codegen-units causes performance problems

[pcwalton]

See rust-lang/rust#33522, for example. codegen-units is well known to cause performance problems due to missed IPO such as inlining.

Can we enable this only on bots?

[pcwalton]

I've seen mem::forget not being inlined, and gw has seen performance problems in euclid suspected to be caused by this.

[mbrubeck]

Now that Cargo supports $RUSTFLAGS, we could add -C codegen-units=4 to RUSTFLAGS in the buildbot environment, instead of putting it in the Cargo manifests.

[mbrubeck]

cc #9987

[metajack]

If this really causes big performance issues, this is likely going to inflate our test times even as it saves our build times. The only way to know whether this is worth it in automation is to try it and see.

[mbrubeck]

#9987 (comment) found no major change to test times. However, Servo performance may not be the dominating factor in our test runs.

[larsbergstrom]

I'm OK with disabling codegen-units. As I mentioned on IRC, though, if we do that then I'd like to do so consistently so that what we're developing on is what we're testing and what we're (eventually) shipping. I would be really iffy about "just using codegen-units" in CI to cut down the build time, as we've had concurrency issues and even crashes that were specific to the optimization level.

[glennw]

I tested the mean frame time in the backend WR2 prototype thread on one of the stress tests I profile with:

codegen-units: 8.9ms
without: 6.2ms

So it does cause a dramatic performance issue in the WR2 prototype - which is unfortunate, since the build times are very noticeably faster with codegen-units :(