Avoid several bad show_default invalidations

[timholy]

This fixes 53 quite critical invalidations from loading StaticArrays. At least by our old standards, 53 invalidations would be considered a modest number, but in this case they manage to hit the package-loading code, the docstring code, and the MethodError display code. Since particularly the first two are used by everything, these invalidations are about as bad as they could be.

The invalidation of the package-loading code is due to a rather hilarious chain of events. This method definition
ends up invalidating show_default, which got called because stale_cachefile used an outdated variable name in a @debug macro that ended up grabbing the global mod function. Since show(::IO, ::Function) can call show_default, this ended up invalidating basically all the package-loading code.

The docstring code got invalidated because show(::IO, ::Pair{Symbol,Any}) calls show_default. This PR provides a dedicated path that circumvents show_default.

Finally, an innocuous-seeming function-print in showerror(::IO, ::MethodError) also led to quite a chain of invalidation.

EDIT: as pointed out below by @KristofferC, this fixes #36005. (If master hadn't already fixed it, I haven't checked.)

[timholy]

Just to show the overall consequences: on this branch,

julia> using StaticArrays

julia> @time using Example
  0.002901 seconds (2.90 k allocations: 185.891 KiB)

On Julia 1.5,

julia> using StaticArrays

julia> @time using Example
  0.656540 seconds (1.33 M allocations: 66.060 MiB, 7.79% gc time)

That's gonna make a difference in package loading times.

There are a couple of show_default stragglers:

• we lack a show method for generic IPAddr: we have them for all subtypes but when inference is lost we end up inferring a call to show_default.
• something is calling print(::ProcessException) (we have a showerror method but not a show method)
• of course, show(::IO, ::Function) gets invalidated

For all of these, the trail of backedges truncates (#37046) so it is not easy to figure out where these are coming from.

[KristofferC]

Fixes #36005?

[timholy]

Heck yeah!

julia> @time using Plots
  3.217109 seconds (7.83 M allocations: 544.377 MiB, 4.78% gc time)

julia> @time using Plots
  0.000421 seconds (717 allocations: 38.922 KiB)

That 3.2s load time is pretty awesome (CCing @daschw, @BeastyBlacksmith). For reference, here's what I get on Julia 1.5:

julia> @time using Plots
  8.466600 seconds (18.09 M allocations: 1021.804 MiB, 3.44% gc time)

julia> @time using Plots
  0.685273 seconds (1.43 M allocations: 70.920 MiB, 11.50% gc time)

julia> @time using Plots
  0.000393 seconds (761 allocations: 44.250 KiB)

Both with Plots 1.6.0, FWIW. My guess is that on 1.5 the loading code was being invalidated repeatedly by several different dependencies of Plots (that extra time for the second using is just the most recent of these invalidations). Most of those were probably fixed in other PRs though.

I also pushed a proposed fix for the IPAddr issue described above, although it has the downside of blocking any additional IPAddr subtypes. Such types would surely require other changes to the Sockets library so on balance I think it's a worthy fix, but let's see what others think. (It's also not a big deal either way, it only affects 2 additional invalidations.)

[daschw]

Wow, that's awesome!

[KristofferC]

That 3.2s load time is pretty awesome (CCing @daschw, @BeastyBlacksmith). For reference, here's what I get on Julia 1.5:

FWIW. My guess is that the loading code was being invalidated repeatedly by several different dependencies of Plots (that extra time for the second using is just the most recent of these invalidations).

Out of curiosity, what do you get on master without this PR?

[giordano]

@timholy what version of FFMPEG.jl are you using? How long does it take to load?

[timholy]

@KristofferC, here's master (which of course has a lot of relevant recent work from you, Jeff, Jameson, Elliot, me, and likely others I'm not naming):

julia> @time using Plots
  3.370194 seconds (7.59 M allocations: 528.619 MiB, 6.43% gc time)

julia> @time using Plots
  0.377861 seconds (588.13 k allocations: 33.855 MiB, 4.82% gc time)

julia> @time using Plots
  0.000394 seconds (717 allocations: 38.922 KiB)

So most of the benefit over 1.5 is due to things we've recently added, but this PR adds further progress on the scale of ~0.5s if you're loading more than just Plots.

@giordano, here's what I get on master:

julia> @time using FFMPEG
  0.457856 seconds (994.93 k allocations: 63.036 MiB, 3.36% gc time)

I doubt this PR changes that (loading FFMPEG produces zero invalidations).

[KristofferC]

I doubt this PR changes that (loading FFMPEG produces zero invalidations).

@staticfloat has work in progress that will reduce this time quite significantly though.

[giordano]

What version of FFMPEG.jl is that? Does it use FFMPEG_jll?

[timholy]

(@v1.6) pkg> st FFMPEG
Status `~/.julia/environments/v1.6/Project.toml`
  [c87230d0] FFMPEG v0.3.0