The Lemon Parser Generator

Lemon is an LALR(1) parser generator for C. It does the same job as "bison" and "yacc". But lemon is not a bison or yacc clone. Lemon uses a different grammar syntax which is designed to reduce the number of coding errors. Lemon also uses a parsing engine that is faster than yacc and bison and which is both reentrant and threadsafe. (Update: Since the previous sentence was written, bison has also been updated so that it too can generate a reentrant and threadsafe parser.) Lemon also implements features that can be used to eliminate resource leaks, making is suitable for use in long-running programs such as graphical user interfaces or embedded controllers.

This document is an introduction to the Lemon parser generator.

Theory of Operation

The main goal of Lemon is to translate a context free grammar (CFG) for a particular language into C code that implements a parser for that language. The program has two inputs:

Typically, only the grammar specification is supplied by the programmer. Lemon comes with a default parser template which works fine for most applications. But the user is free to substitute a different parser template if desired.

Depending on command-line options, Lemon will generate between one and three files of outputs.

By default, all three of these output files are generated. The header file is suppressed if the "-m" command-line option is used and the report file is omitted when "-q" is selected.

The grammar specification file uses a ".y" suffix, by convention. In the examples used in this document, we'll assume the name of the grammar file is "gram.y". A typical use of Lemon would be the following command:

   lemon gram.y
This command will generate three output files named "gram.c", "gram.h" and "gram.out". The first is C code to implement the parser. The second is the header file that defines numerical values for all terminal symbols, and the last is the report that explains the states used by the parser automaton.

Command Line Options

The behavior of Lemon can be modified using command-line options. You can obtain a list of the available command-line options together with a brief explanation of what each does by typing

   lemon -?
As of this writing, the following command-line options are supported:

The Parser Interface

Lemon doesn't generate a complete, working program. It only generates a few subroutines that implement a parser. This section describes the interface to those subroutines. It is up to the programmer to call these subroutines in an appropriate way in order to produce a complete system.

Before a program begins using a Lemon-generated parser, the program must first create the parser. A new parser is created as follows:

   void *pParser = ParseAlloc( malloc );
The ParseAlloc() routine allocates and initializes a new parser and returns a pointer to it. The actual data structure used to represent a parser is opaque — its internal structure is not visible or usable by the calling routine. For this reason, the ParseAlloc() routine returns a pointer to void rather than a pointer to some particular structure. The sole argument to the ParseAlloc() routine is a pointer to the subroutine used to allocate memory. Typically this means malloc().

After a program is finished using a parser, it can reclaim all memory allocated by that parser by calling

   ParseFree(pParser, free);
The first argument is the same pointer returned by ParseAlloc(). The second argument is a pointer to the function used to release bulk memory back to the system.

After a parser has been allocated using ParseAlloc(), the programmer must supply the parser with a sequence of tokens (terminal symbols) to be parsed. This is accomplished by calling the following function once for each token:

   Parse(pParser, hTokenID, sTokenData, pArg);
The first argument to the Parse() routine is the pointer returned by ParseAlloc(). The second argument is a small positive integer that tells the parse the type of the next token in the data stream. There is one token type for each terminal symbol in the grammar. The gram.h file generated by Lemon contains #define statements that map symbolic terminal symbol names into appropriate integer values. A value of 0 for the second argument is a special flag to the parser to indicate that the end of input has been reached. The third argument is the value of the given token. By default, the type of the third argument is integer, but the grammar will usually redefine this type to be some kind of structure. Typically the second argument will be a broad category of tokens such as "identifier" or "number" and the third argument will be the name of the identifier or the value of the number.

The Parse() function may have either three or four arguments, depending on the grammar. If the grammar specification file requests it (via the extra_argument directive), the Parse() function will have a fourth parameter that can be of any type chosen by the programmer. The parser doesn't do anything with this argument except to pass it through to action routines. This is a convenient mechanism for passing state information down to the action routines without having to use global variables.

A typical use of a Lemon parser might look something like the following:

   01 ParseTree *ParseFile(const char *zFilename){
   02    Tokenizer *pTokenizer;
   03    void *pParser;
   04    Token sToken;
   05    int hTokenId;
   06    ParserState sState;
   08    pTokenizer = TokenizerCreate(zFilename);
   09    pParser = ParseAlloc( malloc );
   10    InitParserState(&sState);
   11    while( GetNextToken(pTokenizer, &hTokenId, &sToken) ){
   12       Parse(pParser, hTokenId, sToken, &sState);
   13    }
   14    Parse(pParser, 0, sToken, &sState);
   15    ParseFree(pParser, free );
   16    TokenizerFree(pTokenizer);
   17    return sState.treeRoot;
   18 }
This example shows a user-written routine that parses a file of text and returns a pointer to the parse tree. (All error-handling code is omitted from this example to keep it simple.) We assume the existence of some kind of tokenizer which is created using TokenizerCreate() on line 8 and deleted by TokenizerFree() on line 16. The GetNextToken() function on line 11 retrieves the next token from the input file and puts its type in the integer variable hTokenId. The sToken variable is assumed to be some kind of structure that contains details about each token, such as its complete text, what line it occurs on, etc.

This example also assumes the existence of structure of type ParserState that holds state information about a particular parse. An instance of such a structure is created on line 6 and initialized on line 10. A pointer to this structure is passed into the Parse() routine as the optional 4th argument. The action routine specified by the grammar for the parser can use the ParserState structure to hold whatever information is useful and appropriate. In the example, we note that the treeRoot field of the ParserState structure is left pointing to the root of the parse tree.

The core of this example as it relates to Lemon is as follows:

      pParser = ParseAlloc( malloc );
      while( GetNextToken(pTokenizer,&hTokenId, &sToken) ){
         Parse(pParser, hTokenId, sToken);
      Parse(pParser, 0, sToken);
      ParseFree(pParser, free );
Basically, what a program has to do to use a Lemon-generated parser is first create the parser, then send it lots of tokens obtained by tokenizing an input source. When the end of input is reached, the Parse() routine should be called one last time with a token type of 0. This step is necessary to inform the parser that the end of input has been reached. Finally, we reclaim memory used by the parser by calling ParseFree().

There is one other interface routine that should be mentioned before we move on. The ParseTrace() function can be used to generate debugging output from the parser. A prototype for this routine is as follows:

   ParseTrace(FILE *stream, char *zPrefix);
After this routine is called, a short (one-line) message is written to the designated output stream every time the parser changes states or calls an action routine. Each such message is prefaced using the text given by zPrefix. This debugging output can be turned off by calling ParseTrace() again with a first argument of NULL (0).

Differences With YACC and BISON

Programmers who have previously used the yacc or bison parser generator will notice several important differences between yacc and/or bison and Lemon.

These differences may cause some initial confusion for programmers with prior yacc and bison experience. But after years of experience using Lemon, I firmly believe that the Lemon way of doing things is better.

Updated as of 2016-02-16: The text above was written in the 1990s. We are told that Bison has lately been enhanced to support the tokenizer-calls-parser paradigm used by Lemon, and to obviate the need for global variables.

Input File Syntax

The main purpose of the grammar specification file for Lemon is to define the grammar for the parser. But the input file also specifies additional information Lemon requires to do its job. Most of the work in using Lemon is in writing an appropriate grammar file.

The grammar file for lemon is, for the most part, free format. It does not have sections or divisions like yacc or bison. Any declaration can occur at any point in the file. Lemon ignores whitespace (except where it is needed to separate tokens) and it honors the same commenting conventions as C and C++.

Terminals and Nonterminals

A terminal symbol (token) is any string of alphanumeric and/or underscore characters that begins with an upper case letter. A terminal can contain lowercase letters after the first character, but the usual convention is to make terminals all upper case. A nonterminal, on the other hand, is any string of alphanumeric and underscore characters than begins with a lower case letter. Again, the usual convention is to make nonterminals use all lower case letters.

In Lemon, terminal and nonterminal symbols do not need to be declared or identified in a separate section of the grammar file. Lemon is able to generate a list of all terminals and nonterminals by examining the grammar rules, and it can always distinguish a terminal from a nonterminal by checking the case of the first character of the name.

Yacc and bison allow terminal symbols to have either alphanumeric names or to be individual characters included in single quotes, like this: ')' or '$'. Lemon does not allow this alternative form for terminal symbols. With Lemon, all symbols, terminals and nonterminals, must have alphanumeric names.

Grammar Rules

The main component of a Lemon grammar file is a sequence of grammar rules. Each grammar rule consists of a nonterminal symbol followed by the special symbol "::=" and then a list of terminals and/or nonterminals. The rule is terminated by a period. The list of terminals and nonterminals on the right-hand side of the rule can be empty. Rules can occur in any order, except that the left-hand side of the first rule is assumed to be the start symbol for the grammar (unless specified otherwise using the %start directive described below.) A typical sequence of grammar rules might look something like this:

  expr ::= expr PLUS expr.
  expr ::= expr TIMES expr.
  expr ::= LPAREN expr RPAREN.
  expr ::= VALUE.

There is one non-terminal in this example, "expr", and five terminal symbols or tokens: "PLUS", "TIMES", "LPAREN", "RPAREN" and "VALUE".

Like yacc and bison, Lemon allows the grammar to specify a block of C code that will be executed whenever a grammar rule is reduced by the parser. In Lemon, this action is specified by putting the C code (contained within curly braces {...}) immediately after the period that closes the rule. For example:

  expr ::= expr PLUS expr.   { printf("Doing an addition...\n"); }

In order to be useful, grammar actions must normally be linked to their associated grammar rules. In yacc and bison, this is accomplished by embedding a "$$" in the action to stand for the value of the left-hand side of the rule and symbols "$1", "$2", and so forth to stand for the value of the terminal or nonterminal at position 1, 2 and so forth on the right-hand side of the rule. This idea is very powerful, but it is also very error-prone. The single most common source of errors in a yacc or bison grammar is to miscount the number of symbols on the right-hand side of a grammar rule and say "$7" when you really mean "$8".

Lemon avoids the need to count grammar symbols by assigning symbolic names to each symbol in a grammar rule and then using those symbolic names in the action. In yacc or bison, one would write this:

  expr -> expr PLUS expr  { $$ = $1 + $3; };
But in Lemon, the same rule becomes the following:
  expr(A) ::= expr(B) PLUS expr(C).  { A = B+C; }
In the Lemon rule, any symbol in parentheses after a grammar rule symbol becomes a place holder for that symbol in the grammar rule. This place holder can then be used in the associated C action to stand for the value of that symbol.

The Lemon notation for linking a grammar rule with its reduce action is superior to yacc/bison on several counts. First, as mentioned above, the Lemon method avoids the need to count grammar symbols. Secondly, if a terminal or nonterminal in a Lemon grammar rule includes a linking symbol in parentheses but that linking symbol is not actually used in the reduce action, then an error message is generated. For example, the rule

  expr(A) ::= expr(B) PLUS expr(C).  { A = B; }
will generate an error because the linking symbol "C" is used in the grammar rule but not in the reduce action.

The Lemon notation for linking grammar rules to reduce actions also facilitates the use of destructors for reclaiming memory allocated by the values of terminals and nonterminals on the right-hand side of a rule.

Precedence Rules

Lemon resolves parsing ambiguities in exactly the same way as yacc and bison. A shift-reduce conflict is resolved in favor of the shift, and a reduce-reduce conflict is resolved by reducing whichever rule comes first in the grammar file.

Just like in yacc and bison, Lemon allows a measure of control over the resolution of paring conflicts using precedence rules. A precedence value can be assigned to any terminal symbol using the %left, %right or %nonassoc directives. Terminal symbols mentioned in earlier directives have a lower precedence that terminal symbols mentioned in later directives. For example:

   %left AND.
   %left OR.
   %nonassoc EQ NE GT GE LT LE.
   %left PLUS MINUS.
   %right EXP NOT.

In the preceding sequence of directives, the AND operator is defined to have the lowest precedence. The OR operator is one precedence level higher. And so forth. Hence, the grammar would attempt to group the ambiguous expression

     a AND b OR c
like this
     a AND (b OR c).
The associativity (left, right or nonassoc) is used to determine the grouping when the precedence is the same. AND is left-associative in our example, so
     a AND b AND c
is parsed like this
     (a AND b) AND c.
The EXP operator is right-associative, though, so
     a EXP b EXP c
is parsed like this
     a EXP (b EXP c).
The nonassoc precedence is used for non-associative operators. So
     a EQ b EQ c
is an error.

The precedence of non-terminals is transferred to rules as follows: The precedence of a grammar rule is equal to the precedence of the left-most terminal symbol in the rule for which a precedence is defined. This is normally what you want, but in those cases where you want to precedence of a grammar rule to be something different, you can specify an alternative precedence symbol by putting the symbol in square braces after the period at the end of the rule and before any C-code. For example:

   expr = MINUS expr.  [NOT]

This rule has a precedence equal to that of the NOT symbol, not the MINUS symbol as would have been the case by default.

With the knowledge of how precedence is assigned to terminal symbols and individual grammar rules, we can now explain precisely how parsing conflicts are resolved in Lemon. Shift-reduce conflicts are resolved as follows:

Reduce-reduce conflicts are resolved this way:

Special Directives

The input grammar to Lemon consists of grammar rules and special directives. We've described all the grammar rules, so now we'll talk about the special directives.

Directives in lemon can occur in any order. You can put them before the grammar rules, or after the grammar rules, or in the mist of the grammar rules. It doesn't matter. The relative order of directives used to assign precedence to terminals is important, but other than that, the order of directives in Lemon is arbitrary.

Lemon supports the following special directives:

Each of these directives will be described separately in the following sections:

The %code directive

The %code directive is used to specify addition C code that is added to the end of the main output file. This is similar to the %include directive except that %include is inserted at the beginning of the main output file.

%code is typically used to include some action routines or perhaps a tokenizer or even the "main()" function as part of the output file.

The %default_destructor directive

The %default_destructor directive specifies a destructor to use for non-terminals that do not have their own destructor specified by a separate %destructor directive. See the documentation on the %destructor directive below for additional information.

In some grammers, many different non-terminal symbols have the same datatype and hence the same destructor. This directive is a convenience way to specify the same destructor for all those non-terminals using a single statement.

The %default_type directive

The %default_type directive specifies the datatype of non-terminal symbols that do no have their own datatype defined using a separate %type directive.

The %destructor directive

The %destructor directive is used to specify a destructor for a non-terminal symbol. (See also the %token_destructor directive which is used to specify a destructor for terminal symbols.)

A non-terminal's destructor is called to dispose of the non-terminal's value whenever the non-terminal is popped from the stack. This includes all of the following circumstances:

The destructor can do whatever it wants with the value of the non-terminal, but its design is to deallocate memory or other resources held by that non-terminal.

Consider an example:

   %type nt {void*}
   %destructor nt { free($$); }
   nt(A) ::= ID NUM.   { A = malloc( 100 ); }
This example is a bit contrived but it serves to illustrate how destructors work. The example shows a non-terminal named "nt" that holds values of type "void*". When the rule for an "nt" reduces, it sets the value of the non-terminal to space obtained from malloc(). Later, when the nt non-terminal is popped from the stack, the destructor will fire and call free() on this malloced space, thus avoiding a memory leak. (Note that the symbol "$$" in the destructor code is replaced by the value of the non-terminal.)

It is important to note that the value of a non-terminal is passed to the destructor whenever the non-terminal is removed from the stack, unless the non-terminal is used in a C-code action. If the non-terminal is used by C-code, then it is assumed that the C-code will take care of destroying it. More commonly, the value is used to build some larger structure and we don't want to destroy it, which is why the destructor is not called in this circumstance.

Destructors help avoid memory leaks by automatically freeing allocated objects when they go out of scope. To do the same using yacc or bison is much more difficult.

The %extra_argument directive

The %extra_argument directive instructs Lemon to add a 4th parameter to the parameter list of the Parse() function it generates. Lemon doesn't do anything itself with this extra argument, but it does make the argument available to C-code action routines, destructors, and so forth. For example, if the grammar file contains:

    %extra_argument { MyStruct *pAbc }

Then the Parse() function generated will have an 4th parameter of type "MyStruct*" and all action routines will have access to a variable named "pAbc" that is the value of the 4th parameter in the most recent call to Parse().

The %fallback directive

The %fallback directive specifies an alternative meaning for one or more tokens. The alternative meaning is tried if the original token would have generated a syntax error.

The %fallback directive was added to support robust parsing of SQL syntax in SQLite. The SQL language contains a large assortment of keywords, each of which appears as a different token to the language parser. SQL contains so many keywords, that it can be difficult for programmers to keep up with them all. Programmers will, therefore, sometimes mistakenly use an obscure language keyword for an identifier. The %fallback directive provides a mechanism to tell the parser: "If you are unable to parse this keyword, try treating it as an identifier instead."

The syntax of %fallback is as follows:

%fallback ID TOKEN... .

In words, the %fallback directive is followed by a list of token names terminated by a period. The first token name is the fallback token - the token to which all the other tokens fall back to. The second and subsequent arguments are tokens which fall back to the token identified by the first argument.

The %ifdef, %ifndef, and %endif directives.

The %ifdef, %ifndef, and %endif directives are similar to #ifdef, #ifndef, and #endif in the C-preprocessor, just not as general. Each of these directives must begin at the left margin. No whitespace is allowed between the "%" and the directive name.

Grammar text in between "%ifdef MACRO" and the next nested "%endif" is ignored unless the "-DMACRO" command-line option is used. Grammar text betwen "%ifndef MACRO" and the next nested "%endif" is included except when the "-DMACRO" command-line option is used.

Note that the argument to %ifdef and %ifndef must be a single preprocessor symbol name, not a general expression. There is no "%else" directive.

The %include directive

The %include directive specifies C code that is included at the top of the generated parser. You can include any text you want -- the Lemon parser generator copies it blindly. If you have multiple %include directives in your grammar file, their values are concatenated so that all %include code ultimately appears near the top of the generated parser, in the same order as it appeared in the grammer.

The %include directive is very handy for getting some extra #include preprocessor statements at the beginning of the generated parser. For example:

   %include {#include <unistd.h>}

This might be needed, for example, if some of the C actions in the grammar call functions that are prototyed in unistd.h.

The %left directive

The %left directive is used (along with the %right and %nonassoc directives) to declare precedences of terminal symbols. Every terminal symbol whose name appears after a %left directive but before the next period (".") is given the same left-associative precedence value. Subsequent %left directives have higher precedence. For example:

   %left AND.
   %left OR.
   %nonassoc EQ NE GT GE LT LE.
   %left PLUS MINUS.
   %right EXP NOT.

Note the period that terminates each %left, %right or %nonassoc directive.

LALR(1) grammars can get into a situation where they require a large amount of stack space if you make heavy use or right-associative operators. For this reason, it is recommended that you use %left rather than %right whenever possible.

The %name directive

By default, the functions generated by Lemon all begin with the five-character string "Parse". You can change this string to something different using the %name directive. For instance:

   %name Abcde

Putting this directive in the grammar file will cause Lemon to generate functions named

The %name directive allows you to generator two or more different parsers and link them all into the same executable.

The %nonassoc directive

This directive is used to assign non-associative precedence to one or more terminal symbols. See the section on precedence rules or on the %left directive for additional information.

The %parse_accept directive

The %parse_accept directive specifies a block of C code that is executed whenever the parser accepts its input string. To "accept" an input string means that the parser was able to process all tokens without error.

For example:

   %parse_accept {
      printf("parsing complete!\n");

The %parse_failure directive

The %parse_failure directive specifies a block of C code that is executed whenever the parser fails complete. This code is not executed until the parser has tried and failed to resolve an input error using is usual error recovery strategy. The routine is only invoked when parsing is unable to continue.

   %parse_failure {
     fprintf(stderr,"Giving up.  Parser is hopelessly lost...\n");

The %right directive

This directive is used to assign right-associative precedence to one or more terminal symbols. See the section on precedence rules or on the %left directive for additional information.

The %stack_overflow directive

The %stack_overflow directive specifies a block of C code that is executed if the parser's internal stack ever overflows. Typically this just prints an error message. After a stack overflow, the parser will be unable to continue and must be reset.

   %stack_overflow {
     fprintf(stderr,"Giving up.  Parser stack overflow\n");

You can help prevent parser stack overflows by avoiding the use of right recursion and right-precedence operators in your grammar. Use left recursion and and left-precedence operators instead, to encourage rules to reduce sooner and keep the stack size down. For example, do rules like this:

   list ::= list element.      // left-recursion.  Good!
   list ::= .
Not like this:
   list ::= element list.      // right-recursion.  Bad!
   list ::= .

The %stack_size directive

If stack overflow is a problem and you can't resolve the trouble by using left-recursion, then you might want to increase the size of the parser's stack using this directive. Put an positive integer after the %stack_size directive and Lemon will generate a parse with a stack of the requested size. The default value is 100.

   %stack_size 2000

The %start_symbol directive

By default, the start-symbol for the grammar that Lemon generates is the first non-terminal that appears in the grammar file. But you can choose a different start-symbol using the %start_symbol directive.

   %start_symbol  prog

The %token_destructor directive

The %destructor directive assigns a destructor to a non-terminal symbol. (See the description of the %destructor directive above.) This directive does the same thing for all terminal symbols.

Unlike non-terminal symbols which may each have a different data type for their values, terminals all use the same data type (defined by the %token_type directive) and so they use a common destructor. Other than that, the token destructor works just like the non-terminal destructors.

The %token_prefix directive

Lemon generates #defines that assign small integer constants to each terminal symbol in the grammar. If desired, Lemon will add a prefix specified by this directive to each of the #defines it generates. So if the default output of Lemon looked like this:

    #define AND              1
    #define MINUS            2
    #define OR               3
    #define PLUS             4
You can insert a statement into the grammar like this:
    %token_prefix    TOKEN_
to cause Lemon to produce these symbols instead:
    #define TOKEN_AND        1
    #define TOKEN_MINUS      2
    #define TOKEN_OR         3
    #define TOKEN_PLUS       4

The %token_type and %type directives

These directives are used to specify the data types for values on the parser's stack associated with terminal and non-terminal symbols. The values of all terminal symbols must be of the same type. This turns out to be the same data type as the 3rd parameter to the Parse() function generated by Lemon. Typically, you will make the value of a terminal symbol by a pointer to some kind of token structure. Like this:

   %token_type    {Token*}

If the data type of terminals is not specified, the default value is "int".

Non-terminal symbols can each have their own data types. Typically the data type of a non-terminal is a pointer to the root of a parse-tree structure that contains all information about that non-terminal. For example:

   %type   expr  {Expr*}

Each entry on the parser's stack is actually a union containing instances of all data types for every non-terminal and terminal symbol. Lemon will automatically use the correct element of this union depending on what the corresponding non-terminal or terminal symbol is. But the grammar designer should keep in mind that the size of the union will be the size of its largest element. So if you have a single non-terminal whose data type requires 1K of storage, then your 100 entry parser stack will require 100K of heap space. If you are willing and able to pay that price, fine. You just need to know.

The %wildcard directive

The %wildcard directive is followed by a single token name and a period. This directive specifies that the identified token should match any input token.

When the generated parser has the choice of matching an input against the wildcard token and some other token, the other token is always used. The wildcard token is only matched if there are no other alternatives.

Error Processing

After extensive experimentation over several years, it has been discovered that the error recovery strategy used by yacc is about as good as it gets. And so that is what Lemon uses.

When a Lemon-generated parser encounters a syntax error, it first invokes the code specified by the %syntax_error directive, if any. It then enters its error recovery strategy. The error recovery strategy is to begin popping the parsers stack until it enters a state where it is permitted to shift a special non-terminal symbol named "error". It then shifts this non-terminal and continues parsing. But the %syntax_error routine will not be called again until at least three new tokens have been successfully shifted.

If the parser pops its stack until the stack is empty, and it still is unable to shift the error symbol, then the %parse_failed routine is invoked and the parser resets itself to its start state, ready to begin parsing a new file. This is what will happen at the very first syntax error, of course, if there are no instances of the "error" non-terminal in your grammar.