110. No Recovered-Error Rollback Without Recovery

• Status: Implemented
• Date: 2026-10-22

#Problem and Mechanism

Every backtracking point of a generated parser saved len(p.Recovered) next to the position and truncated the list when it went back. p.Recovered is only appended by recovery (#recover); memo entries copy only errors that exist. In a grammar without #recover, the list stays empty, so saving its length and truncating it are no-ops.

The generator now decides once per program whether any rule can recover (a #recover attribute or a recovery expression). If none can, the saved variable, the p.Restore(start, rec) calls and the length stores are not generated; a rollback is p.Pos = start. Grammars with #recover generate what they did before. TestGeneratedRecoveredRollbackOnlyWithRecovery checks both forms, and TestGeneratedParsersMatchEngine, TestGeneratedTypes, TestGeneratedDocumentsMatchBatchOnCorpus, TestGeneratedUnsharedParsersMatchEngine and TestGeneratedConstructedFieldChecks fail when the rollback is dropped under recovery.

#Results

parser.go with -typed-only, bytes and lines before and after:

The duckdb test binary shrinks from 17,656,658 to 17,078,738 bytes (-3.3%); its cold compile time (best of three, clean build cache) is 7.0 s before and 7.4 s after, which is the spread of that measurement. Parse speed (30 interleaved rounds, minimum, an identical copy as the control) is within noise: python -1.25% (control -1.21%), cue -2.01% (-0.29%), duckdb -1.76% (-0.03%), cel +1.14% (+1.05%); cue and duckdb are 1.7 points faster than their controls, python and cel within 0.1 point.

#Decision and Limits

Adopted. The optimization applies to the whole program: one #recover anywhere keeps the rollback everywhere.