inference failure on array relabeling

[stevengj]

The following function (boiled down from this example seems to have a spurious type-inference failure:

function foo!(u::T,v::T) where {T<:AbstractVector}
    axes(u,1) == axes(v,1) || error("axis mismatch")
    for t = 1:10, i = firstindex(u)+1:lastindex(u)-1
        v[i] = -2u[i] + sum(k -> u[i+k]+u[i-k], (1,))
        u,v = v,u
    end
    return u,v
end

as evinced by a huge number of allocations:

julia> u = rand(100); v = copy(u);

julia> @btime foo!($u,$v);
  345.500 μs (8834 allocations: 168.83 KiB)

If I comment out the u,v = v,u line, which should be a type-stable operation since u and v have the same type, then the number of allocations goes down to zero.

[simeonschaub]

u is captured by a closure and assigned to multiple times, so this is just another instance of #15276 again.