List functions that are not type stable?

General Usage Performance
1

Hi all,

Is there a quick way to list functions that are not type stable during execution in a given module? I’m aware of the @code_warntype macro, or the Cthulu package, but it can be quite cumbersome to descend into all functions and test them one by one.

Thanks!

2

Not really answering your question but this recent paper on “Type Stability in Julia” might interest you. Apart from the formal part (which, frankly, I didn’t try to understand), they also seem to have systematically analyzed the type stability of several Julia packages. I would assume that the tools / the code with which they did it is accessible somewhere.

1 Like
3

@ulysses (first author of the paper): First of all, cool analysis! Do you think that the tool that you used to analyze the type (in)stability of methods of a package could be made a Julia package? Ideally, this would make it easy for people to analyze their own packages.

2 Likes
4

For analysis of a toplevel call, I usually use JET.jl and its @report_call macro.

3 Likes
5

Hm…

julia> @code_warntype rand([1, 2.2, "string"])
MethodInstance for rand(::Vector{Any})
  from rand(X) in Random at /some/path/share/julia/stdlib/v1.7/Random/src/Random.jl:259
Arguments
  #self#::Core.Const(rand)
  X::Vector{Any}
Body::Any
1 ─ %1 = Random.default_rng()::Core.Const(Random.TaskLocalRNG())
│   %2 = Random.rand(%1, X)::Any
└──      return %2

julia> using JET

julia> @report_opt rand([1, 2.2, "string"])
No errors !
1 Like
6

That function seems perfectly stable to me - you put in a Vector{Any} after all. The output type is perfectly well defined by the input type.

3 Likes
7

@carstenbauer thanks for mentioning it! I do think that the tool can be useful for this scenario. One caveat is that we were aiming at bulk data processing so the interface is geared towards: we generate some CSV files and you go and figure it. But that’s may be fine, actually. Writing up how one would analyze a single package using the tool was always on my TODO list but I never got to it. Maybe I should.

Btw the right GitHub repo to reference is https://github.com/prl-julia/julia-type-stability

4 Likes
8

Thank you all for your suggestions!

Jet.jl and @report_opt definitely looks like the right tool for my use case, however it is a bit verbose and doesn’t seem to catch all the issues with type-unstable functions in my package (I haven’t tried to narrow it down to a MWE though).

@ulysses impressive work! A user-friendly tool for analyzing packages based on this would be amazing!

So at the moment I’ll keep using Cthulhu…