[PATCH] D134961: [AMDGPU][MC][GFX11] Correct v_fmac_.*_e64_dpp
Joe Nash via Phabricator via llvm-commits
llvm-commits at lists.llvm.org
Fri Sep 30 10:38:56 PDT 2022
Joe_Nash added a comment.
In D134961#3827685 <https://reviews.llvm.org/D134961#3827685>, @dp wrote:
> In D134961#3827656 <https://reviews.llvm.org/D134961#3827656>, @Joe_Nash wrote:
>
>> Unfortunately there is a codegen issue with this patch. Since the Instruction would no longer have an old operand, we couldn't create the DPP version (See GCNDPPCombine.cpp:244). I have just pushed a test for that 50dfd3e9e4930834c3c77a289000fd57bbc16727 <https://reviews.llvm.org/rG50dfd3e9e4930834c3c77a289000fd57bbc16727>
>
> So for `_e64_dpp` we need both `old` and `src2` and both these operands must be tied to `dst`, correct?
Yes, it seems that would be correct to me.
The FIXME hints the presence of src2 at all is a hack, but I'm not sure how practical it is to remove src2 completely.
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D134961/new/
https://reviews.llvm.org/D134961
More information about the llvm-commits
mailing list