[llvm] [AMDGPU][GlobalISel] Improve combining on v_ldexp (PR #190236)

[Petar Avramovic via llvm-commits]

petar-avramovic wrote:

> just to confirm, you want me to just rip out this fmul combine from pre and post legalizer and just move it completely to regbankcombine?

No, it is fine as it is right now, just make it more obvious which cases happen before legalizer and which in regbankcombiner.
I am fine with checking what happens if s16/s32 combine if fully moved to regbankcombiner, no need to move s64 combine just comment/make it obvious it happens before regbank-select/legalize.

For context, for S64 case it will be harder to match since S64 select will be lowered to two S32 selects, and I doubt it will generate better code. Although I tried removing this combine from from pre/post legalizer-combiner and some cases were actually better, obviously S64 cases did not combine. But did not really check what happened. So if you want, try moving S64 combine later just to see if it is better or worse


https://github.com/llvm/llvm-project/pull/190236