[macports-ports] branch master updated: Add optimizations variant to python{27, 35, 36, 37}

[Perry E. Metzger]

On Tue, 17 Jul 2018 00:15:43 +0100 Chris Jones
<jonesc at hep.phy.cam.ac.uk> wrote:
> > On 16 Jul 2018, at 1:00 am, Perry E. Metzger
> > <pmetzger at macports.org> wrote:
> > 
> > I want to be clear about something. In theory, we could turn on
> > runtime trace guided optimization for _every_ binary. Python isn't
> > particularly special on this. It would also, in my opinion, be a
> > horrible mistake -- the benefits are fairly small compared to the
> > astonishing build times. There are people for whom this is
> > important, but they're not normal users.  
> 
> Its also important for the buildbots. We don’t, I suspect, have the
> spare resources available there to handle suddenly having all ports
> take an order of magnitude longer to build. This is important. For
> instance when a large number if builds come in at once. It
> currently already takes weeks to build all ports once a new OS is
> released. Increasing this time by an order of magnitude is probably
> not acceptable.
> 
> I also suspect the vast majority of ports will not gain from this
> by any meaningful, real world, degree. Do we, for instance, have
> firm numbers on what this brings to the python port ?

It's a modest improvement. I can't remember the exact numbers, but for
some programs it's nothing, for some benchmarks it's around 10-12%? I
imagine it has a lot to do with what your workload looks like and
whether it matches the tests that get profiled.

In truth, if you have some really large Python load, your best bet is
probably to use numba or something like that.

Perry
-- 
Perry E. Metzger		pmetzger at macports.org