RFR: 8332842: Optimize empty CopyOnWriteArrayList allocations

[Doug Lea]

On Mon, 3 Jun 2024 16:47:20 GMT, jengebr <duke at openjdk.org> wrote:

> Improve `java/util/concurrent/CopyOnWriteArrayList` by eliminating needless cloning of Object[0] instances. This cloning is intended to prevent callers from changing array contents, but many `CopyOnWriteArrayList`s are allocated to size zero, or are otherwise maintained empty, so cloning is unnecessary.
> 
> Results from the included JMH benchmark:
> Before:
> 
> Benchmark                                                    Mode  Cnt   Score   Error  Units
> CopyOnWriteArrayListBenchmark.clear                          avgt    5  74.487 ± 1.793  ns/op
> CopyOnWriteArrayListBenchmark.clearEmpty                     avgt    5  27.918 ± 0.759  ns/op
> CopyOnWriteArrayListBenchmark.createInstanceArray            avgt    5  16.656 ± 0.375  ns/op
> CopyOnWriteArrayListBenchmark.createInstanceArrayEmpty       avgt    5  15.415 ± 0.489  ns/op
> CopyOnWriteArrayListBenchmark.createInstanceCollection       avgt    5  21.608 ± 0.363  ns/op
> CopyOnWriteArrayListBenchmark.createInstanceCollectionEmpty  avgt    5  15.374 ± 0.260  ns/op
> CopyOnWriteArrayListBenchmark.createInstanceDefault          avgt    5  15.688 ± 0.350  ns/op
> 
> 
> After:
> 
> Benchmark                                                    Mode  Cnt   Score   Error  Units
> CopyOnWriteArrayListBenchmark.clear                          avgt    5  75.365 ± 2.092  ns/op
> CopyOnWriteArrayListBenchmark.clearEmpty                     avgt    5  20.803 ± 0.539  ns/op
> CopyOnWriteArrayListBenchmark.createInstanceArray            avgt    5  16.808 ± 0.582  ns/op
> CopyOnWriteArrayListBenchmark.createInstanceArrayEmpty       avgt    5  12.980 ± 0.418  ns/op
> CopyOnWriteArrayListBenchmark.createInstanceCollection       avgt    5  21.627 ± 0.173  ns/op
> CopyOnWriteArrayListBenchmark.createInstanceCollectionEmpty  avgt    5  12.864 ± 0.408  ns/op
> CopyOnWriteArrayListBenchmark.createInstanceDefault          avgt    5  12.931 ± 0.255  ns/op

The jmh benchmark checks only the empty case, you need to also show lack of impact on non-empty cases.

Assuming you demonstrate this, it seems basically OK, (Deja vu previous cases including hash maps). It is only a small band-aid -- programs generating lots of them still have to allocate the COWAL object, so the savings are small compared to not generating them at all unless needed.

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

PR Comment: https://git.openjdk.org/jdk/pull/19527#issuecomment-2145896233