Stream: helpdesk (published)

Topic: Make a struct act like a Tuple under iteration


view this post on Zulip Mason Protter (Jan 21 2021 at 01:31):

Hey, so suppose I have a struct

struct Foo
    a::Int
    b::Float64
end

and I basically want this thing to behave like a Tuple under iteration. e.g. I'd like this to 'just work' and carry no runtime performance overhead.

f((a, b)) = a + b

What's a clean way to do this right?

view this post on Zulip Mason Protter (Jan 21 2021 at 01:37):

If I do the naive thing, it's very slow:

Base.iterate(f::Foo, s=1) = s <= length(fieldnames(Foo)) ? (getfield(f, s), s+1) : nothing

let foo = Ref(Foo(1,2))
    @btime f($foo[])
end

#+RESULTS:
   802.830 ns (4 allocations: 112 bytes)
 3.0

I feel there must be some clean way to basically say "iterate this thing as if it were a Tuple"

view this post on Zulip Mason Protter (Jan 21 2021 at 01:55):

Okay, so if I don't define the above iterate method and instead define

Base.indexed_iterate(f::Foo, i::Int, state=1) = (Base.@_inline_meta; (getfield(f, i), i+1))

everything is fast. Good to know I suppose.

view this post on Zulip Simeon Schaub (Jan 21 2021 at 06:53):

You should probably still define iterate as well though, since indexed_iterate is just an implementation detail and is always assumed to return the same results as regular iteration

view this post on Zulip Pablo Zubieta (Jan 21 2021 at 16:52):

Still slower that indexed_iterate, but better that the OP

@inline Base.iterate(f::Foo, s = 1) = s  fieldcount(Foo) ? (getfield(f, s), s + 1) : nothing

view this post on Zulip Fredrik Bagge Carlson (Jan 21 2021 at 17:18):

Some LinearAlgebra types does this, I thing QR does, you can perhaps see how they do it?

view this post on Zulip Mason Protter (Jan 21 2021 at 17:20):

Hm, it seems that @Pablo Zubieta's method is just as fast as indexed_iterate for me on 1.6.

view this post on Zulip Mason Protter (Jan 21 2021 at 17:21):

julia> struct Foo
           a::Int
           b::Float64
       end

julia> f((a, b)) = a + b
f (generic function with 1 method)

julia> @inline Base.iterate(f::Foo, s = 1) = s  fieldcount(Foo) ? (getfield(f, s), s + 1) : nothing

julia> let foo = Ref(Foo(1,2))
           @btime f($foo[])
       end
  1.859 ns (0 allocations: 0 bytes)
3.0

and

julia> struct Foo
           a::Int
           b::Float64
       end

julia> f((a, b)) = a + b
f (generic function with 1 method)

julia> Base.indexed_iterate(f::Foo, i::Int, state=1) = (Base.@_inline_meta; (getfield(f, i), i+1))

julia> let foo = Ref(Foo(1,2))
           @btime f($foo[])
       end
  1.849 ns (0 allocations: 0 bytes)
3.0

view this post on Zulip Pablo Zubieta (Jan 21 2021 at 18:07):

Oh. I was trying on 1.5.3, great to see that the compiler is also better here in 1.6.

view this post on Zulip Eric Hanson (Jan 23 2021 at 01:10):

I think with https://github.com/JuliaLang/julia/pull/39285 you wouldn’t even need an iterate method, just a semicolon: f((;a, b)) = a + b

view this post on Zulip Mason Protter (Jan 23 2021 at 01:48):

Wow, that’s very neat. Kinda niche though, since you need to bind the variables to the same name as the property names.

view this post on Zulip Simeon Schaub (Jan 23 2021 at 09:51):

There was some discussion in https://github.com/JuliaLang/julia/issues/28579 about having some more advanced matching as well. Wouldn't be too hard to add, but would need to come to a decission on how exactly that should work.


Last updated: Nov 22 2024 at 04:41 UTC