DAGCombine: Remove redundant NaN checks around ISD::FSQRT

[• tstellarAMD]

Summary

This folds:

(select (setcc x, -0.0, olt), NaN, (fsqrt x)) -> ( fsqrt x)

[resistor]

LGTM

—Owen

[mehdi_amini]

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

12444

s/becuase/because

12467

Since fsqrt(NaN) = NaN (I believe), SETULT should be fine as well here right? And SETLT too?

[• tstellarAMD]

Here is an updated patch with the typo fix and with an expanded pattern that patches all of the *LT flavors of condition codes. I forgot to mention before that this patch depends on http://reviews.llvm.org/D8469

[mehdi_amini]

LGTM.

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

     }
   }
 
+  if (SimplifySelectOps(N, N1, N2))
+    return SDValue(N, 0);  // Don't revisit N.
+
   // If the VSELECT result requires splitting and the mask is provided by a
   // SETCC, then split both nodes and its operands before legalization. This
   // prevents the type legalizer from unrolling SETCC into scalar comparisons

 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.
+  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;
+
+      if (TheSelect->getOpcode() == ISD::SELECT_CC) {
+        CC = dyn_cast<CondCodeSDNode>(TheSelect->getOperand(4))->get();
+        CmpLHS = TheSelect->getOperand(0);
+        NegZero = 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));
+        }
+      }
+      if (NegZero && NegZero->isNegative() && NegZero->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))
+        CombineTo(TheSelect, Sqrt);
+        return true;
+      }
+    }
+  }
   // Cannot simplify select with vector condition
   if (TheSelect->getOperand(0).getValueType().isVector()) return false;
 

test/CodeGen/R600/llvm.sqrt.ll

   ret void
 }
 
+; SI-LABEL: {{^}}elim_redun_check:
+; SI: v_sqrt_f32_e32
+; SI-NOT: v_cndmask
+define void @elim_redun_check(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:
+; SI: v_sqrt_f32_e32
+; SI-NOT: v_cndmask
+define void @elim_redun_check_ult(float addrspace(1)* %out, float %in) {
+entry:
+  %sqrt = call float @llvm.sqrt.f32(float %in)
+  %cmp = fcmp ult 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_v2:
+; SI: v_sqrt_f32_e32
+; SI: v_sqrt_f32_e32
+; SI-NOT: v_cndmask
+define void @elim_redun_check_v2(<2 x float> addrspace(1)* %out, <2 x float> %in) {
+entry:
+  %sqrt = call <2 x float> @llvm.sqrt.v2f32(<2 x float> %in)
+  %cmp = fcmp olt <2 x float> %in, <float -0.000000e+00, float -0.000000e+00>
+  %res = select <2 x i1> %cmp, <2 x float> <float 0x7FF8000000000000, float 0x7FF8000000000000>, <2 x float> %sqrt
+  store <2 x float> %res, <2 x float> addrspace(1)* %out
+  ret void
+}
+
+; SI-LABEL: {{^}}elim_redun_check_v2_ult
+; SI: v_sqrt_f32_e32
+; SI: v_sqrt_f32_e32
+; SI-NOT: v_cndmask
+define void @elim_redun_check_v2_ult(<2 x float> addrspace(1)* %out, <2 x float> %in) {
+entry:
+  %sqrt = call <2 x float> @llvm.sqrt.v2f32(<2 x float> %in)
+  %cmp = fcmp ult <2 x float> %in, <float -0.000000e+00, float -0.000000e+00>
+  %res = select <2 x i1> %cmp, <2 x float> <float 0x7FF8000000000000, float 0x7FF8000000000000>, <2 x float> %sqrt
+  store <2 x float> %res, <2 x float> addrspace(1)* %out
+  ret void
+}
+
 declare float @llvm.sqrt.f32(float %in)
 declare <2 x float> @llvm.sqrt.v2f32(<2 x float> %in)
 declare <4 x float> @llvm.sqrt.v4f32(<4 x float> %in)