[Haifux] GCC Conundrum

[Ilya A. Volynets-Evenbakh]

Were results different (as in "the way one would expect") with earlier
versions of
gcc?

Shlomi Fish wrote:
> I'm getting strange results (and counter-intuitive ones) when trying to 
> compile a program using gcc-4.4.0-3mnb2 on Mandriva Linux Cooker.
>
> The program in question is:
>
> https://svn.berlios.de/svnroot/repos/fc-solve/trunk/fc-solve/source
>
> (you can checkout it using "svn co").
>
> And I'm getting the following benchmark timings: (in seconds)
>
> {{{{{{
> dump:126.027262926102
> dump-2:124.866523981094
> dump-3:125.149952888489
> dump-4:125.010930776596
> dump-aft-reboot:127.221956968307
> dump-flip:123.10272693634
> dump-flip-2:121.934360027313
> dump-flip-3:121.98522400856
> dump-old:130.334769010544
> dump-old-aft-reboot:122.411857128143
> }}}}}}
>
> dump-old is of an old revision before I rebooted. dump-old-aft-reboot is the 
> same version after I reboot the computer.
>
> The dump-flip's are after doing "./Tatzer -l p4b" and then "make" (and "make 
> install") and then running:
>
>  ./freecell-solver-range-parallel-solve 1 32000 500 -l gi
>
> The dump-[234] are after doing "./Tatzer -l p4b --without-flip". What this 
> flag does is define the FCS_WITHOUT_CARD_FLIPPING macro, which in turn toggles 
> some #ifndef's in the code that eliminate a lot of code. I verified that with 
> "--without-flip" binaries are smaller.
>
> My question is: why is the --without-flip binaries slower than the ones with 
> the flipping logic? So far my best bet is that with the flipping logic enabled 
> (but underused) the code aligns to have fewer cache misses and that it is just 
> a co-incidence that this is the case.
>
> I'll appreciate any insights.
>
> Regards,
>
> 	Shlomi Fish
>
>   


-- 
Ilya A. Volynets-Evenbakh
http://www.total-knowledge.com