Papercuts with #[bench]

[catamorphism]

1. #[bench] functions that have the wrong type signature get silently ignored. This should be an error.
2. Running a compiled thing with #[bench]s in it yields:

running 1 test
test g ... bench: 16 ns/iter (+/- 0)
result: ok. 0 passed; 0 failed; 0 ignored

The last line (0 passed/0 failed/0 ignored) is confusing here, since it did run something.

1. I don't think there's much documentation of #[bench] -- specifically that you have to compile with --test and run with --bench.

[pnkfelix]

visiting for triage, email from 2013 sep 09. I agree, the above problems do indeed look like papercuts.

[pnkfelix]

also, it might be nice if the benchmark output right-aligned the decimal points in the numbers, for ease of quick comparison by eye.

Doing the above in full generality is probably not worthwhile, since it requires knowing how many digits the maximal output has before you present any results. (It also requires knowing the longest test name, but that is easier to predict accurately. ;)

But even just guessing a reasonable value for the max number of digits for ns/iter, e.g. 9 or 10, and then using that for format width padding, would be step forward.

[pnkfelix]

Also, the presentation of 0 ns/iter for benchmark runs that actually timed out is a pretty serious papercut; see for example discussion on #9653

[emberian]

The testing guide now has docs. The new result line is test result: ok. 0 passed; 0 failed; 0 ignored; 1 measured. I've opened #12997 for the typechecking thing, closing.