underscores still allowed as static parameter names

[timholy]

In combination with Revise, some code (in ImageFiltering) that was initially written prior to the Julia-0.6 type system overhaul exposed the following inconsistency:

julia> foo(A::AbstractArray{_,N}, flags::NTuple{N,Bool}) where {_,N} = 1
foo (generic function with 1 method)                                                                                                                                                                                                                               
                                                                                                                                                                                                                                                                   
julia> foo(rand(2,2), (true, true))   # just to prove that the method works
1                                                                                                                                                                                                                                                                  
                                                                                                                                                                                                                                                                   
julia> Tuple{AbstractArray{_,N}, NTuple{N,Bool}} where {_,N}   # create the signature type
ERROR: syntax: all-underscore identifier used as rvalue

Obviously, post-0.6 the better way to write this is

julia> foo(A::AbstractArray{<:Any,N}, flags::NTuple{N,Bool}) where N = 1
foo (generic function with 1 method)

julia> Tuple{AbstractArray{<:Any,N}, NTuple{N,Bool}} where N
Tuple{AbstractArray{#s3,N} where #s3,Tuple{Vararg{Bool,N}}} where N

but the inconsistency should be fixed. In principle it seems that either option would be OK (throw an error on the method definition, or allow the Tuple-type to be created), with the obvious caveat that one of these is breaking.

[Keno]

Also while we're at it, I don't like the printing of

julia> Tuple{AbstractArray{<:Any,N}, NTuple{N,Bool}} where N
Tuple{AbstractArray{#s3,N} where #s3,Tuple{Vararg{Bool,N}}} where N

because you can't paste it back into the REPL. I've had to go through more than one time tracking down bugs and delete all the # in the types in my test case.

[JeffBezanson]

The issue here is just that _ is still allowed as a static parameter name, since the uses of _ in signatures are not quite normal rvalues. The safest thing is probably to disallow it, but that could be quite breaking since this seems like something people might actually write.

[ararslan]

See also #29875, which specifically requests being able to use _ in a similar way.

[StefanKarpinski]

And #24990 which proposes a conflicting meaning 😬

[JeffBezanson]

From triage: disallow in 1.x and see what happens.

[StefanKarpinski]

This was just an oversight when _ as an r-value was deprecated. We should try disallowing and see what PkgEval says and make further decisions based on that.