[PATCH] D135900: [AMDGPU][MC][GFX8+] Correct v_cndmask modifiers

[Dmitry Preobrazhensky via Phabricator via llvm-commits]

dp added a comment.

In D135900#3856534 <https://reviews.llvm.org/D135900#3856534>, @Joe_Nash wrote:

> I guess CNDMASK_B16 should be a True16 instruction. I shall plan to do that. With that in mind, this patch appears even less parameterized for (i32 vs i16) input types than before. Is there any way to get some of that back? IE before this patch CNDMASK_B16_T16 wouldn't need to override InsDPP, but after this it does.

I assume that True16 instructions may use first 128 VGPRs only. v_cndmask_b16 is a native VOP3 instruction, it does not have such limitations. Also, it has no e32_dpp variant, so I do not understand why InsDPP may need a correction. What am I missing here?


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D135900/new/

https://reviews.llvm.org/D135900