polymorphic specializations field: either svec or value

[vtjnash]

Attempt an experiment: remove arrays that only have one value. Seems to save about 0.8% (1MB) in the sysimg, since we were rather significantly over-sizing these (about 16 slots initialized) before.

Also a small lookup-failure optimization for the hashing-failed case to avoid duplicated work.

[aviatesk]

Can we make m.specializations::Union{SingleSpec, SimpleVector} and define iterate for SingleSpec to land this as non-breaking change? It would probably better to run pkgeval to assess the impact of this change on the ecosystem.

[vtjnash]

I don't think that is worthwhile. I have made significant changes to this field in most releases (and likely will again), so anyone depending on it without going through the API is already in trouble.

[timholy]

What is the API? I seem to be guilty of not using it: https://juliahub.com/ui/Search?type=code&q=.specializations&w=true

The MethodAnalysis package defines what I'd call a real API: mi = methodinstance(f, tt). But I'm not aware of one in Base.

[vtjnash]

AFAIK, it is only defined as Core.Compiler.specialize_method (which is also unstable, but slightly less so)

[timholy]

OK, I was trying to avoid depending on compiler internals but perhaps that's better in this case.

[timholy]

There's more: MethodAnalysis.jl can traverse the entire system (or specific modules) and collect all the existing MethodInstances. Is there an approved API for this? I riffed off of Base.visit but we don't (or didn't when I last checked) have a method for visiting all specializations of a Method. That's needed for quite a few analyses & test-suites (https://github.com/JuliaDebug/Cthulhu.jl/blob/master/TypedSyntax/test/exhaustive.jl is one recent example). Many of our internal representations are not fully documented, so the only way to know if, e.g., packages like JuliaInterpreter/CodeTracking/TypedSyntax/Revise can handle all the different constructs is to try a large collection of examples.

[vtjnash]

MethodInstance is sufficiently internal, that it does not really have a usable API except for that query

[timholy]

Ok. Assuming this is sticking around, I'll adapt to the new approach.

[vtjnash]

If you do want to add an API for that in Base somewhere as well, I will try to maintain it

[timholy]

Will do. I think it works well as a visit method for m.specializations.

[ulysses4ever]

I'm both using Tim's MethodAnalysis and occasionally index into specializations (including in tests: kinda what this patch had to handle in https://github.com/JuliaLang/julia/pull/49071/files#diff-7e050c69723b3de58379c7ac90882676badb1dcd11cd74866a0d2d7373a93a7fL289), so a good amount of my code will be broken. Oh well, I'm looking forward to an API and to an update to MehtodAnalysis, 'cause currently my CI fails on nightly inside MethodAnalysis.visit(::Module).

@timholy would it make sense to open a ticket on MethodAnalysis so that interested parties could get a notification when support for nightly will be restored?

[timholy]

@timholy would it make sense to open a ticket on MethodAnalysis so that interested parties could get a notification when support for nightly will be restored?

Already fixed (v0.4.12)

[ulysses4ever]

@timholy thank you!