further latency reductions

[ranocha]

Additionally, this tests whether renaming master to main worked as expected.

[sloede]

I think it would be good to document the use of @nospecialize somewhere in the docs: Why it's there, what it does, and when/how to use it yourself. Maybe in the performance section?

Also, did you time how much this saves?

[ranocha]

Good suggestion! I added some explanation to the docs.

Here are some timings I get from running

julia16 --threads=1 --check-bounds=no -e '@time using Trixi; @time trixi_include("examples/2d/elixir_advection_amr.jl"); @time trixi_include("examples/2d/elixir_euler_ec.jl"); @time trixi_include("examples/2d/elixir_mhd_blast_wave.jl", tspan=(0.0, 1.0e-3))'

three times.

	main	This PR
using Trixi	3.4 ± 0.3 s	3.4 ± 0.3 s
Advection	29.2 ± 0.3 s	28.6 ± 0.1 s
Euler	6.6 ± 0.1 s	6.4 ± 0.2 s
MHD	13.4 ± 0.1 s	13.1 ± 0.1 s

[sloede]

Not trying to be unpopular, but I am wondering if this optimization is really worth it. As far as I can tell, it's at best a ~3% improvement in startup time, while the code noise grows considerably with these @nospecialize scattered around. My initial feeling is that effort-vs-impact is not really that good... What do you think, @trixi-framework/principal-developers?

[ranocha]

I don't insist on merging this as it is. Just another data point: Precompiling Trixi on main takes 23.8 s while this PR reduces that to 20.5 s on my system (measured using julia16 --threads=1 --check-bounds=no -e '@time using Trixi'). Since some of you complained about the time it takes to precompile Trixi after modifying something, you might want to consider this, too.

Alternatively, we can also just delete the @nospecialize statements. Nevertheless, some of the additional/changed precompile statement should still be considered (since the output format changed after they were added).

[sloede]

Hm. Yeah, I also noticed the slowdown in precompilation. The above improvement sounds better, but I am still not sure if it's worth it. I'll try to get a hold of the others today.

[ranocha]

I will move @nospecialize to the function body comments like

@nospecialize stuff # reduce precompilation time

[ranocha]

Could you please review this again, @sloede?