BigFloat QR leads to ERROR: cannot resize array with shared data

[dlfivefifty]

This is a bit surprising:

julia> A = BigFloat.(randn(5,5)); b = BigFloat.(randn(5));

julia> resize!(qr(A) \ b, 3)
ERROR: cannot resize array with shared data
Stacktrace:
 [1] _deleteend!
   @ ./array.jl:893 [inlined]
 [2] resize!(a::Vector{BigFloat}, nl::Int64)
   @ Base ./array.jl:1109
 [3] top-level scope
   @ REPL[9]:1

I tried to debug, the culprit seems to be somewhere in the orthogonal lmul!:

https://github.com/JuliaLang/julia/blob/efad4e3134540c7a45d385b047f145cb0da6d385/stdlib/LinearAlgebra/src/qr.jl#L634

But nothing stands out...

[MikaelSlevinsky]

A workaround is to copy the output

ulia> A = BigFloat.(randn(5,5)); b = BigFloat.(randn(5));

julia> resize!(copy(qr(A) \ b), 3)
3-element Array{BigFloat,1}:
 1.128343225039452563013632216725243690076281918797619929418640609900994969010822
 0.3046898873582031740087773316219913340227916402714182242361382185620764253197802
 0.9309977144595598360049286776484760070368732518629144855630406146909686632942985

I could be wrong, but I think this is because the output vector is shared with a temporary vector in the QR factorization (maybe a Householder reflector?) that doesn't get cleared before its use in the solve phase. Copying the output places a memory barrier between it and a temporary array. (Not saying this is ideal.)

Note that lu works fine (perhaps because it implements elementary row operations in-place, i.e. without any temporary storage.)

julia> resize!(lu(A) \ b, 3)
3-element Array{BigFloat,1}:
 1.128343225039452563013632216725243690076281918797619929418640609900994969010788
 0.3046898873582031740087773316219913340227916402714182242361382185620764253198018
 0.9309977144595598360049286776484760070368732518629144855630406146909686632942726