Traits for julia: dispatch on whatever you want using where syntax
69 Stars
Updated Last
6 Months Ago
Started In
July 2020


Stable Dev Build Status Coverage

Welcome to WhereTraits.jl. This package exports one powerful macro @traits with which you can extend Julia's where syntax in order to support traits definitions.

In addition, WhereTraits comes with a standardized way how to resolve ambiguities among traits, by defining an order among the traits with @traits_order.

Installation & Import

Install by running

using Pkg
pkg"add WhereTraits"

Then use this package by loading

using WhereTraits

which brings @traits into your namespace, and in addition also @traits_order for resolving ambiguities.

Usage @traits

@traits supports the following three extensions to Julia's where-syntax:

  • dispatch on functions returning Bool
@traits f(a) where {isodd(a)} = (a+1)/2
@traits f(a) where {!isodd(a)} = a/2
f(4) # 2.0
f(5) # 3.0
  • dispatch on functions returning anything
@traits g(a) where {Base.IteratorSize(a)::Base.HasShape} = 43
@traits g(a) = 1
g([1,2,3]) # 43
g(Iterators.repeated(1)) # 1
  • dispatch on upperbounds on functions returning Types
@traits h(a) where {eltype(a) <: Number} = true
@traits h(a) = false
h([1.0]) # true
h([""]) # false

And all this works with arbitrary many where expressions and creates optimal code where possible via standard Julia compiler.

Usage @traits_order - Resolving Ambiguities

WhereTraits comes with special support for resolving ambiguities among traits dispatch.

Under the hood @traits uses normal function dispatch to achieve the speed and flexibility, however, julia function dispatch can lead to ambiguities. With traits these can easily happen if someone defines @traits for the same standard dispatch but using different traits. Let's take a look how this looks like

using WhereTraits
# let's say someone defined this version
@traits conflict(a) where {eltype(a) <: Number} = "eltype wins"
# and another this one
@traits conflict(a) where {Base.IteratorSize(a) :: Base.HasShape} = "IteratorSize wins"

You can still use these traits definition as long as there is no ambiguity.

julia> conflict(Iterators.countfrom(42))
"eltype wins"

julia> conflict(["hello", "world"])
"IteratorSize wins"

If you use something ambiguous, e.g. a Vector of Numbers, you get a proper ambiguity error, stating what you can do in order to fix it.

julia> conflict([1,2,3,4])
ERROR: Disambiguity found. Please specify an ordering between traits, like the following.

    @traits_order (Main).conflict(a1::T1) where T1 begin


What is needed in order to resolve the ambiguity is an order between the traits. This can be defined with @traits_order, which takes the respective function signature followed by a begin-end block of ordered traits (most dominant one should be at the top).

Hence just executing the example @traits_order will make eltype be the winning trait.

@traits_order (Main).conflict(a1::T1) where T1 begin

Let's take a look that everything is resolved

julia> conflict([1,2,3,4])
"eltype wins"

Alternatively to the @traits_order you can always define your own custom resolution

@traits function conflict(a) where {eltype(a) <: Number, Base.IteratorSize(a) :: Base.HasShape}
    "custom implementation"

which immediately will resolve correctly

julia> conflict([1,2,3,4])
"custom implementation"

For more details, take a look at the documentation.


Optimal Code

Warning: While the dispatch works for dynamic functions, it will only be able to create optimal code if your traits function supports proper type-inference. E.g. you can use Base.isempty, however type-inference cannot see whether it will return true or false by static inspection. Hence it will use slower dynamic code.

Keyword arguments

Keyword arguments are at the moment not support for WhereTraits dispatch. They are just passed through.

Symbol Level

The extended where syntax is currently implemented on symbol level, which is why traits functions like Base.IteratorSize and the non-qualified IteratorSize (assuming you imported import Base: IteratorSize) are treated as two different functions, despite being the same. So for now try to only use the one style or the other.

Top Level Only

Currently only top-level functions are supported, as the syntax stores and needs information about previous function definitions, which it stores globally. If macros would get informed about whether they are defined within another function, WhereTraits could also support innerfunctions.

Test package

The @traits macro currently does not work well within the Test.@testset macro. Usually you won't encounter this, as standard dispatch is probably enough for your tests.

Nevertheless there is a workaround. WhereTraits.jl exports a @traits_test macro variant which works better, but still might have cases where it fails. This needs to be investigated further, and maybe needs a change on Test.@testset.

Other traits packages

There are many different attempts to add traits to Julia. Everyone puts a different emphasis on different aspects of traits interfaces.

  • SimpleTraits.jl
  • BinaryTraits.jl
  • CanonicalTraits.jl

Used By Packages