Will Newton [Fri, 23 May 2014 14:24:30 +0000 (15:24 +0100)]
ARM: Fix handling of concurrent TLS descriptor resolution
The current code for handling concurrent resolution says that the
ABI for _dl_tlsdesc_resolve_hold is the same as that of
_dl_tlsdesc_lazy_resolver. However _dl_tlsdesc_resolve_hold is
called from the trampoline directly rather than the lazy resolver
stub so, for example, r2 has not been pushed so does not needed
to be restored.
This fixes an intermittent failure in nptl/tst-tls3 when building
glibc for arm-linux-gnueabihf with -mtls-dialect=gnu2.
ChangeLog:
2014-05-27 Will Newton <will.newton@linaro.org>
[BZ #16990]
* sysdeps/arm/dl-tlsdesc.S (_dl_tlsdesc_resolve_hold): Save
and restore r2 rather than just restoring.
Fix offset computation for append+ mode on switching from read (BZ #16724)
The offset computation in write mode uses the fact that _IO_read_end
is kept in sync with the external file offset. This however is not
true when O_APPEND is in effect since switching to write mode ought to
send the external file offset to the end of file without making the
necessary adjustment to _IO_read_end.
Hence in append mode, offset computation when writing should only
consider the effect of unflushed writes, i.e. from _IO_write_base to
_IO_write_ptr.
The wiki has a detailed document that describes the rationale for
offsets returned by ftell in various conditions:
Carlos O'Donell [Mon, 26 May 2014 13:54:59 +0000 (09:54 -0400)]
Remove nested function mi_arena from malloc_info.
The nested function mi_arena was removed from malloc_info
and made into a non-nested static inline function of the same
name with the correct set of arguments passed from malloc_info.
This enables building glibc with compilers that don't support
nested functions. Future work on malloc_info should remove these
functions entirely to support JSON format output. Therefore we
do the minimum required to remove the nested function.
benchtests: Add new directive for benchmark initialization hook
Add a new 'init' directive that specifies the name of the function to
call to do function-specific initialization. This is useful for
benchmarks that need to do a one-time initialization before the
functions are executed.
Kyle McMartin [Mon, 26 May 2014 07:03:22 +0000 (12:33 +0530)]
[AARCH64] correct alignment of TLS_TCB_ALIGN (BZ #16796)
This fixes a variety of testsuite failures for me:
tststatic.out Error 1
tststatic2.out Error 1
tst-tls9-static.out Error 1
tst-audit8.out Error 127
tst-audit9.out Error 127
tst-audit1.out Error 127
and also has the added benefit of making LD_AUDIT/sotruss work on
AArch64.
Otherwise, we bail out early in _dl_try_allocate_static_tls as the
alignment requirement of the PT_TLS section in libc is 16.
Use NSS_STATUS_TRYAGAIN to indicate insufficient buffer (BZ #16878)
The netgroups nss modules in the glibc tree use NSS_STATUS_UNAVAIL
(with errno as ERANGE) when the supplied buffer does not have
sufficient space for the result. This is wrong, because the canonical
way to indicate insufficient buffer is to set the errno to ERANGE and
the status to NSS_STATUS_TRYAGAIN, as is used by all other modules.
This fixes nscd behaviour when the nss_ldap module returns
NSS_STATUS_TRYAGAIN to indicate that a netgroup entry is too long to
fit into the supplied buffer.
2005-11-16 Daniel Jacobowitz <dan@codesourcery.com>
but not applied to the (still separate) eabi port so necro'd
when the eabi port superceded the old abi. It was thence
copied into the new AArch64 port.
Joseph Myers [Fri, 23 May 2014 20:37:40 +0000 (20:37 +0000)]
Define TSVTX in tar.h for older POSIX (bug 16978).
As noted in bug 16978, older POSIX versions include
in the specified contents of <tar.h>, with only the 2001 edition
introducing the notion of XSI-conditional definitions and conditioning
that definition. Thus, this macro should be defined for
!__USE_XOPEN2K as well as for __USE_XOPEN, and this patch duly defines
it in that case. Tested x86_64.
Joseph Myers [Fri, 23 May 2014 12:07:50 +0000 (12:07 +0000)]
Fix log10 (1) in round-downward mode (bug 16977).
As with various other issues of this kind, bug 16977 is log10 (1)
wrongly returning -0 rather than +0 in round-downward mode because of
an implementation effectively in terms of log1p (x - 1). This patch
fixes the issue in the same way used for log.
Tested x86_64 and x86 and ulps updated accordingly. Also tested for
mips64 to confirm a fix was needed for ldbl-128 and to validate that
fix (also applied to ldbl-128ibm since that version of logl is
essentially the same as the ldbl-128 one).
[BZ #16977]
* sysdeps/i386/fpu/e_log10.S (__ieee754_log10): Take absolute
value when x - 1 is zero.
* sysdeps/i386/fpu/e_log10f.S (__ieee754_log10f): Likewise.
* sysdeps/i386/fpu/e_log10l.S (__ieee754_log10l): Likewise.
* sysdeps/ieee754/ldbl-128/e_log10l.c (__ieee754_log10l): Return
0.0L for an argument of 1.0L.
* sysdeps/ieee754/ldbl-128ibm/e_log10l.c (__ieee754_log10l):
Likewise.
* sysdeps/x86_64/fpu/e_log10l.S (__ieee754_log10l): Take absolute
value when x - 1 is zero.
* math/libm-test.inc (log10_test): Use ALL_RM_TEST.
* sysdeps/i386/fpu/libm-test-ulps: Update.
* sysdeps/x86_64/fpu/libm-test-ulps: Likewise.
This patch is the first in the series of patches that remove nested
functions from glibc.
Rationale: nested functions is a non-standard language feature;
removing nested functions
will allow to compile glibc with compilers other than GCC and thus
benefit from other compilers
and code analysis tools.
This patch fixes a similar issue to 736c304a1ab4cee36a2f3343f1698bc0abae4608, where for PPC32 if the symbol
is defined as hidden (memchr) then compiler will create a local branc
(symbol@local) and the linker will not create a required PLT call to
make the ifunc work. It changes the default hidden symbol (__GI_memchr)
to default memchr symbol for powerpc32 (__memchr_ppc32).
Joseph Myers [Wed, 21 May 2014 16:53:11 +0000 (16:53 +0000)]
Don't mention linuxthreads in Depend files.
I noticed that some of the Depend files, used to determine the
subdirectory build order in sysd-sorted, still mentioned linuxthreads,
although it hasn't been supported for many years. This patch removes
those references. In the case of nscd, it substitutes an nptl
reference, since I believe there is a fact a thread library dependence
there; the others already mentioned nptl.
Note that I am not at all confident in the completeness of these
Depend files.
Note also that references to linuxthreads remain in a comment in
sysdeps/unix/sysv/linux/ia64/Versions, and in manual/maint.texi,
manual/signal.texi and scripts/documented.sh.
Tested x86_64 that the installed shared libraries are unchanged by the
patch (as is sysd-sorted).
Joseph Myers [Wed, 21 May 2014 16:52:08 +0000 (16:52 +0000)]
Consistently use $(elf-objpfx).
As previously noted
<https://sourceware.org/ml/libc-alpha/2013-05/msg00696.html>,
$(elf-objpfx) and $(elfobjdir) are redundant and should be
consolidated. This patch consolidates on $(elf-objpfx) (for
consistency with $(csu-objpfx)), also changing direct uses of
$(common-objpfx)elf/ to use $(elf-objpfx).
Tested x86_64, including that installed shared libraries are unchanged
by the patch.
* Makeconfig [$(build-hardcoded-path-in-tests) = yes]
(rtld-tests-LDFLAGS): Use $(elf-objpfx) instead of
$(common-objpfx)elf/.
(link-libc-before-gnulib): Likewise.
(elfobjdir): Remove variable.
* Makefile (install): Use $(elf-objpfx) instead of
$(common-objpfx)elf/.
* Makerules (link-libc-args): Use $(elf-objpfx) instead of
$(elfobjdir)/.
(link-libc-deps): Likewise.
($(common-objpfx)libc.so): Likewise.
($(common-objpfx)linkobj/libc.so): Likewise.
[$(cross-compiling) = no] (symbolic-link-prog): Use $(elf-objpfx)
instead of $(common-objpfx)elf/.
(symbolic-link-list): Likewise.
* iconvdata/Makefile ($(inst_gconvdir)/gconv-modules)
[$(cross-compiling) = no]: Likewise.
* sysdeps/arm/Makefile (gnulib-arch): Use $(elf-objpfx) instead of
$(elfobjdir)/.
(static-gnulib-arch): Likewise.
* sysdeps/s390/s390-64/Makefile ($(inst_gconvdir)/gconv-modules)
[$(cross-compiling) = no]: Use $(elf-objpfx) instead of
$(common-objpfx)elf/.
localedata/ChangeLog:
* Makefile (LOCALEDEF): Use $(elf-objpfx) instead of
$(common-objpfx)elf/.
aarch64: Merge __local_multiple_threads offset with memory reference
This also highlights that we'd been loading 64-bits instead of
the proper 32-bits. Caught by the linker as a relocation error,
since the variable happened to be unaligned for 64-bits.
Joseph Myers [Tue, 20 May 2014 21:27:13 +0000 (21:27 +0000)]
Fix ARM build with GCC trunk.
sysdeps/unix/sysv/linux/arm/unwind-resume.c and
sysdeps/unix/sysv/linux/arm/unwind-forcedunwind.c have static
variables that are written in C code but only read from toplevel asms.
Current GCC trunk now optimizes away such apparently write-only static
variables, so causing a build failure. This patch marks those
variables with __attribute_used__ to avoid that optimization.
Tested that this fixes the build for ARM.
* sysdeps/unix/sysv/linux/arm/unwind-forcedunwind.c
(libgcc_s_resume): Use __attribute_used__.
* sysdeps/unix/sysv/linux/arm/unwind-resume.c (libgcc_s_resume):
Likewise.
This patch fixes the __copysignf optimized macro meant to internal libm
usage when used with constant value. Without the explicit cast to
float, if it is used with const double value (for instance, on
s_casinhf.c) double constants will be used and it may lead to precision
issues in some algorithms.
It fixes the following failures on PPC64/POWER7:
Failure: Test: Real part of: cacos_downward (inf + 0 i)
Result:
is: 1.19209289550781250000e-07 0x1.00000000000000000000p-23
should be: 0.00000000000000000000e+00 0x0.00000000000000000000p+0
Failure: Test: Real part of: cacos_downward (inf - 0 i)
Result:
is: 1.19209289550781250000e-07 0x1.00000000000000000000p-23
should be: 0.00000000000000000000e+00 0x0.00000000000000000000p+0
Failure: Test: Real part of: cacos_downward (inf + 0.5 i)
Result:
is: 1.19209289550781250000e-07 0x1.00000000000000000000p-23
should be: 0.00000000000000000000e+00 0x0.00000000000000000000p+0
Failure: Test: Real part of: cacos_downward (inf - 0.5 i)
Result:
is: 1.19209289550781250000e-07 0x1.00000000000000000000p-23
should be: 0.00000000000000000000e+00 0x0.00000000000000000000p+0
Failure: Test: Real part of: cacos_towardzero (inf + 0 i)
Result:
is: 1.19209289550781250000e-07 0x1.00000000000000000000p-23
should be: 0.00000000000000000000e+00 0x0.00000000000000000000p+0
Failure: Test: Real part of: cacos_towardzero (inf - 0 i)
Result:
is: 1.19209289550781250000e-07 0x1.00000000000000000000p-23
should be: 0.00000000000000000000e+00 0x0.00000000000000000000p+0
Failure: Test: Real part of: cacos_towardzero (inf + 0.5 i)
Result:
is: 1.19209289550781250000e-07 0x1.00000000000000000000p-23
should be: 0.00000000000000000000e+00 0x0.00000000000000000000p+0
Failure: Test: Real part of: cacos_towardzero (inf - 0.5 i)
Result:
is: 1.19209289550781250000e-07 0x1.00000000000000000000p-23
should be: 0.00000000000000000000e+00 0x0.00000000000000000000p+0
Aurelien Jarno [Tue, 20 May 2014 12:41:44 +0000 (14:41 +0200)]
Fix strtold on 32-bit sparc (and probably others) (BZ #16965)
This patch fixes an issue observed running the tst-strtod-round test on
32 bit sparc. In some conditions, strtold calls round_and_return, which in
turn calls __mpn_rshift with cnt = 0, while stdlib/rshift.c explicitly says
that cnts should satisfy 0 < CNT < BITS_PER_MP_LIMB. In this case, the code
end up doing a logical shift right of the same amount than the register,
which is undefined in the C standard.
Due to this bug, 32-bit sparc does not correctly convert the value
"0x1p-16446", but it is likely that other architectures are also
affected for other input values.
Aurelien Jarno [Thu, 15 May 2014 22:06:54 +0000 (00:06 +0200)]
fix nl_langinfo with static linking (BZ #16915)
For static linking the locale code avoids linking code and data for
unused categories. However for nl_langinfo we know only at runtime which
categories are used, so direct reference to every nl_current_CATEGORY
symbol should be done.
This was broken by commit bc3e1c127392da88d0c8bf2ae728147982a3d1bc where
nl_langinfo_l and nl_langinfo have been merged and some code has been
lost in the process.
In order to detect locales issues with static linking, compile a version
of tst-langinfo with static linking.
Note: this is Debian bug#747103 reported by Raphael <raphael.astier@eliot-sa.com>
Using the default header instead. This matches the kernel, which also
uses the generic header. Fixes the sys/wait.h conform issue, where
si_band had the wrong type.
Will Newton [Mon, 19 May 2014 13:38:30 +0000 (14:38 +0100)]
AArch64: Fix handling of nocancel syscall failures
The current code for nocancel syscalls does not do a comparison of
the system call return value. This leads to code being generated
where the b.cs follows the svc instruction directly without setting
the flags on which the branch depends.
ChangeLog:
2014-05-20 Will Newton <will.newton@linaro.org>
* sysdeps/unix/sysv/linux/aarch64/nptl/sysdep-cancel.h (PSEUDO):
Test the return value of the system call in the nocancel case.
Yvan Roux [Tue, 20 May 2014 12:45:22 +0000 (13:45 +0100)]
AArch64: Remove asm/ptrace.h inclusion in sys/user.h and sys/procfs.h
This patch fixes an issue observed by the Xen project, where including
signal.h exposes various PSR_MODE #defines. This is due to the usage
in sys/user.h and sys/procfs.h of the struct user_pt_regs and
user_fpsimd_state included via asm/ptrace.h. The namespace pollution
this inclusion introduce is already partially fixed with some #undef
of the PTRACE_* symbols, but other symbols like the PSR_MODE ones are
still present, and undefining them is not safe since a user can
include ptrace.h before user.h.
My proposition is to define the 2 structures we need in user.h and get
rid of the asm/ptrace.h inclusion.
Build and make check are clean on AArch64.
2014-05-20 Will Newton <will.newton@linaro.org>
Yvan Roux <yvan.roux@linaro.org>
* sysdeps/unix/sysv/linux/aarch64/sys/user.h: Remove unused
#include of asm/ptrace.h.
(PTRACE_GET_THREAD_AREA): Remove #undef.
(PTRACE_GETHBPREGS): Likewise.
(PTRACE_SETHBPREGS): Likewise.
(struct user_regs_struct): New structure.
(struct user_fpsimd_struct): New structure.
* sysdeps/unix/sysv/linux/aarch64/sys/procfs.h: Remove unused
#include of asm/ptrace.h and second #include of sys/user.h.
(PTRACE_GET_THREAD_AREA): Remove #undef.
(PTRACE_GETHBPREGS): Likewise.
(PTRACE_SETHBPREGS): Likewise.
(ELF_NGREG): Use new struct user_regs_struct.
(elf_fpregset_t): Use new struct user_fpsimd_struct.
Commit 7d92b78723848ae616709eb8f0191ea067025b18 [Fix ARM NAN fraction
bits.] removed all the bits set from NANFRAC macros and, when propagated
to libgcc, regressed gcc.dg/torture/builtin-math-7.c on soft-fp arm-eabi
targets, currently ARMv6-M (`-march=armv6-m -mthumb') only. This is
because when used to construct a NaN in the semi-raw mode, they now
build an infinity instead. Consequently operations such as (Inf - Inf)
now produce Inf rather than NaN. The change worked for the original
test case, posted with PR libgcc/60166, because division is made in the
canonical mode, where the quiet bit is set separately, from the fp
class.
This change brings the quiet bit back to these macros, making semi-raw
mode calculations produce the expected results again.
Joseph Myers [Fri, 16 May 2014 21:38:08 +0000 (21:38 +0000)]
Use existing makefile variables for dependencies on glibc libraries.
glibc's Makeconfig defines some variables such as $(libm) and $(libdl)
for linking with libraries built by glibc, and nptl/Makeconfig
(included by the toplevel Makeconfig) defines others such as
$(shared-thread-library).
In some places glibc's Makefiles use those variables when linking
against the relevant libraries, but in other places they hardcode the
location of the libraries in the build tree. This patch cleans up
various places to use the variables that already exist (in the case of
libm, replacing several duplicate definitions of a $(link-libm)
variable in subdirectory Makefiles). (It's not necessarily exactly
equivalent to what the existing code does - in particular,
$(shared-thread-library) includes libpthread_nonshared, but is
replacing places that just referred to libpthread.so. But I think
that change is desirable on the general principle of linking things as
close as possible to the way in which they would be linked with an
installed library, unless there is a clear reason not to do so.)
To support running tests with an installed copy of glibc without
needing the full build tree from when that copy was built, I think it
will be useful to use such variables more generally and systematically
- every time the rules for building a test refer to some file from the
build tree that's also installed by glibc, use a makefile variable so
that the installed-testing case can point those variables to installed
copies of the files. This patch just deals with straightforward cases
where such variables already exist.
It's quite possible some uses of $(shared-thread-library) should
actually be a new $(thread-library) variable that's set appropriately
in the --disable-shared case, if those uses would in fact work without
shared libraries. I didn't change the status quo that those cases
hardcode use of a shared library whether or not it's actually needed
(but other uses such as $(libm) and $(libdl) would now get the static
library if the shared library isn't built, when some previously
hardcoded use of the shared library - if they actually need shared
libraries, the test itself needs an enable-shared conditional anyway).
Tested x86_64.
* benchtests/Makefile
($(addprefix $(objpfx)bench-,$(bench-math))): Depend on $(libm),
not $(common-objpfx)math/libm.so.
($(addprefix $(objpfx)bench-,$(bench-pthread))): Depend on
$(shared-thread-library), not $(common-objpfx)nptl/libpthread.so.
* elf/Makefile ($(objpfx)noload): Depend on $(libdl), not
$(common-objpfx)dlfcn/libdl.so.
($(objpfx)tst-audit8): Depend on $(libm), not
$(common-objpfx)math/libm.so.
* malloc/Makefile ($(objpfx)libmemusage.so): Depend on $(libdl),
not $(common-objpfx)dlfcn/libdl.so.
* math/Makefile
($(addprefix $(objpfx),$(filter-out $(tests-static),$(tests)))):
Depend on $(libm), not $(objpfx)libm.so. Do not condition on
[$(build-shared) = yes].
($(objpfx)test-fenv-tls): Depend on $(shared-thread-library), not
$(common-objpfx)nptl/libpthread.so.
* misc/Makefile ($(objpfx)tst-tsearch): Depend on $(libm), not
$(common-objpfx)math/libm.so$(libm.so-version) or
$(common-objpfx)math/libm.a depending on [$(build-shared) = yes].
* nptl/Makefile ($(objpfx)tst-unload): Depend on $(libdl), not
$(common-objpfx)dlfcn/libdl.so.
* setjmp/Makefile (link-libm): Remove variable.
($(objpfx)tst-setjmp-fp): Depend on $(libm), not $(link-libm).
* stdio-common/Makefile (link-libm): Remove variable.
($(objpfx)tst-printf-round): Depend on $(libm), not $(link-libm).
* stdlib/Makefile (link-libm): Remove variable.
($(objpfx)bug-getcontext): Depend on $(libm), not $(link-libm).
($(objpfx)tst-strtod-round): Likewise.
($(objpfx)tst-tininess): Likewise.
($(objpfx)tst-strtod-underflow): Likewise.
($(objpfx)tst-strtod6): Likewise.
($(objpfx)tst-tls-atexit): Depend on $(shared-thread-library) and
$(libdl), not $(common-objpfx)nptl/libpthread.so and
$(common-objpfx)dlfcn/libdl.so.
If the fd refers to a terminal device, but not a pty master, the
TIOCGPTN ioctl returns with ENOTTY. This error is not caught, and the
possibly undefined buffer passed to ptsname_r is sent directly to the
stat64 syscall.
Fix this by using a fallback to the old method only if the TIOCGPTN
ioctl fails with EINVAL. This also fix the return value in that specific
case (it return ENOENT without this patch).
Also add tests to the ptsname_r function (and ptsname at the same time).
Note: this is Debian bug#741482, reported by Jakub Wilk <jwilk@debian.org>
Return EAI_AGAIN for AF_UNSPEC when herrno is TRY_AGAIN (BZ #16849)
getaddrinfo correctly returns EAI_AGAIN for AF_INET and AF_INET6
queries. For AF_UNSPEC however, an older change
(a682a1bf553b1efe4dbb03207fece5b719cec482) broke the check and due to
that the returned error was EAI_NONAME.
This patch fixes the check so that a non-authoritative not-found is
returned as EAI_AGAIN to the user instead of EAI_NONAME.
Joseph Myers [Wed, 14 May 2014 12:38:56 +0000 (12:38 +0000)]
Fix log1pl (LDBL_MAX) in FE_UPWARD mode (bug 16564).
Bug 16564 is spurious overflow of log1pl (LDBL_MAX) in FE_UPWARD mode,
resulting from log1pl adding 1 to its argument (for arguments not
close to 0), which overflows in that mode. This patch fixes this by
avoiding adding 1 to large arguments (precisely what counts as large
depends on the floating-point format).
Tested x86_64 and x86, and spot-checked log1pl tests on mips64 and
powerpc64.
[BZ #16564]
* sysdeps/i386/fpu/s_log1pl.S (__log1pl): Do not add 1 to positive
arguments with exponent 65 or above.
* sysdeps/ieee754/ldbl-128/s_log1pl.c (__log1pl): Do not add 1 to
arguments 0x1p113L or above.
* sysdeps/ieee754/ldbl-128ibm/s_log1pl.c (__log1pl): Do not add 1
to arguments 0x1p107L or above.
* sysdeps/x86_64/fpu/s_log1pl.S (__log1pl): Do not add 1 to
positive arguments with exponent 65 or above.
* math/auto-libm-test-in: Add more tests of log1p.
* math/auto-libm-test-out: Regenerated.
Joseph Myers [Wed, 14 May 2014 12:37:24 +0000 (12:37 +0000)]
Fix cacos (+Inf + finite*i) in round-downward mode (bug 16928).
According to C99/C11 Annex G, cacos applied to a value with real part
+Inf and finite imaginary part should produce a result with real part
+0. glibc wrongly produces a result with real part -0 in FE_DOWNWARD
mode. This patch fixes this by checking for zero results in the
relevant case of non-finite arguments (where there should never be a
result with -0 real part), and converts the tests of cacos to
ALL_RM_TEST.
Tested x86_64 and x86 and ulps updated accordingly.
[BZ #16928]
* math/s_cacos.c (__cacos): Ensure zero real part of result from
non-finite arguments is +0.
* math/s_cacosf.c (__cacosf): Likewise.
* math/s_cacosl.c (__cacosl): Likewise.
* math/libm-test.inc (cacos_test): Use ALL_RM_TEST.
* sysdeps/i386/fpu/libm-test-ulps: Update.
* sysdeps/x86_64/fpu/libm-test-ulps: Likewise.
Joseph Myers [Wed, 14 May 2014 12:35:40 +0000 (12:35 +0000)]
Fix acosh (1) in round-downward mode (bug 16927).
According to C99 and C11 Annex F, acosh (1) should be +0 in all
rounding modes. However, some implementations in glibc wrongly return
-0 in round-downward mode (which is what you get if you end up
computing log1p (-0), via 1 - 1 being -0 in round-downward mode).
This patch fixes the problem implementations, by correcting the test
for an exact 1 value in the ldbl-96 implementation to allow for the
explicit high bit of the mantissa, and by inserting fabs instructions
in the i386 implementations; tests of acosh are duly converted to
ALL_RM_TEST. I believe all the other sysdeps/ieee754 implementations
are already OK (I haven't checked the ia64 versions, but if buggy then
that will be obvious from the results of test runs after this patch is
in).
Tested x86_64 and x86 and ulps updated accordingly.
[BZ #16927]
* sysdeps/i386/fpu/e_acosh.S (__ieee754_acosh): Use fabs on x-1
value.
* sysdeps/i386/fpu/e_acoshf.S (__ieee754_acoshf): Likewise.
* sysdeps/i386/fpu/e_acoshl.S (__ieee754_acoshl): Likewise.
* sysdeps/ieee754/ldbl-96/e_acoshl.c (__ieee754_acoshl): Correct
for explicit high bit of mantissa when testing for argument equal
to 1.
* math/libm-test.inc (acosh_test): Use ALL_RM_TEST.
* sysdeps/i386/fpu/libm-test-ulps: Update.
* sysdeps/x86_64/fpu/libm-test-ulps: Likewise.
Joseph Myers [Wed, 14 May 2014 12:34:03 +0000 (12:34 +0000)]
Fix erf underflow handling near 0 (bug 16516).
Bug 16516 reports spurious underflows from erf (for all floating-point
types), when the result is close to underflowing but does not actually
underflow.
erf (x) is about (2/sqrt(pi))*x for x close to 0, so there are
subnormal arguments for which it does not underflow. The various
implementations do (x + efx*x) (for efx = 2/sqrt(pi) - 1), for greater
accuracy than if just using a single multiplication by an
approximation to 2/sqrt(pi) (effectively, this way there are a few
more bits in the approximation to 2/sqrt(pi)). This can introduce
underflows when efx*x underflows even though the final result does
not, so a scaled calculation with 8*efx is done in these cases - but 8
is not a big enough scale factor to avoid all such underflows. 16 is
(any underflows with a scale factor of 16 would only occur when the
final result underflows), so this patch changes the code to use that
factor. Rather than recomputing all the values of the efx8 variable,
it is removed, leaving it to the compiler's constant folding to
compute 16*efx. As such scaling can also lose underflows when the
final scaling down happens to be exact, appropriate checks are added
to ensure underflow exceptions occur when required in such cases.
Tested x86_64 and x86; no ulps updates needed. Also spot-checked for
powerpc32 and mips64 to verify the changes to the ldbl-128ibm and
ldbl-128 implementations.
[BZ #16516]
* sysdeps/ieee754/dbl-64/s_erf.c (efx8): Remove variable.
(__erf): Scale by 16 instead of 8 in potentially underflowing
case. Ensure exception if result actually underflows.
* sysdeps/ieee754/flt-32/s_erff.c (efx8): Remove variable.
(__erff): Scale by 16 instead of 8 in potentially underflowing
case. Ensure exception if result actually underflows.
* sysdeps/ieee754/ldbl-128/s_erfl.c: Include <float.h>.
(efx8): Remove variable.
(__erfl): Scale by 16 instead of 8 in potentially underflowing
case. Ensure exception if result actually underflows.
* sysdeps/ieee754/ldbl-128ibm/s_erfl.c: Include <float.h>.
(efx8): Remove variable.
(__erfl): Scale by 16 instead of 8 in potentially underflowing
case. Ensure exception if result actually underflows.
* sysdeps/ieee754/ldbl-96/s_erfl.c: Include <float.h>.
(efx8): Remove variable.
(__erfl): Scale by 16 instead of 8 in potentially underflowing
case. Ensure exception if result actually underflows.
* math/auto-libm-test-in: Add more tests of erf.
* math/auto-libm-test-out: Regenerated.