[llvm] [LangRef] inline asm: the instructions are treated opaquely (PR #157080)

[James Y Knight via llvm-commits]

https://github.com/jyknight approved this pull request.

The current wording LGTM.

>From pushback from arsenm on PR 155491, I don't think the proposal to add a magic comment hack to AMDGPU looks destined to be accepted regardless of what happens here. Given that, I don't think there's any reason to have a debate in this PR about whether we ought to accommodate it.

In general, storing structured data in inline-asm comments is not a good design -- and at the _IR level_, source-code compatibility with other compilers is not at all a valid rationale. That is to say, if passing this data in a inline-asm comment string was the best way to spell this functionality in C (which I am skeptical of, but just for argument's sake...), then the _frontend_ may parse the magic comment, and emit a more reasonable representation in IR.

https://github.com/llvm/llvm-project/pull/157080