Fix log (1) in round-downward mode (bug 16731)
Commit Message
According to ISO C Annex F, log (1) should be +0 in all rounding
modes, but some implementations in glibc wrongly return -0 in
round-downward mode (mapping to log1p (x - 1) is problematic because 1
- 1 is -0 in round-downward mode, and log1p (-0) is -0). This patch
fixes this. (It helps with some implementations of other functions
such as acosh, log2 and log10 that call out to log, but not enough to
enable all-rounding-modes testing for those functions without further
fixes to other implementations of them.)
Tested x86_64 and x86 and ulps updated accordingly, and did spot tests
for mips64 for the ldbl-128 fix, and i586 for the sysdeps/i386/fpu
implementations shadowed by those in sysdeps/i386/i686/fpu.
2014-03-21 Joseph Myers <joseph@codesourcery.com>
[BZ #16731]
* sysdeps/i386/fpu/e_log.S (__ieee754_log): Take absolute value
when x - 1 is zero.
* sysdeps/i386/fpu/e_logf.S (__ieee754_logf): Likewise.
* sysdeps/i386/fpu/e_logl.S (__ieee754_logl): Likewise.
* sysdeps/i386/i686/fpu/e_logl.S (__ieee754_logl): Likewise.
* sysdeps/ieee754/dbl-64/e_log.c (__ieee754_log): Return +0 when
argument is 1.
* sysdeps/ieee754/ldbl-128/e_logl.c (__ieee754_logl): Likewise.
* sysdeps/x86_64/fpu/e_logl.S: Take absolute value when x - 1 is
zero.
* math/libm-test.inc (log_test): Use ALL_RM_TEST.
* sysdeps/i386/fpu/libm-test-ulps: Update.
* sysdeps/x86_64/fpu/libm-test-ulps: Likewise.
Comments
On 03/21/2014 01:47 PM, Joseph S. Myers wrote:
> According to ISO C Annex F, log (1) should be +0 in all rounding
> modes, but some implementations in glibc wrongly return -0 in
> round-downward mode (mapping to log1p (x - 1) is problematic because 1
> - 1 is -0 in round-downward mode, and log1p (-0) is -0). This patch
> fixes this. (It helps with some implementations of other functions
> such as acosh, log2 and log10 that call out to log, but not enough to
> enable all-rounding-modes testing for those functions without further
> fixes to other implementations of them.)
>
> Tested x86_64 and x86 and ulps updated accordingly, and did spot tests
> for mips64 for the ldbl-128 fix, and i586 for the sysdeps/i386/fpu
> implementations shadowed by those in sysdeps/i386/i686/fpu.
>
> 2014-03-21 Joseph Myers <joseph@codesourcery.com>
>
> [BZ #16731]
> * sysdeps/i386/fpu/e_log.S (__ieee754_log): Take absolute value
> when x - 1 is zero.
> * sysdeps/i386/fpu/e_logf.S (__ieee754_logf): Likewise.
> * sysdeps/i386/fpu/e_logl.S (__ieee754_logl): Likewise.
> * sysdeps/i386/i686/fpu/e_logl.S (__ieee754_logl): Likewise.
> * sysdeps/ieee754/dbl-64/e_log.c (__ieee754_log): Return +0 when
> argument is 1.
> * sysdeps/ieee754/ldbl-128/e_logl.c (__ieee754_logl): Likewise.
> * sysdeps/x86_64/fpu/e_logl.S: Take absolute value when x - 1 is
> zero.
> * math/libm-test.inc (log_test): Use ALL_RM_TEST.
> * sysdeps/i386/fpu/libm-test-ulps: Update.
> * sysdeps/x86_64/fpu/libm-test-ulps: Likewise.
Looks good to me.
> diff --git a/math/libm-test.inc b/math/libm-test.inc
> index 7abc7c1..5e50f0e 100644
> --- a/math/libm-test.inc
> +++ b/math/libm-test.inc
> @@ -7786,9 +7786,7 @@ static const struct test_f_f_data log_test_data[] =
> static void
> log_test (void)
> {
> - START (log, 0);
> - RUN_TEST_LOOP_f_f (log, log_test_data, );
> - END;
> + ALL_RM_TEST (log, 0, log_test_data, RUN_TEST_LOOP_f_f, END);
OK. Nice to see another function using ALL_RM_TEST.
> }
>
>
> diff --git a/sysdeps/i386/fpu/e_log.S b/sysdeps/i386/fpu/e_log.S
> index 0877924..3fa32aa 100644
> --- a/sysdeps/i386/fpu/e_log.S
> +++ b/sysdeps/i386/fpu/e_log.S
> @@ -46,7 +46,13 @@ ENTRY(__ieee754_log)
> fnstsw // x-1 : x : log(2)
> andb $0x45, %ah
> jz 2f
> - fstp %st(1) // x-1 : log(2)
> + fxam
> + fnstsw
> + andb $0x45, %ah
> + cmpb $0x40, %ah
> + jne 5f
> + fabs // log(1) is +0 in all rounding modes.
> +5: fstp %st(1) // x-1 : log(2)
OK.
> fyl2xp1 // log(x)
> ret
>
> diff --git a/sysdeps/i386/fpu/e_logf.S b/sysdeps/i386/fpu/e_logf.S
> index 485180e..ca83d39 100644
> --- a/sysdeps/i386/fpu/e_logf.S
> +++ b/sysdeps/i386/fpu/e_logf.S
> @@ -47,7 +47,13 @@ ENTRY(__ieee754_logf)
> fnstsw // x-1 : x : log(2)
> andb $0x45, %ah
> jz 2f
> - fstp %st(1) // x-1 : log(2)
> + fxam
> + fnstsw
> + andb $0x45, %ah
> + cmpb $0x40, %ah
> + jne 5f
> + fabs // log(1) is +0 in all rounding modes.
> +5: fstp %st(1) // x-1 : log(2)
OK.
> fyl2xp1 // log(x)
> ret
>
> diff --git a/sysdeps/i386/fpu/e_logl.S b/sysdeps/i386/fpu/e_logl.S
> index d7a459a..edae1d7 100644
> --- a/sysdeps/i386/fpu/e_logl.S
> +++ b/sysdeps/i386/fpu/e_logl.S
> @@ -47,7 +47,13 @@ ENTRY(__ieee754_logl)
> fnstsw // x-1 : x : log(2)
> andb $0x45, %ah
> jz 2f
> - fstp %st(1) // x-1 : log(2)
> + fxam
> + fnstsw
> + andb $0x45, %ah
> + cmpb $0x40, %ah
> + jne 5f
> + fabs // log(1) is +0 in all rounding modes.
> +5: fstp %st(1) // x-1 : log(2)
OK.
> fyl2xp1 // log(x)
> ret
>
> diff --git a/sysdeps/i386/fpu/libm-test-ulps b/sysdeps/i386/fpu/libm-test-ulps
> index 3be1806..b7b2e12 100644
> --- a/sysdeps/i386/fpu/libm-test-ulps
> +++ b/sysdeps/i386/fpu/libm-test-ulps
> @@ -1096,6 +1096,18 @@ Function: "log1p":
> ildouble: 1
> ldouble: 1
>
> +Function: "log_downward":
> +ildouble: 1
> +ldouble: 1
> +
> +Function: "log_towardzero":
> +ildouble: 1
> +ldouble: 1
> +
> +Function: "log_upward":
> +ildouble: 1
> +ldouble: 1
> +
OK.
> Function: "pow":
> ildouble: 1
> ldouble: 1
> diff --git a/sysdeps/i386/i686/fpu/e_logl.S b/sysdeps/i386/i686/fpu/e_logl.S
> index 8a86222..53a3d10 100644
> --- a/sysdeps/i386/i686/fpu/e_logl.S
> +++ b/sysdeps/i386/i686/fpu/e_logl.S
> @@ -46,7 +46,13 @@ ENTRY(__ieee754_logl)
> fcomip %st(1) // |x-1| : x-1 : x : log(2)
> fstp %st(0) // x-1 : x : log(2)
> jc 2f
> - fstp %st(1) // x-1 : log(2)
> + fxam
> + fnstsw
> + andb $0x45, %ah
> + cmpb $0x40, %ah
> + jne 4f
> + fabs // // log(1) is +0 in all rounding modes.
> +4: fstp %st(1) // x-1 : log(2)
OK.
> fyl2xp1 // log(x)
> ret
>
> diff --git a/sysdeps/ieee754/dbl-64/e_log.c b/sysdeps/ieee754/dbl-64/e_log.c
> index 05d318b..4ecd372 100644
> --- a/sysdeps/ieee754/dbl-64/e_log.c
> +++ b/sysdeps/ieee754/dbl-64/e_log.c
> @@ -96,6 +96,10 @@ __ieee754_log (double x)
> if (__glibc_likely (ABS (w) > U03))
> goto case_03;
>
> + /* log (1) is +0 in all rounding modes. */
> + if (w == 0.0)
> + return 0.0;
OK.
> +
> /*--- Stage I, the case abs(x-1) < 0.03 */
>
> t8 = MHALF * w;
> diff --git a/sysdeps/ieee754/ldbl-128/e_logl.c b/sysdeps/ieee754/ldbl-128/e_logl.c
> index 3d1034d..cb43816 100644
> --- a/sysdeps/ieee754/ldbl-128/e_logl.c
> +++ b/sysdeps/ieee754/ldbl-128/e_logl.c
> @@ -240,6 +240,8 @@ __ieee754_logl(long double x)
> /* On this interval the table is not used due to cancellation error. */
> if ((x <= 1.0078125L) && (x >= 0.9921875L))
> {
> + if (x == 1.0L)
> + return 0.0L;
OK.
> z = x - 1.0L;
> k = 64;
> t.value = 1.0L;
> diff --git a/sysdeps/x86_64/fpu/e_logl.S b/sysdeps/x86_64/fpu/e_logl.S
> index a8e3108..315afc0 100644
> --- a/sysdeps/x86_64/fpu/e_logl.S
> +++ b/sysdeps/x86_64/fpu/e_logl.S
> @@ -45,7 +45,13 @@ ENTRY(__ieee754_logl)
> fnstsw // x-1 : x : log(2)
> andb $0x45, %ah
> jz 2f
> - fstp %st(1) // x-1 : log(2)
> + fxam
> + fnstsw
> + andb $0x45, %ah
> + cmpb $0x40, %ah
> + jne 5f
> + fabs // log(1) is +0 in all rounding modes.
> +5: fstp %st(1) // x-1 : log(2)
OK.
> fyl2xp1 // log(x)
> ret
>
> diff --git a/sysdeps/x86_64/fpu/libm-test-ulps b/sysdeps/x86_64/fpu/libm-test-ulps
> index 5f4ab06..301eaa6 100644
> --- a/sysdeps/x86_64/fpu/libm-test-ulps
> +++ b/sysdeps/x86_64/fpu/libm-test-ulps
> @@ -1170,6 +1170,22 @@ ifloat: 1
> ildouble: 1
> ldouble: 1
>
> +Function: "log_downward":
> +float: 1
> +ifloat: 1
> +ildouble: 1
> +ldouble: 1
> +
> +Function: "log_towardzero":
> +ildouble: 1
> +ldouble: 1
> +
> +Function: "log_upward":
> +float: 1
> +ifloat: 1
> +ildouble: 1
> +ldouble: 1
> +
> Function: "pow":
> float: 1
> ifloat: 1
>
OK.
Cheers,
Carlos.
On Fri, Mar 21, 2014 at 05:47:45PM +0000, Joseph S. Myers wrote:
> According to ISO C Annex F, log (1) should be +0 in all rounding
> modes, but some implementations in glibc wrongly return -0 in
> round-downward mode (mapping to log1p (x - 1) is problematic because 1
> - 1 is -0 in round-downward mode, and log1p (-0) is -0). This patch
> fixes this. (It helps with some implementations of other functions
> such as acosh, log2 and log10 that call out to log, but not enough to
> enable all-rounding-modes testing for those functions without further
> fixes to other implementations of them.)
This reminds me (in context) of this patch we're carrying in Debian:
http://anonscm.debian.org/viewvc/pkg-glibc/glibc-package/trunk/debian/patches/powerpc/local-math-logb.diff?revision=5951&view=markup
If I recall, Aurelien said we're carrying it locally because it was
rejected upstream as not a bug, due to it only affecting older HW
that no one cares about.
Should we be cleaning this up and re-submitting, to go along with
your batch of fixes here?
... Adam
On Fri, 21 Mar 2014, Adam Conrad wrote:
> On Fri, Mar 21, 2014 at 05:47:45PM +0000, Joseph S. Myers wrote:
> > According to ISO C Annex F, log (1) should be +0 in all rounding
> > modes, but some implementations in glibc wrongly return -0 in
> > round-downward mode (mapping to log1p (x - 1) is problematic because 1
> > - 1 is -0 in round-downward mode, and log1p (-0) is -0). This patch
> > fixes this. (It helps with some implementations of other functions
> > such as acosh, log2 and log10 that call out to log, but not enough to
> > enable all-rounding-modes testing for those functions without further
> > fixes to other implementations of them.)
>
> This reminds me (in context) of this patch we're carrying in Debian:
>
> http://anonscm.debian.org/viewvc/pkg-glibc/glibc-package/trunk/debian/patches/powerpc/local-math-logb.diff?revision=5951&view=markup
>
> If I recall, Aurelien said we're carrying it locally because it was
> rejected upstream as not a bug, due to it only affecting older HW
> that no one cares about.
>
> Should we be cleaning this up and re-submitting, to go along with
> your batch of fixes here?
What I think should be done here is:
* Submit a bug to GCC Bugzilla about conversion from integer to floating
point being incorrect for older hard-float powerpc with -frounding-math.
* Reopen bug 887 as not in fact fixed, and remove it from the list of bugs
fixed in 2.16 in NEWS.
* Close bug 16423 as a duplicate of the newly reopened bug 887.
* Work around the GCC bug in glibc conditional on a configure test for
whether (building for powerpc and) $CC $CFLAGS $CPPFLAGS -frounding-math
generates the problem code sequence (i.e. the one involving certain magic
constants and floating-point addition / subtraction), not using fcfid, not
using any instructions to save / restore the rounding mode (the last bit
may be more difficult to test for in a future-proof way, since you don't
know what any fix for the GCC bug might look like). Make sure comments
reference the GCC bug in question.
* Having got such a workaround in glibc, bug 887 can then be closed as
fixed.
@@ -7786,9 +7786,7 @@ static const struct test_f_f_data log_test_data[] =
static void
log_test (void)
{
- START (log, 0);
- RUN_TEST_LOOP_f_f (log, log_test_data, );
- END;
+ ALL_RM_TEST (log, 0, log_test_data, RUN_TEST_LOOP_f_f, END);
}
@@ -46,7 +46,13 @@ ENTRY(__ieee754_log)
fnstsw // x-1 : x : log(2)
andb $0x45, %ah
jz 2f
- fstp %st(1) // x-1 : log(2)
+ fxam
+ fnstsw
+ andb $0x45, %ah
+ cmpb $0x40, %ah
+ jne 5f
+ fabs // log(1) is +0 in all rounding modes.
+5: fstp %st(1) // x-1 : log(2)
fyl2xp1 // log(x)
ret
@@ -47,7 +47,13 @@ ENTRY(__ieee754_logf)
fnstsw // x-1 : x : log(2)
andb $0x45, %ah
jz 2f
- fstp %st(1) // x-1 : log(2)
+ fxam
+ fnstsw
+ andb $0x45, %ah
+ cmpb $0x40, %ah
+ jne 5f
+ fabs // log(1) is +0 in all rounding modes.
+5: fstp %st(1) // x-1 : log(2)
fyl2xp1 // log(x)
ret
@@ -47,7 +47,13 @@ ENTRY(__ieee754_logl)
fnstsw // x-1 : x : log(2)
andb $0x45, %ah
jz 2f
- fstp %st(1) // x-1 : log(2)
+ fxam
+ fnstsw
+ andb $0x45, %ah
+ cmpb $0x40, %ah
+ jne 5f
+ fabs // log(1) is +0 in all rounding modes.
+5: fstp %st(1) // x-1 : log(2)
fyl2xp1 // log(x)
ret
@@ -1096,6 +1096,18 @@ Function: "log1p":
ildouble: 1
ldouble: 1
+Function: "log_downward":
+ildouble: 1
+ldouble: 1
+
+Function: "log_towardzero":
+ildouble: 1
+ldouble: 1
+
+Function: "log_upward":
+ildouble: 1
+ldouble: 1
+
Function: "pow":
ildouble: 1
ldouble: 1
@@ -46,7 +46,13 @@ ENTRY(__ieee754_logl)
fcomip %st(1) // |x-1| : x-1 : x : log(2)
fstp %st(0) // x-1 : x : log(2)
jc 2f
- fstp %st(1) // x-1 : log(2)
+ fxam
+ fnstsw
+ andb $0x45, %ah
+ cmpb $0x40, %ah
+ jne 4f
+ fabs // // log(1) is +0 in all rounding modes.
+4: fstp %st(1) // x-1 : log(2)
fyl2xp1 // log(x)
ret
@@ -96,6 +96,10 @@ __ieee754_log (double x)
if (__glibc_likely (ABS (w) > U03))
goto case_03;
+ /* log (1) is +0 in all rounding modes. */
+ if (w == 0.0)
+ return 0.0;
+
/*--- Stage I, the case abs(x-1) < 0.03 */
t8 = MHALF * w;
@@ -240,6 +240,8 @@ __ieee754_logl(long double x)
/* On this interval the table is not used due to cancellation error. */
if ((x <= 1.0078125L) && (x >= 0.9921875L))
{
+ if (x == 1.0L)
+ return 0.0L;
z = x - 1.0L;
k = 64;
t.value = 1.0L;
@@ -45,7 +45,13 @@ ENTRY(__ieee754_logl)
fnstsw // x-1 : x : log(2)
andb $0x45, %ah
jz 2f
- fstp %st(1) // x-1 : log(2)
+ fxam
+ fnstsw
+ andb $0x45, %ah
+ cmpb $0x40, %ah
+ jne 5f
+ fabs // log(1) is +0 in all rounding modes.
+5: fstp %st(1) // x-1 : log(2)
fyl2xp1 // log(x)
ret
@@ -1170,6 +1170,22 @@ ifloat: 1
ildouble: 1
ldouble: 1
+Function: "log_downward":
+float: 1
+ifloat: 1
+ildouble: 1
+ldouble: 1
+
+Function: "log_towardzero":
+ildouble: 1
+ldouble: 1
+
+Function: "log_upward":
+float: 1
+ifloat: 1
+ildouble: 1
+ldouble: 1
+
Function: "pow":
float: 1
ifloat: 1