[llvm] [X86] Fix CCMP miscompile for flag-source-folded SETCC leaves (PR #217682)

[Phoebe Wang via llvm-commits]

================
@@ -25736,23 +25739,52 @@ static SDValue emitConjunctionForCCMPRec(SDValue Val, X86::CondCode &OutCC,
     assert(X86CC != X86::COND_INVALID);
     OutCC = X86CC;
 
-    SDValue Flags = EmitCmp(LHS, RHS, X86CC, DL, DAG, Subtarget);
-    // Produce a normal comparison if we are first in the chain.
+    // Produce a normal comparison if we are first in the chain. Only there may
+    // EmitCmp fold the comparison into the EFLAGS of an existing arithmetic or
+    // logic node: those flags are unconditional, which is exactly what the root
+    // of the chain needs.
     if (!CCOp)
-      return Flags;
+      return EmitCmp(LHS, RHS, X86CC, DL, DAG, Subtarget);
----------------
phoebewang wrote:

Looks like we can use it in non-root as long as the node is a SUB. Anyway, let's fix miscompilation first.

https://github.com/llvm/llvm-project/pull/217682