TTFP regression since 1.7

[jakobnissen]

I discovered a 25% regression in compilation time from 1.7.2 to master (0b48b91) when running a workflow for some work code. My work code is really hard to boil down to a MWE because it makes use of packages from a local registry, reference data sets, and operates on specific files in a workflow.

But at least some of the regression is due to a regression in Plots.jl.

The following code:

@time begin
    using Plots
    Plots.savefig(plot([1,2,3,4], [1,2,3,4]), "path.pdf")
end

Has the following timings:

• 1.7.2: 10.7 seconds
• Master: 14.3 seconds

But wait, Plots.jl makes use of Base.Experimental.@max_methods, which is only available in 1.8. What happens if we simply comment that out?

• Master (with default max_methods): 19.5 seconds.

[KristofferC]

Looks like JuliaPlots/Plots.jl#4079.

[aviatesk]

Maybe is this a duplicate of #43157 (comment)?

[timholy]

If you launch Julia with --compiled-modules=no you get 54.6s on 1.7 and 44.5s on 1.8. Because of lines like this, that may be a reasonably-fair comparison. And of course that's not even including #43990.

[jakobnissen]

Does that mean it's nearly certainly not a true regression, but instead, as Kristoffer said, due to Plots not precompiling on master? If so, this issue can happily be closed