similar to the code in forgetLoop, except that
// it handles SCEVUnknown PHI nodes specially.
- if (BECount.hasAnyInfo()) {
+ if (Result.hasAnyInfo()) {
SmallVector<Instruction *, 16> Worklist;
PushLoopPHIs(L, Worklist);
PushDefUseChildren(I, Worklist);
}
}
- return Pair.first->second;
+
+ // Re-lookup the insert position, since the call to
+ // ComputeBackedgeTakenCount above could result in a
+ // recusive call to getBackedgeTakenInfo (on a different
+ // loop), which would invalidate the iterator computed
+ // earlier.
+ return BackedgeTakenCounts.find(L)->second = Result;
}
/// forgetLoop - This method should be called by the client when it has
/// compute a trip count, or if the loop is deleted.
void ScalarEvolution::forgetLoop(const Loop *L) {
// Drop any stored trip count value.
- BackedgeTakenCounts.erase(L);
+ DenseMap<const Loop*, BackedgeTakenInfo>::iterator BTCPos =
+ BackedgeTakenCounts.find(L);
+ if (BTCPos != BackedgeTakenCounts.end()) {
+ BTCPos->second.clear();
+ BackedgeTakenCounts.erase(BTCPos);
+ }
// Drop information about expressions based on loop-header PHIs.
SmallVector<Instruction *, 16> Worklist;
}
}
+/// getExact - Get the exact loop backedge taken count considering all loop
+/// exits. A computable result can only be return for loops with a single exit.
+/// Returning the minimum taken count among all exits is incorrect because one
+/// of the loop's exit limit's may have been skipped. HowFarToZero assumes that
+/// the limit of each loop test is never skipped. This is a valid assumption as
+/// long as the loop exits via that test. For precise results, it is the
+/// caller's responsibility to specify the relevant loop exit using
+/// getExact(ExitingBlock, SE).
+const SCEV *
+ScalarEvolution::BackedgeTakenInfo::getExact(ScalarEvolution *SE) const {
+ // If any exits were not computable, the loop is not computable.
+ if (!ExitNotTaken.isCompleteList()) return SE->getCouldNotCompute();
+
+ // We need exactly one computable exit.
+ if (!ExitNotTaken.ExitingBlock) return SE->getCouldNotCompute();
+ assert(ExitNotTaken.ExactNotTaken && "uninitialized not-taken info");
+
+ const SCEV *BECount = 0;
+ for (const ExitNotTakenInfo *ENT = &ExitNotTaken;
+ ENT != 0; ENT = ENT->getNextExit()) {
+
+ assert(ENT->ExactNotTaken != SE->getCouldNotCompute() && "bad exit SCEV");
+
+ if (!BECount)
+ BECount = ENT->ExactNotTaken;
+ else if (BECount != ENT->ExactNotTaken)
+ return SE->getCouldNotCompute();
+ }
+ assert(BECount && "Invalid not taken count for loop exit");
+ return BECount;
+}
+
+/// getExact - Get the exact not taken count for this loop exit.
+const SCEV *
+ScalarEvolution::BackedgeTakenInfo::getExact(BasicBlock *ExitingBlock,
+ ScalarEvolution *SE) const {
+ for (const ExitNotTakenInfo *ENT = &ExitNotTaken;
+ ENT != 0; ENT = ENT->getNextExit()) {
+
+ if (ENT->ExitingBlock == ExitingBlock)
+ return ENT->ExactNotTaken;
+ }
+ return SE->getCouldNotCompute();
+}
+
+/// getMax - Get the max backedge taken count for the loop.
+const SCEV *
+ScalarEvolution::BackedgeTakenInfo::getMax(ScalarEvolution *SE) const {
+ return Max ? Max : SE->getCouldNotCompute();
+}
+
+/// Allocate memory for BackedgeTakenInfo and copy the not-taken count of each
+/// computable exit into a persistent ExitNotTakenInfo array.
+ScalarEvolution::BackedgeTakenInfo::BackedgeTakenInfo(
+ SmallVectorImpl< std::pair<BasicBlock *, const SCEV *> > &ExitCounts,
+ bool Complete, const SCEV *MaxCount) : Max(MaxCount) {
+
+ if (!Complete)
+ ExitNotTaken.setIncomplete();
+
+ unsigned NumExits = ExitCounts.size();
+ if (NumExits == 0) return;
+
+ ExitNotTaken.ExitingBlock = ExitCounts[0].first;
+ ExitNotTaken.ExactNotTaken = ExitCounts[0].second;
+ if (NumExits == 1) return;
+
+ // Handle the rare case of multiple computable exits.
+ ExitNotTakenInfo *ENT = new ExitNotTakenInfo[NumExits-1];
+
+ ExitNotTakenInfo *PrevENT = &ExitNotTaken;
+ for (unsigned i = 1; i < NumExits; ++i, PrevENT = ENT, ++ENT) {
+ PrevENT->setNextExit(ENT);
+ ENT->ExitingBlock = ExitCounts[i].first;
+ ENT->ExactNotTaken = ExitCounts[i].second;
+ }
+}
+
+/// clear - Invalidate this result and free the ExitNotTakenInfo array.
+void ScalarEvolution::BackedgeTakenInfo::clear() {
+ ExitNotTaken.ExitingBlock = 0;
+ ExitNotTaken.ExactNotTaken = 0;
+ delete[] ExitNotTaken.getNextExit();
+}
+
/// ComputeBackedgeTakenCount - Compute the number of times the backedge
/// of the specified loop will execute.
ScalarEvolution::BackedgeTakenInfo
L->getExitingBlocks(ExitingBlocks);
// Examine all exits and pick the most conservative values.
- const SCEV *BECount = getCouldNotCompute();
const SCEV *MaxBECount = getCouldNotCompute();
- bool CouldNotComputeBECount = false;
+ bool CouldComputeBECount = true;
+ SmallVector<std::pair<BasicBlock *, const SCEV *>, 4> ExitCounts;
for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
- BackedgeTakenInfo NewBTI =
- ComputeBackedgeTakenCountFromExit(L, ExitingBlocks[i]);
-
- if (NewBTI.Exact == getCouldNotCompute()) {
+ ExitLimit EL = ComputeExitLimit(L, ExitingBlocks[i]);
+ if (EL.Exact == getCouldNotCompute())
// We couldn't compute an exact value for this exit, so
// we won't be able to compute an exact value for the loop.
- CouldNotComputeBECount = true;
- BECount = getCouldNotCompute();
- } else if (!CouldNotComputeBECount) {
- if (BECount == getCouldNotCompute())
- BECount = NewBTI.Exact;
- else
- BECount = getUMinFromMismatchedTypes(BECount, NewBTI.Exact);
- }
+ CouldComputeBECount = false;
+ else
+ ExitCounts.push_back(std::make_pair(ExitingBlocks[i], EL.Exact));
+
if (MaxBECount == getCouldNotCompute())
- MaxBECount = NewBTI.Max;
- else if (NewBTI.Max != getCouldNotCompute())
- MaxBECount = getUMinFromMismatchedTypes(MaxBECount, NewBTI.Max);
+ MaxBECount = EL.Max;
+ else if (EL.Max != getCouldNotCompute()) {
+ // We cannot take the "min" MaxBECount, because non-unit stride loops may
+ // skip some loop tests. Taking the max over the exits is sufficiently
+ // conservative. TODO: We could do better taking into consideration
+ // that (1) the loop has unit stride (2) the last loop test is
+ // less-than/greater-than (3) any loop test is less-than/greater-than AND
+ // falls-through some constant times less then the other tests.
+ MaxBECount = getUMaxFromMismatchedTypes(MaxBECount, EL.Max);
+ }
}
- return BackedgeTakenInfo(BECount, MaxBECount);
+ return BackedgeTakenInfo(ExitCounts, CouldComputeBECount, MaxBECount);
}
-/// ComputeBackedgeTakenCountFromExit - Compute the number of times the backedge
-/// of the specified loop will execute if it exits via the specified block.
-ScalarEvolution::BackedgeTakenInfo
-ScalarEvolution::ComputeBackedgeTakenCountFromExit(const Loop *L,
- BasicBlock *ExitingBlock) {
+/// ComputeExitLimit - Compute the number of times the backedge of the specified
+/// loop will execute if it exits via the specified block.
+ScalarEvolution::ExitLimit
+ScalarEvolution::ComputeExitLimit(const Loop *L, BasicBlock *ExitingBlock) {
// Okay, we've chosen an exiting block. See what condition causes us to
// exit at this block.
}
// Proceed to the next level to examine the exit condition expression.
- return ComputeBackedgeTakenCountFromExitCond(L, ExitBr->getCondition(),
- ExitBr->getSuccessor(0),
- ExitBr->getSuccessor(1));
+ return ComputeExitLimitFromCond(L, ExitBr->getCondition(),
+ ExitBr->getSuccessor(0),
+ ExitBr->getSuccessor(1));
}
-/// ComputeBackedgeTakenCountFromExitCond - Compute the number of times the
+/// ComputeExitLimitFromCond - Compute the number of times the
/// backedge of the specified loop will execute if its exit condition
/// were a conditional branch of ExitCond, TBB, and FBB.
-ScalarEvolution::BackedgeTakenInfo
-ScalarEvolution::ComputeBackedgeTakenCountFromExitCond(const Loop *L,
- Value *ExitCond,
- BasicBlock *TBB,
- BasicBlock *FBB) {
+ScalarEvolution::ExitLimit
+ScalarEvolution::ComputeExitLimitFromCond(const Loop *L,
+ Value *ExitCond,
+ BasicBlock *TBB,
+ BasicBlock *FBB) {
// Check if the controlling expression for this loop is an And or Or.
if (BinaryOperator *BO = dyn_cast<BinaryOperator>(ExitCond)) {
if (BO->getOpcode() == Instruction::And) {
// Recurse on the operands of the and.
- BackedgeTakenInfo BTI0 =
- ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(0), TBB, FBB);
- BackedgeTakenInfo BTI1 =
- ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(1), TBB, FBB);
+ ExitLimit EL0 = ComputeExitLimitFromCond(L, BO->getOperand(0), TBB, FBB);
+ ExitLimit EL1 = ComputeExitLimitFromCond(L, BO->getOperand(1), TBB, FBB);
const SCEV *BECount = getCouldNotCompute();
const SCEV *MaxBECount = getCouldNotCompute();
if (L->contains(TBB)) {
// Both conditions must be true for the loop to continue executing.
// Choose the less conservative count.
- if (BTI0.Exact == getCouldNotCompute() ||
- BTI1.Exact == getCouldNotCompute())
+ if (EL0.Exact == getCouldNotCompute() ||
+ EL1.Exact == getCouldNotCompute())
BECount = getCouldNotCompute();
else
- BECount = getUMinFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
- if (BTI0.Max == getCouldNotCompute())
- MaxBECount = BTI1.Max;
- else if (BTI1.Max == getCouldNotCompute())
- MaxBECount = BTI0.Max;
+ BECount = getUMinFromMismatchedTypes(EL0.Exact, EL1.Exact);
+ if (EL0.Max == getCouldNotCompute())
+ MaxBECount = EL1.Max;
+ else if (EL1.Max == getCouldNotCompute())
+ MaxBECount = EL0.Max;
else
- MaxBECount = getUMinFromMismatchedTypes(BTI0.Max, BTI1.Max);
+ MaxBECount = getUMinFromMismatchedTypes(EL0.Max, EL1.Max);
} else {
// Both conditions must be true at the same time for the loop to exit.
// For now, be conservative.
assert(L->contains(FBB) && "Loop block has no successor in loop!");
- if (BTI0.Max == BTI1.Max)
- MaxBECount = BTI0.Max;
- if (BTI0.Exact == BTI1.Exact)
- BECount = BTI0.Exact;
+ if (EL0.Max == EL1.Max)
+ MaxBECount = EL0.Max;
+ if (EL0.Exact == EL1.Exact)
+ BECount = EL0.Exact;
}
- return BackedgeTakenInfo(BECount, MaxBECount);
+ return ExitLimit(BECount, MaxBECount);
}
if (BO->getOpcode() == Instruction::Or) {
// Recurse on the operands of the or.
- BackedgeTakenInfo BTI0 =
- ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(0), TBB, FBB);
- BackedgeTakenInfo BTI1 =
- ComputeBackedgeTakenCountFromExitCond(L, BO->getOperand(1), TBB, FBB);
+ ExitLimit EL0 = ComputeExitLimitFromCond(L, BO->getOperand(0), TBB, FBB);
+ ExitLimit EL1 = ComputeExitLimitFromCond(L, BO->getOperand(1), TBB, FBB);
const SCEV *BECount = getCouldNotCompute();
const SCEV *MaxBECount = getCouldNotCompute();
if (L->contains(FBB)) {
// Both conditions must be false for the loop to continue executing.
// Choose the less conservative count.
- if (BTI0.Exact == getCouldNotCompute() ||
- BTI1.Exact == getCouldNotCompute())
+ if (EL0.Exact == getCouldNotCompute() ||
+ EL1.Exact == getCouldNotCompute())
BECount = getCouldNotCompute();
else
- BECount = getUMinFromMismatchedTypes(BTI0.Exact, BTI1.Exact);
- if (BTI0.Max == getCouldNotCompute())
- MaxBECount = BTI1.Max;
- else if (BTI1.Max == getCouldNotCompute())
- MaxBECount = BTI0.Max;
+ BECount = getUMinFromMismatchedTypes(EL0.Exact, EL1.Exact);
+ if (EL0.Max == getCouldNotCompute())
+ MaxBECount = EL1.Max;
+ else if (EL1.Max == getCouldNotCompute())
+ MaxBECount = EL0.Max;
else
- MaxBECount = getUMinFromMismatchedTypes(BTI0.Max, BTI1.Max);
+ MaxBECount = getUMinFromMismatchedTypes(EL0.Max, EL1.Max);
} else {
// Both conditions must be false at the same time for the loop to exit.
// For now, be conservative.
assert(L->contains(TBB) && "Loop block has no successor in loop!");
- if (BTI0.Max == BTI1.Max)
- MaxBECount = BTI0.Max;
- if (BTI0.Exact == BTI1.Exact)
- BECount = BTI0.Exact;
+ if (EL0.Max == EL1.Max)
+ MaxBECount = EL0.Max;
+ if (EL0.Exact == EL1.Exact)
+ BECount = EL0.Exact;
}
- return BackedgeTakenInfo(BECount, MaxBECount);
+ return ExitLimit(BECount, MaxBECount);
}
}
// With an icmp, it may be feasible to compute an exact backedge-taken count.
// Proceed to the next level to examine the icmp.
if (ICmpInst *ExitCondICmp = dyn_cast<ICmpInst>(ExitCond))
- return ComputeBackedgeTakenCountFromExitCondICmp(L, ExitCondICmp, TBB, FBB);
+ return ComputeExitLimitFromICmp(L, ExitCondICmp, TBB, FBB);
// Check for a constant condition. These are normally stripped out by
// SimplifyCFG, but ScalarEvolution may be used by a pass which wishes to
}
// If it's not an integer or pointer comparison then compute it the hard way.
- return ComputeBackedgeTakenCountExhaustively(L, ExitCond, !L->contains(TBB));
-}
-
-static const SCEVAddRecExpr *
-isSimpleUnwrappingAddRec(const SCEV *S, const Loop *L) {
- const SCEVAddRecExpr *SA = dyn_cast<SCEVAddRecExpr>(S);
-
- // The SCEV must be an addrec of this loop.
- if (!SA || SA->getLoop() != L || !SA->isAffine())
- return 0;
-
- // The SCEV must be known to not wrap in some way to be interesting.
- if (!SA->hasNoUnsignedWrap() && !SA->hasNoSignedWrap())
- return 0;
-
- // The stride must be a constant so that we know if it is striding up or down.
- if (!isa<SCEVConstant>(SA->getOperand(1)))
- return 0;
- return SA;
+ return ComputeExitCountExhaustively(L, ExitCond, !L->contains(TBB));
}
-/// getMinusSCEVForExitTest - When considering an exit test for a loop with a
-/// "x != y" exit test, we turn this into a computation that evaluates x-y != 0,
-/// and this function returns the expression to use for x-y. We know and take
-/// advantage of the fact that this subtraction is only being used in a
-/// comparison by zero context.
-///
-static const SCEV *getMinusSCEVForExitTest(const SCEV *LHS, const SCEV *RHS,
- const Loop *L, ScalarEvolution &SE) {
- // If either LHS or RHS is an AddRec SCEV (of this loop) that is known to not
- // wrap (either NSW or NUW), then we know that the value will either become
- // the other one (and thus the loop terminates), that the loop will terminate
- // through some other exit condition first, or that the loop has undefined
- // behavior. This information is useful when the addrec has a stride that is
- // != 1 or -1, because it means we can't "miss" the exit value.
- //
- // In any of these three cases, it is safe to turn the exit condition into a
- // "counting down" AddRec (to zero) by subtracting the two inputs as normal,
- // but since we know that the "end cannot be missed" we can force the
- // resulting AddRec to be a NUW addrec. Since it is counting down, this means
- // that the AddRec *cannot* pass zero.
-
- // See if LHS and RHS are addrec's we can handle.
- const SCEVAddRecExpr *LHSA = isSimpleUnwrappingAddRec(LHS, L);
- const SCEVAddRecExpr *RHSA = isSimpleUnwrappingAddRec(RHS, L);
-
- // If neither addrec is interesting, just return a minus.
- if (RHSA == 0 && LHSA == 0)
- return SE.getMinusSCEV(LHS, RHS);
-
- // If only one of LHS and RHS are an AddRec of this loop, make sure it is LHS.
- if (RHSA && LHSA == 0) {
- // Safe because a-b === b-a for comparisons against zero.
- std::swap(LHS, RHS);
- std::swap(LHSA, RHSA);
- }
-
- // Handle the case when only one is advancing in a non-overflowing way.
- if (RHSA == 0) {
- // If RHS is loop varying, then we can't predict when LHS will cross it.
- if (!SE.isLoopInvariant(RHS, L))
- return SE.getMinusSCEV(LHS, RHS);
-
- // If LHS has a positive stride, then we compute RHS-LHS, because the loop
- // is counting up until it crosses RHS (which must be larger than LHS). If
- // it is negative, we compute LHS-RHS because we're counting down to RHS.
- const ConstantInt *Stride =
- cast<SCEVConstant>(LHSA->getOperand(1))->getValue();
- if (Stride->getValue().isNegative())
- std::swap(LHS, RHS);
-
- return SE.getMinusSCEV(RHS, LHS, true /*HasNUW*/);
- }
-
- // If both LHS and RHS are interesting, we have something like:
- // a+i*4 != b+i*8.
- const ConstantInt *LHSStride =
- cast<SCEVConstant>(LHSA->getOperand(1))->getValue();
- const ConstantInt *RHSStride =
- cast<SCEVConstant>(RHSA->getOperand(1))->getValue();
-
- // If the strides are equal, then this is just a (complex) loop invariant
- // comparison of a and b.
- if (LHSStride == RHSStride)
- return SE.getMinusSCEV(LHSA->getStart(), RHSA->getStart());
-
- // If the signs of the strides differ, then the negative stride is counting
- // down to the positive stride.
- if (LHSStride->getValue().isNegative() != RHSStride->getValue().isNegative()){
- if (RHSStride->getValue().isNegative())
- std::swap(LHS, RHS);
- } else {
- // If LHS's stride is smaller than RHS's stride, then "b" must be less than
- // "a" and "b" is RHS is counting up (catching up) to LHS. This is true
- // whether the strides are positive or negative.
- if (RHSStride->getValue().slt(LHSStride->getValue()))
- std::swap(LHS, RHS);
- }
-
- return SE.getMinusSCEV(LHS, RHS, true /*HasNUW*/);
-}
-
-/// ComputeBackedgeTakenCountFromExitCondICmp - Compute the number of times the
+/// ComputeExitLimitFromICmp - Compute the number of times the
/// backedge of the specified loop will execute if its exit condition
/// were a conditional branch of the ICmpInst ExitCond, TBB, and FBB.
-ScalarEvolution::BackedgeTakenInfo
-ScalarEvolution::ComputeBackedgeTakenCountFromExitCondICmp(const Loop *L,
- ICmpInst *ExitCond,
- BasicBlock *TBB,
- BasicBlock *FBB) {
+ScalarEvolution::ExitLimit
+ScalarEvolution::ComputeExitLimitFromICmp(const Loop *L,
+ ICmpInst *ExitCond,
+ BasicBlock *TBB,
+ BasicBlock *FBB) {
// If the condition was exit on true, convert the condition to exit on false
ICmpInst::Predicate Cond;
// Handle common loops like: for (X = "string"; *X; ++X)
if (LoadInst *LI = dyn_cast<LoadInst>(ExitCond->getOperand(0)))
if (Constant *RHS = dyn_cast<Constant>(ExitCond->getOperand(1))) {
- BackedgeTakenInfo ItCnt =
- ComputeLoadConstantCompareBackedgeTakenCount(LI, RHS, L, Cond);
+ ExitLimit ItCnt =
+ ComputeLoadConstantCompareExitLimit(LI, RHS, L, Cond);
if (ItCnt.hasAnyInfo())
return ItCnt;
}
switch (Cond) {
case ICmpInst::ICMP_NE: { // while (X != Y)
// Convert to: while (X-Y != 0)
- BackedgeTakenInfo BTI = HowFarToZero(getMinusSCEVForExitTest(LHS, RHS, L,
- *this), L);
- if (BTI.hasAnyInfo()) return BTI;
+ ExitLimit EL = HowFarToZero(getMinusSCEV(LHS, RHS), L);
+ if (EL.hasAnyInfo()) return EL;
break;
}
case ICmpInst::ICMP_EQ: { // while (X == Y)
// Convert to: while (X-Y == 0)
- BackedgeTakenInfo BTI = HowFarToNonZero(getMinusSCEV(LHS, RHS), L);
- if (BTI.hasAnyInfo()) return BTI;
+ ExitLimit EL = HowFarToNonZero(getMinusSCEV(LHS, RHS), L);
+ if (EL.hasAnyInfo()) return EL;
break;
}
case ICmpInst::ICMP_SLT: {
- BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, true);
- if (BTI.hasAnyInfo()) return BTI;
+ ExitLimit EL = HowManyLessThans(LHS, RHS, L, true);
+ if (EL.hasAnyInfo()) return EL;
break;
}
case ICmpInst::ICMP_SGT: {
- BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
+ ExitLimit EL = HowManyLessThans(getNotSCEV(LHS),
getNotSCEV(RHS), L, true);
- if (BTI.hasAnyInfo()) return BTI;
+ if (EL.hasAnyInfo()) return EL;
break;
}
case ICmpInst::ICMP_ULT: {
- BackedgeTakenInfo BTI = HowManyLessThans(LHS, RHS, L, false);
- if (BTI.hasAnyInfo()) return BTI;
+ ExitLimit EL = HowManyLessThans(LHS, RHS, L, false);
+ if (EL.hasAnyInfo()) return EL;
break;
}
case ICmpInst::ICMP_UGT: {
- BackedgeTakenInfo BTI = HowManyLessThans(getNotSCEV(LHS),
+ ExitLimit EL = HowManyLessThans(getNotSCEV(LHS),
getNotSCEV(RHS), L, false);
- if (BTI.hasAnyInfo()) return BTI;
+ if (EL.hasAnyInfo()) return EL;
break;
}
default:
#endif
break;
}
- return
- ComputeBackedgeTakenCountExhaustively(L, ExitCond, !L->contains(TBB));
+ return ComputeExitCountExhaustively(L, ExitCond, !L->contains(TBB));
}
static ConstantInt *
return cast<SCEVConstant>(Val)->getValue();
}
-/// GetAddressedElementFromGlobal - Given a global variable with an initializer
-/// and a GEP expression (missing the pointer index) indexing into it, return
-/// the addressed element of the initializer or null if the index expression is
-/// invalid.
-static Constant *
-GetAddressedElementFromGlobal(GlobalVariable *GV,
- const std::vector<ConstantInt*> &Indices) {
- Constant *Init = GV->getInitializer();
- for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
- uint64_t Idx = Indices[i]->getZExtValue();
- if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
- assert(Idx < CS->getNumOperands() && "Bad struct index!");
- Init = cast<Constant>(CS->getOperand(Idx));
- } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
- if (Idx >= CA->getNumOperands()) return 0; // Bogus program
- Init = cast<Constant>(CA->getOperand(Idx));
- } else if (isa<ConstantAggregateZero>(Init)) {
- if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
- assert(Idx < STy->getNumElements() && "Bad struct index!");
- Init = Constant::getNullValue(STy->getElementType(Idx));
- } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Init->getType())) {
- if (Idx >= ATy->getNumElements()) return 0; // Bogus program
- Init = Constant::getNullValue(ATy->getElementType());
- } else {
- llvm_unreachable("Unknown constant aggregate type!");
- }
- return 0;
- } else {
- return 0; // Unknown initializer type
- }
- }
- return Init;
-}
-
-/// ComputeLoadConstantCompareBackedgeTakenCount - Given an exit condition of
+/// ComputeLoadConstantCompareExitLimit - Given an exit condition of
/// 'icmp op load X, cst', try to see if we can compute the backedge
/// execution count.
-ScalarEvolution::BackedgeTakenInfo
-ScalarEvolution::ComputeLoadConstantCompareBackedgeTakenCount(
- LoadInst *LI,
- Constant *RHS,
- const Loop *L,
- ICmpInst::Predicate predicate) {
+ScalarEvolution::ExitLimit
+ScalarEvolution::ComputeLoadConstantCompareExitLimit(
+ LoadInst *LI,
+ Constant *RHS,
+ const Loop *L,
+ ICmpInst::Predicate predicate) {
+
if (LI->isVolatile()) return getCouldNotCompute();
// Check to see if the loaded pointer is a getelementptr of a global.
// Okay, we allow one non-constant index into the GEP instruction.
Value *VarIdx = 0;
- std::vector<ConstantInt*> Indexes;
+ std::vector<Constant*> Indexes;
unsigned VarIdxNum = 0;
for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i)
if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
Indexes.push_back(0);
}
+ // Loop-invariant loads may be a byproduct of loop optimization. Skip them.
+ if (!VarIdx)
+ return getCouldNotCompute();
+
// Okay, we know we have a (load (gep GV, 0, X)) comparison with a constant.
// Check to see if X is a loop variant variable value now.
const SCEV *Idx = getSCEV(VarIdx);
// Form the GEP offset.
Indexes[VarIdxNum] = Val;
- Constant *Result = GetAddressedElementFromGlobal(GV, Indexes);
+ Constant *Result = ConstantFoldLoadThroughGEPIndices(GV->getInitializer(),
+ Indexes);
if (Result == 0) break; // Cannot compute!
// Evaluate the condition for this iteration.
/// specified type, assuming that all operands were constants.
static bool CanConstantFold(const Instruction *I) {
if (isa<BinaryOperator>(I) || isa<CmpInst>(I) ||
- isa<SelectInst>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I))
+ isa<SelectInst>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I) ||
+ isa<LoadInst>(I))
return true;
if (const CallInst *CI = dyn_cast<CallInst>(I))
return false;
}
-/// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node
-/// in the loop that V is derived from. We allow arbitrary operations along the
-/// way, but the operands of an operation must either be constants or a value
-/// derived from a constant PHI. If this expression does not fit with these
-/// constraints, return null.
-static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) {
- // If this is not an instruction, or if this is an instruction outside of the
- // loop, it can't be derived from a loop PHI.
- Instruction *I = dyn_cast<Instruction>(V);
- if (I == 0 || !L->contains(I)) return 0;
+/// Determine whether this instruction can constant evolve within this loop
+/// assuming its operands can all constant evolve.
+static bool canConstantEvolve(Instruction *I, const Loop *L) {
+ // An instruction outside of the loop can't be derived from a loop PHI.
+ if (!L->contains(I)) return false;
- if (PHINode *PN = dyn_cast<PHINode>(I)) {
+ if (isa<PHINode>(I)) {
if (L->getHeader() == I->getParent())
- return PN;
+ return true;
else
// We don't currently keep track of the control flow needed to evaluate
// PHIs, so we cannot handle PHIs inside of loops.
- return 0;
+ return false;
}
// If we won't be able to constant fold this expression even if the operands
- // are constants, return early.
- if (!CanConstantFold(I)) return 0;
+ // are constants, bail early.
+ return CanConstantFold(I);
+}
+
+/// getConstantEvolvingPHIOperands - Implement getConstantEvolvingPHI by
+/// recursing through each instruction operand until reaching a loop header phi.
+static PHINode *
+getConstantEvolvingPHIOperands(Instruction *UseInst, const Loop *L,
+ DenseMap<Instruction *, PHINode *> &PHIMap) {
// Otherwise, we can evaluate this instruction if all of its operands are
// constant or derived from a PHI node themselves.
PHINode *PHI = 0;
- for (unsigned Op = 0, e = I->getNumOperands(); Op != e; ++Op)
- if (!isa<Constant>(I->getOperand(Op))) {
- PHINode *P = getConstantEvolvingPHI(I->getOperand(Op), L);
- if (P == 0) return 0; // Not evolving from PHI
- if (PHI == 0)
- PHI = P;
- else if (PHI != P)
- return 0; // Evolving from multiple different PHIs.
+ for (Instruction::op_iterator OpI = UseInst->op_begin(),
+ OpE = UseInst->op_end(); OpI != OpE; ++OpI) {
+
+ if (isa<Constant>(*OpI)) continue;
+
+ Instruction *OpInst = dyn_cast<Instruction>(*OpI);
+ if (!OpInst || !canConstantEvolve(OpInst, L)) return 0;
+
+ PHINode *P = dyn_cast<PHINode>(OpInst);
+ if (!P)
+ // If this operand is already visited, reuse the prior result.
+ // We may have P != PHI if this is the deepest point at which the
+ // inconsistent paths meet.
+ P = PHIMap.lookup(OpInst);
+ if (!P) {
+ // Recurse and memoize the results, whether a phi is found or not.
+ // This recursive call invalidates pointers into PHIMap.
+ P = getConstantEvolvingPHIOperands(OpInst, L, PHIMap);
+ PHIMap[OpInst] = P;
}
-
+ if (P == 0) return 0; // Not evolving from PHI
+ if (PHI && PHI != P) return 0; // Evolving from multiple different PHIs.
+ PHI = P;
+ }
// This is a expression evolving from a constant PHI!
return PHI;
}
+/// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node
+/// in the loop that V is derived from. We allow arbitrary operations along the
+/// way, but the operands of an operation must either be constants or a value
+/// derived from a constant PHI. If this expression does not fit with these
+/// constraints, return null.
+static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) {
+ Instruction *I = dyn_cast<Instruction>(V);
+ if (I == 0 || !canConstantEvolve(I, L)) return 0;
+
+ if (PHINode *PN = dyn_cast<PHINode>(I)) {
+ return PN;
+ }
+
+ // Record non-constant instructions contained by the loop.
+ DenseMap<Instruction *, PHINode *> PHIMap;
+ return getConstantEvolvingPHIOperands(I, L, PHIMap);
+}
+
/// EvaluateExpression - Given an expression that passes the
/// getConstantEvolvingPHI predicate, evaluate its value assuming the PHI node
/// in the loop has the value PHIVal. If we can't fold this expression for some
/// reason, return null.
-static Constant *EvaluateExpression(Value *V, Constant *PHIVal,
- const TargetData *TD) {
- if (isa<PHINode>(V)) return PHIVal;
+static Constant *EvaluateExpression(Value *V, const Loop *L,
+ DenseMap<Instruction *, Constant *> &Vals,
+ const TargetData *TD,
+ const TargetLibraryInfo *TLI) {
+ // Convenient constant check, but redundant for recursive calls.
if (Constant *C = dyn_cast<Constant>(V)) return C;
- Instruction *I = cast<Instruction>(V);
+ Instruction *I = dyn_cast<Instruction>(V);
+ if (!I) return 0;
+
+ if (Constant *C = Vals.lookup(I)) return C;
+
+ // An instruction inside the loop depends on a value outside the loop that we
+ // weren't given a mapping for, or a value such as a call inside the loop.
+ if (!canConstantEvolve(I, L)) return 0;
+
+ // An unmapped PHI can be due to a branch or another loop inside this loop,
+ // or due to this not being the initial iteration through a loop where we
+ // couldn't compute the evolution of this particular PHI last time.
+ if (isa<PHINode>(I)) return 0;
std::vector<Constant*> Operands(I->getNumOperands());
for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
- Operands[i] = EvaluateExpression(I->getOperand(i), PHIVal, TD);
- if (Operands[i] == 0) return 0;
+ Instruction *Operand = dyn_cast<Instruction>(I->getOperand(i));
+ if (!Operand) {
+ Operands[i] = dyn_cast<Constant>(I->getOperand(i));
+ if (!Operands[i]) return 0;
+ continue;
+ }
+ Constant *C = EvaluateExpression(Operand, L, Vals, TD, TLI);
+ Vals[Operand] = C;
+ if (!C) return 0;
+ Operands[i] = C;
}
- if (const CmpInst *CI = dyn_cast<CmpInst>(I))
+ if (CmpInst *CI = dyn_cast<CmpInst>(I))
return ConstantFoldCompareInstOperands(CI->getPredicate(), Operands[0],
- Operands[1], TD);
- return ConstantFoldInstOperands(I->getOpcode(), I->getType(),
- &Operands[0], Operands.size(), TD);
+ Operands[1], TD, TLI);
+ if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
+ if (!LI->isVolatile())
+ return ConstantFoldLoadFromConstPtr(Operands[0], TD);
+ }
+ return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Operands, TD,
+ TLI);
}
/// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
ScalarEvolution::getConstantEvolutionLoopExitValue(PHINode *PN,
const APInt &BEs,
const Loop *L) {
- std::map<PHINode*, Constant*>::const_iterator I =
+ DenseMap<PHINode*, Constant*>::const_iterator I =
ConstantEvolutionLoopExitValue.find(PN);
if (I != ConstantEvolutionLoopExitValue.end())
return I->second;
Constant *&RetVal = ConstantEvolutionLoopExitValue[PN];
+ DenseMap<Instruction *, Constant *> CurrentIterVals;
+ BasicBlock *Header = L->getHeader();
+ assert(PN->getParent() == Header && "Can't evaluate PHI not in loop header!");
+
// Since the loop is canonicalized, the PHI node must have two entries. One
// entry must be a constant (coming in from outside of the loop), and the
// second must be derived from the same PHI.
bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
- Constant *StartCST =
- dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
- if (StartCST == 0)
- return RetVal = 0; // Must be a constant.
+ PHINode *PHI = 0;
+ for (BasicBlock::iterator I = Header->begin();
+ (PHI = dyn_cast<PHINode>(I)); ++I) {
+ Constant *StartCST =
+ dyn_cast<Constant>(PHI->getIncomingValue(!SecondIsBackedge));
+ if (StartCST == 0) continue;
+ CurrentIterVals[PHI] = StartCST;
+ }
+ if (!CurrentIterVals.count(PN))
+ return RetVal = 0;
Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
- if (getConstantEvolvingPHI(BEValue, L) != PN &&
- !isa<Constant>(BEValue))
- return RetVal = 0; // Not derived from same PHI.
// Execute the loop symbolically to determine the exit value.
if (BEs.getActiveBits() >= 32)
unsigned NumIterations = BEs.getZExtValue(); // must be in range
unsigned IterationNum = 0;
- for (Constant *PHIVal = StartCST; ; ++IterationNum) {
+ for (; ; ++IterationNum) {
if (IterationNum == NumIterations)
- return RetVal = PHIVal; // Got exit value!
+ return RetVal = CurrentIterVals[PN]; // Got exit value!
- // Compute the value of the PHI node for the next iteration.
- Constant *NextPHI = EvaluateExpression(BEValue, PHIVal, TD);
- if (NextPHI == PHIVal)
- return RetVal = NextPHI; // Stopped evolving!
+ // Compute the value of the PHIs for the next iteration.
+ // EvaluateExpression adds non-phi values to the CurrentIterVals map.
+ DenseMap<Instruction *, Constant *> NextIterVals;
+ Constant *NextPHI = EvaluateExpression(BEValue, L, CurrentIterVals, TD,
+ TLI);
if (NextPHI == 0)
return 0; // Couldn't evaluate!
- PHIVal = NextPHI;
+ NextIterVals[PN] = NextPHI;
+
+ bool StoppedEvolving = NextPHI == CurrentIterVals[PN];
+
+ // Also evaluate the other PHI nodes. However, we don't get to stop if we
+ // cease to be able to evaluate one of them or if they stop evolving,
+ // because that doesn't necessarily prevent us from computing PN.
+ SmallVector<std::pair<PHINode *, Constant *>, 8> PHIsToCompute;
+ for (DenseMap<Instruction *, Constant *>::const_iterator
+ I = CurrentIterVals.begin(), E = CurrentIterVals.end(); I != E; ++I){
+ PHINode *PHI = dyn_cast<PHINode>(I->first);
+ if (!PHI || PHI == PN || PHI->getParent() != Header) continue;
+ PHIsToCompute.push_back(std::make_pair(PHI, I->second));
+ }
+ // We use two distinct loops because EvaluateExpression may invalidate any
+ // iterators into CurrentIterVals.
+ for (SmallVectorImpl<std::pair<PHINode *, Constant*> >::const_iterator
+ I = PHIsToCompute.begin(), E = PHIsToCompute.end(); I != E; ++I) {
+ PHINode *PHI = I->first;
+ Constant *&NextPHI = NextIterVals[PHI];
+ if (!NextPHI) { // Not already computed.
+ Value *BEValue = PHI->getIncomingValue(SecondIsBackedge);
+ NextPHI = EvaluateExpression(BEValue, L, CurrentIterVals, TD, TLI);
+ }
+ if (NextPHI != I->second)
+ StoppedEvolving = false;
+ }
+
+ // If all entries in CurrentIterVals == NextIterVals then we can stop
+ // iterating, the loop can't continue to change.
+ if (StoppedEvolving)
+ return RetVal = CurrentIterVals[PN];
+
+ CurrentIterVals.swap(NextIterVals);
}
}
-/// ComputeBackedgeTakenCountExhaustively - If the loop is known to execute a
+/// ComputeExitCountExhaustively - If the loop is known to execute a
/// constant number of times (the condition evolves only from constants),
/// try to evaluate a few iterations of the loop until we get the exit
/// condition gets a value of ExitWhen (true or false). If we cannot
/// evaluate the trip count of the loop, return getCouldNotCompute().
-const SCEV *
-ScalarEvolution::ComputeBackedgeTakenCountExhaustively(const Loop *L,
- Value *Cond,
- bool ExitWhen) {
+const SCEV *ScalarEvolution::ComputeExitCountExhaustively(const Loop *L,
+ Value *Cond,
+ bool ExitWhen) {
PHINode *PN = getConstantEvolvingPHI(Cond, L);
if (PN == 0) return getCouldNotCompute();
// That's the only form we support here.
if (PN->getNumIncomingValues() != 2) return getCouldNotCompute();
+ DenseMap<Instruction *, Constant *> CurrentIterVals;
+ BasicBlock *Header = L->getHeader();
+ assert(PN->getParent() == Header && "Can't evaluate PHI not in loop header!");
+
// One entry must be a constant (coming in from outside of the loop), and the
// second must be derived from the same PHI.
bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
- Constant *StartCST =
- dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge));
- if (StartCST == 0) return getCouldNotCompute(); // Must be a constant.
-
- Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
- if (getConstantEvolvingPHI(BEValue, L) != PN &&
- !isa<Constant>(BEValue))
- return getCouldNotCompute(); // Not derived from same PHI.
+ PHINode *PHI = 0;
+ for (BasicBlock::iterator I = Header->begin();
+ (PHI = dyn_cast<PHINode>(I)); ++I) {
+ Constant *StartCST =
+ dyn_cast<Constant>(PHI->getIncomingValue(!SecondIsBackedge));
+ if (StartCST == 0) continue;
+ CurrentIterVals[PHI] = StartCST;
+ }
+ if (!CurrentIterVals.count(PN))
+ return getCouldNotCompute();
// Okay, we find a PHI node that defines the trip count of this loop. Execute
// the loop symbolically to determine when the condition gets a value of
// "ExitWhen".
- unsigned IterationNum = 0;
+
unsigned MaxIterations = MaxBruteForceIterations; // Limit analysis.
- for (Constant *PHIVal = StartCST;
- IterationNum != MaxIterations; ++IterationNum) {
+ for (unsigned IterationNum = 0; IterationNum != MaxIterations;++IterationNum){
ConstantInt *CondVal =
- dyn_cast_or_null<ConstantInt>(EvaluateExpression(Cond, PHIVal, TD));
+ dyn_cast_or_null<ConstantInt>(EvaluateExpression(Cond, L, CurrentIterVals,
+ TD, TLI));
// Couldn't symbolically evaluate.
if (!CondVal) return getCouldNotCompute();
return getConstant(Type::getInt32Ty(getContext()), IterationNum);
}
- // Compute the value of the PHI node for the next iteration.
- Constant *NextPHI = EvaluateExpression(BEValue, PHIVal, TD);
- if (NextPHI == 0 || NextPHI == PHIVal)
- return getCouldNotCompute();// Couldn't evaluate or not making progress...
- PHIVal = NextPHI;
+ // Update all the PHI nodes for the next iteration.
+ DenseMap<Instruction *, Constant *> NextIterVals;
+
+ // Create a list of which PHIs we need to compute. We want to do this before
+ // calling EvaluateExpression on them because that may invalidate iterators
+ // into CurrentIterVals.
+ SmallVector<PHINode *, 8> PHIsToCompute;
+ for (DenseMap<Instruction *, Constant *>::const_iterator
+ I = CurrentIterVals.begin(), E = CurrentIterVals.end(); I != E; ++I){
+ PHINode *PHI = dyn_cast<PHINode>(I->first);
+ if (!PHI || PHI->getParent() != Header) continue;
+ PHIsToCompute.push_back(PHI);
+ }
+ for (SmallVectorImpl<PHINode *>::const_iterator I = PHIsToCompute.begin(),
+ E = PHIsToCompute.end(); I != E; ++I) {
+ PHINode *PHI = *I;
+ Constant *&NextPHI = NextIterVals[PHI];
+ if (NextPHI) continue; // Already computed!
+
+ Value *BEValue = PHI->getIncomingValue(SecondIsBackedge);
+ NextPHI = EvaluateExpression(BEValue, L, CurrentIterVals, TD, TLI);
+ }
+ CurrentIterVals.swap(NextIterVals);
}
// Too many iterations were needed to evaluate.
return C;
}
+/// This builds up a Constant using the ConstantExpr interface. That way, we
+/// will return Constants for objects which aren't represented by a
+/// SCEVConstant, because SCEVConstant is restricted to ConstantInt.
+/// Returns NULL if the SCEV isn't representable as a Constant.
+static Constant *BuildConstantFromSCEV(const SCEV *V) {
+ switch (V->getSCEVType()) {
+ default: // TODO: smax, umax.
+ case scCouldNotCompute:
+ case scAddRecExpr:
+ break;
+ case scConstant:
+ return cast<SCEVConstant>(V)->getValue();
+ case scUnknown:
+ return dyn_cast<Constant>(cast<SCEVUnknown>(V)->getValue());
+ case scSignExtend: {
+ const SCEVSignExtendExpr *SS = cast<SCEVSignExtendExpr>(V);
+ if (Constant *CastOp = BuildConstantFromSCEV(SS->getOperand()))
+ return ConstantExpr::getSExt(CastOp, SS->getType());
+ break;
+ }
+ case scZeroExtend: {
+ const SCEVZeroExtendExpr *SZ = cast<SCEVZeroExtendExpr>(V);
+ if (Constant *CastOp = BuildConstantFromSCEV(SZ->getOperand()))
+ return ConstantExpr::getZExt(CastOp, SZ->getType());
+ break;
+ }
+ case scTruncate: {
+ const SCEVTruncateExpr *ST = cast<SCEVTruncateExpr>(V);
+ if (Constant *CastOp = BuildConstantFromSCEV(ST->getOperand()))
+ return ConstantExpr::getTrunc(CastOp, ST->getType());
+ break;
+ }
+ case scAddExpr: {
+ const SCEVAddExpr *SA = cast<SCEVAddExpr>(V);
+ if (Constant *C = BuildConstantFromSCEV(SA->getOperand(0))) {
+ if (C->getType()->isPointerTy())
+ C = ConstantExpr::getBitCast(C, Type::getInt8PtrTy(C->getContext()));
+ for (unsigned i = 1, e = SA->getNumOperands(); i != e; ++i) {
+ Constant *C2 = BuildConstantFromSCEV(SA->getOperand(i));
+ if (!C2) return 0;
+
+ // First pointer!
+ if (!C->getType()->isPointerTy() && C2->getType()->isPointerTy()) {
+ std::swap(C, C2);
+ // The offsets have been converted to bytes. We can add bytes to an
+ // i8* by GEP with the byte count in the first index.
+ C = ConstantExpr::getBitCast(C,Type::getInt8PtrTy(C->getContext()));
+ }
+
+ // Don't bother trying to sum two pointers. We probably can't
+ // statically compute a load that results from it anyway.
+ if (C2->getType()->isPointerTy())
+ return 0;
+
+ if (C->getType()->isPointerTy()) {
+ if (cast<PointerType>(C->getType())->getElementType()->isStructTy())
+ C2 = ConstantExpr::getIntegerCast(
+ C2, Type::getInt32Ty(C->getContext()), true);
+ C = ConstantExpr::getGetElementPtr(C, C2);
+ } else
+ C = ConstantExpr::getAdd(C, C2);
+ }
+ return C;
+ }
+ break;
+ }
+ case scMulExpr: {
+ const SCEVMulExpr *SM = cast<SCEVMulExpr>(V);
+ if (Constant *C = BuildConstantFromSCEV(SM->getOperand(0))) {
+ // Don't bother with pointers at all.
+ if (C->getType()->isPointerTy()) return 0;
+ for (unsigned i = 1, e = SM->getNumOperands(); i != e; ++i) {
+ Constant *C2 = BuildConstantFromSCEV(SM->getOperand(i));
+ if (!C2 || C2->getType()->isPointerTy()) return 0;
+ C = ConstantExpr::getMul(C, C2);
+ }
+ return C;
+ }
+ break;
+ }
+ case scUDivExpr: {
+ const SCEVUDivExpr *SU = cast<SCEVUDivExpr>(V);
+ if (Constant *LHS = BuildConstantFromSCEV(SU->getLHS()))
+ if (Constant *RHS = BuildConstantFromSCEV(SU->getRHS()))
+ if (LHS->getType() == RHS->getType())
+ return ConstantExpr::getUDiv(LHS, RHS);
+ break;
+ }
+ }
+ return 0;
+}
+
const SCEV *ScalarEvolution::computeSCEVAtScope(const SCEV *V, const Loop *L) {
if (isa<SCEVConstant>(V)) return V;
const SCEV *OpV = getSCEVAtScope(OrigV, L);
MadeImprovement |= OrigV != OpV;
- Constant *C = 0;
- if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(OpV))
- C = SC->getValue();
- if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(OpV))
- C = dyn_cast<Constant>(SU->getValue());
+ Constant *C = BuildConstantFromSCEV(OpV);
if (!C) return V;
if (C->getType() != Op->getType())
C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
Constant *C = 0;
if (const CmpInst *CI = dyn_cast<CmpInst>(I))
C = ConstantFoldCompareInstOperands(CI->getPredicate(),
- Operands[0], Operands[1], TD);
- else
+ Operands[0], Operands[1], TD,
+ TLI);
+ else if (const LoadInst *LI = dyn_cast<LoadInst>(I)) {
+ if (!LI->isVolatile())
+ C = ConstantFoldLoadFromConstPtr(Operands[0], TD);
+ } else
C = ConstantFoldInstOperands(I->getOpcode(), I->getType(),
- &Operands[0], Operands.size(), TD);
+ Operands, TD, TLI);
if (!C) return V;
return getSCEV(C);
}
for (++i; i != e; ++i)
NewOps.push_back(getSCEVAtScope(AddRec->getOperand(i), L));
- AddRec = cast<SCEVAddRecExpr>(getAddRecExpr(NewOps, AddRec->getLoop()));
+ const SCEV *FoldedRec =
+ getAddRecExpr(NewOps, AddRec->getLoop(),
+ AddRec->getNoWrapFlags(SCEV::FlagNW));
+ AddRec = dyn_cast<SCEVAddRecExpr>(FoldedRec);
+ // The addrec may be folded to a nonrecurrence, for example, if the
+ // induction variable is multiplied by zero after constant folding. Go
+ // ahead and return the folded value.
+ if (!AddRec)
+ return FoldedRec;
break;
}
}
llvm_unreachable("Unknown SCEV type!");
- return 0;
}
/// getSCEVAtScope - This is a convenience function which does
// Compute the two solutions for the quadratic formula.
// The divisions must be performed as signed divisions.
APInt NegB(-B);
- APInt TwoA( A << 1 );
+ APInt TwoA(A << 1);
if (TwoA.isMinValue()) {
const SCEV *CNC = SE.getCouldNotCompute();
return std::make_pair(CNC, CNC);
return std::make_pair(SE.getConstant(Solution1),
SE.getConstant(Solution2));
- } // end APIntOps namespace
+ } // end APIntOps namespace
}
/// HowFarToZero - Return the number of times a backedge comparing the specified
/// value to zero will execute. If not computable, return CouldNotCompute.
-ScalarEvolution::BackedgeTakenInfo
+///
+/// This is only used for loops with a "x != y" exit test. The exit condition is
+/// now expressed as a single expression, V = x-y. So the exit test is
+/// effectively V != 0. We know and take advantage of the fact that this
+/// expression only being used in a comparison by zero context.
+ScalarEvolution::ExitLimit
ScalarEvolution::HowFarToZero(const SCEV *V, const Loop *L) {
// If the value is a constant
if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
R2->getValue()))) {
if (CB->getZExtValue() == false)
std::swap(R1, R2); // R1 is the minimum root now.
-
+
// We can only use this value if the chrec ends up with an exact zero
// value at this index. When solving for "X*X != 5", for example, we
// should not accept a root of 2.
const SCEV *Start = getSCEVAtScope(AddRec->getStart(), L->getParentLoop());
const SCEV *Step = getSCEVAtScope(AddRec->getOperand(1), L->getParentLoop());
- // If the AddRec is NUW, then (in an unsigned sense) it cannot be counting up
- // to wrap to 0, it must be counting down to equal 0. Also, while counting
- // down, it cannot "miss" 0 (which would cause it to wrap), regardless of what
- // the stride is. As such, NUW addrec's will always become zero in
- // "start / -stride" steps, and we know that the division is exact.
- if (AddRec->hasNoUnsignedWrap())
- // FIXME: We really want an "isexact" bit for udiv.
- return getUDivExpr(Start, getNegativeSCEV(Step));
-
// For now we handle only constant steps.
+ //
+ // TODO: Handle a nonconstant Step given AddRec<NUW>. If the
+ // AddRec is NUW, then (in an unsigned sense) it cannot be counting up to wrap
+ // to 0, it must be counting down to equal 0. Consequently, N = Start / -Step.
+ // We have not yet seen any such cases.
const SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step);
if (StepC == 0)
return getCouldNotCompute();
- // First, handle unitary steps.
- if (StepC->getValue()->equalsInt(1)) // 1*N = -Start (mod 2^BW), so:
- return getNegativeSCEV(Start); // N = -Start (as unsigned)
-
- if (StepC->getValue()->isAllOnesValue()) // -1*N = -Start (mod 2^BW), so:
- return Start; // N = Start (as unsigned)
+ // For positive steps (counting up until unsigned overflow):
+ // N = -Start/Step (as unsigned)
+ // For negative steps (counting down to zero):
+ // N = Start/-Step
+ // First compute the unsigned distance from zero in the direction of Step.
+ bool CountDown = StepC->getValue()->getValue().isNegative();
+ const SCEV *Distance = CountDown ? Start : getNegativeSCEV(Start);
+
+ // Handle unitary steps, which cannot wraparound.
+ // 1*N = -Start; -1*N = Start (mod 2^BW), so:
+ // N = Distance (as unsigned)
+ if (StepC->getValue()->equalsInt(1) || StepC->getValue()->isAllOnesValue()) {
+ ConstantRange CR = getUnsignedRange(Start);
+ const SCEV *MaxBECount;
+ if (!CountDown && CR.getUnsignedMin().isMinValue())
+ // When counting up, the worst starting value is 1, not 0.
+ MaxBECount = CR.getUnsignedMax().isMinValue()
+ ? getConstant(APInt::getMinValue(CR.getBitWidth()))
+ : getConstant(APInt::getMaxValue(CR.getBitWidth()));
+ else
+ MaxBECount = getConstant(CountDown ? CR.getUnsignedMax()
+ : -CR.getUnsignedMin());
+ return ExitLimit(Distance, MaxBECount);
+ }
+ // If the recurrence is known not to wraparound, unsigned divide computes the
+ // back edge count. We know that the value will either become zero (and thus
+ // the loop terminates), that the loop will terminate through some other exit
+ // condition first, or that the loop has undefined behavior. This means
+ // we can't "miss" the exit value, even with nonunit stride.
+ //
+ // FIXME: Prove that loops always exhibits *acceptable* undefined
+ // behavior. Loops must exhibit defined behavior until a wrapped value is
+ // actually used. So the trip count computed by udiv could be smaller than the
+ // number of well-defined iterations.
+ if (AddRec->getNoWrapFlags(SCEV::FlagNW)) {
+ // FIXME: We really want an "isexact" bit for udiv.
+ return getUDivExpr(Distance, CountDown ? getNegativeSCEV(Step) : Step);
+ }
// Then, try to solve the above equation provided that Start is constant.
if (const SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start))
return SolveLinEquationWithOverflow(StepC->getValue()->getValue(),
/// HowFarToNonZero - Return the number of times a backedge checking the
/// specified value for nonzero will execute. If not computable, return
/// CouldNotCompute
-ScalarEvolution::BackedgeTakenInfo
+ScalarEvolution::ExitLimit
ScalarEvolution::HowFarToNonZero(const SCEV *V, const Loop *L) {
// Loops that look like: while (X == 0) are very strange indeed. We don't
// handle them yet except for the trivial case. This could be expanded in the
case ICmpInst::ICMP_SLE:
if (!getSignedRange(RHS).getSignedMax().isMaxSignedValue()) {
RHS = getAddExpr(getConstant(RHS->getType(), 1, true), RHS,
- /*HasNUW=*/false, /*HasNSW=*/true);
+ SCEV::FlagNSW);
Pred = ICmpInst::ICMP_SLT;
Changed = true;
} else if (!getSignedRange(LHS).getSignedMin().isMinSignedValue()) {
LHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), LHS,
- /*HasNUW=*/false, /*HasNSW=*/true);
+ SCEV::FlagNSW);
Pred = ICmpInst::ICMP_SLT;
Changed = true;
}
case ICmpInst::ICMP_SGE:
if (!getSignedRange(RHS).getSignedMin().isMinSignedValue()) {
RHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), RHS,
- /*HasNUW=*/false, /*HasNSW=*/true);
+ SCEV::FlagNSW);
Pred = ICmpInst::ICMP_SGT;
Changed = true;
} else if (!getSignedRange(LHS).getSignedMax().isMaxSignedValue()) {
LHS = getAddExpr(getConstant(RHS->getType(), 1, true), LHS,
- /*HasNUW=*/false, /*HasNSW=*/true);
+ SCEV::FlagNSW);
Pred = ICmpInst::ICMP_SGT;
Changed = true;
}
case ICmpInst::ICMP_ULE:
if (!getUnsignedRange(RHS).getUnsignedMax().isMaxValue()) {
RHS = getAddExpr(getConstant(RHS->getType(), 1, true), RHS,
- /*HasNUW=*/true, /*HasNSW=*/false);
+ SCEV::FlagNUW);
Pred = ICmpInst::ICMP_ULT;
Changed = true;
} else if (!getUnsignedRange(LHS).getUnsignedMin().isMinValue()) {
LHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), LHS,
- /*HasNUW=*/true, /*HasNSW=*/false);
+ SCEV::FlagNUW);
Pred = ICmpInst::ICMP_ULT;
Changed = true;
}
case ICmpInst::ICMP_UGE:
if (!getUnsignedRange(RHS).getUnsignedMin().isMinValue()) {
RHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), RHS,
- /*HasNUW=*/true, /*HasNSW=*/false);
+ SCEV::FlagNUW);
Pred = ICmpInst::ICMP_UGT;
Changed = true;
} else if (!getUnsignedRange(LHS).getUnsignedMax().isMaxValue()) {
LHS = getAddExpr(getConstant(RHS->getType(), 1, true), LHS,
- /*HasNUW=*/true, /*HasNSW=*/false);
+ SCEV::FlagNUW);
Pred = ICmpInst::ICMP_UGT;
Changed = true;
}
switch (Pred) {
default:
llvm_unreachable("Unexpected ICmpInst::Predicate value!");
- break;
case ICmpInst::ICMP_SGT:
Pred = ICmpInst::ICMP_SLT;
std::swap(LHS, RHS);
assert(!isKnownNegative(Step) &&
"This code doesn't handle negative strides yet!");
- const Type *Ty = Start->getType();
+ Type *Ty = Start->getType();
+
+ // When Start == End, we have an exact BECount == 0. Short-circuit this case
+ // here because SCEV may not be able to determine that the unsigned division
+ // after rounding is zero.
+ if (Start == End)
+ return getConstant(Ty, 0);
+
const SCEV *NegOne = getConstant(Ty, (uint64_t)-1);
const SCEV *Diff = getMinusSCEV(End, Start);
const SCEV *RoundUp = getAddExpr(Step, NegOne);
if (!NoWrap) {
// Check Add for unsigned overflow.
// TODO: More sophisticated things could be done here.
- const Type *WideTy = IntegerType::get(getContext(),
+ Type *WideTy = IntegerType::get(getContext(),
getTypeSizeInBits(Ty) + 1);
const SCEV *EDiff = getZeroExtendExpr(Diff, WideTy);
const SCEV *ERoundUp = getZeroExtendExpr(RoundUp, WideTy);
/// HowManyLessThans - Return the number of times a backedge containing the
/// specified less-than comparison will execute. If not computable, return
/// CouldNotCompute.
-ScalarEvolution::BackedgeTakenInfo
+ScalarEvolution::ExitLimit
ScalarEvolution::HowManyLessThans(const SCEV *LHS, const SCEV *RHS,
const Loop *L, bool isSigned) {
// Only handle: "ADDREC < LoopInvariant".
return getCouldNotCompute();
// Check to see if we have a flag which makes analysis easy.
- bool NoWrap = isSigned ? AddRec->hasNoSignedWrap() :
- AddRec->hasNoUnsignedWrap();
+ bool NoWrap = isSigned ?
+ AddRec->getNoWrapFlags((SCEV::NoWrapFlags)(SCEV::FlagNSW | SCEV::FlagNW)) :
+ AddRec->getNoWrapFlags((SCEV::NoWrapFlags)(SCEV::FlagNUW | SCEV::FlagNW));
if (AddRec->isAffine()) {
unsigned BitWidth = getTypeSizeInBits(AddRec->getType());
// The maximum backedge count is similar, except using the minimum start
// value and the maximum end value.
- const SCEV *MaxBECount = getBECount(MinStart, MaxEnd, Step, NoWrap);
+ // If we already have an exact constant BECount, use it instead.
+ const SCEV *MaxBECount = isa<SCEVConstant>(BECount) ? BECount
+ : getBECount(MinStart, MaxEnd, Step, NoWrap);
- return BackedgeTakenInfo(BECount, MaxBECount);
+ // If the stride is nonconstant, and NoWrap == true, then
+ // getBECount(MinStart, MaxEnd) may not compute. This would result in an
+ // exact BECount and invalid MaxBECount, which should be avoided to catch
+ // more optimization opportunities.
+ if (isa<SCEVCouldNotCompute>(MaxBECount))
+ MaxBECount = BECount;
+
+ return ExitLimit(BECount, MaxBECount);
}
return getCouldNotCompute();
if (!SC->getValue()->isZero()) {
SmallVector<const SCEV *, 4> Operands(op_begin(), op_end());
Operands[0] = SE.getConstant(SC->getType(), 0);
- const SCEV *Shifted = SE.getAddRecExpr(Operands, getLoop());
+ const SCEV *Shifted = SE.getAddRecExpr(Operands, getLoop(),
+ getNoWrapFlags(FlagNW));
if (const SCEVAddRecExpr *ShiftedAddRec =
dyn_cast<SCEVAddRecExpr>(Shifted))
return ShiftedAddRec->getNumIterationsInRange(
// Range.getUpper() is crossed.
SmallVector<const SCEV *, 4> NewOps(op_begin(), op_end());
NewOps[0] = SE.getNegativeSCEV(SE.getConstant(Range.getUpper()));
- const SCEV *NewAddRec = SE.getAddRecExpr(NewOps, getLoop());
+ const SCEV *NewAddRec = SE.getAddRecExpr(NewOps, getLoop(),
+ // getNoWrapFlags(FlagNW)
+ FlagAnyWrap);
// Next, solve the constructed addrec
std::pair<const SCEV *,const SCEV *> Roots =
this->F = &F;
LI = &getAnalysis<LoopInfo>();
TD = getAnalysisIfAvailable<TargetData>();
+ TLI = &getAnalysis<TargetLibraryInfo>();
DT = &getAnalysis<DominatorTree>();
return false;
}
FirstUnknown = 0;
ValueExprMap.clear();
+
+ // Free any extra memory created for ExitNotTakenInfo in the unlikely event
+ // that a loop had multiple computable exits.
+ for (DenseMap<const Loop*, BackedgeTakenInfo>::iterator I =
+ BackedgeTakenCounts.begin(), E = BackedgeTakenCounts.end();
+ I != E; ++I) {
+ I->second.clear();
+ }
+
BackedgeTakenCounts.clear();
ConstantEvolutionLoopExitValue.clear();
ValuesAtScopes.clear();
AU.setPreservesAll();
AU.addRequiredTransitive<LoopInfo>();
AU.addRequiredTransitive<DominatorTree>();
+ AU.addRequired<TargetLibraryInfo>();
}
bool ScalarEvolution::hasLoopInvariantBackedgeTakenCount(const Loop *L) {
return LoopInvariant;
case scCouldNotCompute:
llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
- return LoopVariant;
- default: break;
+ default: llvm_unreachable("Unknown SCEV kind!");
}
- llvm_unreachable("Unknown SCEV kind!");
- return LoopVariant;
}
bool ScalarEvolution::isLoopInvariant(const SCEV *S, const Loop *L) {
return ProperlyDominatesBlock;
case scCouldNotCompute:
llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
- return DoesNotDominateBlock;
- default: break;
+ default:
+ llvm_unreachable("Unknown SCEV kind!");
}
- llvm_unreachable("Unknown SCEV kind!");
- return DoesNotDominateBlock;
}
bool ScalarEvolution::dominates(const SCEV *S, const BasicBlock *BB) {
return false;
case scCouldNotCompute:
llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
- return false;
- default: break;
+ default:
+ llvm_unreachable("Unknown SCEV kind!");
}
- llvm_unreachable("Unknown SCEV kind!");
- return false;
}
void ScalarEvolution::forgetMemoizedResults(const SCEV *S) {