[LUGOS-PROG] intlov compiler

[Janez Trenz]

Pero wrote:
> 
> Tole me je tolk zanimal da sem sel zadevo instalirat in skompajlat BladeEnc
> (mp3 encoder)
> 
> tule so rezulati(gcc je bil nastavljen kot ga je nastavil configure, za icc
> sem samo zamenjal gcc z icc v makefilih), cas sem meril z time
> 
> wav 44.1kHz 16bit stereo 4:20m -> mp3 128kbps
> 
> icc      gcc
> 26.058s  37.572a
> 
> mislim da je razlika dovolj ocitna. bladeend skompajlan z icc je zadevo
> opravil v manj kot 70% casa ki ga je porabil bladeenc skompajlan z gcc
> 
> ko bom imel cas bom poskusil se qt in kde mal pospest, al pa avifile
> 
> Pri intelu pravijo da se kernela ne da ker icc  ne podpira gcc inline
> assambly. Module bi se pa verjetno dalo, ce bo kdo to poskusil.

gcc pod verzijo 3.0 ima probleme z fpu. zadeve (double predvsem) na
stacku niso "poravnane"
Na comp.sys.intel je bil cel thread o "pathetic gcc performance".

==============from Daniel Spĺngberg===================================
I have had some alignment problems using gcc in my simulation code.
I inserted the following code into main before calling any other
routines to resolve the issue:

#if defined(__GNUC__) && defined(__i386__) && !defined(NO_X86_ASM)
#ifdef USE_X86_ALIGN_HACK
  void *p;
  int enabled_x86_align_hack=0;
  __asm__ __volatile__ (" movl %%esp,%0" : "=m" (p) : );
  if (((unsigned long)p) & 0x7)
  {
    __asm__ __volatile__ (" subl $4,%%esp " : :);
    enabled_x86_align_hack=1;
  }
#endif
#endif

This improved the speed of the core routines from 500 MFlops to 800
MFlops on an 1 GHz Athlon TB.
======================================================================

-- 
Janez Trenz  AKA  der_Alte               o      _     _         _
janez.trenz at bigfoot.com  __o     __o    /\_   _ \\o  (_)\__/o  (_)
ICQ #4215165           _`\<,   _`\<,   _>(_) (_)/<_    \_| \   _|/' \/
GNU/Linux zealot ;-)  (_)/(_) (_)/(_) (_)        (_)   (_)    (_)'  _\o_