<i18n dev> RFR: 8261254: Initialize charset mapping data lazily

[Claes Redestad]

On Sun, 7 Feb 2021 19:08:18 GMT, Claes Redestad <redestad at openjdk.org> wrote:

> This patch refactor JDK internal charsets to initialize charset mapping data lazily when needed via holder classes. This means both a startup improvement in some cases, and possible throughput improvements for all DoubleByte-based Charsets.
> 
> Testing: tier1-3

Applications calling `Charset.availableCharsets()` might see a decent startup improvement (125-190ms -> 57-95ms first call overhead).

All `DoubleByte`-derived charsets can see a throughput improvement on benchmarks thanks to removing DCL initialization checks with a Holder pattern implementation, which allows static data to be declared final and be more amenable to optimizations:

Before:
Benchmark                                   (charsetName)  Mode  Cnt    Score    Error  Units
StringDecode.WithCharset.decodeCharset              MS932  avgt   15  280.072 ± 17.217  ns/op
StringDecode.WithCharset.decodeCharsetName          MS932  avgt   15  289.368 ± 18.174  ns/op

After:
Benchmark                                   (charsetName)  Mode  Cnt    Score    Error  Units
StringDecode.WithCharset.decodeCharset              MS932  avgt   15  254.906 ± 15.095  ns/op
StringDecode.WithCharset.decodeCharsetName          MS932  avgt   15  268.241 ± 15.725  ns/op

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

PR: https://git.openjdk.java.net/jdk/pull/2449