On Fri, 15 Dec 2023 01:16:55 GMT, Joshua Cao <d...@openjdk.org> wrote:
> ConcurrentHashMap's copy constructor calls `putAll()` -> `tryPresize()` -> > `transfer()`. When coming from the copy constructor, the Map is empty, so > there is nothing to transfer. But `transfer()` will still copy all the empty > nodes from the old table to the new table. > > This patch avoids this work by only calling `tryPresize()` if the table is > already initialized. If `table` is null, the initialization is deferred to > `putVal()`, which calls `initTable()`. > > --- > > ### JMH results for testCopyConstructor > > before patch: > > > Result "org.openjdk.bench.java.util.concurrent.Maps.testCopyConstructor": > 937395.686 ±(99.9%) 99074.324 ns/op [Average] > (min, avg, max) = (825732.550, 937395.686, 1072024.041), stdev = 92674.184 > CI (99.9%): [838321.362, 1036470.010] (assumes normal distribution) > > > after patch: > > > Result "org.openjdk.bench.java.util.concurrent.Maps.testCopyConstructor": > 620871.469 ±(99.9%) 59195.406 ns/op [Average] > (min, avg, max) = (545304.633, 620871.469, 689013.573), stdev = 55371.419 > CI (99.9%): [561676.063, 680066.875] (assumes normal distribution) > > > Average time is decreased by about 33%. We need @DougLea to take a look :) ------------- PR Comment: https://git.openjdk.org/jdk/pull/17116#issuecomment-1875046493
