This is a subtle point I wanted to learn a bit more about. I want to know all reasons why I might wish to write: # assume we have `abstract type MyType end` MyType[x...] Rather than Any[x...] Assuming we have: abstract type MyType end struct MyTypeImpl1 <: MyType; end struct MyTypeImpl2 <: My…

I’ll start by saying this is one of the areas where I think Julia will change and improve in some nearby versions. So both answering how things are and how things may be changing towards. [image] MilesCranmer: Static checking of a constraint. For example, all <:MyType structs might implement a…

Thanks Chris! Regarding this: [image] ChrisRackauckas: Because of that, simply giving up and going to ::Any can be nicer on the compiler, meaning that using an abstract type can slow down type inference. This is really interesting. I’m trying to suss out the real-world impact of these thing…

[image] MilesCranmer: Basically, what performance characteristics are attached to “nicer on the compiler” in this context? Are these things a one-time cost at compile time, or are they a permanent runtime penalty? In those paragraphs, @ChrisRackauckas was discussing a specific narrow topic, t…

[image] MilesCranmer: But maybe you have a specific example of this with some clear performance difference before/after the ::Any? Here’s an idea: you’re moving elements from one container to another. The first container’s element types are typed Any. Thus, moving to the second container shou…

[image] nsajko: One effect in particular that counters the argument by @ChrisRackauckas is something I don’t completely understand, but I’ve read it said many times by the compiler developers, paraphrasing: abstract interpretation is slow compared to concrete evaluation I believe a specific …

[image] ChrisRackauckas: [image] nsajko: but it’s configurable with experimental features in Base.Experimental, e.g., @max_methods No that’s union splitting. I’m pretty sure @max_methods is for world splitting, not union splitting. help?> Base.Experimental.@max_methods │ Warning │…

Ehh it’s a little bit of both, so it’s probably better to just be exact with what it’s doing. [image] Mason: Set the maximum number of potentially-matching methods considered when running inference for methods defined in the current module. This setting affects inference of calls with inc…

[image] MilesCranmer: Here are the reasons I can think of: One very important but kinda trivial reason is: You can have methods foo(x::Vector{MyType}); these would apply to MyType[...] and not to Any[...]. In other words: Non-erasure of parametric types is not just an optimization / perform…

Yeah, maybe this is just semantics, but I’d say that yes there’s a union splitting component there, but that union splitting only occurs because world splitting happened first. Specifically, I think this distinction is important because union splitting is typically understood as not being vulernabl…

[image] MilesCranmer: Correct me if im wrong, but it might even be slower sometimes, because the runtime has to do an extra type assertion at parts of the code. there has been quite a few related discussion on union splitting / world age splitting in relation to performence, you might be inte…

Efficiency of abstract types vs `Any`

General Usage

Topic		Replies	Views
Struct with abstract-type field General Usage question	22	8047	June 24, 2020
Avoiding Vectors of Abstract Types Performance question , type-stability	22	4909	February 17, 2022
How does Any actually work? New to Julia question	46	1763	December 25, 2024
Performance problems when dealing with deliberately type-unstable code. How to use type knowledge better? Performance	33	2590	December 30, 2020
Dramatic performance change by adding additional unused method Performance	33	2389	November 23, 2021

Efficiency of abstract types vs `Any`

Related topics