This trivial macro was suggested (or a similar macro was) by @tim.holy and @tkf . It is not registered. Provides the macro @unionsplit to dispatch “manually” on types rather than rely on runtime multiple dispatch. If type information is hidden from the compiler, then the dispatch system has to se…

Thanks for sharing the package! Regarding the syntax, how about something like @unionsplit(func(y::Union{Int, Float64})) so that it can be used for arbitrary argument position?

Yes, that would be much more flexible. It currently only manually dispatches on the first argument.

Thanks for creating this! This is a usable manual version of the core of my JIT proposal, and it seems we have worked on the two completely parallel. The weekend of manual dispatch! :smiley: I did not know of previous work in this area (at least not consciously) but I guessed that the idea is not …

Thanks @jlapeyre or the idea and package! [image] tkf: Regarding the syntax, how about something like @unionsplit(func(y::Union{Int, Float64})) so that it can be used for arbitrary argument position? Expanding on this idea, I’m not sure why we should restrict the use of such a macro to a sin…

I thought the whole point of union splitting was to do this automatically, as explained in this blog post: Union-splitting: what it is, and why you should care . Why is this faster?

If I understand correctly, this package is useful when inference can’t tell anything about the type, i.e. the type is Any. In the README example, would there be no advantage to the macro if the vector was typed with Union{Float64,Int} elements?

Yes, that’s also my understanding. Also, automatic union splitting by the compiler kicks in for small unions. With manual union splitting, you can get the same kind of benefits in the case of larger unions, for which the benefit/cost ratio of automatic compiler optimizations was deemed too low.

[image] jebej: would there be no advantage to the macro if the vector was typed with Union{Float64,Int} elements? using BenchmarkTools using ManualDispatch func(x::Int) = 1 func(x::Float64) = 1 function ex_union_split(x) for y in x @unionsplit((Int, Float64), func(y)) end en…

It seems things are being constant prop’ed away, 1ns is too quick!

[image] ffevotte: Expanding on this idea, I’m not sure why we should restrict the use of such a macro to a single function call? Couldn’t we allow arbitrary expressions? Absolutely! (In fact, I wrote a half of the sentence suggesting this but I then I thought it might be better to keep the co…

[ANN] ManualDispatch.jl

Package Announcements

j-fu February 7, 2022, 10:24pm 16

Also, there is #37378 which removed the union size limit for “single dispatch sites”. Also, see Avoiding Vectors of Abstract Types

Topic		Replies	Views
Union splitting vs C++ Performance	22	4417	July 9, 2021
[ANN]: WrappedUnions.jl: Wrap a Union for enhanced type-stability Package Announcements package , announcement , type , type-stability , union	11	824	March 24, 2026
Performance benefits of multiple dispatch Performance question , multidispatch	20	2672	December 19, 2021
Performance drawback with subtyping Performance	34	3472	August 26, 2021
Slow code with Union and Box General Usage	17	399	March 6, 2024

[ANN] ManualDispatch.jl

Related topics