Re: PowerPC floating point little-endian [4 of 15]

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

On Wed, 10 Jul 2013, Alan Modra wrote:

> Another batch of ieee854 macros and union replacement.  These four
> files also have bugs fixed with this patch.  The fact that the two
> doubles in an IBM long double may have different signs means that
> negation and absolute value operations can't just twiddle one sign bit
> as you can with ieee864 style extended double.  fmodl, remainderl,
> erfl and erfcl all had errors of this type.  erfl also returned +1 for
> large magnitude negative input where it should return -1.  The hypotl
> error is innocuous since the value adjusted twice is only used as a
> flag.

When fixing user-visible bugs, please file them in Bugzilla, and if 
possible include testcases for them (that fail before, pass after the 
patch) in libm-test.inc.

My guess is that fmod/remainder tests would be conditional on "#if defined 
TEST_LDOUBLE && LDBL_MANT_DIG >= 106"; I'm less sure about erf/erfc, but 
presumably by knowing where the error is, you should be able to make the 
incorrect negation give errors of up to around 2^52 ulps, and cases with 
large errors make suitable testcases.  The erf issue with large negative 
input is certainly easy to test - libm-test.inc is currently lacking any 
coverage of finite negative values.

>  	if(__builtin_expect(hx<=hy,0)) {
> -	    if((hx<hy)||(lx<ly)) return x;	/* |x|<|y| return x */
> +	    if (hx < hy
> +		|| (int64_t) (lx ^ sx) < (int64_t) (ly ^ sy))
> +		return x;			/* |x|<|y| return x */
>  	    if(lx==ly)
> -		return Zero[(u_int64_t)sx>>63];	/* |x|=|y| return x*0*/
> +		return Zero[(uint64_t)sx>>63];	/* |x|=|y| return x*0*/

It looks to me like the (lx==ly) conditional needs to check (lx^sx) and 
(ly^sy) instead to be correct, further illustrating the value of a range 
of new testcases.

-- 
Joseph S. Myers
joseph@codesourcery.com