The current approach tracks math maximum supported errors by explicitly
setting them per function and architecture. On newer implementations or
new compiler versions, the file is updated with newer values if it
shows higher results. The idea is to track the maximum known error, to
update the manual with the obtained values.
The constant libm-test-ulps shows little value, where it is usually a
mechanical change done by the maintainer, for past releases it is
usually ignored whether the ulp change resulted from a compiler
regression, and the math tests already have a maximum ulp error that
triggers a regression.
It was shown by a recent update after the new acosf [1] implementation
that is correctly rounded, where the libm-test-ulps was indeed from a
compiler issue.
This patch removes all arch-specific libm-test-ulps, adds system generic
libm-test-ulps where applicable, and changes its semantics. The generic
files now track specific implementation constraints, like if it is
expected to be correctly rounded, or if the system-specific has
different error expectations.
Now multiple libm-test-ulps can be defined, and system-specific
overrides generic implementation. This is for the case where
arch-specific implementation might show worse precision than generic
implementation, for instance, the cbrtf on i686.
Regressions are only reported if the implementation shows larger errors
than 9 ulps (13 for IBM long double) unless it is overridden by
libm-test-ulps and the maximum error is not printed at the end of tests.
The regen-ulps rule is also removed since it does not make sense to
update the libm-test-ulps automatically.
The manual error table is also removed, Paul Zimmermann and others have
been tracking libm precision with a more comprehensive analysis for some
releases; so link to his work instead.
[1] https://sourceware.org/git/?p=glibc.git;a=commit;h=9cc9f8e11e8fb8f54f1e84d9f024917634a78201
C23 adds various <math.h> function families originally defined in TS
18661-4. Add the rsqrt functions (1/sqrt(x)). The test inputs are
taken from those for sqrt.
Tested for x86_64 and x86, and with build-many-glibcs.py.
Single-precision remainderf() and quad-precision remainderl()
implementation derived from Sun is affected by an issue when the result
is +-0. IEEE754 requires that if remainder(x, y) = 0, its sign shall be
that of x regardless of the rounding direction.
The implementation seems to have assumed that x - x = +0 in all
rounding modes, which is not the case. When rounding direction is
roundTowardNegative the sign of an exact zero sum (or difference) is −0.
Regression tests that triggered this erroneous behavior are added to
math/libm-test-remainder.inc.
Tested for cross riscv64 and powerpc.
Original fix by: Bruce Evans <bde@FreeBSD.org> in FreeBSD's
a2ddfa5ea726c56dbf825763ad371c261b89b7c7.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
A number of fma tests started to fail on hppa when gcc was changed to
use Ranger rather than EVRP. Eventually I found that the value of
a1 + u.d in this is block of code was being computed in FE_TOWARDZERO
mode and not the original rounding mode:
if (TININESS_AFTER_ROUNDING)
{
w.d = a1 + u.d;
if (w.ieee.exponent == 109)
return w.d * 0x1p-108;
}
This caused the exponent value to be wrong and the wrong return path
to be used.
Here we add an optimization barrier after the rounding mode is reset
to ensure that the previous value of a1 + u.d is not reused.
Signed-off-by: John David Anglin <dave.anglin@bell.net>
GCC aligns global data to 16 bytes if their size is >= 16 bytes. This patch
changes the exp_data struct slightly so that the fields are better aligned
and without gaps. As a result on targets that support them, more load-pair
instructions are used in exp. Exp10 is improved by moving invlog10_2N later
so that neglog10_2hiN and neglog10_2loN can be loaded using load-pair.
The exp benchmark improves 2.5%, "144bits" by 7.2%, "768bits" by 12.7% on
Neoverse V2. Exp10 improves by 1.5%.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
The libm size improvement built with "--enable-stack-protector=strong
--enable-bind-now=yes --enable-fortify-source=2":
Before:
text data bss dec hex filename
585192 860 12 586064 8f150 aarch64-linux-gnu/math/libm.so
960775 1068 12 961855 ead3f x86_64-linux-gnu/math/libm.so
1189174 5544 368 1195086 123c4e powerpc64le-linux-gnu/math/libm.so
After:
text data bss dec hex filename
584952 860 12 585824 8f060 aarch64-linux-gnu/math/libm.so
960615 1068 12 961695 eac9f x86_64-linux-gnu/math/libm.so
1189078 5544 368 1194990 123bee powerpc64le-linux-gnu/math/libm.so
The are small code changes for x86_64 and powerpc64le, which do not
affect performance; but on aarch64 with gcc-14 I see a slight better
code generation due the usage of ldq for floating point constant loading.
Reviewed-by: Andreas K. Huettel <dilfridge@gentoo.org>
The libm size improvement built with "--enable-stack-protector=strong
--enable-bind-now=yes --enable-fortify-source=2":
Before:
text data bss dec hex filename
587304 860 12 588176 8f990 aarch64-linux-gnu-master/math/libm.so
962855 1068 12 963935 eb55f x86_64-linux-gnu-master/math/libm.so
1191222 5544 368 1197134 12444e powerpc64le-linux-gnu-master/math/libm.so
After:
text data bss dec hex filename
585192 860 12 586064 8f150 aarch64-linux-gnu/math/libm.so
960775 1068 12 961855 ead3f x86_64-linux-gnu/math/libm.so
1189174 5544 368 1195086 123c4e powerpc64le-linux-gnu/math/libm.so
The are small code changes for x86_64 and powerpc64le, which do not
affect performance; but on aarch64 with gcc-14 I see a slight better
code generation due the usage of ldq for floating point constant loading.
Reviewed-by: Andreas K. Huettel <dilfridge@gentoo.org>
The CORE-MATH implementation is correctly rounded (for any rounding mode),
although it should worse performance than current one. The current
implementation performance comes mainly from the internal usage of
the optimize expf implementation, and shows a maximum ULPs of 2 for
FE_TONEAREST and 3 for other rounding modes.
The code was adapted to glibc style and to use the definition of
math_config.h (to handle errno, overflow, and underflow).
Benchtest on x64_64 (Ryzen 9 5900X, gcc 14.2.1), aarch64 (Neoverse-N1,
gcc 13.3.1), and powerpc (POWER10, gcc 13.2.1):
Latency master patched improvement
x86_64 40.6995 49.0737 -20.58%
x86_64v2 40.5841 44.3604 -9.30%
x86_64v3 39.3879 39.7502 -0.92%
i686 112.3380 129.8570 -15.59%
aarch64 (Neoverse) 18.6914 17.0946 8.54%
power10 11.1343 9.3245 16.25%
reciprocal-throughput master patched improvement
x86_64 18.6471 24.1077 -29.28%
x86_64v2 17.7501 20.2946 -14.34%
x86_64v3 17.8262 17.1877 3.58%
i686 64.1454 86.5645 -34.95%
aarch64 (Neoverse) 9.77226 12.2314 -25.16%
power10 4.0200 5.3316 -32.63%
Signed-off-by: Alexei Sibidanov <sibid@uvic.ca>
Signed-off-by: Paul Zimmermann <Paul.Zimmermann@inria.fr>
Signed-off-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
Reviewed-by: DJ Delorie <dj@redhat.com>
ieee_long_double_shape_type has
typedef union
{
long double value;
struct
{
...
int sign_exponent:16;
...
} parts;
} ieee_long_double_shape_type;
Clang issues an error:
../sysdeps/ieee754/ldbl-96/test-totalorderl-ldbl-96.c:49:2: error: implicit truncation from 'int' to bit-field changes value from 65535 to -1 [-Werror,-Wbitfield-constant-conversion]
49 | SET_LDOUBLE_WORDS (ldnx, 0xffff,
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
50 | tests[i] >> 32, tests[i] & 0xffffffffULL);
|
Use -1, instead of 0xffff, to silence Clang.
Signed-off-by: H.J. Lu <hjl.tools@gmail.com>
Reviewed-by: Sam James <sam@gentoo.org>
C23 adds various <math.h> function families originally defined in TS
18661-4. Add the atan2pi functions (atan2(y,x)/pi).
Tested for x86_64 and x86, and with build-many-glibcs.py.
C23 adds various <math.h> function families originally defined in TS
18661-4. Add the atanpi functions (atan(x)/pi).
Tested for x86_64 and x86, and with build-many-glibcs.py.
C23 adds various <math.h> function families originally defined in TS
18661-4. Add the asinpi functions (asin(x)/pi).
Tested for x86_64 and x86, and with build-many-glibcs.py.
C23 adds various <math.h> function families originally defined in TS
18661-4. Add the acospi functions (acos(x)/pi).
Tested for x86_64 and x86, and with build-many-glibcs.py.
C23 adds various <math.h> function families originally defined in TS
18661-4. Add the tanpi functions (tan(pi*x)).
Tested for x86_64 and x86, and with build-many-glibcs.py.
C23 adds various <math.h> function families originally defined in TS
18661-4. Add the sinpi functions (sin(pi*x)).
Tested for x86_64 and x86, and with build-many-glibcs.py.
C23 adds various <math.h> function families originally defined in TS
18661-4. Add the cospi functions (cos(pi*x)).
Tested for x86_64 and x86, and with build-many-glibcs.py.
Some CORE-MATH routines uses roundeven and most of ISA do not have
an specific instruction for the operation. In this case, the call
will be routed to generic implementation.
However, if the ISA does support round() and ctz() there is a better
alternative (as used by CORE-MATH).
This patch adds such optimization and also enables it on powerpc.
On a power10 it shows the following improvement:
expm1f master patched improvement
latency 9.8574 7.0139 28.85%
reciprocal-throughput 4.3742 2.6592 39.21%
Checked on powerpc64le-linux-gnu and aarch64-linux-gnu.
Reviewed-by: DJ Delorie <dj@redhat.com>
The k>>31 in signgam = 1 - (((k&(k>>31))&1)<<1); is not portable:
* The ISO C standard says "If E1 has a signed type and a negative
value, the resulting value is implementation-defined." (this is
still in C23).
* If the int type is larger than 32 bits (e.g. a 64-bit type),
then k = INT_MAX; line 144 will make k>>31 put 1 in bit 0
(thus signgam will be -1) while 0 is expected.
Moreover, instead of the fx >= 0x1p31f condition, testing fx >= 0
is probably better for 2 reasons:
The signgam expression has more or less a condition on the sign
of fx (the goal of k>>31, which can be dropped with this new
condition). Since fx ≥ 0 should be the most common case, one can
get signgam directly in this case (value 1). And this simplifies
the expression for the other case (fx < 0).
This new condition may be easier/faster to test on the processor
(e.g. by avoiding a load of a constant from the memory).
This is commit d41459c731865516318f813cf4c966dafa0eecbf from CORE-MATH.
Checked on x86_64-linux-gnu.
The CORE-MATH implementation is correctly rounded (for any rounding mode)
and shows better performance to the generic tanf.
The code was adapted to glibc style, to use the definition of
math_config.h, to remove errno handling, and to use a generic
128 bit routine for ABIs that do not support it natively.
Benchtest on x64_64 (Ryzen 9 5900X, gcc 14.2.1), aarch64 (neoverse1,
gcc 13.2.1), and powerpc (POWER10, gcc 13.2.1):
latency master patched improvement
x86_64 82.3961 54.8052 33.49%
x86_64v2 82.3415 54.8052 33.44%
x86_64v3 69.3661 50.4864 27.22%
i686 219.271 45.5396 79.23%
aarch64 29.2127 19.1951 34.29%
power10 19.5060 16.2760 16.56%
reciprocal-throughput master patched improvement
x86_64 28.3976 19.7334 30.51%
x86_64v2 28.4568 19.7334 30.65%
x86_64v3 21.1815 16.1811 23.61%
i686 105.016 15.1426 85.58%
aarch64 18.1573 10.7681 40.70%
power10 8.7207 8.7097 0.13%
Signed-off-by: Alexei Sibidanov <sibid@uvic.ca>
Signed-off-by: Paul Zimmermann <Paul.Zimmermann@inria.fr>
Signed-off-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
Reviewed-by: DJ Delorie <dj@redhat.com>
The commit 9247f53219 triggered some regressions on loongarch and
riscv:
math/test-float-log10
math/test-float32-log10
And it is due a wrong sync with CORE-MATH for special 0.0/-0.0
inputs.
Checked on aarch64-linux-gnu and loongarch64-linux-gnu-lp64d.
Adhemerval Zanella
Joseph Myers
Sergei Zimmerman
John David Anglin
Wilco Dijkstra
Paul Eggert
H.J. Lu
Vincent Lefevre