[X86][TargetLowering] Avoid deleting temporary nodes in getNegatedExpression (#139029)

In the original case, the third call to `getCheaperNegatedExpression`
deletes the SDNode returned by the first call.
Similar to 74e6030, this patch uses
`HandleSDNodes` to prevent nodes from being deleted by subsequent calls.
Closes #138944.

Author: dtcxzyw

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -54726,12 +54726,19 @@ SDValue X86TargetLowering::getNegatedExpression(SDValue Op, SelectionDAG &DAG,
     if (!Flags.hasNoSignedZeros())
       break;
 
+    // Because getCheaperNegatedExpression can delete nodes we need a handle to
+    // keep temporary nodes alive.
+    std::list<HandleSDNode> Handles;
+
     // This is always negatible for free but we might be able to remove some
     // extra operand negations as well.
     SmallVector<SDValue, 4> NewOps(Op.getNumOperands(), SDValue());
-    for (int i = 0; i != 3; ++i)
+    for (int i = 0; i != 3; ++i) {
       NewOps[i] = getCheaperNegatedExpression(
           Op.getOperand(i), DAG, LegalOperations, ForCodeSize, Depth + 1);
+      if (!!NewOps[i])
+        Handles.emplace_back(NewOps[i]);
+    }
 
     bool NegA = !!NewOps[0];
     bool NegB = !!NewOps[1];

llvm/test/CodeGen/X86/pr138982.ll

@@ -0,0 +1,23 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc < %s -mtriple=x86_64 -mattr=+fma | FileCheck %s
+
+define <4 x float> @pr138982(<4 x float> %in_vec) {
+; CHECK-LABEL: pr138982:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    vxorps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm1
+; CHECK-NEXT:    vrcpps %xmm0, %xmm2
+; CHECK-NEXT:    vrcpps %xmm1, %xmm1
+; CHECK-NEXT:    vxorps %xmm3, %xmm3, %xmm3
+; CHECK-NEXT:    vcmpneqps %xmm0, %xmm3, %xmm0
+; CHECK-NEXT:    vbroadcastss {{.*#+}} xmm4 = [1.0E+0,1.0E+0,1.0E+0,1.0E+0]
+; CHECK-NEXT:    vblendvps %xmm0, %xmm1, %xmm4, %xmm0
+; CHECK-NEXT:    vfnmadd231ps {{.*#+}} xmm0 = -(xmm3 * xmm2) + xmm0
+; CHECK-NEXT:    retq
+entry:
+  %fneg = fneg <4 x float> %in_vec
+  %rcp = tail call <4 x float> @llvm.x86.sse.rcp.ps(<4 x float> %fneg)
+  %cmp = fcmp une <4 x float> zeroinitializer, %in_vec
+  %sel = select <4 x i1> %cmp, <4 x float> %rcp, <4 x float> splat (float 1.000000e+00)
+  %fma = call nsz <4 x float> @llvm.fma.v4f32(<4 x float> %rcp, <4 x float> zeroinitializer, <4 x float> %sel)
+  ret <4 x float> %fma
+}