Practical LR error recovery

We present a practical, language independent mechanism for error recovery in LR parsers. The method is easy to implement in existing parser generators. It uses only the normal parse tables and a small amount of symbol cost information. It is possible to use compressed parse tables with the method, as well as other LR augmentations such as precedence and associativity declarations. The method is efficient enough to be used in environments where space and time are at a premium. Our method utilizes a number of previous error recovery ideas: forward moves in the input after detection of errors to incorporate right context, recovery choice based on weighted costs, and the use of special productions to indicate major productions in the grammar for error recovery. The method also makes use of semantic information in choosing between possible recovery actions. We have implemented the method in a production Pascal translator which has been in use for instruction at a number of universities for two years. We report here on measurements of our system on a standard data base of Pascal programs with errors.