On Thu, 30 May 2024 16:15:22 GMT, Maurizio Cimadamore <mcimadam...@openjdk.org> wrote:
> This PR restores a var handle cache in `Utils::makeSegmentViewVarHandle`. The > cache was moved to `ValueLayouts::varHandle` as part of > [pull/19251](https://git.openjdk.org/jdk/pull/19251), on the basis that we > want to optimize the common case like: > > > ValueLayout layout = ... > layout.varHandle().get(...) > > > And that caching more complex var handles didn't seem to add value, given > that, for these var handles, the logic in `LayoutPath` needs to adapt the > returned var handle anyways. > > But, `TestAccessModes` revealed a different picture - w/o any cache in > `Utils` the test end up allocating 8963 var handle instances (instead of just > 4), in each of the 4 runs the test includes. While this is admittedly a > stress test, it seems nice to restore the level of sharing we had before > [pull/19251](https://git.openjdk.org/jdk/pull/19251). A different approach could be to find a way to add a per-layout cache. E.g. if the root layout is a group layout, that layout could have a cache for all the different var handles generated from it. The key would be the path element used to retrieve the var handle. ------------- PR Comment: https://git.openjdk.org/jdk/pull/19485#issuecomment-2140148973
