My super-optimizer noticed that we weren't folding this expression to
true: (x *nsw x) sgt 0, where x = (y | 1). This occurs in 464.h264ref.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@143028 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Analysis/ValueTracking.cpp b/lib/Analysis/ValueTracking.cpp
index e80ee65..f2740a6 100644
--- a/lib/Analysis/ValueTracking.cpp
+++ b/lib/Analysis/ValueTracking.cpp
@@ -201,9 +201,36 @@
ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero, KnownOne, TD,Depth+1);
ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, TD,
Depth+1);
- assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
- assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
-
+ assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
+ assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
+
+ bool isKnownNegative = false;
+ bool isKnownNonNegative = false;
+ // If the multiplication is known not to overflow, compute the sign bit.
+ if (Mask.isNegative() && cast<BinaryOperator>(I)->hasNoSignedWrap()) {
+ Value *Op1 = I->getOperand(1), *Op2 = I->getOperand(0);
+ if (Op1 == Op2) {
+ // The product of a number with itself is non-negative.
+ isKnownNonNegative = true;
+ } else {
+ bool isKnownNonNegative1 = KnownZero.isNegative();
+ bool isKnownNonNegative2 = KnownZero2.isNegative();
+ bool isKnownNegative1 = KnownOne.isNegative();
+ bool isKnownNegative2 = KnownOne2.isNegative();
+ // The product of two numbers with the same sign is non-negative.
+ isKnownNonNegative = (isKnownNegative1 && isKnownNegative2) ||
+ (isKnownNonNegative1 && isKnownNonNegative2);
+ // The product of a negative number and a non-negative number is either
+ // negative or zero.
+ isKnownNegative = (isKnownNegative1 && isKnownNonNegative2 &&
+ isKnownNonZero(Op2, TD, Depth)) ||
+ (isKnownNegative2 && isKnownNonNegative1 &&
+ isKnownNonZero(Op1, TD, Depth));
+ assert(!(isKnownNegative && isKnownNonNegative) &&
+ "Sign bit both zero and one?");
+ }
+ }
+
// If low bits are zero in either operand, output low known-0 bits.
// Also compute a conserative estimate for high known-0 bits.
// More trickiness is possible, but this is sufficient for the
@@ -220,6 +247,12 @@
KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
APInt::getHighBitsSet(BitWidth, LeadZ);
KnownZero &= Mask;
+
+ if (isKnownNonNegative)
+ KnownZero.setBit(BitWidth - 1);
+ else if (isKnownNegative)
+ KnownOne.setBit(BitWidth - 1);
+
return;
}
case Instruction::UDiv: {
@@ -767,7 +800,7 @@
}
// The remaining tests are all recursive, so bail out if we hit the limit.
- if (Depth++ == MaxDepth)
+ if (Depth++ >= MaxDepth)
return false;
unsigned BitWidth = getBitWidth(V->getType(), TD);
@@ -851,6 +884,15 @@
if (YKnownNonNegative && isPowerOfTwo(X, TD, Depth))
return true;
}
+ // X * Y.
+ else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
+ BinaryOperator *BO = cast<BinaryOperator>(V);
+ // If X and Y are non-zero then so is X * Y as long as the multiplication
+ // does not overflow.
+ if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
+ isKnownNonZero(X, TD, Depth) && isKnownNonZero(Y, TD, Depth))
+ return true;
+ }
// (C ? X : Y) != 0 if X != 0 and Y != 0.
else if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
if (isKnownNonZero(SI->getTrueValue(), TD, Depth) &&