[vpaturet]

Summary

This PR adds FastZonedDateTimeParser, a position-based parser for yyyy-MM-ddTHH:mm:ss[.fffffffff][Z|±HH:MM]. It reads the fields at fixed positions, uses no regular expressions and allocates nothing but the result. ZonedDateTimeAdapter.parse tries it first and falls back to the previous ZonedDateTime.parse / LocalDateTime.parse logic for any other ISO form (missing seconds, ±HH:MM:SS offsets, [Region] suffixes, lowercase t/z, extended years, ...) and for any out-of-range field, so results and exception types are unchanged. Local times without zone designator are resolved in the fast path exactly as the fallback does (ZonedDateTime.ofLocal with a UTC preferred offset).

The same approach is in use in netex-java-model: entur/netex-java-model#310.

Results

Measured on a SIRI-ET document (282 MB, 1.8 million xsd:dateTime values), JDK 25, -Duser.timezone=Europe/Oslo, JMH average time per parse:

	before	after
ns/op, all values	1079 ± 118	74 ± 7
ns/op, values with fraction	1156 ± 273	95 ± 12
time in ZonedDateTimeAdapter.parse for the whole dump	≈ 1.94 s	≈ 0.13 s

The fast path handles 100 % of the values in that dump (95.7 % plain ±HH:MM, 4.3 % with a 1–9 digit fraction). A regex-based variant was also measured and only halved the cost, so it was dropped.

End-to-end test

Full SiriXml.parseXml(InputStream) of the 282 MB tst dump (18,874 EstimatedVehicleJourney, 1.8 M dateTimes), -Duser.timezone=Europe/Oslo, -Xms6G -Xmx6G, 4 pinned cores, 5 interleaved JVM launches per arm × 5 measured parses after 3 warm-up:

metric	master	PR	delta
wall per unmarshal, mean ± 95 % CI	7539 ± 224 ms	4849 ± 254 ms	−35.7 % (×1.55)
wall, median	7413 ms	4554 ms
main-thread CPU, mean	7104 ms	4614 ms	−35.1 %
young-gen allocation per unmarshal (GC log, approx.)	4.6 GB	1.0 GB	−77 %
G1 pauses over 6 unmarshals	23 / 9.65 s	14 / 6.95 s

ZonedDateTimeAdapter.parse alone over the 1,795,022 real strings (5 JVMs per arm × 10 iterations): 906 ± 23 ns/op → 118 ± 10 ns/op (×7.7), consistent with the JMH figures in the description.

The end-to-end gain (≈ 2.7 s) is about twice the adapter-only saving (1.8 M × 0.79 µs ≈ 1.4 s); the rest comes from the removed allocation pressure (fewer and shorter G1 pauses while the whole document tree is live) and the formatter running cold inside JAXB rather than hot in a micro loop. Per-round medians were consistent across all five rounds (master 6983–8127 ms, PR 4290–5745 ms).