On Sat, Oct 23, 2010 at 3:44 AM, Altair Sythos <syt...@gmail.com> wrote: > On Sat, 23 Oct 2010 03:19:24 -0700 > leliel <leliel.mir...@gmail.com> wrote: > > >> >> > the other (imho better) way to select right march/mtune is use >> > "generic" and declare *all* parameters, sort of: >> >> This is the proper fix imho. >> >> > for 32bit: >> > gcc -march=generic -mtune=generic -m32 -mmmx -msse -msse2 >> > -mfpmath=both <etc.etc.> >> > >> > for 64bit: >> > gcc -march=generic -mtune=generic -m64 -mmmx -msse -msse2 >> > -mfpmath=both <etc.etc.> >> >> Should use -mfpmath=sse, gcc isn't very good at doing both x87 and sse >> fp math at the same time. > > let GCC to choose the faster way to assembly machine code... (done > under GCC4.5 a lot of test in KV enviroment, but on intel up)
SSE2 is faster, mfpmath=both tells gcc to use x87 & SSE at the same time which is not an easy task given how they work completely different. And it may not even give any speed up, many CPUs share execution hardware between x87 & SSE. > is time for a libs update too, especially on multimedia (opeonal, > pulse, SSL0.9.7 is buged, etc.etc.), as change compilefarm upgrading > GCC (mainly now, mesh need high performance compilated code) The first step towards that is to stop using the archaic and down right weird x87 ABI. _______________________________________________ Policies and (un)subscribe information available here: http://wiki.secondlife.com/wiki/OpenSource-Dev Please read the policies before posting to keep unmoderated posting privileges