[Python-3000] Python 3.0a2 release
Gregory P. Smith
greg at krypto.org
Thu Nov 22 07:32:11 CET 2007
On 11/21/07, Christian Heimes <lists at cheimes.de> wrote:
> Gregory P. Smith wrote:
> > I would not hold up a compiler decision based on the mingw project. Always
> > use the latest MS compiler released at the time for a x.0 build of any
> > python. Doing otherwise costs the world a fortune in lost performance
> > (higher power consumption via lower efficiency code). mingw will
> > undoubtedly catch up on its own.
> That raises another question. Should the official releases use PGO
> (profile guided optimization) and how should the profiles be created?
> PGO is an optimization system which works in two steps. At first a PGO
> instrument build is compiled. The instrument build is linked against an
> additional library. Next one or several typical applications are run to
> gather profiling information. Finally the project is compiled against
> using the profiles to optimize the new build.
I'd say yes, it should demonstrably produce faster code (easy to test
that). I don't know what workload to suggest, pystone could be one
place to start. Possibly even just a small subset of the test suite
would be enough. The idea is to exercise the basics so it can
optimize with knowledge of actual common control flow paths and hot
spots in mind, it may not require much.
More information about the Python-3000