harmony-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Egor Pasko <egor.pa...@gmail.com>
Subject Re: [drlvm][jit][opt] HARMONY-3243 It's possible to fill array with constant much faster
Date Sat, 17 Mar 2007 15:32:19 GMT
On the 0x29C day of Apache Harmony Mikhail Fursov wrote:
> On 17 Mar 2007 12:46:38 +0300, Egor Pasko <egor.pasko@gmail.com> wrote:
> >
> > On the 0x29B day of Apache Harmony Mikhail Fursov wrote:
> > > Egor,
> > >
> > > On 16 Mar 2007 20:02:15 +0300, Egor Pasko <egor.pasko@gmail.com> wrote:
> > > >
> > > > I would love to implement versioning
> > >
> > >
> > > Since we come into agreement about arrays fill optimizations, I want to
> > ask
> > > you about versioning.
> > > One of my student asked me if it worth to start working on this
> > optimization
> > > (he wants to join it with value profiler input).
> > >
> > > Do you have any plans to implement it by yourself in the nearest month
> > or
> > > two?
> >
> > cool, I am starting to work on it
> >
> > Mikhail, you probably have some ideas/expectations on
> > versioning+vaueProfiler?
> >
> > Is it that we want to make a version of a loop according to the
> > target class prediction (from value profiler)?
> 
> The idea is to use aggresive specialization for variables to simplify (or to
> reduce strength) of operations inside of a loop.

sounds very general. No doubt, specialization in the form of (for
example) bounds checking elimination and devirtualization becomes more
applicable after versioning. But versioning should not include these
specializations. All it should do is transform loops.

Versioning by type info is a good idea. Thanks! I'm not sure if it is
better t devirtualize right from within versioning or as a separate
pass.. 

-- 
Egor Pasko


Mime
View raw message