Enable deoptimisation of inlined constants.

[ptersilie]

I'm raising this as a DRAFT, as I'm not quite sold on the design yet and would like to have some additional eyes on this. There's two main issues: 1) We need to introduce a LocalAlloc::Const, which doesn't make much sense conceptually. 2) We currently don't adjust for the size of the constant (this can be easily fixed by adding the size to the enum variant)

---

Due to function inlining (and later other optimisations) it can happen that a variable in AOT becomes a constant in the JIT. Currently, neither the trace builder nor deoptimisation can handle such cases, which this commit fixes, by inserting a ProxyConst instruction and reference this instead of the direct constant.

[ltratt]

I have a couple of comments that are probably worth addressing, but overall I'd say: let's just do this. It might turn out to be imperfect, and then we can iterate on it in-tree.

[ltratt]

The other thing we'll want to add to this PR is at least one unit test to x86_64/mod.rs for the const case.

[ptersilie]

The other thing we'll want to add to this PR is at least one unit test to x86_64/mod.rs for the const case.

Sure. Can the JIT IR parser deal with ProxyConsts or do I need to add that first?

[ltratt]

Can the JIT IR parser deal with ProxyConsts or do I need to add that first?

Good point. No, it doesn't deal with them because I originally thought "they can't appear in the output therefore they shouldn't be parsed". But I think that was wrong: they should be parsed even though they can't appear in the output. I would suggest %0: i8 = 1i8esque syntax for this.

[I think we're tending in a direction where ProxyConst will become AssignConst, but we can worry about that later.]

[ptersilie]

Yeah I'm also starting to think a normal assign instruction isn't a bad idea. It would make the special LocalAlloc::Const case go away too.

[ltratt]

For this PR I think I would stick with ProxyConst just because there will be more knock-on effects from changing it than are first apparent, and such a change is probably best done in isolation.

[ltratt]

I think it's just Const -> ConstInt and the unit test (and JIT IR parser) change needed to get this in.

[ptersilie]

This should address the remaining issues.

[ltratt]

Please squash and mark the PR as "ready for review".

[ptersilie]

Squashed.

[ptersilie]

Force pushed some C formatting changes.

[ptersilie]

Sorry, forgot to run the other tests. Doing so revealed that the instruction printing needed changing to deal with proxy stuff.

[ltratt]

Hmm, not totally convinced by that new patch, but I'm also unsure what the right thing should be! Let's get it in and work out what to do later.

Please squash.

[ptersilie]

Yeah, I must admit I wasn't too sure either.

Squashed.