[Python-Dev] Micro-optimizations by adding special-case bytecodes?

[Ben Hoyt]

Thanks for the pointers -- I wasn't aware of perf timeit. I'll try to
get some more stable benchmark results. -Ben

On Wed, Jun 28, 2017 at 10:20 AM, Victor Stinner
<victor.stinner at gmail.com> wrote:
> (Victor wears his benchmark hat.)
>
> 2017-06-28 15:50 GMT+02:00 Ben Hoyt <benhoyt at gmail.com>:
>> ../test_none.sh
>> x = 1234; x is None -- 20000000 loops, best of 5: 19.8 nsec per loop
>> x = 1234; x is not None -- 10000000 loops, best of 5: 20 nsec per loop
>> x = None; x is None -- 10000000 loops, best of 5: 20.7 nsec per loop
>> x = None; x is not None -- 10000000 loops, best of 5: 20.8 nsec per loop
>> avg 20.3 nsec per loop
>
> Hum, please use perf timeit instead of timeit, it's more reliable. See also:
>
> "How to get reproductible benchmark results"
> http://perf.readthedocs.io/en/latest/run_benchmark.html#how-to-get-reproductible-benchmark-results
>
>> [2] Benchmarks comparing master and is_none_bytecode patch (each
>> compiled with --enable-optimizations) using python/performance:
>>
>> +-------------------------+------------+------------------------------+
>> | Benchmark               | master_opt | is_none_bytecode_opt         |
>> +=========================+============+==============================+
>> | 2to3                    | 617 ms     | 541 ms: 1.14x faster (-12%)  |
>> +-------------------------+------------+------------------------------+
>> | chameleon               | 19.9 ms    | 18.6 ms: 1.07x faster (-7%)  |
>> +-------------------------+------------+------------------------------+
>> | crypto_pyaes            | 208 ms     | 201 ms: 1.04x faster (-3%)   |
>> +-------------------------+------------+------------------------------+
>> | deltablue               | 13.8 ms    | 12.9 ms: 1.07x faster (-7%)  |
>
> FYI you can add the -G option to perf compare_to to sort results by
> speed (group faster & slower). It gives a more readable table. I also
> like using --min-speed=5 to ignore changes smaller than 5%, it reduces
> the noise and makes the stable more readable.
>
> Victor