`dot` could be faster for `SArray`s

[simeonschaub]

This just came up on discourse:

julia> a1 = rand(9, 9)
9×9 Array{Float64,2}:

julia> a2 = rand(9, 9)
9×9 Array{Float64,2}:

julia> s1 = SArray{Tuple{9,9}}(a1)
9×9 SArray{Tuple{9,9},Float64,2,81}:

julia> s2 = SArray{Tuple{9,9}}(a2)
9×9 SArray{Tuple{9,9},Float64,2,81}:

julia> @btime dot($a1, $a2)
  22.762 ns (0 allocations: 0 bytes)
18.782035065611815

julia> @btime dot($s1, $s2)
  70.213 ns (0 allocations: 0 bytes)
18.782035065611826

julia> @btime dot($(vec(a1)), $(vec(a2)))
  23.527 ns (0 allocations: 0 bytes)
18.782035065611815

julia> @btime dot($(vec(s1)), $(vec(s2)))
  15.570 ns (0 allocations: 0 bytes)
18.782035065611815

Is there any reason that dot is special-cased for multi-dimensional SArrays and not just implemented as dot(vec(s1), vec(s2))?

[tkoolen]

Is there any reason that dot is special-cased for multi-dimensional SArrays and not just implemented as dot(vec(s1), vec(s2))?

The problem is in fact that it's not special-cased at all (while dot for StaticVectors is). It falls back to the generic code in the LinearAlgebra stdlib. A PR would be very welcome.