Avoid constructing an array only to splat it to make a tuple in tuple_tfunc

[oxinabox]

No description provided.

[vtjnash]

Those both still construct it (can't avoid that as it's needed for dispatch), they would just also make a closure and allocate some other objects too.

[oxinabox]

I made a similar case to time in the REPL, not sure that it is completely valid, but it seems like it is much faster to use ntuple.

julia> foo_vecsplat(x) = tuple(Any[x[i] for i in 1:length(x)]...)
foo_vecsplat (generic function with 1 method)

julia> foo_ntuple(x) = ntuple(i->x[i], length(x))
foo_ntuple (generic function with 1 method)

julia> foo_generator(x) = Tuple(x[i] for i in 1:length(x))
foo_generated (generic function with 1 method)

julia> @btime foo_vecsplat([1,2,3,4,5,6,7,8]);
  244.023 ns (3 allocations: 368 bytes)

julia> @btime foo_ntuple([1,2,3,4,5,6,7,8]);
  32.596 ns (2 allocations: 224 bytes)

julia> @btime foo_generator([1,2,3,4,5,6,7,8]);
  286.795 ns (3 allocations: 368 bytes)

It seems like it still holds with @nospecialize

julia> bar_ntuple(@nospecialize x) = ntuple(i->x[i], length(x))
bar_ntuple (generic function with 1 method)

julia> @btime bar_ntuple([1,2,3,4,5,6,7,8]);
  32.676 ns (2 allocations: 224 bytes)

[vtjnash]

That's not quite valid. Using ntuple is a bit faster, but not by quite that much (it's the same implementation as vecsplat, but additionally has optimizations for less than 10 elements):

julia> @btime foo_ntuple($(Any[1,2,3,4,5,6,7,8]));
  157.234 ns (1 allocation: 80 bytes)
julia> @btime foo_vecsplat($(Any[1,2,3,4,5,6,7,8]));
  304.480 ns (2 allocations: 224 bytes)

[oxinabox]

... but additionally has optimizations for less than 10 elements ...

Are those optimisations relevent here?
Trying to familarise myself more with this code

[JeffBezanson]

Tuples are usually small, so that's the case we're likely to hit here. But, this code probably does not run often enough in the compiler to have an impact.

[vtjnash]

Yep. But since you've done the work to identify it, seems sensible to merge.