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

[Phoebe Wang via llvm-commits]

================
@@ -25518,16 +25518,36 @@ static bool canEmitConjunctionForCCMP(SelectionDAG &DAG, SDValue Val,
     return false;
   unsigned Opcode = Val.getOpcode();
   if (Opcode == ISD::SETCC) {
-    EVT VT = Val.getOperand(0).getValueType();
+    SDValue LHS = Val.getOperand(0), RHS = Val.getOperand(1);
+    EVT VT = LHS.getValueType();
     if (!VT.isInteger())
       return false;
     CanNegate = true;
     MustBeFirst = false;
+    // A SETCC leaf whose flag source will be folded by EmitTest onto an
+    // OR/ADD/XOR node cannot appear in a non-root slot of a CCMP chain:
+    // CCMP is a subtract and CTEST is an AND, so neither can reproduce this
+    // leaf's SF/ZF. Force such a leaf to the chain root, where its flag
+    // source becomes a plain flag-setting instruction and its EFLAGS feed
+    // the next CCMP. This happens for icmp-vs-0 (and the icmp-vs-{-1,1}
+    // forms TranslateX86CC rewrites into icmp-vs-0) whose LHS is an
+    // OR/ADD/XOR the DAG will re-use for flags.
+    if (auto *C = dyn_cast<ConstantSDNode>(RHS)) {
+      ISD::CondCode CC = cast<CondCodeSDNode>(Val.getOperand(2))->get();
+      bool VsZero = C->isZero() || (CC == ISD::SETGT && C->isAllOnes()) ||
+                    (CC == ISD::SETLT && C->isOne());
+      if (VsZero && LHS.hasOneUse()) {
+        unsigned LOpc = LHS.getOpcode();
+        MustBeFirst =
+            (LOpc == ISD::OR || LOpc == ISD::ADD || LOpc == ISD::XOR) &&
+            isProfitableToUseFlagOp(LHS);
+      }
+    }
     // Designate this operation as a preferred first operation if the flags of a
     // corresponding SUB node can be reused. The root comparison is emitted as a
     // plain CMP, which can share EFLAGS with an existing SUB; a CCMP cannot.
-    PreferFirst = DAG.doesNodeExist(ISD::SUB, DAG.getVTList(VT),
-                                    {Val.getOperand(0), Val.getOperand(1)});
+    PreferFirst = MustBeFirst ||
+                  DAG.doesNodeExist(ISD::SUB, DAG.getVTList(VT), {LHS, RHS});
----------------
phoebewang wrote:

I'm a bit worry about it. PreferFirst is week than MustBeFirst. Will the assertion triggered in some complex case?

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