[llvm] [X86] Use GFNI for vXi8 shifts/rotates (PR #89115)

[via llvm-commits]

goldsteinn wrote:

This looks net positive, but looks like there are some regressions related to demanded bits where we can actually just use a single 16-bit shift.

Maybe add support analyzing these patterns in the x86 knownbits functions?

Or maybe make this a tablgen pattern that matches the complete patterns?

https://github.com/llvm/llvm-project/pull/89115