[llvm] [X86][ISel] Improve logic for optimizing `movmsk(bitcast(shuffle(x)))` (PR #68369)

[via llvm-commits]

================
@@ -45836,18 +45836,52 @@ static SDValue combineSetCCMOVMSK(SDValue EFLAGS, X86::CondCode &CC,
   // MOVMSK(SHUFFLE(X,u)) -> MOVMSK(X) iff every element is referenced.
   SmallVector<int, 32> ShuffleMask;
   SmallVector<SDValue, 2> ShuffleInputs;
+  SDValue BaseVec = peekThroughBitcasts(Vec);
   if (NumElts <= CmpBits &&
-      getTargetShuffleInputs(peekThroughBitcasts(Vec), ShuffleInputs,
-                             ShuffleMask, DAG) &&
+      getTargetShuffleInputs(BaseVec, ShuffleInputs, ShuffleMask, DAG) &&
       ShuffleInputs.size() == 1 && !isAnyZeroOrUndef(ShuffleMask) &&
       ShuffleInputs[0].getValueSizeInBits() == VecVT.getSizeInBits()) {
----------------
goldsteinn wrote:

The thing is its overly conservative. As long as all the "high" elements stay "high", we are fine. We don't need the proper words to stay together.

https://github.com/llvm/llvm-project/pull/68369