Optimize basetype caches #9608
Optimize basetype caches #9608
Conversation
Their base types change, so they should get new denotations.
Phase-ids should not change base type caches. Instead we make sure that class denotations that contain base type caches are copied whenever base types might change.
odersky
force-pushed
the
optimize-basedata
branch
from
1bfab07 to
6729e1c
Compare
|
test performance please |
|
performance test scheduled: 14 job(s) in queue, 1 running. |
|
Performance test finished successfully: Visit http://dotty-bench.epfl.ch/9608/ to see the changes. Benchmarks is based on merging with master (1431be2) |
| valid && createdAt.runId == ctx.runId | ||
| // Note: We rely on the fact that whenever base types of classes change, | ||
| // the affected classes will get new denotations with new basedata caches. | ||
| // So basedata caches can become invalid only if the run changes. |
There was a problem hiding this comment.
I'm not familiar with code in this area. I'm wondering will the cache be always valid for classes defined in libraries, regardless of runs?
There was a problem hiding this comment.
Yes, but libraries might be redefined in source in the next run, so you need to re-check.
|
We get a ~5% improvement for dotc and a 8% improvement for re2. No change for scalap and 1-2% worse for stdlib. That could be in the noise though. Still I wonder why stdlib does not do better here. Maybe a low cache hit rate for base type caches? That would make sense as we have lots of small methods with their own type parameters. With a low hit rate, maybe fewer cache invalidations (which is what this PR achieves) means larger caches that are under-utilized. But that's speculation. |
Main optimization: Invalidate base type caches only if runId changes
Phase-ids should not change base type caches. Instead we make sure
that class denotations that contain base type caches are copied
whenever base types might change.