[llvm] [ISel][AArch64] Add CodeGen support for partial sub reductions. (PR #186809)

[Paul Walker via llvm-commits]

https://github.com/paulwalker-arm commented:

I'm not sure of the merits of this suggestion but is it worth having a single canonicalisation for `smla(A,-B, C) -> -smla(-A, B, C)`, based on the assumption the latter  is easier to optimise later? rather than all later combines having to specifically handle the negation case. I can think of downsides like "what if later combines do nothing" and given my floating-point comment below I think we'd only want this for integer types so the later combines would still need special handling for floating-point. I guess I've talking myself out of it but am leaving here just in case there's a nugget of value.

https://github.com/llvm/llvm-project/pull/186809