Regex match performance: ~4x slower than native Perl #1904

Problem

The regex benchmark (dev/bench/benchmark_regex.pl) running $i =~ /42/ in a tight loop shows PerlOnJava at ~10,453 iterations/s vs native Perl at ~39,683/s — a 3.8x performance gap.

Analysis

Async-profiler reveals the Joni regex engine itself accounts for only ~2% of CPU. The overhead is dominated by per-match ceremony in the runtime plumbing:

• ThreadLocal operations: ~21 per match (save/restore package, debug mode, warning state, source lifecycle)
• Object allocations: ~10 per match (RuntimeList, Region, Matcher, ResolvedRegex)
• Redundant calls: resolveRegex called twice, emitWarningsOnUse reads 4 ThreadLocals even when the warnings list is empty

Benchmark

# Native Perl
timethis 100000: @ 39683/s

# PerlOnJava (before)
timethis 100000: @ 10453/s

# PerlOnJava (after PR)
timethis 100000: @ 17430/s  (+67%)

Remaining gap

A ~2.3x gap remains vs native Perl. Further gains would require deeper changes:

• Integer-to-string caching (dual PV/IV slot in RuntimeScalar)
• Moving bytecode-emitted warning state from ThreadLocal to PerlRuntime fields
• Joni Matcher reuse/reset API

These are deferred to future work.