mirror of
https://github.com/c64scene-ar/llvm-6502.git
synced 2024-08-22 10:29:35 +00:00
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
This commit is contained in:
parent
a921a46854
commit
e8ec225e77
@ -201,9 +201,36 @@ void llvm::ComputeMaskedBits(Value *V, const APInt &Mask,
|
||||
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 @@ void llvm::ComputeMaskedBits(Value *V, const APInt &Mask,
|
||||
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 @@ bool llvm::isKnownNonZero(Value *V, const TargetData *TD, unsigned Depth) {
|
||||
}
|
||||
|
||||
// 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 @@ bool llvm::isKnownNonZero(Value *V, const TargetData *TD, unsigned Depth) {
|
||||
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) &&
|
||||
|
@ -323,3 +323,34 @@ define i1 @and1(i32 %X) {
|
||||
ret i1 %B
|
||||
; CHECK: ret i1 false
|
||||
}
|
||||
|
||||
define i1 @mul1(i32 %X) {
|
||||
; CHECK: @mul1
|
||||
; Square of a non-zero number is non-zero if there is no overflow.
|
||||
%Y = or i32 %X, 1
|
||||
%M = mul nuw i32 %Y, %Y
|
||||
%C = icmp eq i32 %M, 0
|
||||
ret i1 %C
|
||||
; CHECK: ret i1 false
|
||||
}
|
||||
|
||||
define i1 @mul2(i32 %X) {
|
||||
; CHECK: @mul2
|
||||
; Square of a non-zero number is positive if there is no signed overflow.
|
||||
%Y = or i32 %X, 1
|
||||
%M = mul nsw i32 %Y, %Y
|
||||
%C = icmp sgt i32 %M, 0
|
||||
ret i1 %C
|
||||
; CHECK: ret i1 true
|
||||
}
|
||||
|
||||
define i1 @mul3(i32 %X, i32 %Y) {
|
||||
; CHECK: @mul3
|
||||
; Product of non-negative numbers is non-negative if there is no signed overflow.
|
||||
%XX = mul nsw i32 %X, %X
|
||||
%YY = mul nsw i32 %Y, %Y
|
||||
%M = mul nsw i32 %XX, %YY
|
||||
%C = icmp sge i32 %M, 0
|
||||
ret i1 %C
|
||||
; CHECK: ret i1 true
|
||||
}
|
||||
|
Loading…
Reference in New Issue
Block a user