GSP
Quick Navigator

Search Site

Unix VPS
A - Starter
B - Basic
C - Preferred
D - Commercial
MPS - Dedicated
Previous VPSs
* Sign Up! *

Support
Contact Us
Online Help
Handbooks
Domain Status
Man Pages

FAQ
Virtual Servers
Pricing
Billing
Technical

Network
Facilities
Connectivity
Topology Map

Miscellaneous
Server Agreement
Year 2038
Credits
 

USA Flag

 

 

Man Pages
Marpa::XS::Semantics(3) User Contributed Perl Documentation Marpa::XS::Semantics(3)

Marpa::XS::Semantics - How Marpa evaluates parses

    my $grammar = Marpa::XS::Grammar->new(
        {   start          => 'Expression',
            actions        => 'My_Actions',
            default_action => 'first_arg',
            rules          => [
                { lhs => 'Expression', rhs => [qw/Term/] },
                { lhs => 'Term',       rhs => [qw/Factor/] },
                { lhs => 'Factor',     rhs => [qw/Number/] },
                { lhs => 'Term', rhs => [qw/Term Add Term/], action => 'do_add' },
                {   lhs    => 'Factor',
                    rhs    => [qw/Factor Multiply Factor/],
                    action => 'do_multiply'
                },
            ],
        }
    );

    sub My_Actions::do_add {
        my ( undef, $t1, undef, $t2 ) = @_;
        return $t1 + $t2;
    }

    sub My_Actions::do_multiply {
        my ( undef, $t1, undef, $t2 ) = @_;
        return $t1 * $t2;
    }

    sub My_Actions::first_arg { shift; return shift; }

    my $value_ref = $recce->value;
    my $value = $value_ref ? ${$value_ref} : 'No Parse';

Marpa's semantics will be familiar to those who have used traditional methods to evaluate parses. A parse is seen as a parse tree. Nodes on the tree are evaluated recursively, bottom-up. Once the values of all its child nodes are known, a parent node is ready to be evaluated. The value of a parse is the value of the top node of the parse tree.

Marpa's semantics operate on the application's original rules and symbols. The internals of the parser are hidden from the semantics. Nodes in Marpa's virtual parse trees are of three kinds:

  • Token Nodes, which have a constant semantics, initialized when the token is read.
  • Rule Nodes, which have a dynamic semantics, based on action names and actions, as described below.
  • Null Nodes, which have a constant semantics, initialized when the recognizer is created.

When a Marpa grammar is created, its dynamic semantics is specified indirectly, as action names. To implement its semantics of action names and actions, Marpa must do three things:
  • Determine the action name.
  • Resolve the action name to an action. An action is a Perl closure.
  • Call the Perl closure to produce the actual result.

An action name and action is also used to create the per-parse-tree variable, as described below. The per-parse-tree variable is a special case, but it is intended to be used as part of the semantics.

For every input token, there is an associated token node. Token nodes are leaf nodes in the parse tree. Tokens always have a token symbol. At lexing time, they can be assigned a token value. If no token value is assigned at lex time, the token value defaults to a Perl "undef".

Nodes which are ancestors of token nodes are called rule nodes. Rule nodes are always associated with a rule. The value of a rule node is computed at Node Evaluation Time. Applications can specify, on a per-rule basis, Perl closures to evaluate rule nodes. The Perl closures which produce the value of rule nodes are called value actions.

A value action's arguments will be a per-parse-tree variable followed by the values of its child nodes in lexical order. The return value of a value action becomes the value of the node. A value action is always evaluated in scalar context. If there is no value action for a rule node, the value of the rule node is a Perl "undef".

Some rules are sequence rules. Sequence rule nodes are also rule nodes. Everything said above about rule nodes applies to sequence rule nodes. Specifically, the arguments to the value actions for sequence rules are the per-parse-tree variable followed by the values of the child nodes in lexical order.

The difference (and it is a big one) is that in an ordinary rule, the right hand side is fixed in length, and that length is known when you are writing the code for the value action. In a sequence rule, the number of right hand side symbols is not known until node evaluation time. The value action of a sequence rule node must be a Perl closure capable of dealing with a variable number of arguments.

Sequence semantics work best when every child node in the sequence has the same semantics. When that is not the case, writing the sequence using ordinary non-sequence rules should be considered as an alternative.

By default, if a sequence rule has separators, the separators are thrown away before the value action is called. This means that separators do not appear in the @_ array of the Perl closure which is the value action. If the value of the "keep" rule property is a Perl true value, separators are kept, and do appear in the value action's @_ array.

A null node is a node which derives the zero-length, or empty string. By default, the value of a null node is a Perl "undef". This is adequate for many applications, but Marpa allows other values to be specified for null nodes. In fact, Marpa allows allows an arbitrarily complex null semantics. Readers interested in null nodes with values other than "undef", or who would like to read a more detailed account of how Marpa's null semantics works, should turn to the document on null semantics.

Most applications will find that the order in which Marpa executes its semantics "just works". This section describes that order in detail. These details can matter in some applications, for example, those which exploit side effects.

When the semantics are applied to a parse tree, it produces a value called a parse result. Because Marpa allows ambiguous parsing, each parse can produce a parse series -- a series of zero or more parse trees, each with its own parse result. The first call to the the recognizer's "value" method after the recognizer is created is the start of the first parse series. The first parse series continues until there is a call to the the "reset_evaluation" method or until the recognizer is destroyed. Usually, an application is only interested in a single parse series.

When the "reset_evaluation" method is called for a recognizer, it begins a new parse series. The new parse series continues until there is another call to the the "reset_evaluation" method, or until the recognizer is destroyed.

While processing a recognizer, we have
  • A Recognizer Setup Phase, which occurs during the call of a recognizer's "new" method. It is followed by
  • the processing of zero or more parse series.

While processing a parse series, we have:

  • A Series Setup Phase, which occurs during the first call of the recognizer's "value" method for that series. It is followed by
  • the processing of zero or more parse trees.

While processing a parse tree, we have:

  • A Tree Setup Phase, which occurs during the call of the recognizer's "value" method for that parse tree. It is followed by
  • a Tree Traveral Phase.

Node Evaluation Time is the Tree Traversal Phase, as seen from the point of view of each rule node. It is not a separate phase.

In the Recognizer Setup Phase, the null values of the symbols are determined. The null values of symbols never change -- they are in effect properties of the grammar.

During the Series Setup Phase all value action names are resolved to Perl closures -- the value actions. The value actions are never called in the Series Setup Phase. Value actions are called in the Tree Traversal Phase. Also during the Series Setup Phase, the logic which ranks parse trees is executed.

In the Tree Setup Phase, the per-parse-tree variable is created. If a constructor was found for the "action_object", it is run at this point, and the per-parse-tree variable is its return value. Exactly one Tree Setup Phase occurs for each parse tree.

During the Tree Traversal Phase, the value actions are called. Node Evaluation Time is the Tree Traversal Phase, as seen from the point of view of the individual nodes of the parse tree. If a node has a value action, it is called at Node Evaluation Time.

Marpa finds the action for each rule based on rule and symbol properties and on the grammar named arguments. Specifically, Marpa attempts the following, in order:
  • Resolving an action based on the "action" property of the rule, if one is defined.
  • Resolving an action based on the "lhs" property of the rule.
  • Resolving an action based on the "default_action" named argument of the grammar, if one is defined.
  • Defaulting to a virtual action, one which returns a Perl "undef".

Resolution of action names is described below. If the "action" property or the "default_action" named argument is defined, but does not resolve successfully, Marpa throws an exception. Marpa prefers to "fast fail" in these cases, because they often indicate a mistake in writing the application.

Action names come from these sources:
  • The "default_action" named argument of Marpa's grammar.
  • The "action" property of Marpa's rules.
  • The "new" constructor in the package specified by the "action_object" named argument of the Marpa grammar.
  • The "lhs" property of Marpa's rules.

The recognizer's "closures" named argument allows the user to directly control the mapping from action names to actions. The value of the "closures" named argument is a reference to a hash whose keys are action names and whose hash values are CODE refs.

If an action name is the key of an entry in the "closures" hash, it resolves to the closure referenced by the value part of that hash entry. Resolution via the "closures" named argument is called explicit resolution.

When explicit resolution is the only kind of resolution that is wanted, it is best to pick a name that is very unlikely to be the name of a Perl closure. Many of Marpa::HTML's action names are intended for explicit resolution only. In Marpa::HTML those action names begin with an exclamation mark ("!"), and that convention is recommended.

If explicit resolution fails, Marpa transforms the action name into a fully qualified Perl name. An action name that contains a double colon (""::"") or a single quote (""'"") is considered to be a fully qualified name. Any other action name is considered to be a bare action name.

If the action name to be resolved is already a fully qualified name, it is not further transformed. It will be resolved in the form it was received, or not at all.

For bare action names, Marpa tries to qualify them by adding a package name. If the "actions" grammar named argument is defined, Marpa uses it as the package name. Otherwise, if the "action_object" grammar named argument is defined, Marpa uses it as the package name. Once Marpa has fully qualified the action name, Marpa looks for a Perl closure with that name.

Marpa will not attempt to resolve an action name that it cannot fully qualify. This implies that, for an action name to resolve successfully, one of these four things must be the case:

  • The action name resolves explicitly.
  • The action name is fully qualified to begin with.
  • The "actions" named argument is defined.
  • The "action_object" named argument is defined.

In all but one circumstance, failure to resolve an action name is thrown as an exception. Marpa is more lenient when a rule attempts to use the "lhs" rule property as an action name. That is the one case in which Marpa will look at other alternatives.

Marpa's philosophy is to require that the programmer be specific about action names. This can be an inconvenience, but Marpa prefers this to silently executing unintended code.

Generally it is a good practice to keep the semantic Perl closures in their own namespace. But if, for example, the user wants to leave the semantic closures in the "main" namespace, she can specify "main" as the value of the "actions" named argument.

In the Tree Setup Phase, Marpa creates a per-parse-tree variable. This becomes the first argument of the semantic Perl closures for the rule nodes. If the grammar's "action_object" named argument is not defined, the per-parse-tree variable is initialized to an empty hash ref.

Most data for the value actions of the rules will be passed up the parse tree. The actions will see the values of the rule node's child nodes as arguments, and will return their own value to be seen as an argument by their parent node. The per-parse-tree variable can be used for data which does not conveniently fit this model.

The lifetime of the per-parse-tree variable extends into the Tree Traversal Phase. During the Tree Traversal Phase, Marpa's internals never alter the per-parse-tree variable -- it is reserved for use by the application.

If the grammar's "action_object" named argument has a defined value, that value is treated as the name of a class. The action object constructor is the "new" method in the "action_object" class.

The action object constructor is called in the Tree Setup Phase. The return value of the action object constructor becomes the per-parse-tree variable. It is a fatal error if the grammar's "action_object" named argument is defined, but does not name a class with a "new" method.

Resolution of the action object constructor is resolution of the action object constructor name. The action object constructor name is formed by concatenating the literal string ""::new"" to the value of the "action_object" named argument.

All standard rules apply when resolving the action object constructor name. In particular, bypass via explicit resolution applies to the action object constructor. If the action object constructor name is a hash key in the evaluator's "closures" named argument, then the value referred to by that hash entry becomes the action object constructor.

If a grammar has both the "action" and the "action_object" named arguments defined, all action names except for the action object constructor will be resolved in the "action" package or not at all. The action object constructor is always in the "action_object" class, if it is anywhere.

If a parse is ambiguous, all parses are returned, with no duplication. By default, the order is arbitrary, but it is also possible to control the order. Details are in the document on parse order.

Marpa allows grammars with infinite loops. These are universally considered useless in practical applications. Marpa supports all the semantics for these that should be necessary and more. Because it can handle infinite loops, Marpa can accurately claim to support every grammar that can be written in BNF.

Marpa applies semantics to infinite loops by breaking the loop at some convenient point, so that an infinite regress is prevented. The exact location at which the loop is broken varies among Marpa implementations. If an infinite loop is given a null semantics, the location of the break will not matter. A null semantics is the default.

More could be done. In particular, a precise definition of where loops are broken would allow applications to depend on the details of the structure of infinite loops. But practical interest in this seems non-existent. For those who want to know more, the details are in a separate document.

  Copyright 2012 Jeffrey Kegler
  This file is part of Marpa::XS.  Marpa::XS is free software: you can
  redistribute it and/or modify it under the terms of the GNU Lesser
  General Public License as published by the Free Software Foundation,
  either version 3 of the License, or (at your option) any later version.
  
  Marpa::XS is distributed in the hope that it will be useful,
  but WITHOUT ANY WARRANTY; without even the implied warranty of
  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  Lesser General Public License for more details.
  
  You should have received a copy of the GNU Lesser
  General Public License along with Marpa::XS.  If not, see
  http://www.gnu.org/licenses/.
2022-04-12 perl v5.32.1

Search for    or go to Top of page |  Section 3 |  Main Index

Powered by GSP Visit the GSP FreeBSD Man Page Interface.
Output converted with ManDoc.