# Copyright 2022 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/.

=head1 NAME

Marpa::R2::HTML - High-level HTML Parser

=head1 SYNOPSIS

B<Delete all tables>:

=for Marpa::R2::Display
name: 'HTML Synopsis: Delete Tables'
normalize-whitespace: 1

    use Marpa::R2::HTML qw(html);

    my $with_table = 'Text<table><tr><td>I am a cell</table> More Text';
    my $no_table   = html( \$with_table,
        { table => sub { return q{} } });

=for Marpa::R2::Display::End

B<Delete everything but tables>:

=for Marpa::R2::Display
name: 'HTML Synopsis: Delete Everything But Tables'
perltidy: '-dcsc -sil=0'

    my %handlers_to_keep_only_tables = (
        table  => sub { return Marpa::R2::HTML::original() },
        ':TOP' => sub { return \( join q{}, @{ Marpa::R2::HTML::values() } ) }
    );
    my $only_table = html( \$with_table, \%handlers_to_keep_only_tables );

=for Marpa::R2::Display::End

The above works by turning the original text of the HTML into values
and concatenating the values at the top of the parse.
The same logic works even if a B<table is very defective>:

=for Marpa::R2::Display
name: 'HTML Synopsis: Defective Tables'
perltidy: '-dcsc -sil=0'

    my $with_bad_table = 'Text<tr>I am a cell</table> More Text';
    my $only_bad_table =
        html( \$with_bad_table, \%handlers_to_keep_only_tables );

=for Marpa::R2::Display::End

B<Delete all comments>:

=for Marpa::R2::Display
name: 'HTML Synopsis: Delete Comments'
normalize-whitespace: 1

    my $with_comment = 'Text <!-- I am a comment --> I am not a comment';
    my $no_comment   = html( \$with_comment,
        { ':COMMENT' => sub { return q{} } }
    );

=for Marpa::R2::Display::End

By default, text is passed through unchanged, so that the user
need only specify semantic actions for those components she
wants changed.  To B<change the title of a document>:

=for Marpa::R2::Display
name: 'HTML Synopsis: Change Title'
perltidy: '-dcsc -sil=0'

    my $old_title = '<title>Old Title</title>A little html text';
    my $new_title = html(
        \$old_title,
        {   'title' => sub { return '<title>New Title</title>' }
        }
    );

=for Marpa::R2::Display::End

B<Delete all elements with a class attribute
of "C<delete_me>">:

=for Marpa::R2::Display
name: 'HTML Synopsis: Delete by Class'
normalize-whitespace: 1

    my $stuff_to_be_edited = '<p>A<p class="delete_me">B<p>C';
    my $edited_stuff       = html( \$stuff_to_be_edited,
        { '.delete_me' => sub { return q{} } });

=for Marpa::R2::Display::End

L<Marpa::R2::HTML> recognizes elements even if they have missing
start and/or end tags.
L<Marpa::R2::HTML> can B<supply missing tags>:

=for Marpa::R2::Display
name: 'HTML Synopsis: Supply Missing Tags'
perltidy: '-dcsc -sil=0'

    sub supply_missing_tags {
        my $tagname = Marpa::R2::HTML::tagname();
        return if Marpa::R2::HTML::is_empty_element($tagname);
        return ( Marpa::R2::HTML::start_tag() // "<$tagname>\n" )
            . Marpa::R2::HTML::contents() .
            ( Marpa::R2::HTML::end_tag() // "</$tagname>\n" );
    }
    my $html_with_just_a_title = '<title>I am a title and That is IT!';
    my $valid_html_with_all_tags =
        html( \$html_with_just_a_title, { q{*} => \&supply_missing_tags } );

=for Marpa::R2::Display::End

L<Marpa::R2::HTML> understands the hierarchical structure of an HTML document.
B<Finding the maximum nesting depth in elements> is straightforward:

=for Marpa::R2::Display
name: 'HTML Synopsis: Maximum Element Depth'
perltidy: '-dcsc -sil=0'

    sub depth_below_me {
        return List::Util::max( 0, @{ Marpa::R2::HTML::values() } );
    }
    my %handlers_to_calculate_maximum_element_depth = (
        q{*}   => sub { return 1 + depth_below_me() },
        ':TOP' => sub { return depth_below_me() },
    );
    my $maximum_depth_with_just_a_title = html( \$html_with_just_a_title,
        \%handlers_to_calculate_maximum_element_depth );

=for Marpa::R2::Display::End

L<Marpa::R2::HTML> tracks actual elements, however tagged.
The above code returns the same depth for C<$valid_html_with_all_tags>
as for C<$html_with_just_a_title>.

=head1 DESCRIPTION

L<Marpa::R2::HTML> does "high-level" parsing of HTML.
It allows handlers to be specified for elements, terminals and other
components in the hierarchical structure of an HTML document.
L<Marpa::R2::HTML> is an extremely liberal HTML parser.
L<Marpa::R2::HTML> does not reject any documents,
no mater how poorly they fit the HTML standards.

=head1 THE Marpa::R2::HTML::html STATIC METHOD

The interface to
L<Marpa::R2::HTML> is through the
C<Marpa::R2::HTML::html>
static method.
It is
the only
L<Marpa::R2::HTML>
method not part of the API for the
semantic actions.

C<html> takes one or more arguments.
The first argument is required, and must be a reference to
a string to be parsed as HTML.
The second and
subsequent arguments (all optional) are hash references
with handler descriptions.
(See L<the synopsis|/"SYNOPSIS"> for several examples of
calls using the C<html> method.)

=head2 CSS-style Handler Options

Handler descriptions in L<Marpa::R2::HTML> are key-value pairs
in a hash.
In each pair, the key is a CSS-style handler specifier,
and the value is a closure,
which is called the B<action> for the
handler.

Specifiers are "CSS-style" --
their syntax imitates some of the basic cases
of CSS specifiers.
No attempt is planned
to implement the full CSS specifier syntax.

Supported specifier syntaxes are as follows:

=over

=item Tagname Specifiers

=for Marpa::R2::Display
name: 'HTML Synopsis: Delete Everything But Tables'
partial: 1
normalize-whitespace: 1
flatten: 1

        table  => sub { return Marpa::R2::HTML::original() },

=for Marpa::R2::Display::End

If a specifier contains no special characters
it is taken
as the name of an element.
(A "special" character is
anything except an alphanumeric, a hyphen or an underscore.)
Consistent with L<HTML::Parser>'s default behavior,
element names must be specified in lowercase.

=item Class Specifiers

A specifier which is a dot or period followed by a name
will match any element whose class attribute is the same as the specified name.
For example, the specifier "C<.delete_me>" will match any element whose
class attribute is "C<delete_me>".

=item Tagname-Class Pair Specifiers

A specifier which contains a dot or period somewhere other than
the first position (such as "C<span.label>")
is treated as a dotted tagname-class pair.
Its action will be called for
any component whose tagname and class attribute both match
the specifiers.

=item The Tagname Wildcard Specifier

A specifier of just an asterisk ("C<*>") matches all elements.
Be careful to note that matching all elements is B<not> the same
as matching all components.
The element wildcard specifier will not match any pseudoclasses.

=item Pseudoclass Specifiers

=for Marpa::R2::Display
name: 'HTML Synopsis: Delete Comments'
partial: 1
normalize-whitespace: 1
flatten: 1

     ':COMMENT' => sub { return q{} }

=for Marpa::R2::Display::End

A specifier which begins with a colon ("C<:>") matches a pseudoclass.
L<Marpa::R2::HTML> defines
pseudoclasses to deal with terminals and other non-element
components of the HTML hierarchy.

=back

=head2 Conflicting Specifiers

At most one semantic action is called for each component.
Where an element component matches several specifiers,
the action is picked based on the B<most specific match>.

=over

=item 1. Matches by tagname-class pair are the most specific.

=item 2. Matches by class are the next most specific.

=item 3. Matches by tagname are considered less specific than matches by class.

=item 4. The wildcard match is the least specific.

=back

Here's an example:

=for Marpa::R2::Display
name: 'HTML Pod: Handler Precedence'
perltidy: '-dcsc -sil=0'
remove-display-indent: 1

    my $html = <<'END_OF_HTML';
    <span class="high">High Span</span>
    <span class="low">Low Span</span>
    <div class="high">High Div</div>
    <div class="low">Low Div</div>
    <div class="oddball">Oddball Div</div>
    END_OF_HTML

    our @RESULTS = ();
    Marpa::R2::HTML::html(
        \$html,
        {   q{*} => sub {
                push @RESULTS, 'wildcard handler: ' . Marpa::R2::HTML::contents();
            },
            'div' => sub {
                push @RESULTS, '"div" handler: ' . Marpa::R2::HTML::contents();
            },
            '.high' => sub {
                push @RESULTS, '".high" handler: ' . Marpa::R2::HTML::contents();
            },
            'div.high' => sub {
                push @RESULTS,
                    '"div.high" handler: ' . Marpa::R2::HTML::contents();
            },
            '.oddball' => sub {
                push @RESULTS,
                    '".oddball" handler: ' . Marpa::R2::HTML::contents();
            },
            'body' => sub {undef},
            'head' => sub {undef},
            'html' => sub {undef},
            'p'    => sub {undef},
        }
    );

=for Marpa::R2::Display::End

Here is what C<$result> would contain after the above code was run:

=for Marpa::R2::Display
name: 'HTML Pod: Handler Precedence Result'
normalize-whitespace: 1

    ".high" handler: High Span
    wildcard handler: Low Span
    "div.high" handler: High Div
    "div" handler: Low Div
    ".oddball" handler: Oddball Div

=for Marpa::R2::Display::End

=head2 Details of the Specifier Syntax

For elements and class names only alphanumerics, hyphens and underscores are supported.
Elements must be specified in lowercase, but they will match tagnames in the original
document on a case-insensitive basis.

Forcing element names to be lowercase follows the default behavior of
L<HTML::Parser>, which coerces all tagnames to lowercase.
This is consistent with the HTML standards.
It is B<not> consistent with the XML standards,
and an option to configure this behavior may be added in
the future.

Pseudoclass names special to
L<Marpa::R2::HTML> are case-sensitive, and must be all uppercase.
Lowercase is reserved for CSS pseudoclasses.
The CSS standard specifies that its pseudoclass names are case-indifferent.
No CSS pseudoclasses are supported at this writing.

=head1 PSEUDOCLASSES

L<Marpa::R2::HTML> uses
L<HTML::Parser> to do its low-level parsing.
L<HTML::Parser> "events"
become the terminals for L<Marpa::R2::HTML>.

Besides terminals and elements,
three other HTML components are recognized:
the SGML prolog (C<:PROLOG>),
the SGML trailer (C<:TRAILER>),
and the HTML document as a whole (C<:TOP>).

=head2 :CDATA

The C<:CDATA> pseudoclass specifies the action for
CDATA terminals.
Its action is called once for each non-whitespace raw C<text> event
that is not reclassed as cruft.
(B<Raw text>
is text in which any markup and entities should be left as is.)

More precisely,
a C<:CDATA> terminal is created from any
L<HTML::Parser> C<text> event that has the C<is_cdata> flag on;
that contains a non-whitespace character
as defined in the HTML 4.01 specification
(L<http://www.w3.org/TR/html4/struct/text.html#h-9.1>);
and that is not reclassed as cruft.

=head2 :COMMENT

The C<:COMMENT> pseudoclass specifies the action for HTML comments.
Its action is called once for every C<HTML::Parser> C<comment> event that
is not reclassed as cruft.

=head2 :CRUFT

The C<:CRUFT> pseudoclass specifies the action for cruft.
Its action is called once for every C<HTML::Parser> event that
L<Marpa::R2::HTML> reclasses as cruft.

L<Marpa::R2::HTML> reclasses terminals as cruft when
they do not fit the structure of an HTML document.
One example of a terminal that
L<Marpa::R2::HTML> would reclass as cruft is a
C<< </head> >> end tag in the HTML body.

Reclassing terminals as cruft is only done as the last resort.
When it can,
L<HTML::Parser> forgives
violations of the HTML standards and accepts terminals as non-cruft.

Cruft is treated in much the same way as comments.
It is preserved, untouched, in the original text view.

=head2 :DECL

The C<:DECL> pseudoclass specifies the action for SGML declarations.
Its action is called once for every C<HTML::Parser> C<declaration> event that
is not reclassed as cruft.

=head2 :PCDATA

The C<:PCDATA> pseudoclass specifies the action for
PCDATA terminals.
Its action is called once for each non-whitespace non-raw C<text> event
that is not reclassed as cruft.

More precisely,
a C<:PCDATA> terminal is created from any
L<HTML::Parser> C<text> event that has the C<is_cdata> flag B<off>;
that contains a non-whitespace character
as defined in the HTML 4.01 specification
(L<http://www.w3.org/TR/html4/struct/text.html#h-9.1>);
and that is not reclassed as cruft.

Markup and entities in
C<:PCDATA> text
are expected to be interpreted eventually,
but it can be counter-productive to do this
during parsing.
An application may, for example,
be rewriting a document for display on the web.
In that case it will often
want to leave markup and entities for the client's browser
to interpret.

L<Marpa::R2::HTML> leaves interpretation of markup and entities entirely to
the application.
An application which chooses to do the interpretation itself
may do it in the actions,
or deal with it in post-processing.
CPAN has excellent tools for this,
some of which are part of L<HTML::Parser>.

=head2 :PI

The C<:PI> pseudoclass specifies the action for SGML processing instructions.
Its action is called once for every L<HTML::Parser> C<process> event that
is not reclassed as cruft.

=head2 :PROLOG

The C<:PROLOG> pseudoclass specifies the action for SGML prolog.
This is the part of the HTML document which precedes the HTML root element.
Components valid in the
prolog include SGML comments, processing instructions and whitespace.

=head2 :TOP

The action specified for the C<:TOP> pseudoclass will be called
once and only once in every parse,
and will be the last action called in every parse.
The C<:TOP> component is the entire
physical document, including
the SGML prolog,
the root element,
and the SGML trailer.
All the other HTML components in a document
will be descendants of the C<:TOP> component.

The C<:TOP> action is unique, in that there is always an action
for it, even if one is not specified.
The C<html> method returns the value
returned by the C<:TOP> action.
The default C<:TOP> action returns a B<reference> to a string
with the literal text value of all
of its descendants.

=head2 :TRAILER

The C<:TRAILER> pseudoclass specifies the action for SGML trailer.
This is the part of the HTML document which follows the HTML root element.
Components valid in the
trailer include SGML comments, processing instructions, and whitespace.
Cruft can also be found here, though for L<Marpa::R2::HTML> that is a
last resort.

=head2 :WHITESPACE

A L<Marpa::R2::HTML> C<:WHITESPACE> terminal is created for every
L<HTML::Parser> C<text> event that is entirely whitespace
as defined in the HTML 4.01 specification
(L<http://www.w3.org/TR/html4/struct/text.html#h-9.1>)
and that is not reclassed as cruft.
Whitespace is acceptable in places where non-whitespace is not,
and the difference can be very significant structurally.

=head1 VIEWS

I hope L<the synopsis|/"SYNOPSIS"> convinces the reader
that the action semantics of L<Marpa::R2::HTML> are natural.
This naturalness is achieved at the price of some novelty.
This section explains the ideas behind the semantic action API.
Depending on taste,
readers may want to skip this section and go straight to
the API.

The components of an HTML document form a hierarchy,
with the C<:TOP> component on top, and the terminals on the bottom.
The traditional syntax tree method requires semantic actions
to know precisely
what children every component will have.
This processing model is not a good fit to HTML.
L<Marpa::R2::HTML> gives the writer of semantic actions
"views" of each component that better fit situations
where the number and type of
children is unknown or vaguely defined.

L<Marpa::R2::HTML>'s semantics
focus more widely --
on a component's descendants instead of
just its direct children.
(The terms ancestor and descendant are used in the standard way:
If a component X is
above Y in the hierarchy,
X is an B<ancestor> of Y; and
Y is a B<descendant> of the X.)

=head2 The Original View

The B<original view> sees the text of a component as it was
originally passed to the parser.
The original view never changes.
The original view is seen
through the L</"Marpa::R2::HTML::original"> API method.

=head2 The Terminals View

The B<terminals view> sees the terminals corresponding to the
original text of a component.
The terminals view never changes.
The terminals view is usually seen as part of other views.

At this writing the API does not contain a "pure" terminals view method.
For a terminals view of the whole HTML document,
L<HTML::Parser> does the job with significantly lower overhead.
For views and sections of views with no values defined,
the descendants view (L<described below|/"The Descendants View">)
is equivalent to the terminals view.

=head2 The Values View

When actions are called, they return a value.
If that value is defined, it becomes visible to the B<values view> of
its ancestors.
The values view of a component sees the visible values for its descendants.

The B<values view> is an array, with the values ordered according to the
lexical order of the components whose actions returned them.
If no descendants have visible values,
then the values view is a zero-length array.

The values view is hierarchical.
When a component produces a visible value,
it makes the values of its descendants disappear.
That is, whenever the semantic action for a component X
returns anything other than a Perl C<undef>, it has two effects:

=over

=item * That return value becomes the visible value associated with component X.

=item * All the values previously visible due to semantic actions
for the descendants of component X disappear.

=back

Values which disappear are gone forever.
There is no mechanism to make them "reappear".

As a special case, if an action for a component returns a Perl C<undef>,
not only do the values of all its descendants disappear,
the component for the action also will not appear in the values view.
When its semantic action returns C<undef>, a component permanently "drops out" of the values view
taking all descendants with it.
The original view is seen
through the L</"Marpa::R2::HTML::values"> API method.

=head2 The Literal View

The literal view can be thought of as a mix between the original view
and the values view.
It sees a text string, like the original view.
But unlike the original view, the literal view includes the visible values.

Values appear in the B<literal view> in stringized form.
For sections of the original text without visible values,
the literal view is the same as the original
view.
In all L<Marpa::R2::HTML>'s views,
whether descendants are seen as text or values,
they
are seen in the original lexical order.
The literal view is seen
through the L</"Marpa::R2::HTML::literal"> API method.

=head2 The Descendants View

Just as the literal view can be thought of as a mix between the original view
and the values view,
the descendants view can be thought of a mix between the terminals view and the
values view.

The B<descendants view> sees an array of elements
with data for each
of the component's descendants,
in lexical order.
Where a value is visible, the descendants view sees data for the component with the
visible value.
Where no value is visible, the descendants view sees data for the terminals.
This means that
when no values are visible, the descendants view is the same as the terminals view.

The descendants view is implemented via the L</"Marpa::R2::HTML::descendants"> method.
It is the most fine-grained and detailed way to look at the descendants of a component.
The descendants view can do anything that the other views can do,
but the other views should be preferred when they fit the application.
Other views are typically more intuitive and efficient.

=head2 Views versus Syntax Trees

Views are a generalization of the traditional method
for processing semantics: syntax trees.
The values view is the view that most
closely resembles a syntax tree.
But there are important differences.

In its purest form,
the syntax tree model
required the semantic actions to
define exactly how many and what kind of immediate children
each node had.
Each node in a syntax tree worked with its immediate children.
Children in a syntax tree appeared as values.

The values view, on the other hand, sees all its descendants,
not just its immediate children, but only if
they make themselves visible.
Because of this,
the values view lends itself to being mixed with other views.
The values view allows pieces of the tree to decide when they will
come into sight
and when they will fall out of view.

=head2 Views and Efficiency

In most applications,
views are more efficient than syntax trees.
In terms of L<Marpa::R2::HTML> views,
traditional syntax tree processing
corresponds most closely to the values view
when every component in the parse has a visible value.
For L<Marpa::R2::HTML> this is close to the worst case.

L<Marpa::R2::HTML> optimizes for unvalued components.
Unvalued components are represented as terminal spans.
Adjacent descendant spans are automatically merged.
This means the size and time required do not increase as
processing rises up the component hierarchy.

Terminals views are calculated on a just-in-time basis
when they are requested through the action API.
The terminals view is produced quickly from the merged terminal span.

Original views are also calculated on a just-in-time basis
as requested.
Each terminal tracks the text it represents as a
character span.
The original text can be quickly reconstructed
as the text in the source document from
the first character location of its component's first terminal
to the last character location of the component's last terminal.

When a handler does not need to return a value,
the most efficient thing to do is to return C<undef>.
This reverts that component and all its descendants to
the efficient unvalued representation.

=head1 THE SEMANTIC ACTION API

L<Marpa::R2::HTML>'s semantic action API is implemented
mainly through context-aware static methods.
No arguments are passed to the
user's semantics action callbacks.
Instead the semantic actions get whatever data they need
by calling these static methods.

=head2 API Static Methods

=over

=item Marpa::R2::HTML::attributes

Returns a hash ref to the attributes of the start tag.
This hash ref is exactly the hash ref returned
for the C<attr> arg specification of L<HTML::Parser>.
The C<attributes> API method
returns an empty hash
if there were no attributes,
if there was no start tag for this element,
or if the current component is not an element.

=item Marpa::R2::HTML::contents

For an element, returns the literal view of the contents.
The contents of an element are its entire text
except for its start tag and its end tag.
For an non-element component, returns undef.

=item Marpa::R2::HTML::descendants

This static method implements the descendants view.
It takes one argument, the "dataspec".
The B<dataspec> is a string specifying
the data to be returned for each descendant.
The C<descendants> method
returns a reference to an array with one element per descendant,
in lexical order.
Each element in the array is a reference to an array whose
elements are
the per-descendant data requested in the string.

The descendant data specification string has a syntax
similar to that of the C<argspec> strings of L<HTML::Parser>.
Details of that syntax L<are given below|/"Dataspecs">

=item Marpa::R2::HTML::end_tag

For an element with an explicit end tag,
returns the original text of the end tag.
For non-element components, returns undef.
For elements with no end tag, returns undef.

=item Marpa::R2::HTML::is_empty_element

For an element, returns a Perl true value if the element
is empty,
a defined Perl false value otherwise.
For non-element components, returns undef.

=item Marpa::R2::HTML::literal

The C<Marpa::R2::HTML::literal> method implements the literal view.
Returns a string containing the literal view of the component --
its text as modified by any the visible values of its
descendants.

=item Marpa::R2::HTML::literal_ref

Returns a reference to a string containing the literal view of the
component.  This can be useful for very long strings.

=item Marpa::R2::HTML::offset

Returns the start offset of the component.
This is a zero-based location in the source document.
Some components are zero-length,
containing none of the tokens in the physical input.
The C<Marpa::R2::HTML::offset> method return C<undef>
for these.

=item Marpa::R2::HTML::original

The C<Marpa::R2::HTML::original> method implements the original view.
Returns a string containing the original view of the component --
its text unchanged from the source document.

=item Marpa::R2::HTML::start_tag

For an element with an explicit start tag,
returns the original text of the start tag.
For non-element components, returns undef.
For elements with no explicit start tag, returns undef.

=item Marpa::R2::HTML::tagname

For an element component,
returns its tagname.
There is a tagname even if there are no
explicit tags.
Tagname is determined based on structure.
For non-element components, returns undef.

=item Marpa::R2::HTML::title

Returns the value of the title attribute.
For a non-element component, returns undef.
If there was no explicit start tag, returns undef.
If there was no title attribute, returns undef.

=item Marpa::R2::HTML::token_type

For a token, returns the token type.
The token types
are the event types from L<HTML::Parser>:
"C<T>" for text,
"C<S>" for a start tag,
"C<E>" for an end tag,
"C<PI>" for a processing instruction,
"C<D>" for an SGML declaration,
and "C<C>" for a comment.
If the component is an element or some other
non-token, returns undef.

=item Marpa::R2::HTML::values

The C<Marpa::R2::HTML::values> method implements the values view.
It returns a reference to an array of the descendant
values visible from this component,
in lexical order.
No elements of this array will be undefined.
The array will be zero length if no descendant
has a visible value.

=back

=head2 Dataspecs

=for Marpa::R2::Display
name: dataspec example
perltidy: '-dcsc -sil=0'

    Marpa::R2::HTML::descendants('token_type,literal,element')

=for Marpa::R2::Display::End

The data specification string, or dataspec,
is a comma separated list of B<descendant data specifiers>.
The C<Marpa::R2::HTML::descendants> method takes a dataspec
as its argument.
The C<Marpa::R2::HTML::descendants> method returns a reference
to an array of references to arrays of per-descendant data.
The contents of the per-descendant data arrays
and their order is as specified
by the dataspec.
These are the valid descendant data specifiers:

=over

=item C<element>

For an element descendant, returns the tagname.
A valid tagname is returned even if there were no explicit tags.
For non-element descendants, returns undef.

=item C<literal>

Returns a string containing the literal view of the
descendant.

=item C<original>

Returns a string containing the original view of the
descendant.

=item C<token_type>

If the descendant is a terminal, returns the token type.
Token types are as described for the
L</"Marpa::R2::HTML::token_type"> API method.
For components with visible values, returns undef.

=item C<value>

For element descendants with a value, returns that
value.
In all other cases, returns undef.

=back

=head2 The Instance Hash

Each L<Marpa::R2::HTML> instance
makes available
a per-instance variable
as a scratchpad for the application:
C<$Marpa::R2::HTML::INSTANCE>.
Each call to
L<Marpa::R2::HTML::html|/"THE Marpa::R2::HTML::html STATIC METHOD">
creates a C<$Marpa::R2::HTML::INSTANCE>
variable which is
reserved for that application using the C<local> keyword.
L<Marpa::R2::HTML::html|/"THE Marpa::R2::HTML::html STATIC METHOD">
initializes it to an empty hash,
but after that does not touch it.
When programming via side effects
is more natural than
passing data up the parse
tree (and it often is),
C<$Marpa::R2::HTML::INSTANCE> can be used to
store the data.

Ordinarily, C<$Marpa::R2::HTML::INSTANCE> is destroyed,
with the rest of the parse instance,
when C<Marpa::R2::HTML::html> returns.
But it can be useful
for the C<:TOP> semantic action to
return
C<$Marpa::R2::HTML::INSTANCE> as the value of the parse.

=head2 Undefined Actions versus Actions Which Return C<undef>

It is worth emphasizing that
the effect of not defining a semantic action for a component
is different from the effect of defining a semantic action which
returns a Perl C<undef>.
The difference lies in what happens to any visible values
of the descendants of that component.

Where no action is defined for a component,
it leaves all that component's views as they were before.
That is, all values which were visible remain visible and
no new values become visible.
When an action is defined for a component, but that action returns undef,
no new values become visible, and
all descendant values which were visible B<disappear>.

=head2 Root Element versus :TOP Pseudoclass

It is important to understand the very
special function of the C<:TOP>
component,
and to avoid confusing it with the HTML root element.
The most important distinctions are that

=over

=item *

The semantic action
for C<:TOP> pseudoclass is always the last action
to be called in a parse.

=item *

The C<:TOP> component is B<always> the entire HTML document.
This can be true of the root element, but it is not true in
all cases.

=item *

The value that the action for the C<:TOP> component
returns becomes the value that
the
L<Marpa::R2::HTML::html|/"THE Marpa::R2::HTML::html STATIC METHOD">
method returns.

=back

The root element is the HTML element whose tagname is "C<html>", though
its start and end tags are optional
and can be omitted even in strictly valid HTML.
Tags or no tags, every HTML document has a
root element.
(The C<:TOP> component is not an element, so it does not have a tagname and
never has tags.)

The root element is always a descendant of the C<:TOP>
component.
The SGML prolog and SGML trailer are always descendants of the C<:TOP>
component.
The SGML prolog and SGML trailer
are never descendants of the root element.

If an action for the root element is specified,
it will also be called
once and only once in every parse.
An action for the root element can be specified in same way as actions
for other elements, using its tagname of "C<html>".
An element wildcard action also becomes the action for the root element,
if no more specific handler declaration takes precedence.

A C<:TOP> action will be called once and only once in every parse.
The C<:TOP> action is unique in that there is a default action.
No other component has a default action.

=head2 Tags versus Structure

Where tags conflict with structure,
L<HTML::Parser> follows structure.
"Following structure" means that, for example,
if semantic actions for the C<html>, C<head>,
and C<body> elements exist,
they will be called once and only once during every parse.

Consider this short and very defective HTML document:

=for Marpa::R2::Display
name: 'HTML Pod: Structure vs. Element Example'
normalize-whitespace: 1

    <title>Short</title><p>Text</head><head>

=for Marpa::R2::Display::End

L<HTML::Parser> starts the HTML document's body
when it encounters the C<< <p> >> start tag.
That means that, even if they were in the right order,
the two C<head> tags cannot be fit into any reasonable parse
structure.

If an action is specified for the C<head> element,
it will be called for the actual header,
and the original view of the C<head> element component
will be the text "C<< <title>Short</title> >>".
The action for the C<head> element will not be called again.
The two stray tags, C<< </head> >> and C<< <head> >>,
will be treated as descendants of
the C<body> element, and reclassed as
"cruft" terminals.

=head2 Explicit and Implicit Elements

If a semantic action
is specified for a tagname, it is called
whenever an element is found with that tagname,
even if there are no explicit tags for
that element.
The HTML standards allow both start and end tags
to be missing
for
C<html>,
C<head>,
C<body> and
C<tbody> elements.
L<Marpa::R2::HTML> is more liberal,
and will recognize virtual tags for
C<table>, C<tr>, and C<td> elements
as required to repair a defective table.

L<Marpa::R2::HTML> is more even
liberal about recognizing virtual end tags
than it is about start tags.
Virtual start tags are recognized only for the specific
elements listed above.
For any non-empty HTML element, there is some circumstance
under which
L<Marpa::R2::HTML> will recognize a virtual end tag.
At end of file,
as one example,
L<Marpa::R2::HTML> will do its best to produce a balanced
HTML structure by
creating a virtual end tag for every element
in the stack of
currently active elements.

=head1 EXPORTS

L<Marpa::R2::HTML> exports nothing by default.
Optionally,
L<Marpa::R2::HTML::html|/"THE Marpa::R2::HTML::html STATIC METHOD">
may be exported.

=head1 Copyright and License

=for Marpa::R2::Display
ignore: 1

  Copyright 2022 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/.

=for Marpa::R2::Display::End

=cut

# Local Variables:
#   mode: cperl
#   cperl-indent-level: 4
#   fill-column: 100
# End:
# vim: expandtab shiftwidth=4: