[llvm] [RFC] IR: Define noalias.addrspace metadata (PR #102461)

Alex Voicu via llvm-commits llvm-commits at lists.llvm.org
Fri Aug 9 04:44:01 PDT 2024


AlexVlx wrote:

> > Is the MD a promise to the backend that the attached memory operation is not in a specific address space?
> 
> Yes, for the underlying object definition.
> 
> > Who will attach the MD to the IR? FE or some LLVM passes? FE may not be able to figure out the real address space of some memory operations.
> 
> The main user is the frontend since OpenCL/CUDA give the guarantee. We can also infer this in many cases in backend passes.

Just a quick check: in the absence of the MD (neither the FE nor inference are deterministic when it comes to figuring out the concrete AS), we'd still fall back to the run time check on things like atomics on private, correct?

https://github.com/llvm/llvm-project/pull/102461


More information about the llvm-commits mailing list