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

[Simon Pilgrim via llvm-commits]

================
@@ -265,7 +265,6 @@ define i32 @ptestz_v2i64_signbits(<2 x i64> %c, i32 %a, i32 %b) {
 ; SSE41-LABEL: ptestz_v2i64_signbits:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movl %edi, %eax
-; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
----------------
RKSimon wrote:

This is definitely wrong as we've gone from just needing the signbits of a <2 x i64> to the signbits of a <4  x i32>

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