[pypy-dev] PGO Optimized PyPy Built

[Armin Rigo]

Hi,

On 2 August 2016 at 08:16, Patrascu, Alecsandru
<alecsandru.patrascu at intel.com> wrote:
> number of benchmarks that have a speedup and slowdown are about even, in the
> range +/-3%, as measured on our Haswell systems. These results are not
> unexpected or weird, since doing PGO/LTO on other projects, such as Node.JS
> or HHVM, which also feature a JIT, got to the same outcome.

PyPy played with profile-guided optimizations long ago, before it had
a fully-working JIT.  It gave some speed-ups, but I don't remember the
numbers (at that time they were certainly not as impressive as 26%).

With the JIT, these optimizations can still in theory help a bit for
parts that are never JITted (like ``long`` integer computations), for
the JIT warm-up itself, or for special code like the GC.  Alecsandru's
results are that it doesn't help at all, but maybe his benchmark suite
doesn't cover the kind of code where it would, like warm-up-intensive
code.  You'd need to try and measure on your own examples to be sure.
I think it can all be done by generating the C files and then tweaking
manually the Makefile.


A bientôt,

Armin.