[pypy-dev] Pypy jit and (meta) genetic algorithms

[Armin Rigo]

Hi David,

On Tue, Sep 27, 2011 at 22:43, David Naylor <naylor.b.david at gmail.com> wrote:
> It occurred to me that with the many options available for jit (such as
> inlining, function_threshold) there may be some merit to optimising those
> values.

You are correct in that it makes sense to try more to optimize,
notably the "trace_limit" parameter, but we try to do it from time to
time, manually.  There is also the "retrace_limit" about which I am
not sure --- on obscure cases I'm sure that increasing it makes sense.
 On long-running processes, all other parameters should have a much
smaller impact (threshold, function_threshold, trace_eagerness), are
about reusing memory on large programs with different phases
(loop_longevity), or are for debugging (enable_opts, inlining).

All in all we don't really have more than this one parameter, the
trace_limit, to optimize heuristically (and one is even too much,
we're trying to think of ways to avoid it).


A bientôt,

Armin.