Base.setindex return type inference quick fix

[nsajko]

Xref #54462

[kimikage]

In the first place, is it appropriate to use ntuple for code which do not return NTuple?
It is misleading, even if it is intentional and follows the documentation.

(I think it is also misleading that the function _setindex exists in Base. Can't we use _setindex_tuple or something like that?)

I think it would be appropriate to specialize only for true NTuple cases, and naively use Tuple() or (itr...,) without the tricky Vararg for the rest cases. Since the setindex creates a new object and checks the bounds, I suspect the optimization effect by specialization is minor.

Of course, I have no objection to this PR (quick fix) per se.

[kimikage]

E.g.

import Base: setindex

function setindex(x::Tuple, v, i::Integer)
    @boundscheck 1 <= i <= length(x) || throw(BoundsError(x, i))
    @inline
    _setindex_tuple(x, v, i)::Tuple
end

function _setindex_tuple(x::NTuple{N, T}, v::T, i::Integer) where {N, T}
    @inline
    return ntuple(j -> ifelse(j == i, v, x[j]), Val{N}())
end

function _setindex_tuple(x::Tuple, v, i::Integer)
    @inline
    return ((ifelse(j == i, v, xj) for (j, xj) in enumerate(x))...,)::Tuple
end

[nsajko]

I've got a better fix locally, a recursive implementation of _setindex, not based on ntuple. It's got way better inference. E.g.:

julia> Core.Compiler.return_type(Base.setindex, Tuple{Tuple{Vararg{Int}},Int,Int})
Tuple{Vararg{Int64}}