speed of mpz_odd_p and lazy allocation

[Marc Glisse]

On Tue, 21 Aug 2012, Marc Glisse wrote:

> I guess I should simulate the gains for std::vector, since those would 
> also apply to gmpxx and not just external wrappers.

I made a very quick test. All I did was change the C++ wrapper so that 
when I move-construct from a value, instead of calling mpz_init for the 
moved-from object, I just set _mp_d to 0 (I should have used _mp_alloc 
according to the conversation, but that's just a test). And in the 
destructor, I call mpz_clear only if _mp_d!=0. Clearly this modification 
is incomplete, it is just to get a feeling for potential gain.

I create a vector<mpz_class> and append a 0 (resize) to it a few million 
times. With -std=c++11 and the current code, I get a time of 1. With 
-std=c++03, I get 1.5. And with the small patch described above, I get 
0.5.

I am a bit surprised that c++11 helps so much even with an old gmpxx.h 
without move constructors/assignment. It seems to be because libstdc++ 
copies the 0s in c++03 and constructs them in place in C++11.

The factor 2 is probably never reached on realistic code, but the 
difference should still be noticable in real code using small numbers.

Note that by default, C++11 assumes that destructors don't throw. If we 
want to allow users to throw from their deallocation functions (why?), we 
might want to specify it (which disables the optimization discussed here).

-- 
Marc Glisse