benchmarks that take longer than 1ms per iteration aren't run and report 0 ns/iter

[toffaletti]

I was randomly seeing benchmarks report 0 ns/iter and had no idea why until @alexcrichton clued me in on the 1ms limit per iter. I'd like to add a warning when the initial iteration takes longer than 1ms so the user knows why the benchmark is "failing". Instead of just reporting 0 ns/iter.

I've got a question on the intent of part of the code before I continue trying to make this change.

https://github.com/mozilla/rust/blob/master/src/libextra/test.rs#L1060

        if self.ns_per_iter() == 0 {
            n = 1_000_000;
        } else {
            n = 1_000_000 / self.ns_per_iter().max(&1);
        }

The .max(&1) seems needless because self.ns_per_iter can only return positive integers and it was just checked for 0. I'm wondering if the line should actually be:

n = (1_000_000 / self.ns_per_iter()).max(&1);

To guarantee at least one iteration is run. If zero iterations is fine, then I'd like to add a check for that and print a warning instead.

[toffaletti]

repro:

extern mod extra;
use std::libc::sleep;
use extra::test::BenchHarness;

#[bench]
#[fixed_stack_segment] 
fn sleep_2_seconds(bh: &mut BenchHarness) {
    do bh.iter {
        unsafe { sleep(2); }
    }
}

running 1 test
test sleep_2_seconds ... bench: 0 ns/iter (+/- 0)

test result: ok. 0 passed; 0 failed; 0 ignored; 1 measured

[bluss]

Duplicate of #9532

[toffaletti]

Hah, I swear I searched for this issue before making one. Oh well.