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



Ok. I will get through this.
I tried to change the imports.c code before. But it broke the execution. It doesn't find the templates correctly. I am not sure for what reason. Do i miss something while filling the hash table?

Thanks!

On Thu, Aug 27, 2015 at 2:32 PM, Nick Wellnhofer <wellnhofer aevum de> wrote:
On 20/08/2015 19:21, Christian Ceelen wrote:
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

Some quick comments:

- It's nice that you create the hash lazily.
- The helper function xsltHasNamedTemplate doesn't seem to be
  necessary since it's used only once.
- xmlHashCreate(1024) creates a hash table with 1024 initial slots.
  This value should be reduced to something sensible like 4.

I think reordering structs in xsltInternals.h doesn't break ABI compatibility. So `namedTemplatesHash` could be moved to the other template fields. But I'm not 100% sure.

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.

Put your test files in:

- tests/REC/test-6-2.xml
- tests/REC/test-6-2.xsl

Then run `make tests`. This will create files

- tests/REC/test-6-2.out
- tests/REC/test-6-2.err (if there's stderr output)

Inspect the contents of the `.out` and `.err` files. If everything's OK, add them to version control.

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?

I don't think it's necessary to add any helper functions. But please change the lookup code in imports.c to make use of the named templates hash.

Nick



[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]