The HLL Performance Gap

[Andrew Whitworth]

This is quite excellent analysis and a good summation of important points.

On Tue, Jul 21, 2009 at 7:31 PM, Geoffrey Broadwell<geoff at broadwell.org> wrote:
> 1. Add past_optimize and post_optimize passes to PCT.  Start with just a
> single useful optimization in each.
>
> 2. Make it easy to plug in additional optimizations in a way that allows
> people to work in parallel on different optimizations without blocking
> on each other unduly.

Last summer this was actually a suggested GSOC project, although I
don't think there were any takers for it. So, there is some prior
thought put into it.

Good optimization steps are orthogonal and non-interacting. Any well
designed optimization system should make it perfectly possible for
people to develop stages in parallel and plug them in arbitrarily.

> 5. Analyze and micro-optimize Rakudo's per-scope boilerplate.  Focus
> first on optimizations that other HLLs may be able to use for their own
> particular boilerplate hell.

A lot about the Rakudo boiler plate could definitely be optimized,
although I want to be careful that we worry about getting Rakudo more
feature complete before we start aggressive optimization.

That said, I think there are a lot of cases where special variables
could be created more lazily (and in some scopes not created at all if
not needed).

> 6. Reduce the overhead of PIR sub calls.  I believe Allison is already
> working on refactoring that will help with this, but I suspect there
> will be more improvement possible even after her work is merged.
>
> 7. Allow lexical scopes to be separate from PIR subs.  If sub calls
> become really cheap, this will be unnecessary.  If sub calls remain
> expensive, this should be considered.

It may worthwhile to create a separate Sub-like object that is much
lighter weight (does not pass parameters or returns for example) and
only serves as a delimiter for lexical scopes. It would be nice if
Subs became so inexpensive, although having a separate such object
would enable Subs to do more then was strictly needed for a simple
lexical scope. One possible optimization step could be inlining of
subroutines like this where scope is not necessary.

All very good information and some very good ideas. I would like to
see much of this come to fruition sooner rather then later.

--Andrew Whitworth