On Fri, 23 Jan 2026 10:20:37 GMT, Tobias Hartmann <[email protected]> wrote:

> Since [JDK-8247299](https://bugs.openjdk.org/browse/JDK-8247299), value 
> objects no longer use `markWord::always_locked_pattern` (see description of 
> [JDK-8247298](https://bugs.openjdk.org/browse/JDK-8247298) for more details). 
> Therefore, the limitations around storing the hash code in the mark word that 
> are mentioned in [JDK-8244975](https://bugs.openjdk.org/browse/JDK-8244975) 
> no longer exist. Let's enable this optimization.
> 
> Running `System.identityHashCode(obj)` in a loop and passing various 
> primitive boxes leads to a **17-18x speedup** on my machine. I'm also seeing 
> a **7% improvement** in the score of the SPECjvm2008 serial benchmark but we 
> are still a bit behind baseline without `--enable-preview`. More improvements 
> to come.
> 
> It's worth noting that value objects will lose their buffer object when 
> stored in a flat container or sometimes when being scalarized 
> ([JDK-8372268](https://bugs.openjdk.org/browse/JDK-8372268) will improve 
> this). The hash then needs to be re-computed, which is expensive.
> 
> I'm working on ways to further improve this, see 
> [JDK-8252185](https://bugs.openjdk.org/browse/JDK-8252185).
> 
> Thanks,
> Tobias

Thanks for taking a look at this Marc!

> A question nevertheless: I see that now we have & markWord::hash_mask, so I 
> take that this might change the hash (if result.get_int() highest bit is 1, 
> so negative) to make it fit on 31 bits?

Yes, we only have 31 bits in the mark word reserved for the hash code:
https://github.com/openjdk/valhalla/blob/258bd5840bb7b029b9877155c53b6b1122b0e9dc/src/hotspot/share/oops/markWord.hpp#L52-L55

Existing code for identity objects already does the same truncation.

-------------

PR Comment: https://git.openjdk.org/valhalla/pull/1954#issuecomment-3790043174

Reply via email to