properly handle colon in offset reshape

[johnnychen94]

As a placeholder for a slice, Colon at the k-th position also preserves the offset of the axes(A, k). The default offset for k>=ndims(A) is assumed to be 0.

The implementation is based on https://github.com/JuliaLang/julia/blob/d29126a43ee289fc5ab8fcb3dc0e03f514605950/base/reshapedarray.jl#L119-L137

closes #220

cc: @jishnub

[codecov]

Codecov Report

Merging #228 (bf24c31) into master (b74ae55) will decrease coverage by 0.14%.
The diff coverage is 96.55%.

@@            Coverage Diff             @@
##           master     #228      +/-   ##
==========================================
- Coverage   94.40%   94.26%   -0.15%     
==========================================
  Files           5        5              
  Lines         322      349      +27     
==========================================
+ Hits          304      329      +25     
- Misses         18       20       +2     

Impacted Files	Coverage Δ
src/utils.jl	96.15% <96.00%> (-3.85%)	⬇️
src/OffsetArrays.jl	98.17% <100.00%> (+0.01%)	⬆️

---

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 b74ae55...bf24c31. Read the comment docs.

[johnnychen94]

This is to let utils.jl know the OffsetAxisKnownLength and OffsetAxis.

[jishnub]

Since this changes the axes of the reshaped array, this should perhaps be treated as a breaking change?

[johnnychen94]

Since we never successfully supported Colon in reshape before (had we?), I would say this is just a new feature. A minor version should be sufficient.

[jishnub]

This works currently on master:

julia> A = zeros(1:2, 3:4)
2×2 OffsetArray(::Matrix{Float64}, 1:2, 3:4) with eltype Float64 with indices 1:2×3:4:
 0.0  0.0
 0.0  0.0

julia> reshape(A, 2:3, :)
2×2 OffsetArray(::Matrix{Float64}, 2:3, 1:2) with eltype Float64 with indices 2:3×1:2:
 0.0  0.0
 0.0  0.0

After this PR:

julia> reshape(A, 2:3, :)
2×2 OffsetArray(::Matrix{Float64}, 2:3, 3:4) with eltype Float64 with indices 2:3×3:4:
 0.0  0.0
 0.0  0.0

[johnnychen94]

This is equivalent to

    return OffsetArray(AR, map(_offset, axes(AR), _offset_reshape_uncolon(A, inds)))

By removing the map part, it calls another constructor.

OffsetArrays.jl/src/OffsetArrays.jl

Lines 210 to 219 in b74ae55

	# convert ranges to offsets
	@eval @inline function $FT(A::AbstractArray, inds::Tuple{AbstractUnitRange,Vararg{AbstractUnitRange}})
	_checkindices(A, inds, "indices")
	# Performance gain by wrapping the error in a function: see https://github.com/JuliaLang/julia/issues/37558
	throw_dimerr(lA, lI) = throw(DimensionMismatch("supplied axes do not agree with the size of the array (got size $lA for the array and $lI for the indices"))
	lA = size(A)
	lI = map(length, inds)
	lA == lI || throw_dimerr(lA, lI)
	$FT(A, map(_offset, axes(A), inds))
	end

[johnnychen94]

Ahhhh... Didn't realize that it worked before. I think we need triage here on which solution we should take: #220 or this. @mbauman @timholy

[jishnub]

Perhaps some of the other methods need to be altered too to get this to be consistent:

julia> reshape(OffsetArray(3:4, 2), :) # preserves offset
3:4 with indices 3:4

julia> reshape(OffsetArray(3:4, 2), :, 1) # does not preserve offset
2×1 reshape(::UnitRange{Int64}, 2, 1) with eltype Int64:
 3
 4