Forward aliasing check for ReshapedArrays to the parent

[jishnub]

SubArrays have a specialized aliasing check that uses both the parent and the indices to detect potential overlaps. However, wrappers of SubArrays use only the dataids, which is usually determined only by the parent, ignoring the indices. This PR extends the aliasing check to reshaped subarrays as well, by forwarding the check to the parent subarray. This makes the following not allocate:

On master:

julia> C = rand(2000,2000);

julia> @btime $C[2:end, :] .= permutedims(@view $C[1, :]);
  2.816 ms (3 allocations: 15.70 KiB)

This PR

julia> @btime $C[2:end, :] .= permutedims(@view $C[1, :]);
  2.882 ms (0 allocations: 0 bytes)

[LilithHafner]

That benchmark is not particularly convincing because in this case it is actually faster to make an unaliased copy of the RHS because that makes the copy faster. On my computer the PR is a 2x runtime regression in that case even though it eliminates the allocations.

This benchmark is a bit more convincing:

julia> C = rand(2000,2000);

julia> @b reshape(@view($C[:, 1:1000]), 2_000_000) .= reshape(@view($C[:, 1001:2000]), 2_000_000)
1.149 ms (3 allocs: 15.259 MiB)

julia> @eval Base begin
       mightalias(A::ReshapedArray, B::ReshapedArray) = mightalias(parent(A), parent(B))
       # special handling for reshaped SubArrays that dispatches to the subarray aliasing check
       mightalias(A::ReshapedArray, B::SubArray) = mightalias(parent(A), B)
       mightalias(A::SubArray, B::ReshapedArray) = mightalias(A, parent(B))
       end
mightalias (generic function with 6 methods)

julia> @b reshape(@view($C[:, 1:1000]), 2_000_000) .= reshape(@view($C[:, 1001:2000]), 2_000_000)
374.877 μs

---

My conclusion here is that on some architectures and use-cases it makes sense to copy even when unnecessary. This does not mean we should keep lower quality mightalias checking. LGTM.

[jishnub]

I agree, and changing a non-copying version to a copying one is easy at the call site.