Refs for BLAS calls causes performance hit for small matrices.

[KristofferC]

With this benchmark:

function mmult(N, n)
    A = rand(n,n)
    B = similar(A)
    @time for i = 1:N
        A_mul_B!(B, A, A)
    end
end

I get the following on master:

julia> mmult(10^6, 6)
 621.537 milliseconds (10000 k allocations: 153 MB, 0.90% gc time)

With JuliaLang/julia@62e5942 reverted:

julia> mmult(10^6, 6)
 208.683 milliseconds

Sort of dup of JuliaLang/julia#11531 except this is in Base. Imho it is not acceptable to double the time for matrix multiplication for smallish matrices.

[pao]

cc @simonbyrne

[andreasnoack]

@vtjnash Is this expected?

[jakebolewski]

I would think that this is due mostly to Array construction being special cased in the system. I noticed a similar performance regression when the special casing of Expr object construction was removed.

[JeffBezanson]

Doesn't it have to do with Refs being constructed, not Arrays?

[vtjnash]

it has to do with Ref objects being actually constructed rather than going through the painfully complex & handling (which does stack allocation).

this should be fixed by cleaning up JuliaLang/julia#8134 so it can be merged

[simonbyrne]

We should probably revert the changes until JuliaLang/julia#8134 is fixed

[ViralBShah]

+1 to reverting.