<i18n dev> RFR: 8302877: Speed up latin1 case conversions [v2]

[Eirik Bjorsnos]

On Tue, 21 Feb 2023 06:59:47 GMT, Eirik Bjorsnos <duke at openjdk.org> wrote:

>> This PR suggests we speed up Character.toUpperCase and Character.toLowerCase for latin1 code points by applying the 'oldest ASCII trick in the book'.
>> 
>> This takes advantage of the fact that latin1 uppercase code points are always 0x20 lower than their lowercase (with the exception of two code points which uppercase out of latin1).
>> 
>> To verify the correctness of the new implementation, the test `Latin1CaseConversion` is added with an exhaustive verification of toUpperCase/toLowerCase for all latin1 code points.
>> 
>> The implementation needs to balance the performance of the various ranges in latin1. An effort has been made to favour operations on ASCII code points, without causing excessive regression for higher code points.
>> 
>> Performance is benchmarked for 7 chosen sample code points, each representing a range or a special-case.  Results in the first comment.
>
> Eirik Bjorsnos has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Spell fix for 'exhaustive' in comments in sun/text/resources

A site note: Early and crude experiements using the Vector API indicate that the 'oldest ASCII trick in the book' vectorizes pretty well.

Here's a benchmark comparing the strings "helloworld" and "HelloWorld" repeated 1024 times, followed by either 'A' or 'B' (to force a an expensive mismatch):


Benchmark                    (size)  Mode  Cnt     Score    Error  Units
EqualsIgnoreCase.scalar        1024  avgt   15  6225.624 ± 89.182  ns/op
EqualsIgnoreCase.vectorized    1024  avgt   15  1246.110 ± 14.767  ns/op


I have the feeling that most case-insensitive comparisons are pretty short, so not sure how useful this is IRL.

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

PR: https://git.openjdk.org/jdk/pull/12623