Re: PowerPC: Update long double IEEE 128-bit tests.

[Michael Meissner <meissner@linux.ibm.com>]

On Mon, Nov 02, 2020 at 07:00:15PM -0600, Segher Boessenkool wrote:
> Hi!
> 
> On Thu, Oct 22, 2020 at 06:07:14PM -0400, Michael Meissner wrote:
> > This patch fixes 3 tests in the testsuite that fail if long double is set
> > to IEEE 128-bit.
> 
> > 	* c-c++-common/dfp/convert-bfp-11.c: If long double is IEEE
> > 	128-bit, skip the test.
> 
> > --- a/gcc/testsuite/c-c++-common/dfp/convert-bfp-11.c
> > +++ b/gcc/testsuite/c-c++-common/dfp/convert-bfp-11.c
> > @@ -5,6 +5,7 @@
> >     Don't force 128-bit long doubles because runtime support depends
> >     on glibc.  */
> >  
> > +#include <float.h>
> >  #include "convert.h"
> >  
> >  volatile _Decimal32 sd;
> > @@ -39,6 +40,12 @@ main ()
> >    if (sizeof (long double) != 16)
> >      return 0;
> >  
> > +  /* This test is written to test IBM extended double, which is a pair of
> > +     doubles.  If long double can hold a larger value than a double can, such
> > +     as when long double is IEEE 128-bit, just exit immediately.  */
> 
> A double-double can hold bigger values than a double can, as well
> (if X is the biggest double, then X+Y is a valid double-double whenever
> you take Y small enough).
> 
> > +  if (LDBL_MAX_10_EXP > DBL_MAX_10_EXP)
> > +    return 0;

Yes a double-double can hold more mantissa bits than a double, but the exponent
size is the same (which is what I'm testing).

> This is testing something different though: whether the base-10
> logarithm of the maximum finite double is different from that of the
> maximum finite double-double.
> 
> Is there no more direct test you can do?  Just test __FLOAT128__ maybe?
> The test is not even compiled if not powerpc*-linux, so you can test
> such macros just fine.

I will have to look at it.

> > 	* gcc.dg/nextafter-2.c: On PowerPC, if long double is IEEE
> > 	128-bit, include math.h to get the built-in mapped correctly.
> 
> > diff --git a/gcc/testsuite/gcc.dg/nextafter-2.c b/gcc/testsuite/gcc.dg/nextafter-2.c
> > index e51ae94be0c..64e9e3c485f 100644
> > --- a/gcc/testsuite/gcc.dg/nextafter-2.c
> > +++ b/gcc/testsuite/gcc.dg/nextafter-2.c
> > @@ -13,4 +13,14 @@
> >  #  define NO_LONG_DOUBLE 1
> >  # endif
> >  #endif
> > +
> > +#if defined(_ARCH_PPC) && defined(__LONG_DOUBLE_IEEE128__)
> > +/* On PowerPC systems, long double uses either the IBM long double format, or
> > +   IEEE 128-bit format.  The compiler switches the long double built-in
> > +   function names and glibc switches the names when math.h is included.
> > +   Because this test is run with -fno-builtin, include math.h so that the
> > +   appropriate nextafter functions are called.  */
> > +#include <math.h>
> > +#endif
> > +
> >  #include "nextafter-1.c"
> 
> Please explain *what* mappings are made?  And why is it okay to do this
> in the testsuite, when all "normal" code (that does not do this) will
> just fail?

I can put in a better comment.  However, this test fails because it explicitly
does not include math.h and it uses -fno-builtin.  So the compiler can't
effectively map the nextafter math function.


> > 	* gcc.target/powerpc/pr70117.c: Add support for long double being
> > 	IEEE 128-bit.
> 
> That is not what the patch does -- it instead changes the code because
> it does not work correctly with long double ieee128 (which it already
> did claim to support!)
> 
> So what are the actual changes doing, why are they correct, why was the
> original not correct?
> 
> (It is easy to make a test not fail anymore: just delete it!  Something
> here should be better than that :-) )

I will have to look into it.

-- 
Michael Meissner, IBM
IBM, M/S 2506R, 550 King Street, Littleton, MA 01460-6245, USA
email: meissner@linux.ibm.com, phone: +1 (978) 899-4797