Add a new const eval perf problem to the test suite #349
Conversation
|
2 seconds seems fine; we try to stay under a hard limit of 30. I'm not so sure about modifying an existing benchmark though - it'll show up as a regression despite no changes to the compiler. @nnethercote, what do you think? |
|
could we rename the benchmark? so the old name just stops having data and the new one starts fresh? |
|
Yeah, thinking on this more, that seems like the best option. It might be that we want to long-term add some form of versioning to benchmarks (so we can update whenever and graphs are properly disjoint). |
|
I agree that renaming is a good idea. Simply adding a '-2' suffix might suffice? |
|
I've renamed the benchmark |
|
I will try to deploy this later today, thanks! |
| // copying all these zeros and the corresponding definedness bits can be expensive and is probably | ||
| // prone to regressions. | ||
| // 24 is an exponent that makes the repeat expression take less than two seconds to compute | ||
| const FOO: [i32; 1 << 24] = [0; 1 << 24]; |
There was a problem hiding this comment.
I tried hard to give more or less reasonable names to the constants above; I think we can do better than FOO ;)
This takes around two seconds on my machine, but with a few optimizations I've been able to push it below one second (rust-lang/rust#58556) and I still see many more avenues to optimize it. We can lower the exponent to decrease the compilation time further.
Since this entire test used to take around 16 seconds, I hope the additional 2 seconds are ok for now.