ReinterpretArray and IndexStyle

[timholy]

Cartesian indexing has some overhead:

julia> a = rand(3,3,3,3,3,3);

julia> v1 = view(a,:,:,:,:,:,:);

julia> v2 = view(a,1:3,:,:,:,:,:);

julia> IndexStyle(v1)
IndexLinear()

julia> IndexStyle(v2)
IndexCartesian()

julia> using BenchmarkTools

julia> @btime fill!(v1, 0);
  79.435 ns (0 allocations: 0 bytes)

julia> @btime fill!(v2, 0);
  1.813 μs (0 allocations: 0 bytes)

This is why we set IndexStyle(a) == IndexLinear() whenever possible. This PR turns on IndexLinear() for appropriate (I hope!) ReinterpretArrays. Demo:

julia> r = reinterpret(Int64, a);

julia> @btime fill!(r, 0);       # on current master
  34.229 μs (0 allocations: 0 bytes)

julia> Revise.track(Base)        # this PR

julia> @btime fill!(r, 0);
  83.182 ns (0 allocations: 0 bytes)

That's a 400× improvement. Some of this seems likely to be due to the performance optimizations planned in #27213, but as the example above illustrates, I suspect we'd be seeing some substantial performance gains even if all the LLVM optimizations had been applied.

[Keno]

I think we should be able to turn on IndexLinear for any array as long as the parent is IndexLinear.

[timholy]

Yeah, makes sense. Try this one.