x = yield; makes generators larger than they need to be

[jonas-schievink]

This generator has a size of 4 Bytes:

    let mut gen = |mut x| {
        loop {
            drop(x);
            yield;
            x = makeit();
            useit(&x);
        }
    };

This one is 16 Bytes in size, even though it also does not need to keep x alive across the yield:

    let mut gen = |mut x| {
        loop {
            drop(x);
            x = yield;
            useit(&x);
        }
    };

(where makeit is a fn returning a usize, and useit is a fn taking &usize)

This seems to be fallout from #69302. Either the layout calculation soundness fix exposed it, or the visitor changes in there caused it.

This means that #69033 will increase the size of futures created from async fns, unless this bug is fixed.

[hanna-kruppe]

FWIW the evaluation order of <place> = <expr> (at MIR level) is currently: first evaluate <expr> (into a temporary), then drop <place>. For types without drop glue (like usize) that shouldn't matter, and even for types with drop glue there's reasons to change that if we can (to be able to avoid temporaries). But still, the evaluation order being this way currently means x would, in fact, have to be kept alive across the yield if it had drop glue or might have drop glue (since MIR is polymorphic, not monomorphized, and the generator->state machine transform happens on MIR).

[jonas-schievink]

True, but shouldn't the unconditional drop(x); before the yield prevent that? The value wouldn't be live anymore when the yield happens.

(there's also a separate bug where a type being Copy can actually make a generator larger, but that also shouldn't happen here)

[hanna-kruppe]

Oh, I missed that somehow. That liveness is not taken into account for generator captures is a problem independent of resume arguments, see #69663. Could async fns after #69033 have such an explicit drop too or would they still be affected if #69663 was fixed?

[jonas-schievink]

#69663 is about typeck's view of captures from what I can tell. The transform itself (which determines the memory layout) runs on optimized MIR and should already be able to tell whether variables are really live across yield points.

[tmandry]

@rustbot claim

Assigning myself to think about this issue and leave notes on the underlying issues involved.

[jonas-schievink]

I've got a fix up in #69716 that should cover most cases here.