[llvm] r262125 - DAGCombiner: Relax sqrt NaN folding check
Matt Arsenault via llvm-commits
llvm-commits at lists.llvm.org
Sat Feb 27 01:38:05 PST 2016
Author: arsenm
Date: Sat Feb 27 03:38:05 2016
New Revision: 262125
URL: http://llvm.org/viewvc/llvm-project?rev=262125&view=rev
Log:
DAGCombiner: Relax sqrt NaN folding check
This is OK for +0 since compares to +/-0 give the same result.
Modified:
llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
llvm/trunk/test/CodeGen/AMDGPU/llvm.sqrt.ll
Modified: llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp?rev=262125&r1=262124&r2=262125&view=diff
==============================================================================
--- llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp (original)
+++ llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp Sat Feb 27 03:38:05 2016
@@ -13814,33 +13814,33 @@ SDValue DAGCombiner::SimplifySelect(SDLo
bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
SDValue RHS) {
- // fold (select (setcc x, -0.0, *lt), NaN, (fsqrt x))
- // The select + setcc is redundant, because fsqrt returns NaN for X < -0.
+ // fold (select (setcc x, [+-]0.0, *lt), NaN, (fsqrt x))
+ // The select + setcc is redundant, because fsqrt returns NaN for X < 0.
if (const ConstantFPSDNode *NaN = isConstOrConstSplatFP(LHS)) {
if (NaN->isNaN() && RHS.getOpcode() == ISD::FSQRT) {
// We have: (select (setcc ?, ?, ?), NaN, (fsqrt ?))
SDValue Sqrt = RHS;
ISD::CondCode CC;
SDValue CmpLHS;
- const ConstantFPSDNode *NegZero = nullptr;
+ const ConstantFPSDNode *Zero = nullptr;
if (TheSelect->getOpcode() == ISD::SELECT_CC) {
CC = dyn_cast<CondCodeSDNode>(TheSelect->getOperand(4))->get();
CmpLHS = TheSelect->getOperand(0);
- NegZero = isConstOrConstSplatFP(TheSelect->getOperand(1));
+ Zero = isConstOrConstSplatFP(TheSelect->getOperand(1));
} else {
// SELECT or VSELECT
SDValue Cmp = TheSelect->getOperand(0);
if (Cmp.getOpcode() == ISD::SETCC) {
CC = dyn_cast<CondCodeSDNode>(Cmp.getOperand(2))->get();
CmpLHS = Cmp.getOperand(0);
- NegZero = isConstOrConstSplatFP(Cmp.getOperand(1));
+ Zero = isConstOrConstSplatFP(Cmp.getOperand(1));
}
}
- if (NegZero && NegZero->isNegative() && NegZero->isZero() &&
+ if (Zero && Zero->isZero() &&
Sqrt.getOperand(0) == CmpLHS && (CC == ISD::SETOLT ||
CC == ISD::SETULT || CC == ISD::SETLT)) {
- // We have: (select (setcc x, -0.0, *lt), NaN, (fsqrt x))
+ // We have: (select (setcc x, [+-]0.0, *lt), NaN, (fsqrt x))
CombineTo(TheSelect, Sqrt);
return true;
}
Modified: llvm/trunk/test/CodeGen/AMDGPU/llvm.sqrt.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/AMDGPU/llvm.sqrt.ll?rev=262125&r1=262124&r2=262125&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/AMDGPU/llvm.sqrt.ll (original)
+++ llvm/trunk/test/CodeGen/AMDGPU/llvm.sqrt.ll Sat Feb 27 03:38:05 2016
@@ -50,16 +50,28 @@ entry:
ret void
}
-; SI-LABEL: {{^}}elim_redun_check:
+; SI-LABEL: {{^}}elim_redun_check_neg0:
; SI: v_sqrt_f32_e32
; SI-NOT: v_cndmask
-define void @elim_redun_check(float addrspace(1)* %out, float %in) {
+define void @elim_redun_check_neg0(float addrspace(1)* %out, float %in) {
entry:
%sqrt = call float @llvm.sqrt.f32(float %in)
%cmp = fcmp olt float %in, -0.000000e+00
%res = select i1 %cmp, float 0x7FF8000000000000, float %sqrt
store float %res, float addrspace(1)* %out
ret void
+}
+
+; SI-LABEL: {{^}}elim_redun_check_pos0:
+; SI: v_sqrt_f32_e32
+; SI-NOT: v_cndmask
+define void @elim_redun_check_pos0(float addrspace(1)* %out, float %in) {
+entry:
+ %sqrt = call float @llvm.sqrt.f32(float %in)
+ %cmp = fcmp olt float %in, 0.000000e+00
+ %res = select i1 %cmp, float 0x7FF8000000000000, float %sqrt
+ store float %res, float addrspace(1)* %out
+ ret void
}
; SI-LABEL: {{^}}elim_redun_check_ult:
More information about the llvm-commits
mailing list