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

[Nikita Popov via llvm-commits]

nikic wrote:

Just to clarify my position on this: LLVM should absolutely not be looking at magic comments in inline assembly to derive semantics. And I think that everyone apart from @linuxrocks123 agrees with that -- including AMDGPU maintainers, who have already rejected this approach in the linked PR. We should not be changing the wording to accommodate this use case.

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