extra::test: define an opaque function to assist with accurate benchmarking. #12105
Conversation
|
I'll file a bug replacing #8261 about fixing the various overly-optimised benchmarks in the source tree when this lands. |
|
Is there any way you could use |
|
Probably, but I don't think that's particularly better than this solution. (It might maybe work better in the face of LTO; I'll test.) |
|
I've added two new commits; the second of which is an example of how to fix broken benchmarks. |
|
It seems a little unfortunate to require that most benchmarks use |
|
I think @alexcrichton's suggestion is the right way to go. |
benchmarking. This allows a result to be marked as "used" by passing it to a function LLVM cannot see inside. By making `iter` generic and using this `black_box` on the result benchmarks can get this behaviour simply by returning their computation.
Before:
test test::bench_nonpod_nonarena ... bench: 62 ns/iter (+/- 6)
test test::bench_pod_nonarena ... bench: 0 ns/iter (+/- 0)
After:
test test::bench_nonpod_nonarena ... bench: 158 ns/iter (+/- 11)
test test::bench_pod_nonarena ... bench: 48 ns/iter (+/- 2)
The other tests show no change, but are adjusted to use the generic
return value of `.iter` anyway so that this doesn't change in future.
|
Done. |
This allows a result to be marked as "used" by passing it to a function LLVM cannot see inside (unless LTO is enabled). Closes #8261.
… r=llogiq Extend `useless_asref` lint on `map(clone)` If you have code like: ```rust Some(String::new()).as_ref().map(Clone::clone) ``` the `as_ref` call is unneeded. Interestingly enough, this lint and `map_clone` are starting to share a same "space" where both lints warn about different things for the same code. Not sure what's the policy about such cases though... r? `@llogiq` changelog: Extend `useless_asref` lint on `map(clone)`
This allows a result to be marked as "used" by passing it to a function
LLVM cannot see inside (unless LTO is enabled).
Closes #8261.