[PATCH] D124192: [AMDGPU] Callee must always spill writelane VGPRs

Mon Jun 20 09:06:50 PDT 2022

cdevadas added a comment.

In D124192#3577496 <https://reviews.llvm.org/D124192#3577496>, @cdevadas wrote:

> In D124192#3534493 <https://reviews.llvm.org/D124192#3534493>, @nhaehnle wrote:
>
>> I don't think the premise of this patch is correct. A source-level `writelane` intrinsic is **not** allowed to overwrite inactive lanes. If it does, then that's UB. If you want wave-wide `writelane` in source, the way to achieve that would be to use WWM.
>
> Are you saying the usage of writelane intrinsic as specified in test `kernel_writelane_intrinsic` added with llvm/test/CodeGen/AMDGPU/spill-writelane-vgprs.ll doesn't really cause any inactive lane overwrite?
> I remember having such usages being mentioned for a potential inactive lane overwrite issue.
>
> Sorry about the late response (was behind a priority work).

Ping @nhaehnle

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D124192/new/

https://reviews.llvm.org/D124192