GNOME.org
 

Re: [xslt] XSLT compilation is slow for big XSLTs

On Mon, Sep 07, 2015 at 09:47:40AM +0200, Christian Ceelen wrote:

Hi,

Did anyone find the time to look into this or test this?


  Sorry Christian,

not forgotten but not done yet, will try this week !

  thanks,

Daniel


Cheers,
Christian

On Mon, Aug 31, 2015 at 11:47 AM, Christian Ceelen <
christian ceelen gmail com> wrote:


Hi,

i think i no know what i did wrong while filling the hash table. And yes.
My bad ... :

xslt.c: 5448
+ xmlHashAddEntry2(style->namedTemplatesHash, ret->name, ret->nameURI,
template);

This adds a pointer to the xmlNode in the parsed stylesheet to the hash
table. For just detecting collisions that seems to be ok.
But it is utterly useless for using it in xsltFindTemplate in 'import.c'
as that expects a pointer to the compiled template struct 'xsltTemplate'.
So my initial patch for reusing it doing a bad pointer cast here:
--- a/libxslt/imports.c
+++ b/libxslt/imports.c
@@ -400,19 +400,12 @@ xsltFindTemplate(xsltTransformContextPtr ctxt, const
xmlChar *name,
        return(NULL);
     style = ctxt->style;
     while (style != NULL) {
-       cur = style->templates;
-       while (cur != NULL) {
-           if (xmlStrEqual(name, cur->name)) {
-               if (((nameURI == NULL) && (cur->nameURI == NULL)) ||
-                   ((nameURI != NULL) && (cur->nameURI != NULL) &&
-                    (xmlStrEqual(nameURI, cur->nameURI)))) {
-                   return(cur);
-               }
-           }
-           cur = cur->next;
-       }
-
-       style = xsltNextImport(style);
+        if(style->namedTemplatesHash != NULL){
+            cur = (xsltTemplatePtr)
xmlHashLookup2(style->namedTemplatesHash, name, nameURI);
+            if(cur != NULL)
+                return (cur);
+        }
+        style = xsltNextImport(style);
     }
     return(NULL);
 }

This explains why it was not working as i had hoped. The hash table
returns the xmlNodePtr instead of the xsltTemplatePtr expected by
xsltFindTemplate. The patch Daniel does the same, it expects
xsltTemplatePtr.

So changing in xslt.c: 5448
+ xmlHashAddEntry2(style->namedTemplatesHash, ret->name, ret->nameURI,
template);
to
+ xmlHashAddEntry2(style->namedTemplatesHash, ret->name, ret->nameURI,
ret);
Makes it possible to reuse the hash table for the named templates also in
xsltFindTemplate.

I pushed the updated changes to github and attached a clean patch to the
email. It is created against 1.1.28.

The patch adds the hash table initialization and population to
xsltAddTemplate (pattern.c), sets the initial size to a more reasonable 8
entries, reuses it in xsltFindTemplate (imports.c), and adds a test case to
check that an error is emitted during compilation for template names that
are not unique.

So far i haven't replaced the helper functions as Nico suggested, as i was
thinking of moving the look up code to imports.c and share part of a lookup
local to a single stylesheet between xsltFindTemplate and
xsltParseStylesheetTemplate. I am not sure if you rather go for a small
code duplication or for having small helper classes. In terms of
performance i doubt that there is a noticeable difference.

Best regards,
Christian


On Mon, Aug 31, 2015 at 8:27 AM, Daniel Veillard <veillard redhat com>
wrote:


On Thu, Aug 20, 2015 at 07:21:50PM +0200, Christian Ceelen wrote:

Hi,

Ok. I think i found the answer by trial and error myself. I was not

able to

look up the named templates in the hash table for the (pre-) compiled
templates.

So i followed Nick's suggest and came up with this first draft:
   https://github.com/GNOME/libxslt/pull/1


  So I applied the patch, this looks reasonable, we need to add to the
end of the structure, to minimize risk of ABI breakage. Should change
the 1024 value that Nick pointed out.


I tested locally and didn't get any surprises. I also created a simple

test

with 2 named templates that have the same name to test for collision
detection. The new code behaves in that case as the previous one. But so
far I have no idea yet where and how to add it to the tests in the test
suite as this is a test for failure detection.

Looking at a second glance at the imports.c and xsltFindTemplate i

think it

would be worthwhile to add the new function i created for lookup in the
hash table in the file imports.c and try to encapsulate the behavior

there.

Do you think i should refactor the code in that direction?


  That's where things started to go wrong, I modified xsltFindTemplate()
to look in the namedTemplatesHash instead of digging the templates list,
patch is attached, and that raised a number of errors in the regression
tests.
So it seems to me that namedTemplatesHash isn't initialized with all the
named templates from the stylesheet and hence the collision detection
is also likely wrong even if it passed an initial simple test.

  So I think we need to give it more scrutiny, i.e. each time a template
is
added to style->templates we need to make sure we also add it to
style->namedTemplatesHash.



Best regards,

Christian




On Thu, Aug 20, 2015 at 10:28 AM, Christian Ceelen <
christian ceelen gmail com> wrote:


Hi Nick,

Thank you very much for the answer. That sounds like a very good

idea. Yet

would the validation still be correct, if the hashtable is local to

the

stylesheet in which the template was defined? I was wondering what

would

happen if the same template name is defined in two files, where one

of the

files then imports or includes the other. Concerning the imports i

think

the standard is clear, the defined template has lower preference, so

having

two with the same name sounds legit to me. But what happens when the

file

is included?
My current understanding is that real validation requires to walk the
include tree for each template name. Would that be correctly

replicating

the existing behavior?

Btw. In xsltInernals.h:L1514 i found in struct _xsltStylesheet:
void *templatesHash;
Is this table already containing what we want?

Best regards,

Christian

On Wed, Aug 19, 2015 at 1:00 PM, Nick Wellnhofer <wellnhofer aevum de



wrote:


On 19/08/2015 10:47, Christian Ceelen wrote:


At the moment i do not have a deep understanding of the libXSLT

code. My

current guess is, that this point at which the template name is
validated to
be unique in the stylesheet. The current algorithm seems to be

walking

linearly through the list of all already created templates and is
comparing
the name with each. Given that the compilation process is walking
through all
templates this loop means that we have an O(n^2) algorithm (with n

being

the
number of template instances in the stylesheet to compile).
The huge number of templates in my XSLTs are just so far over the

edge,

that
the compilation takes 35s. I ran a test in which i skipped the

loop. This

reduced the compilation time to below 2.5 s.

Would anyone let me know if i have understood the code? What can i

do to

improve the code that would get easily accepted and released? I am

open

to any
kind of suggestions on what to do to speed this validation step up

with

a data
structure or mechanism already existing ?



Yes, template names are verified by searching a linked list which is
O(n^2). Template lookup by name uses the same list:



https://github.com/GNOME/libxslt/blob/master/libxslt/imports.c#L375


It shouldn't be too hard to change the code to use an xmlHashTable:



https://github.com/GNOME/libxml2/blob/master/include/libxml/hash.h


Simply add a the hash table to struct _xsltStylesheet




https://github.com/GNOME/libxslt/blob/master/libxslt/xsltInternals.h#L1509


add some initialization and cleanup, then use xmlHashAddEntry2 and
xmlHashLookup2 with the template's name and namespace URI. This

should make

name verification O(n) overall and lookup of a single template by

name O(1).


What is now the developer or working repository for libxslt?

I found the git repository on git.gnome.org <http://git.gnome.org>,

the

GNOME/libxslt on github, plus several forks which seem to be working
repos.
Can you point me please to the repo against which i should use for
further
investigation or patches?



The official repo is the one at git.gnome.org:

    https://git.gnome.org/browse/libxslt/

Nick

_______________________________________________
xslt mailing list, project page http://xmlsoft.org/XSLT/
xslt gnome org
https://mail.gnome.org/mailman/listinfo/xslt







_______________________________________________
xslt mailing list, project page http://xmlsoft.org/XSLT/
xslt gnome org
https://mail.gnome.org/mailman/listinfo/xslt



--
Daniel Veillard      | Open Source and Standards, Red Hat
veillard redhat com  | libxml Gnome XML XSLT toolkit  http://xmlsoft.org/
http://veillard.com/ | virtualization library  http://libvirt.org/







_______________________________________________
xslt mailing list, project page http://xmlsoft.org/XSLT/
xslt gnome org
https://mail.gnome.org/mailman/listinfo/xslt



-- 
Daniel Veillard      | Open Source and Standards, Red Hat
veillard redhat com  | libxml Gnome XML XSLT toolkit  http://xmlsoft.org/
http://veillard.com/ | virtualization library  http://libvirt.org/
[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]