New structurizer: now with ∞% more φ!

[eddyb]

Not sure this is ready, but it seems to handle everything I've thrown at it.

In an effort to test it, I came up with this silly example:

(click to open example code + graph rendering)

pub fn marker() {}

pub fn shortcircuit_and(a: bool, b: bool) {
    if a && b {
        marker();
    }
}

pub fn nested_while_loop(a: bool, b: bool) {
    while a {
        let mut i = 0;
        loop {
            marker();
            match i {
                0 | 1 | 2 | 4 => {}
                5 => break,
                _ => continue,
            }
            marker();
            i += 1;
        }
        if a & b {
            break;
        }
        if a | b {
            return;
        }
        while b {
            marker();
        }
    }
}

I was inspired by the loop-specific solution of "threading all exits through one break (with a tag indicating which to take when outside the loop)" and tried to extend it to handle non-structural acyclic branches as well - the result is this PR.

You can see all the OpPhis, which are used to decide whether a "deferred exit" executes, when it's finally "fully accounted for" (or "owned"? not sure what the best terminology is) by a "region".

Most of the implementation is a SEME -> SESE propagation (with the aforementioned "deferred exits"), which is then used to handle both acyclic and cyclic "regions".

[khyperia]

I do not believe I am qualified to review new_structurizer.rs, nor am I going to try to other than a quick glance-over, because wow okay that's a lot of algorithm. Other than that, though, I think this is good to merge! (not hitting the approve button because mergify scares me)

[khyperia]

👍 ✅ 🆗 💯

[DJMcNab]

Either<Cow, SmallVec>

[VZout]

I assume cargo test was successful. Seems way better than my code& algorithm. Haven't ran spirv-cfg to see if I can see any weird things.

[khyperia]

heck yee, love yeeting #[ignore]s