Since we have loaded address of PREINIT_FUNCTION into %rax, we can
avoid extra branch to PLT slot.
[BZ #19745]
* sysdeps/x86_64/crti.S (_init): Replace PREINIT_FUNCTION@PLT
with *%rax in call.
Since __libc_start_main is called very early, lazy binding isn't relevant
here. Use indirect branch via GOT to avoid extra branch to PLT slot.
[BZ #19745]
* sysdeps/x86_64/start.S (_start): __libc_start_main@PLT
with *__libc_start_main@GOTPCREL(%rip) in call.
Chek Fast_Unaligned_Load, instead of Slow_BSF, and also check for
Fast_Copy_Backward to enable __memcpy_ssse3_back. Existing selection
order is updated with following selection order:
1. __memcpy_avx_unaligned if AVX_Fast_Unaligned_Load bit is set.
2. __memcpy_sse2_unaligned if Fast_Unaligned_Load bit is set.
3. __memcpy_sse2 if SSSE3 isn't available.
4. __memcpy_ssse3_back if Fast_Copy_Backward bit it set.
5. __memcpy_ssse3
[BZ #18880]
* sysdeps/x86_64/multiarch/memcpy.S: Check Fast_Unaligned_Load,
instead of Slow_BSF, and also check for Fast_Copy_Backward to
enable __memcpy_ssse3_back.
We should turn on bit_Prefer_MAP_32BIT_EXEC in EXTRA_LD_ENVVARS without
overriding other bits.
[BZ #19758]
* sysdeps/unix/sysv/linux/x86_64/64/dl-librecon.h
(EXTRA_LD_ENVVARS): Or bit_Prefer_MAP_32BIT_EXEC.
No need to compile x86_64 _mcount.S with -pg. We can just copy the
normal static object.
* gmon/Makefile (noprof): Add $(sysdep_noprof).
* sysdeps/x86_64/Makefile (sysdep_noprof): Add _mcount.
Since __mcount_internal and __sigjmp_save are internal to x86-64 libc.so:
3532: 0000000000104530 289 FUNC LOCAL DEFAULT 13 __mcount_internal
3391: 0000000000034170 38 FUNC LOCAL DEFAULT 13 __sigjmp_save
they can be called directly without PLT.
* sysdeps/x86_64/_mcount.S (C_LABEL(_mcount)): Call
__mcount_internal directly.
(C_LABEL(__fentry__)): Likewise.
* sysdeps/x86_64/setjmp.S __sigsetjmp): Call __sigjmp_save
directly.
Now we require Linux 3.2 or later kernel headers everywhere, the
configure test for <linux/fanotify.h> is obsolete; this patch removes
it.
Tested for x86_64.
* sysdeps/unix/sysv/linux/configure.ac (linux/fanotify.h): Do not
test for header.
* sysdeps/unix/sysv/linux/configure: Regenerated.
* config.h.in (HAVE_LINUX_FANOTIFY_H): Remove #undef.
* sysdeps/unix/sysv/linux/tst-fanotify.c [!HAVE_LINUX_FANOTIFY_H]:
Remove conditional code.
[HAVE_LINUX_FANOTIFY_H]: Make code unconditional.
In <https://sourceware.org/ml/libc-alpha/2016-01/msg00885.html> I
proposed a minimum Linux kernel version of 3.2 for glibc 2.24, since
Linux 2.6.32 has reached EOL.
In the discussion in February, some concerns were expressed about
compatibility with OpenVZ containers. It's not clear that these are
real issues, given OpenVZ backporting kernel features and faking the
kernel version for guest software, as discussed in
<https://sourceware.org/ml/libc-alpha/2016-02/msg00278.html>. It's
also not clear that supporting running GNU/Linux distributions from
late 2016 (at the earliest) on a kernel series from 2009 is a sensible
expectation. However, as an interim step, this patch increases the
requirement everywhere except x86 / x86_64 (since the controversy was
only about those architectures); the special caveats and settings can
easily be removed later when we're ready to increase the requirements
on x86 / x86_64 (and if someone would like to raise the issue on LWN
as suggested in the previous discussion, that would be welcome). 3.2
kernel headers are required everywhere by this patch.
(x32 already requires 3.4 or later, so is unaffected by this patch.)
As usual for such a change, this patch only changes the configure
scripts and associated documentation. The intent is to follow up with
removal of dead __LINUX_KERNEL_VERSION conditionals. Each __ASSUME_*
or other macro that becomes dead can then be removed independently.
Tested for x86_64 and x86.
* sysdeps/unix/sysv/linux/configure.ac (LIBC_LINUX_VERSION):
Define to 3.2.0.
(arch_minimum_kernel): Likewise.
* sysdeps/unix/sysv/linux/configure: Regenerated.
* sysdeps/unix/sysv/linux/i386/configure.ac (arch_minimum_kernel):
Define to 2.6.32.
* sysdeps/unix/sysv/linux/i386/configure: Regenerated.
* sysdeps/unix/sysv/linux/x86_64/64/configure.ac
(arch_minimum_kernel): Define to 2.6.32.
* sysdeps/unix/sysv/linux/x86_64/64/configure: Regenerated.
* README: Document Linux 3.2 requirement.
* manual/install.texi (Linux): Document Linux 3.2 headers
requirement.
* INSTALL: Regenerated.
Due to GCC bug:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=58066
__tls_get_addr may be called with 8-byte stack alignment. Although
this bug has been fixed in GCC 4.9.4, 5.3 and 6, we can't assume
that stack will be always aligned at 16 bytes. Since SSE optimized
memory/string functions with aligned SSE register load and store are
used in the dynamic linker, we must set DL_RUNTIME_UNALIGNED_VEC_SIZE
to 8 so that _dl_runtime_resolve_sse will align the stack before
calling _dl_fixup:
Dump of assembler code for function _dl_runtime_resolve_sse:
0x00007ffff7deea90 <+0>: push %rbx
0x00007ffff7deea91 <+1>: mov %rsp,%rbx
0x00007ffff7deea94 <+4>: and $0xfffffffffffffff0,%rsp
^^^^^^^^^^^ Align stack to 16 bytes
0x00007ffff7deea98 <+8>: sub $0x100,%rsp
0x00007ffff7deea9f <+15>: mov %rax,0xc0(%rsp)
0x00007ffff7deeaa7 <+23>: mov %rcx,0xc8(%rsp)
0x00007ffff7deeaaf <+31>: mov %rdx,0xd0(%rsp)
0x00007ffff7deeab7 <+39>: mov %rsi,0xd8(%rsp)
0x00007ffff7deeabf <+47>: mov %rdi,0xe0(%rsp)
0x00007ffff7deeac7 <+55>: mov %r8,0xe8(%rsp)
0x00007ffff7deeacf <+63>: mov %r9,0xf0(%rsp)
0x00007ffff7deead7 <+71>: movaps %xmm0,(%rsp)
0x00007ffff7deeadb <+75>: movaps %xmm1,0x10(%rsp)
0x00007ffff7deeae0 <+80>: movaps %xmm2,0x20(%rsp)
0x00007ffff7deeae5 <+85>: movaps %xmm3,0x30(%rsp)
0x00007ffff7deeaea <+90>: movaps %xmm4,0x40(%rsp)
0x00007ffff7deeaef <+95>: movaps %xmm5,0x50(%rsp)
0x00007ffff7deeaf4 <+100>: movaps %xmm6,0x60(%rsp)
0x00007ffff7deeaf9 <+105>: movaps %xmm7,0x70(%rsp)
[BZ #19679]
* sysdeps/x86_64/dl-trampoline.S (DL_RUNIME_UNALIGNED_VEC_SIZE):
Renamed to ...
(DL_RUNTIME_UNALIGNED_VEC_SIZE): This. Set to 8.
(DL_RUNIME_RESOLVE_REALIGN_STACK): Renamed to ...
(DL_RUNTIME_RESOLVE_REALIGN_STACK): This. Updated.
(DL_RUNIME_RESOLVE_REALIGN_STACK): Renamed to ...
(DL_RUNTIME_RESOLVE_REALIGN_STACK): This.
* sysdeps/x86_64/dl-trampoline.h
(DL_RUNIME_RESOLVE_REALIGN_STACK): Renamed to ...
(DL_RUNTIME_RESOLVE_REALIGN_STACK): This.
The ldbl-128ibm implementation of nextafterl / nexttowardl returns -0
in FE_DOWNWARD mode when taking the next value below the least
positive subnormal, when it should return +0. This patch fixes it to
check explicitly for this case.
Tested for powerpc.
[BZ #19678]
* sysdeps/ieee754/ldbl-128ibm/s_nextafterl.c (__nextafterl):
Ensure +0.0 is returned when taking the next value below the least
positive value.
The ldbl-128ibm implementation of powl has some problems in the case
of overflow or underflow, which are mainly visible in non-default
rounding modes.
* When overflow or underflow is detected early, the correct sign of an
overflowing or underflowing result is not allowed for. This is
mostly hidden in the default rounding mode by the errno-setting
wrappers recomputing the result (except in non-default
error-handling modes such as -lieee), but visible in other rounding
modes where a result that is not zero or infinity causes the
wrappers not to do the recomputation.
* The final scaling is done before the sign is incorporated in the
result, but should be done afterwards for correct overflowing and
underflowing results in directed rounding modes.
This patch fixes those problems. Tested for powerpc.
[BZ #19674]
* sysdeps/ieee754/ldbl-128ibm/e_powl.c (__ieee754_powl): Include
sign in overflowing and underflowing results when overflow or
underflow is detected early. Include sign in result before rather
than after scaling.
The ldbl-128ibm implementations of remainderl and remquol have logic
resulting in incorrect tests for equality of the absolute values of
the arguments. Equality is tested based on the integer
representations of the high and low parts, with the sign bit masked
off the high part - but when this changes the sign of the high part,
the sign of the low part needs to be changed as well, and failure to
do this means arguments are wrongly treated as equal when they are
not.
This patch fixes the logic to adjust signs of low parts as needed.
Tested for powerpc.
[BZ #19603]
* sysdeps/ieee754/ldbl-128ibm/e_remainderl.c
(__ieee754_remainderl): Adjust sign of integer version of low part
when taking absolute value of high part.
* sysdeps/ieee754/ldbl-128ibm/s_remquol.c (__remquol): Likewise.
* math/libm-test.inc (remainder_test_data): Add another test.
(remquo_test_data): Likewise.
The ldbl-128ibm implementation of fmodl has logic to detect when the
first argument has absolute value less than or equal to the second.
This logic is only correct for nonzero low parts; if the high parts
are equal and the low parts are zero, then the signs of the low parts
(which have no semantic effect on the value of the long double number)
can result in equal values being wrongly treated as unequal, and an
incorrect result being returned from fmodl. This patch fixes this by
checking for the case of zero low parts.
Although this does show up in tests from libm-test.inc (both tests of
fmodl, and, indirectly, of remainderl / dreml), the dependence on
non-semantic zero low parts means that test shouldn't be expected to
reproduce it reliably; thus, this patch adds a standalone test that
sets up affected values using unions.
Tested for powerpc.
[BZ #19602]
* sysdeps/ieee754/ldbl-128ibm/e_fmodl.c (__ieee754_fmodl): Handle
equal high parts and both low parts zero specially.
* sysdeps/ieee754/ldbl-128ibm/test-fmodl-ldbl-128ibm.c: New test.
* sysdeps/ieee754/ldbl-128ibm/Makefile [$(subdir) = math] (tests):
Add test-fmodl-ldbl-128ibm.
The ldbl-128ibm implementation of fmodl has completely bogus logic for
subnormal results (in this context, that means results for which the
result is in the subnormal range for double, not results with absolute
value below LDBL_MIN), based on code used for ldbl-128 that is correct
in that case but incorrect in the ldbl-128ibm use. This patch fixes
it to convert the mantissa into the correct form expected by
ldbl_insert_mantissa, removing the other cases of the code that were
incorrect and in one case unreachable for ldbl-128ibm. A correct
exponent value is then passed to ldbl_insert_mantissa to reflect the
shifted result.
Tested for powerpc.
[BZ #19595]
* sysdeps/ieee754/ldbl-128ibm/e_fmodl.c (__ieee754_fmodl): Use
common logic for all cases of shifting subnormal results. Do not
insert sign bit in shifted mantissa. Always pass -1023 as biased
exponent to ldbl_insert_mantissa in subnormal case.
The ldbl-128ibm implementation of roundl is only correct in
round-to-nearest mode (in other modes, there are incorrect results and
overflow exceptions in some cases). This patch reimplements it along
the lines used for floorl, ceill and truncl, using __round on the high
part, and on the low part if the high part is an integer, and then
adjusting in the cases where this is incorrect.
Tested for powerpc.
[BZ #19594]
* sysdeps/ieee754/ldbl-128ibm/s_roundl.c (__roundl): Use __round
on high and low parts then adjust result and use
ldbl_canonicalize_int if needed.
The ldbl-128ibm implementation of truncl is only correct in
round-to-nearest mode (in other modes, there are incorrect results and
overflow exceptions in some cases). It is also unnecessarily
complicated, rounding both high and low parts to the nearest integer
and then adjusting for the semantics of trunc, when it seems more
natural to take the truncation of the high part (__trunc optimized
inline versions can be used), and the floor or ceiling of the low part
(depending on the sign of the high part) if the high part is an
integer, as was done for floorl and ceill. This patch makes it use
that simpler approach.
Tested for powerpc.
[BZ #19593]
* sysdeps/ieee754/ldbl-128ibm/s_truncl.c (__truncl): Use __trunc
on high part and __floor or __ceil on low part then use
ldbl_canonicalize_int if needed.
The ldbl-128ibm implementation of ceill is only correct in
round-to-nearest mode (in other modes, there are incorrect results and
overflow exceptions in some cases). It is also unnecessarily
complicated, rounding both high and low parts to the nearest integer
and then adjusting for the semantics of ceil, when it seems more
natural to take the ceiling of the high part (__ceil optimized inline
versions can be used), and that of the low part if the high part is an
integer, as was done for floorl. This patch makes it use that simpler
approach.
Tested for powerpc.
[BZ #19592]
* sysdeps/ieee754/ldbl-128ibm/s_ceill.c (__ceill): Use __ceil on
high and low parts then use ldbl_canonicalize_int if needed.
The ldbl-128ibm implementation of floorl is only correct in
round-to-nearest mode (in other modes, there are incorrect results and
overflow exceptions in some cases going beyond the incorrect signs of
zero results noted in bug 17899). It is also unnecessarily
complicated, rounding both high and low parts to the nearest integer
and then adjusting for the semantics of floor, when it seems more
natural to take the floor of the high part (__floor optimized inline
versions can be used), and that of the low part if the high part is an
integer. This patch makes it use that simpler approach, with a
canonicalization that works in all rounding modes (given that the only
way the result can be noncanonical is if taking the floor of a
negative noninteger low part increased its exponent).
Tested for powerpc, where over a thousand failures are removed from
test-ldouble.out (floorl problems affect many powl tests).
[BZ #17899]
* sysdeps/ieee754/ldbl-128ibm/math_ldbl.h (ldbl_canonicalize_int):
New function.
* sysdeps/ieee754/ldbl-128ibm/s_floorl.c (__floorl): Use __floor
on high and low parts then use ldbl_canonicalize_int if needed.
As discussed in
https://sourceware.org/ml/libc-alpha/2015-10/msg00403.html
the setting of _STRING_ARCH_unaligned currently controls the external
GLIBC ABI as well as selecting the use of unaligned accesses withing
GLIBC.
Since _STRING_ARCH_unaligned was recently changed for AArch64, this
would potentially break the ABI in GLIBC 2.23, so split the uses and add
_STRING_INLINE_unaligned to select the string ABI. This setting must be
fixed for each target, while _STRING_ARCH_unaligned may be changed from
release to release. _STRING_ARCH_unaligned is used unconditionally in
glibc. But <bits/string.h>, which defines _STRING_ARCH_unaligned, isn't
included with -Os. Since _STRING_ARCH_unaligned is internal to glibc and
may change between glibc releases, it should be made private to glibc.
_STRING_ARCH_unaligned should defined in the new string_private.h heade
file which is included unconditionally from internal <string.h> for glibc
build.
[BZ #19462]
* bits/string.h (_STRING_ARCH_unaligned): Renamed to ...
(_STRING_INLINE_unaligned): This.
* include/string.h: Include <string_private.h>.
* string/bits/string2.h: Replace _STRING_ARCH_unaligned with
_STRING_INLINE_unaligned.
* sysdeps/aarch64/bits/string.h (_STRING_ARCH_unaligned): Removed.
(_STRING_INLINE_unaligned): New.
* sysdeps/aarch64/string_private.h: New file.
* sysdeps/generic/string_private.h: Likewise.
* sysdeps/m68k/m680x0/m68020/string_private.h: Likewise.
* sysdeps/s390/string_private.h: Likewise.
* sysdeps/x86/string_private.h: Likewise.
* sysdeps/m68k/m680x0/m68020/bits/string.h
(_STRING_ARCH_unaligned): Renamed to ...
(_STRING_INLINE_unaligned): This.
* sysdeps/s390/bits/string.h (_STRING_ARCH_unaligned): Renamed
to ...
(_STRING_INLINE_unaligned): This.
* sysdeps/sparc/bits/string.h (_STRING_ARCH_unaligned): Renamed
to ...
(_STRING_INLINE_unaligned): This.
* sysdeps/x86/bits/string.h (_STRING_ARCH_unaligned): Renamed
to ...
(_STRING_INLINE_unaligned): This.
Since libmvec_nonshared.a may be linked into shared objects, ALIAS_IMPL
should use PIC relocation.
[BZ #19590]
* sysdeps/x86_64/fpu/svml_finite_alias.S (ALIAS_IMPL): Use PIC
relocation.
The handling of negative offsets in MIPS mmap is inconsistent with
other architectures, as shown by failure of the test
posix/tst-mmap-offset for o32 and n32. The MIPS mmap syscall uses a
signed argument and does a signed arithmetic shift on it, whereas the
glibc semantics expected by that test are for the offset to be
considered as a large positive offset. This patch makes MIPS
consistent with other architectures as far as possible by using the
mmap2 syscall on o32 (#including the generic implementation), and
making mmap not an alias for mmap64 for n32, with a custom
implementation for n32 that zero-extends the offset argument to 64-bit
before calling the mmap syscall.
Tested for MIPS64 (o32, n32, n64).
[BZ #19550]
* sysdeps/unix/sysv/linux/mips/mips32/mmap.c: New file.
* sysdeps/unix/sysv/linux/mips/mips64/mmap64.c: Move to ....
* sysdeps/unix/sysv/linux/mips/mips64/n64/mmap64.c: ... here.
* sysdeps/unix/sysv/linux/mips/mips64/n32/mmap.c: New file.
* sysdeps/unix/sysv/linux/mips/mips64/n32/syscalls.list (mmap64):
New syscall entry.
* sysdeps/unix/sysv/linux/mips/mips64/n64/syscalls.list (mmap):
New syscall entry.
* sysdeps/unix/sysv/linux/mips/mips64/syscalls.list (mmap): Remove
syscall entry.
The MIPS memcpy optimizations at
<https://sourceware.org/ml/libc-alpha/2015-10/msg00597.html>
introduced a bug causing many string function tests to fail with
segfaults for n32 and n64:
FAIL: string/stratcliff
FAIL: string/test-bcopy
FAIL: string/test-memccpy
FAIL: string/test-memcmp
FAIL: string/test-memcpy
FAIL: string/test-memmove
FAIL: string/test-mempcpy
FAIL: string/test-stpncpy
FAIL: string/test-strncmp
FAIL: string/test-strncpy
(Some failures in other directories could also be caused by this bug.)
The problem is that after the check for whether a word of input is
left that can be copied as a word before moving to byte copies, a load
can occur in the branch delay slot, resulting in a segfault if we are
at the end of a page and the following page is unmapped. I don't see
how this would have passed the tests as reported in the original patch
posting (different kernel configurations affecting the code setting up
unmapped pages, maybe?), since the tests in question don't appear to
have changed recently.
This patch moves a later instruction into the delay slot, as suggested
at <https://sourceware.org/ml/libc-alpha/2016-01/msg00584.html>.
Tested for n32 and n64.
2016-01-28 Steve Ellcey <sellcey@imgtec.com>
Joseph Myers <joseph@codesourcery.com>
* sysdeps/mips/memcpy.S (MEMCPY_NAME) [USE_DOUBLE]: Avoid word
load in branch delay slot when less than a word of input left.
Complement the addition of the required kernel support, present upstream
as from commit 2b5e869ecfcb3112f7e1267cb0328f3ff6d49b18 ("MIPS: ELF:
Interpret the NAN2008 file header flag") and released with Linux 4.5-rc1
on Jan 24th, 2016.
* sysdeps/unix/sysv/linux/mips/configure.ac: Set
`arch_minimum_kernel' to 4.5.0 if 2008 NaN encoding is used.
* sysdeps/unix/sysv/linux/mips/configure: Regenerate.
The changes to restrict implementation-namespace symbol aliases such
as __finitel to compat symbols used code for __finitel in libm
analogous to that for __finitel in libc. However, the versions for
the two symbols are actually different, GLIBC_2.0 in libc and
GLIBC_2.1 in libm. This patch fixes the handling of the libm compat
symbol.
Tested for mips (o32), where it fixes an ABI test failure.
* sysdeps/ieee754/dbl-64/s_finite.c
[NO_LONG_DOUBLE && LDBL_CLASSIFY_COMPAT] (__finitel): Define
compat symbol at version GLIBC_2_1 and use GLIBC_2_1 in
SHLIB_COMPAT condition for libm, not GLIBC_2_0.
* sysdeps/ieee754/dbl-64/wordsize-64/s_finite.c
[NO_LONG_DOUBLE && LDBL_CLASSIFY_COMPAT] (__finitel): Likewise.
Testing for powerpc-nofpu showed that localplt.data was out of date.
Two new soft-fp functions showed up in the list: __gtsf2 and
__unordsf2; this patch adds these as optional. __signbit and
__signbitl no longer appear as local PLT entries; given the move to
__builtin_signbit* for all GCC versions supported for building glibc
(and given the use of the type-generic signbit macro within glibc),
those can safely be removed from the list, which this patch does.
Tested for powerpc-nofpu.
* sysdeps/unix/sysv/linux/powerpc/powerpc32/nofpu/localplt.data
(__gtsf2): Add as optional for libc.so.
(__unordsf2): Likewise.
(__signbit): Remove for libc.so.
(__signbitl): Likewise.
On running tests after from-scratch ulps regeneration, I found that
some libm tests failed with ulps in excess of those recorded in the
from-scratch regeneration, which should never happen unless those ulps
exceed the limit on ulps that can go in libm-test-ulps files.
Failure: Test: atan2_upward (inf, -inf)
Result:
is: 2.35619498e+00 0x1.2d97ccp+1
should be: 2.35619450e+00 0x1.2d97c8p+1
difference: 4.76837159e-07 0x1.000000p-21
ulp : 2.0000
max.ulp : 1.0000
Maximal error of `atan2_upward'
is : 2 ulp
accepted: 1 ulp
Failure: Test: carg_upward (-inf + inf i)
Result:
is: 2.35619498e+00 0x1.2d97ccp+1
should be: 2.35619450e+00 0x1.2d97c8p+1
difference: 4.76837159e-07 0x1.000000p-21
ulp : 2.0000
max.ulp : 1.0000
Maximal error of `carg_upward'
is : 2 ulp
accepted: 1 ulp
The problem comes from the addition of tests for the finite-math-only
versions of libm functions. Those tests share ulps with the default
function variants. make regen-ulps runs the default tests before the
finite-math-only tests, concatenating the resulting ulps before
feeding them to gen-libm-test.pl to generate a new libm-test-ulps
file. But gen-libm-test.pl always takes the last ulps value given for
any (function, type) pair. So, if the largest ulps for a function
come from non-finite inputs, a from-scratch regeneration loses those
ulps.
This patch fixes gen-libm-test.pl, in the case where there are
multiple ulps values for a (function, type) pair - which can only
happen as part of a regeneration - to take the largest ulps value
rather than the last one.
Tested for ARM / MIPS / powerpc-nofpu.
* math/gen-libm-test.pl (parse_ulps): Do not reduce
already-recorded ulps.
* sysdeps/arm/libm-test-ulps: Regenerated.
* sysdeps/mips/mips32/libm-test-ulps: Likewise.
* sysdeps/mips/mips64/libm-test-ulps: Likewise.
* sysdeps/powerpc/nofpu/libm-test-ulps: Likewise.
I've regenerated ulps from scratch for s390/s390x.
All math testcases are passing afterwards.
ChangeLog:
* sysdeps/s390/fpu/libm-test-ulps: Regenerated.
