Iterate and foldl for humans™
Author JuliaFolds
21 Stars
Updated Last
2 Years Ago
Started In
April 2020

GeneratorsX: iterate and foldl for humans™

GitHub Actions

NOTE: The mechanism for defining foldl is factored out as a more simple and robust package FGenerators.jl.

GeneratorsX.jl is a package for defining iterate and foldl with a single easy-to-read source code. An example for creating an ad-hoc iterable:

julia> using GeneratorsX

julia> @generator function generate123()
           @yield 1
           @yield 2
           @yield 3

julia> collect(generate123())
3-element Array{Int64,1}:

It is also possible to use it to define the iteration protocols for existing type:

julia> struct Count

julia> Base.length(itr::Count) = max(0, itr.stop - itr.start + 1);

julia> Base.eltype(::Type{<:Count}) = Int;

julia> @generator function (itr::Count)
           i = itr.start
           i > itr.stop && return
           while true
               @yield i
               i == itr.stop && break
               i += 1

julia> collect(Count(0, 2))
3-element Array{Int64,1}:

GeneratorsX.jl uses IRTools.jl to define Julia's iterate-based iteration protocol. It also derives foldl-based iteration protocol for Transducers.jl using a simple AST transformation.


Defining iterate for a collection is hard because the programmer has to come up with an adequate state machine and code it carefully. Both of these processes are hard. Furthermore, the julia and LLVM compilers do not produce optimal machine code from the loop involving iterate. More importantly, it is hard for the programmer to directly control what would happen in the end result by using iterate since there are many complex transformations from their mental of the collection and the final machine code.

GeneratorsX.jl solves the first problem by providing a syntax sugar for defining iterate. Since this can use arbitrary Julia control flow constructs, the programmer can write down what they mean by using the natural Julia syntax.

The second problem (sub-optimal performance) is solved by generating foldl from the same syntax sugar that generates iterate. Since the syntax sugar used by GeneratorsX.jl directly translates to the foldl definition, it can be optimized much easily by the compiler and it is much easier for the programmer to control the performance characteristics. This is vital for defining fast iteration over blocked/nested data structures as well as collections with heterogeneously typed elements.

However, this foldl-based solution applied alone without generating iterate would have created the third problem: zip can be implemented by iterate but not with foldl. More in general, the new collection wouldn't work with iterate-based existing code. This is why GeneratorsX.jl defines iterate and foldl from the same expression.


GeneratorsX.jl is still a proof-of-concept. As of writing, it works only with Julia 1.5 beta due to IRTools.jl#55. Furthermore, the performance is likely to be awful when consuming the collection without foldl-based frameworks such as Transducers.jl and FLoops.jl.

See also

  • Continuables.jl takes an approach very similar to foldl portion of GeneratorsX.jl. An important difference is that it uses foreach-like function instead of foldl as the basic building block of the iterations. Consequently, it relies on Ref for constructing stateful accumulation. This approach can introduce performance problems if the compiler cannot elide the heap-allocations of the state and it is not applicable for type-changing state. Another difference to GeneratorsX.jl is that Continuables.jl does not provide iterate hence supporting zip is impossible (without extending the compiler). On the other hand, it also has very similar mechanisms to Transducers.jl version of foldl. For example, it uses an approach similar to InitialValues.jl to implement robust initial value handling. See also its README which includes benchmarks and discussions, especially in contrasts with ResumableFunctions.jl and Channel.
  • ResumableFunctions.jl can be used to create more flexible full-blown coroutine. However, since its implementation is based on mutation, it's not the best choice for performance, especially for type-changing state. The mutation-based mechanism also does not play nicely with parallelism.
  • PyGen is a Python style generator that has a similar syntax. However, its implementation is based on Channel and thus not adequate for high-performance iteration.
  • Transducers & Effects – Mike Innes: Exploration of a similar idea by the author of IRTools.jl. See also a discussion in Discourse: Comments on "Transducers & Effects – Mike Innes" - Internals & Design / Internals - JuliaLang