AOT Loop unrolling breaks Yk.

[vext01]

If you crank up the optimisation level on the promote_expr.c test it fails with:

ld.lld: error: yk_mt_control_point() must be called inside a loop.

And if you look at the AOT IR, you can see that the interpreter loop was unrolled and the calls to the control point have been duplicated (which may cause some related issues).

The safest thing to do in the short term would be to disable loop unrolling loops containing a call to the control point.

(We don't want to disable all loop unrolling though)

[ltratt]

The safest thing to do in the short term would be to disable loop unrolling loops

I think you mean "disable AOT loop unrolling in functions containing a control point"? [Obviously we still need to unroll such loops at tracing time!]

[vext01]

disable AOT loop unrolling in functions containing a control point

From reading about it, it's not a function attribute, but rather metadata attached to an instruction in the loop itself. I assume you annotate the loop header, but I've not looked deeper yet.

If it pans out this way, this is actually good, as it's more flexible.

It means in a scenario like:

while(...) {
   control_point();
    ...
}

while(...) {
   some other random loop
}

The second loop could still be unrolled.