Resolve aliasing problems in getindex

[bkamins]

Life is hard. I am not sure how to write a fully generic code that would work on all supported versions of Julia and be fast.

Therefore I propose half-measure that resolves the following problem:

Timing on 1.2.0 release

julia> @time x = PooledArray(1:10^5);
  0.007226 seconds (59 allocations: 6.718 MiB)

julia> @time v = view(x, [1:10^5;]);
  0.000352 seconds (3 allocations: 781.375 KiB)

julia> @time s = zeros(10^5);
  0.000396 seconds (2 allocations: 781.328 KiB)

julia> @time copyto!(s, v);
  2.714555 seconds (298.98 k allocations: 7.614 MiB)

and this is timing with this PR:

julia> @time x = PooledArray(1:10^5);
  0.006874 seconds (59 allocations: 6.718 MiB)

julia> @time v = view(x, [1:10^5;]);
  0.000355 seconds (3 allocations: 781.375 KiB)

julia> @time s = zeros(10^5);
  0.000390 seconds (2 allocations: 781.328 KiB)

julia> @time copyto!(s, v);
  0.000191 seconds

What is the problem - we cannot use view for DataAPI.refarray in general if we want to be fast as if indices are a general AbstractVector then view must do unaliasing and this is extremely slow.

Therefore what I propose is that we use fast path for PooledArray or its view only if the first index in getindex is AbstractVector (it should cover 99.99% of use-cases I can think of - especially given that PooledArray did not work correctly for more than 1 dimension anyway in the past an no-one reported it before). In other cases we fall-back to mechanics of Julia Base that avoid allocations (and rewriting them in PooledArrays.jl would be an overkill especially that internals changed since Julia 1.0).

[codecov]

Codecov Report

Merging #60 (b523437) into main (040053c) will increase coverage by 0.06%.
The diff coverage is 90.90%.

@@            Coverage Diff             @@
##             main      #60      +/-   ##
==========================================
+ Coverage   88.04%   88.10%   +0.06%     
==========================================
  Files           1        1              
  Lines         276      269       -7     
==========================================
- Hits          243      237       -6     
+ Misses         33       32       -1     

Impacted Files	Coverage Δ
src/PooledArrays.jl	88.10% <90.90%> (+0.06%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 040053c...d1983de. Read the comment docs.

[quinnj]

LGTM

[bkamins]

I need to resolve one more thing before merging. I will report.

[bkamins]

I incorrectly used extrema for empty PooledArray - a case that was not tested.

[bkamins]

@nalimilan - I will merge it to fix #61. Let us then later discuss if we can have a cleaner code please.