[llvm] [AMDGPU] Remove generic-hsa GPU name (PR #149526)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> Why, and/or why can't it just default to one of the real CPUs (like gfx600 or gfx700) instead of being something different and special?

This is what the device library bitcode uses. It very specially should not be any real target. It needs to be devoid of all features, otherwise transforms will incorrectly trigger based on the assumption of the wrong target. e.g. wavesize intrinsics will now be treated as definitely wave64

> 
> > which also gets used as a dummy target that the faux-portable bitcode libraries use which has the minimal set of features.
> 
> Why can't we build them with an explicit -mcpu=generic{,-hsa} so we don't need extra code in AMDGPUTargetMachine to implement this?

I think spreading this hack into the downstream builds is worse than just dealing with it here 



https://github.com/llvm/llvm-project/pull/149526