Damian McGuckin <damianm at esi dot com dot au> noted that the accuracy
claims in the code for cbrt(3) and cbrtl(3) were incorrect. Fix the
comments to more accurately describe the accuracies.
PR: 265603
MFC after: 3 days
Since https://github.com/llvm/llvm-project/commit/ca75ac5f04f2, clang 15
has a new warning about _Generic selection expressions, such as used in
math.h:
lib/libc/gdtoa/_ldtoa.c:82:10: error: due to lvalue conversion of the controlling expression, association of type 'volatile float' will never be selected because it is qualified [-Werror,-Wunreachable-code-generic-assoc]
switch (fpclassify(u.e)) {
^
lib/msun/src/math.h:109:2: note: expanded from macro 'fpclassify'
__fp_type_select(x, __fpclassifyf, __fpclassifyd, __fpclassifyl)
^
lib/msun/src/math.h:85:14: note: expanded from macro '__fp_type_select'
volatile float: f(x), \
^
This is because the controlling expression always undergoes lvalue
conversion first, dropping any cv-qualifiers. The 'const', 'volatile',
and 'volatile const' associations will therefore never be used.
MFC after: 1 week
Reviewed by: theraven
Differential Revision: https://reviews.freebsd.org/D35815
Summary:
These functions are missing from the library itself, and exist solely in
the header. This breaks a few ports that expect libm to have the
symbols in the library itself.
Questions on MFC-ability: Can this be MFC'd to 13.2, and how?
Reviewers: imp, emaste, kib
Reviewed By: kib
Differential Revision: https://reviews.freebsd.org/D35204
There are some sections which could be improved
and work to do so is on going. The work will be
covered via 'X-MFC-WITH' commits.
Obtained from: OpenBSD
MFC after: 1 month
Differential Revision: https://reviews.freebsd.org/D34759
All supported compilers (modern versions of GCC and clang) support
this.
Many places didn't have an #else so would just silently do the wrong
thing. Ancient versions of icc (the original motivation for this) are
no longer a compiler FreeBSD supports.
PR: 263102 (exp-run)
Reviewed by: brooks, imp
Differential Revision: https://reviews.freebsd.org/D34797
. Disconnect imprecise.c from the build. This file can be deleted.
. Add b_tgammal.c to the build for ld80 and ld128 targets. The ld128
is a 'git mv' of imprecise.c to ld128/b_tgammal.c.
* lib/msun/ld80/b_expl.c:
. New file. Implement __exp__D for ld80 targets. This is based on
bsdsrc/b_exp.c.
* lib/msun/ld80/b_logl.c:
. New file. Implement __log__D for ld80 targets. This is based on
bsdsrc/b_log.c.
* lib/msun/ld80/b_tgammal.c b/lib/msun/ld80/b_tgammal.c
. New file. Implement tgammal(x) for ld80 targets.
Submitted by: Steve Kargl
Differential Revision: https://reviews.freebsd.org/D33444
Reviewed by: pfg
. Disconnect b_exp.c and b_log.c from the build.
* lib/msun/bsdsrc/b_exp.c:
. Replace scalb() usage with C99's ldexp().
. Replace finite(x) usage with C99's isfinite().
. Whitespace changes towards style(9).
. Remove include of "mathimpl.h". It is no longer needed.
. Remove #if 0 ... #endif code, which has been present since svn r93211
(2002-03-26).
. New minimax polynomial coefficients.
. Add comments to explain origins of some constants.
. Use ansi-C prototype. Remove K&R prototype. Add static to prototype.
* lib/msun/bsdsrc/b_log.c:
. Remove include of "mathimpl.h". It is no longer needed.
. Fix comments to actually describe the code.
. Reduce minimax polynomial from degree 4 to degree 3.
This uses newly computed coefficients.
. Use ansi-C prototype. Remove K&R prototype. Add static to prototype.
. Remove volatile in declaration of u1.
. Alphabetize decalaration list.
. Whitespace changes towards style(9).
. In argument reduction of x to g and m, replace use of logb() and
ldexp() with a single call to frexp(). Add code to get 1 <= g < 2.
. Remove #if 0 ... #endif code, which has been present since svn r93211
(2002-03-26).
. The special case m == -1022, replace logb() with ilogb().
* lib/msun/bsdsrc/b_tgamma.c:
. Update comments. Fix comments where needed.
. Add float.h to get LDBL_MANT_DIG for weak reference of tgammal to tgamma.
. Remove include of "mathimpl.h". It is no longer needed.
. Use "math.h" instead of <math.h>.
. Add '#include math_private.h"
. Add struct Double from mathimpl.h and include b_log.c and b_exp.c.
. Remove forward declarations of neg_gam(), small_gam(), smaller_gam,
large_gam() and ratfun_gam() by re-arranging the code to move these
function above their first reference.
. New minimax coefficients for polynomial in large_gam().
. New splitting of a0 into a0hi nd a0lo, which include additional
bits of precision.
. Use ansi-C prototype. Remove K&R prototype.
. Replace the TRUNC() macro with a simple cast of a double entities
to float before assignment (functional changes).
. Replace sin(M_PI*z) with sinpi(z) and cos(M_PI*(0.5-z)) with cospi(0.5-z).
Submitted by: Steve Kargl
Differential Revision: https://reviews.freebsd.org/D33444
Reviewed by: pfg
This look like a copy and paste leftover.
Reported by: enh@google.com (via freebsd-numerics@)
Reviewed by: Steve Kargl <sgk@troutmask.apl.washington.edu>
MFC after: 3 days
Some of the functions in msun can be implemented using a compiler
builtin function to generate a small number of instructions. Implement
this support in fma, fmax, fmin, and sqrt on arm64.
Care must be taken as the builtin can be implemented as a function
call on some architectures that lack direct support. In these cases
we need to use the original code path.
As we don't set errno on failure build with -fno-math-errno so the
toolchain doesn't convert a builtin into a function call when it
detects a failure, e.g. gcc will add a call to sqrt when the input
is negative leading to an infinite loop.
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D32801
In ac76bc1145, I added a few volatiles to work around ctrig_test
failures with {inf,inf}. This is not necessary anymore now, since in
3b00222f15 we added -fp-exception-behavior=maytrap for clang >= 10 in
libm's Makefile. (The flag tells clang to use stricter floating point
semantics, which libm depends on.)
PR: 244732, 254911
Fixes: ac76bc1145
MFC after: 3 days
The change implements cexpl() for both ld80 and ld128 architectures.
Testing was done on x86_64 and aarch64 systems.
Along the way sincos[fl]() use an optimization that reduces the argument
to being done one rather than twice. This optimization actually pointed
to a bug in the ld128 version of sincosl(), which is now fixed. In
addition, the minmax polynomial coefficients for sincosl() have been
updated.
A concise log of the file-by-file changes follows.
* include/complex.h:
. Add a prototype for cexpl().
* lib/msun/Makefile:
. Add s_cexpl.c to the build.
. Setup a link for cexpl.3 to cexp.3.
* lib/msun/Symbol.map:
. Expose cexpl symbol in libm shared library.
* lib/msun/ld128/s_cexpl.c:
* Implementation of cexpl() for 128-bit long double architectures.
Tested on an aarch64 system.
* lib/msun/ld80/s_cexpl.c:
* Implementation of cexpl() for Intel 80-bit long double.
* lib/msun/man/cexp.3:
. Document cexpl().
* lib/msun/man/complex.3:
. Add a BUGS section about cpow[fl].
* lib/msun/src/s_cexp.c:
. Include float.h for weak references on 53-bit long double targets.
. Use sincos() to reduce argument reduction cost.
* lib/msun/src/s_cexpf.c:
. Use sincosf() to reduce argument reduction cost.
* lib/msun/src/k_sincosl.h:
. Catch up with the new minmax polynomial coefficients for the kernel for
the 128-bit cosl() implementation.
. BUG FIX: *cs was used where *sn should have been. This means that sinl()
was no computed correctly when iy != 0.
* lib/msun/src/s_cosl.c:
. Include fpmath.h to get access to IEEEl2bits.
. Replace M_PI_4 with pio4, a 64-bit or 113-bit approximation for pi / 4.
PR: 216862
MFC after: 1 week
As mention previously, the minmax polynomial approximation
in the kernel for cosl() seem to have a bad set of coefficients.
In testing, cosl() in the interval [0.785, pi/4] for 1 million
values and pi/4 written to 37 decimal digits. The old version
on an aarch64 system gave
% tlibm/tlibm_lmath -l -x 0.78 -X
7.85398163397448309615660845819875721e-1L cos
Interval tested for cosl: [0.78,0.785398]
count: 1000000
xm = 7.80213913234863919029058821396125599e-01L
libm = 7.10763080972549562455058499280609083e-01L
mpfr = 7.10763080972549562455058499280608983e-01L
ULP = 1.04431
The max ULP exceeds 1, which is not good. So, I rinsed off a 10
year code and recomputed coefficients. The new minmax polynomial
now yields
% tlibm/tlibm_lmath -l -x 0.78 -X
7.85398163397448309615660845819875721e-1L cos
Interval tested for cosl: [0.78,0.785398]
count: 1000000
xm = 7.82916198746768272588844890973704219e-01L
libm = 7.08859615479571058183956453286628396e-01L
mpfr = 7.08859615479571058183956453286628469e-01L
ULP = 0.75407
which is very good.
PR: 218514
MFC after: 1 week
Mark Murray graciously provided access to an aarch64 system
to test the ld128 implementations. This patch address
* Misuses of copysignl() in sinpil() and tanpil().
* Redo the splitting of argument 'x' into an integer part and
remainder. The remainder must satify 0 <= r < 1.
* Update the reduction of the integer part to something that can
easily be seen as even or odd, e.g., sin(pi*x) = (-1)^n*sin(pi*r)
with n <= 2^112 and we an reduce n by subtracting integer powers
of 2.
* In s_cospil.c, fix typos where 'x' is used where 'ax', the
remainder, is required.
* In tanpil(), fix the use of an uninitialized variable, ax = fabsl(ax),
ax should be x in fabsl().
One item of note, in the limited tested on aarch64, the max ULP
for sinpil() and cospil() were less than 1.1 ULP, which is higher
that the desired max ULP less than 1. This was traced to the
kernel for cosl() in the fundamental interval [0,pi/4].
The coefficients in the minmax polynomial likely need refinement.
PR: 218514
MFC after: 1 week
The patch fixes the omission of '#include <float.h>', which is needed for
the weak reference on systems with LDBL_MANT_DIG == DBL_MANT_DIG.
PR: 218514
MFC after: 1 week
Both IEEE-754 2008 and ISO/IEC TS 18661-4 define the half-cycle
trignometric functions cospi, sinpi, and tanpi. The attached
patch implements cospi[fl], sinpi[fl], and tanpi[fl]. Limited
testing on the cospi and sinpi reveal a max ULP less than 0.89;
while tanpi is more problematic with a max ULP less than 2.01
in the interval [0,0.5]. The algorithms used in these functions
are documented in {ks}_cospi.c, {ks}_sinpi.c, and s_tanpi.c.
Note. I no longer have access to a system with ld128 and
adequate support to compile and test the ld128 implementations
of these functions. Given the almost complete lack of input from
others on improvements to libm, I doubt that anyone cares. If
someone does care, the ld128 files contain a number of FIXME comments,
and in particular, while the polynomial coefficients are given
I did not update the polynomial algorithms to properly use the
coefficients.
PR: 218514
MFC after: 2 weeks
These files were copied from MUSL. Add the standard copyright notice and
SPDX-License-Identifier: MIT consistent with our new draft license
policy. It reads word for word the same as the MIT license on the SPDX
web site. Add a pointer to the MUSL COPYIRGHT file which contains a list
of all authors of MUSL.
Sponsored by: Netflix
Noticed by: Steve Kargl
Summary:
From Steve Kargl:
Paul Zimmermann has identified a bug in Openlibm's powf(),
which is identical to FreeBSD's libm. Both derived from
fdlibm. https://github.com/JuliaMath/openlibm/issues/212.
Consider
% cat h.c
int
main(void)
{
float x, y, z;
x = 0x1.ffffecp-1F;
y = -0x1.000002p+27F;
z = 0x1.557a86p115F;
printf("%e %e %e <-- should be %e\n", x, y, powf(x,y), z);
return 0;
}
% cc -o h -fno-builtin h.c -lm && ./h
9.999994e-01 -1.342177e+08 inf <-- should be 5.540807e+34
Reviewers: manu
Subscribers: imp, andrew, emaste
Differential Revision: https://reviews.freebsd.org/D31865
.Fa is the suitable macro for functions in comparsion to the
.Ar macro, which should be used for commandline arguments.
While here, fix some mandoc warnings.
Reviewed by: imp (earlier version)
Obtained from: OpenBSD (in partial)
MFC after: 3 days
Differential Revision: https://reviews.freebsd.org/D31090
After df3b437c1e, older gcc's such as
4.2.1 (still used on earlier branches for e.g. mips and powerpc) and
6.3.0 (still used for some cross-builds) started throwing bogus errors
like:
In file included from /workspace/src/lib/msun/src/s_llround.c:11:0:
/workspace/src/lib/msun/src/s_lround.c:54:31: error: initializer element is not constant
static const type dtype_min = type_min - 0.5;
^~~~~~~~
/workspace/src/lib/msun/src/s_lround.c:55:31: error: initializer element is not constant
static const type dtype_max = type_max + 0.5;
^~~~~~~~
Since 'type_min' and 'type_max' are constants declared just above these
lines this error is nonsensical, but older gcc's are not smart enough.
Work around the error by reusing the (type)DTYPE_MIN and (type)DTYPE_MAX
macros, so I can MFC this right away, unbreaking a few stable builds.
MFC after: immediately
It turned out that the (type)DTYPE_MAX conversions at the top of
s_lround.c are now emitted as cvtsi2sd instructions, at least on SSE
capable CPUs. This caused the FE_INEXACT flag to always be set, at least
for the double and float variants. Under clang 11, the whole INRANGE()
comparisons were still optimized away, but this has "improved" in clang
12, due to stricter adherence to the -ffp-exception-behavior=maytrap
compiler flag.
To avoid run-time integer to float conversions, use static constants
instead, so they are computed at compile time, and the INRANGE()
statements are optimized away again, if applicable.
While here, use an integer instead of a floating type to store the test
results in lround_test.c, as this is more appropriate, and we can also
drop the volatile hack.
Reported by: arichardson
MFC after: 3 days
Summary:
Fix FPU exception management for powerpcspe. Bits are in a different place from
the standard FPSCR, so we need to handle the shifting differences. Also,
there's no concept of a "software exception" raise, so we need to do exceptional
math to trigger the exception from software.
Reviewed By: alfredo
Differential Revision: https://reviews.freebsd.org/D22824
For some reason the ld128 log1pl() implementation is less accurate than
logl(), but does at least guarantee precision >= the ld80 implementation.
Mark log1p_accuracy_tests as XFAIL for ld128 and increase the log1p tolerance
to the ld80 equivalent in accuracy_tests to avoid losing test coverage for
the other functions.
PR: 253984
Reviewed By: ngie, dim
Differential Revision: https://reviews.freebsd.org/D29039
This avoids build failures due to the clang 12 warning:
'#pragma FENV_ACCESS' is not supported on this target - ignored
Clang 12 currently emits this warning for all non-x86 architectures.
While this can result in incorrect code generation (e.g. on AArch64 some
exceptions are not raised as expected), this is a pre-existing issue and
we should not fail the build due to this warning.
Reviewed By: dim, emaste
MFC after: 1 week
Differential Revision: https://reviews.freebsd.org/D29743
After 3b00222f15, it turns out that clang only supports strict
floating point semantics for SystemZ and x86 at the moment, while for
other architectures it is still experimental.
Therefore, only use -fp-exception-behavior=maytrap on x86 for now,
otherwise this option results in "error: overriding currently
unsupported use of floating point exceptions on this target
[-Werror,-Wunsupported-floating-point-opt]" on other architectures.
Fixes: 3b00222f15
PR: 254911
MFC after: 1 week
When using clang with x86_64 CPUs that support AVX, some floating point
transformations may raise exceptions that would not have been raised by
the original code. To avoid this, use the -fp-exception-behavior=maytrap
flag, introduced in clang 10.0.0.
In particular, this fixes a number of test failures with ctanhf(3) and
ctanf(3), when libm is compiled with -mavx. An unexpected FE_INVALID
exception is then raised, because clang emits vdivps instructions to
perform certain divides. (The vdivps instruction operates on multiple
single-precision float operands simultaneously, but the exceptions may
be influenced by unused parts of the XMM registers. In this particular
case, it was calculating 0 / 0, which results in FE_INVALID.)
If -fp-exception-behavior=maytrap is specified however, clang uses
vdivss instructions instead, which work on one operand, and should not
raise unexpected exceptions.
Reported by: olivier
Reviewed by: arichardson
PR: 254911
MFC after: 1 week
Differential Revision: https://reviews.freebsd.org/D29686
When compiling parts of math.h with clang using a C standard before C11,
and using -pedantic, it will result in warnings similar to:
bug254714.c:5:11: warning: '_Generic' is a C11 extension [-Wc11-extensions]
return !isfinite(1.0);
^
/usr/include/math.h:111:21: note: expanded from macro 'isfinite'
^
/usr/include/math.h:82:39: note: expanded from macro '__fp_type_select'
^
This is because the block that enables use of _Generic is conditional
not only on C11, but also on whether the compiler advertises support for
C generic selections via __has_extension(c_generic_selections).
To work around the warning without having to pessimize the code, use the
__extension__ keyword, which is supported by both clang and gcc. While
here, remove the check for __clang__, as _Generic has been supported for
a long time by gcc too now.
Reported by: yuri
PR: 254714
MFC after: 1 week
The man page says "The feenableexcept(), fedisableexcept(), and
fegetexcept() functions return a bitmap of the exceptions that were
unmasked prior to the call.", so we should return zero not -1.
Reviewed By: mhorne
MFC after: 3 days
Differential Revision: https://reviews.freebsd.org/D29386
After increasing the lib/msun/tests WARNS to 6, this triggers a
compilation error for RISC-V.
Fixes: 87d65c747a ("lib/msun: Allow building tests with WARNS=6")
Reported by: Jenkins
The LLVM bug was fixed a long time ago and with D29076 this test actually
passes now.
Reviewed By: andrew
Differential Revision: https://reviews.freebsd.org/D29092
If long double has more than 64 mantissa bits, using uint64_t to hold the
mantissa bits will truncate the value and result in test failures. To fix
this problem use __uint128_t since all platforms that have
__LDBL_MANT_DIG__ > 64 also have compiler support for 128-bit integers.
Reviewed By: rlibby
Differential Revision: https://reviews.freebsd.org/D29076
I only tested the WARNS=6 change on AArch64 and AMD64, but this file has
unused functions for architectures with LDBL_PREC == 53.
While touching this file change the LDBL_PREC == 53 checks to i386 checks.
The long double tests should only be disabled for i386 (due to the rather
odd rounding mode that it uses) not all architectures where long double
is the same as double.
PR: 205449
Fixes: 87d65c747a ("lib/msun: Allow building tests with WARNS=6")
Reported by: Jenkins
Some CPUs (e.g. AArch64 QEMU) cannot trap on floating point exceptions and
therefore ignore the writes to the floating point control register inside
feenableexcept(). If no exceptions are enabled after
feenableexcept(FE_ALL_EXCEPT), we can assume that the CPU does not
support exceptions and we can then skip the test.
Reviewed By: dim
Differential Revision: https://reviews.freebsd.org/D29095
These appears to have been resolved by compiling the test with -fno-builtin
and/or using a newer compiler.
PR: 208703
Reviewed By: ngie
Differential Revision: https://reviews.freebsd.org/D28884
This provides better error messages that just an assertion failure and
also makes it easier to mark individual tests as XFAIL.
It was also helpful when coming up with D28786 and D28787.
Differential Revision: https://reviews.freebsd.org/D28798
Apparently GCC only supports arithmetic expressions that use static
const variables in initializers starting with GCC8. To keep older
versions happy use a macro instead.
Fixes: 221622ec0c ("lib/msun: Avoid FE_INEXACT for x86 log2l/log10l")
Reported by: Jenkins
Reviewed By: imp
Differential Revision: https://reviews.freebsd.org/D29233
The existing code masked off all bits that it didn't know about. To be
future-proof, we should save and restore the entire fpcr/fpsr registers.
Additionally, the existing fesetenv() was incorrectly setting the rounding
mode in fpsr instead of fpcr.
This patch stores fpcr in the high 32 bits of fenv_t and fpsr in the low
bits instead of trying to interleave them in a single 32-bit field.
Technically, this is an ABI break if you re-compile parts of your code or
pass a fenv_t between DSOs that were compiled with different versions
of fenv.h. However, I believe we should fix this since the existing code
was broken and passing fenv_t across DSOs should rarely happen.
Reviewed By: andrew
Differential Revision: https://reviews.freebsd.org/D29160
This fixes tests/lib/msun/logarithm_test after compiling the test with
-fno-builtin (D28577). Adding invln10_lo + invln10_10 results in
FE_INEXACT (for all inputs) and the same for the log2l invln2_lo + invln2_hi.
This patch avoids FE_INEXACT (for exact results such as 0) by defining a
constant and using that.
Reviewed By: dim
Differential Revision: https://reviews.freebsd.org/D28786
This caused LDBL_MANT_DIG to not be defined and therefore the scalbnl
alias was not being emitted for double==long double platforms.
Fixes: 760b2ffc ("Update scalbn* functions to the musl versions")
Reported by: Jenkins