Re: Fix IBM long double division inaccuracy (glibc bug 15396)

[David Edelsohn <dje dot gcc at gmail dot com>]

On Thu, Jan 2, 2014 at 4:46 PM, Joseph S. Myers <joseph@codesourcery.com> wrote:
> This patch fixes inaccuracy of IBM long double division that causes
> large errors in glibc testing
> <https://sourceware.org/bugzilla/show_bug.cgi?id=15396>.  The
> implementation computes a first approximation to the result as t = a /
> c, dividing the high parts of the arguments, then does adjustments
> involving calculating c * t as the exact sum of two double values.  In
> the problem cases, the low part of c * t underflows, so resulting in
> loss of precision compared to the precision available for a number
> with the exponent of the final result.  This patch scales the
> arguments up in the problem case.
>
> Tested with no regressions for cross to powerpc-linux-gnu.  OK to
> commit?
>
> (Note that there remain other bugs in the IBM long double code, some
> causing glibc test failures, at least (a) invalid results in rounding
> modes other than FE_TONEAREST, (b) spurious overflow and underflow
> exceptions, mainly but not entirely where discontiguous mantissa bits
> are involved.)
>
> libgcc:
> 2014-01-02  Joseph Myers  <joseph@codesourcery.com>
>
>         * config/rs6000/ibm-ldouble.c (__gcc_qdiv): Scale up arguments in
>         case of small numerator and finite nonzero result.
>
> gcc/testsuite:
> 2014-01-02  Joseph Myers  <joseph@codesourcery.com>
>
>         * gcc.target/powerpc/rs6000-ldouble-3.c: New test.

Okay.

Thanks, David