[X86] combineAndNotOrIntoAndNotAnd - don't fold other constant operands

[Saldivarcher]

Looks like having a constant in Z also caused infinite loops. This fixes #113240.

[llvmbot]

@llvm/pr-subscribers-backend-x86

Author: Miguel Saldivar (Saldivarcher)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/113264.diff

2 Files Affected:

• (modified) llvm/lib/Target/X86/X86ISelLowering.cpp (+2-2)
• (modified) llvm/test/CodeGen/X86/pr108731.ll (+20)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index bcb84add65d83e..5ae5a7dbe0f6da 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -50049,8 +50049,8 @@ static SDValue combineAndNotOrIntoAndNotAnd(SDNode *N, SelectionDAG &DAG) {
   SDValue X, Y, Z;
   if (sd_match(N, m_And(m_Value(X),
                         m_OneUse(m_Or(m_Value(Y), m_Not(m_Value(Z))))))) {
-    // Don't fold if Y is a constant to prevent infinite loops.
-    if (!isa<ConstantSDNode>(Y))
+    // Don't fold if Y and Z are constants to prevent infinite loops.
+    if (!isa<ConstantSDNode>(Y) && !isa<ConstantSDNode>(Z))
       return DAG.getNode(
           ISD::AND, DL, VT, X,
           DAG.getNOT(
diff --git a/llvm/test/CodeGen/X86/pr108731.ll b/llvm/test/CodeGen/X86/pr108731.ll
index 473b4f7f4da2e3..2983d108eaeddc 100644
--- a/llvm/test/CodeGen/X86/pr108731.ll
+++ b/llvm/test/CodeGen/X86/pr108731.ll
@@ -192,3 +192,23 @@ define void @PR112347(ptr %p0, ptr %p1, ptr %p2) {
   ret void
 }
 
+define void @PR113240(i64 %a) {
+; CHECK-LABEL: PR113240:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    movq %rdi, %rax
+; CHECK-NEXT:    notq %rax
+; CHECK-NEXT:    movabsq $8796093022206, %rcx # imm = 0x7FFFFFFFFFE
+; CHECK-NEXT:    notq %rcx
+; CHECK-NEXT:    orq %rax, %rcx
+; CHECK-NEXT:    andq %rdi, %rcx
+; CHECK-NEXT:    movq %rcx, 0
+; CHECK-NEXT:    retq
+entry:
+  %and = and i64 %a, 8796093022206
+  %bf.value = and i64 8796093022206, 0
+  %not = xor i64 %and, -1
+  %and4 = and i64 %a, %not
+  store i64 %and4, ptr null, align 8
+  ret void
+}
+

[Saldivarcher]

@RKSimon let me know what you think.

[Saldivarcher]

@phoebewang and @goldsteinn would you be able to review this? @RKSimon might be away.

[RKSimon]

Sorry, been travelling for the devmtg

[Saldivarcher]

Sorry, been travelling for the devmtg

It's ok, sorry for bugging you! Safe travels btw!

[goldsteinn]

LGTM, although this type of re-ordering is liable to creating inf loops and this might be the only case. If this continues to be an issue we should probably drop the fold.

[Saldivarcher]

@goldsteinn would you be able to rebase and merge for me? I don't have write access. I'd be ok with reverting if this keeps causing issues.

Thanks for taking a look btw. 👍