Does profiling remove some optimisations?

[andrew cooke]

Gabriel Genellina wrote:
> En Thu, 05 Mar 2009 07:39:20 -0200, andrew cooke <andrew at acooke.org>
> escribió:
>
>> I have a core loop that is critical to performance.  The code is at
>> http://www.acooke.org/lepl/api/lepl.parser-pysrc.html#trampoline
>>
>> If I write a separate "optimised" version of that function for when
>> "monitor" is empty, with all the "if monitor" tests removed, the
>> profiler
>> (cProfile) indicates a 10% reduction in time spent in the loop.
>>
>> But if I run the same code 100 times under timeit, without profiling, I
>> see no difference in total time.  The process is CPU bound.
>
> How do you measure that? The profiler is able to isolate the time spent in
> your function, and the time spent in the remaining code. But timeit isn't.
> Perhaps the 10% reduction in trampoline() is minimized against the time
> spent in the rest of the code?

Yeah, I wasn't very clear, and when you do the maths it's borderline in
the noise from the timeit results.

Basically the argument is something like that routine was taking 1/10 of
the total time, and I'd improved it by 10% so should see a 1% change in
total speed.  Which I thought I should be able to see (by using a large
number of iterations in timeit).

But really, worrying about a 1% total speedup is silly.  I should be
looking elsewhere.  Which is good news, I guess, because having two
versions of a core routine was a bit worrying.

Cheers,
Andrew