[SelectionDAG] Remove unused FP constant in getNegatedExpression

[qiucf]

Summary

960cbc53 immediately removes nodes that won't be used to avoid compilation time explosion. This patch adds the removal to constants to fix PR47517.

[RKSimon]

Have you tried to reduce the test case any further?

[qiucf]

Reduce test case

[steven.zhang]

llvm/test/CodeGen/X86/pr47517.ll

2

Please update the test with FileCheck

[qiucf]

Update test using script

[steven.zhang]

LGTM and thank you for doing this. But please for a while to see if RKSimon has more comments.

[RKSimon]

LGTM

[qiucf]

After running sanitizer, I think the bug is not totally covered by this patch.

In D86689, we introduced such pattern:

// Negate the X if its cost is less or equal than Y.
if (NegX && (CostX <= CostY)) {
  Cost = CostX;
  SDValue N = DAG.getNode(ISD::FSUB, DL, VT, NegX, Y, Flags);
  RemoveDeadNode(NegY);
  return N;
}

Somewhere in this test case, N=3.0, NegY=3.0, NegY->use_size()=0, so NegY is just deleted but N also gets deleted.

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp

       break;
 
     APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF();
     V.changeSign();
     SDValue CFP = DAG.getConstantFP(V, DL, VT);
 
     // If we already have the use of the negated floating constant, it is free
     // to negate it even it has multiple uses.
-    if (!Op.hasOneUse() && CFP.use_empty())
+    if (!Op.hasOneUse() && CFP.use_empty()) {
+      RemoveDeadNode(CFP);
       break;
+    }
     Cost = NegatibleCost::Neutral;
     return CFP;
   }
   case ISD::BUILD_VECTOR: {
     // Only permit BUILD_VECTOR of constants.
     if (llvm::any_of(Op->op_values(), [&](SDValue N) {
           return !N.isUndef() && !isa<ConstantFPSDNode>(N);
         }))

llvm/test/CodeGen/X86/pr47517.ll

+; RUN: llc -mtriple x86_64 < %s
+
+@a = external global { float, float }
+
+; To ensure unused floating point constant is removed in negation
+define void @test() {
+entry:
+  br label %for.cond
+
+for.cond:
+  %b.0 = phi i16 [ undef, %entry ], [ %conv87, %for.cond ]
+  store float 0.000000e+00, float* getelementptr inbounds ({ float, float }, { float, float }* @a, i16 0, i32 0)
+  store float 0.000000e+00, float* getelementptr inbounds ({ float, float }, { float, float }* @a, i16 0, i32 1)
+  %a.real3 = load float, float* getelementptr inbounds ({ float, float }, { float, float }* @a, i16 0, i32 0)
+  %conv = sitofp i16 %b.0 to float
+  %mul_ac = fmul fast float %a.real3, %conv
+  %add.r = fadd fast float %mul_ac, %a.real3
+  store float %add.r, float* getelementptr inbounds ({ float, float }, { float, float }* @a, i16 0, i32 0)
+  %a.real8 = load float, float* getelementptr inbounds ({ float, float }, { float, float }* @a, i16 0, i32 0)
+  %factor = fmul fast float %a.real8, 2.000000e+00
+  store float %factor, float* getelementptr inbounds ({ float, float }, { float, float }* @a, i16 0, i32 0)
+  store float 0.000000e+00, float* getelementptr inbounds ({ float, float }, { float, float }* @a, i16 0, i32 1)
+  %a.real14 = load float, float* getelementptr inbounds ({ float, float }, { float, float }* @a, i16 0, i32 0)
+  %mul_ac18 = fmul fast float %a.real14, %a.real14
+  %mul_ac40 = fmul fast float %a.real14, 2.000000e+00
+  %reass.add = fadd fast float %mul_ac40, -3.000000e+00
+  %reass.mul = fmul fast float %reass.add, %a.real14
+  %0 = fadd fast float %a.real14, %conv
+  %sub.r71 = fadd fast float %0, %reass.mul
+  %mul_ac73 = fmul fast float %mul_ac18, %sub.r71
+  %conv87 = fptosi float %mul_ac73 to i16
+  br label %for.cond
+}