[PATCH] D140537: SIInsertWait: Skip tied source of d16 buffer instruction

[Ruiling, Song via Phabricator via llvm-commits]

ruiling added inline comments.


================
Comment at: llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp:1181
+
+        // D16 buffer instructions does not actually read the tied source
+        // operand, so we can skip the source operand.
----------------
arsenm wrote:
> ruiling wrote:
> > foad wrote:
> > > Should we do the same thing for _d16 DS, FLAT, SCRATCH and GLOBAL instructions?
> > For _d16 DS/FLAT, skipping the tied source does not change the code generation. As we always need a s_waitcnt for two successive _d16 ds/flat load because they may return out of order. So, I would rather not handle them here. I am actually setting the D16Buf bit for global/scratch load in parent change.  Should I change `D16Buf` to something else? I will update the comment.
> Since when can't they return out of order?
> 
> Could you demonstrate a change by combining a ds load with a global load? The two halves don't need to access the same address space
> Since when can't they return out of order?
Sorry I don't know what you are asking. Could you help explain?
> 
> Could you demonstrate a change by combining a ds load with a global load? The two halves don't need to access the same address space

A ds load followed by a global_load to the same VGPR will always have a WAW dependency. The change here is to remove a false read-after-write dependency.




Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D140537/new/

https://reviews.llvm.org/D140537