FYI: IBM is working on LLVM support for arbitrary aliasing graphs (for Fortran)

I just stumbled upon a recent talk from 2020 LLVM Developers' Meeting that I figured might be of interest to the Rust compiler team?

Towards a representation of arbitrary alias graph in LLVM IR for Fortran code by Kelvin Li (IBM) & Tarique Islam (IBM).

The talk didn't make any mention of Rust specifically (they only briefly mention "other languages"), so I was wondering if either team is aware of these potentially symbiotic efforts and/or might want to provide feedback to or collaborate with them in order to ensure their API also covers Rust's language model nicely, enabling optimization of non-aliasing borrows on a par with Fortran?

At around 22:13 minute mark they have a slide specifically asking for feedback and/or collaboration with links and schedule for a regular round table. The next scheduled meeting is on November 3rd, apparently.

Just thought I'd leave this here to make sure you folks are aware of these possibly symbiotic efforts.

6 Likes

How about they (LLVM) first fix their noalias tags after several years, so that they are actually usable :grimacing:

3 Likes

Perhaps this new effort will serve as the motivator for that long-awaited fix. It can't hurt to point out that old issue to this new focus group.

1 Like

AFAIK there already is an ongoing effort for "better noalias" in LLVM; I wonder if that is related or a separate project?

1 Like