[AMDGPU][MC][GFX10+] Warn when disassembling a VOP3 v_cmpx* opcode with dst!=EXEC

[dpreobra]

This is a follow up on https://reviews.llvm.org/D130345.

Joe Nash wrote:

it also occurs to me we could use a warning when disassembling something with non-exec (not 0x7e) in the destination. The point of this was to avoid that potential error, after all, so it would help fixing suspect shaders.

[llvmbot]

@llvm/issue-subscribers-backend-amdgpu

[gandhi56]

@dpreobra Should the warning be implemented in the AMDGPUDisassembler?

[dpreobra]

@dpreobra Should the warning be implemented in the AMDGPUDisassembler?

Yes. Also, I think we should not warn when the value is 0. The message should be like "unexpected dst value".

[jayfoad]

Is there any precedent for the disassembler emitting warnings?

[dpreobra]

llvm-project/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp

Line 1111 in 07bfbce

*CommentStream << "Warning: " << getRegClassName(SRegClassID)

[gandhi56]

gandhi56@6072366

*CommentStream does not seem to help with generating the warning. Is there a different diagnostic API I should use instead?

[dpreobra]

You are right, CommentStream is hardcoded to output to nulls() and we cannot change this. Sorry for bad advice.
But we can use outs() or errs() as a replacement, for example:

errs() << "; Warning: " << ... ;

There is no guarantee that our warnings will be in sync with disassembler output, so the warnings should probably include the corresponding dumps.

[gandhi56]

https://reviews.llvm.org/D133334

[jayfoad]

So there isn't a precedent. It seems like target disassemblers do not emit warnings and don't have the diagnostic machinery to do it properly.

[gandhi56]

https://reviews.llvm.org/D147424

[gandhi56]

That's right. I am looking for a way to give AMDGPUDisassembler access to the stream where the instructions are emitted through annotations hopefully.

[gandhi56]

Posted an RFC: https://discourse.llvm.org/t/rfc-diagnostics-support-in-disassembler/69761