- About this document
- Additional reserved symbol names
- The return value of the read() method has changed
- Rule LHS's are no longer a source of action names
- Different rules with the same rank now appear in arbitrary order
- Null actions now come from the rules
- Actions can now be constants
- Actions names beginning with "::" are reserved
- The "default_null_value" named argument for grammars has been removed
- The "null_value" symbol property has been removed
- The token value argument of read() has changed
- The token value argument of alternative() has changed
- Marpa::R2::Recognizer::value() does not accept named arguments
- Marpa's grammar rewriting is now invisible
- The semantics now defaults to "whatever" values
- By default, the non-LHS symbols are the terminals
- The lhs_terminals grammar named argument has been eliminated
- Nulling symbols cannot be terminals
- A sequence must have a unique LHS
- The terminal status of a symbol is locked once set
- Evaluation of infinite loops has been changed
- The range of values allowed for ranks has been clarified
- Copyright and License
Marpa::R2::Changes - Differences between Marpa::R2 and Marpa::XS
This document describes the incompatible differences between Marpa::XS and Marpa::R2. (Differences that do not give rise to incompatibility are outside of its scope.) It is intended for readers already familiar with Marpa::XS, who are writing new applications for Marpa::R2, and for readers migrating Marpa::XS applications and tools to Marpa::R2.
Marpa::XS reserved, for its internal use, all symbol names ending with the right square bracket ("
]"). In addition, Marpa::RS reserved symbols ending with the right parenthesis ("
)"), the right angle bracket ("
>"), and the right curly bracket ("
}"). Any other valid Perl string remains an acceptable symbol name.
The return value of the Marpa::R2 recognizer's
read() method differs from its Marpa::XS equivalent. In Marpa::XS it returned the number of distinct terminals (by symbol ID) allowed in the next
read(). In Marpa::R2 it returns the number of recognizer events that occurred during the read. Examples of recognizer events are exhaustion, the Earley sets exceeding a designated "warning" level, and other circumstances settable by the user. For more detail, see the documentation of recognizer's
In Marpa::XS, if there was no explicit action name for a rule, Marpa would try to find a closure that had the same name as the rule's LHS. The use of rule LHS's as action names had a potential for unpleasant surprises. A surprise could occur if the rule's LHS coincided with a function name without the prorgrammer realizing or intending it. This kind of 'action at a distance' bug can be very hard to detect and trace.
It was originally thought that implicitly using the LHS as the name of an action would be convenient enough to outweigh the dangers. But in fact, this feature wound up being little used. And accidental resolution via a rule LHS was a danger for all users, whether they used the feature or not. For these reasons, as well as potential optimization and efficiency considerations, Marpa::R2 no longer does implicit action resolution using a rule LHS.
In ranking parse trees, if two rule instances are for different rules but have the same rule rank, they will now appear in arbitrary order. This is probably the behavior that programmers have always expected.
In Marpa::XS, when the
null_ranking named argument of rules was in use for one of the rules, specific guarantees were made for the order in some of the cases. The intent was to be orthogonal with the guarantees made for the ranking of null variants within the same rule. These additional guarantees proved useless in practice, cumbersome to implement, and, when documented, opaque and unintuitive. In Marpa::R2 they have been dropped.
In Marpa::XS null actions were specified by symbol. This created a dual semantics -- one for non-nulled rules, and another for nulled rules. The conventions and behaviors of the two semantics were quite dissimilar. The rules for their coordination were complicated, and it was possible for a programmer expecting one semantics, to be surprised by a result from the other.
In Marpa::R2 the semantics of nulled rules is the same as that of non-nulled rules, and the semantics of nulled symbols comes from the semantics of the nulled rules. This requires rule evaluation closures to be aware they might be called for nulled rules. But it greatly simplifies the semantics conceptually. For more detail, see Marpa::R2::Semantics::Null.
If an action name resolves to a constant, that constant is the action. The effect is the same as if the action name resolved to a function that returned that constant, except that it is more efficient.
Perl cannot reliably distinguish between non-existent symbols and symbols whose value is
undef, so constants whose value is
undef are not allowed. The
::undef reserved action name can be used instead.
Action names which start with "
::" are reserved. "
::whatever" explicitly requests "whatever" semantics. "
::undef" is a safe way of specify a constant whose value is
undef. Use of a reserved name which has not yet been defined causes an exception to be thrown.
Symbols no longer have null values, so the "default_null_value" named argument of grammars has been removed.
Symbols no longer have null values. Use of the
null value symbol property now causes an exception.
The Marpa::R2 recognizer's
read() method differs from its Marpa::XS equivalent. In Marpa::R2, If
read()'s token value argument is omitted, then the token value will be a "whatever" value. If
read()'s token value is given explicitly, then that explicit value will be the value of the token. In particular, an explicit
undef token value argument will behave differently from an omitted token value argument. For details, see the documentation of recognizer's
The Marpa::R2 recognizer's
alternative() method differs from its Marpa::XS equivalent. Its token value argument must now be a reference to the token value, not the token value itself, as in Marpa::XS. If alternative's token value argument is omitted or a
undef, then the token has a "whatever" value. If alternative's token value argument is reference to
undef, then the value of the token is a Perl
undef. For details, see the documentation of the
In the Marpa::XS recognizer, the
set() and value() methods all accepted named arguments. As of Marpa::R2, the
value() method will no longer do so.
Allowing named arguments for the
value() was a holdover from a previous interface, which also seemed like it might be a convenience. But, since it was even more important that the
value() method be convenient as the termination test controlling a loop over the parse results, a lot of special logic was added to deal with arguments which only made sense before the first pass of the loop, etc., etc.
Eliminating named arguments from the
value() method eliminates a variety of special cases and, as a result, the documentation of the
value() method is now simpler, shorter and clearer. Anything that could be done by providing named arguments to the
value() method can be done more using the recognizer's
set() method, and the code will be clearer for it.
Internally, Marpa rewrites its grammars. In Marpa::XS, most details of these rewrites were invisible, but not all. In Marpa::R2, all internal rules and symbols are now completely invisible to the user, even in the tools for debugging grammars.
In Marpa::XS, the default value for rules, null values, and token values, was a Perl
undef. In Marpa::R2, rules, null values and token values now default to a "whatever" value. A "whatever" value is guaranteed to be one of the Perl values which is assignable to a scalar. Otherwise, its value is unspecified. "Whatever" values will usually only be appropriate when the application simply does not care what the value is.
The motivation for the change is efficiency. When Marpa::R2 knows that a value on its evaluation stack is a "whatever" value, it implements the logic to create it as a no-op. Real applications allow "whatever" values surprisingly often. According to the author's sense of the typical application mix, this one change makes Marpa::R2 20% faster than Marpa::XS.
Traditionally, a symbol has been a terminal if it is not on the LHS of any rule, and vice versa. This is now the default in Marpa::R2, replacing the more complicated, and less intuitive, scheme that was in Marpa::XS. Marpa::R2 still allows the user to use any non-nulling symbol as a terminal, including those symbols that appear on the LHS of a rule, but this is now an option, and never the default. For more, see "Terminal symbols" in Marpa::R2::Grammar.
The lhs_terminals named argument of grammar objects implemented what is now the default behavior. Since it no longer performs a function, its use is now a fatal error.
In Marpa::XS, it was possible for a symbol to be both nulling and a terminal. In practice that meant that the symbol was nulling, but that, on input, that property could be overriden, and a specific instance of the nulling symbol could be made non-nulling. This behavior was worse than useless and non-intuitive -- it was dangerous and logically inconsistent.
Marpa::R2 will not allow a nulling symbol to be used as a terminal. To the extent that the Marpa::XS behavior made sense, it can be duplicated by creating a symbol which is the LHS of two rules, one empty, and the other rule with a RHS consisting of exactly one terminal symbol.
The LHS of a sequence rule may not be on the LHS of any other rule, whether another sequence rule, or a BNF rule. This is not as severe a restriction as it might sound -- while sequences cannot share the same LHS with other rules directly, they can do so indirectly. For details, see "Duplicate rules" in Marpa::R2::Grammar.
In Marpa::XS, the definition of when a sequence was a duplicate was more liberal, but it was also complicated and non-intuitive. The new definition is simpler and more intuitive, and its greater restrictiveness is easy to work around.
Once a symbol is marked as a terminal or a non-terminal, its terminal status cannot be changed. We doubt this will affect any actual applications. It would only affect an application that changes symbols from their default status to non-terminal, and then only if they attempted to mark the same symbol as a terminal at another point. Few Marpa::R2 applications change symbols from their default terminal status, and none to my knowledge mark symbols as non-terminals.
Infinite loops (cycles) are still, by default, fatal errors. For those considering programming with them, and evaluating parses from grammars with cycles, the semantics of cycles is now more closely specified. For details of the new semantics, see Marpa::R2::Semantics::Infinite.
Symbols and rules have numeric ranks. Previously, no mention was made of range of values allowed. This is implemented-defined, except that the minimum magnitudes will always be the 28th power of 2, less 1. That is, numbers in the range between -134,217,727 and 134,217,727 will always be allowed as ranks.
Copyright 2012 Jeffrey Kegler This file is part of Marpa::R2. Marpa::R2 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::R2 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::R2. If not, see http://www.gnu.org/licenses/.