The JIT compiles hot code paths to machine code and helps almost nothing in a typical web request, where the time is in I/O.
opcache.jit_buffer_size=64M
opcache.jit=tracing
; measured, on a real application:
; a checkout request: -1% to +2%
; a Mandelbrot benchmark: -70%
;
; the request spends its time in the database, in Redis
; and in serialisation. none of that is JIT-able.
Reporting a benchmark improvement as an application improvement is the mistake this exists to prevent, and it happened widely when 8.0 shipped. Where it genuinely helps is a long-running CPU-bound process — image processing, a large export, a numerical job — which is a small and identifiable set of workloads worth enabling it for specifically.