#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/ValueTracking.h"
#include "llvm/Assembly/Writer.h"
-#include "llvm/Target/TargetData.h"
+#include "llvm/DataLayout.h"
+#include "llvm/Target/TargetLibraryInfo.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/ConstantRange.h"
#include "llvm/Support/Debug.h"
"derived loop"),
cl::init(100));
+// FIXME: Enable this with XDEBUG when the test suite is clean.
+static cl::opt<bool>
+VerifySCEV("verify-scev",
+ cl::desc("Verify ScalarEvolution's backedge taken counts (slow)"));
+
INITIALIZE_PASS_BEGIN(ScalarEvolution, "scalar-evolution",
"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;
// Implementation of the SCEV class.
//
+#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
void SCEV::dump() const {
print(dbgs());
dbgs() << '\n';
}
+#endif
void SCEV::print(raw_ostream &OS) const {
switch (getSCEVType()) {
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: {
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) {}
}
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,
- Type* ResultTy) {
+ Type *ResultTy) {
// Handle the simplest case efficiently.
if (K == 1)
return SE.getTruncateOrZeroExtend(It, ResultTy);
// Fold if the operand is constant.
if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
return getConstant(
- cast<ConstantInt>(ConstantExpr::getTrunc(SC->getValue(),
- getEffectiveSCEVType(Ty))));
+ cast<ConstantInt>(ConstantExpr::getTrunc(SC->getValue(), Ty)));
// trunc(trunc(x)) --> trunc(x)
if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op))
return getAddRecExpr(Operands, AddRec->getLoop(), SCEV::FlagAnyWrap);
}
- // As a special case, fold trunc(undef) to undef. We don't want to
- // know too much about SCEVUnknowns, but this special case is handy
- // and harmless.
- if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Op))
- if (isa<UndefValue>(U->getValue()))
- return getSCEV(UndefValue::get(Ty));
-
// The cast wasn't folded; create an explicit cast node. We can reuse
// the existing insert position since if we get here, we won't have
// made any changes which would invalidate it.
// Fold if the operand is constant.
if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
return getConstant(
- cast<ConstantInt>(ConstantExpr::getZExt(SC->getValue(),
- getEffectiveSCEVType(Ty))));
+ cast<ConstantInt>(ConstantExpr::getZExt(SC->getValue(), Ty)));
// zext(zext(x)) --> zext(x)
if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
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);
+ const SCEV *ZAdd = getZeroExtendExpr(getAddExpr(Start, ZMul), WideTy);
+ const SCEV *WideStart = getZeroExtendExpr(Start, WideTy);
+ const SCEV *WideMaxBECount =
+ getZeroExtendExpr(CastedMaxBECount, WideTy);
const SCEV *OperandExtendedAdd =
- getAddExpr(getZeroExtendExpr(Start, WideTy),
- getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
+ getAddExpr(WideStart,
+ getMulExpr(WideMaxBECount,
getZeroExtendExpr(Step, WideTy)));
- if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd) {
+ if (ZAdd == 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.
}
// Similar to above, only this time treat the step value as signed.
// This covers loops that count down.
- const SCEV *SMul = getMulExpr(CastedMaxBECount, Step);
- Add = getAddExpr(Start, SMul);
OperandExtendedAdd =
- getAddExpr(getZeroExtendExpr(Start, WideTy),
- getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
+ getAddExpr(WideStart,
+ getMulExpr(WideMaxBECount,
getSignExtendExpr(Step, WideTy)));
- if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd) {
+ if (ZAdd == 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);
// Check for a simple looking step prior to loop entry.
const SCEVAddExpr *SA = dyn_cast<SCEVAddExpr>(Start);
- if (!SA || SA->getNumOperands() != 2 || SA->getOperand(0) != Step)
+ 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 = SA->getOperand(1);
+ const SCEV *PreStart = SE->getAddExpr(DiffOps, SA->getNoWrapFlags());
const SCEVAddRecExpr *PreAR = dyn_cast<SCEVAddRecExpr>(
SE->getAddRecExpr(PreStart, Step, L, SCEV::FlagAnyWrap));
// Fold if the operand is constant.
if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
return getConstant(
- cast<ConstantInt>(ConstantExpr::getSExt(SC->getValue(),
- getEffectiveSCEVType(Ty))));
+ cast<ConstantInt>(ConstantExpr::getSExt(SC->getValue(), Ty)));
// sext(sext(x)) --> sext(x)
if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
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);
+ const SCEV *SAdd = getSignExtendExpr(getAddExpr(Start, SMul), WideTy);
+ const SCEV *WideStart = getSignExtendExpr(Start, WideTy);
+ const SCEV *WideMaxBECount =
+ getZeroExtendExpr(CastedMaxBECount, WideTy);
const SCEV *OperandExtendedAdd =
- getAddExpr(getSignExtendExpr(Start, WideTy),
- getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
+ getAddExpr(WideStart,
+ getMulExpr(WideMaxBECount,
getSignExtendExpr(Step, WideTy)));
- if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd) {
+ if (SAdd == 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.
}
// 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);
- Add = getAddExpr(Start, UMul);
OperandExtendedAdd =
- getAddExpr(getSignExtendExpr(Start, WideTy),
- getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy),
+ getAddExpr(WideStart,
+ getMulExpr(WideMaxBECount,
getZeroExtendExpr(Step, WideTy)));
- if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd) {
+ if (SAdd == 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(Ops, AR->getLoop(), SCEV::FlagNW);
}
- // As a special case, fold anyext(undef) to undef. We don't want to
- // know too much about SCEVUnknowns, but this special case is handy
- // and harmless.
- if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Op))
- if (isa<UndefValue>(U->getValue()))
- return getSCEV(UndefValue::get(Ty));
-
// If the expression is obviously signed, use the sext cast value.
if (isa<SCEVSMaxExpr>(Op))
return SExt;
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 absence 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,
// multiplied together. If so, we can fold them.
for (unsigned OtherIdx = Idx+1;
OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
- ++OtherIdx)
- if (AddRecLoop == cast<SCEVAddRecExpr>(Ops[OtherIdx])->getLoop()) {
- // {A,+,B}<L> * {C,+,D}<L> --> {A*C,+,A*D + B*C + B*D,+,2*B*D}<L>
- for (; OtherIdx != Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
- ++OtherIdx)
- if (const SCEVAddRecExpr *OtherAddRec =
- dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx]))
- if (OtherAddRec->getLoop() == AddRecLoop) {
- const SCEV *A = AddRec->getStart();
- const SCEV *B = AddRec->getStepRecurrence(*this);
- const SCEV *C = OtherAddRec->getStart();
- const SCEV *D = OtherAddRec->getStepRecurrence(*this);
- const SCEV *NewStart = getMulExpr(A, C);
- const SCEV *BD = getMulExpr(B, D);
- const SCEV *NewStep = getAddExpr(getMulExpr(A, D),
- getMulExpr(B, C), BD);
- const SCEV *NewSecondOrderStep =
- getMulExpr(BD, getConstant(BD->getType(), 2));
-
- SmallVector<const SCEV *, 3> AddRecOps;
- AddRecOps.push_back(NewStart);
- AddRecOps.push_back(NewStep);
- AddRecOps.push_back(NewSecondOrderStep);
- 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;
+ ++OtherIdx) {
+ if (AddRecLoop != cast<SCEVAddRecExpr>(Ops[OtherIdx])->getLoop())
+ continue;
+
+ // {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) {
+ const SCEVAddRecExpr *OtherAddRec =
+ dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx]);
+ if (!OtherAddRec || OtherAddRec->getLoop() != AddRecLoop)
+ continue;
+
+ 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));
}
- return getMulExpr(Ops);
+ }
+ AddRecOps.push_back(Term);
+ }
+ if (!Overflow) {
+ const SCEV *NewAddRec = getAddRecExpr(AddRecOps, AddRec->getLoop(),
+ SCEV::FlagAnyWrap);
+ if (Ops.size() == 2) return NewAddRec;
+ Ops[Idx] = NewAddRec;
+ Ops.erase(Ops.begin() + OtherIdx); --OtherIdx;
+ OpsModified = true;
+ AddRec = dyn_cast<SCEVAddRecExpr>(NewAddRec);
+ if (!AddRec)
+ break;
+ }
}
+ if (OpsModified)
+ return getMulExpr(Ops);
+ }
// Otherwise couldn't fold anything into this recurrence. Move onto the
// next one.
return getNotSCEV(getUMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS)));
}
-const SCEV *ScalarEvolution::getSizeOfExpr(Type *AllocTy) {
- // If we have TargetData, we can bypass creating a target-independent
+const SCEV *ScalarEvolution::getSizeOfExpr(Type *AllocTy, Type *IntPtrTy) {
+ // If we have DataLayout, we can bypass creating a target-independent
// constant expression and then folding it back into a ConstantInt.
// This is just a compile-time optimization.
if (TD)
- return getConstant(TD->getIntPtrType(getContext()),
- TD->getTypeAllocSize(AllocTy));
+ return getConstant(IntPtrTy, TD->getTypeAllocSize(AllocTy));
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;
Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
return getTruncateOrZeroExtend(getSCEV(C), Ty);
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;
Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
return getTruncateOrZeroExtend(getSCEV(C), Ty);
}
-const SCEV *ScalarEvolution::getOffsetOfExpr(StructType *STy,
+const SCEV *ScalarEvolution::getOffsetOfExpr(StructType *STy, Type *IntPtrTy,
unsigned FieldNo) {
- // If we have TargetData, we can bypass creating a target-independent
+ // If we have DataLayout, we can bypass creating a target-independent
// constant expression and then folding it back into a ConstantInt.
// This is just a compile-time optimization.
if (TD)
- return getConstant(TD->getIntPtrType(getContext()),
+ return getConstant(IntPtrTy,
TD->getStructLayout(STy)->getElementOffset(FieldNo));
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;
Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(STy));
return getTruncateOrZeroExtend(getSCEV(C), Ty);
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;
Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(CTy));
return getTruncateOrZeroExtend(getSCEV(C), Ty);
uint64_t ScalarEvolution::getTypeSizeInBits(Type *Ty) const {
assert(isSCEVable(Ty) && "Type is not SCEVable!");
- // If we have a TargetData, use it!
+ // If we have a DataLayout, use it!
if (TD)
return TD->getTypeSizeInBits(Ty);
if (Ty->isIntegerTy())
return Ty->getPrimitiveSizeInBits();
- // The only other support type is pointer. Without TargetData, conservatively
+ // The only other support type is pointer. Without DataLayout, conservatively
// assume pointers are 64-bit.
assert(Ty->isPointerTy() && "isSCEVable permitted a non-SCEVable type!");
return 64;
// The only other support type is pointer.
assert(Ty->isPointerTy() && "Unexpected non-pointer non-integer type!");
- if (TD) return TD->getIntPtrType(getContext());
+ if (TD) return TD->getIntPtrType(Ty);
- // Without TargetData, conservatively assume pointers are 64-bit.
+ // Without DataLayout, conservatively assume pointers are 64-bit.
return Type::getInt64Ty(getContext());
}
const SCEV *ScalarEvolution::getSCEV(Value *V) {
assert(isSCEVable(V->getType()) && "Value is not SCEVable!");
- ValueExprMapType::const_iterator I = ValueExprMap.find(V);
+ ValueExprMapType::const_iterator I = ValueExprMap.find_as(V);
if (I != ValueExprMap.end()) return I->second;
const SCEV *S = createSCEV(V);
if (!Visited.insert(I)) continue;
ValueExprMapType::iterator It =
- ValueExprMap.find(static_cast<Value *>(I));
+ ValueExprMap.find_as(static_cast<Value *>(I));
if (It != ValueExprMap.end()) {
const SCEV *Old = It->second;
if (BEValueV && StartValueV) {
// While we are analyzing this PHI node, handle its value symbolically.
const SCEV *SymbolicName = getUnknown(PN);
- assert(ValueExprMap.find(PN) == ValueExprMap.end() &&
+ assert(ValueExprMap.find_as(PN) == ValueExprMap.end() &&
"PHI node already processed?");
ValueExprMap.insert(std::make_pair(SCEVCallbackVH(PN, this), SymbolicName));
// 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);
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);
+ const SCEV *FieldOffset = getOffsetOfExpr(STy, IntPtrTy, FieldNo);
// Add the field offset to the running total offset.
TotalOffset = getAddExpr(TotalOffset, FieldOffset);
} else {
// For an array, add the element offset, explicitly scaled.
- const SCEV *ElementSize = getSizeOfExpr(*GTI);
+ const SCEV *ElementSize = getSizeOfExpr(*GTI, IntPtrTy);
const SCEV *IndexS = getSCEV(Index);
// Getelementptr indices are signed.
IndexS = getTruncateOrSignExtend(IndexS, IntPtrTy);
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 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,
// 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);
// 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);
//
/// getSmallConstantTripCount - Returns the maximum trip count of this loop as a
-/// normal unsigned value, if possible. 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)
-unsigned ScalarEvolution::getSmallConstantTripCount(Loop *L,
- BasicBlock *ExitBlock) {
+/// 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, ExitBlock));
+ dyn_cast<SCEVConstant>(getExitCount(L, ExitingBlock));
if (!ExitCount)
return 0;
/// 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).
-unsigned ScalarEvolution::getSmallConstantTripMultiple(Loop *L,
- BasicBlock *ExitBlock) {
- const SCEV *ExitCount = getExitCount(L, ExitBlock);
+///
+/// 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;
ConstantInt *Result = MulC->getValue();
- // Guard against huge trip counts.
- if (!Result || Result->getValue().getActiveBits() > 32)
+ // Guard against huge trip counts (this requires checking
+ // for zero to handle the case where the trip count == -1 and the
+ // addition wraps).
+ if (!Result || Result->getValue().getActiveBits() > 32 ||
+ Result->getValue().getActiveBits() == 0)
return 1;
return (unsigned)Result->getZExtValue();
if (!Visited.insert(I)) continue;
ValueExprMapType::iterator It =
- ValueExprMap.find(static_cast<Value *>(I));
+ ValueExprMap.find_as(static_cast<Value *>(I));
if (It != ValueExprMap.end()) {
const SCEV *Old = It->second;
Instruction *I = Worklist.pop_back_val();
if (!Visited.insert(I)) continue;
- ValueExprMapType::iterator It = ValueExprMap.find(static_cast<Value *>(I));
+ ValueExprMapType::iterator It =
+ ValueExprMap.find_as(static_cast<Value *>(I));
if (It != ValueExprMap.end()) {
forgetMemoizedResults(It->second);
ValueExprMap.erase(It);
I = Worklist.pop_back_val();
if (!Visited.insert(I)) continue;
- ValueExprMapType::iterator It = ValueExprMap.find(static_cast<Value *>(I));
+ ValueExprMapType::iterator It =
+ ValueExprMap.find_as(static_cast<Value *>(I));
if (It != ValueExprMap.end()) {
forgetMemoizedResults(It->second);
ValueExprMap.erase(It);
}
/// getExact - Get the exact loop backedge taken count considering all loop
-/// exits. If all exits are computable, this is the minimum computed count.
+/// 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 at least one computable exit.
+ // We need exactly one computable exit.
if (!ExitNotTaken.ExitingBlock) return SE->getCouldNotCompute();
assert(ExitNotTaken.ExactNotTaken && "uninitialized not-taken info");
if (!BECount)
BECount = ENT->ExactNotTaken;
- else
- BECount = SE->getUMinFromMismatchedTypes(BECount, ENT->ExactNotTaken);
+ else if (BECount != ENT->ExactNotTaken)
+ return SE->getCouldNotCompute();
}
assert(BECount && "Invalid not taken count for loop exit");
return BECount;
if (MaxBECount == getCouldNotCompute())
MaxBECount = EL.Max;
- else if (EL.Max != getCouldNotCompute())
- MaxBECount = getUMinFromMismatchedTypes(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(ExitCounts, CouldComputeBECount, MaxBECount);
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 (StructType *STy = dyn_cast<StructType>(Init->getType())) {
- assert(Idx < STy->getNumElements() && "Bad struct index!");
- Init = Constant::getNullValue(STy->getElementType(Idx));
- } else if (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;
-}
-
/// ComputeLoadConstantCompareExitLimit - Given an exit condition of
/// 'icmp op load X, cst', try to see if we can compute the backedge
/// execution count.
// 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 (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 DataLayout *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, 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
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);
}
}
/// 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::ComputeExitCountExhaustively(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, TD);
+ Operands, TD, TLI);
if (!C) return V;
return getSCEV(C);
}
}
llvm_unreachable("Unknown SCEV type!");
- return 0;
}
/// getSCEVAtScope - This is a convenience function which does
SqrtTerm *= B;
SqrtTerm -= Four * (A * C);
+ if (SqrtTerm.isNegative()) {
+ // The loop is provably infinite.
+ const SCEV *CNC = SE.getCouldNotCompute();
+ return std::make_pair(CNC, CNC);
+ }
+
// Compute sqrt(B^2-4ac). This is guaranteed to be the nearest
// integer value or else APInt::sqrt() will assert.
APInt SqrtVal(SqrtTerm.sqrt());
// 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
// 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)
+ if (StepC == 0 || StepC->getValue()->equalsInt(0))
return getCouldNotCompute();
// For positive steps (counting up until unsigned overflow):
// 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())
- return Distance;
+ 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
// 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))
+ 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(),
/// predicate Pred. Return true iff any changes were made.
///
bool ScalarEvolution::SimplifyICmpOperands(ICmpInst::Predicate &Pred,
- const SCEV *&LHS, const SCEV *&RHS) {
+ const SCEV *&LHS, const SCEV *&RHS,
+ unsigned Depth) {
bool Changed = false;
+ // If we hit the max recursion limit bail out.
+ if (Depth >= 3)
+ return false;
+
// Canonicalize a constant to the right side.
if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
// Check for both operands constant.
default: llvm_unreachable("Unexpected ICmpInst::Predicate value!");
case ICmpInst::ICMP_EQ:
case ICmpInst::ICMP_NE:
+ // Fold ((-1) * %a) + %b == 0 (equivalent to %b-%a == 0) into %a == %b.
+ if (!RA)
+ if (const SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(LHS))
+ if (const SCEVMulExpr *ME = dyn_cast<SCEVMulExpr>(AE->getOperand(0)))
+ if (AE->getNumOperands() == 2 && ME->getNumOperands() == 2 &&
+ ME->getOperand(0)->isAllOnesValue()) {
+ RHS = AE->getOperand(1);
+ LHS = ME->getOperand(1);
+ Changed = true;
+ }
break;
case ICmpInst::ICMP_UGE:
if ((RA - 1).isMinValue()) {
// TODO: More simplifications are possible here.
+ // Recursively simplify until we either hit a recursion limit or nothing
+ // changes.
+ if (Changed)
+ return SimplifyICmpOperands(Pred, LHS, RHS, Depth+1);
+
return Changed;
trivially_true:
switch (Pred) {
default:
llvm_unreachable("Unexpected ICmpInst::Predicate value!");
- break;
case ICmpInst::ICMP_SGT:
Pred = ICmpInst::ICMP_SLT;
std::swap(LHS, RHS);
return false;
}
+/// RAII wrapper to prevent recursive application of isImpliedCond.
+/// ScalarEvolution's PendingLoopPredicates set must be empty unless we are
+/// currently evaluating isImpliedCond.
+struct MarkPendingLoopPredicate {
+ Value *Cond;
+ DenseSet<Value*> &LoopPreds;
+ bool Pending;
+
+ MarkPendingLoopPredicate(Value *C, DenseSet<Value*> &LP)
+ : Cond(C), LoopPreds(LP) {
+ Pending = !LoopPreds.insert(Cond).second;
+ }
+ ~MarkPendingLoopPredicate() {
+ if (!Pending)
+ LoopPreds.erase(Cond);
+ }
+};
+
/// isImpliedCond - Test whether the condition described by Pred, LHS,
/// and RHS is true whenever the given Cond value evaluates to true.
bool ScalarEvolution::isImpliedCond(ICmpInst::Predicate Pred,
const SCEV *LHS, const SCEV *RHS,
Value *FoundCondValue,
bool Inverse) {
+ MarkPendingLoopPredicate Mark(FoundCondValue, PendingLoopPredicates);
+ if (Mark.Pending)
+ return false;
+
// Recursively handle And and Or conditions.
if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FoundCondValue)) {
if (BO->getOpcode() == Instruction::And) {
return getCouldNotCompute();
// Check to see if we have a flag which makes analysis easy.
- bool NoWrap = isSigned ? AddRec->getNoWrapFlags(SCEV::FlagNSW) :
- AddRec->getNoWrapFlags(SCEV::FlagNUW);
+ 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());
bool ScalarEvolution::runOnFunction(Function &F) {
this->F = &F;
LI = &getAnalysis<LoopInfo>();
- TD = getAnalysisIfAvailable<TargetData>();
+ TD = getAnalysisIfAvailable<DataLayout>();
+ TLI = &getAnalysis<TargetLibraryInfo>();
DT = &getAnalysis<DominatorTree>();
return false;
}
I->second.clear();
}
+ assert(PendingLoopPredicates.empty() && "isImpliedCond garbage");
+
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 getBlockDisposition(S, BB) == ProperlyDominatesBlock;
}
-bool ScalarEvolution::hasOperand(const SCEV *S, const SCEV *Op) const {
- switch (S->getSCEVType()) {
- case scConstant:
- return false;
- case scTruncate:
- case scZeroExtend:
- case scSignExtend: {
- const SCEVCastExpr *Cast = cast<SCEVCastExpr>(S);
- const SCEV *CastOp = Cast->getOperand();
- return Op == CastOp || hasOperand(CastOp, Op);
- }
- case scAddRecExpr:
- case scAddExpr:
- case scMulExpr:
- case scUMaxExpr:
- case scSMaxExpr: {
- const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(S);
- for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
- I != E; ++I) {
- const SCEV *NAryOp = *I;
- if (NAryOp == Op || hasOperand(NAryOp, Op))
- return true;
- }
- return false;
- }
- case scUDivExpr: {
- const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(S);
- const SCEV *LHS = UDiv->getLHS(), *RHS = UDiv->getRHS();
- return LHS == Op || hasOperand(LHS, Op) ||
- RHS == Op || hasOperand(RHS, Op);
- }
- case scUnknown:
- return false;
- case scCouldNotCompute:
- llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!");
- return false;
- default: break;
+namespace {
+// Search for a SCEV expression node within an expression tree.
+// Implements SCEVTraversal::Visitor.
+struct SCEVSearch {
+ const SCEV *Node;
+ bool IsFound;
+
+ SCEVSearch(const SCEV *N): Node(N), IsFound(false) {}
+
+ bool follow(const SCEV *S) {
+ IsFound |= (S == Node);
+ return !IsFound;
}
- llvm_unreachable("Unknown SCEV kind!");
- return false;
+ bool isDone() const { return IsFound; }
+};
+}
+
+bool ScalarEvolution::hasOperand(const SCEV *S, const SCEV *Op) const {
+ SCEVSearch Search(Op);
+ visitAll(S, Search);
+ return Search.IsFound;
}
void ScalarEvolution::forgetMemoizedResults(const SCEV *S) {
UnsignedRanges.erase(S);
SignedRanges.erase(S);
}
+
+typedef DenseMap<const Loop *, std::string> VerifyMap;
+/// getLoopBackedgeTakenCounts - Helper method for verifyAnalysis.
+static void
+getLoopBackedgeTakenCounts(Loop *L, VerifyMap &Map, ScalarEvolution &SE) {
+ for (Loop::reverse_iterator I = L->rbegin(), E = L->rend(); I != E; ++I) {
+ getLoopBackedgeTakenCounts(*I, Map, SE); // recurse.
+
+ std::string &S = Map[L];
+ if (S.empty()) {
+ raw_string_ostream OS(S);
+ SE.getBackedgeTakenCount(L)->print(OS);
+ }
+ }
+}
+
+void ScalarEvolution::verifyAnalysis() const {
+ if (!VerifySCEV)
+ return;
+
+ ScalarEvolution &SE = *const_cast<ScalarEvolution *>(this);
+
+ // Gather stringified backedge taken counts for all loops using SCEV's caches.
+ // FIXME: It would be much better to store actual values instead of strings,
+ // but SCEV pointers will change if we drop the caches.
+ VerifyMap BackedgeDumpsOld, BackedgeDumpsNew;
+ for (LoopInfo::reverse_iterator I = LI->rbegin(), E = LI->rend(); I != E; ++I)
+ getLoopBackedgeTakenCounts(*I, BackedgeDumpsOld, SE);
+
+ // Gather stringified backedge taken counts for all loops without using
+ // SCEV's caches.
+ SE.releaseMemory();
+ for (LoopInfo::reverse_iterator I = LI->rbegin(), E = LI->rend(); I != E; ++I)
+ getLoopBackedgeTakenCounts(*I, BackedgeDumpsNew, SE);
+
+ // Now compare whether they're the same with and without caches. This allows
+ // verifying that no pass changed the cache.
+ assert(BackedgeDumpsOld.size() == BackedgeDumpsNew.size() &&
+ "New loops suddenly appeared!");
+
+ for (VerifyMap::iterator OldI = BackedgeDumpsOld.begin(),
+ OldE = BackedgeDumpsOld.end(),
+ NewI = BackedgeDumpsNew.begin();
+ OldI != OldE; ++OldI, ++NewI) {
+ assert(OldI->first == NewI->first && "Loop order changed!");
+
+ // Compare the stringified SCEVs. We don't care if undef backedgetaken count
+ // changes.
+ // FIXME: We currently ignore SCEV changes towards CouldNotCompute. This
+ // means that a pass is buggy or SCEV has to learn a new pattern but is
+ // usually not harmful.
+ if (OldI->second != NewI->second &&
+ OldI->second.find("undef") == std::string::npos &&
+ NewI->second != "***COULDNOTCOMPUTE***") {
+ dbgs() << "SCEVValidator: SCEV for Loop '"
+ << OldI->first->getHeader()->getName()
+ << "' from '" << OldI->second << "' to '" << NewI->second << "'!";
+ std::abort();
+ }
+ }
+
+ // TODO: Verify more things.
+}
projects / oota-llvm.git / blobdiff