Add builtin derive benchmarks

[rylev]

tl;dr Using #[derive] for built in traits incurs significant overhead at scale

I've created a benchmark which has 5,000 structs with one i32 field. I then #[derive(Copy, Clone, Default, Eq, PartialEq, Debug)] for all the structs. This builds using the current nightly compiler (f2bbdd0a3 2020-11-04) in ~5 seconds. When I manually implement these traits, the code compiles in ~2.5 seconds.

This issue is being discussed on Zulip.

[Mark-Simulacrum]

It may make sense to delete the Copy derive (at least in say half the cases); Copy as a derive makes the Clone impl much simpler and likely much cheaper too.

[joshtriplett]

for the sake of making it easier to see optimizations or pessimizations that affect a series of derives on the same type, could you:

1. Add Ord, PartialOrd, and Hash, which rounds out the complete set of derivable traits.
2. Change half of these to write the derives separately (#[derive(Debug)] #[derive(Default)] #[derive(PartialEq)] ...)?

(This is in addition to @Mark-Simulacrum's suggestion.)

[rylev]

@joshtriplett rustsmt collapses separate derives into one line. Obviously not everyone uses rustfmt, but maybe this makes it a bit less compelling to test that case? Also, it does not seem to have any performance impact to keep them separate.

By the way, I am using the this script to generate the benchmark.

[joshtriplett]

@rylev My only concern was catching potential regressions, but it isn't critical, and as you said, rustfmt makes it unlikely anyone will hit that case.