On Wed, 31 Oct 2018, Rajalakshmi Srinivasaraghavan wrote:
> On 10/25/2018 09:40 PM, Joseph Myers wrote:
> > On Thu, 25 Oct 2018, Rajalakshmi Srinivasaraghavan wrote:
> >
> > > + if (__builtin_unpack_longdouble (a, 0) < TWO53)
> > > + {
> > > + /* In this case the integer portion is completely contained
> > > + within the high double. So use the hardware convert to
> > > + integer doubleword, and then extend to int. */
> > > + l1 = __builtin_unpack_longdouble (a, 0);
> > > + result = l1;
> >
> > But if the high double is a positive integer, and the low double is
> > strictly negative (not -0), you need to subtract 1 to get a result that's
> > correctly truncated towards zero.
>
> Do you suggest like this?
>
> if (__builtin_unpack_longdouble (a, 0) < TWO53)
> {
> l1 = __builtin_unpack_longdouble (a, 0);
> if (__builtin_unpack_longdouble (a, 1) < 0.0)
> result = l1 - 1;
> else
> result = l1;
> }
Yes, but there's another case to consider as well.
As I understand it, this part of the code also applies when the IBM long
double value is negative. If that value is <= -1, the code is fine
regardless of any adjustment for the low part (the result of the
conversion is unspecified, with INVALID exception). But you also need to
consider:
* High part equal to -1, and low part positive. In that case you must
produce an integer result of 0, with no INVALID exception.
* High part in open interval (-1, 0). In that case, regardless of the low
part, you must produce an integer result of 0, with no INVALID exception -
you must not adjust an intermediate result of 0 (from converting the high
part to integer) into one of -1 - so subtracting 1 based on the low part
being negative, without regard to whether the high part is also negative,
isn't correct.
I strongly advise adding tests to the testsuite covering all these cases
and the various cases in the code (some may be covered by the existing
gcc.dg/torture/fp-int-convert* tests, but those are generic rather than
aimed at the peculiarities of IBM long double).
--
Joseph S. Myers
jos...@codesourcery.com
