Staging
v0.5.1
https://github.com/python/cpython
Raw File
Tip revision: 2f6c878a1402ddd83ef913928f8355d491a03d1c authored by cvs2svn on 14 April 1998, 02:16:23 UTC
This commit was manufactured by cvs2svn to create tag 'r151'.
Tip revision: 2f6c878
l2hinit.perl
#LaTeX2HTML Version 96.1 : dot.latex2html-init		-*- perl -*-
#
#  Significantly revised by Fred L. Drake, Jr. <fdrake@acm.org> for use
#  with the Python documentation.
#
#  New name to avoid distributing "dot" files with the Python documentation.
#

$INFO = 1;              # 0 = do not make a "About this document..." section
$MAX_LINK_DEPTH = 3;

$NUMBERED_FOOTNOTES = 1;

# Python documentation uses section numbers to support references to match
# in the printed and online versions.
#
$SHOW_SECTION_NUMBERS = 1;

$ICONSERVER = '../icons';

$CHILDLINE = "\n<p><hr>\n";
$VERBOSITY = 0;

# Locate a file that's been "require"d.  Assumes that the file name of interest
# is unique within the set of loaded files, after directory names have been
# stripped.  Only the directory is returned.
#
sub find_my_file{
    my($myfile,$key,$tmp,$mydir) = (@_[0], '', '', '');
    foreach $key (keys %INC) {
	$tmp = "$key";
	$tmp =~ s|^.*/||o;
	if ($tmp eq $myfile) {
	    $mydir = $INC{$key};
	}
    }
    $mydir =~ s|/[^/]*$||;
    $mydir;
}


# A little painful, but lets us clean up the top level directory a little,
# and not be tied to the current directory (as far as I can tell).
#
use Cwd;
use File::Basename;
($myname, $mydir, $myext) = fileparse(__FILE__, '\..*');
chop $mydir;			# remove trailing '/'
$mydir = getcwd() . "$dd$mydir"
  unless $mydir =~ s|^/|/|;
$LATEX2HTMLSTYLES = "$mydir$envkey$LATEX2HTMLSTYLES";

($myrootname, $myrootdir, $myext) = fileparse($mydir, '\..*');
chop $myrootdir;


sub make_nav_panel{
    ($NEXT_TITLE ? "$NEXT\n" : '')
      . ($UP_TITLE ? "$UP\n" : '')
      . ($PREVIOUS_TITLE ? "$PREVIOUS\n" : '')
      . "$CONTENTS\n$INDEX"
#      . " $CUSTOM_BUTTONS"
      . "<br>\n"
      . ($NEXT_TITLE ? "<b> Next:</b> $NEXT_TITLE\n" : '')
      . ($UP_TITLE ? "<b>Up:</b> $UP_TITLE\n" : '')
      . ($PREVIOUS_TITLE ? "<b>Previous:</b> $PREVIOUS_TITLE\n" : '');
}

sub top_navigation_panel {
    "<div class=navigation>\n"
      . &make_nav_panel
      . "<br><hr><p></div>"
}

sub bot_navigation_panel {
    "<p>\n<div class=navigation><hr>"
      . &make_nav_panel
      . "</div>"
}


sub gen_index_id {
    # this is used to ensure common index key generation and a stable sort
    my($str,$extra) = @_;
    sprintf("%s###%s%010d", $str, $extra, ++$global{'max_id'});
}

sub make_index_entry {
    my($br_id,$str) = @_;
    # If TITLE is not yet available (i.e the \index command is in the title of the
    # current section), use $ref_before.
    $TITLE = $ref_before unless $TITLE;
    # Save the reference
    $str = gen_index_id($str, '');
    $index{$str} .= &make_half_href("$CURRENT_FILE#$br_id");
    "<a name=\"$br_id\">$anchor_invisible_mark<\/a>";
}

# use this instead with the buildindex.py tool
sub add_idx{
    print "\nDoing the index ...";
    close(IDXFILE);
    my $index = `$myrootdir/tools/buildindex.py index.dat`;
    s/$idx_mark/$index/;
}


$idx_module_mark = '<tex2html_idx_module_mark>';
$idx_module_title = 'Module Index';

sub add_module_idx{
    print "\nDoing the module index ...";
    my $key;
    my $index = "<p>";
    open(MODIDXFILE, ">modindex.dat") || die "\n$!\n";
    foreach $key (keys %Modules) {
	# dump the line in the data file; just use a dummy seqno field
	print MODIDXFILE "$Modules{$key}" . $IDXFILE_FIELD_SEP . "$key###\n";
    }
    close(MODIDXFILE);
    $index = `$myrootdir/tools/buildindex.py modindex.dat`;
    s/$idx_module_mark/$index<p>/;
}

# replace both indexes as needed:
sub add_idx_hook{
    &add_idx if (/$idx_mark/);
    &add_module_idx if (/$idx_module_mark/);
}


# In addition to the standard stuff, add label to allow named node files.
sub do_cmd_tableofcontents {
    local($_) = @_;
    $TITLE = $toc_title;
    $tocfile = $CURRENT_FILE;
    my($closures,$reopens) = &preserve_open_tags();
    &anchor_label("contents",$CURRENT_FILE,$_);		# this is added
    join('', "<BR>\n", $closures
	 , &make_section_heading($toc_title, "H2"), $toc_mark
	 , $reopens, $_);
}
# In addition to the standard stuff, add label to allow named node files.
sub do_cmd_listoffigures {
    local($_) = @_;
    $TITLE = $lof_title;
    $loffile = $CURRENT_FILE;
    my($closures,$reopens) = &preserve_open_tags();
    &anchor_label("lof",$CURRENT_FILE,$_);		# this is added
    join('', "<BR>\n", $closures
	 , &make_section_heading($lof_title, "H2"), $lof_mark
	 , $reopens, $_);
}
# In addition to the standard stuff, add label to allow named node files.
sub do_cmd_listoftables {
    local($_) = @_;
    $TITLE = $lot_title;
    $lotfile = $CURRENT_FILE;
    my($closures,$reopens) = &preserve_open_tags();
    &anchor_label("lot",$CURRENT_FILE,$_);		# this is added
    join('', "<BR>\n", $closures
	 , &make_section_heading($lot_title, "H2"), $lot_mark
	 , $reopens, $_);
}
# In addition to the standard stuff, add label to allow named node files.
sub do_cmd_textohtmlinfopage {
    local($_) = @_;
    if ($INFO) {					# 
	&anchor_label("about",$CURRENT_FILE,$_);	# this is added
    }							#
    ( ($INFO == 1)
     ? join('', $close_all
	, "<STRONG>$t_title</STRONG><P>\nThis document was generated using the\n"
	, "<A HREF=\"$TEX2HTMLADDRESS\"><STRONG>LaTeX</STRONG>2<tt>HTML</tt></A>"
	, " translator Version $TEX2HTMLVERSION\n"
	, "<P>Copyright &#169; 1993, 1994, 1995, 1996, 1997,\n"
	, "<A HREF=\"$AUTHORADDRESS\">Nikos Drakos</A>, \n"
	, "Computer Based Learning Unit, University of Leeds.\n"
	, "<P>The command line arguments were: <BR>\n "
	, "<STRONG>latex2html</STRONG> <tt>$argv</tt>.\n"
	, "<P>The translation was initiated by $address_data[0] on $address_data[1]"
	, $open_all, $_)
      : join('', $close_all, $INFO,"\n", $open_all, $_))
}

# $idx_mark will be replaced with the real index at the end
sub do_cmd_textohtmlindex {
    local($_) = @_;
    $TITLE = $idx_title;
    $idxfile = $CURRENT_FILE;
    if (%index_labels) { make_index_labels(); }
    if (($SHORT_INDEX) && (%index_segment)) { make_preindex(); }
    else { $preindex = ''; }
    my $heading = make_section_heading($idx_title, 'h2') . $idx_mark;
    my($pre,$post) = minimize_open_tags($heading);
    anchor_label('genindex',$CURRENT_FILE,$_);		# this is added
    '<br>\n' . $pre . $_;
}

# $idx_module_mark will be replaced with the real index at the end
sub do_cmd_textohtmlmoduleindex {
    local($_) = @_;
    $TITLE = $idx_module_title;
    &anchor_label("modindex",$CURRENT_FILE,$_);
    '<p>' . make_section_heading($idx_module_title, "h2")
      . $idx_module_mark . $_;
}

# The bibliography and the index should be treated as separate sections
# in their own HTML files. The \bibliography{} command acts as a sectioning command
# that has the desired effect. But when the bibliography is constructed 
# manually using the thebibliography environment, or when using the
# theindex environment it is not possible to use the normal sectioning 
# mechanism. This subroutine inserts a \bibliography{} or a dummy 
# \textohtmlindex command just before the appropriate environments
# to force sectioning.

# XXX	This *assumes* that if there are two {theindex} environments, the
#	first is the module index and the second is the standard index.  This
#	is sufficient for the current Python documentation, but that's about
#	it.

sub add_bbl_and_idx_dummy_commands {
    my $id = $global{'max_id'};

    s/([\\]begin\s*$O\d+$C\s*thebibliography)/$bbl_cnt++; $1/eg;
    s/([\\]begin\s*$O\d+$C\s*thebibliography)/$id++; "\\bibliography$O$id$C$O$id$C $1"/geo
      #if ($bbl_cnt == 1)
    ;
    #}
    #----------------------------------------------------------------------
    # (FLD) This was added
    my(@parts) = split(/\\begin\s*$O\d+$C\s*theindex/);
    if (scalar(@parts) == 3) {
	# Be careful to re-write the string in place, since $_ is *not*
	# returned explicity;  *** nasty side-effect dependency! ***
	print "\nadd_bbl_and_idx_dummy_commands ==> adding module index";
	my $rx = "([\\\\]begin\\s*$O\\d+$C\\s*theindex[\\s\\S]*)"
	         . "([\\\\]begin\\s*$O\\d+$C\\s*theindex)";
	s/$rx/\\textohtmlmoduleindex \1 \\textohtmlindex \2/o;
    }
    else {
	$global{'max_id'} = $id; # not sure why....
	s/([\\]begin\s*$O\d+$C\s*theindex)/\\textohtmlindex $1/o;
	s/[\\]printindex/\\textohtmlindex /o;
    }
    #----------------------------------------------------------------------
    &lib_add_bbl_and_idx_dummy_commands()
        if defined(&lib_add_bbl_and_idx_dummy_commands);
}

# The bibliographic references, the appendices, the lists of figures and tables
# etc. must appear in the contents table at the same level as the outermost
# sectioning command. This subroutine finds what is the outermost level and
# sets the above to the same level;

sub set_depth_levels {
    # Sets $outermost_level
    my $level;
    #RRM:  do not alter user-set value for  $MAX_SPLIT_DEPTH
    foreach $level ("part", "chapter", "section", "subsection",
		    "subsubsection", "paragraph") {
	last if (($outermost_level) = /\\($level)$delimiter_rx/);
    }
    $level = ($outermost_level ? $section_commands{$outermost_level} :
	      do {$outermost_level = 'section'; 3;});

    #RRM:  but calculate value for $MAX_SPLIT_DEPTH when a $REL_DEPTH was given
    if ($REL_DEPTH && $MAX_SPLIT_DEPTH) { 
	$MAX_SPLIT_DEPTH = $level + $MAX_SPLIT_DEPTH;
    } elsif (!($MAX_SPLIT_DEPTH)) { $MAX_SPLIT_DEPTH = 1 };

    %unnumbered_section_commands = (
          'tableofcontents', $level
	, 'listoffigures', $level
	, 'listoftables', $level
	, 'bibliography', $level
	, 'textohtmlindex', $level
	, 'textohtmlmoduleindex', $level
        );
    $section_headings{'textohtmlmoduleindex'} = "h1";

    %section_commands = ( 
	  %unnumbered_section_commands
        , %section_commands
        );

    make_sections_rx();
}


# Fix from Ross Moore for ']' in \item[...]; this can be removed once the next
# patch to LaTeX2HTML is released and tested ... if the patch gets included.
# Be very careful to keep this around, just in case things break again!
#
sub protect_useritems {
    local(*_) = @_;
    local($preitems,$thisitem);
    while (/\\item\s*\[/) {
        $preitems .= $`;
	$_ = $';
        $thisitem = $&.'<<'.++$global{'max_id'}.'>>';
        s/^(((($O|$OP)\d+($C|$CP)).*\3|<[^<>]*>|[^\]<]+)*)\]/$thisitem.=$1;''/e;
        $preitems .= $thisitem . '<<' . $global{'max_id'} . '>>]';
	s/^]//;
    }
    $_ = $preitems . $_;
}

# This changes the markup used for {verbatim} environments, and is the
# best way I've found that ensures the <dl> goes one the outside of the
# <pre>...</pre>.
#
# Note that this *must* be done in the init file, not the python.perl
# style support file.  The %declarations must be set before initialize()
# is called in the main script.
#
%declarations = ('preform', '<dl><dd><pre></pre></dl>',
		 %declarations);

1;	# This must be the last line
back to top