build: Reduce test build times by tweaking build config

[osiewicz]

It seems that with cargo t --no-run we're quite blocked on frontend producing CGUs. The following changes reduced incremental build time of bors in test scenarios by 40%.

[osiewicz]

r? @Kobzol

[Kobzol]

Thanks for taking a look! Sorry, I'm swamped at the moment, so it might take me some time before I can test this. I'll try to take a look after RustMeet.

[Kobzol]

I have some bad news =D

Some notes:

• The default duration for me at the time of testing was around 17-17.5s.
• Setting the CGU count to 32 increases the number of crates compiled in cargo test from 455 to 529. Using this:

  [profile.test]
  codegen-units = 32
  
  [profile.test.build-override]
  codegen-units = 32

  brings it back to 455.
• With CGU=32, the rebuild time is essentially the same, even slightly worse, but that is probably noise.
• With CGU=32 and build-override CGU=32, it's the same as above.
• Setting debug = "limited" is what really helps, as it brings the time from ~17s down to ~10s. But I often debug and step through code when testing, so the debuginfo = "limited" change is kinda limiting ( :) ) for me. I might use some kind of fast-test profile though for the case where I don't need to debug.

It is clear that debuginfo generation is a big contributor to the slow compile time here (though 10s is still very bad 😆).

Moving tests to a separate file, so that only tests (and not the bors lib target) is recompiled after a test change, gets the time down to 9s. So the (re-)compilation of tests themselves is just slow, even ignoring the Cargo inefficiency of recompiling also the lib.