A Scannerless GLR parser/parser generator

Details

DParser is an simple but powerful tool for parsing. You can specify the form of the text to be parsed using a combination of regular expressions and grammar productions. Because of the parsing technique (technically a scannerless GLR parser based on the Tomita algorithm) there are no restrictions. The grammar can be ambiguous, right or left recursive, have any number of null productions, and because there is no separate tokenizer, can include whitespace in terminals and have terminals which are prefixes of other terminals. DParser handles not just well formed computer languages and data files, but just about any wacky situation that occurs in the real world.

Features

Powerful GLR parsing
Simple EBNF-style grammars and regular expression terminals
Priorities and associativities for token and rules
Built-in error recovery
Speculative actions (for semantic disambiguation)
Auto-building of parse tree (optionally)
Final actions as you go, or on the complete parse tree
Tree walkers and default actions (multi-pass compilation support)
Symbol table built for ambiguous parsing
Partial parses, recursive parsing, parsing starting with any non-terminal
Whitespace can be specified as a subgrammar
External (C call interface) tokenizers and external terminal scanners
Good asymptotically efficiency
Comes with ANSI-C, Python and Verilog grammars
Comes with full source
Portable C for easy compilation and linking
BSD license, so you can included it in your application without worrying about licensing

The result is natural grammars and powerful parsing.

The R based tree parsing in dparse creates dlls on the fly based on C code. By default the echoing of the compile is disabled, but you can change this by options(dparser.echo.compile=FALSE)

Garbage collection

There are two user options that control if the dlls for the grammars created by dparser will be deleted upon garbage collection or R exit if they are not associated with any active objects. These are:

dpaser.rm.unnamed.parser.dll:: when TRUE, this remove parsers that are created from strings, or other memory-based items in R.
dpaser.rm.unnamed.parser.dll:: when TRUE, this removes parsers created from grammar files.

Creating a Grammar for Parsing

Grammar Comments:

Grammars can include C/C++ style comments

Example:


// My first grammar
E: E '+' E | "[abc]";
/* is this right? */

Grammar Productions:

A production is the parts of your language your are trying to parse and are tpyically named. See https://en.wikipedia.org/wiki/Top-down_parsing

The first production is the root of your grammar (what you will be trying to parse).
Productions start with the non-terminal being defined followed by a colon ':', a set of right hand sides separated by '|' (or) consisting of elements (non-terminals or terminals).
Elements can be grouped with parens '(', and the normal regular expression symbols can be used ('+' '*' '?' '|').
Elements can be grouped with parens '(', and the normal regular expression symbols can be used ('+' '*' '?' '|').
Elements can be repeated using '@', for example elem@3 or elem@1:3 for repeat 3 or between 1 and 3 times respectively.

Example:


 program: statements+ |  comment* (function |  procedure)?;

Note: Instead of using '[' ']' for optional elements we use the more familar and consistent '?' operator. The square brackets are reserved for speculative actions (below).

Global C code in Grammars:

Since the main parsing of the language grammar is in C, intermixing C code with the grammar can be useful.

Global (or static) C code can be intermixed with productions by surrounding the code with brackets '{}'.

Example:


  { void dr_s() { printf("Dr. S\n"); }
   S: 'the' 'cat' 'and' 'the' 'hat' { dr_s(); } | T;
   { void twain() { printf("Mark Twain\n"); }
       T: 'Huck' 'Finn' { twain(); };

Note: When parsing the grammar using mkdparse, the option use_r_header = TRUE will redefine printf to Rprintf to better comply with R packages.

Terminals

The terminals are the peices of the language that are being parsed, like language keywords.

Strings terminals are surrounded with single quotes. For example:


block: '{' statements* '}';
whileblock: 'while' '(' expression ')' block;

Unicode literals can appear in strings or as charaters with U+ or u+. For example:


U+03c9 { printf("omega\n"); }

Priorities and Associativity:

Priorities can very from MININT to MAXINT and are specified as integers. Associativity can take the values:


assoc : '$unary_op_right' | '$unary_op_left' | '$binary_op_right'
| '$binary_op_left' | '$unary_right' | '$unary_left'
| '$binary_right' | '$binary_left' | '$right' | '$left' ;

Token Prioritites

Termininal priorities apply after the set of matching strings has been found and the terminal(s) with the highest priority is selected.

Terminal priorities are introduced after a terminal by the specifier $term. We saw an example of token priorities with the definition of ident.

Example:


S : 'if' '(' S ')' S ';' | 'do' S 'while' '(' S ')' ';' | ident;
ident: "[a-z]+" $term -1;

Operator Priorities

Operator priorities specify the priority of a operator symbol (either a terminal or a non-terminal). This corresponds to the yacc or bison doesn't require a global tokenizer, operator priorities and associativities are specified on the reduction which creates the token. Moreover, the associativity includes the operator usage as well since it cannot be infered from rule context. Possible operator associativies are:


operator_assoc : '$unary_op_right' | '$unary_op_left' | '$binary_op_right'
                    | '$binary_op_left' | '$unary_right' | '$unary_left'
                    | '$binary_right' | '$binary_left';

Example:


E: ident op ident;
ident: '[a-z]+';
op: '*' $binary_op_left 2 |
    '+' $binary_op_left 1;

Rule Priorities

Rule priorities specify the priority of the reduction itself and have the possible associativies:


rule_assoc: '$right' | '$left';

Rule and operator priorities can be intermixed and are interpreted at run time (not when the tables are built). This make it possible for user-defined scanners to return the associativities and priorities of tokens.

Actions:

Actions are the bits of code which run when a reduction occurs. Example


 S: this | that;
 this: 'this' { printf("got this\n"); };
 that: 'that' { printf("got that\n"); };

Speculative Action

Speculative actions occur when the reduction takes place during the speculative parsing process. It is possible that the reduction will not be part of the final parse or that it will occur a different number of times. For example:


S: this | that;
this: hi 'mom';
that: ho 'dad';
ho: 'hello' [ printf("ho\n"); ];
hi: 'hello' [ printf("hi\n"); ];

Will print both 'hi' and 'ho' when given the input 'hello dad' because at the time hello is reduced, the following token is not known.

Final Actions

Final actions occur only when the reduction must be part of any legal final parse (committed). It is possible to do final actions during parsing or delay them till the entire parse tree is constructed (see Options). Final actions are executed in order and in number according the the single final unambiguous parse.


 S: A S 'b' | 'x';
 A: [ printf("speculative e-reduce A\n"); ]
    { printf("final e-reduce A\n"); };

On input:


 xbbb

Will produce:


 speculative e-reduce A
 final e-reduce A
 final e-reduce A
 final e-reduce A

Embedded Actions

Actions can be embedded into rule. These actions are executed as if they were replaced with a synthetic production with a single null rule containing the actions. For example:


 S: A { printf("X"); } B;
 A: 'a' { printf("a"); };
 B: 'b' { printf("b"); };

On input:

ab

Will produce:

aXb

Note that in the above example, the print("X") is evaluated in a context null rule context while in:


 S: A (A B { printf("X"); }) B;

The print is evalated in the context of the "A B" subrule because it appears at the end of the subrule and is therefor treated as a normal action for the subrule.

Pass Actions

DParser supports multiple pass compilation. The passes are declared at the top of the grammar, and the actions are associated with individual rules.

Example;


 ${pass sym for_all postorder}
 ${pass gen for_all postorder}

 translation_unit: statement*;

 statement
   : expression ';' {
     d_pass(${parser}, &$n, ${pass sym});
     d_pass(${parser}, &$n, ${pass gen});
   }
   ;

 expression :  integer
   gen: { printf("gen integer\n"); }
   sym: { printf("sym integer\n"); }
   | expression '+' expression $right 2
   sym: { printf("sym +\n"); }
   ;

A pass name then a colon indicate that the following action is associated with a particular pass. Passes can be either for_all or for_undefined (which means that the automatic traversal only applies to rules without actions defined for this pass). Furthermore, passes can be postorder, preorder, and manual (you have to call d_pass yourself). Passes can be initiated in the final action of any rule.

Default Actions

The special production "_" can be defined with a single rule whose actions become the default when no other action is specified. Default actions can be specified for speculative, final and pass actions and apply to each separately.

Example


 _: { printf("final action"); }
     gen: { printf("default gen action"); }
     sym: { printf("default sym action"); }
     ;

Attributes and Action Specifiers

Each of the language parser can have some global atrributes and actions associated with each part of the parsed code.

Global State ($g)

Global state is declared by define'ing D_ParseNode_Globals (see the ANSI C grammar for a similar declaration for symbols). Global state can be accessed in any action with $g. Because DParser handles ambiguous parsing global state can be accessed on different speculative parses. In the future automatic splitting of global state may be implemented (if there is demand). Currently, the global state can be copied and assigned to $g to ensure that the changes made only effect subsequent speculative parses derived from the particular parse.

Example


  [ $g = copy_globals($g);
  $g->my_variable = 1;
  ]

The symbol table (see below) can be used to manage state information safely for different speculative parses.

Parse Node State

Each parse node includes a set of system state variables and can have a set of user-defined state variables. User defined parse node state is declared by define'ing D_ParseNodeUser. The size of the parse node state must be passed into new_D_Parser() to ensure that the appropriate amount of space is allocated for parse nodes. Parse node state is accessed with:

$#: number of child nodes
$$: user parse node state for parent node (non-terminal defined by the production)
$X (where X is a number): the user parse node state of element X of the production
$n: the system parse node state of the rule node
$nX: the system parse node state of element X of the production

The system parse node state is defined in dparse.h which is installed with DParser. It contains such information as the symbol, the location of the parsed string, and pointers to the start and end of the parsed string.

Misc

${scope}: the current symbol table scope
${reject}: in speculative actions permits the current parse to be rejected

Symbol Table

The symbol table can be updated down different speculative paths while sharing the bulk of the data. It defines the following functions in the file (dsymtab.h):


  struct D_Scope *new_D_Scope(struct D_Scope *st);
  struct D_Scope *enter_D_Scope(struct D_Scope *current, struct D_Scope *scope);
  D_Sym *NEW_D_SYM(struct D_Scope *st, char *name, char *end);
  D_Sym *find_D_Sym(struct D_Scope *st, char *name, char *end);
  D_Sym *UPDATE_D_SYM(struct D_Scope *st, D_Sym *sym);
  D_Sym *current_D_Sym(struct D_Scope *st, D_Sym *sym);
  D_Sym *find_D_Sym_in_Scope(struct D_Scope *st, char *name, char *end);

'new_D_Scope' creates a new scope below 'st' or NULL for a 'top level' scope. 'enter_D_Scope' returns to a previous scoping level. NOTE: do not simply assign ${scope} to a previous scope as any updated symbol information will be lost. 'commit_D_Scope' can be used in final actions to compress the update list for the top level scope and improve efficiency.

'find_D_Sym' finds the most current version of a symbol in a given scope. 'UPDATE_D_SYM' updates the value of symbol (creates a difference record on the current speculative parse path). 'current_D_Sym' is used to retrive the current version of a symbol, the pointer to which may have been stored in some other attribute or variable. Symbols with the same name should not be created in the same scope. The function 'find_D_Sym_in_Scope' is provided to detect this case.

User data can be attached to symbols by define'ing D_UserSym. See the ANSI C grammar for an example.

Here is a full example of scope usage (from tests/g29.test.g):



  #include <stdio.h>

  typedef struct My_Sym {
    int value;
  } My_Sym;
  #define D_UserSym My_Sym
  typedef struct My_ParseNode {
    int value;
    struct D_Scope *scope;
  } My_ParseNode;
  #define D_ParseNode_User My_ParseNode
}

ranslation_unit: statement*;

tatement
: expression ';'
{ printf("
| '{' new_scope statement* '}'
[ ${scope} = enter_D_Scope(${scope}, $n0.scope); ]
{ ${scope} = commit_D_Scope(${scope}); }
;

ew_scope: [ ${scope} = new_D_Scope(${scope}); ];

xpression
: identifier ':' expression
[
_Sym *s;
f (find_D_Sym_in_Scope(${scope}, $n0.start_loc.s, $n0.end))
rintf("duplicate identifier line 
 = NEW_D_SYM(${scope}, $n0.start_loc.s, $n0.end);
->user.value = $2.value;
$.value = s->user.value;
]
| identifier '=' expression
[ D_Sym *s = find_D_Sym(${scope}, $n0.start_loc.s, $n0.end);
 = UPDATE_D_SYM(${scope}, s);
->user.value = $2.value;
$.value = s->user.value;
]
| integer
[ $$.value = atoi($n0.start_loc.s); ]
| identifier
[ D_Sym *s = find_D_Sym(${scope}, $n0.start_loc.s, $n0.end);
f (s)
$.value = s->user.value;
]
| expression '+' expression
[ $$.value = $0.value + $1.value; ]
;

nteger: "-?([0-9]|0(x|X))[0-9]*(u|U|b|B|w|W|L|l)*" $term -1;
dentifier: "[a-zA-Z_][a-zA-Z_0-9]*";

Whitespace

Whitespace can be specified two ways: C function which can be user-defined, or as a subgrammar. The default whitespace parser is compatible with C/C++ #line directives and comments. It can be replaced with any user specified function as a parsing option (see Options).

Additionally, if the (optionally) reserved production whitespace is defined, the subgrammar it defines will be used to consume whitespace for the main grammar. This subgrammar can include normal actions.

Example


: 'a' 'b' 'c';
hitespace: "[ \t\n]*";

Whitespace can be accessed on a per parse node basis using the unctions: d_ws_before and d_ws_after, which return the tart of the whitespace before start_loc.s and after end respectively.

Ambiguities

Ambiguities are resolved automatically based on priorities and associativities. In addition, when the other resolution techniques fail, user defined ambiguity resolution is possible. The default ambiguity handler produces a fatal error on an unresolved ambiguity. This behavior can be replaced with a user defined resolvers the signature of which is provided in dparse.h.

If the verbose_level flag is set, the default ambiguity andler will print out parenthesized versions of the ambiguous parse rees. This may be of some assistence in disambiguating a grammar.

Error Recovery

DParser implements an error recovery scheme appropriate to scannerless parsers. I haven't had time to investigate all the prior work in this area, so I am not sure if it is novel. Suffice for now that it is optional and works well with C/C++ like grammars.

Parsing Options

Parser are instantiated with the function new_D_Parser. The resulting data structure contains a number of user configurable options (see dparser.h). These are provided reasonable default values and include:

initial_globals - the initial global variables accessable through $g
initial_skip_space_fn - the initial whitespace function
initial_scope - the initial symbol table scope
syntax_error_fn - the function called on a syntax error
ambiguity_fn - the function called on an unresolved ambiguity
loc - the initial location (set on an error).

In addtion, there are the following user configurables:

sizeof_user_parse_node - the sizeof D_ParseNodeUser
save_parse_tree - whether or not the parse tree should be save once the final actions have been executed
dont_fixup_internal_productions - to not convert the Kleene star into a variable number of children from a tree of reductions
dont_merge_epsilon_trees - to not automatically remove ambiguities which result from trees of epsilon reductions without actions
dont_use_greediness_for_disambiguation - do not use the rule that the longest parse which reduces to the same token should be used to disambiguate parses. This rule is used to handle the case (if then else?) relatively cleanly.
dont_use_height_for_disambiguation - do not use the rule that the least deep parse which reduces to the same token should be used to disabiguate parses. This rule is used to handle recursive grammars relatiively cleanly.
dont_compare_stacks - disables comparing stacks to handle certain exponential cases during ambiguous operator priority resolution.
commit_actions_interval - how often to commit final actions (0 is immediate, MAXINT is essentially not till the end of parsing)
error_recovery - whether or not to use error recovery (defaults ON)

An the following result values:

syntax_errors - how many syntax errors (if error_recovery was on)

This final value should be checked to see if parse was successful.

Author

Maintainer: Matthew Fidler matthew.fidler@gmail.com

Authors:

John Plevyak jplevyak@gmail.com [copyright holder]

Details

Features

Garbage collection

Creating a Grammar for Parsing

See also

Author