[PATCH] D141717: [Clang] Only emit textual LLVM-IR in device only mode

Fri Jan 13 14:31:06 PST 2023

jhuber6 added a comment.

In D141717#4052753 <https://reviews.llvm.org/D141717#4052753>, @tra wrote:

> In D141717#4052587 <https://reviews.llvm.org/D141717#4052587>, @jhuber6 wrote:
>
>> Well you'll get textual output for the host output, but the device code embedded in the host module will be bitcode instead. So the final output from the compiler is still textual IR. It just won't be some weird global like this
>>
>> This is bad because it can't be handled by LTO or anything else. It makes the resulting IR file difficult to use for its intended purpose.
>
> I understand your use case and I do agree that you do need the embedded IR to be binary.
>
> It appears to me that the real problem here is that "-S"  should not have been propagated to the GPU sub-compilation.
>
> The fact that producing binary IR works for you seems to be just a coincidence. What if the user would specify "-E" instead? I think the right fix would be to tell compiler to generate the right kind of output, not try to tell it to produce textual assembly and undo the "textual" part somewhere downstream.

For `-E` we don't embed anything, so I think we just ignore the device portion unless they used `--offload-device-only`. And I don't think we have access to any "per-toolchain" options at this point in the compilation phase.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D141717/new/

https://reviews.llvm.org/D141717