The SQL language parser for SQLite is generated using a code-generator program called "Lemon". The Lemon program reads a grammar of the input language and emits C-code to implement a parser for that language.
Lemon does not have its own source repository. Rather, Lemon consists of a few files in the SQLite source tree:
lemon.html → The original detailed usage documentation and programmers reference for Lemon.
lemon.c → The source code for the utility program that reads a grammar file and generates corresponding parser C-code.
lempar.c → A template for the generated parser C-code. The "lemon" utility program reads this template and inserts additional code in order to generate a parser.
Lemon generates an LALR(1) parser. Its operation is similar to the more familiar tools Yacc and Bison, but Lemon adds important improvements, including:
The grammar syntax is less error prone - using symbolic names for semantic values rather that the "$1"-style positional notation of Yacc.
In Lemon, the tokenizer calls the parser. Yacc operates the other way around, with the parser calling the tokenizer. The Lemon approach is reentrant and threadsafe, whereas Yacc uses global variables and is therefore neither. Reentrancy is especially important for SQLite since some SQL statements make recursive calls to the parser. For example, when parsing a CREATE TABLE statement, SQLite invokes the parser recursively to generate an INSERT statement to make a new entry in the sqlite_schema table.
Lemon has the concept of a non-terminal destructor that can be used to reclaim memory or other resources following a syntax error or other aborted parse.
Lemon is used in two places in SQLite.
The primary use of Lemon is to create the SQL language parser. A grammar file (parse.y) is compiled by Lemon into parse.c and parse.h. The parse.c file is incorporated into the amalgamation without further modification.
Lemon is also used to generate the parser for the query pattern expressions in the FTS5 extension. In this case, the input grammar file is fts5parse.y.
One of the advantages of hosting code generator tools as part of the project is that the tools can be optimized to serve specific needs of the overall project. Lemon has benefited from this effect. Over the years, the Lemon parser generator has been extended and enhanced to provide new capabilities and improved performance to SQLite. A few of the specific enhancements to Lemon that are specifically designed for use by SQLite include:
Lemon has the concept of a "fallback" token. The SQL language contains a large number of keywords and these keywords have the potential to collide with identifier names. Lemon has the ability to designate some keywords as being able to "fallback" to an identifier. If the keyword appears in the input token stream in a context that would otherwise be a syntax error, the token is automatically transformed into its fallback before the syntax error is raised. This feature allows the parser to be very forgiving of reserved words used as identifiers, which is a problem that comes up frequently in the SQL language.
In support of the 100% MC/DC testing goal for SQLite, the parser code generated by Lemon has no unreachable branches, and contains extra (compile-time selected) instrumentation useful for measuring test coverage.
Lemon supports conditional compilation of grammar file rules, so that a different parser can be generated depending on compile-time options.
As a performance optimization, reduce actions in the Lemon input grammar are allowed to contain comments of the form "/*A-overwrites-Z*/" to indicate that the semantic value "A" on the right-hand side of the rule is allowed to directly overwrite the semantic value "Z" on the left-hand side. This simple optimization reduces the number of stack operations in the push-down automaton used to parse the input grammar, and thus improve performance of the parser. It also makes the generated code a little smaller.
The parsing of SQL statements is a significant consumer of CPU cycles in any SQL database engine. On-going efforts to optimize SQLite have caused the developers to spend a lot of time tweaking Lemon to generate faster parsers. These efforts have benefited all users of the Lemon parser generator, not just SQLite. But if Lemon had been a separately maintained tool, it would have been more difficult to make coordinated changes to both SQLite and Lemon, and as a result not as much optimization would have been accomplished. Hence, the fact that the parser generator tool is included in the source tree for SQLite has turned out to be a net benefit for both the tool itself and for SQLite.
Lemon was originally written by D. Richard Hipp (also the creator of SQLite) while he was in graduate school at Duke University between 1987 and 1992. The original creation date of Lemon has been lost, but was probably sometime around 1990. Lemon generates an LALR(1) parser. There was a companion LL(1) parser generator tool named "Lime", but the source code for Lime has been lost.
The Lemon source code was originally written as separate source files, and only later merged into a single "lemon.c" source file.
The author of Lemon and SQLite (Hipp) reports that his C programming skills were greatly enhanced by studying John Ousterhout's original source code to Tcl. Hipp discovered and studied Tcl in 1993. Lemon was written before then, and SQLite afterwards. There is a clear difference in the coding styles of these two products, with SQLite seeming to be cleaner, more readable, and easier to maintain.
SQLite is in the Public Domain.
https://sqlite.org/lemon.html