[PATCH] D147245: [AMDGPU] Make amdgcn.s.buffer.load a memory-reading intrinsic

[Krzysztof Drewniak via Phabricator via llvm-commits]

krzysz00 added a comment.

Ok, so, just to clarify:

1. There's an undocumented invariant that the buffer you pass in to `s.buffer.load` is constant for the duration of the program.
2. A bunch of folks are relying on that invariant for performance.
3. That aside, `s.buffer.load` is, more or less, `raw.buffer.load`

Given that those things are true, I'm going to back out of this change to `s.buffer.load`, and we'll leave all these codepaths untouched during the buffer intrinsic updates for now. Once address space 7 properly exists, folks'll be able to migrate to that system and use a typical `load` with the appropriate "this points to constant memory" annotations, so we may end up not needing this special intrinsic.

So, I'll go ahead and file a new revision to update the docs.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D147245/new/

https://reviews.llvm.org/D147245