Performance regression on master in views

[KristofferC]

#29795 was backported to the 1.0.2 backport branch and was found to give performance regressions (see report in #29444 (comment)). These also exist on master, @vtjnash.

Reproduced with

julia> function perf_laplace_iter_sub(N)
           u = zeros(N, N)
           u[1,:] .= 1
           Niter = 2^10
           dx2 = dy2 = 0.1*0.1
           u0 = view(u, 2:N-1, 2:N-1)
           u1 = view(u, 1:N-2, 2:N-1)
           u2 = view(u, 3:N,   2:N-1)
           u3 = view(u, 2:N-1, 1:N-2)
           u4 = view(u, 2:N-1, 3:N)
           for i = 1:Niter
               u0 .= ((u1 .+ u2).*dy2 .+ (u3 .+ u4).*dx2) .* (1 / (2*(dx2+dy2)))
           end
           return u
       end
perf_laplace_iter_sub (generic function with 1 method)

julia> using BenchmarkTools

julia> @btime perf_laplace_iter_sub(100);
  5.321 s (157364231 allocations: 4.98 GiB)

On julia 1.0.1 this takes:

julia> @btime perf_laplace_iter_sub(100);
  127.267 ms (12295 allocations: 300.89 MiB)