RFR: 8322149: ConcurrentHashMap copy constructor should not transfer from old table on presizing

[Aleksey Shipilev]

On Fri, 15 Dec 2023 01:16:55 GMT, Joshua Cao <duke at 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