enhancement: default to opcache.jit=tracing for CPU-bound workloads

[denerFernandes]

Summary

Current default opcache.jit=function gives us ~42k req/s on the Mandelbrot benchmark (ZTS, 8w). A flamegraph from profile-hetzner.yml shows that ~70% of CPU in CPU-bound workloads is already in the JIT buffer (addresses 0x8000XXXX), with only ~20% in libphp (interpreter/runtime) and ~0.2% in our Rust code. That confirms JIT is firing — but function JIT compiles each hot function individually and cannot speculate across calls or inline.

opcache.jit=tracing records runtime type profiles and compiles hot traces across function boundaries, inlining and specialising. For tight inner loops (Mandelbrot is the canonical case) it typically wins 20–40% over function JIT.

Proposal

1. Add an INI control (already wired via overrides.jit_buffer_size) that also lets users pick jit=tracing|function|off, default tracing.
2. Rebuild with tracing and rerun the full benchmark matrix. Pay special attention to:
   • Mandelbrot (pure CPU-bound): primary target, expected 20–40% gain.
   • Raw Hello World: likely neutral (no hot loops), but measure for regressions.
   • Laravel/Symfony worker mode: could gain 5–15% on request bootstrap.
   • Binary 50KB: likely neutral.
3. Compare Firefox Profiler call-trees before/after to verify more time in tighter native code (fewer 0x8000 sites but each hotter) and less in libphp helpers.

Risks

• tracing has a higher warmup cost — first dozens of requests per route may be slower while traces are recorded. Our worker mode amortises this trivially.
• Buffer size: tracing typically consumes 2–4× more JIT memory than function mode. Current default jit_buffer_size=64M should handle it but re-check on Laravel.
• Regressions on certain codepaths have been reported historically. Always run the full matrix.

Acceptance criteria

• Config exposes php.opcache_jit (off|function|tracing), default tracing.
• Benchmark matrix rerun committed to benchmarks/results/ showing Mandelbrot ZTS 8w ≥ 50k req/s.
• No workload regresses by >5% vs current numbers.

Related

• perf: per-request mode with frameworks ~3× slower than FPM (Laravel 500ms p50) #17, bug: worker mode returns all errors for 50KB binary responses (NTS/ZTS × 4w/8w flaky) #18, perf: worker mode ~2× slower than per-request for large (50KB) binary responses #19 — other perf/bug work derived from the same benchmark sweep.
• Profile data: turbine-profile-zts-php8.5 artifact in run 24897714344.