Aloha,
since I have quite a bit of XS experience, I thought I'd tackle
implementing a few methods in PPI::XS to speed things up. To find
bottlenecks I used the module I talked about in the earlier message and
let it analyze a ~3500 line program under the supervision of a
profiler.[1] The profiles unveiled a few hotspots:
* PPI::Node::find_first, 9 calls, ~9 seconds
* PPI::Tokenizer::_process_next_char, 44305 calls, ~3.5 seconds
* PPI::Lexer::_lex_statement, 2908 calls, ~3 seconds
PPI::Node::find_first calls many methods and does little computation on
its own, so implementing it in XS wouldn't provide much benefit. The
same is true for PPI::Lexer::_lex_statement.
PPI::Tokenizer::_process_next_char looked more promising. So I wrote an
XS port of it. Patch attached.
Well, the bad news is that this didn't improve performance measurably.
I looked at a few other candidates in the profiles but the all seem to
have the same issue: not much computation, lots of method calls. I
don't think you can improve code like this very much by simply porting
it to XS verbatim. perl already does a pretty good job when it creates
the optree, it seems.
So, any suggestions on what might be a good candidate for porting?
--
Bye,
-Torsten
[1] I tried Devel::DProf, Devel::Profile, and Devel::AutoProfiler. All
yielded similar results.
|
