[llvm] [AArch64] Use `AArch64ISD::UADDLP` for manual widening adjacent arithmetic (zext/shuffle combination) (PR #189255)

[Rajveer Singh Bharadwaj via llvm-commits]

Rajveer100 wrote:

I don't think TableGen is optimal here, from my view it would handle basic cases but would struggle with complex ones.

Regarding UZP1/UZP2, maybe this could be done since we would avoid potentially checking for the even/odd indices as we already know this, but what about the inverse case?

Also, we would be doing two optimisations which would be an extra step, since we detect the shuffle pattern before any other conversions happen.

https://github.com/llvm/llvm-project/pull/189255