[llvm] [X86] Prefer VPMOVMSKB for nonnegative byte comparison masks (PR #216884)

[Oscar Priego via llvm-commits]

================
@@ -46518,14 +46518,22 @@ static SDValue combineBitcastvxi1(SelectionDAG &DAG, EVT VT, SDValue Src,
                        Src.getOperand(0).getValueType() == MVT::v32i8 ||
                        Src.getOperand(0).getValueType() == MVT::v64i8);
 
-  // Prefer movmsk for AVX512 for (bitcast (setlt X, 0)) which can be handled
-  // directly with vpmovmskb/vmovmskps/vmovmskpd.
-  if (Src.getOpcode() == ISD::SETCC && Src.hasOneUse() &&
-      cast<CondCodeSDNode>(Src.getOperand(2))->get() == ISD::SETLT &&
-      ISD::isBuildVectorAllZeros(Src.getOperand(1).getNode())) {
+  // Prefer movmsk for AVX512 for bitcasts of signed sign-bit tests, which can
+  // be handled directly with vpmovmskb/vmovmskps/vmovmskpd.
----------------
Opriego wrote:

Updated the comment to keep the existing SETLT case separate from the new byte-specific cases.

https://github.com/llvm/llvm-project/pull/216884