circshift is not performant

[JeffreySarnoff]

using BenchmarkTools
v = [i for i=1:4];
W = [v v v v];

bench_det = @benchmarkable( det(W) ) ;tune!(bench_det); run(bench_det)
BenchmarkTools.Trial: 
  samples:          10000
  evals/sample:     195
  time tolerance:   5.00%
  memory tolerance: 1.00%
  memory estimate:  848.00 bytes
  allocs estimate:  12
  minimum time:     484.00 ns (0.00% GC)
  median time:      506.00 ns (0.00% GC)
  mean time:        683.73 ns (11.35% GC)
  maximum time:     27.55 μs (0.00% GC)

bench_cshift = @benchmarkable( circshift(v,2) ); tune!(bench_cshift); run(bench_cshift)
BenchmarkTools.Trial: 
  samples:          10000
  evals/sample:     194
  time tolerance:   5.00%
  memory tolerance: 1.00%
  memory estimate:  640.00 bytes
  allocs estimate:  10
  minimum time:     493.00 ns (0.00% GC)
  median time:      519.00 ns (0.00% GC)
  mean time:        632.77 ns (9.90% GC)
  maximum time:     27.16 μs (0.00% GC)

uncovered by Michael Prange: https://groups.google.com/forum/#!topic/julia-users/yOFYiNPtWWw

issue submitted at the suggestion of @StefanKarpinski

[stevengj]

Here is a more appropriate comparison. The 1d case is the easiest to make fast:

function mycircshift(a::AbstractVector, shift::Integer)
       b = similar(a)
       n = length(a)
       n == 0 && return b
       s = mod(shift, n)
       copy!(b, 1, a, n - s + 1, s)
       copy!(b, s + 1, a, 1, n - s)
end

x = rand(1000);
@time circshift(x, 3);
@time mycircshift(x, 3);

gives (after a couple repetitions):

  0.000042 seconds (14 allocations: 32.094 KB)
  0.000008 seconds (5 allocations: 8.094 KB)

i.e. almost 6 times faster, and 1/4 the memory.

See also #16032.