Travis: Enable cpu-specific optimizations.

[c0gent]

This is off by default so that binaries can be shared. Enabling this
flag could improve test speed on Travis (will check).

[afck]

🏁 12 min 39 sec 🏁
That was among the fastest test runs. 🏆
(Unfortunately there is a high variance because of the randomness in the tests, so it's hard to tell how much of a difference the optimizations make.)

[c0gent]

Indeed.

[c0gent]

I'm going to wait until the mlock issues are sorted before rebasing and merging this.

[afck]

This could probably be rebased now. (Which will give us another data point. ⏱️)

[c0gent]

Done.

[afck]

Needs to be rebased again; then it can be merged, I think.

[mbr]

I'm going to wait until the mlock issues are sorted before rebasing and merging this.

I'd +1 this. Should we give the SIGILL a bit more time to crop up?

[mbr]

Also: rust-lang/rust#53322 (only nightly - but will that envvar mess with clippy compilation?).

[c0gent]

This is rebased. Will merge if tests pass.

[afck]

SIGILL 😩
Did MLOCK_SECRETS somehow not work (not exported? outdated threshold_crypto dependency?) or is this a separate issue…?

[c0gent]

The dependency should always be up to date on travis. I'm not sure what the deal is.

[c0gent]

Added --test-threads 1. I'm not sure if this is being added elsewhere or what has been tried or discussed about it.

[c0gent]

I'm going to rerun the CI check a few times just to see what happens. Do we want to keep the --test-threads 1 part?

[c0gent]

We may want to experiment with setting it to 2 as well since the Travis env uses 2 cores.

We could also use a mutex to selectively prevent the tests that use the most mlock memory from running in parallel.

[afck]

In either case, the SIGILL isn't related to -cpu=native at all, is it? I've also seen it in other PRs; so this is good to merge?