[llvm] [AMDGPU] Re-enable closed-world assumption as an opt-out feature (PR #115371)

[Johannes Doerfert via llvm-commits]

jdoerfert wrote:

> > I'm not sure that is related. Essentially what you are describing is ABI linking, which is not officially supported. That's why we only allow RDC for those programming models we support, and in RDC mode I don't think it's gonna happen.
> 
> "ABI" linking works just fine for globals, it's only an issue for functions and functions that reference LDS. We still define global constructors in "non-RDC" mode, the distinction isn't really important once we're at the backend.


We have had this conversation many times before. 
As said before: please point out the documentation that tells me this is expected to work.

Long story short, you cannot make the argument it "happens to work", hence we have to continue to support it.
There is no reason to assume this breaks code outside of the inside LLVM use case, at least I'm not aware of one. The LLVM use case can opt out of this.
There is no alternative to this, and holding this back for the LLVM use case is not helpful to anyone, especially since the feature is not "officially" supported to begin with.


If there is any other objection to this, please let us know, otherwise, I'll advocate to accept this and update the internal use case.




https://github.com/llvm/llvm-project/pull/115371