Sizzle.matches passes null context, causes document to be set to global document

[catrope]

Related to #292

If you call, say, $foo.children( 'a' ) , checking which children of $foo match the 'a' selector is done by jQuery.filter. If $foo has exactly one child, jQuery.filter will call Sizzle.matchesSelector which invokes setDocument( elem.ownerDocument ) if necessary, and then passes that document to the Sizzle constructor.

However, if the number of children of $foo isn't exactly one, jQuery.filter will call Sizzle.matches instead, which just calls the Sizzle constructor directly, passing in null for the context, causing an invocation of setDocument( window.document ) regardless of what document $foo is really in.

This means that the following code causes the document context to thrash: setDocument is invoked from every .children() call, which is detrimental to performance, since setDocument contains 300+ lines of feature/bug detection code.

var doc = document.implementation.createHTMLDocument( '' ),
    $foo = $( '<p><b>Hi</b></p>', doc ),
    $bar = $( '<p><b>Hello</b><i>world</i></p>', doc ),
    $baz = $( '<p>', doc );

// Context is doc due to $( '....', doc ) above
$baz.children( 'b' ); // Sets context to window.document
$foo.children( 'b' ); // Sets context to doc
$bar.children( 'b' ); // Sets context to window.document
$foo.children( 'i' ); // Sets context to doc
$bar.children( 'i' ); // Sets context to window.document

[timmywil]

Thanks for opening an issue. We actually had a discussion about this recently and meant to open one.

[catrope]

Yeah, one of my coworkers started that discussion, but I was the one writing our workarounds for it and I needed a URL to use to refer to this problem in my commit messages :)

I've also noticed anecdotally that if you do something like $foo.find( 'a' ).filter( '[href]' ) where $foo is in a foreign document, two context switches occur for every call. That's probably the same underlying issue as this one, since .filter() calls the same Sizzle internals as .children()

[timmywil]

Implementation discussion here. @gibson042 or I will get to it as soon as one of us is free.