Stream: cranelift

Topic: custom passes in cranelift?

fitzgen (he/him) (Dec 16 2019 at 21:38):

We don't currently have any way for a cranelift embedder to register custom passes over a function, right? cc @Dan Gohman

Sean Stangl (Dec 16 2019 at 21:39):

@Nick Fitzgerald we don't, and the IR isn't stable.

fitzgen (he/him) (Dec 16 2019 at 21:41):

thanks

stability is a bit of an orthogonal issue, since if you are translating to clif in the first place, you're already dealing with the instability, but I get the point of why we might not want to commit to supporting custom passes

Sean Stangl (Dec 16 2019 at 21:43):

There's an open question atm about whether we should keep the legalized form of instructions as the IR or have an explicitly MIR-like IR that doesn't require selecting an Encoding

Sean Stangl (Dec 16 2019 at 21:46):

The current design was made sort of on the assumption that there wouldn't be any real passes other than regalloc and instruction selection: all the LICM/GVN/etc things were intended to be done by the entity that emits the CLIF.

Benjamin Bouvier (Dec 17 2019 at 09:16):

Switching from one IR (wasm) to another (CLIF) gives new optimization opportunities, so LICM/GVN etc. may be worth doing again, if data suggests they're relevant.

Last updated: Nov 22 2024 at 16:03 UTC