performance improvements for 1.3

I'm going to check into 1.3 a set of performance improvements based on
work at Amazon over the past month. Some methods have been significantly
rewritten and there are a few minor backward incompatibilities, but I
believe it'll be well worth it. The main focus has been to reduce
component call overhead as much as possible.

Don't have any reportable numbers yet wrt improvements. I'll try our
checked-in benchmarks before and after the changes, but really, the
speed-up will vary greatly from site to site depending on the quantity
and pattern of component calls.

Summary of changes:

* Replace the Buffer object with a simple scalar, and replace the
generated $m->print() call in components with in-line concatenation
unless autoflush is enabled.
* Delay computation of base_comp for each stack frame until it is needed.
* Create a cache from the uncanonicalized paths used to call components
(e.g. <& foo/bar.m &>) into actual component objects when in
static_source mode.
* Make the request stack a dynamically scoped array, so that it does not
have to be manually popped after each component.
* Make each request stack frame an array rather than a hash.
* Allow preallocation of output buffer to a large string to reduce 
unnecessary mallocs.
* Determine if we are in Perl debugger at compile time, and don't call
debug hooks unless we are.

Backward incompatibilities:

* Can no longer print a list inside <% %>, e.g. <% @foo %>, because 
these tags are now translated into simple in-line concatenations. This 
is no great loss since you can just use a join. If people think it is 
warranted, we can continue to support this syntax (and disable the 
optimization) for a limited time via a special compiler flag.
* A new parameter, enable_autoflush, must be passed to the compiler or 
interpreter in order for autoflush to work. Again, this is to support 
simple in-line concatenations. If you use autoflush on just a few pages 
on your site, it is probably better to manually call $m->flush_buffer 
than incur this penalty for all components.

This is being tracked in todo item #555. Watch for the code check-ins
soon, and please comment as you see fit.

Jon




-------------------------------------------------------
This SF.Net email is sponsored by: YOU BE THE JUDGE. Be one of 170
Project Admins to receive an Apple iPod Mini FREE for your judgement on
who ports your project to Linux PPC the best. Sponsored by IBM.
Deadline: Sept. 24. Go here: http://sf.net/ppc_contest.php