Techniques

Templatized Functors

For the sake of genericity, it is often better to make the functor's member operator() a template. That way, we do not have to concern ourselves with the type of the argument to expect as long as the behavior is appropriate. For instance, rather than hard-coding char const* as the argument of a generic semantic action, it is better to make it a template member function. That way, it can accept any type of iterator:

    struct my_functor
    {
        template <typename IteratorT>
        void operator()(IteratorT first, IteratorT last) const;
    };

Take note that this is only possible with functors. It is not possible to pass in template functions as semantic actions unless you cast it to the correct function signature; in which case, you monomorphize the function. This clearly shows that functors are superior to plain functions.

Rule With Multiple Scanners

As of v1.8.0, rules can use one or more scanner types. There are cases, for instance, where we need a rule that can work on the phrase and character levels. Rule/scanner mismatch has been a source of confusion and is the no. 1 FAQ. To address this issue, we now have multiple scanner support.

Here is an example of a grammar with a rule r that can be called with 3 types of scanners (phrase-level, lexeme, and lower-case). See the rule, grammar, lexeme_scanner and as_lower_scanner for more information.

Here's the grammar (see multiple_scanners.cpp):

    struct my_grammar : grammar<my_grammar>
    {
        template <typename ScannerT>
        struct definition
        {
            definition(my_grammar const& self)
            {
                r = lower_p;
                rr = +(lexeme_d[r] >> as_lower_d[r] >> r);
            }

            typedef scanner_list<
                ScannerT
              , typename lexeme_scanner<ScannerT>::type
              , typename as_lower_scanner<ScannerT>::type
            > scanners;

            rule<scanners> r;
            rule<ScannerT> rr;
            rule<ScannerT> const& start() const { return rr; }
        };
    };

By default support for multiple scanners is disabled. The macro BOOST_SPIRIT_RULE_SCANNERTYPE_LIMIT must be defined to the maximum number of scanners allowed in a scanner_list. The value must be greater than 1 to enable multiple scanners. Given the example above, to define a limit of three scanners for the list, the following line must be inserted into the source file before the inclusion of Spirit headers:

    #define BOOST_SPIRIT_RULE_SCANNERTYPE_LIMIT 3

Look Ma' No Rules

You use grammars and you use lots of 'em? Want a fly-weight, no-cholesterol, super-optimized grammar? Read on...

I have a love-hate relationship with rules. I guess you know the reasons why. A lot of problems stem from the limitation of rules. Dynamic polymorphism and static polymorphism in C++ do not mix well. There is no notion of virtual template functions in C++; at least not just yet. Thus, the rule is tied to a specific scanner type. This results in problems such as the scanner business, our no. 1 FAQ. Apart from that, the virtual functions in rules slow down parsing, kill all meta-information, and kills inlining, hence bloating the generated code, especially for very tiny rules such as:

    r = ch_p('x') >> uint_p;

The rule's limitation is the main reason why the grammar is designed the way it is now, with a nested template definition class. The rule's limitation is also the reason why subrules exists. But do we really need rules? Of course! Before C++ adopts some sort of auto-type deduction, such as that proposed by David Abrahams in clc++m:

    auto r = ...definition ...

we are tied to the rule as RHS placeholders. However.... in some occasions we can get by without rules! For instance, rather than writing:

    rule<> x = ch_p('x');

It's better to write:

    chlit<> x = ch_p('x');

That's trivial. But what if the rule is rather complicated? Ok, let's proceed stepwise... I'll investigate a simple skip_parser based on the C grammar from Hartmut Kaiser. Basically, the grammar is written as (see no_rule1.cpp):

    struct skip_grammar : grammar<skip_grammar>
    {
        template <typename ScannerT>
        struct definition
        {
            definition(skip_grammar const& /*self*/)
            {
                skip
                    =   space_p
                    |   "//" >> *(anychar_p - '\n') >> '\n'
                    |   "/*" >> *(anychar_p - "*/") >> "*/"
                    ;
            }

            rule<ScannerT> skip;

            rule<ScannerT> const&
            start() const { return skip; }
        };
    };

Ok, so far so good. Can we do better? Well... since there are no recursive rules there (in fact there's only one rule), you can expand the type of rule's RHS as the rule type (see no_rule2.cpp):

    struct skip_grammar : grammar<skip_grammar>
    {
        template <typename ScannerT>
        struct definition
        {
            definition(skip_grammar const& /*self*/)
            : skip
                (       space_p
                    |   "//" >> *(anychar_p - '\n') >> '\n'
                    |   "/*" >> *(anychar_p - "*/") >> "*/"
                )
            {
            }

            typedef
               alternative<alternative<space_parser, sequence<sequence<
               strlit<const char*>, kleene_star<difference<anychar_parser,
               chlit<char> > > >, chlit<char> > >, sequence<sequence<
               strlit<const char*>, kleene_star<difference<anychar_parser,
               strlit<const char*> > > >, strlit<const char*> > >
            skip_t;
            skip_t skip;

            skip_t const&
            start() const { return skip; }
        };
    };

Ughhh! How did I do that? How was I able to get at the complex typedef? Am I insane? Well, not really... there's a trick! What you do is define the typedef skip_t first as int:

    typedef int skip_t;

Try to compile. Then, the compiler will generate an obnoxious error message such as:

    "cannot convert boost::spirit::alternative<... blah blah...to int".

THERE YOU GO! You got it's type! I just copy and paste the correct type (removing explicit qualifications, if preferred).

Can we still go further? Yes. Remember that the grammar was designed for rules. The nested template definition class is needed to get around the rule's limitations. Without rules, I propose a new class called sub_grammar, the grammar's low-fat counterpart:

    namespace boost { namespace spirit
    {
        template <typename DerivedT>
        struct sub_grammar : parser<DerivedT>
        {
            typedef sub_grammar self_t;
            typedef DerivedT const& embed_t;

            template <typename ScannerT>
            struct result
            {
                typedef typename parser_result<
                    typename DerivedT::start_t, ScannerT>::type
                type;
            };

            DerivedT const& derived() const
            { return *static_cast<DerivedT const*>(this); }

            template <typename ScannerT>
            typename parser_result<self_t, ScannerT>::type
            parse(ScannerT const& scan) const
            {
                return derived().start.parse(scan);
            }
        };
    }}

With the sub_grammar class, we can define our skipper grammar this way (see no_rule3.cpp):

    struct skip_grammar : sub_grammar<skip_grammar>
    {
        typedef
           alternative<alternative<space_parser, sequence<sequence<
           strlit<const char*>, kleene_star<difference<anychar_parser,
           chlit<char> > > >, chlit<char> > >, sequence<sequence<
           strlit<const char*>, kleene_star<difference<anychar_parser,
           strlit<const char*> > > >, strlit<const char*> > >
        start_t;

        skip_grammar()
        : start
            (
                space_p
            |   "//" >> *(anychar_p - '\n') >> '\n'
            |   "/*" >> *(anychar_p - "*/") >> "*/"
            )
        {}

        start_t start;
    };

But what for, you ask? You can simply use the start_t type above as-is. It's already a parser! We can just type:

    skipper_t skipper =
        space_p
        |   "//" >> *(anychar_p - '\n') >> '\n'     
| "/*" >> *(anychar_p - "*/") >> "*/" ;

and use skipper just as we would any parser? Well, a subtle difference is that skipper, used this way will be embedded by value when you compose more complex parsers using it. That is, if we use skipper inside another production, the whole thing will be stored in the composite. Heavy!

The proposed sub_grammar OTOH will be held by reference. Note:

    typedef DerivedT const& embed_t;

The proposed sub_grammar does not have the inherent limitations of rules, is very lighweight, and should be blazingly fast (can be fully inlined and does not use virtual functions). Perhaps this class will be part of a future spirit release.

The no-rules result

So, how much did we save? On MSVCV7.1, the original code: no_rule1.cpp compiles to 28k. Eliding rules, no_rule2.cpp, we got 24k. Not bad, we shaved off 4k amounting to a 14% reduction. But you'll be in for a surprise. The last version, using the sub-grammar: no_rule3.cpp, compiles to 5.5k! That's a whopping 80% reduction.

no_rule1.cpp 28k standard rule and grammar
no_rule2.cpp 24k standard grammar, no rule
no_rule3.cpp 5.5k sub_grammar, no rule, no grammar

typeof

Some compilers already support the typeof keyword. Examples are g++ and Metrowerks CodeWarrior. Someday, typeof will become commonplace. It is worth noting that we can use typeof to define non-recursive rules without using the rule class. To give an example, we'll use the skipper example above; this time using typeof. First, to avoid redundancy, we'll introduce a macro RULE:

    #define RULE(name, definition) typeof(definition) name = definition

Then, simply:

    RULE(
        skipper,
        (       space_p
            |   "//" >> *(anychar_p - '\n') >> '\n'
            |   "/*" >> *(anychar_p - "*/") >> "*/"
        )
    );

(see typeof.cpp)

That's it! Now you can use skipper just as you would any parser. Be reminded, however, that skipper above will be embedded by value when you compose more complex parsers using it (see sub_grammar rationale above). You can use the sub_grammar class to avoid this problem.

Nabialek trick

This technique, I'll call the "Nabialek trick" (from the name of its inventor, Sam Nabialek), can improve the rule dispatch from linear non-deterministic to deterministic. The trick applies to grammars where a keyword (operator, etc), precedes a production. There are lots of grammars similar to this:

    r =
            keyword1 >> production1
        |   keyword2 >> production2
        |   keyword3 >> production3
        |   keyword4 >> production4
        |   keyword5 >> production5
        /*** etc ***/
        ;

The cascaded alternatives are tried one at a time through trial and error until something matches. The Nabialek trick takes advantage of the symbol table's search properties to optimize the dispatching of the alternatives. For an example, see nabialek.cpp. The grammar works as follows. There are two rules (one and two). When "one" is recognized, rule one is invoked. When "two" is recognized, rule two is invoked. Here's the grammar:

    one = name;
    two = name >> ',' >> name;
    
    continuations.add
        ("one", &one)
        ("two", &two)
    ;
    
    line = continuations[set_rest<rule_t>(rest)] >> rest;

where continuations is a symbol table with pointer to rule_t slots. one, two, name, line and rest are rules:

    rule_t name;
    rule_t line;
    rule_t rest;
    rule_t one;
    rule_t two;
    
    symbols<rule_t*> continuations;

set_rest, the semantic action attached to continuations is:

    template <typename Rule>
    struct set_rest
    {
        set_rest(Rule& the_rule)
        : the_rule(the_rule) {}

        void operator()(Rule* newRule) const
        { m_theRule = *newRule; }

        Rule& the_rule;
    };

Notice how the rest rule gets set dynamically when the set_rule action is called. The dynamic grammar parses inputs such as:

"one only"
"one again"
"two first, second"

The cool part is that the rest rule is set (by the set_rest action) depending on what the symbol table got. If it got a "one" then rest = one. If it got "two", then rest = two. Very nifty! This technique should be very fast, especially when there are lots of keywords. It would be nice to add special facilities to make this easy to use. I imagine:

    r = keywords >> rest;

where keywords is a special parser (based on the symbol table) that automatically sets its RHS (rest) depending on the acquired symbol. This, I think, is mighty cool! Someday perhaps...

Also, see the switch parser for another deterministic parsing trick for character/token prefixes.