Improve state machine codegen

[nikomatsakis]

Metadata
Point of contact	@folkertdev
Team(s)	compiler, lang
Goal document	2025h1/improve-rustc-codegen

Summary

We want to improve rustc codegen, based on this initialive by the Trifecta Tech Foundation. The work focusses on improving state machine code generation, and finding (and hopefully fixing) cases where clang produces better code than rustc for roughly equivalent input.

Tasks and status

• Discussion and moral support (compiler )
• Lang-team experiment (lang )
• Lang-team champion (lang )
• Refine RFC 3720 (@folkertdev)
• Implementation (@folkertdev, @bjorn3)
• Standard reviews (compiler )

[nikomatsakis]

This issue is intended for status updates only.

For general questions or comments, please contact the owner(s) directly.

[traviscross]

This is, I believe, mostly waiting on us on the lang team to have a look, probably in a design meeting, to feel out what's in the realm of possibility for us to accept.

[folkertdev]

@traviscross how would we make progress on that? So far we've mostly been talking to @joshtriplett, under the assumption that a #[loop_match] attribute on loops combined with a #[const_continue] attribute on "jumps to the next iteration" will be acceptable as a language experiment.

Our current implementation handles the following

#![feature(loop_match)]

enum State {
    A,
    B,
}

fn main() {
    let mut state = State::A;
    #[loop_match]
    'outer: loop {
        state = 'blk: {
            match state {
                State::A =>
                {
                    #[const_continue]
                    break 'blk State::B
                }
                State::B => break 'outer,
            }
        }
    }
}

Crucially, this does not add syntax, only the attributes and internal logic in MIR lowering for statically performing the pattern match to pick the right branch to jump to.

The main challenge is then to implement this in the compiler itself, which we've been working on (I'll post our tl;dr update shortly)