Re: some odd problems with MD5 on Alpha

[Quant-X Alpha Linux Support <stefan sime com>]

On Sat, Jun 26, 1999 at 10:33:21PM -0700, Richard Henderson wrote:
> On Sat, Jun 26, 1999 at 12:28:44AM -0700, designer seven wrote:
> > By the way, in time tests, I gained 5-fold increase in
> > speed by including -mcpu=ev56 -Wa,-m21164a. What is so
> > special about these switches and the MD5 algorithm
> > that would make such a drastic performance increase?
> 
> Byte load/store instructions.

Hmm yes, especially byte store is much faster. 

Richard, I'm not saying that I found a bug in old egcs-1.0.3 release;
but if I replace "-mcpu=ev5" by "-mcpu=ev56" when compiling nagratv-2.10
I get a "different" runtime behavior. Here the working compile-options
and the code that breaks (the outcommented for loop breaks): 

CFLAGS= -Wall -O6 -funroll-all-loops -fomit-frame-pointer -mcpu=ev5 \
        -Wa,-m21164a

----------------------------------------------------------------------
void ClearScreen(DispInfo *d)
{
  int i,s;
  unsigned char *base;

  s=2*d->width*d->height;
  base=d->base;
#if 0
  for(i=0;i<s;i+=2) *(short *)(base+i)=0;
#else
  memset(base, 0, s);
#endif
}
----------------------------------------------------------------------

-- 
  ciao -
    Stefan Traby
    Quant-X Alpha Linux Support

   Quant-X Service &                         Ph:  (+43) 4212 90555-0
   Consulting Ges.m.b.H.                     Fax: (+43) 4212 90555-555
   http://www.quant-x.com                    Email: support@quant-x.com