Add a test for shadow stack allocation before an external call.

[vext01]

An external call can "call-back" to interpreter code, so we still need to allocate shadow space.

Requires ykjit/ykllvm#217

[vext01]

hold off on this for this and the companion for a moment actually.

[vext01]

Force pushed test fixes that I didn't anticipate. Go ahead.

[vext01]

Oh this will need a ykllvm sync once the compiler PR is merged.

[ltratt]

Please force push when that's ready.

[vext01]

synced and ready.

(I note that the ykllvm sync also pulled in some SWT-related stuff)

[ltratt]

Yes the SWT parts are expected.

[vext01]

Ready when you are then.

[ltratt]

11:58:56 ***** FILE 'events.lua'*****
11:58:56 testing metatables
11:58:56 ..Segmentation fault (core dumped)
11:58:59 + status=139
11:58:59 + [ 1 -eq 1 ]

Do we think that's a non-deterministic failure, or perhaps evidence of a problem caused the ykllvm change?

[vext01]

Hrm, hard to say. I have a try_repeat going to see if I can repro it locally.

Maybe retry it?

[ltratt]

Let's give it a minute: if you can get to 100 without crashes (or if you believe any crashes you do see are unrelated), we can retry.

[vext01]

My run of 100 serialised debug runs just finished without hitting this seg 🤔

[ltratt]

OK, well let's retry, and see what happens.

[ltratt]

OK, it's crashed in the same place, so I think we can assume we've hit an actual problem.

[vext01]

OK, i'll try harder to repro.

[vext01]

I've pushed an assertion that may save us some debugging effort in the future.

With ykjit/ykllvm#219 I think this should merge (if the benchmarks turn out to be ok).

[vext01]

This will need an llvm sync after ykjit/ykllvm#219 is in.

[vext01]

synced llvm.