Re: [PATCH v4 1/5] sim: cgen: add remainder functions (needed for OR1K lf.rem.[sd])

[Stafford Horne <shorne@gmail.com>]

On Thu, Aug 31, 2017 at 11:10:47PM +0200, Simon Marchi wrote:
> On 2017-05-29 16:47, Stafford Horne wrote:
> > diff --git a/sim/common/sim-fpu.c b/sim/common/sim-fpu.c
> > index 0d4d08a..1a79e71 100644
> > --- a/sim/common/sim-fpu.c
> > +++ b/sim/common/sim-fpu.c
> > @@ -41,6 +41,7 @@ along with this program.  If not, see
> > <http://www.gnu.org/licenses/>.  */
> >  #include "sim-io.h"
> >  #include "sim-assert.h"
> > 
> > +#include <math.h> /* for drem, remove when soft-float version is
> > implemented */
> > 
> >  /* Debugging support.
> >     If digits is -1, then print all digits.  */
> > @@ -1551,6 +1552,68 @@ sim_fpu_div (sim_fpu *f,
> > 
> > 
> >  INLINE_SIM_FPU (int)
> > +sim_fpu_rem (sim_fpu *f,
> > +	     const sim_fpu *l,
> > +	     const sim_fpu *r)
> > +{
> > +  if (sim_fpu_is_snan (l))
> > +    {
> > +      *f = *l;
> > +      f->class = sim_fpu_class_qnan;
> > +      return sim_fpu_status_invalid_snan;
> > +    }
> > +  if (sim_fpu_is_snan (r))
> > +    {
> > +      *f = *r;
> > +      f->class = sim_fpu_class_qnan;
> > +      return sim_fpu_status_invalid_snan;
> > +    }
> > +  if (sim_fpu_is_qnan (l))
> > +    {
> > +      *f = *l;
> > +      f->class = sim_fpu_class_qnan;
> > +      return 0;
> > +    }
> > +  if (sim_fpu_is_qnan (r))
> > +    {
> > +      *f = *r;
> > +      f->class = sim_fpu_class_qnan;
> > +      return 0;
> > +    }
> > +  if (sim_fpu_is_infinity (l))
> > +    {
> > +      *f = sim_fpu_qnan;
> > +      return sim_fpu_status_invalid_irx;
> > +    }
> > +  if (sim_fpu_is_zero (r))
> > +    {
> > +      *f = sim_fpu_qnan;
> > +      return sim_fpu_status_invalid_div0;
> > +    }
> > +  if (sim_fpu_is_zero (l))
> > +    {
> > +      *f = *l;
> > +      return 0;
> > +    }
> > +  if (sim_fpu_is_infinity (r))
> > +    {
> > +      *f = *l;
> > +      return 0;
> > +    }
> > +  {
> > +    /* TODO: Implement remainder here.  */
> > +
> > +    sim_fpu_map lval, rval, fval;
> > +    lval.i = pack_fpu(l, 1);
> > +    rval.i = pack_fpu(r, 1);
> > +    fval.d = remainder(lval.d, rval.d);
> > +    unpack_fpu(f, fval.i, 1);
> > +    return 0;
> > +  }
> 
> I can't tell for sure because I'm not maintainer of sim/, but I suppose that
> we would need a proper implementation that doesn't use the host fpu here.

Right, as mentioned in the summary, this is the one place that is a bit
controversial.

I was thinking its kind of strange to not allow using libmath, since
integer math runs on the host system, why not FPU as well?
(probably to implement this I would just copy from libmath in the end:)

  https://sourceware.org/git/?p=glibc.git;a=blob_plain;f=sysdeps/ieee754/dbl-64/e_remainder.c;hb=HEAD

There are actually no OpenRISC cores that implement the remainder
instruction (as its a bit complicated to do in hardware and not really used
much). I could remove it if the implementation is beyond the scope of this
series.

> > +}
> > +
> > +
> > +INLINE_SIM_FPU (int)
> >  sim_fpu_max (sim_fpu *f,
> >  	     const sim_fpu *l,
> >  	     const sim_fpu *r)
> > diff --git a/sim/common/sim-fpu.h b/sim/common/sim-fpu.h
> > index d27d80a..c108f1f 100644
> > --- a/sim/common/sim-fpu.h
> > +++ b/sim/common/sim-fpu.h
> > @@ -151,6 +151,7 @@ typedef enum
> >    sim_fpu_status_overflow = 4096,
> >    sim_fpu_status_underflow = 8192,
> >    sim_fpu_status_denorm = 16384,
> > +  sim_fpu_status_invalid_irx = 32768, /* (inf % X) */
> >  } sim_fpu_status;
> 
> I think it would make sense to put the new entry with the other "invalid"
> ones and shift the others.

OK.

Thanks for the review and picking this up.

-Stafford

> 
> Simon