Performance regression in function calls

[KristofferC]

Sorry for the general issue name but im not sure yet what is actually slower.

I was looking into doing matrix multiplication with tuples and stumbled upon a performance regression that might be interesting.

Script:

typealias Mat{M, N, T} NTuple{M, NTuple{N, T}}

# Generates expression for an unrolled matrix multiply
@generated function unrolled_matmult_noinline{T, M, N, K}(A::Mat{M, N, T}, B::Mat{N, K, T})
    ex = Expr(:tuple, [Expr(:tuple, [:(loopdot_noinline(A, B, $i, $j)) for j=1:N]...) for i=1:N]...)
    return ex
end

# Does the contraction on K
@noinline function loopdot_noinline{M, N, K, T}(A::Mat{M,K,T}, B::Mat{K,N,T}, Arow, Bcol)
    s = zero(T)
    for k = 1:K
        s += A[Arow][k] * B[k][Bcol]
    end
    s
end

# Creates a random tuple
@generated function rand_tuple{N, T}(::Type{Mat{N,N,T}})
    return  ex = Expr(:tuple, [Expr(:tuple, [rand(T) for j=1:N]...) for i=1:N]...)
end


function bench_noinline{N}(::Type{Val{N}})
    A = rand_tuple(Mat{N,N, Float64})
    @time for i in 1:10^5
        unrolled_matmult_noinline(A, A)
    end
end

for i in 1:9
    bench_noinline(Val{i})
end

0.4.3:

  0.000263 seconds
  0.002243 seconds
  0.006109 seconds
  0.017329 seconds
  0.039429 seconds
  0.077035 seconds
  0.137714 seconds
  0.231859 seconds
  0.355290 seconds

Master:

  0.000316 seconds
  0.002214 seconds
  0.009517 seconds
  0.028667 seconds
  0.061345 seconds
  0.142360 seconds
  0.392535 seconds
  0.719159 seconds
  1.347999 seconds

Note that the inlined version is about the same on both versions.

[KristofferC]

The FixedSizeArrays.jl package also seems slower on 0.5 so maybe it is something with tuples in general?

function bench_FSA{N}(::Type{Val{N}})
    A = rand(Mat{N,N, Float64})
    @time for i in 1:10^7
        A * A
    end
end

for i in 1:4
    bench_FSA(Val{i})
end

0.4.3:

  0.019101 seconds
  0.028643 seconds
  0.077164 seconds
  0.187649 seconds

Master:

  0.016788 seconds
  0.038135 seconds
  0.114322 seconds
  0.281677 seconds

[KristofferC]

Apparently a dup of #15277