Re: [PATCH] ldbl-128ibm nexttowar[d/f/l] generates spurious inexact exception

["Joseph S. Myers" <joseph at codesourcery dot com>]

On Mon, 20 May 2013, Adhemerval Zanella wrote:

> This is triggered by glibc testcase "nexttowar[d/f/l] (qNaN, 1.1)". The source
> of the issue is sysdeps/ieee754/ldbl-128ibm/s_nexttoward.c,
> sysdeps/ieee754/ldbl-128ibm/s_nextafterl.c, and sysdeps/ieee754/ldbl-128ibm/s_nexttowardf.c
> all return 'x+y' as return if x or y is a NaN.
> 
> This fix is to return __builtin_nan[f|l]  instead. Ok to commit?

No, that's wrong.  x+y for x or y a NaN should never raise "inexact"; if 
it does, that's a bug in the libgcc implementation of addition.  Your fix 
would lose the NaN payload and would fail to raise "invalid" if the NaN is 
signaling.

There are, it seems, at least three libgcc problems needing fixing for IBM 
long double to avoid failures in glibc tests:

* This one about spurious exceptions with NaNs.

* Operations on long doubles that are in the normal range, but with 
discontiguous mantissa bits so the low parts of the long doubles are 
subnormal doubles, can raise spurious "underflow" exceptions even when the 
final result they return is within the normal long double range.

* Division is inaccurate when subnormal values are involved (see 
<http://sourceware.org/bugzilla/show_bug.cgi?id=15396>).

It would of course be desirable to get those fixes into all active GCC 
release branches so they are more likely to be available for glibc builds.

Care will of course be needed to minimise performance impact of fixes for 
the common case where everything involved is a normal finite double.

In directed rounding modes there is a fourth issue: overflow in 
FE_TOWARDZERO mode (at least) can produce an invalid long double value; 
see <http://sourceware.org/bugzilla/show_bug.cgi?id=13304#c38>.

-- 
Joseph S. Myers
joseph@codesourcery.com