netrik hacker's manual
>========================<
[This file contains a description of the layouting module. See hacking.txt or hacking.html for an overview of the manual.]
0. Overview
The whole layouting is split up into several, fairly simple passes, which are executed one after the other. See the notes in hacking.* for a discussion of this approach.
The first pass is the parse_syntax() function, which creates a Syntax Tree of the document. This tree contains all HTML elements and their content, but the elements have no special meaning yet.
dump_tree() can be used to output the syntax tree.
In the next pass (parse_elements()), all element and attribute names are looked up in tables and stored as enums to facialiate further processing.
If not compiled with -DXHTML_ONLY, an additional pass is inserted after element parsing: In sgml_rework(), the syntax tree is modified to fix the wrong element nesting caused by missing end tags in SGML documents.
dump_tree() can be used again to dump all element and attribute types as found in the lookup, and the possibly modified tree structure.
The third pass is the central processing step. parse_struct() interprets the elements and their attributes, and creates a Structure Tree, which contains all the items that will be visible on the output page.
The fourth pass prepares the page for rendering. In pre_render(), all items created in parse_struct() are assigned actual sizes and positions in the output page. Also, a structure "page_map[]" is created, needed for fast lookup what items are present in any given line.
All of the passes mentioned above are necessary to prepare the rendering, and are executed from layout().
The actual rendering is done in render.c. However, this isn't done for the whole page like the other layouting passes. Instead, every time some region of the output page needs to be displayed, render() is called to render exactly that region.
Alternatively, The whole page can be dumped to the terminal line by line, using dump().
The third function in render.c is dump_items(). This is not really a rendering function; it only dumps the item tree, including the (coloured) text.
1. layout.c
This file forms the framework for the layouting process. It contains functions to load a file and prepare it for rendering, but also to free the memory used by a document when it is no longer needed.
layout()
layout() is given a URL of a file or web resource to load, and does all actions necessary to be able to render the corresponding page.
Before starting any of the loading or layouting operations, a descriptor is allocated where all the data structures created inside layout() will be stored.
The descriptor is a "struct Layout" pointer. It contains the following data:
- A pointer to the input ressource descriptor ("input")
- An additional pointer "url" to the effective page URL, necessary to hold the URL after the input resource descriptor is freed
- Pointers to all data structures necessary for the layouting ("syntax_tree", "item_tree", "page_map[]")
- Pointers to the "link_list" and "anchor_list" data structures
After allocating the descriptor, layout() first opens the resource with init_load(). (Described in hacking-load.*)
Afterwards, parse_syntax(), parse_elements(), sgml_rework(), parse_struct(), and pre_render() are called in sequence. These functions are responsible for preparing the page for rendering.
The file loading itself is done inside parse_syntax(), which uses the load() function from load.c (see hacking-load.*) to read a data block every time the input buffer is empty. It processes the data in the buffer character by character (keeping track of the current read position by "input->buf_ptr"), and when it reaches the end it calls load() again to get the next data block.
After parse_struct(), the syntax tree is no longer needed. It is freed by free_syntax().
At this point also the "link_list" and "anchor_list" data structures are created using make_link_list() and make_anchor_list(). (See hacking-links.* )
free_layout()
When a page is unloaded (usally before loading a new page), this function is called to free all the data structures created by the layouting process to allow rendering (Item tree, page map, link list, anchor list), i.e. all data stored in the "Layout" descriptor, except for "input" and "syntax_tree", which are already freed during the layouting process. (s.a.) The descriptor itself is also freed.
2. parse-syntax.c
The first thing to be done when layouting is parsing the syntax of the input file. This is done by parse_syntax(). This function creates a syntax tree. The pointer to the head of this tree is returned to layout() and stored as "layout->syntax_tree" there.
Syntax Tree
Every node of "syntax_tree" is a structure of the type "Element" (defined in "syntax.h") and corresponds to one HTML element. (An element in an HTML document is represented by an HTML-tag, and the corresponding end tag, if any.) End tags do not create tree nodes, as they only close elements already stored.
For the supplied "test/0.html":
header text
<html><head> </head>
<body>
<h1> heading </h1>
<p>
first paragraph of text;
includes multiple spaces and newlines,
<em> emphasized text </em>and
<strong> strong text </strong>
</p>
<p>
<center>starting with an evil center tag,</center>
this very long second paragraph contains some special characters (including a simple space...):
&; <>"=/ plus a big gap and two unicode escapes
(decimal: ¡ and hexal: ¿)
but also an anchor em<a name="anchor" href="">bed</a>ded inside a word
(this anchor also is the only tag with parameters);
and finally a blank row <br /> (a single tag)
</p>
</body>
</html>
the syntax tree looks like this:
++>NULL
+
+---+
| ! |-. <++
+---+ | +
v +
,-----------.
("header text")
`-+------+--'
| html |-. <++++++++
+------+ | <+ +
v + +
+------+ +------+ <+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
| head |->| body |-. <++++++++++++++ +
+------+ +------+ | <+ + +
v + + + <+++++++++++++++++++++++++++++++++++++++++++++++++
+----++ +---+ <++++++++++++++++++++ +---+ <++++++++++++++++++++++++++++++++ +
| h1 |-. <+ ,->| p |---. <++ + ,->| p |-. <++++++++++++++++++ + + (back
+----+ | + | +---+ | + + | +---+ | <+ + + + to top)
v + | v + + | | + + + + ^
,--------. | ,------------------. ,----. | | + ,----------. ,---------. ,---------------. |
(" heading") | (" first...newlines,") (" and") | v + (" this...em") ("ded...row") (" (a single tag)") |
`--+---+-' | `-----+----+-------' +-+----+-+ | +--------+ `---+---+--' `--+----+-' `-----+---+-----' |
| ? |----' | em |-. <++ ,->| strong |-. <++ | | center |-. <++ ,->| a |-. <+ ,->| br |---------->| ? |--------'
+---+ +----+ | + | +--------+ | + | +--------+ | + | +---+ | + | +----+ +---+
v + | v + | v + | v + |
,----------------. | ,------------. | ,----------------. | ,---.+ |
(" emphasized text") | (" strong text") | (" starting...tag,") | ("bed") |
`--------+---+---' | `-----+---+--' | `--------+---+---' | +---+ |
+++> "parent" | ? |------' | ? |-----' | ? |------' | ? |--'
---> "list_next" +---+ +---+ +---+ +---+
(I'm really curious if anyone can read this ;-) )
The "Element" structure includes:
- The pointers "list_next" and "parent" describe the tree structure. "parent" points to the element (node) which contains this one in its content (text) area. "list_next" points to the next element as it appears in the input stream.
- The "closed" flag is a helper flag for sgml_rework() and has no meaning outside of it.
- The union "name" describes what kind of element this node represents. ("html", "head" etc.) It can store the element name either as a pointer to a string (as appears in the input stream), or as an enum number.
- "attr_count" stores the number of attributes of this element. (Attributes are the parameters of an element, which appear inside the start tag, like: href="foo" etc.)
- "attr" points to an array of "Attr" structures. Each of these structures contains the data for one attribute; it consists of a union of type "Attr_name", which, like "Element_name", stores the attribute name either as a string or as an enum; and a union of type "Attr_value", which stores the value of the attribute. (String or number.)
- The "content" string stores the content. (The text between the tags.) Every element stores the content between the previous tag and the start tag of this element. Thus it does not store the content of the element itself, but part of the content of the *parent* element. This simplifies processing a bit, because this way no facility for storing content blocks divided by sub-elements is needed -- the sub-elements store the content themselves. The caveat is that a lot of dummy elements are needed to store the content if no further sub-element follows them. This is quite a big inefficiency, as nearly every real element also needs a dummy element to store its content. This should change in the future -- if we won't drop the syntax-tree in its present form at all... Which we will :-)
Initialization
Before starting parsing, we have to create the tree top. (We call it the global element.) This is done by setting "cur_el" to NULL and calling add_element().
add_element()
This function creates a new node and inserts it into the syntax tree; thus it has to set the "parent" and "list_next" pointers too, and adjust some pointes of other nodes to point to this one.
"parent" is set to "cur_el", as any new tag is created while parsing the content area of its parent. "list_next" is set to NULL, as the new node is always the last one in the list. "list_next" of "last_el" (the last node in the list up to now) is set to point to the new node; this is omitted if "cur_el" is NULL, indicating that there are no other nodes yet.
Parsing
The parser itself works in a very simple way. It is some kind of state machine. For every input character, one action is taken, selected by a dispatcher depending on the current state (stored in "parse_mode") and the input char itself. Several combinations (e.g. tag start) change the current state, thus the following character(s) are parsed in a different mode. (Other actions are taken.)
Sometimes a character that causes a mode switch has to be parsed in the new mode itself. In this case the flag "recycle" is set after the mode change, causing the dispatch to be repeated for the same char, but in the new mode.
Again, the parsing is not very efficient in the present implementation. (In fact, it is by far the most time consuming part of the whole layouting.) Especially the huge switch is quite slow. (Good compilers have a fairly efficient implementation of the switch itself; however, it still causes many unpredictable branches.) There are some possibilities to optimize this. The bigger problem is that the inner loop is quite big, and may not fit into the processor's instruction cache, thus making it terribly slow. Maybe splitting the parsing into several simpler passes would help. However, we are planning to switch to a completely different, (hopefully) much more efficient parser system in the next major release...
The default parsing mode is "PM_CONTENT", which is the mode for parsing element content. Any normal character encountered in this mode is simply added to "text_buf" by "buf_add_char()". A ' ', '\t', '\n', '\r' or '\f' aren't stored; we switch to "PM_BLANK" instead. Any following blank space is ignored. As soon as a normal character occurs again, we store a single ' ' and swich back to "PM_CONTENT".
input:
first paragraph of text;
includes multiple spaces and newlines,
^
file position
text_buf: " first paragraph of text; includes mul"
<pre> Blocks
After a <pre> tag, the mode isn't switched back to "PM_CONTENT", but to "PM_PRE". In this mode all blank space characters are stored to "text_buf" as non-breakable spaces, except newlines which are stored directly.
The mode is ended and switched to "PM_CONTENT" again when a closing "</pre>" tag is encountered.
References
An '&' indicates a character reference (unicode escape) or entity reference (named escape), and starts the reference parsing mode "PM_AMP". On entering this mode, the current write position in "text_buf" is saved to "amp_pos".
input: [...] < [...]
^
text_buf: "[...] &"
^^
text_buf_len
amp_pos
In this mode characters are added to "text_buf", too. The mode is switched back to the previous mode (saved in "prev_mode_amp") when a ';' occurs.
input: [...] < [...]
^
text_buf: "[...] <"
^ ^
text_buf_len
amp_pos
The text between the saved start position of the escape sequence and the current positon is converted then. First the string is looked up in "ref_table[]", which is a table of named characters, defined in facilities.c. If it isn't there, but it seems to be a numerical character reference, it is converted to an integer (ASCII value). If a replacement char was found either in the table or by the number conversion, the reference is removed from "text_buf" and the replacement char is inserted instead.
text_buf: "[...] <"
^
text_buf_len
If no replacement was found, the string is left unchanged. (We consider this behaviour better then of some other browsers, which simply leave unknown characters out, and the user only wonders that something is missing...) Probably we will mark unknown escapes by some attribute in the future.
Tags
A '<' starts tag parsing. There is a a whole bunch of tag parsing modes. The one entered after the '<' is "PM_TAG_START", which indicates that the tag name should follow next.
input: [...] but also an anchor em<a name="anchor" href="">bed</a>ded inside a word
^
text_buf: "[...] but also an anchor em"
tree:
+
+
+---+
-->| p |-. <== cur_el
+---+ | <+
| +
| +
v +
+--------+
| center |-. <++
+--------+ | +
v +
,----------------.
(" starting...tag,")
`--------+---+---' <-- last_el
| ? |->NULL
+---+
Start Tags
If the following character is an normal char, it's a start tag. (Or a single tag, which is treated the same way for now.) "PM_TAG_NAME" is entered. A new element node is created by add_element(). Any content in front of this new element, which was stored in "text_buf" up to now, is stored to the new node's "text" field by "insert_buf".
input: [...] but also an anchor em<a name="anchor" href="">bed</a>ded inside a word
^
text_buf: ""
tree:
+
+---+
->| p |-. <++++++++++++++++++
+---+ | <+ +
| + <== +
| + ,----------.
v + (" this...em")
+--------+ `---+---+--' <--
| center |-. <++ ,->| |->NULL
+--------+ | + | +---+
v + |
,----------------. |
(" starting...tag,") |
`--------+---+---' |
| ? |------'
+---+
Normal characters encounterd in "PM_TAG_NAME" mode (including the one that started the mode) are stored to "text_buf".
input: [...] but also an anchor em<a name="anchor" href="">bed</a>ded inside a word
^
text_buf: "a"
A blank space character ends "PM_TAG_NAME" and switches to "PM_TAG", which indicates that attributes may follow. "text_buf" is stored as the element name.
input: [...] but also an anchor em<a name="anchor" href="">bed</a>ded inside a word
^
text_buf: ""
tree:
+
+---+
->| p |-. <++++++++++++++++++
+---+ | <+ +
| + <== +
| + ,----------.
v + (" this...em")
+--------+ `---+---+--' <--
| center |-. <++ ,->| a |->NULL
+--------+ | + | +---+
v + |
,----------------. |
(" starting...tag,") |
`--------+---+---' |
| ? |------'
+---+
A following normal char is the beginning of an attribute name, and switches to "PM_ATTR_NAME". Characters encounterd in "PM_ATTR_NAME" mode are also stored to "text_buf".
input: [...] but also an anchor em<a name="anchor" href="">bed</a>ded inside a word
^
text_buf: "name"
The attribute name ends with an '=' or a blank char. A new entry is created in the "attr[]" array, and "text_buf" is stored as the attribute name. Mode is switched to "PM_ATTR_NAME_END" first, which indicates that the attribue value should follow.
input: [...] but also an anchor em<a name="anchor" href="">bed</a>ded inside a word
^
text_buf: ""
attr: name: data:
"name" ""
If the attribute name was ended by an '=', mode is switched immediately to "PM_ATTR_VALUE", otherwise as soon as an '=' is encountered. (After any amount of whitespace.)
White space in "PM_ATTR_VALUE" mode (after the '=') is ignored too.
Next char must be either a '"' or a '\'', and switches to "PM_ATTR_DATA_QUOT" or "PM_ATTR_DATA_APOS", respectively. In this modes characters are stored to "text_buf" again.
input: [...] but also an anchor em<a name="anchor" href="">bed</a>ded inside a word
^
text_buf: "anchor"
attr: name: data:
"name" ""
A second '"' (or '\'', respectively) ends this mode. "text_buf" is stored as the attribute value for the new "attr" entry.
input: [...] but also an anchor em<a name="anchor" href="">bed</a>ded inside a word
^
text_buf: ""
attr: name: data:
"name" "anchor"
Mode is swiched back to "PM_TAG". Now blank space may follow (which is ignored), followed by another attribute.
input: [...] but also an anchor em<a name="anchor" href="">bed</a>ded inside a word
^
In "PM_TAG" mode also a '>' may occur, ending tag parsing and switching back to the mode before tag parsing had begun. (PM_CONTENT or PM_BLANK.) In this case, "cur_el" is set to "last_el"; this means descending in the syntax tree to the newly created node.
input: [...] but also an anchor em<a name="anchor" href="">bed</a>ded inside a word
^
text_buf: ""
attr: name: data:
"name" "anchor"
"href" ""
tree:
+
+---+
->| p |-. <++++++++++++++++++
+---+ | <+ +
| + +
| + ,----------.
v + (" this...em")
+--------+ `---+---+--' <==
| center |-. <++ ,->| a |->NULL <--
+--------+ | + | +---+
v + |
,----------------. |
(" starting...tag,") |
`--------+---+---' |
| ? |------'
+---+
A '>' may also occur in "PM_TAG_NAME" mode, meaning the element has no attributes.
input: <html> <head> [...]
^
In this case creating the new node and storing the name, and descending into the element are done in one step. (By "recycle".)
End Tags
If the first character after the '<' is a '/', the tag is an end tag, and we switch to "PM_END_TAG_NAME". If any text is pending in "text_buf", we have to store it somewhere. As an end tag normally does not create a new element node, we have to create a dummy node for this. (Very inefficient, s.a.)
input: [...] but also an anchor em<a name="anchor" href="">bed</a>ded inside a word
^
tree:
+
+---+
->| p |-. <++++++++++++++++++
+---+ | <+ +
| + +
| + ,----------.
v + (" this...em")
+--------+ `---+---+--' <==
| center |-. <++ ,->| a |-. <+
+--------+ | + | +---+ | +
v + | v +
,----------------. | ,---.+
(" starting...tag,") | ("bed")
`--------+---+---' | +---+ <--
| ? |------' | ? |->NULL
+---+ +---+
Normal chars in "PM_END_TAG_NAME" mode are stored to "text_buf", too. "PM_END_TAG_NAME" can be ended immediately by a '>', or by blank space (switching to "PM_END_TAG_SPACE") followed by '>'.
The tag name extracted to "text_buf" is compared against the element name of the current element, to see if the end tag matches, and then abdannoned. The element is closed by ascending to the parent.
input: [...] but also an anchor em<a name="anchor" href="">bed</a>ded inside a word
^
tree:
+
+---+
->| p |-. <++++++++++++++++++
+---+ | <+ +
| + <== +
| + ,----------.
v + (" this...em")
+--------+ `---+---+--'
| center |-. <++ ,->| a |-. <+
+--------+ | + | +---+ | +
v + | v +
,----------------. | ,---.+
(" starting...tag,") | ("bed")
`--------+---+---' | +---+ <--
| ? |------' | ? |->NULL
+---+ +---+
Single Tags
If a '/' appears instead of an attribute name in "PM_TAG" mode, "parse_mode" is set to "PM_SINGLE_TAG", indicating an (XHTML) single tag.
input: [...] and finally a blank row <br /> (a single tag)
^
The '/' can also immediately follow the element name. (In "PM_TAG_NAME" mode.)
input: <hr/>
^
In this case, creating the node and switching to a single tag are done in one step by "recycle".
In any case, a '>' has to follow, and switches back to normal mode just like in a start tag, only it does not descend (set the new node as "cur_el") -- a single tag has no content area; the content following a single tag still belongs to the parent.
input: [...] and finally a blank row <br /> (a single tag)
^
tree:
+
+---+ <++++++++++++++++++++++++++++++++
->| p |-. <++++++++++++++++++ +
+---+ | <+ + +
| + <== + +
| + ,----------. ,---------.
v + (" this...em") ("ded...row")
+--------+ `---+---+--' `--+----+-' <--
| center |-. <++ ,->| a |-. <+ ,->| br |->NULL
+--------+ | + | +---+ | + | +----+
v + | v + |
,----------------. | ,---.+ |
(" starting...tag,") | ("bed") |
`--------+---+---' | +---+ |
| ? |------' | ? |--'
+---+ +---+
Comments
In "PM_TAG_START" mode (after the '<'), also an '!' can follow, indicating that we have not any tag at all, but either a comment, a DOCTYPE declaration, or a CDATA section. "parse_mode" is set to "PM_EXCLAM" in this case.
input: some text <!--a test-comment--> and more text
^
If a '-' follows, it's a comment. Mode is switched to "PM_COMMENT_START".
input: some text <!--a test-comment--> and more text
^
Now a second '-' has to follow, switching to "PM_COMMENT". In this mode any characters but a '-' are simply ignored.
input: some text <!--a test-comment--> and more text
^
text_buf: "some text"
A '-' switches to "PM_COMMENT_END1", which means that it *may* be the comment end.
input: some text <!--a test-comment--> and more text
^
However, if it is followed by any other char than a second '-', mode is switched back to "PM_COMMENT".
input: some text <!--a test-comment--> and more text
^
A second '-' in "PM_COMMENT_END1" switches to "PM_COMMENT_END2", which means that now the comment really ends.
input: some text <!--a test-comment--> and more text
^
Now the '>' has to follow, and switches back to parsing mode before the beginning of the comment ("prev_mode_tag").
input: some text <!--a test-comment--> and more text
^
DOCTYPE Declarations
If a normal char occurs in "PM_EXCLAM", we assume it is the "D" in "<!DOCTYPE".
input: garbage <!DOCTYPE somedoc> more garbage
^
We treat DOCTYPE declarations as comments. Any characters but '>' are ignored.
input: garbage <!DOCTYPE somedoc> more garbage
^
text_buf: "garbage"
A '>' returns to normal mode.
input: garbage <!DOCTYPE somedoc> more garbage
^
This isn't a very reliable detection, as according to the grammer, an unescaped '>' may appear in some system literal inside the declaration. However, we assume that this won't happen... (We would have to parse the whole declaration otherwise.)
CDATA Sections
A '[' in "PM_EXCLAM" mode starts a CDATA section, indicated by "PM_CDATA_START". If there is a pending blank ("prev_mode_tag" is "PM_BLANK"), it has to be stored *before* the CDATA.
input: some text <![CDATA[a tricky ]> CDATA section]]> and more text
^
text_buf: "some text "
Following normal chars (should) belong to the "CDATA" string, and are ignored.
input: some text <![CDATA[a tricky ]> CDATA section]]> and more text
^
text_buf: "some text "
A second '[' in "PM_CDATA_START" mode switches to "PM_CDATA", indicating that the actual data will follow.
input: some text <![CDATA[a tricky ]> CDATA section]]> and more text
^
Any characters in "PM_CDATA" mode but '>' are stored directly to "text_buf".
input: some text <![CDATA[a tricky ]> CDATA section]]> and more text
^
text_buf: "some text a tricky ]"
When a '>' occurs, the previous two chars (in "text_buf") are tested against "]]". If they do not match, the '>' is simply stored just as any other character.
input: some text <![CDATA[a tricky ]> CDATA section]]> and more text
^
text_buf: "some text a tricky ]>"
If they match, the last two characters are removed from "text_buf" (they belong to the CDATA terminator), and mode is switched back to "PM_CONTENT". (It doesn't need to be switched back to the mode before the CDATA section, as any pending blanks already have been stored, and a CDATA section can't start in other modes than "PM_TEXT" or "PM_BLANK".)
input: some text <![CDATA[a tricky ]> CDATA section]]> and more text
^
text_buf: "some text a tricky ]> CDATA section]]"
input: some text <![CDATA[a tricky ]> CDATA section]]> and more text
^
text_buf: "some text a tricky ]> CDATA section"
Processing Instructions
The '<' may also be followed by a '?', indicating a processing instruction. Mode is switched from "PM_TAG_START" to "PM_INSTR".
input: some text <?a fake? processing instruction??> more text
^
Processing instructions are also treated as comments. Any chars but '?' are ignored in "PM_INSTR".
input: some text <?a fake? processing instruction??> more text
^
text_buf: "some text"
A '?' switches to "PM_INSTR_END", indicating this *may* be the end of the processing instruction.
input: some text <?a fake? processing instruction??> more text
^
If a normal char follows the '?', mode is switched back to "PM_INSTR".
input: some text <?a fake? processing instruction??> more text
^
If a second '?' follows, "PM_INSTR_END" is kept, as the first one isn't the end of the processing instruction, but the new one could be.
input: some text <?a fake? processing instruction??> more text
^
A '>' in PM_INSTR_END really ends the processing instruction, and switches to "prev_mode_tag".
input: some text <?a fake? processing instruction??> more text
^
SGML Mode
When compiled without the "-DXHTML_ONLY" option, a few cases more are possible.
Unclosed Tags
In SGML, not every element has to have an end tag.
When an end tag is encountered, we ascend in the syntax tree not only once, but until an element is found that matches the end tag. Thus, all elements in between are automatically closed.
input:
<body>
<p>
some text
<hr>
</p>
^
+------+
| body |-.<+
+------+ | +
v +
+---+
| p |-.<+
+---+ | +
v +
,-------.
(some text)
`-+----+'
--> | hr |->NULL
==> +----+
input:
<body>
<p>
some text
<hr>
</p>
^
+------+
| body |-.<+ <==
+------+ | +
v +
+---+
| p |-.<+
+---+ | +
v +
,-------.
(some text)
`-+----+'
--> | hr |->NULL
+----+
Unquoted Attribute Values
When a normal char occurs in "PM_ATTR_VALUE" mode, "PM_ATTR_DATA_NOQUOTE" is entered.
input: <sometag someattribute=somevalue minimized third="nothing">
^
This mode is just like "PM_ATTR_DATA_QUOT" or "PM_ATTR_DATA_APOS", only it is ended by a blank or the tag end.
input: <sometag someattribute=somevalue minimized third="nothing">
^
Mimimized Attributes
In SGML, attributes without a value are possible. This is recognized when a normal char or the tag end occurs in "PM_ATTR_NAME_END" mode instead of the '='.
input: <sometag someattribute=somevalue minimized third="nothing">
^
The attribute is ended immediately. "text_buf[]" (which is empty in this case) is stored just like at the end of an unquoted attribute value. Mode is set to "PM_TAG", and the current character (the tag end or beginning of next attribute) is processed in this mode.
SGML Comments
Comments also allow more complicated syntax. For one, blank space is possible between the "--" ending the comment string and the '>' ending the declaration. Thus, blank space in PM_COMMENT_END2 is ignored.
Moreover, another comment string may follow the end. Thus, a '-' in PM_COMMENT_END2 switches back to PM_COMMENT_START, similary to the '-' after the "<!".
input: <!--comment start-- --second comment string in same declaration-- >
^
Finally, SGML also allows empty declarations ("<!>"), which are also a kind of comment. Thus a '>' in PM_EXCLAM switches immediately to PM_COMMENT_END2 and recycles.
input: <!>
^
Finishing
Parsing is ended by EOF. This should only appear in "PM_CONTENT" or "PM_BLANK" mode (not inside some tag, comment, CDATA section or chracter/entity reference), and only if the current element is the global one (not while parsing some element's content).
The "list_next" pointer of the last node is set to point back to the tree top. This faciliates easier processing in the following steps.
Error Handling
When using -DXHTML_ONLY, every syntax error encountered causes netrik to print an error message and immediately quit. (The XML standard requires this.)
Without -DXHTML_ONLY, netrik is more tolerant.
Workarounds
Netrik uses simple workarounds for some of the most common cases of broken HTML:
If some illegal char occurs in a entity/character reference, it's probably not really a reference, but an unescaped '&'. We keep the whole sequence literally and switch back to "prev_mode_amp".
input: x = a & b
^
text_buf: "x = a & "
Similar for illegal characters in "PM_TAG_START", which indicate an unescaped '<'. We store a '<' and switch back to "prev_mode_tag".
input: if(a < b)
^
text_buf: "< "
As SGML comments have a quite complicated syntax, reasonable error handling is also a bit more complicated.
If someting else then '>' (end of comment declaration), '-' (beginning of second comment string), or blank space follows in PM_COMMENT_END2 (after a "--"), then the "--" was probably not intended to have any special meaning, but simply to be part of the comment. Thus, mode is switched back to PM_COMMENT.
input: <!-- some broken -- comment -->
^
The same is done for unexpected characters in PM_COMMENT_START mode, which is most common for "---" inside a comment.
input: <!-- some broken --- comment -->
^
There is one exception to this, however: If a '>' follows in PM_COMMENT_START mode, and it was preceeded not only by one '-' (the one which started PM_COMMENT_START) but two or more, then the the '>' together with the last two '-' was probably intended as as an XML-like "-->" comment end.
input: <!--- anything --->
^
parse_mode: PM_COMMENT_START
dash_count: 3
The "dash_count" variable keeps track of how many dashes have been encountered in a row; it is incremented every time a '-' apprears in some of the comment parsing modes, and is reset to 0 every time some other character is encountered.
"dash_count" is also used in another situation: If a '>' follows in PM_COMMENT or PM_COMMENT_END1, normally it is part of the comment. The '>' is ignored and mode stays PM_COMMENT. (Or is switched back from PM_COMMENT_END1.)
input: <!-- comment with > and -> in it -->
^
However, if there were two or more dashes in front of the '>', this "-->" combination was probably also intended as a comment end. A comment consisting of a series of dashes is a typical example:
input: <!------>
^
parse_mode=PM_COMMENT
dash_count=6
However, only a little warning can be printed in this case -- this is valid SGML, and *has* to be treated as part of the comment, even if it's probably not what the page author intended! Printing an error and using a workaround would mean presumably to violate the standard in favour of broken pages, which is probably not a very good idea...
All other syntax errors are simply ignored, hoping for the best.
html_error()
Whenever some syntax error is detected (no matter whether workarounds are available), html_error() is called, with several parameters describing the error. This function is responsible for everything that needs to be done when an error occurs.
First, it prints an error message. The message text is passed from parse_syntax(), and used as the format string for printf(). If the error message requires additional arguments, they are passed at the end of the parameter list when calling html_error().
If the parsing mode requires that, it quits immediately afterwards. The mode is determined by the config variable "cfg.parser", which is an "enum Parser_mode", with the possible values CORRECT_HTML, NORMAL_HTML, BROKEN_HTML and IGNORE_BROKEN. The parser quits only in CORRECT_HTML mode, or when -DXHTML_ONLY is enabled. If the input resource from which the page is loaded is a pipe from wget (see hacking-load.*), the pipe is closed before quitting to assure a cleaner exit.
In all other modes, an additonal message passed from parse_syntax() is printed afterwards, informing in which way netrik will handle the error. (workaround, ignore etc.)
Finally the error level passed from parse_syntax() is compared against the highest error level up to know, and the new higest level is returned.
$$Warning messages
parse_syntax keeps track of most severe syntax error that was found while parsing the page in "err_level", which is of type "enum Syntax_error" and can have the following values:
- SE_NO: no errors were found
- SE_BREAK: the user issued an interrupt (SIGINT) while loading the document. This isn't really an error, but can be handled very convenient this way...
- SE_WORKAROUND: only errors were found which seem uncritical and workarounds were applied
- SE_HEAVY: there were heavy syntax errors, for which no workaround is known, and quite likely may cause the page to be layouted incorrectly (some parts of text may be missing etc.)
- SE_CRITICAL: the parser was not in normal text mode at the page end; this means that some important char was missing (e.g the '"' at the end of attribute values or the '>' closing a tag), and the following text was misinterpreted in will be rendered incorrectly.
- SE_FAIL: This isn't really a syntax error. It is not used inside parse_syntax() itself; it's only set before returning when a file loading error was detected, for the sake of the calling function.
- SE_NODATA: Similar to SE_FAIL. This is set if EOF is returned by load() before *any* data has been read.
After the whole page is parsed, a warning message is printed if some error was found. The message text depends on the error level. The error level is also passed back to main(), which then waits for a keypress before starting the pager, so the message will be seen.
In IGNORE_BROKEN mode the warning is suppressed, and "err_level" is reset. In BROKEN_HTML mode only SE_WORKAROUND errors are ignored, while more critical errors still cause a warning.
SE_BREAK is set if EOF is returned, but at the same time "input->user_break" has been set, indicating that it's not really EOF, but transfer was interrupted by the user. Other errors are supressed in this mode, as a user break during loading might cause several syntax errors (unclosed elements etc.) with the page itself being not to blame for.
SE_NODATA is set if EOF is returned by load() before any data has been read. (This can be caused by failure to open the resource, but also by an empty file/http response.) It's handled like a normal syntax error; the only difference is that it can't be masked even by IGNORE_BROKEN. The syntax tree consists only of the global element; it will be correctly rendered to an empty page.
SE_FAIL isn't set during parsing. Before returning, parse_syntax() checks whether "input->type" is RES_FAIL; if it is, an error message is printed, and SE_FAIL is set (so main() knows an error occured). However, this test is only necessary if SE_NODATA isn't set; otherwise, an error message has already been printed and an error code would be returned anyways. As EOF is returned by load() also when an error occured, SE_NODATA is already set for most errors; SE_FAIL is only used if the error occurs after some data could be read.
free_syntax()
This function is responsible for freeing the memory used by the syntax tree when it is no longer needed.
The whole tree is traversed by "list_next", and the element nodes are freed one by one.
As the "list_next" pointer is necesary to find the next node, but not longer available after freeing the current node, it is saved in "next_el" before freeing. At the the beginning of the next iteration this is copied to "cur_el".
Before freeing the element node itself, all dynamic data belonging to the node has to be freed.
3. dump-tree.c
dump_tree() is primarily used for dumping the syntax tree generated by parse_sytax() for debugging purposes. The reason it resides in an own file is that it could be easily modified to be a really useful function for dumping a HTML document's structure. This may be implemented in the future, if someone shows interest...
The implementation of dump_tree() is quite straightforward, as the function only needs to print every node in the order it occured in the HTML file.
For every node, first the text is printed. (If "dump_content" is given.) The reason it is printed in front of the node itself is because it's also the content in front of the element in the original HTML file.
Next, the current tree depth is shown by a number of '|'. The current depth is always stored in "depth".
Afterwards the element name is printed, and all attributes with their values. Depending on "elements_parsed", either the raw values extracted from the document are printed, or the transformed values generated by parse_elements().
The next node is reached by the "list_next" pointer of the current node. Before going to the next node the tree depth of the new node needs to be calculated. This is done by assuming that the next node is below the current one, and than going up, until we find the parent of the new node. This idea is explained in more detail in 6. parse-struct.c.
4. parse-elements.c
parse_elements() is responsible for making out the elements and attributes from the syntax tree. (Extraced by parse_syntax().)
All elements but the first one (which is always ELEMENT_GLOBAL) are processed one by one; the tree is traversed via "list_next". For every element, the name is looked up in "element_table[]" by comparing to all entries, in a loop.
"element_table[]" contains all names of ordinary elements, then the "?" representing ELEMENT_NO, and finally "!" representing ELEMENT_GLOBAL. (It also contains other properties of the elements; more on this in Processing in 6. parse-struct.c.)
The last two aren't checked against the element name.
As soon as a match is found, the loop is left and the entry number is stored to "syntax_tree" in place of the string. The entry number is an "enum Element_type", defined in syntax.h; it tells the element type in the following processing passes.
If none of the ordinary entries matched, the entry nuber, which now is ELEMENT_NO (as this one follows after the ordinary entries), is stored anyhow, indicating that the element is unknown. If no element name was stored ("cur_el->name.str" is NULL), indicating a dummy tag, ELEMENT_NO is set also.
After the element name, all attribute names are processed in a loop. They are looked up in "attr_table[]" the same way the element name is.
The attribute value isn't processed at all yet.
5. sgml_rework()
Before the syntax tree is further processed, sgml_rework() (from sgml.c) is applied. (Unless compiled with -DXHTML_ONLY.)
This function is responsible for fixing the problems arising from the fact that SGML allows certain end tags to be left out; thus the syntax parser doesn't recognize the elements' ends, and stores all following elements as children, even if they should actually be at the same level. (e.g. the list items in a list.) sgml_rework() goes over the complete (broken) tree, finds such situations, and unnests the elements, thus creating a correct syntax tree.
It won't be covered in too much detail here, as this is only a temporary solution; it will become obsolete with the planned new parser(s).
The recognition of the missing element ends is done by the "element_group" enum in "element_table[]". This Enum has the values GROUP_SINGLE for single tag elements (elements which mustn't have any content), GROUP_OBLIGATE for all elements where the end tag can't be left out, and several others for various kinds of elements with optional end tag.
The whole tree is scanned element by element (using "list_next"), and each one is tested to fullfill one of the offending conditions. Two things have to be handled: Unclosed single tag elements, and unclosed optional end tag elements.
The second situation is more complicated. If the element is of some type with optional end tag, it could terminate a previous (unclosed) element from the same group; e.g. a <li> will terminate the previous <li>. It doesn't terminate elements from other groups, though; a <td> inside a <tr> doesn't terminate the <tr>, for example. Thus, the group of the current element needs to be tested against the group of the parent (all elements following an unclosed one will be stored as its children!); if they are the same, the parent is actually an element that should be terminated at the position where the child starts, and the child should follow it, at the same depth. This means that the child has to be "lifted" out of the parent. However, we don't do that immediately; we only set the "closed" flag of the parent element for now, and the lifting will be done later.
However, it's not enough to test only the immediate parent: The element may follow some other unclosed element, and thus be a child of it, e.g. a <tr> following a <td>, which is inside the previous <tr>. This also needs to be recognized, and *both* the previous <tr> and the <td> have to be closed. Thus, not only the immediate parent's group is compared, but all ancestors are scanned. The scanning only stops on an element with obligate end tag -- as the element's end is always known for these, nothing will be ever stored inside that element that doesn't belong there, and nothing should be lifted out. (In nested tables for example, the <tr>s and <td>s of the inner table shouldn't mess with the ones of the outer table -- this is ensured by the scanning of the inner table's rows and columns stopping at the inner <table> element.) The "closed" flag is set for all the closed ancestors.
Some more handling is necessary due to the fact that an element node always stores the content which appears *before* the element. When an element is lifted, the content musn't be lifted also -- it appeared *before* the element, and thus also before the previous element's end, so it has to stay where it is. We have to create a new dummy tag inside the closed element therefore, taking the place of the lifted element and storing its content.
However, this isn't done when the parent was already closed. This happens if the parent is a single tag element. These elements end right where they start, not at the beginning of the next element; the content also has to be lifted out. (Nothing is allowed to stay inside a single tag element!)
The actual lifting is done after processing the element: If the parent is closed, we have to "leave" it. (This is done by setting the "parent" to the previous grandparent -- this way, the element is no longer a child of the old parent, but a sibling.) Thus the element that closed it's parent is lifted right afterwards; all following elements of the parent will be lifted also, after being processed. Any preceeding elements (as well as the possibly created new dummy) won't be lifted on the other hand, as they won't be processed anymore.
Single tag elements are handled more or less the other way round: They are not closed by some child (which turns out to actually be a sibling), but close *themselfs*, as soon as they are encountered. This way all children will be lifted out, no matter what.
No other processing is necessary for single tag elements, as they won't ever terminate some other element.
6. parse-struct.c
After the syntax tree was generated by parse_syntax(), we have to "understand" it. This is done by parse_struct(), which is the central pass of the layouting process. In this function the syntax tree, which contains a nearly 1:1 reproduction of the HTML file, is converted to an item tree, which contains a representation of what will be actually shown as the output of the browser -- text blocks, blank rows, boxes grouping severel other items.
Structure Tree
For 0.html, we have to convert the syntax tree:
++>NULL
+
+---+
| ! |-. <++
+---+ | +
v +
,-----------.
("header text")
`-+------+--'
| html |-. <++++++++
+------+ | <+ +
v + +
+------+ +------+ <+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
| head |->| body |-. <++++++++++++++ +
+------+ +------+ | <+ + +
v + + + <+++++++++++++++++++++++++++++++++++++++++++++++++
+----++ +---+ <++++++++++++++++++++ +---+ <++++++++++++++++++++++++++++++++ +
| h1 |-. <+ ,->| p |---. <++ + ,->| p |-. <++++++++++++++++++ + + (back
+----+ | + | +---+ | + + | +---+ | <+ + + + to top)
v + | v + + | | + + + + ^
,--------. | ,------------------. ,----. | | + ,----------. ,---------. ,---------------. |
(" heading") | (" first...newlines,") (" and") | v + (" this...em") ("ded...row") (" (a single tag)") |
`--+---+-' | `-----+----+-------' +-+----+-+ | +--------+ `---+---+--' `--+----+-' `-----+---+-----' |
| ? |----' | em |-. <++ ,->| strong |-. <++ | | center |-. <++ ,->| a |-. <+ ,->| br |---------->| ? |--------'
+---+ +----+ | + | +--------+ | + | +--------+ | + | +---+ | + | +----+ +---+
v + | v + | v + | v + |
,----------------. | ,------------. | ,----------------. | ,---.+ |
(" emphasized text") | (" strong text") | (" starting...tag,") | ("bed") |
`--------+---+---' | `-----+---+--' | `--------+---+---' | +---+ |
| ? |------' | ? |-----' | ? |------' | ? |--'
+---+ +---+ +---+ +---+
to this structure tree:
***> "string" (back <+++++++++
xxx> "first_child" to first +
+++> "parent" item) +-----+
===> "next" ,->| box |-->NULL
---> "list_next" | +-----+<==#
| x ^ #===#
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx +
x | +
x ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
x + + + + + + + + + |
v + + + + + + + + + |
+------+ +-------+ +------+ +-------+ +------+ +-------+ +------+ +------+ +------+ |
| text |-->| blank |-->| text |-->| blank |-->| text |-->| blank |-->| text |-->| text |-->| text |-'
+----*-+==>+-------+==>+----*-+==>+-------+==>+----*-+==>+-------+==>+----*-+==>+----*-+==>+----*-+==>NULL
x * x x * x x * x x * x * x *
v * v v * v v * v v * v * v * ,--------------.
NULL * NULL NULL * NULL NULL * NULL NULL * NULL * NULL **>("(a single tag)")
v v v v * `--------------'
,-----------. ,-------. ,------------. ,---------------. * ,----------------.
("header text") ("heading") ("first...text") ("starting...tag,") **>("this...blank row")
`-----------' `-------' `------------' `---------------' `----------------'
which, in turn, is a representation of this output page:
+-------------------------------------------------------------------------------------------------+ |+-----------+ | ||header text| | |+-----------+ | | | |+-------+ | ||heading| | |+-------+ | | | |+-----------------------------------------------------------------------------------------------+| ||first paragraph of text; includes multiple spaces and newlines, emphasized text and strong text|| |+-----------------------------------------------------------------------------------------------+| | | |+---------------------------------+ | ||starting with an evil center tag,| | |+---------------------------------+ | |+-----------------------------------------------------------------------------------------------+| ||this very long second paragraph contains some special characters (including a simple space...):|| || &; <>"=/ plus a big gap###and two unicode escapes (decimal: ¡ and hexal: ¿) but also an anchor|| || embedded inside a word (this anchor also is the only tag with parameters); and finally a blank|| || row || |+-----------------------------------------------------------------------------------------------+| |+--------------+ | ||(a single tag)| | |+--------------+ | +-------------------------------------------------------------------------------------------------+
Note that there are no actual sizes or postions for the items, and no line breaks inside the text items; this is all done in a later processing step (pre-render.c). The item tree at this point only represents the structure of the output page. (The line breaks in the fifth text block aren't really there; we have inserted them in the figure because the text block is a bit too long to put in a single line...)
The item tree looks complicated at first, but it's a quite trivial example when taking a closer look. (This is becuase at the time of creating the 0.html file used here, netrik wasn't able to do anything more complicated...) However, it should be sufficient to get the idea...
Every node of the item tree consists of an "Item" structure. This structure is declared in "items.h". It contains:
- The pointers "list_next", "next", "parent", and "first_child" connect the nodes inside the tree. "list_next" points to the next node in the order they are generated. Note that in contrast to the element tree, in the item tree any children are generated *before* the parent. "next" points to the next item at the same tree depth and in the same branch, i.e. the next sibling. "parent" points to the parent item, and "first_child" to the first sub-item. More pointers are necessary than for the element tree, because the item tree is traversed in several differnt ways while processing.
- The "center" flag indicates whether the item is centered. As not all items use this, and the exact meaning varies between different item types, it may be reasonable to move this to the item specific data. We'll decide on this as soon as enought HTML facilities are implemented.
- "x_start", "x_end", "y_start" and "y_end" define a square area inside the layouted page, in which the item is displayed. In some processing steps "x_end" end "y_end" are also "abused" to store the minimal size of the item.
- "type" is an enum storing what kind of item this node represents. (Currently ITEM_TEXT, ITEM_BOX, ITEM_FORM, ITEM_BLANK, ITEM_BLOCK_ANCHOR or ITEM_INLINE_ANCHOR.)
- "data" is an union storing all data specific to different item types. Currently this can be the pointer to a text string for text items, to an (block or inline) anchor struct, or to a form paramters struct.
add_item()
New items are created by add_item(). This function allocates a new "Item" structure, and sets some pointers. The item isn't inserted into the tree directly; it's only inserted into a list of all items at the current tree depth. This is a single linked list maintained by the "first_item" and "last_item" pointers of "state" (there is one such list for each tree depth), and the "next" pointers of the item structures. The only other pointers set are "list_next". The "parent" and "first_child" pointers aren't set; this is done later when the items are actually inserted into the tree, while ascending from the current depth. (For the first item in every tree depth the "parent" is explicitly set to NULL, indicating that there is no parent yet.)
This function is called directly to create box items, and from add_string() to create text items. Under certain conditions it also creates a blank item before the actual text item or box item; more on this later, under Blank Lines.
When called with the "virtual" flag, this function behaves slightly different: No line break/blank line handling is done; the status remains unchanged. This is for creating the Virtual Boxes used for anchors.
String
The actual text data of text items is stored in a different place. Every text item points to a "String" structure (also declared in "items.h"). This structure consists of a normal C-string containing the text itself, and an array of "Div" structures, containing all attribute information. (Color etc.) A "String" can consist of several divisions with different attributes. Every "Div" structure stores the attributes for one such division, and the ending position of the division inside the string. (More exactly: the position *after* the end of the div -- which is the starting position of the next div.) The end of the last division is also used to find out the string length. This is quite inefficient...
The "String" structure also contains "line_table[]", which holds the positions of all line breaks inside the string; more on this in 7. pre-render.c.
Finally, it contains an array of "Link" structures, which describe all the links (and form elements) inside this text block. See hacking-links.* for this.
Processing
Many properties of the various element types are data-driven. This presently includes the line break/blank line handling, and elements creating a box around all children. This properties are stored to the same "element_table[]" as the name strings used by parse_elements(). In future probably more properties will be data driven, making the code simpler, and also necessary for handling style sheets, which allow changing of almost all formatting properties.
The item tree is generated while traversing the element tree. Processing of each element is done in two steps: One step is done before entering an element (descending in the element tree), and the second step is done after leaving the element (ascending). Between those two steps, the same is done for all sub-elements. You guess it: This is a recursive algorithm. Only we haven't implemented it recursively, as mentioned in chapter 0. In this function the pseudo-recursive implementation is most evident; we even have to use a pseudo-stack.
The processing could also be split into a couple of much simpler passes using some temporary data structures, e.g. one generating the "normal" items, one generating blank items, one generating the strings, one "optimizing" the tree (lifting items where possible). This would be much easier to understand, and probably it would have been a good idea for the beginning. However, it would be much less efficient; that's why we won't step back to such an implementation after already having the present one.
Pre-processing
The first action in every iteration of the outer loop is to do the first processing step (s.a.) for the current element ("cur_el") -- in every iteration exactly one element is pre-processed.
++>NULL
+
+---+
| ! |-. <++
+---+ | +
v +
,-----------.
("header text")
`-+------+--' <-- cur_el <== depth
| html |-. <++++++++
+------+ | <+ +
v + +
+------+ +------+
| head |->| body |-.
+------+ +------+ |
First, we store any text from the current element node to the current open text item by add_string().
state[0]->first_item -->NULL (depth 0)
state[0]->last_item **>NULL
---
+------+ <-- first_item
| text |-->NULL <** last_item <== depth=1
+----*-+==>NULL
x *
v *
NULL *
v
,-----------.
("header text")
`-----------'
The next thing is processing of line breaks and paragraph breaks, depending what kind of element we have. More on this later, in Text Blocks and Blank Lines.
Recursing
Then we recurse into the element (descend in the element tree).
++>NULL
+
+---+
| ! |-. <++
+---+ | +
v +
,-----------.
("header text")
`-+------+--' <-- cur_el
| html |-. <++++++++
+------+ | <+ +
v + +
+------+ +------+
| head |->| body |-. <== depth
+------+ +------+ |
state[0]->first_item -->NULL (depth 0)
state[0]->last_item **>NULL
---
+------+ <-- state[1].first_item (depth 1)
| text |-->NULL <** state[1].last_item
+----*-+==>NULL
x *
v *
NULL *
v
,-----------.
("header text")
`-----------'
---
first_item -->NULL <== depth=2
last_item **>NULL
This is done by push_state(). This function doubles the top of stack, and returns a pointer to the newly created entry, which is used as the current state. ("first_item" and "last_item" aren't copied, but set to NULL; "id_attr" and "link_type" are set to -1.)
The stack stores all variables which are specific to every tree depth. Note that the stack uses the depths from the element tree, not from the item tree. The depths in the item tree are completely different, and change while processing -- which is one of the most tricky parts about parse_struct().
Currently the data stored is:
- Text attributes ("text_mode" and "high").
- The nesting depth "list_depth" of item lists. (Determines the indent of list items.)
- The two pointers "first_item" and "last_item", necessary to maintain the list of all items at a given depth.
- The type of the link or form control created by the element in the "link_type" enum (a value of -1 indicates there is no link at all)
- The URL of links or value of form elements in the "link_value" string
- The "form_enabled" flag indicates whether a form element is to be submitted to the server
- The name of a possible <select> element, bequeathed to its <option> elements
- The kind of the <select> element, also bequeathed
- "link_start" stores the position where a link or inline anchor beginns inside the current string
- "link_item" stores the text item in which a link/anchor beginns
- For elements with an anchor, "id_attr" stores which attribute contains the anchor id (or name)
After descending, first "link_start" is set to the current string end, so any text generated inside this element will become part of the link or anchor, if the element creates one. If there is no string item open, 0 is stored, so the link will begin at the start of the string if a new string beginns inside the element.
Aferwards, some element type specific handling is done. Mostly this is outputing of special element indicators. (This could be made data-driven, and probably it will do so soon.)
For some element types, also values of the current state are modified; this would be the argument passing in a real recursive implementation.
Ascending
The last step of every outer loop iteration is returning from recursion (ascending in the element tree). But in contrast to descending, ascending isn't done once per outer loop iteration. Instead, there is an inner loop, that ascends as often as needed to reach the level of the next element -- this can be zero, once or several times.
We know how long we need to ascend by starting at the level of the current element and looking for the parent of the new element; as soon as we find it, we know we needn't ascend any more. "depth" is adjusted every time we ascend, and all actions for leaving the element (unrecursing) are taken. ("depth" is always one below the depth of "new_el".)
If the next element is a child of the current one, no ascending is necessary. We have descended one step in the pre-processing step of the current iteration, and this is ok; we keep it. ("cur_el" is alredy the parent of "list_next".)
,-----------.
("header text")
`-+------+--'
| html |-. <++++++++
+------+ | <+ +
v + +
+------+ +------+
| head |->| body |-. <++++++++++++++
+------+ +------+ | <+ +
v + +
+----++ +---+
cur_el --> | h1 |-. <+ ,->| p |->
new_el xx> +----+ | + | +---+
v + |
,--------. |
(" heading") |
`--+---+-' | <== depth
list_next **> | ? |----'
+---+
If the next element is at the same level as the current one (single tags or other elements with no sub-elements), we need to ascend exactly one time, to get back to the level of the current element, after we have descended in the pre-processing step.
,-----------.
("header text")
`-+------+--'
| html |-. <++++++++
+------+ | <+ +
v + +
+------+ +------+ <**
--> | head |->| body |-. <++++++++++++++
xx> +------+ +------+ | <+ +
v + +
+----++ +---+
| h1 |-. <+ ,->| p |-> <==
+----+ | + | +---+
v + |
,--------. |
(" heading") |
`--+---+-' |
| ? |----'
+---+
,-----------.
("header text") <xx
`-+------+--'
| html |-. <++++++++
+------+ | <+ +
v + +
+------+ +------+ <**
--> | head |->| body |-. <++++++++++++++ <==
+------+ +------+ | <+ +
v + +
+----++ +---+
| h1 |-. <+ ,->| p |->
+----+ | + | +---+
v + |
,--------. |
(" heading") |
`--+---+-' |
| ? |----'
+---+
Of course it's a bit of overhead first to descend into an element, just to ascend from it right after. But it saves a lot of code for special handling of such childless elements. Many of the actions of both the first step and the second step have to be done for them also -- putting them together and leaving out the descending and ascending wouldn't save that much, while complicating the code quite a lot. We may consider some way in the future if profiling shows this would be rewarding.
If the next element is above the current one, we have to ascend more than once. (Once to get back to the level of the current one, the others to ascend to the new level.)
,-----------.
("header text")
`-+------+--'
| html |-. <++++++++
+------+ | <+ +
v + +
+------+ +------+
| head |->| body |-. <++++++++++++++
+------+ +------+ | <+ +
v + +
+----++ +---+
| h1 |-. <+ ,->| p |-> <**
+----+ | + | +---+
v + |
,--------. |
(" heading") |
`--+---+-' |
--> | ? |----'
xx> +---+
<==
,-----------.
("header text")
`-+------+--'
| html |-. <++++++++
+------+ | <+ +
v + +
+------+ +------+
| head |->| body |-. <++++++++++++++
+------+ +------+ | <+ +
v + +
+----++ +---+
xx> | h1 |-. <+ ,->| p |-> <**
+----+ | + | +---+
v + |
,--------. |
(" heading") |
`--+---+-' | <==
--> | ? |----'
+---+
,-----------.
("header text")
`-+------+--'
| html |-. <++++++++
+------+ | <+ +
v + +
+------+ +------+ <xx
| head |->| body |-. <++++++++++++++
+------+ +------+ | <+ +
v + +
+----++ +---+
| h1 |-. <+ ,->| p |-> <** <==
+----+ | + | +---+
v + |
,--------. |
(" heading") |
`--+---+-' |
--> | ? |----'
+---+
In every ascending iteration, first we do some element type specific handling again, mostly outputing element end indicators. Then we pop the previous state from the stack.
++>NULL
+
+ .--------------------------------.
+ v |
+---+ <xx |
| ! |-. <++ <** |
+---+ | + |
v + |
,-----------. |
("header text") | <==
`-+------+--' |
| html |-. <+++++++++++ |
+------+ | + |
| [...] [...]
[...] + |
| ,---------------. |
| (" (a single tag)") |
| `-----+---+-----' | <--
`--------->| ? |--------'
+---+
state[0]->first_item -->NULL (depth 0)
state[0]->last_item **>NULL
--- ,-- first_item
+------+ <--' +------+ +------+ <** last_item
| text |--[...]-->| text |-->| text |-->NULL <== depth=1
+----*-+==[...]==>+----*-+==>+----*-+==>NULL
x * x * x *
v * v * v * ,--------------.
NULL * NULL * NULL **>("(a single tag)")
v * `--------------'
,-----------. * ,----------------.
("header text") **>("this...blank row")
`-----------' `----------------'
++>NULL <xx
+
+ .--------------------------------.
+ v |
+---+ |
| ! |-. <++ <** | <==
+---+ | + |
v + |
,-----------. |
("header text") |
`-+------+--' |
| html |-. <+++++++++++ |
+------+ | + |
| [...] [...]
[...] + |
| ,---------------. |
| (" (a single tag)") |
| `-----+---+-----' | <--
`--------->| ? |--------'
+---+
first_item -->NULL <== depht=0
last_item **>NULL
--- ,-- state[1].first_item
+------+ <--' +------+ +------+ <** state[1].last_item (depth 1)
| text |--[...]-->| text |-->| text |-->NULL
+----*-+==[...]==>+----*-+==>+----*-+==>NULL
x * x * x *
v * v * v * ,--------------.
NULL * NULL * NULL **>("(a single tag)")
v * `--------------'
,-----------. * ,----------------.
("header text") **>("this...blank row")
`-----------' `----------------'
Inserting into Item Tree
Afterwards, the probably most interesting part follows: The sub-items created inside the element we are just leaving, are inserted into the item tree properly.
If the element we are leaving enforces a box (looked up in "element_table[]"), a new box item is created. (Box items are always created when leaving the element, and thus *after* all items inside the box.)
--> +-----+
**> ,->| box |-->NULL <== depht=0
| +-----+==>NULL
|
|
|
|
|
|
+------+ <-- +------+ +------+ | <** (depth 1)
| text |--[...]-->| text |-->| text |-'
+----*-+==[...]==>+----*-+==>+----*-+==>NULL
x * x * x *
v * v * v * ,--------------.
NULL * NULL * NULL **>("(a single tag)")
v * `--------------'
,-----------. * ,----------------.
("header text") **>("this...blank row")
`-----------' `----------------'
The "parent" pointers of all immediate children are set to the new box item, and "first_child" of the new item is set to the first of them.
--> +-----+
**> ,->| box |-->NULL <== depht=0
| +-----+==>NULL
| x ^
xxxxxxxx[...]xxxxxxxxxxxxxxxxxxxxxxxxxxx +
x | +
x +++++[...]++++++++++++++++++++++++++++++
x + + + |
v + + + |
+------+ <-- +------+ +------+ | <** (depth 1)
| text |--[...]-->| text |-->| text |-'
+----*-+==[...]==>+----*-+==>+----*-+==>NULL
x * x * x *
v * v * v * ,--------------.
NULL * NULL * NULL **>("(a single tag)")
v * `--------------'
,-----------. * ,----------------.
("header text") **>("this...blank row")
`-----------' `----------------'
If the element does not create a box, things are more tricky: We have to "lift" all sub-elements to the new level. This is done by concatenating the list of elements of the depth we are leaving to the list of elements of the depth we are entering.
state[0]->first_item -->NULL (depth 0)
state[0]->last_item **>NULL
---
+------+ <-- state[1].first_item (depth 1)
| text |-->NULL <** state[1].last_item
+----*-+==>NULL
x *
v *
NULL *
v
,-----------.
("header text")
`-----------'
---
state[2]->first_item -->NULL (depth 2)
state[2]->last_item **>NULL
--- ,-- state[3].first_item
+------+<-'+-------+ <** state[3].last_item (depth 3)
| text |-->| blank |-->NULL
+----*-+==>+-------+==>NULL
x * x
v * v
NULL * NULL
v
,-------.
("heading")
`-------'
---
+------+ <-- first_item
| text |-->NULL <** last_item <== depth=4
+----*-+==>NULL
x *
v *
NULL *
v
,------------.
("first...text")
`------------'
+------+ <+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
->| body |-. <++++++++++++++ +
+------+ | <+ + +
xx^ v + + +
+----++ +---+ <++++++++++++++++++++ +---+
==> | h1 |-. <+ ,->| p |---. <++ + **> ,->| p |->
depth=3 +----+ | + | +---+ | + + | +---+
v + | v + + |
,--------. | ,------------------. ,----. |
(" heading") | (" first...newlines,") (" and") |
`--+---+-' | `-----+----+-------' +-+----+-+ |
| ? |----' | em |-. <++ ,->| strong |-. <++ |
+---+ +----+ | + | +--------+ | + |
v + | v + |
,----------------. | ,------------. |
(" emphasized text") | (" strong text") |
`--------+---+---' | `-----+---+--' |
| ? |------' --> | ? |-----'
+---+ +---+
state[0]->first_item -->NULL (depth 0)
state[0]->last_item **>NULL
---
+------+ <-- state[1].first_item (depth 1)
| text |-->NULL <** state[1].last_item
+----*-+==>NULL
x *
v *
NULL *
v
,-----------.
("header text")
`-----------'
---
state[2]->first_item -->NULL (depth 2)
state[2]->last_item **>NULL
--- ,-- first_item
+------+<-'+-------+ +------+ <** last_item
| text |-->| blank |-->| text |-->NULL <== depth=3
+----*-+==>+-------+==>+----*-+==>NULL
x * x x *
v * v v *
NULL * NULL NULL *
v v
,-------. ,------------.
("heading") ("first...text")
`-------' `------------'
In the second case, we need an additional step if the element we are leaving was a <center> element (or some element with an align="center" attribute): In this case all sub-elements have to be centered one by one. (Elements creating a box are centered as a whole.)
Finally, we have to do some more line break/blank line handling.
Text Blocks
One single text block can be created by several elements. Every series of text parts not interrupted by elements requirering line breaks (or blank lines) around them is stored to a single text item; it can even contain newline characters created by <br> elements. This is done by calling add_string() every time text data is encountered.
The color for the text added by some element is determined from the current "state->text_mode" and "state->high". The normal color for the text mode is looked up in "color_map[]", and than its bit 3 is negated if the text is highlighted. This is more or less a hack; we'll have to replace this by some serious attribute handling at some point...
add_string()
This function concatenates the new text to the "string" of "string_item", which is a global variable pointing to the current open text item. If the new text has other attributes than the last division of the "string" so far, a new division is created for the new text; otherwise, the text is simply added to the last division.
text_item->string:
div[0].end div[1].end div[2].end
v v v
text: "[...]newlines, emphasized text and"
`~~~~~v~~~~~~'`~~~~~~~v~~~~~~'`~v'
div[1].color=MAGENTA div[2].color=WHITE
div[0].color=WHITE
+------------------------------------------------+ |+-----------+ | ||header text| | |+-----------+ | | | |+-------+ | ||heading| | |+-------+ | | | |+----------------------------------+ | ||[...]newlines, emphasized text and| |
,------------------. ,----.
(" first...newlines,") (" and")
`-----+----+-------' +-+----+-+
| em |-. <++ ,->| strong |-. <++
+----+ | + | +--------+ | +
v + | v +
,----------------. | ,------------.
(" emphasized text") | (" strong text")
`--------+---+---' | `-----+---+--'
| ? |------' --> | ? |-->
+---+ +---+
text_item->string:
div[0].end div[1].end div[2].end div[3].end
v v v v
text: "[...]newlines, emphasized text and strong text"
`~~~~~v~~~~~~'`~~~~~~~v~~~~~~'`~v'`~~~~~v~~~~'
div[1].color=MAGENTA div[3].color=STRONG WHITE
div[0].color=WHITE div[2].color=WHITE
+------------------------------------------------+ |+-----------+ | ||header text| | |+-----------+ | | | |+-------+ | ||heading| | |+-------+ | | | |+----------------------------------------------+| ||[...]newlines, emphasized text and strong text||
A new division can also be enforced, by an additional call of add_text() with NULL as text before adding the next text part. (This is necessary to prevent multiple consecutive links from being merged into one div, as link highlighting is done div-wise.)
Appending to an existing text item is only possible if the last created item was a text item, and there was no breaking element. (Neither the last one, nor the new one, nor anyone in between.) Otherwise, a new text item has to be created, and a new "String" structure for it. This is the only place where new text items are created.
| | |+---------------------------------+| ||starting with an evil center tag,|| |+---------------------------------+|
+---+
->| p |-. <++++++++++++++++++
+---+ | <+ +
| + +
| + ,----------.
v + (" this...em")
+--------+ `---+---+--' <--
| center |-. <++ ,->| a |->
+--------+ | + | +---+
v + |
,----------------. |
(" starting...tag,") |
`--------+---+---' |
| ? |------'
+---+
| | |+---------------------------------+| ||starting with an evil center tag,|| |+---------------------------------+| |+---------------------+ | ||this very...anchor em| |
If there is a space at the beginning of a string, it is discarded while creating the string, for a text block always starts with a word.
Line Breaks
As soon as an element forcing a line break is either entered or left (there are no elements creating a break only before or only behind it), "string_item" is set to NULL, indicating that no more text can be added to the last text item, and new text has to create a new one.
+---+
->| p |-. <++++++++++++++++++
+---+ | <+ +
| + +
| + ,----------.
v + (" this...em")
+--------+ <xx `---+---+--' <**
| center |-. <++ ,->| a |->
+--------+ | + | +---+
v + |
,----------------. |
(" starting...tag,") |
`--------+---+---' | <-- <==
| ? |------'
+---+
| | |+---------------------------------+| ||starting with an evil center tag,||
+---+ <xx
->| p |-. <++++++++++++++++++
+---+ | <+ +
| + +
| + ,----------.
v + (" this...em")
+--------+ `---+---+--' <** <==
| center |-. <++ ,->| a |->
+--------+ | + | +---+
v + |
,----------------. |
(" starting...tag,") |
`--------+---+---' | <--
| ? |------'
+---+
| | |+---------------------------------+| ||starting with an evil center tag,|| |+---------------------------------+|
Blank Lines
Probably the most tricky part is handling of blank lines. Similar to line breaks, items needing blank lines have them before *and* after them. However, when two such items meet, they have only *one* blank line between them. Furthermore, a blank line is *never* inserted before the first or after the last item inside a box. That's why blank lines cannot be simply inserted when entering an element causing blank lines, or when leaving it. Instead, only "requests" for blank lines are stored in "para_blank", and if certain conditions are met, a blank line is inserted before the next item (inside add_item()).
Actually, the blank line is inserted not in front of the new item, but after the last item, at the same tree depth as that one. This is important, to ensure that a blank line generated in front of a box is actually inserted *outside* of the box, not inside it, as would be the case if it was inserted at the current depth. The global "blank_depth" variable is responsible for this, and is set to the current depth every time a blank line request is generated. The blank item is then inserted directly to the "state" structure at "blank_depth" in add_item(). This is surely very bad programming style ;-)
The blank line requests are managed by the global "para_blank" variable. A value of 1 indicates that a blank line is needed, and will be generated by the next add_item(). A value of 0 indicates that no blank line is needed in any case; this situation can only occur when there are no items inside the current box yet. A value of -1 indicates that a blank line *may* be necessary. That is the case when we already have some items inside the current box, but the last item does not need a blank line; a blank line needs to be inserted only when the following item wants one.
When entering an element needing a blank line, a request is generated ("para_blank" set to 1 and "blank_depth" is stored) if "para_blank" was -1, indicating that there are already items in the current box.
+
+------+
->| body |-. <++++++++++++++
+------+ | <+ +
v + +
+----++ +---+
--> | h1 |-. <+ ,->| p |->
cur_ +----+ | + | +---+
tag v + |
,--------. |
(" heading") |
`--+---+-' |
| ? |----'
+---+
+------------------+ |+-----------+ | ||header text| | |+-----------+ | ??? (para_blank=-1)
+
+------+
->| body |-. <++++++++++++++
+------+ | <+ +
v + +
+----++ +---+
--> | h1 |-. <+ ,->| p |->
==> +----+ | + | +---+
blank_ v + |
depth ,--------. |
(" heading") |
`--+---+-' |
**> | ? |----'
list_next +---+
(new "cur_el")
+------------------+
|+-----------+ |
||header text| |
|+-----------+ |
| (para_blank=1) |
...
|+-------+ | (will be created later)
||heading| |
|+-------+ |
If "para_blank" is already 1, the request is left unchanged; a blank line will be inserted already.
| | |+-------+ | ||heading| | |+-------+ | | (para_blank=1) |
|
v
+----+ <== +---+ <--
| h1 |-. <+ ,->| p |---. <++
+----+ | + | +---+ | +
v + | v +
,--------. | ,------------------.
(" heading") | (" first...newlines,")
`--+---+-' | `-----+----+-------'
| ? |----' **> | em |-.
+---+ +----+ |
| |
|+-------+ |
||heading| |
|+-------+ |
| (para_blank=1) |
...
|+------------+ |
||first...text| |
|+------------+ |
If "para_blank" is 0 nothing is done, even if the element would normally need a blank line: No blank line is ever inserted in front of the first item inside a box.
+-----------+ (para_blank=0)
+---+
| ! |-. <+
+---+ | +
v +
+---++
--> | p |-. <+
+---+ | +
v +
,---------.
("some text")
`--+---+--'
**> | ? |->
+---+
+-----------+
...
|+---------+|
||some text||
|+---------+|
When some item (text or box) is inserted by add_item(), "para_blank" is always set to -1, as now there is at least one item in the current box.
+------------------+ |+-----------+ | ||header text| | |+-----------+ | | (para_blank=1) |
+
+------+
->| body |-. <++++++++++++++
+------+ | <+ +
v + +
+----++ +---+
| h1 |-. <+ ,->| p |->
+----+ | + | +---+
v + |
,--------. |
(" heading") |
`--+---+-' |
--> | ? |----'
+---+
+------------------+ |+-----------+ | ||header text| | |+-----------+ | | | |+-------+ | ||heading| | |+-------+ | ??? (para_blank=-1)
When leaving an element needing a blank line, a request is generated also, and will be handled in the next add_item(). Otherwise, the current state is kept.
+
+------+
->| body |-. <++++++++++++++
+------+ | <+ +
v + +
+----++ +---+
xx> | h1 |-. <+ ,->| p |-> <**
new_el +----+ | + | +---+
v + |
,--------. |
(" heading") |
`--+---+-' |
--> | ? |----'
+---+
+------------------+ |+-----------+ | ||header text| | |+-----------+ | | | |+-------+ | ||heading| | |+-------+ | ??? (para_blank=-1)
+
+------+ <xx
->| body |-. <++++++++++++++
+------+ | <+ +
v + +
+----++ +---+
==> | h1 |-. <+ ,->| p |-> <**
+----+ | + | +---+
v + |
,--------. |
(" heading") |
`--+---+-' |
--> | ? |----'
+---+
+------------------+ |+-----------+ | ||header text| | |+-----------+ | | | |+-------+ | ||heading| | |+-------+ | | (para_blank=1) |
When entering an element generating a box, "para_blank" is reset to 0.
| | |+---------+ | ||some text| | |+---------+ | ??? (para_blank=-1)
+--------+ +------+
->| center |-. ,->| form |-. <--
+--------+ | | +------+ |
v | |
,---------. | |
("some text") | v
`--+---+--' | +---+
| ? |-----' **> | p |->
+---+ +---+
| | |+---------+ | ||some text| | |+---------+ | |+--------------+| (para_blank=0)
However, it's left unchanged if there is already a request, indicating that a blank line will be inserted *in front* of the box by the next add_item().
| | |+---------+ | ||some text| | |+---------+ | | (para_blank=1) |
+---+ +------+
->| p |-. <== ,->| form |-. <--
+---+ | | +------+ |
v | |
,---------. | |
("some text") | v
`--+---+--' | +---+
| ? |-----' **> | p |->
+---+ +---+
| | |+---------+ | ||some text| | |+---------+ | | (para_blank=1) | |+--------------+|
When leaving a box creating element, any request generated inside the box is discarded. We never insert a blank line after the last item of a box.
++>NULL
+ .--------------------------------.
+ v |
+---+ <xx |
| ! |-. <++ <** |
+---+ | + |
v + |
,-----------. |
("header text") |
`-+------+--' |
| html |-. <+++++++++++ |
+------+ | + |
| [...] [...]
[...] + |
| ,---------------. |
| (" (a single tag)") |
| `-----+---+-----' | <--
`--------->| ? |--------'
+---+
|+--------------+ | ||(a single tag)| | |+--------------+ | | (para_blank=1) |
++>NULL <xx
+ .--------------------------------.
+ v |
+---+ |
| ! |-. <++ <** |
+---+ | + |
v + |
,-----------. |
("header text") |
`-+------+--' |
| html |-. <+++++++++++ |
+------+ | + |
| [...] [...]
[...] + |
| ,---------------. |
| (" (a single tag)") |
| `-----+---+-----' | <--
`--------->| ? |--------'
+---+
|+--------------+ | ||(a single tag)| | |+--------------+ | +------------------+ ??? (para_blank=-1)
However, if the last request was generated *before* descending into the element creating the box (only possible if the box is empty), it is kept, and the blank is inserted when the box is added.
| | |+---------+ | ||some text| | |+---------+ | | (para_blank=1) | |+--------------+|
+---+ +------+ <xx +---+
->| p |-. <== ,->| form |-. ,->| p |-> <**
+---+ | | +------+ | | +---+
v | | |
,---------. | | |
("some text") | v |
`--+---+--' | +------+ |
| ? |-----' --> | span |-'
+---+ +------+
| | |+---------+ | ||some text| | |+---------+ | | | |+--------------+| |+--------------+|
<br> elements
<br> elements used to break text blocks as described above. This however turned out to be a bug: some elements (e.g. <a>) can span over a <br>, and these were handled incorrectly with that approach.
Now simply a '\n' is added to the text block when <br> is encountered, and then correctly handled while Breaking String into Lines.
Links
When a link element is encountered (any <a> element having a "href"), "cur_state->link_type" is set (to FORM_NO), indicatating that the element creates a link. The URL (from the "href" attribute) is saved in "cur_state->link_value".
The link is then stored when leaving the link element. (After processing all sub-elements.) The link start is set to "cur_state->link_start", which was the string end position while entering the link element; the link end is set to the current string end position. This way the link spans all text generated inside the element.
There is some additional handlig necessary to workaround broken links, however. (This used to be important to make forms work with SGML at all; now that full SGML support is implemented, it probably only helps a few broken pages...)
When a link doesn't end in the same string as it started (checked by "cur_state->link_item", which was the current string item when entering the link element), it is stored in the starting string, not the current one. However, if there was no active text item when the link started (indicating the link starts at the beginning of a text block), we can't determine the starting string. In this case, we store it in the current string -- normally, this is the right thing to do; if the link spans multiple strings, on the other hand, at least the last part ist stored this way.
Some magic is necessary to handle nested links: The link which is stored later is the outer one, i.e. it starts *before* the previously stored inner link... To handle this in a useful fashion, we also have to *store* it before the inner link. Thus, instead of simply appending it at the end of the string's link list, we have to shift all inner links one position behind, and put the new (outer) link at the free position created in front of them.
Forms
Form elements are handled very similar, and by the same code. The only difference is that "link_type" is set to some form type instead of FORM_NO, and the initial "value" of the element is stored in the structure tree so it will be submitted to the server, and (for some form control types) displayed on the page via set_form() (see hacking-links.*).
Also, "form_enabled" is set for all elements that will be submitted to the server. It is always set for elements that are submitted unconditionally (text/password/hidden input fields), and set to the initial state for elements that are submitted depending on their state (radio buttons, checkboxes, <select> options). Submit buttons initially aren't enabled.
Some special handling is necessary for <select> options: They do not have an own "name" attribute to store in "Link->name" like other form elements; the "name" for all options is given in the <select> element instead. Thus it needs to be passed to all the options by the state stack, in "select_name". The same for the link type (FORM_OPTION or FORM_MULTIOPTION) of the option links: It depends on the presence of the "multiple" attribute in the <select> element, and is passed in "select_type" for that reason.
Virtual Items
There are two kinds of "virtual" Items: ITEM_BLOCK_ANCHOR and ITEM_INLINE_ANCHOR. As the names suggest, both types are presently used for anchors. The difference is that block anchors are created by block elements with an "id" attribute, and span one or more block elements, while inline anchors are created inside text blocks by the classical "a" element or by any inline element with an "id", and span only a text part.
"virtual" means that these items do not affect layouting like other items; they only create some additional structure.
Block Anchors
As block anchors span multiple other elements, they have to act as boxes, containing a series of virtual children. In contrast to real box items, they do not get their size assigned by the parent, and do not assign size to their virtual children. The children get their size assigned by the real parent directly, and the size of the virtual box is determined afterwards, from the outer bounds of all virtual children.
In the present implementation, virtual box items aren't inserted normally as parents of their virtual children into the item tree. They are simply inserted as normal items after their virtual children, at the same tree depth. The virtual box is created only by special pointers, handled only in the necessary places; the normal tree traversal functions do not know about them. This is a hack; the idea was to integrate them into the existing system changing as little as possible. Most probably this will be replaced by a clean implementation using real parent/children relations in the future.
virtual tree:
+-----+
,->| box |
| +-----+
| x ^
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx +
x +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
v + + + |
+------+ +--------------+ +------+ |
| text |-. ,->| block anchor |-->| text |-'
+------+=|======================|=>+--------------+==>+------+==>NULL
| | x ^
| xxxxxxxxxxxxxxxxxxxxxxxx +
| x +++++++++++++++++++++++++++
| v + + |
| +------+ +------+ |
`->| text |-->| text |-'
+------+ +------+==>NULL
real tree:
°°°> virtual child +-----+
,->| box |
virtual box | +-----+
,----------^----------. | x ^
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx +
x +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
v + + + + + |
+------+ +------+ +------+ +--------------+ +------+ |
| text |-->| text |-->| text |-->| block anchor |-->| text |-'
+------+==>+------+==>+------+==>+--------------+==>+------+==>NULL
^ °
° °
°°°°°°°°°°°°°°°°°°°°°°°°°°°°
Inline Anchors
As inline anchors are contained within text elements, they have to act as children of text items. Text item however can't have children in the present implementation; thus, inline anchors presently are also only virtual children.
Similar to the virtual boxes created by block anchors, these virtual children are stored after their virtual parent, at the same tree depth. This is a hack just as the virtual boxes, and will be replaced also.
Additionally to the virtual parent (and the name of the anchor), inline anchors need to store the position of their start and end inside the parent string.
virtual tree:
%%%> anchor_start,anchor_end +-----+
.->| box |
| +-----+
| x ^
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx +
x +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
v + + + |
+------+ +------+ +------+ |
| text |-. .->| text |-->| text |-'
+------+=|========================================|=>+------+==>+------+==>NULL
| | x ^ *
| | x + * ,--------------------.
| xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx + **>("anchor1 and anchor2.")
| x ++++++++++++++++++++++++++++++++++++++ `--------------------'
| x + + | ^ ^ ^ ^
| x + + | % % % %
| x + %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % %
| x + % + %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
| v + % + % |
| +---------------+ +---------------+ |
`->| inline anchor |-->| inline anchor |-'
+---------------+==>+---------------+==>NULL
real tree:
°°°> virtual parent +-----+
.->| box |
| +-----+
| x ^
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx +
x +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
v + + + + + |
+------+ +------+ +---------------+ +---------------+ +------+ |
| text |-->| text |-->| inline anchor |-->| inline anchor |-->| text |-'
+------+==>+------+==>+---------------+==>+---------------+==>+------+==>NULL
* ^ °% °%
%%%%%%%%%%%%%*%%°%%%%%%%%%%%%%°% °%
% * ° ° °%
% * °°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°%
% v %
% ,--------------------. %
% ("anchor1 and anchor2.") %
% `--------------------' %
% ^ ^ ^ ^ %
% % % % % %
%%%%%%%%%%%% %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
Links are very similar to inline anchors, and probably will be stored the same way as anchors in the future, once ITEM_INLINE is cleanly implemented. The present different handling of links and anchors is by purpose -- to see which method turns out to be better. (Presently, it looks much like the anchor-method is better.)
Finishing
After all elements of the syntax tree are traversed, the structure tree is finalized by setting "next" of the tree top (the global item) to point back to itself, and "parent" point to the first item.
parse_struct() returns a "struct Item" pointer to the tree top. This is passed as argument to the following passes by main().
Is is a bit strange that "parent" of the tree top points to the first item. However, it is important to have some simple way of accessing the first item, as some of the following passes start with the fist item; but only the pointer to the tree top is passed on the following passes. Maybe we should use some function that finds the first item by following "first_child" instead -- as this is done only once in every processing pass, it wouldn't be too big an unefficiency.
free_items()
This function is responsible for freeing the memory used by the item tree.
Just as free_syntax(), the tree is traversed by "list_next", and every node is freed, including all data belonging to it.
For text items, the associated strings are freed also; for anchor items, the anchor data is freed.
7. pre-render.c
So far, the item tree only represents the structure of the page, i.e. the dependencies of the items. The Items have no actual sizes nor positions yet.
# ################################################################################################ # +----------- # |header text # *** # * # +------- # |heading # *** # * # +----------------------------------------------------------------------------------------------- # |first paragraph of text; includes multiple spaces and newlines, emphasized text and strong text # *** # * # +--------------------------------- # |starting with an evil center tag, # +----------------------------------------------------------------------------------------------- # |this very long second paragraph contains some special characters (including a simple space...): # | &; <>"=/ plus a big gap###and two unicode escapes (decimal: ¡ and hexal: ¿) but also an anchor # | embedded inside a word (this anchor also is the only tag with parameters); and finally a blank # | row # +-------------- # |(a single tag)
pre_render() assigns coordinates inside the layouted page to all items. It also breaks the text blocks into lines.
This process is reversible (it doesn't alter any data stored before; it only creates additional data), and thus resizing the output area of the viewer will be possible without regenerating the item tree. (And without reloading the file.)
The coordinates are stored to "x_start", "x_end", "y_start" and "y_end" of the item structures. For every text item, the positions of all line breaks are stored in "line_table[]". This is the only way line breaks are indicated; the text string itself isn't altered. Thus the rendering function has to look at this table while generating output. (8. render.c)
Besides of assigning the coordinates, pre_render() also generates the "page_map[]". This is a structure telling which items occupy every line of the output page. For every line, a table is generated, containing references to all elements that show up in this line. It's described more thoroughly in create_map().
The processing is split into five smaller passes, executed successively by pre_render(). It also could be done in one single, pseudo-recursive pass. Most probably this would be more efficient; it would be harder to understand, too. Maybe we will change this at some point...
calc_width()
In the first pass, the minimal x-width of all items is calculated. This is done by traversing the item tree by "list_next" (bottom to top). This ensures that the sizes of all children are calculated before the parent. In every iteration one item is processed. The actions taken depends on the type of the item.
For text items, the minimal width (stored in "x_end") is set to the width of the longest word in the text block, as in other browsers. Note that we *can* generate narrower text blocks; netrik can break words that do not fit on a line, and probably also scrolling of items will be possible in the future. (AFAIK no other browser does any of that, although it was recommended by the W3C for years...)
# ###############[...] # +------+ # |header| ^ ^x_end=6 0 # *** # * # +------- # |heading [...]
For blank lines, the minimal width is set to 0 -- blank lines do not need any width; anchor items as well.
# ###############[...] # +------+ # |header| # ** # ** ^x_end=0 # +------- # |heading [...]
Finding the longest word is quite easy. The whole text block is processed char by char in a loop. For each char, the current word length "len" is incremented. When the word ends (space or string end encountered), "len" is reset to 0. "longest" keeps the current maximum, and is stored to "x_end" after the whole text block was processed.
For box items, the minimal width is the one of the widest sub-item.
# ############## # +------+ # # |header| # # ** # # ** # # +-------+ # # |heading| # # ** # # ** # # +----------+ # # |emphasized| # # ** # # ** # # +--------+ # # |starting| # # +------------# # |parameters);# # +------+ # # |single| # ^ ^x_end 0
As all sub-items are processed before the parent, we already know the widths of the sub-items when processing a box item. We simply go trough all immediate children (start with "first_child" and go on by "next") and look for the maximum.
assign_width()
The second pass assigns the x-coordinates (x_start and x_end) to all items. (Presently, this is trivial: all items have the same coordinates as their parent, which is the global item...) For text items, the positions of line breaks are also calculated.
Traversing Item Tree Top to Bottom
In this pass, the tree is traversed top to bottom, as the coordinates of the sub-items depend on the coordinates of the parent. This is a bit more complicated than traversing bottom to top, as there is no equivalent of the "list_next" pointer for this.
If the current item has children, we proceed with the first child. (Descend.)
<+++
+
+-----+
,->| box |-->NULL
| +-----+<==#
| x ^ #===#
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx +
x +++++++++++++++++++++++++++++++++++
v + + |
cur_item xx> +-----+ +-----+ |
,->| box |-. ,->| box |-'
new cur_item | +-----+=|===========|=>+-----+==>NULL
| | x ^ | | x ^
xxxxxxxx|xxxxxxxxxxxxxxx + | xxxxxxxxxxxxx +
x +++++|+++++++++++++++++ | x ++++++++++++
v + | + | | v + |
+------+<-'+------+ | | +------+ |
| text |-->| text |-' `->| text |-'
+------+==>+------+==>NULL +------+==>NULL
x x x
v v v
NULL NULL NULL
If the current item has no children, we go to the "next" item.
<+++
+
+-----+
,->| box |-->NULL
| +-----+<==#
| x ^ #===#
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx +
x +++++++++++++++++++++++++++++++++++
v + + |
+-----+ +-----+ |
,->| box |-. ,->| box |-'
| +-----+=|===========|=>+-----+==>NULL
| x ^ | | x ^
xxxxxxxxxxxxxxxxxxxxxxxx + | xxxxxxxxxxxxx +
x +++++++++++++++++++++++ | x ++++++++++++
v + + | | v + |
+------+ +------+ | <-- | +------+ |
| text |-->| text |-' `->| text |-'
+------+==>+------+==>NULL +------+==>NULL
x ^xx x x
v v v
NULL NULL NULL
If there is no "next" item (we are already at the last item of this depth), we have to ascend before we can go to the next item.
<+++
+
+-----+
,->| box |-->NULL
| +-----+<==#
| x ^ #===#
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx +
x +++++++++++++++++++++++++++++++++++
v + + |
+-----+ +-----+ |
--> ,->| box |-. ,->| box |-'
| +-----+=|===========|=>+-----+==>NULL
| x ^ | | x ^
xxxxxxxxxxxxxxxxxxxxxxxx + | xxxxxxxxxxxxx +
x +++++++++++++++++++++++ | x ++++++++++++
v + + | | v + |
+------+ +------+ | <xx | +------+ |
| text |-->| text |-' `->| text |-'
+------+==>+------+==>NULL +------+==>NULL
x x x
v v v
NULL NULL NULL
<+++
+
+-----+
,->| box |-->NULL
| +-----+<==#
| x ^ #===#
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx +
x +++++++++++++++++++++++++++++++++++
v + + |
+-----+ --> +-----+ |
,->| box |-. ,->| box |-'
| +-----+=|===========|=>+-----+==>NULL
| x ^ | | x ^
xxxxxxxxxxxxxxxxxxxxxxxx + | xxxxxxxxxxxxx +
x +++++++++++++++++++++++ | x ++++++++++++
v + + | | v + |
+------+ +------+ | <xx | +------+ |
| text |-->| text |-' `->| text |-'
+------+==>+------+==>NULL +------+==>NULL
x x x
v v v
NULL NULL NULL
If we still do not have a "next" item after ascending, we ascend again.
<+++
+
+-----+
,->| box |-->NULL
| +-----+<==#
| x ^ #===#
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx +
x +++++++++++++++++++++++++++++++++++
v + + |
+-----+ --> +-----+ |
,->| box |-. ,->| box |-'
| +-----+=|===========|=>+-----+==>NULL
| x ^ | | x ^
xxxxxxxxxxxxxxxxxxxxxxxx + | xxxxxxxxxxxxx +
x +++++++++++++++++++++++ | x ++++++++++++
v + + | | v + |
+------+ +------+ | | +------+ | <xx
| text |-->| text |-' `->| text |-'
+------+==>+------+==>NULL +------+==>NULL
x x x
v v v
NULL NULL NULL
<+++
+
+-----+
--> ,->| box |-->NULL
| +-----+<==#
| x ^ #===#
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx +
x +++++++++++++++++++++++++++++++++++
v + + |
+-----+ +-----+ |
,->| box |-. ,->| box |-'
| +-----+=|===========|=>+-----+==>NULL
| x ^ | | x ^
xxxxxxxxxxxxxxxxxxxxxxxx + | xxxxxxxxxxxxx +
x +++++++++++++++++++++++ | x ++++++++++++
v + + | | v + |
+------+ +------+ | | +------+ | <xx
| text |-->| text |-' `->| text |-'
+------+==>+------+==>NULL +------+==>NULL
x x x
v v v
NULL NULL NULL
After all items have been processed, we'll ascend until we get back to the top item, which's "next" pointer points back to itself; Thus we stay at the tree top after following "next", and the main loop is terminated.
<+++
+
+-----+
--> ,->| box |-->NULL
| +-----+<==#
| x ^ #===#
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx +
x +++++++++++++++++++++++++++++++++++
v + + |
+-----+ +-----+ |
,->| box |-. ,->| box |-'
| +-----+=|===========|=>+-----+==>NULL
| x ^ | | x ^
xxxxxxxxxxxxxxxxxxxxxxxx + | xxxxxxxxxxxxx +
x +++++++++++++++++++++++ | x ++++++++++++
v + + | | v + |
+------+ +------+ | | +------+ | <xx
| text |-->| text |-' `->| text |-'
+------+==>+------+==>NULL +------+==>NULL
x x x
v v v
NULL NULL NULL
Assigning x-Coordinates
When processing an item, not the coordinates of "cur_item" are assigned, but the coordinates of its *sub-items*. "cur_item" has had its coodinates assigned while its parent was processed. This is an exact reversal of calc_width(): Instead of "cur_item" getting its size from its sub-items, the sub-items get their size from "cur_item".
The coordinates for the global item are assigned before calling assign_width(). x-start is set to 0, x-end is set to the output width passed from main(). (The screen width when cfg.term_width is set; otherwise, either a constant value (in --dump mode) or the maximum of screen width and page width.) Play around with them if you like. (If set to a narrow box, you can see the word breaking work.)
################################################################################## #------+ # #header| # #* # #* # [...] #------------+ # #parameters);| # #------+ # #single| # ^x_start=0 ^x_end=80
For box (and form) items, the coordinates of all immediate children are simply set to the ones of the box. (This will have to change at some point, but it's ok for now...) It's done in a simple loop processing the children by "first_child" and "next", just as in calc_width().
################################################################################## +--------------------------------------------------------------------------------+ |header text | ********************************************************************************** * * [...] +--------------------------------------------------------------------------------+ |this very long second paragraph contains some special characters (including a | |simple space...): &; <>"=/ plus a big gap###and two unicode escapes (decimal: ¡ | |and hexal: ¿) but also an anchor embedded inside a word (this anchor also is the| |only tag with parameters); and finally a blank row | +--------------------------------------------------------------------------------+ |(a single tag) |
Blank items have no children. Nothing has to be done. The same for anchor items.
Text items have no children, too; but they need to be broken into lines. (This could be done in a seperate pass...)
Breaking String into Lines
The breaking into lines is a bit complicated, because netrik has the ability to break words that do not fit on a single line. And: It does this smartly, avoiding line breaks whenever possible.
The line breaking does not alter the string structure itself. The only output is "line_table[]" (containing the positions of all line breaks inside the string), and its length (The number of line breaks), which is stored as the height of the text item in "y_end".
Before starting, some constants are calculated: The width of the text block is the difference of x_end and x_start. The pointer to the start of the text string is taken from the string structure and stored in "string_start".
The text block is processed word-wise. In each iteration of the outer loop one word is processed.
First, "word_end" is retrieved. It points to the space, newline or string terminator that ends the word to be processed. "next_word_start" is set to "word_end+1", assuming that we will proceed with the next word (which starts after the space); however, this may by altered later.
line_table[0]
v
string: "I am a very simple and stupid example sentence."
^ ^ ^ ^^
s l w en
string_start line_start next_word_start
word_end
word_start
output:
x_start x_end
|<- width ->|
v 12 v
+------------+
|I am a very |
|simple and stupid example sentence.
^ ^ ^^
l w en
Then we test if the whole word fits into the current open line (which's starting position is stored in "line_start").
v
string: "I am a very simple and stupid example sentence."
^ ^ ^ ^^^
s l w enb
line_start+width (line end/Break position)
|<- width ->|
(12)
+------------+
|I am a very |
|simple and stupid example sentence.
^ ^ ^^^
l w enb
e<b -> no break
If it does, we do nothing, and simply proceed with the next word. "word_start" is set to "next_word_start" before the next iteration.
v
string: "I am a very simple and stupid example sentence."
^ ^ ^
s l n
^
w
Otherwise, we have to generate a line break, and put the word on a new line.
v
string: "I am a very simple and stupid example sentence."
^ ^ ^^ ^^
s l wb en
+------------+
|I am a very |
|simple and stupid example sentence.
^ ^^ ^^
l wb en
e>=b -> break
Normally, we simply break the line at the current word, by setting "line_start" to "word_start", and adding a line break at this position to "line_table[]".
line_table[1]
v v
string: "I am a very simple and stupid example sentence."
^ ^ ^^
s w en
^
l
+------------+
|I am a very |
|simple and |
|stupid example sentence.
^ ^^
w en
^
l
However, things are more complicated due to the ability of breaking too long words. For this, we have to test whether the current word fits on one line. If it does, we simlply insert a line break as described above.
v v
string: "I am a very simple and stupid example sentence."
^ ^ ^^ ^^ ^
s l wb en f
word_start+width (Fit on line)
|<- width ->|
+------------+
|I am a very |
|simple and stupid example sentence.
^ ^^ ^^ ^
l wb en f
e<f -> simple break
If it does not (if it's longer than the width of the text block), we have to break the word.
v v v v
string: "I am a second example sentence with aVeryLongWord in me."
^ ^ ^ ^^^
l w b fen
+------------+
|I am a |
|second |
|example |
|sentence |
|with aVeryLongWord in me.
^ ^ ^ ^^^
l w b fen
e>=f -> break word
In this case, there are two possibilities: Either, we put the beginning of the word on the current line, and the remainder on the next line(s). Or, we put the first part to the new line already. This sometimes saves an unnecessary word break:
+------------+ |I am a third| |example | |sentence, | |with | |anEvenABitL\| |ongerWord in| |me. | +------------+
instead of:
+------------+ |I am a third| |example | |sentence, | |with anEven\| |ABitLongerW\| |ord in me. | +------------+
We decide which way to go, by truncating the word of as many whole line lengths ("width-1", because of the break chars) as possible, and testing if the remaining part fits on the current line.
v v v v
string: "I am a second example sentence with aVeryLongWord in me."
^ ^ ^ ^ ^^
l w t b en
trunc_word_end=word_end-(width-1)
|<- ->|
width-1 (11)
+------------+
|I am a |
|second |
|example |
|sentence |
|with a\ |
^ ^ ^ ^
l w t b
t<=b -> fill line
|VeryLongWord| (truncated part)
^
e
Or:
v v v v v
string: "I am a third example sentence, with anEvenABitLongerWord in me."
^ ^ ^ ^ ^^
l w b t en
+------------+
|I am a third|
|example |
|sentence, |
|with |
|anEvenABitL\|
^ ^ ^ ^
l w t b
t<=b -> fill line
|itLongerWord|
^
e
For a longer word, the truncation is done more than once:
v v v
string: "I am another example sentence, with aTerriblyLongWordThatDoesn'tSeemToEnd in me."
^ ^ ^ ^ ^ ^ ^^
l w T b t t en
(final)
+------------+
|I am another|
|example |
|sentence, |
|with aTe\ |
^ ^ ^ ^
l w T b t<=b -> fill line
|rriblyLongW\|
^
t
|ordThatDoes\|
^
t
|n'tSeemToEnd in me.
^^
en
or:
v
string: "This example has again aTerriblyLongWordThatDoesn'tSeemToEnd in some other sentence."
^ ^ ^ ^ ^ ^ ^^
l w b T t t en
+------------+
|This example|
|has again aTe\
^ ^ ^ ^
l w b T
t>b -> don't fill (put first word part in new line)
This is done in a loop. We also could do it by a single expression, but evaluating a quite complicated expression is probably less efficient than a simple loop, which is almost always executed only once. (Very few words are longer than two lines...) Also, it is easier to understand, isn't it?...
If we decide on the first way, we fill up the current line with the beginning of the word. (The last char is kept empty for the word break indicator displayed in the output.) The line break position and "line_start" are set accordingly, so that the rest of the word will be put on the next line.
v v v v v
string: "I am a second example sentence with aVeryLongWord in me."
^ ^° ^^
w lb en
(old value)
+------------+
|I am a |
|second |
|example |
|sentence |
|with aVeryL\|
|ongWord in me.
^ ^^
l en
We also need to test if the remainder fits on the next line; if it does not, we have to adjust "next_word_start" so that processing in the next loop iteration will not continue with the next word, but with the part of the current word that does not fit on the next line.
v v v v
string: "I am another example sentence, with aTerriblyLongWordThatDoesn'tSeemToEnd in me."
^ ^° ^ ^^
w lb b en
+------------+
|I am another|
|example |
|sentence, |
|with aTerri\|
|blyLongWordThatDoesn'tSeemToEnd in me.
^ ^ ^^
l b en
e>=b -> more breaks necessary
v v v v
string: "I am another example sentence, with aTerriblyLongWordThatDoesn'tSeemToEnd in me."
^ ^ ^^ ^
w l nb e
+------------+
|I am another|
|example |
|sentence, |
|with aTerri\|
|blyLongWordThatDoesn'tSeemToEnd in me.
^ ^^ ^
l nb e
If the second way was chosen, things are quite easy: The line break is inserted at the beginning of the word, and "next_word_start" is *always* adjusted to the end of the new line.
v v v v v
string: "I am a third example sentence, with anEvenABitLongerWord in me."
^ ^^ ^°
w nb en
^
l
+------------+
|I am a third|
|example |
|sentence, |
|with |
|anEvenABitLongerWord in me."
^ ^^ ^°
l nb en
In any case, if "next_line_start" was adjusted we will scan for the word end again in the next iteration; but starting with the part of the word that does not fit on the new line. (This is unefficient of course, but this case is quite rare, and it's not worth adding special code for handling this.)
v v v v v
string: "I am a third example sentence, with anEvenABitLongerWord in me."
^ ^ ^^
l w en
+------------+
|I am a third|
|example |
|sentence, |
|with |
|anEvenABitLongerWord in me."
^ ^^ ^^
l wb en
As the new line is already filled up with the previous word part, another line break will always be inserted just in front of the remainder, and the remainder will be put into another line.
v v v v v
string: "I am a third example sentence, with anEvenABitLongerWord in me."
^ ^^ ^^
l wb en
+------------+
|I am a third|
|example |
|sentence, |
|with |
|anEvenABitLongerWord in me."
^ ^^ ^^
l wb en
e>b -> break
v v v v v
string: "I am a third example sentence, with anEvenABitLongerWord in me."
^° ^^
wb en
^
l
+------------+
|I am a third|
|example |
|sentence, |
|with |
|anEvenABitL\|
|ongerWord in me."
^ ^^
l en
Of course it's also possible that the remainder still doesn't fit on a line, and has to be broken again.
v v v v
string: "I am another example sentence, with aTerriblyLongWordThatDoesn'tSeemToEnd in me."
^ ^^ ^^
l wb en
+------------+
|I am another|
|example |
|sentence, |
|with aTerri\|
|blyLongWordThatDoesn'tSeemToEnd in me.
^ ^^ ^^
l wb en
e>b -> break
v v v v
string: "I am another example sentence, with aTerriblyLongWordThatDoesn'tSeemToEnd in me."
^ ^ ^ ^^
l w f en
+------------+
|I am another|
|example |
|sentence, |
|with aTerri\|
|blyLongWordThatDoesn'tSeemToEnd in me.
^ ^ ^ ^^
l w f en
e>f -> word break
v v v v
string: "I am another example sentence, with aTerriblyLongWordThatDoesn'tSeemToEnd in me."
^ ^^ ^ ^^
l wb t en
+------------+
|I am another|
|example |
|sentence, |
|with aTerri\|
|blyLongWordThatDoes\|
^ ^^ ^
l wb t
t>b -> don't fill
v v v v v
string: "I am another example sentence, with aTerriblyLongWordThatDoesn'tSeemToEnd in me."
^ ^^ ^
w nb e
^
l
+------------+
|I am another|
|example |
|sentence, |
|with aTerri\|
|blyLongWord\|
|ThatDoesn'tSeemToEnd in me.
^ ^^ ^
l nb e
This will be repeated, until the whole word is stored.
v v v v v
string: "I am another example sentence, with aTerriblyLongWordThatDoesn'tSeemToEnd in me."
^ ^^ ^^
l wb en
e>b -> break
v v v v v
string: "I am another example sentence, with aTerriblyLongWordThatDoesn'tSeemToEnd in me."
^ ^ ^^
l w en
^
f
e<=f -> no word break
v v v v v v
string: "I am another example sentence, with aTerriblyLongWordThatDoesn'tSeemToEnd in me."
^ ^^
w en
^
l
+------------+
|I am another|
|example |
|sentence, |
|with aTerri\|
|blyLongWord\|
|ThatDoesn't\|
|SeemToEnd in me.
^ ^^
l en
Of course, a line break is always inserted in front of the word no matter if it would fit on the old line, if the word follows a newline character.
v v v v v
string: "I am a stupid example sentence with a newline/in me."
^ ^ ^ ^^^
s l w enb
+------------+
|I am a |
[...]
|newline/in me.
^ ^ ^^^
l w enb
e<b, but w follows newline -> break
+------------+ |I am a | [...] |newline/ | |in me. ^ ^^ w en ^ l
calc_ywidth()
The third pass is simple again. calc_ywidth() calculates minimal heights (y-widths) for all items.
Presently this is complete overkill; assign_ywidth() doesn't really need this. However, it will be necessary as soon as interesting elements are implemented (tables) -- which hopefully won't be too long now...
Like calc_width(), it traverses the tree bottom to top, and for every item it stores the minimal height to "y_end".
Blank lines always have the height of 1. Anchor items have no height. (They are virtual...)
Text items have their height assigned already in assign_width(), while Breaking String into Lines.
################################################################################## +--------------------------------------------------------------------------------+___ |header text |__0 +--------------------------------------------------------------------------------+ 1<--y_end **********************************************************************************___ * *__0 ********************************************************************************** 1<-- [...] +--------------------------------------------------------------------------------+___ |this very long second paragraph contains some special characters (including a | 0 |simple space...): &; <>"=/ plus a big gap###and two unicode escapes (decimal: ¡ | 1 |and hexal: ¿) but also an anchor embedded inside a word (this anchor also is the| 2 |only tag with parameters); and finally a blank row |__3 +--------------------------------------------------------------------------------+ 4<-- +--------------------------------------------------------------------------------+___ |(a single tag) |__0 +--------------------------------------------------------------------------------+ 1<--
The height of a box item is calculated by summing up the height of all its sub-items. This is done by "first_child" and "next", just as seeking the widest sub-item in calc_width().
##################################################################################
+--------------------------------------------------------------------------------+
|header text |
+--------------------------------------------------------------------------------+ <-- 1
**********************************************************************************
* *
********************************************************************************** <-- +1
[...] [...]
+--------------------------------------------------------------------------------+
|this very long second paragraph contains some special characters (including a |
|simple space...): &; <>"=/ plus a big gap###and two unicode escapes (decimal: ¡ |
|and hexal: ¿) but also an anchor embedded inside a word (this anchor also is the|
|only tag with parameters); and finally a blank row |
+--------------------------------------------------------------------------------+ <-- +4
+--------------------------------------------------------------------------------+
|(a single tag) |
+--------------------------------------------------------------------------------+ <-- +1
################################################################################## <## 13
item_tree->y_end
assign_ywidth()
The fourth pass is also fairly simple. It assigns y-coordinates to all items. There is no initialisation necessary before calling it, as the y_end of the global item is already set by calc_ywidth(), and y_start is always 0.
Like in assign_width(), the tree is traversed top to bottom, and the sizes of all sub-items are assigned while processing the parent.
Thus, text and blank items do not need any processing -- they do not have any children.
For box items, the coordinates of all sub-items are assigned in a loop, like in assign_width(). Every item is put immediately after the previous one, i.e. it starts where the previous one ends. We keep track of the current position by "y_pos". At the beginning it is initialized to "y_start" of the box.
##################################################################################
0 <## <== y_pos
item_tree->y_start
For every item, y_start is set to the current y_pos.
y_end y_start
| |
+--###############################################################################____ v v
|header text |___0 <--
* *___1 <-- <**
|heading |___2 <** <----
* *___3 <---- <****
|first paragraph of text; includes multiple spaces and newlines, emphasized text | 4 <**** <--
|and strong text |___5
* *___6 <-- <**
|starting with an evil center tag, |___7 <** <----
|this very long second paragraph contains some special characters (including a | 8 <---- <-- <==
|simple space...): &; <>"=/ plus a big gap###and two unicode escapes (decimal: ¡ | cur_item->y_start=y_pos=8
|and hexal: ¿) but also an anchor embedded inside a word (this anchor also is the|
|only tag with parameters); and finally a blank row |
+--------------------------------------------------------------------------------+ <-- cur_item->y_end=4
+--------------------------------------------------------------------------------+
|(a single tag) |
+--------------------------------------------------------------------------------+ <---- 1
################################################################################## <## 13
"y_end" of an item is determined by adding the y-size (stored in "y_end" up to now) to "y_start" of the item.
+--###############################################################################____ |header text |___0 <-- * *___1 <-- <** |heading |___2 <** <---- * *___3 <---- <**** |first paragraph of text; includes multiple spaces and newlines, emphasized text | 4 <**** <-- |and strong text |___5 * *___6 <-- <** |starting with an evil center tag, |___7 <** <---- |this very long second paragraph contains some special characters (including a | 8 <---- <-- <== |simple space...): &; <>"=/ plus a big gap###and two unicode escapes (decimal: ¡ | 9 |and hexal: ¿) but also an anchor embedded inside a word (this anchor also is the| 10 |only tag with parameters); and finally a blank row |__11 +--------------------------------------------------------------------------------+ 12 <-- cur_item->y_end=cur_item->ystart+cur_item->y_end=12 +--------------------------------------------------------------------------------+ |(a single tag) | +--------------------------------------------------------------------------------+ <---- 1 ################################################################################## <## 13
"y_pos" is adjusted to the end of the item.
+--###############################################################################____ |header text |___0 <-- * *___1 <-- <** |heading |___2 <** <---- * *___3 <---- <**** |first paragraph of text; includes multiple spaces and newlines, emphasized text | 4 <**** <-- |and strong text |___5 * *___6 <-- <** |starting with an evil center tag, |___7 <** <---- |this very long second paragraph contains some special characters (including a | 8 <---- <-- |simple space...): &; <>"=/ plus a big gap###and two unicode escapes (decimal: ¡ | 9 |and hexal: ¿) but also an anchor embedded inside a word (this anchor also is the| 10 |only tag with parameters); and finally a blank row |__11 +--------------------------------------------------------------------------------+ 12 <-- <== +--------------------------------------------------------------------------------+ |(a single tag) | +--------------------------------------------------------------------------------+<---- 1 ################################################################################## <## 13
In the next iteration, "y_pos" -- which now points to the end of the current item -- is used as the beginning of the new item.
+--###############################################################################____ |header text |___0 <-- * *___1 <-- <** |heading |___2 <** <---- * *___3 <---- <**** |first paragraph of text; includes multiple spaces and newlines, emphasized text | 4 <**** <-- |and strong text |___5 * *___6 <-- <** |starting with an evil center tag, |___7 <** <---- |this very long second paragraph contains some special characters (including a | 8 <---- <-- |simple space...): &; <>"=/ plus a big gap###and two unicode escapes (decimal: ¡ | 9 |and hexal: ¿) but also an anchor embedded inside a word (this anchor also is the| 10 |only tag with parameters); and finally a blank row |__11 |(a single tag) | 12 <-- <---- <== +--------------------------------------------------------------------------------+<---- 1 ################################################################################## <## 13
Link Coordinates
Coordinates of links and anchors (both x and y!) are presently also assigned here. Probably it would be a better idea to do that in extra pass... As soon as the current link implementation is dropped and anchors are cleanly implemented, the assignment can be cleanly and logically split between assign_xwidth() and assign_ywidth(), just as for other item types.
Link coordinates are assigned by the parent text item. The text block is scanned, one link after the other, for the lines containing the link start and end. (The first line that ends after the link start, and the first line that ends at/after the link end.) As an optimization, the search for the next link is started in the line where the previous ended, not from beginning -- links can't be nested.
After having the line, (and thus also the y coordinate), the x coordinate is calculated by adding the link's relative position inside the line (link[].start-line_start) to the x coordinate of the line start. (Which is equal to the item start for normal text, but has to be calculated seperately for every line in centered text items.)
Inline anchors are very similar; however, they get their coordinates while processing the anchor item, not the text item containing them. (The text item doesn't know anything about the anchors.) Anchors can be nested if created by <span> or so; thus the above optimization isn't possible. (It would be harder anyhow, due to the anchors being processed every one on its own.)
Block anchors need another processing, of course: If they are empty, they keep the coordinates assigned to them by the parent box; otherwise, all (immediate) virtual children are scanned for the minimum/maximum for each of the four coordinates; these are assigned as the coordinates of the anchor virtual box.
create_map()
The last sub-pass generates the "page_map[]". (This could be also done inside assign_ywidth()...) This map is necessary to quickly determine which elements show up in the visible area of the page, when it is displayed in the viewer.
The page usage map stores references to all items that show up in any given line of the output page. This is a very simple approach, but it should be perfectly sufficient as long as netrik has no full graphic mode -- and this will be probably for quite a while...
"page_map[]" is an array containing an "Item_list" structure for every line of the output page. This structure is declared in items.h . It contains a count of items in this line, and an array of pointers to the items.
The whole line map is allocated at the beginning. Afterwards, all items are processed in a loop. For every item that is visible on the screen (presently this are only text items), a reference to this item is stored to every line of "page_map[]" between "y_start" and "y_end" of the item.
line|page_map[line]
+--###############################################################################__ ----+--------------
|header text |__ 0 | t0
* *__ 1 | -
|heading |__ 2 | t1
* *______3_|_-____
|first paragraph of text; includes multiple spaces and newlines, emphasized text | 4 | t2 <-- cur_item->y_start
|and strong text |______5_|_t2___
* *__ 6 | - <-- cur_item->y_end
|starting with an evil center tag, |__ 7 | -
|this very long second paragraph contains some special characters (including a | 8 | -
|simple space...): &; <>"=/ plus a big gap###and two unicode escapes (decimal: ¡ | 9 | -
|and hexal: ¿) but also an anchor embedded inside a word (this anchor also is the| 10 | -
|only tag with parameters); and finally a blank row |__ 1 | -
|(a single tag) |__ 12 | -
###############################################################################--+
"page_map[]" is returned to pre_render(), and from there to main(), where it is passed to dump() or to render() (via display(), see hacking-pager.*) along with "item_tree".
free_map()
This function frees the memory allocated for the page map.
First it goes through the table line by line, and frees the associated "list" for each one. Afterwards, it frees the table itself.
8. render.c
With the item tree and the page usage map prepared in pre_render(), we can now actually render the page. There are two different rendering functions:
dump() renders the whole page and dumps the output to the terminal. The output is layouted correctly, using all the coordinates and text attributes.
render() renders only a specified area of the page, and outputs it to the curses screen.
dump()
The page is rendered line by line, using "page_map[]" to determine which items we need to render in every line. Of course this isn't really necessary, as we do not ever have more than one item in a line presently. However, using "page_map[]" here is *not* overkill, just for a change ;-) On the contrary, this is a pragmatic approach. Dumping the page item by item without using "page_map[]" would be more efficient; however, it would be also more complicated than dumping line by line.
In every line, we process all items (from "page_map[line]") one after the other. If there is actually more then one item (which is impossible presently...), they are printed one after the other. As long as they do not overlap (they will at some point, but not too soon...), they shouldn't interfere with each other.
At the beginning of processing an item, some constants are set. "item" and "string" point to the processed item and its text area.
"string_line" is the line number inside the text block of the line we are just printing. That is to say, it's the offset to the first line of the text block.
2 |[...] | 3 |+-----------------------------------------------+| 4 <-- item->y_start=4 0 ||first paragraph of text; includes multiple || 5 <-- line=5 string_ --> 1 ||spaces and newlines, emphasized text and strong|| 6 line=1 2 ||text || 7 |+-----------------------------------------------+| 8 |[...] |
"text_start" tells where this line beginns inside the string. Normally, it is taken directly from "line_table[string_line-1]", which is the position of the line break at which the line starts.
line_table[0]
v
text: "first paragraph of text; includes multiple spaces and newlines, [...]"
^
text_start=line_table[string_line-1]
As line 0 of a text block doesn't start with a line break, its start isn't stored in "line_table[]"; we have to test for this case and set "text_start" to 0.
v
text: "first paragraph of text; includes multiple spaces and newlines, [...]"
^
text_start
"text_end" tells where this line ends; it is set to the beginning of the next line ("line_table[string_line]").
line_table[0] line_table[1]
v v
text: "[...]multiple spaces and newlines, emphasized text and strong text"
^ ^
text_start text_end=line_table[string_line]
As the last line does not end with a line break, we have to test for this case end set "text_end" to the string end.
v v
text: "[...]multiple spaces and newlines, emphasized text and strong text"
^ ^
text_start text_end
Maybe we should store the beginning of the first line and the end of the last line to "line_table[]" also. This would simplify the code (we wouldn't need the tests), and the efficiency loss should be quite irrelevant.
"text_end" isn't declared as a constant, because it may be altered: If the last character in the line turns out to be a blank (or newline), it isn't printed. We accomplish this by decrementing "text_end" by one.
v v
text: "[...]multiple spaces and newlines, emphasized text and strong text"
^ ^
text_start text_end
After knowing start and end of the line, we calculate "x_start", which is the screen column at which the line will start. Normally, this is equal to "x_start" of the text item.
||spaces and newlines, emphasized text and strong||
^
x_start=item->x_start
If the text item is centered, we have to add an individual offset to each line. The offset is half of the unused space inside the line, i.e. the difference between the width of the text block and the width of the text in this line.
text_start text_end
v v
||starting with an evil center tag, ||
^ ^
item->x_start item->x_end
|<---- text_end-text_start ----->|
|<--------------- x_end-x_start -------------->|
|<----------->|
unused
|<---->|
unused/2
v v
|| starting with an evil center tag, ||
^ ^ ^
item->x_start x_start=item->xstart+unused/2
|<---->|
unused/2
Then we set the cursor to the calculated col. This we do by a carrige return (not really needed for the first item in a line...) and than going forward as much character positions as needed.
Now we can output the text itself. We print all attribute divisions in a loop. But first we have to find the first one that shows up in the current line.
div[0].end ... div[3].end
v v v v
text: "first[...]multiple spaces and newlines, emphasized text and strong text"
| ^ | ^
text_start
In every iteration we print all the text between "div_start" end "div_end", which normally point to the start and end of the current division. For the first iteration, "div_start" is set to "text_start" -- we only want the part of the div that shows up in the current text line.
v v v v
text: "first[...]multiple spaces and newlines, emphasized text and strong text"
^ ^
div_start=text_start div_end=div[0].end
output:
first paragraph of text; contains multiple
spaces and newlines,
The next div starts where the current one ends.
v v v v
text: "first[...]multiple spaces and newlines, emphasized text and strong text"
^ ^ ^ ^
text_start div_start div_end
The last division is truncated to "text_end" -- as with the first one, we only want the part that shows up in the current line.
v v v v
text: "first[...]multiple spaces and newlines, emphasized text and strong text"
^ ^ ^
div_start div_end
=text_end
Finally, we test if the line ends with a word break, and print the break character if it does. We know it does when the character at the line end is a word character (not a space, newline, or string end), because every word end is followed by the space separating it from the next word; if there is no space at the line end, we are inside a broken word.
v v
text "Some sentence containing aVeryLongAndThusBrokenWord."
^
output:
Some sentence
containing aVeryLongAn\
...
dThusBrokenWord.
render()
render() works similar to dump(). The main difference is how the output is printed. However, there are a couple of differences in screen position handling and other calculations also.
render() takes the starting positon of the rendered area inside the page, the starting position on the screen, and the size of the rendered area as arguments.
The area is processed line by line, and every line is processed item by item, just like in dump().
"x_start" describes the starting column relative to the beginning of the rendered area, not the screen. (dump() always dumps whole lines, and thus there is no difference.) If it turns out that the line starts before the rendered area, it has to be truncated. the ending position "x_end" is calculated in a similar fashion.
Before the line is printed (div by div, as in dump()), the cursor is set to the start position of the line. The column is the starting position relative to the area ("x_start"), plus the starting position of the area on the screen; likewise the row.
Before anything is printed, we test if some part of the line shows up inside the rendered area at all. The word break indicator is also printed only if the line ends inside the area.
If render() was called with the "overpaint" flag, the requested area is cleaned before rendering anything, so any garbage will be removed. (Areas not containing any text aren't affected otherwise.) This is done by overwriting the desired part of each line with a string of spaces.
dump_items()
dump_items() dumps the item tree, including the text of text items. The text is printed with correct attributes, but ignoring any coordinates and line breaks. This function is for debugging purposes, and may be called anywhere inside or after pre_render() (anywhere after parse_struct()).
The reason this function is in render.c is that it needs the same screen handling functions as dump(). Moreover, it works in a very similar fashion.
The difference is that it does not dump line by line, but item by item (traversing the tree top to bottom). After printing some information about the item itself, it dumps the text division by division in the same way dump() does, only it doesn't need to care about positions or line breaks; it dumps the whole string at once.