Check for comparisons of +/- zero when optimizing less-than-or-equal and

greater-than-or-equal SELECT_CCs to NEON vmin/vmax instructions.  This is
only allowed when UnsafeFPMath is set or when at least one of the operands
is known to be nonzero.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@97065 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Bob Wilson 2010-02-24 22:15:53 +00:00
parent c382bc3c0f
commit e742bb53a3
2 changed files with 47 additions and 22 deletions

View File

@ -3879,50 +3879,59 @@ static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
unsigned Opcode = 0; unsigned Opcode = 0;
bool IsReversed; bool IsReversed;
if (LHS == CondLHS && RHS == CondRHS) { if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
IsReversed = false; // x CC y ? x : y IsReversed = false; // x CC y ? x : y
} else if (LHS == CondRHS && RHS == CondLHS) { } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
IsReversed = true ; // x CC y ? y : x IsReversed = true ; // x CC y ? y : x
} else { } else {
return SDValue(); return SDValue();
} }
bool IsUnordered;
switch (CC) { switch (CC) {
default: break; default: break;
case ISD::SETOLT: case ISD::SETOLT:
case ISD::SETOLE: case ISD::SETOLE:
case ISD::SETLT: case ISD::SETLT:
case ISD::SETLE: case ISD::SETLE:
// This can be vmin if we can prove that the LHS is not a NaN.
// (If either operand is NaN, the comparison will be false and the result
// will be the RHS, which matches vmin if RHS is the NaN.)
if (DAG.isKnownNeverNaN(LHS))
Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
break;
case ISD::SETULT: case ISD::SETULT:
case ISD::SETULE: case ISD::SETULE:
// Likewise, for ULT/ULE we need to know that RHS is not a NaN. // If LHS is NaN, an ordered comparison will be false and the result will
if (DAG.isKnownNeverNaN(RHS)) // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN; // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
break;
// For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
// will return -0, so vmin can only be used for unsafe math or if one of
// the operands is known to be nonzero.
if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
!UnsafeFPMath &&
!(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
break;
Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
break; break;
case ISD::SETOGT: case ISD::SETOGT:
case ISD::SETOGE: case ISD::SETOGE:
case ISD::SETGT: case ISD::SETGT:
case ISD::SETGE: case ISD::SETGE:
// This can be vmax if we can prove that the LHS is not a NaN.
// (If either operand is NaN, the comparison will be false and the result
// will be the RHS, which matches vmax if RHS is the NaN.)
if (DAG.isKnownNeverNaN(LHS))
Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
break;
case ISD::SETUGT: case ISD::SETUGT:
case ISD::SETUGE: case ISD::SETUGE:
// Likewise, for UGT/UGE we need to know that RHS is not a NaN. // If LHS is NaN, an ordered comparison will be false and the result will
if (DAG.isKnownNeverNaN(RHS)) // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX; // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
break;
// For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
// will return +0, so vmax can only be used for unsafe math or if one of
// the operands is known to be nonzero.
if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
!UnsafeFPMath &&
!(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
break;
Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
break; break;
} }

View File

@ -8,6 +8,14 @@ define float @fmin_ole(float %x) nounwind {
ret float %min1 ret float %min1
} }
define float @fmin_ole_zero(float %x) nounwind {
;CHECK: fmin_ole_zero:
;CHECK-NOT: vmin.f32
%cond = fcmp ole float 0.0, %x
%min1 = select i1 %cond, float 0.0, float %x
ret float %min1
}
define float @fmin_ult(float %x) nounwind { define float @fmin_ult(float %x) nounwind {
;CHECK: fmin_ult: ;CHECK: fmin_ult:
;CHECK: vmin.f32 ;CHECK: vmin.f32
@ -32,6 +40,14 @@ define float @fmax_uge(float %x) nounwind {
ret float %max1 ret float %max1
} }
define float @fmax_uge_zero(float %x) nounwind {
;CHECK: fmax_uge_zero:
;CHECK-NOT: vmax.f32
%cond = fcmp uge float %x, 0.0
%max1 = select i1 %cond, float %x, float 0.0
ret float %max1
}
define float @fmax_olt_reverse(float %x) nounwind { define float @fmax_olt_reverse(float %x) nounwind {
;CHECK: fmax_olt_reverse: ;CHECK: fmax_olt_reverse:
;CHECK: vmax.f32 ;CHECK: vmax.f32