The XML/HTML lexer is now capable of processing most invalid XML/HTML
(that I can think of at least). This is achieved by inserting missing
closing tags (where needed) and/or ignoring excessive closing tags. For
example, HTML such as this:
<a></a></p>
Results in the following tokens:
[:T_ELEM_START, nil, 1]
[:T_ELEM_NAME, 'a', 1]
[:T_ELEM_CLOSE, nil, 1]
In turn this HTML:
<a>
Results in these tokens:
[:T_ELEM_START, nil, 1]
[:T_ELEM_NAME, 'a', 1]
[:T_ELEM_CLOSE, nil, 1]
Fixes#84
Previously a single Ragel machine was used for processing HTML
script and style tags. This had the unfortunate side-effect that the
following was not parsed correctly (while being valid HTML):
<script>
var foo = "</style>";
</script>
The same applied to style tags:
<style>
/* </script> */
</style>
By using separate machines we can work around the above issue. The
downside is that this can produce multiple T_TEXT nodes, which have to
be stitched back together in the parser.
Similar to comments (ea8b4aa92f) and CDATA
tags (8acc7fc743) processing instructions
are now lexed in separate chunks _with_ proper support for streaming
input.
Related issue: #93
Instead of using a single token (T_CDATA) for a CDATA tag the lexer now
uses 3 tokens:
1. T_CDATA_START
2. T_CDATA_BODY
3. T_CDATA_END
The T_CDATA_BODY token can occur multiple times and is turned into a
single value in the XML parser. This is similar to the way strings are
lexed.
By changing the way CDATA tags are lexed Oga can now lex CDATA tags
containing newlines when using an IO as input. For example, this would
previously fail:
Oga.parse_xml(StringIO.new("<![CDATA[\nfoo]]>"))
Because IO input reads input per line the input for the lexer would be
as following:
"<![CDATA[\n"
"foo]]>"
Related issues: #93
This cache is flushed whenever Element#register_namespace is called.
When this cache is flushed it's also recursively flushed for all child
elements. This makes calls to Element#register_namespace a bit more
expensive but in turn calls to Element#available_namespaces will be a
lot faster.
The results of these methods is now cached until a Node is moved into
another NodeSet. This reduces the time spent in the
xpath/evaluator/big_xml_average_bench.rb benchmark from roughly 10
seconds to roughly 5 seconds per iteration.
In HTML the text of a script/style tag should be left untouched, no
entities must be converted. Doing so would break Javascript such as the
following:
foo&&bar;
Such code is often the result of minifiers doing their dirty business.
This was broken by introducing the process of lazy decoding of XML/HTML
entities. The new setup works similar to how XML::Text#text decodes any
entities that may be present.
Fixes#91
When querying an XML document that explicitly defines the default XML
namespace the XPath evaluator now correctly matches all nodes within
that namespace if no namespace prefix is given in the query. Previously
this would always return an empty set.
Instead of trying to make this class thread-safe I'm going with the
option of simply declaring it unsafe to mutate instances of XML::Text
while reading it in parallel. This removes the need for Mutex
allocations and keeps the code simple.
Fixes#82
Currently all operators are left-associative with no particular precedence. This
causes a few specs to fail for now. Outside of that the new parser should be
able to parse the same input as the Racc based parser.
Removing this makes the process of parsing larger XML documents a bit faster.
The downside is that NodeSet#initialize will no longer filter out duplicate
nodes, though this is not something Oga itself relies upon.
Methods such as NodeSet#push still do ignore elements already present.
This was utterly broken, mainly due to me overlooking it. There are now 2 new
callbacks to handle this properly:
* on_attribute: to handle a single attribute/value pair
* on_attributes: to handle a collection of attributes (as returned by
on_attribute)
By default on_attribut returns a Hash, on_attributes in turn merges all
attribute hashes into a single one. This ensures that on_element _actually_
receives the attributes as a Hash, instead of an Array with random
nil/XML::Attribute values.
Using Array#+ for large sets (e.g. in the benchmarks) is _really_ slow.
Interesting enough Array#unshift uses as much memory as the Racc parser and is
about as fast, even though it has to move memory around.
Instead of decoding entities in the lexer we'll do this whenever XML::Text#text
is called. This removes the overhead from the parsing phase and ensures the
process is only triggered when actually needed. Note that calling #to_xml and/or
the #inspect methods on a Text (or parent) instance will also trigger the entity
conversion process.
The new entity decoding API supports both regular entities (e.g. &) as well
as codepoint based entities (both regular and hexadecimal codepoints).
To allow safe read-only access to Text instances from multiple threads a mutex
is used. This mutex ensures that only 1 thread can trigger the conversion
process.
Fixes#68
This basically re-applies the technique used for HTML <script> tags. With this
extra addition I decided to rename/normalize a few things so it's easier to add
any extra tags in the future. One downside of this setup is that the following
will not be parsed by Oga:
<style>
</script>
</style>
The same applies to script tags containing a literal </style> tag. Since this
particular case is rather unlikely to occur I'm OK with not supporting it as it
_does_ simplify the lexer quite a bit.
Fixes#80
When lexing input in HTML mode the lexer has to treat _all_ content of a
<script> tag as plain text. This ensures that the lexer can process input such
as "x <y" and "// <foo>" correctly.
Fixes#70.
This changes the behaviour of after_element when parsing documents using the SAX
parsing API. Previously it would always receive a nil argument, which is kinda
pointless. This commit changes that by making sure it receives a namespace name
(if any) and the element name.
This fixes#54.
The old XPath "position() = 1" would work in Nokogiri due to the way they
retrieve descendants. In Oga however this would simply always return the first
node.
To fix this Oga now counts the amount of preceding siblings that match the same
full name.
This means that "foo[1]" uses this AST:
(predicate (test nil "foo") (int 1))
Instead of this AST:
(test nil "foo" (int 1))
This makes it easier for the XPath evaluator to process predicates correctly.
This method can be used to compare two NodeSet instances. By using
XML::NodeSet#equal_nodes?() the need for exposing the "nodes" instance variable
is also removed.
This expression could be used to get all elements that _don't_ have any
namespace. The problem is that this can't be expressed as just a node test,
instead the resulting XPath would have to look something like the following:
X[local-name() = name()]
However, since the XPath predicates are already created for pseudo classes and
such, also injecting the above into it would be a real big pain. As such I've
decided not to support it.
Instead of using "descendant-or-self" Oga will use "descendant". This ensures
that expressions such as "foo *" don't return a set also including the "foo"
element.
Nokogiri solves this problem in a somewhat different way by using //foo//* for
the CSS expression "foo *". While this works in Nokogiri the expression
"descendant-or-self::node()" is slow as a snail in Oga (due to its nature of
retrieving _all_ nodes first). By using "descendant" we can work around this
problem.
When running XPath queries such as "self::node()" the result should be a set
containing the document itself. This in turn fixes expressions such as
descendant-or-self::node()/a.
This removes parsing support for selectors such as :nth-child(-n-6). According
to the CSS spec this isn't valid anyway (confirmed by testing it in Chromium).
As a result there's no point in supporting it in any way.
When lexing multi-line strings everything used to work fine as long as the input
were to be read as a whole. However, when using an IO instance all hell would
break loose. Due to the lexer reading IO instances on a per line basis,
sometimes Ragel would end up setting "ts" to NULL. For example, the following
input would break the lexer:
<foo class="\nbar" />
Due to the input being read per line, the following data would be sent to the
lexer:
<foo class="\n
bar" />
This would result in different (or NULL) pointers being used for building a
string, in turn resulting in memory allocation errors.
To work around this the string lexing setup has been broken into separate
machines for single and double quoted strings. The tokens used have also been
changed so that instead of just "T_STRING" there are now the following tokens:
* T_STRING_SQUOTE
* T_STRING_DQUOTE
* T_STRING_BODY
A string can have multiple T_STRING_BODY tokens (= multi-line strings, only the
case for IO inputs). These strings are stitched back together by the parser.
This fixes#58.
Processing of this axis along with a predicate wouldn't quite work out. Even if
the predicate returned false the node would still be matched (which should not
be the case).
Previously input such as "x > y" would result in the following token sequences:
T_IDENT, T_CHILD, T_IDENT
This commit changes this to the following:
T_IDENT, T_SPACE, T_CHILD, T_IDENT
This allows the parser to use T_SPACE as a terminal token, this in turn prevents
around 16 shift/reduce conflicts from arising.
This does mean that input such as " > y" or " x > y" is now invalid. This
however can be solved by simply _not_ adding leading/trailing whitespace to CSS
queries.
The new setup will not involve a separate transformation stage, instead the CSS
parser will directly emit an XPath AST. This reduces the overhead needed for
parsing/evaluating CSS selectors while also simplifying the code. The downside
is that I basically have to re-write 80% of the parser.
This uses stricter (and more correct) rules in both the lexer and the parser.
The resulting AST has also received a small rework to make it more compact and
less confusing.
Yorick Peterse
Dmitry Krasnoukhov