PEGO
A160: Whole-Input Scratch Pooling
Repeated batch parses can reuse decoded input, offset, memo and collection buffers that no returned result references.
Mechanism and Invariants
Engine Programs and generated Go runtimes acquire/release cleared scratch
through sync.Pool. Result Node/list/field storage is detached from pooled
scratch; token strings may still retain source text. Oversized backing is
discarded under per-buffer limits. Recognition may use another engine Program
pool or a generated shared pool with rule-table counters reset. A pool may be
cleared by GC and is not persistent result ownership.
Scope and State
Default closure/VM whole-input parsing and generated Go Node/recognition/typed parsing. Documents and streams keep their own state; TypeScript creates an invocation-owned parser with no cross-call pool.
Reference Control
No universal same-source pool-disable control exists. Narrow Node frame/sparse controls do not disable all pooling.
Evidence and Implementation
Records 057/058 report lower allocation volume and workload-specific time reductions; JSON generated Parse remains within noise. Typed pooling is also described in 048. These historical results do not promise a warmed pool or zero allocation for every invocation.
Source owners: acquirePooledParser and releaseScratch,
sharedProgram.scratch, generated parser pool, typed pool and release.
The comparison policy distinguishes same-source controls, historical changes and full checkpoints.