Need help optimizing a regexp

Hi.

In JMeter development I've encountered the need to parse some specific 
HTML constructs REALLY FAST. To do this, I was using this regular 
expression:

<BASE(?=\s)[^\>]*\sHREF\s*=\s*"([^">]*)"
|<(?:IMG|SCRIPT)(?=\s)[^\>]*\sSRC\s*=\s*"([^">]*)"
|<APPLET(?=\s)[^\>]*\sCODE(?:BASE)?\s*=\s*"([^">]*)"
|<(?:EMBED|OBJECT)(?=\s)[^\>]*\s(?:SRC|CODEBASE)\s*=\s*"([^">]*)"
|<(?:BODY|TABLE|TR|TD)(?=\s)[^\>]*\sBACKGROUND\s*=\s*"([^">]*)"
|<INPUT(?=\s)(?:[^\>]*\s(?:SRC\s*=\s*"([^">]*)"|TYPE\s*=\s*"image")){2,}
|<LINK(?=\\s)(?:[^\>]*\s(?:HREF\s*=\s*"([^">]*)"|REL\s*=\s*"stylesheet\")){2,}

but my ORO-based parser could hardly compete with Sourceforge's 
HtmlParser for the same job, which was shocking to me, because 
HtmlParser does much more work...

In a random attempt to improve the situation, I tried this:

<(?:
BASE(?=\s)[^\>]*\sHREF\s*=\s*"([^">]*)"
|(?:IMG|SCRIPT)(?=\s)[^\>]*\sSRC\s*=\s*"([^">]*)"
|APPLET(?=\s)[^\>]*\sCODE(?:BASE)?\s*=\s*"([^">]*)"
|(?:EMBED|OBJECT)(?=\s)[^\>]*\s(?:SRC|CODEBASE)\s*=\s*"([^">]*)"
|(?:BODY|TABLE|TR|TD)(?=\s)[^\>]*\sBACKGROUND\s*=\s*"([^">]*)"
|INPUT(?=\s)(?:[^\>]*\s(?:SRC\s*=\s*"([^">]*)"|TYPE\s*=\s*"image")){2,}
|LINK(?=\\s)(?:[^\>]*\s(?:HREF\s*=\s*"([^">]*)"|REL\s*=\s*"stylesheet\")){2,}
)

now my test runs REALLY FAST!

First question (out of sheer curiosity): why is this later regexp faster 
than the earlier one?

Second question: I would like to run the regexps against the HTML 
content as a byte array (byte[]) without having to convert it into a 
string. Can ORO do this?

On the reasons why I don't want to do the byte[]-to-String conversion:
1/ Memory efficiency.
2/ I don't need it: even if there were multi-byte characters in the 
input, they are not part of my problem.
3/ The conversion can cause problems if the input is wrong.

Third question: I've read that byte-based regexp engines use a type of 
state machines which is significantly faster than char-based regexp 
engines. Am I correct? Can ORO take advantage of this? Could you 
recommend a regexp engine which can?

Thanks for your help,

Jordi.