PPC float context only 4 Bytes per register?

[Joel Sherrill]

Till Straumann wrote:
> 
> To quote from a lesson out of my first English textbook:
> 
> "What a shock for Mrs. Dent..."
> 
> I had a look at the code in question and I must agree. This simple
> test program confirms our suspicions:

I am at home and have been off a few days so this response is from
memory.  There used to be an ifdef about the size of the FPU 
registers to save/restore.  AFAIK no CPU had ever set it to
64-bits.  I was under the impression that the architecture could
have at least 3 variants for FPU support:

  + none
  + 32 bit
  + 64 bit  

If the FPU support is present or 64 bit, then the code needs to be
changed.

No big deal from an implementation standpoint, but the solution depends
on what variants are possible.

Could someone please answer this from an architectural standpoint? :)

--joel

> void fpctxt_tst(void *parm)
> {
> volatile double val;
> register double tst;
> rtems_interrupt_level l;
> int     i,j;
> char    mval[50];
> 
>         rtems_interrupt_disable(l);
>         tst = val = sqrt(2.);
>         rtems_interrupt_enable(l);
> 
>         j=0;
>         do {
>                 if (j) rtems_task_wake_after(2);
>                 j++;
>                 rtems_interrupt_disable(l);
>                 i = (tst == val);
>                 rtems_interrupt_enable(l);
>         } while( i );
> 
>         rtems_interrupt_disable(l);
>         sprintf(mval,"%.20g",val);
>         rtems_interrupt_enable(l);
> 
>         printf("FP register mismatch in %ith loop, should be %s but is
> %.20g\n",j,mval,tst);
> }
> 
> I must admit that I'm quite a bit shocked -
> thanks to Thomas for pointing this out. How come that this
> has been known for a while (see Sergej's message) but not fixed?
> IMO, this deserved at least a 'known-issues' entry or similar.
> 
> Luckily, it should be easy - is somebody working on this already?
> 
> -- Till
> 
> Thomas Doerfler wrote:
> 
> > Hi,
> >
> > I have just had some looks into the
> > libcpu/powerpc/new_exception_processing/cpu_asm.S code to
> > save/restore a floating-point context. In there it seems that
> > for every context switch all the float registers are saved
> > into 4 byte locations (in "single" precision format).
> >
> > From my point of view this results in rounding errors for
> > "double precision" (64 bit) float arithmetics. So if I have
> > written some code using "double precision" variables I will
> > get roundoff errors, if there has been a context switch during
> > calculations. (I did not try that out, but it seems logical to
> > me...)
> >
> > Did I miss something there?
> >
> > Any comments welcome...
> >
> >         Thomas Doerfler.
> > --------------------------------------------
> > IMD Ingenieurbuero fuer Microcomputertechnik
> > Thomas Doerfler           Herbststrasse 8
> > D-82178 Puchheim          Germany
> > email:    Thomas.Doerfler at imd-systems.de
> > PGP public key available at: http://www.imd-
> > systems.de/pgp_key.htm

-- 
Joel Sherrill, Ph.D.             Director of Research & Development
joel at OARcorp.com                 On-Line Applications Research
Ask me about RTEMS: a free RTOS  Huntsville AL 35805
   Support Available             (256) 722-9985