release/20.x: [X86][TargetLowering] Avoid deleting temporary nodes in getNegatedExpression (#139029)

[llvmbot]

Backport 143cce7

Requested by: @dtcxzyw

[llvmbot]

@topperc What do you think about merging this PR to the release branch?

[llvmbot]

@llvm/pr-subscribers-backend-x86

Author: None (llvmbot)

Changes

Backport 143cce7

Requested by: @dtcxzyw

---

Full diff: https://github.com/llvm/llvm-project/pull/139356.diff

2 Files Affected:

• (modified) llvm/lib/Target/X86/X86ISelLowering.cpp (+8-1)
• (added) llvm/test/CodeGen/X86/pr138982.ll (+23)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 627cef9ead7ff..4413fbb77f415 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -54147,12 +54147,19 @@ 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];
diff --git a/llvm/test/CodeGen/X86/pr138982.ll b/llvm/test/CodeGen/X86/pr138982.ll
new file mode 100644
index 0000000000000..32346d823a9fe
--- /dev/null
+++ b/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
+}

[RKSimon]

LGTM

[github-actions]

@dtcxzyw (or anyone else). If you would like to add a note about this fix in the release notes (completely optional). Please reply to this comment with a one or two sentence description of the fix. When you are done, please add the release:note label to this PR.