Vector accidentally overwritten in one specialization of /

[daanhb]

Reviewing all occurrences of conversions to the abstract type AbstractArray in LinearAlgebra, I came across this corner-case issue:

julia> using LinearAlgebra

julia> l = lu(rand(3,3)); v = transpose(rand(3));

julia> v
1×3 Transpose{Float64,Array{Float64,1}}:
 0.303153  0.935216  0.228176

julia> v / adjoint(l)
1×3 Adjoint{Float64,Array{Float64,1}}:
 1.78182  -0.065159  0.00787519

julia> v
1×3 Transpose{Float64,Array{Float64,1}}:
 1.78182  -0.065159  0.00787519

The vector v here is accidentally overwritten. The offending line is this one, where in the statement convert(AbstractVector{T}, conj(trA.parent)) neither conj nor the conversion is guaranteed to actually copy the data in all cases. There is a similar issue three lines below it (with equally exotic arguments).

My code review was motivated by JuliaLang/julia#34995 and I'm currently attempting to fix both issues.

[daanhb]

Meanwhile this issue has been fixed:

julia> using LinearAlgebra

julia> l = lu(rand(3,3)); v = transpose(rand(3));

julia> v
1×3 transpose(::Vector{Float64}) with eltype Float64:
 0.14045  0.85708  0.644903

julia> v / adjoint(l)
1×3 adjoint(::Vector{Float64}) with eltype Float64:
 0.128085  1.384  -0.657738

julia> v
1×3 transpose(::Vector{Float64}) with eltype Float64:
 0.14045  0.85708  0.644903

The offending line that said convert(...) now uses copy_oftype in lu.jl, compare this line to the new and improved implementation. It was fixed by JuliaLang/julia#36657.

So I'm closing this issue.