Bench harness

[graydon]

This is scaffolding for the new #[bench] attribute for marking unit tests as benchmarks. They are run with the --bench flag that the test runner now accepts. The runner automatically calibrates a test loop to an appropriate count to get a good per-iteration measurement.

[graydon]

r? @brson

[graydon]

Yeah, sorry, this is still pretty preliminary. You use it like this:

#[bench]
fn bench1(bh: & mut std::test::BenchHarness) {
    let mut x = 1;
    do bh.iter {
        x += 1;
        x ^= 2 * x;
    }
}

The separation is to allow you to do some scaffolding setup in the pre-iter part and then have the benchmark tool just measure the speed of the inner loop passed to .iter.

[brson]

r+

I am wary about codifying this one type of test harness, instead of having an extensible system, but I don't have any concrete suggestions right now for how to get there from here, nor any grand plans to redesign std::test. Perhaps we can revisit later in the year.

[graydon]

Agreed. This is just to get something habitual under way. We can do mindless rewrites to a better harness if someone thinks one up.

[quentusrex]

What type of measurements is the benchharness going to be recording? Just the runtime of the iteration? For runtime can it measure real, user and system time?

[graydon]

No, it just measures userspace nanoseconds. Idea is to be portable and reliable -- "too easy to not-use" much like our unit test harness -- not complete or thorough. We have proper profilers for real performance tuning.