9851873e42
Two byte-identical interpreter-dispatch optimizations (headless golden
n200m unchanged). Together ~13% faster on the -n 200M --gpu-inline
benchmark (4.33s -> ~3.8s, ~46 -> ~53 MIPS).
- scheduler.rs: `decrement_quantum_by(n)` — the superblock epilogue looped
`for _ in 0..executed { decrement_quantum() }` (~one bounds-checked call
per retired guest instruction, the largest fixed per-superblock cost).
QUANTUM_DEFAULT (50k) >> a superblock's instr count, so the quantum
boundary is crossed at most once/call: common path is one subtraction,
the rare boundary step defers to decrement_quantum for exact rotation
(reload + same-priority peer hand-off) semantics. Byte-identical.
- main.rs run_superblock: resolve the running thread's PpcContext ONCE per
block (was three `ctx_mut_ref` double-indexed slot lookups: cycle-before,
the step, cycle-after).
Profiled remaining breakdown (-n 300M --gpu-inline): step_block body 50%
(67 MIPS, the JIT target), block decode/cache 9% (page-version-bound),
kernel HLE 8%, scheduler/lock remainder ~31%. Sub-10% cheap wins remain
(block-linking is page-version-gated so only partial); the 5-10x lift is
the JIT.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>