RFR: 8278302: [s390] Implement fast-path for ASCII-compatible CharsetEncoders

[Lutz Schmidt]

On Tue, 7 Dec 2021 11:20:05 GMT, Lutz Schmidt <lucy at openjdk.org> wrote:

> This pull request contains the s390x variant of the fast-path for ASCII-compatible CharsetEncoders. 
> 
> The intrinsic implementation is pretty much similar to that of PPC64. Compared to the previously existing intrinsic implementation, changes to program logic are minimal. There is a new match rule for ascii strings and the bit mask used to detect invalid bits in the string characters is set depending on the target charset: 0xff00 for ISO, 0xff80 for ascii.
> 
> **Request:** SAP does not maintain a full build and test infrastructure for s390x anymore. Could somebody please test this PR on s390x? At least build and tier1 tests, the more the better. Thank you!
> 
> With my private testing (some microbenchmarks, some SPEC benchmarks) I could not find any issues. With the fast-path active I see a 2x, sometime 2.5x, improvement of CharsetEncodeDecode.encode. 
> 
> **Without the fast-path patch:**
> 
> CharsetEncodeDecode.encode   16384        UTF-8  avgt   30  16.350 ± 0.067  us/op
> CharsetEncodeDecode.encode   16384         BIG5  avgt   30  15.972 ± 0.136  us/op
> CharsetEncodeDecode.encode   16384  ISO-8859-15  avgt   30  14.372 ± 0.110  us/op
> CharsetEncodeDecode.encode   16384        ASCII  avgt   30  14.287 ± 0.051  us/op
> 
> 
> **With the fast-path patch:**
> 
> CharsetEncodeDecode.encode   16384        UTF-8  avgt   30   6.833 ± 0.064  us/op
> CharsetEncodeDecode.encode   16384         BIG5  avgt   30   8.195 ± 0.059  us/op
> CharsetEncodeDecode.encode   16384  ISO-8859-15  avgt   30   6.387 ± 0.149  us/op
> CharsetEncodeDecode.encode   16384        ASCII  avgt   30   6.321 ± 0.049  us/op

Thank you for testing! The results don't look good, though. I'll have to find out what the test expects and why my code doesn't deliver the expected. Will get back to you asap. 
Lutz

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

PR: https://git.openjdk.java.net/jdk/pull/6738