Regression: stride errors on 0-dim array

[mkitti]

Julia 1.7.2:

julia> stride( fill(1), 1)
1

#44027 modified stride.

master-branch

julia> stride( fill(1) , 1)
ERROR: ArgumentError: reducing with add_sum over an empty collection of element type Union{} is not allowed.
You may be able to prevent this error by supplying an `init` value to the reducer.
Stacktrace:
  [1] _empty_reduce_error(f::Any, T::Type)
    @ Base ./reduce.jl:307
  [2] reduce_empty(#unused#::typeof(Base.add_sum), #unused#::Core.TypeofBottom)
    @ Base ./reduce.jl:338
  [3] reduce_empty(op::Base.BottomRF{typeof(Base.add_sum)}, #unused#::Type{Union{}})
    @ Base ./reduce.jl:347
  [4] reduce_empty_iter
    @ ./reduce.jl:371 [inlined]
  [5] reduce_empty_iter
    @ ./reduce.jl:370 [inlined]
  [6] foldl_impl(op::Base.BottomRF{typeof(Base.add_sum)}, nt::Base._InitialValue, itr::Tuple{})
    @ Base ./reduce.jl:49
  [7] mapfoldl_impl(f::typeof(identity), op::typeof(Base.add_sum), nt::Base._InitialValue, itr::Tuple{})
    @ Base ./reduce.jl:44
  [8] mapfoldl(f::Function, op::Function, itr::Tuple{}; init::Base._InitialValue)
    @ Base ./reduce.jl:162
  [9] mapfoldl
    @ ./reduce.jl:162 [inlined]
 [10] #mapreduce#264
    @ ./reduce.jl:294 [inlined]
 [11] mapreduce(f::Function, op::Function, itr::Tuple{})
    @ Base ./reduce.jl:294
 [12] sum(f::Function, a::Tuple{}; kw::Base.Pairs{Symbol, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
    @ Base ./reduce.jl:520
 [13] sum(f::Function, a::Tuple{})
    @ Base ./reduce.jl:520
 [14] sum(a::Tuple{}; kw::Base.Pairs{Symbol, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
    @ Base ./reduce.jl:549
 [15] sum(a::Tuple{})
    @ Base ./reduce.jl:549
 [16] stride(A::Array{Int64, 0}, k::Int64)
    @ Main ./REPL[7]:4
 [17] top-level scope
    @ REPL[9]:1

[N5N3]

We have a daily PkgEval (https://s3.amazonaws.com/julialang-reports/nanosoldier/pkgeval/by_date/2022-02/07/report.html).
Some other PKG seems also failed because of this PR.
I'll take a look to decide which value to use.

Looks like Strided.jl prefer length(A).
(Although I think test the exact value of stride(zeros(60,60), 3) is somehow fragile)

[mkitti]

My question is if we can always assume that length exists, but if this was used before then just use that.

[mkitti]

Please add a test while fixing this.

[N5N3]

Looks like length is the first choice before sum(st .* size(A)). But @mbauman is not happy with it. (#35929 (comment))

I'll recover the original behavior before #44027