compile times degraded on using Hecke

[martinra]

I started to investigate some of my compile times, and tracked down the following issue to Hecke.

Not using Hecke

using Nemo
fm = FreeModule(QQ, 5);
vs = Vector{fmpq}[];
@time sub(fm, map(fm,vs));
0.410914 seconds (1.64 M allocations: 84.424 MiB, 5.92% gc time)

Using Hecke

using Nemo, Hecke
fm = FreeModule(QQ, 5);
vs = Vector{fmpq}[];
julia> @time sub(fm, map(fm,vs));
178.114302 seconds (488.43 M allocations: 23.518 GiB, 3.03% gc time)

I'm not sufficiently deep into the code base to find the source, but I can provide some experience from a similar phenomenon that I discovered in my own code. The compile time exploded in the same way. I had a function of the shape

function fct(m::MatElem{R}) where R
  rref!(m)
end

which caused Julia type interference to almost break down. My fix was

function fct(m::MatElem{R}) where R
  if R == fmpq
    rref!(m::fmpq_mat)
  else
    rref!(m)
  end
end

I believe I didn't have these problems in Julia 1.4, but I haven't double checked this. In any case, the work around is not great, but 180s is neither.

[thofma]

Thanks for the report. Which julia version is this?

P.S.: The FreeModules provided by AbstractAlgebra are inherently type unstable and prone to problems with inference/compilation (try having a look at @code_warntype). But this increase in the compile time should not happen.

[martinra]

This is Julia 1.5.3 on Fedora. But I don't think they modify upstream.

I didn't know about @code_warntype. This seems very use. Thank you!

[thofma]

Thanks for the version.

Be careful, @code_warntype is very addictive. In a similar vein, you might find https://github.com/JuliaDebug/Cthulhu.jl useful. We use it all the time to track down type problems.

[thofma]

The inference fix has been merged in AA. I will close this for now @martinra. Feel free to open another issue if you encounter this again. These slowdowns hint at some non-optimal code on our side.