Apply type depth limit to :new results

[carnaval]

Should fix #14009.

[yuyichao]

👍 =)

[yuyichao]

Any less confusing way of specifying a condition that type inference cannot reason about?

[carnaval]

sure, I happen to really like this but I think I'm alone on this one. I somewhat remember convincing Jeff that it made question-mark-as-a-global a feature but maybe he was just being nice.

[JeffBezanson]

I'm not sure this is a good way to fix this, since new itself is not responsible for any growth in the type. This will penalize all code that might allocate deeply-nested types, even without recursion.

We should try to do the limiting during function calls, perhaps in abstract_call, and ideally only if we detect recursion might be happening.

[carnaval]

Agreed, but I think the quick fix in the PR is analogous to how we do the tuple depth limit

julia> f() = ((((((1,),),),),),)
julia> Base.return_types(f,())[1]
Tuple{Tuple{Tuple{Tuple{Tuple{Tuple}}}}}

since tuple() and new() play the same role ?

edit: actually making it analogous would be to enforce the limit in apply_type which is what this patch should have been. f(a) = f(A{a}) also keeps inference busy.

[JeffBezanson]

Yes, if we want a quick fix then I'd much rather do this in apply_type than in new.

[JeffBezanson]

I will submit a different version of this.