Looking a bit into why <a class="issue-link js-issue-link" data-error-text="Failed to

Yes, <a class="issue-link js-issue-link" data-error-text="Failed to load title" data-i

I guss a question is, is it reasonable for <div class="highlight highlight-source-

The AllocOptPass is potentially slow (and runs too often) in some cases about julia HOT 3 OPEN

KristofferC commented on June 20, 2024 1

The AllocOptPass is potentially slow (and runs too often) in some cases

from julia.

Comments (3)

maleadt commented on June 20, 2024

FWIW, the code generated by LinearAlgebra.exp is pretty large:

julia> write("file.ll", sprint(io->code_llvm(io, LinearAlgebra.exp!, Tuple{Array{Float64, 2}};
             optimize=false, debuginfo=:none, dump_module=true)))

shell> ls -lah *.ll
-rw-r--r-- 1 tim tim 529K May 21 13:51 1.10.ll
-rw-r--r-- 1 tim tim 1.7M May 21 13:54 1.11.ll
-rw-r--r-- 1 tim tim 1.8M May 21 13:53 1.12-broadcast.ll
-rw-r--r-- 1 tim tim 548K May 21 13:53 1.12-loop.ll

from julia.

KristofferC commented on June 20, 2024

Yes, #43200 makes it better:

-rw-r--r--  1 kristoffercarlsson  staff   1.8M May 21 14:44 exp_1.11.ll
-rw-r--r--  1 kristoffercarlsson  staff   491K May 21 14:36 exp_1.11_noinline.ll

but it doesn't seem to impact compilation time much:

# current
julia> @time @eval LinearAlgebra.exp!(rand(2,2))
  1.988244 seconds (8.59 M allocations: 422.911 MiB, 4.09% gc time, 99.49% compilation time)

# no @inline
julia> @time @eval LinearAlgebra.exp!(rand(2,2))
  2.040852 seconds (7.57 M allocations: 370.544 MiB, 3.47% gc time, 99.48% compilation time)

from julia.

KristofferC commented on June 20, 2024

I guss a question is, is it reasonable for

function f(A::StridedMatrix{T}) where T
    CC = T[64764752532480000.,32382376266240000.,7771770303897600.,
           1187353796428800.,  129060195264000.,  10559470521600.,
            670442572800.,      33522128640.,      1323241920.,
            40840800.,           960960.,           16380.,
            182.,                1.]

    tmp1, tmp2 = similar(A), similar(A)

    tmp1 .= CC[14].*A .+ CC[12].*A .+ CC[10].*A
    tmp2 .= CC[8].*A .+ CC[6].*A .+ CC[4].*A

    tmp1 .= CC[13].*A .+ CC[11].*A .+ CC[9].*A
    tmp2 .= CC[7].*A .+ CC[5].*A .+ CC[3].*A

    return tmp1, tmp2
end

write("file.ll", sprint(io->code_llvm(io, f, Tuple{Array{Float64, 2}};
             optimize=false, debuginfo=:none, dump_module=true)))