[PATCH] D92483: AMDGPU - Use MUBUF instructions for global address space access

Wed Dec 23 11:19:12 PST 2020

pvellien added a comment.

If this is fine, place land. I don't have commit access yet.

================
Comment at: llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp:93
   if (isAmdHsaOS()) // Turn on FlatForGlobal for HSA.
     FullFS += "+flat-for-global,+unaligned-access-mode,+trap-handler,";

----------------
t-tye wrote:
> pvellien wrote:
> > t-tye wrote:
> > > Seems +flat-for-global should not be added if SI. Suggest pulling up later logic to here and combine.
> > But the gen member variable is set to correct generation only after the call to parsesubtargetfeatures function, so it would be incorrect to use getGeneration function in this place.  I think it would be better to conditionally select later.
> But could that earlier code be using some other means to determine if the target supports flat operations? Is there no target property for that?
> 
> If not then leaving "as is" is fine with me. Just would be nice if the logic for each aspect was logically organized together rather than spread over different places making it hard to figure out what is happening.
As far as I know there is no target property for that. I  agree that it would be very much clear to have a common place for this mechanism

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D92483/new/

https://reviews.llvm.org/D92483