This is the mail archive of the
gcc-help@gcc.gnu.org
mailing list for the GCC project.
Re: Strange optimization results
Ian Lance Taylor <iant@google.com> writes:
> Дмитрий Оксенчук <oksenchuk89@gmail.com> writes:
>
>> I have tested performance of std::string compare methods and get
>> strange results. With -O2 optimization std::string::compare 30 times
>> faster then operator==. With -Os optimization operator== 5 times
>> faster then std::string::compare. Is this bug or feature?
>>
>> Full table (time for 10000000 comparisons in seconds):
>> -O0 -O1 -O2 -O3 -Os
>> compare 1.069 1.057 0.029 0.025 1.072
>> operator== 0.461 0.990 0.980 1.022 0.201
>>
>> g++ (Debian 4.3.4-6) 4.3.4. Test code attached.
>
> It's pretty hard to tell whether this is a bug or a feature in and of
> itself. -Os optimizes for code size. -O2 optimizes for runtime
> performance. I don't see any particular reason to expect
> std::string::compare or operator== to be faster or slower with either
> one.
>
> If you showed that -Os was faster than -O2, or that -O2 was smaller than
> -Os, then I would say that you have found a bug. However, I would
> qualify that by asking whether you are using the -march and -mtune
> options appropriate for the specific CPU on which you are doing the
> testing. And I would qualify it further by saying that micro-benchmarks
> are notoriously unreliable when it comes to predicting performace in
> real code. They are not always wrong, but they often are.
Second row of results table shows that -O0 was faster than -O2 and -Os
was faster than -O0. I'm not using -march and -mtune options, version
of g++ is 4.3.4 from Debian stable, CPU Intel Core2 Duo T8100.