I did a profile of a SAX parse using Xerces 1.1.0 d05 and Numega's TrueTime and
found what looks to be a very ripe opportunity for a substantial performance
gain.
The most obvious hot-spot was XMLAttr::set() which accounted for 26% of the
total time. I've reproduced the code here:
void XMLAttr::set( const unsigned int uriId
, const XMLCh* const attrName
, const XMLCh* const attrPrefix
, const XMLCh* const attrValue
, const XMLAttDef::AttTypes type)
{
// Clean up the old stuff
delete [] fName;
fName = 0;
delete [] fPrefix;
fPrefix = 0;
delete [] fValue;
fValue = 0;
// And clean up the QName and leave it undone until asked for
delete [] fQName;
fQName = 0;
// And now replicate the stuff into our members
fType = type;
fURIId = uriId;
fName = XMLString::replicate(attrName);
fPrefix = XMLString::replicate(attrPrefix);
fValue = XMLString::replicate(attrValue);
}
XMLAttr seems to be called XMLAttr::XMLAttr() (10 times) and
XMLScanner::scanStartTag() (4564 times) in my sample document. 57% of the
time is spent in the operator []'s and 36% in the replicates.
Since attributes are generally about the same size, it would seem to be much
more efficient to attempt to reuse the existing "stuff" instead of explicitly
deleting it then reallocating it. I don't
think that it requires anything as sophisticated as a string pool. Maybe just
always allocate fName (for example) at least, say, 64 characters. If the new
values is less than 64 characters, copy the
new value into fName otherwise allocate a longer fName.
Another anomaly is that I was not doing a validating parse, but I spent 10% of
the time in DTDValidator::scanDTD().
I wish I could give some more insights, but I haven't found the trick to set
breakpoints on code in Xerces Lib in VC6. I think it has been covered here
before, but I wasn't able to locate it. If
someone wants to send me a pointer, that would be helpful.
