[llvm] [X86] Use the standard cmp+cmov for select (X != 0), -1, Y if we will be setting sbb to 0 anyway (PR #149672)

[Craig Topper via llvm-commits]

================
@@ -24917,6 +24917,30 @@ static SDValue LowerSELECTWithCmpZero(SDValue CmpVal, SDValue LHS, SDValue RHS,
   if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) &&
       (isAllOnesConstant(LHS) || isAllOnesConstant(RHS))) {
     SDValue Y = isAllOnesConstant(RHS) ? LHS : RHS;
+
+    // If CMOV is available, use it instead. Only prefer CMOV when SBB
+    // dependency breaking is not available or when CMOV is likely to be more
+    // efficient
+    if (Subtarget.canUseCMOV() &&
+        (VT == MVT::i16 || VT == MVT::i32 || VT == MVT::i64) &&
+        !Subtarget.hasSBBDepBreaking()) {
+      // Create comparison against zero to set EFLAGS
+      SDValue Zero = DAG.getConstant(0, DL, CmpVT);
+      SDValue Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpVal, Zero);
+
+      // For CMOV: FalseVal is used when condition is false, TrueVal when
+      // condition is true We want: when X==0 return -1, when X!=0 return Y So
----------------
topperc wrote:

```suggestion
      // condition is true. We want: when X==0 return -1, when X!=0 return Y So
```

https://github.com/llvm/llvm-project/pull/149672