Re: Inline round for IA64

[Geoff Keating <geoffk at geoffk dot org>]

Richard Henderson <rth@redhat.com> writes:

> As mentioned by Jim, you need to protect against overflow.
> 
> In addition, one needs to protect against the rounding mode being set
> to something other than round-to-nearest.  This, actually, is probably
> a non-starter when it comes to inlining this function.
> 
> Are you sure you want to implement round, and not rint or nearbyint?
> 
> Finally, there's very little of this that's specific to ia64.  Indeed,
> none of it.  If you were to generate this:
> 
> 	float rint(float orig)
> 	{
> 	  float two_exp_p = 1 << FLT_MANT_DIG;
> 	  float x = fabs(orig);
> 	  if (x < two_exp_p)
> 	    {
> 	      x += two_exp_p;
> 	      x -= two_exp_p;
> 	      x = copysignf (x, orig);
> 	    }
> 	  else
> 	    x = orig;
> 	  return x;
> 	}
> 
> in rtl from a function in builtins.c, every target would benefit.

... well, assuming that the real format is right.  This won't work for
the two-double 'long double' format used on ppc-aix and ppc-darwin.

But it'd be perfectly fine to use it for ieee_double_format
and ieee_single_format and so on.

-- 
- Geoffrey Keating <geoffk@geoffk.org>