[PowerPC64] Correct IBM long double nextafterl
Commit Message
Fix for values near a power of two, and some tidies.
The added tests do cause some new fails, because IBM long double isn't
supposed to support rounding modes other than round-to-nearest.
I believe powerpc should have a math-tests.h like arm and mips that
limits the tested rounding modes.
[BZ #16739]
* sysdeps/ieee754/ldbl-128ibm/s_nextafterl.c (__nextafterl): Correct
output when value is near a power of two. Use int64_t for lx and
remove casts. Use decimal rather than hex exponent constants.
Don't use long double multiplication when double will suffice.
* math/libm-test.inc (nextafter_test_data): Add tests.
Comments
On Tue, 25 Mar 2014, Alan Modra wrote:
> + TEST_ff_f (nextafter, 1.0L, -10.0L, 1.0L-0x1p-106L, NO_EXCEPTION),
> + TEST_ff_f (nextafter, 1.0L, 10.0L, 1.0L+0x1p-105L, NO_EXCEPTION),
> + TEST_ff_f (nextafter, 1.0L-0x1p-106L, 10.0L, 1.0L, NO_EXCEPTION),
> + TEST_ff_f (nextafter, -1.0L, -10.0L, -1.0L-0x1p-105L, NO_EXCEPTION),
> + TEST_ff_f (nextafter, -1.0L, 10.0L, -1.0L+0x1p-106L, NO_EXCEPTION),
> + TEST_ff_f (nextafter, -1.0L+0x1p-106L, -10.0L, -1.0L, NO_EXCEPTION),
You mean NO_INEXACT_EXCEPTION (NO_EXCEPTION makes no assertion about
whether or not "inexact" is present; "inexact" has a different default
from the other exceptions, following ISO C Annex F).
On Tue, Mar 25, 2014 at 12:13:08PM +0000, Joseph S. Myers wrote:
> On Tue, 25 Mar 2014, Alan Modra wrote:
>
> > + TEST_ff_f (nextafter, 1.0L, -10.0L, 1.0L-0x1p-106L, NO_EXCEPTION),
> > + TEST_ff_f (nextafter, 1.0L, 10.0L, 1.0L+0x1p-105L, NO_EXCEPTION),
> > + TEST_ff_f (nextafter, 1.0L-0x1p-106L, 10.0L, 1.0L, NO_EXCEPTION),
> > + TEST_ff_f (nextafter, -1.0L, -10.0L, -1.0L-0x1p-105L, NO_EXCEPTION),
> > + TEST_ff_f (nextafter, -1.0L, 10.0L, -1.0L+0x1p-106L, NO_EXCEPTION),
> > + TEST_ff_f (nextafter, -1.0L+0x1p-106L, -10.0L, -1.0L, NO_EXCEPTION),
>
> You mean NO_INEXACT_EXCEPTION (NO_EXCEPTION makes no assertion about
> whether or not "inexact" is present; "inexact" has a different default
> from the other exceptions, following ISO C Annex F).
I deliberately did not choose NO_INEXACT_EXCEPTION, because nextafterl
on IBM long double *does* set FE_INEXACT on some of these tests. The
reason is the "x + u" or "x - u" operations inside nextafterl which
use __gcc_qadd. As you know, __gcc_qadd spuriously sets FE_INEXACT.
Also, as the description of IBM long double in the PowerPC64 ABI
states, the "Extended precision" format
* Does not support the IEEE status flags for overflow, underflow, and
other conditions. These flag have no meaning in this format.
On 25-03-2014 10:15, Alan Modra wrote:
> On Tue, Mar 25, 2014 at 12:13:08PM +0000, Joseph S. Myers wrote:
>> On Tue, 25 Mar 2014, Alan Modra wrote:
>>
>>> + TEST_ff_f (nextafter, 1.0L, -10.0L, 1.0L-0x1p-106L, NO_EXCEPTION),
>>> + TEST_ff_f (nextafter, 1.0L, 10.0L, 1.0L+0x1p-105L, NO_EXCEPTION),
>>> + TEST_ff_f (nextafter, 1.0L-0x1p-106L, 10.0L, 1.0L, NO_EXCEPTION),
>>> + TEST_ff_f (nextafter, -1.0L, -10.0L, -1.0L-0x1p-105L, NO_EXCEPTION),
>>> + TEST_ff_f (nextafter, -1.0L, 10.0L, -1.0L+0x1p-106L, NO_EXCEPTION),
>>> + TEST_ff_f (nextafter, -1.0L+0x1p-106L, -10.0L, -1.0L, NO_EXCEPTION),
>> You mean NO_INEXACT_EXCEPTION (NO_EXCEPTION makes no assertion about
>> whether or not "inexact" is present; "inexact" has a different default
>> from the other exceptions, following ISO C Annex F).
> I deliberately did not choose NO_INEXACT_EXCEPTION, because nextafterl
> on IBM long double *does* set FE_INEXACT on some of these tests. The
> reason is the "x + u" or "x - u" operations inside nextafterl which
> use __gcc_qadd. As you know, __gcc_qadd spuriously sets FE_INEXACT.
> Also, as the description of IBM long double in the PowerPC64 ABI
> states, the "Extended precision" format
>
> * Does not support the IEEE status flags for overflow, underflow, and
> other conditions. These flag have no meaning in this format.
>
If Joseph any more concerns, I'm ok with this patch.
On Thu, 27 Mar 2014, Adhemerval Zanella wrote:
> If Joseph any more concerns, I'm ok with this patch.
I have no further concerns with this patch.
@@ -8312,6 +8312,14 @@ static const struct test_ff_f_data nextafter_test_data[] =
// XXX Enable once gcc is fixed.
//TEST_ff_f (nextafter, 0x0.00000040000000000000p-16385L, -0.1L, 0x0.0000003ffffffff00000p-16385L),
#endif
+#if defined TEST_LDOUBLE && LDBL_MANT_DIG == 106
+ TEST_ff_f (nextafter, 1.0L, -10.0L, 1.0L-0x1p-106L, NO_EXCEPTION),
+ TEST_ff_f (nextafter, 1.0L, 10.0L, 1.0L+0x1p-105L, NO_EXCEPTION),
+ TEST_ff_f (nextafter, 1.0L-0x1p-106L, 10.0L, 1.0L, NO_EXCEPTION),
+ TEST_ff_f (nextafter, -1.0L, -10.0L, -1.0L-0x1p-105L, NO_EXCEPTION),
+ TEST_ff_f (nextafter, -1.0L, 10.0L, -1.0L+0x1p-106L, NO_EXCEPTION),
+ TEST_ff_f (nextafter, -1.0L+0x1p-106L, -10.0L, -1.0L, NO_EXCEPTION),
+#endif
/* XXX We need the hexadecimal FP number representation here for further
tests. */
@@ -30,8 +30,7 @@ static char rcsid[] = "$NetBSD: $";
long double __nextafterl(long double x, long double y)
{
- int64_t hx,hy,ihx,ihy;
- uint64_t lx;
+ int64_t hx, hy, ihx, ihy, lx;
double xhi, xlo, yhi;
ldbl_unpack (x, &xhi, &xlo);
@@ -79,19 +78,28 @@ long double __nextafterl(long double x, long double y)
u = math_opt_barrier (x);
x -= __LDBL_DENORM_MIN__;
if (ihx < 0x0360000000000000LL
- || (hx > 0 && (int64_t) lx <= 0)
- || (hx < 0 && (int64_t) lx > 1)) {
+ || (hx > 0 && lx <= 0)
+ || (hx < 0 && lx > 1)) {
u = u * u;
math_force_eval (u); /* raise underflow flag */
}
return x;
}
- if (ihx < 0x06a0000000000000LL) { /* ulp will denormal */
- INSERT_WORDS64 (yhi, hx & (0x7ffLL<<52));
- u = yhi;
- u *= 0x1.0000000000000p-105L;
+ /* If the high double is an exact power of two and the low
+ double is the opposite sign, then 1ulp is one less than
+ what we might determine from the high double. Similarly
+ if X is an exact power of two, and positive, because
+ making it a little smaller will result in the exponent
+ decreasing by one and normalisation of the mantissa. */
+ if ((hx & 0x000fffffffffffffLL) == 0
+ && ((lx != 0 && (hx ^ lx) < 0)
+ || (lx == 0 && hx >= 0)))
+ ihx -= 1LL << 52;
+ if (ihx < (106LL << 52)) { /* ulp will denormal */
+ INSERT_WORDS64 (yhi, ihx & (0x7ffLL<<52));
+ u = yhi * 0x1p-105;
} else {
- INSERT_WORDS64 (yhi, (hx & (0x7ffLL<<52))-(0x069LL<<52));
+ INSERT_WORDS64 (yhi, (ihx & (0x7ffLL<<52))-(105LL<<52));
u = yhi;
}
return x - u;
@@ -109,8 +117,8 @@ long double __nextafterl(long double x, long double y)
u = math_opt_barrier (x);
x += __LDBL_DENORM_MIN__;
if (ihx < 0x0360000000000000LL
- || (hx > 0 && (int64_t) lx < 0 && lx != 0x8000000000000001LL)
- || (hx < 0 && (int64_t) lx >= 0)) {
+ || (hx > 0 && lx < 0 && lx != 0x8000000000000001LL)
+ || (hx < 0 && lx >= 0)) {
u = u * u;
math_force_eval (u); /* raise underflow flag */
}
@@ -118,12 +126,21 @@ long double __nextafterl(long double x, long double y)
x = -0.0L;
return x;
}
- if (ihx < 0x06a0000000000000LL) { /* ulp will denormal */
- INSERT_WORDS64 (yhi, hx & (0x7ffLL<<52));
- u = yhi;
- u *= 0x1.0000000000000p-105L;
+ /* If the high double is an exact power of two and the low
+ double is the opposite sign, then 1ulp is one less than
+ what we might determine from the high double. Similarly
+ if X is an exact power of two, and negative, because
+ making it a little larger will result in the exponent
+ decreasing by one and normalisation of the mantissa. */
+ if ((hx & 0x000fffffffffffffLL) == 0
+ && ((lx != 0 && (hx ^ lx) < 0)
+ || (lx == 0 && hx < 0)))
+ ihx -= 1LL << 52;
+ if (ihx < (106LL << 52)) { /* ulp will denormal */
+ INSERT_WORDS64 (yhi, ihx & (0x7ffLL<<52));
+ u = yhi * 0x1p-105;
} else {
- INSERT_WORDS64 (yhi, (hx & (0x7ffLL<<52))-(0x069LL<<52));
+ INSERT_WORDS64 (yhi, (ihx & (0x7ffLL<<52))-(105LL<<52));
u = yhi;
}
return x + u;