__need_getopt is misnamed; what it really means is "we want only the
getopt features specified in POSIX, not the GNU extensions". Because
this code is shared with gnulib, it winds up being cleanest to split
getopt.h into *four* headers. getopt_core.h and getopt_ext.h will
be shared with gnulib, getopt_posix.h will be just for glibc, and
each project will have its own copy of getopt.h.
* posix/bits/getopt_core.h, posix/bits/getopt_ext.h:
New files, intended to be shared with gnulib.
* posix/bits/getopt_posix.h:
New file, not intended to be shared with gnulib.
* posix/getopt.h: Now just includes features.h,
bits/getopt_core.h, and bits/getopt_ext.h. Will
no longer be shared with gnulib.
* include/bits/getopt_core.h, include/bits/getopt_ext.h
* include/bits/getopt_posix.h: New wrappers.
* posix/Makefile: Install new headers.
* posix/unistd.h, libio/stdio.h:
Include bits/getopt_posix.h instead of getopt.h.
gnulib now has annotations on at least some functions to cater to
compilation with -Wunused-parameter. In order to follow suit cleanly,
I've added to libc-symbols.h some of the _GL_* macros that
gnulib-common.m4 puts into config.h. (I don't think they belong in
sys/cdefs.h, at least not without further thought.)
At this point all gnulib-side changes to getopt.c have been merged.
* include/libc-symbols.h: For gnulib compatibility, define
_GL_UNUSED, _GL_UNUSED_LABEL, _GL_ATTRIBUTE_PURE, and
_GL_ATTRIBUTE_CONST.
* posix/getopt.c (_getopt_initialize): Mark argc and argv
parameters with _GL_UNUSED.
In one place, glibc's getopt uses alloca to construct a linked list of
possibilities for an "ambiguous" long option. In gnulib, malloc
should be used instead. Providing for both cases complicates things a
fair bit. Instead of merging straight across, therefore, I have
chosen to rewrite it using a boolean vector instead of a linked list.
There is then only one allocation that might need freeing; in glibc it
can honor __libc_use_alloca as usual, and in gnulib we define
__libc_use_alloca to always be false, so we don't need ifdefs in the
middle of the function. This should also be slightly more efficient
in the normal case of long options being fully spelled out -- I think
most people aren't even aware they _can_ sometimes abbreviate long
options.
One interesting consequence is that the list of possibilities is now
printed in exactly the order they appear in the list of long options,
instead of the first possibility being shuffled to the end. This
shouldn't be a big deal but it did break one test that relied on the
exact text of this error message.
(The reason the previous patch was "in aid of" merging from gnulib is
I didn't want to have to make this change in two places.)
(The patch looks bigger than it really is because there's a fair bit
of reindentation and code rearrangement.)
* posix/getopt.c: When used standalone, define __libc_use_alloca
as always false and alloca to abort if called.
(process_long_option): Rewrite handling of ambiguous long options
to use a single boolean vector, not a linked list; use
__libc_use_alloca to decide whether to allocate this using alloca.
* posix/tst-getopt_long1.c: Adjust text of expected error message.
There were two copies of the bulk of the code to handle long options.
Now there is only one. (Yes, this is in aid of merging from gnulib.)
The change to bug-getopt4.c clarifies the error messages when the test
fails.
* posix/getopt.c (process_long_option): New function split out
from _getopt_internal_r.
(_getopt_internal_r): Replace both copies of the long-option
processing code with calls to process_long_option.
* posix/bug-getopt4.c (one_test): Print argv[0] in error messages.
(do_test): Differentiate argv[0] in the two subtests.
_getopt_data.__posixly_correct is completely redundant to
_getopt_data.__ordering, and some work that logically belongs in
_getopt_initialize was being done by _getopt_internal_r, making the
code harder to understand.
As a side effect, getenv will no longer be called if the first
character of the options string is '+' or '-', which is probably a
Good Thing. (Perhaps we should have a flag character that
specifically asks for the permutation behavior?)
* posix/getopt_int.h (_getopt_data): Remove __posixly_correct field.
* posix/getopt.c (_getopt_internal_r): Move some initialization code...
(_getopt_initialize): ...here. Don't set d->__posixly_correct.
For standards compliance, getopt, getopt_long, and getopt_long_only in
glibc have to take 'char *const *argv' even though they can mutate the
array. gnulib has tried to clean this up as much as possible: all the
internal functions use 'char **argv', and when used standalone, so do
getopt_long and getopt_long_only.
Also brought over are __nonnull annotations, corrections to documentation,
and apparently it is no longer necessary to worry about conflicting
prototypes for getopt. The macroification of the definitions of
getopt and __posix_getopt goes beyond what is currently in gnulib.
At this point getopt1.c and getopt_int.h are identical to their gnulib
versions.
* posix/getopt.h: Add backup definition of __nonnull for
consistency with gnulib. Define __getopt_argv_const to const
if not already defined.
(getopt): Update doc comment from gnulib. Prototype
unconditionally. Add __nonnull annotation.
(__posix_getopt): Add __nonnull annotation.
(getopt_long, getopt_long_only): Use __getopt_argv_const in
prototypes for consistency with gnulib. Add __nonnull
annotations.
* posix/getopt.c (_getopt_initialize, _getopt_internal_r)
(getopt_internal): Change 'argv' argument to type 'char **'.
Remove now-unnecessary casts.
(getopt, __posix_getopt): Eliminate repetition with a macro.
Cast 'argv' to 'char **' when calling _getopt_internal.
* posix/getopt1.c (getopt_long, getopt_long_only):
Use __getopt_argv_const for consistency with gnulib.
Cast 'argv' to 'char **' when calling _getopt_internal.
(_getopt_long_r, _getopt_long_only_r):
Change 'argv' argument to type 'char **'.
(main): Constify 'long_options'.
* posix/getopt_int.h (getopt_internal, _getopt_internal_r)
(_getopt_long_r, _getopt_long_only_r):
Change 'argv' argument to type 'char **'.
getopt can print a whole bunch of error messages, and when used
standalone (from gnulib) it uses fprintf to do that. But fprintf is a
cancellation point and getopt isn't, and also applying fprintf to a
stream in wide-character mode is not allowed.
glibc has an internal function called __fxprintf that writes a narrow
format string to a stream regardless of mode, but it only handles
ASCII format strings, and it's still a cancellation point. getopt's
messages are translated, so they might not be ASCII. So getopt has an
error message to an asprintf buffer, monkeys with internal flag bits
on stderr to disable cancellation, and then calls
__fxprintf(stderr, "%s", buffer). There isn't even a helper function,
the code is duplicated every time.
This patch fixes __fxprintf to handle arbitrary multibyte format
strings, and adds a variant __fxprintf_nocancel that does the same
thing but also isn't a cancellation point. (It still _works_ by
monkeying with internal flag bits on the FILE, but that's not really a
layering violation for code in stdio-common.) All of the #ifdef _LIBC
blocks can then be reduced to their standalone versions with a little
help from some macros at the top of the file.
I also wrote a test case to verify that getopt really isn't a
cancellation point, and I'm glad I did, because it found two bugs, one
of which wasn't even to do with cancellation (see previous patch).
* stdio-common/fxprintf.c (__fxprintf_nocancel): New function.
(locked_vfxprintf): New helper function. Handle arbitrary
multibyte strings, not just ASCII.
* include/stdio.h: Declare __fxprintf_nocancel.
* posix/getopt.c: When _LIBC is defined, define fprintf to
__fxprintf_nocancel, flockfile to _IO_flockfile, and
funlockfile to _IO_funlockfile. When neither _LIBC nor
_POSIX_THREAD_SAFE_FUNCTIONS is defined, define flockfile and
funlockfile as no-ops. (_getopt_internal_r): Remove all
internal #ifdef _LIBC blocks; the standalone error-printing
code can now be used for libc as well. Add an
flockfile/funlockfile pair around one case where the error
message is printed in several chunks. Don't use fputc.
* posix/tst-getopt-cancel.c: New test.
* posix/Makefile: Run it.
getopt_long contains an undocumented (AFAICT) feature in which, if you
put "W;" in the short-options list, then '-W foo' and '-Wfoo' are
treated as equivalent to '--foo'. This is implemented with a partial
second copy of the code for handling long options, and that code
increments optind one too many times when recovering from an ambiguous
abbreviated option, which can cause the main loop to walk past the end
of argv and crash.
I discovered this while writing a test case that tries to exercise all
of getopt's error reporting paths; I wouldn't be surprised to learn
that this feature is never used by real applications.
* posix/getopt.c (_getopt_internal_r): Don't increment
d->optind a second time when reporting ambiguous -W options.
This covers changes with little or no consequences when the code is
used in glibc.
* posix/getopt_int.h: Include getopt.h.
Use impl-namespace names for all arguments to _getopt_internal and
_getopt_internal_r.
Declare __ordering enum outside the struct.
Harmonize comments with gnulib.
* posix/getopt1.c: Simplify #ifdeffage at top of file. Remove
ELIDE_CODE logic entirely. Move inclusion of stdlib.h to
#ifdef TEST block and make unconditional. Do not define NULL.
* posix/getopt.c: Partial merge from gnulib, covering the
initial includes and global declarations, commentary, and
a couple of semantically-neutral code changes.
I'm not sure whether this is official GNU style now, but `...' quotes
haven't looked properly balanced in most people's terminal fonts since
2001ish? and gnulib has chosen to switch over to '...' quotes.
I'm merging this separately from the other changes in gnulib because
it's very mechanical.
* posix/getopt.c, posix/getopt.h, posix/getopt1.c, posix/getopt_int.h:
Use '...' instead of `...' for quotation marks inside
comments and strings.
glibc's implementation of getopt includes code to parse an environment
variable named _XXX_GNU_nonoption_argv_flags_ (where XXX is the
current process's PID in decimal); but all of it has been #ifdefed out
since 2001, with no official way to turn it back on.
According to commentary in our config.h.in, bash version 2.0 set this
environment variable to indicate argv elements that were the result of
glob expansion and therefore should not be treated as options, but the
feature was "disabled later" because "it caused problems". According
to bash's CHANGES file, "later" was release 2.01; it gives no more
detail about what the problems were.
Version 2.0 of bash was released on the last day of 1996, and version
2.01 in June of 1997. Twenty years later, I think it is safe to
assume that this environment variable isn't coming back.
* config.h.in (USE_NONOPTION_FLAGS): Remove.
* csu/init-first.c: Remove all #ifdef USE_NONOPTION_FLAGS blocks.
* sysdeps/mach/hurd/i386/init-first.c: Likewise.
* posix/getopt_int.h: Likewise.
* posix/getopt.c: Likewise. Also remove SWAP_FLAGS and the
__libc_argc and __libc_argv externs, which were only used by
#ifdef USE_NONOPTION_FLAGS blocks.
* posix/getopt_init.c: Remove file.
* posix/Makefile (routines): Remove getopt_init.
* include/getopt.h: Don't declare __getopt_initialize_environment.
* manual/getopt.texi: Remove mention of USE_NONOPTION_FLAGS in
a comment.
With the removal of divdi3 object from sparcv9-linux-gnu build, its
definition came from libgcc and its functions internall calls .udiv.
Since glibc also exports these symbols for compatibility reasons, it
will end up creating PLT calls internally in libc.so.
To avoid it, this patch uses the linker option --wrap to replace all
the internal libc.so .udiv calls to the wrapper __wrap_.udiv. Along
with strong alias in the udiv implementations, it makes linker do
local calls.
Checked on sparcv9-linux-gnu.
* sysdeps/sparc/sparc32/Makefile (libc.so-gnulib): New rule.
* sysdeps/sparc/sparc32/sparcv8/udiv.S (.udiv): Make a strong_alias
to __wrap_.udiv.
* sysdeps/sparc/sparc32/sparcv9/udiv.S (.udiv): Likewise.
* sysdeps/sparc/sparc32/udiv.S (.udiv): Likewise.
As noted in [1], divdi3 object is only exported in a handful ABIs
(i386, m68k, powerpc32, s390-32, and ia64), however it is built
for all current architectures regardless.
This patch refact the make rules for this object to so only the
aforementioned architectures that actually require it builds it.
Also, to avoid internal PLT calls to the exported symbol from the
module, glibc uses an internal header (symbol-hacks.h) which is
unrequired (and in fact breaks the build for architectures that
intend to get symbol definitions from libgcc.a). The patch also
changes it to create its own header (divdi3-symbol-hacks.h) and
adjust the architectures that require it accordingly.
I checked the build/check (with run-built-tests=no) on the
following architectures (which I think must cover all supported
ABI/builds) using GCC 6.3:
aarch64-linux-gnu
alpha-linux-gnu
arm-linux-gnueabihf
hppa-linux-gnu
ia64-linux-gnu
m68k-linux-gnu
microblaze-linux-gnu
mips64-n32-linux-gnu
mips-linux-gnu
mips64-linux-gnu
nios2-linux-gnu
powerpc-linux-gnu
powerpc-linux-gnu-power4
powerpc64-linux-gnu
powerpc64le-linux-gnu
s390x-linux-gnu
s390-linux-gnu
sh4-linux-gnu
sh4-linux-gnu-soft
sparc64-linux-gnu
sparcv9-linux-gnu
tilegx-linux-gnu
tilegx-linux-gnu-32
tilepro-linux-gnu
x86_64-linux-gnu
x86_64-linux-gnu-x32
i686-linux-gnu
I only saw one regression on sparcv9-linux-gnu (extra PLT call to
.udiv) which I address in next patch in the set. It also correctly
build SH with GCC 7.0.1 (without any regression from c89721e25d).
[1] https://sourceware.org/ml/libc-alpha/2017-03/msg00243.html
* sysdeps/i386/symbol-hacks.h: New file.
* sysdeps/m68k/symbol-hacks.h: New file.
* sysdeps/powerpc/powerpc32/symbol-hacks.h: New file.
* sysdeps/s390/s390-32/symbol-hacks.h: New file.
* sysdeps/unix/sysv/linux/i386/Makefile
[$(subdir) = csu] (sysdep_routines): New rule: divdi3 object.
[$(subdir) = csu] (sysdep-only-routines): Likewise.
[$(subdir) = csu] (CFLAGS-divdi3.c): Likewise.
* sysdeps/unix/sysv/linux/m68k/Makefile
[$(subdir) = csu] (sysdep_routines): Likewise.
[$(subdir) = csu] (sysdep-only-routines): Likewise.
[$(subdir) = csu] (CFLAGS-divdi3.c): Likewise.
* sysdeps/unix/sysv/linux/powerpc/powerpc32/Makefile
[$(subdir) = csu] (sysdep_routines): Likewise.
[$(subdir) = csu] (sysdep-only-routines): Likewise.
[$(subdir) = csu] (CFLAGS-divdi3.c): Likewise.
* sysdeps/unix/sysv/linux/s390/s390-32/Makefile
[$(subdir) = csu] (sysdep_routines): Likewise.
[$(subdir) = csu] (sysdep-only-routines): Likewise.
[$(subdir) = csu] (CFLAGS-divdi3.c): Likewise.
* sysdeps/wordsize-32/Makefile: Remove file.
* sysdeps/wordsize-32/symbol-hacks.h: Definitions move to ...
* sysdeps/wordsize-32/divdi3-symbol-hacks.h: ... here.
May be tricky for otimized implementations to handle strings around
page boundary once, for instance, it is performed unaligned loads or
when maxlen is used as a hint for vectorized loops. The test cases
should unveil regression bugs on these cases.
To some extend do_random_tests in string/test-strnlen tests strings
placed at page end but it does not cover all cases. So this change
adds tests which consists of placing strings of varying sizes ending
at the page boundary. It also combines with different values of maxlen.
Tested on ppc64le and x86_64.
* string/test-strnlen.c (do_page_tests): New function
to check length of strings ending at the page boundary.
(test_main): Added call to the do_page_tests function.
Added strnlen POWER8 otimized for long strings. It delivers
same performance as POWER7 implementation for short strings.
This takes advantage of reasonably performing unaligned loads
and bit permutes to check the first 1-16 bytes until
quadword aligned, then checks in 64 bytes strides until unsafe,
then 16 bytes, truncating the count if need be.
Likewise, the POWER7 code is recycled for less than 32 bytes strings.
Tested on ppc64 and ppc64le.
* sysdeps/powerpc/powerpc64/multiarch/Makefile
(sysdep_routines): Add strnlen-power8.
* sysdeps/powerpc/powerpc64/multiarch/ifunc-impl-list.c
(strnlen): Add __strnlen_power8 to list of strnlen functions.
* sysdeps/powerpc/powerpc64/multiarch/strnlen-power8.S:
New file.
* sysdeps/powerpc/powerpc64/multiarch/strnlen.c
(__strnlen): Add __strnlen_power8 to ifunc list.
* sysdeps/powerpc/powerpc64/power8/strnlen.S: New file.
This patch removes CALL_THREAD_FCT macro usage and its defition for
x86. For 32 bits it usage is only for force 16 stack alignment,
however stack is already explicit aligned in clone syscall. For
64 bits and x32 it just a function call and there is no need to
code it with inline assembly.
Checked on i686-linux-gnu, x86_64-linux-gnu, and x86_64-linux-gnu-x32.
* nptl/pthread_create.c (START_THREAD_DEFN): Remove
CALL_THREAD_FCT macro usage.
* sysdeps/i386/nptl/tls.h (CALL_THREAD_FCT): Remove definition.
* sysdeps/x86_64/nptl/tls.h (CALL_THREAD_FCT): Likewise.
* sysdeps/x86_64/32/nptl/tls.h: Remove file.
The new cond var implementation (ed19993b5b) removed all the
__ASSUME_{REQUEUE_PI,FUTEX_LOCK_PI} internal usage so there is no
need to keep defining it. This patch removes all USE_REQUEUE_PI
and __ASSUME_REQUEUE_PI. It is as follow up from BZ#18463.
Checked with a build for x86_64-linux-gnu, arm-linux-gnueabhf,
m68-linux-gnu, mips64-linux-gnu, and sparc64-linux-gnu.
* nptl/pthreadP.h (USE_REQUEUE_PI): Remove ununsed macro.
* sysdeps/unix/sysv/linux/arm/kernel-features.h
(__ASSUME_REQUEUE_PI): Likewise.
* sysdeps/unix/sysv/linux/kernel-features.h
(__ASSUME_REQUEUE_PI): Likewise.
* sysdeps/unix/sysv/linux/m68k/kernel-features.h
(__ASSUME_REQUEUE_PI): Likewise.
* sysdeps/unix/sysv/linux/mips/kernel-features.h
(__ASSUME_REQUEUE_PI): Likewise.
* sysdeps/unix/sysv/linux/sparc/kernel-features.h
(__ASSUME_REQUEUE_PI): Likewise.
This patch also clobbers r14 in TLS_LD, TLS_GD macros on 31bit.
This ensures that r14 isn't used to save and restore r12 while
r14 is clobbered by the bas-instruction.
As note:
r12 can't be added to clobber list as gcc would fail with:
error: PIC register clobbered by ‘12’ in ‘asm’
For 64bit this fix was already done in 2004 in the
commit b80af23ac6.
ChangeLog:
* sysdeps/s390/s390-32/tls-macros.h (TLS_LD, TLS_GD):
Clobber also r14.
When glibc is built with -fstack-check, trying to use posix_spawn can
lead to segfaults due to gcc internally probing stack memory too far.
The new spawn API will allocate a minimum of 1 page, but the stack
checking logic might probe a couple of pages. When it tries to walk
them, everything falls apart.
The gcc internal docs [1] state the default interval checking is one
page. Which means we need two pages (the current one, and the next
probed). No target currently defines it larger.
Further, it mentions that the default minimum stack size needed to
recover from an overflow is 4/8KiB for sjlj or 8/12KiB for others.
But some Linux targets (like mips and ppc) go up to 16KiB (and some
non-Linux targets go up to 24KiB).
Let's create each child with a minimum of 32KiB slack space to support
them all, and give us future breathing room.
No test is added as existing ones crash. Even a simple call is
enough to trigger the problem:
char *argv[] = { "/bin/ls", NULL };
posix_spawn(NULL, "/bin/ls", NULL, NULL, argv, NULL);
[1] https://gcc.gnu.org/onlinedocs/gcc-6.3.0/gccint/Stack-Checking.html
Fixes a typo introduced in commit
be7991c070. This caused
mallopt(M_ARENA_MAX) as well as the environment variable
MALLOC_ARENA_MAX to not work as intended because it set the
wrong internal parameter.
[BZ #21338]
* malloc/malloc.c: Call do_set_arena_max for M_ARENA_MAX
instead of incorrect do_set_arena_test
The condition for declaration of long double functions in
math-finite.h was #ifdef __MATH_DECLARE_LDOUBLE before the
macroization of this file. After the macroization, it was incorreclty
changed to #if __MATH_DECLARE_LDOUBLE, which broke the build for arm.
* math/math.h: Fix check for __MATH_DECLARE_LDOUBLE.
* math/bits/math-finite.h: Likewise.
This patch fixes some test-errno-linux unexpected returns for the
tested syscalls on some older kernels (I saw it on a Linux 3.8 on
armv7l). Basically:
- inotify_add_watch: Linux v3.8 (676a0675c) removed the test to
check at least one valid bit in flags (to return EINVAL). It
was later added back in v3.9 (04df32fa1).
- quotactl: returns ENOSYS for kernels not configured with
CONFIG_QUOTA.
Checked on x86_64-linux-gnu and armv7l-linux-gnueabihf.
* sysdeps/unix/sysv/linux/test-errno-linux.c (do_test): Handle
non expected inotify_add_watch and quotactl return.
The implementation of __ieee754_rem_pio2l in ldbl-128, ldbl-128ibm,
and ldbl-96 return the type int32_t, whereas math_private.h declares
it as returning int. This patch changes the declaration to match the
declaration in thoses directories, as well as it changes the stub
implementation in math/e_rem_pio2l.c, similarly.
* math/e_rem_pio2l.c (__ieee754_rem_pio2l): Change return type
to int32_t.
* sysdeps/generic/math_private.h: Declare __ieee754_rem_pio2l
as returning int32_t.
This patch macroizes the declarations in math/bits/math-finite.h
similarly to what math/bits/mathcalls.h does. For each floating-point
type, the file is included once in math/math.h. This will reduce the
amount of repetitive boilerplate required when adding float128
versions of these declarations.
Tested for powerpc64le and s390x.
* math/math.h: Include bits/math-finite.h once per
floating-point type.
* math/bits/math-finite.h: Macroize all declarations by
floating-point type.
This patch fixes the regression added by 23d2770 for final address
overflow calculation. The subtraction of the considered size (16)
at line 120 is at wrong place, for sizes less than 16 subsequent
overflow check will not take in consideration an invalid size (since
the subtraction will be negative). Also, the lea instruction also
does not raise the carry flag (CF) that is used in subsequent jbe
to check for overflow.
The fix is to follow x86_64 logic from 3daef2c where the overflow
is first check and a sub instruction is issued. In case of resulting
negative size, CF will be set by the sub instruction and a NULL
result will be returned. The patch also add similar tests reported
in bug report.
Checked on i686-linux-gnu and x86_64-linux-gnu.
* string/test-memchr.c (do_test): Add BZ#21182 checks for address
near end of a page.
* sysdeps/i386/i686/multiarch/memchr-sse2.S (__memchr): Fix
overflow calculation.
Fix the incorrect sorting order of a digraph and its geminated variant,
regression introduced by a faulty fix to bug 13547 in commit
b008d4c856.
Fix two inconsistencies in sorting unusual capitalization of digraphs
(bug #18587).
Enable DIACRIT_FORWARD to work around bug #17750.
Sort foreign accents after the Hungarian ones.
Add extensive unittests containing all the examples from The Rules of
Hungarian Orthography and many more, including explanatory comments.
termios.h should define IUCLC for UNIX98 and older XSI standards. The
sysdeps/unix/sysv/linux/alpha version defines it only if __USE_MISC,
so causing some conform/ tests to fail.
Other versions define it unconditionally (I* being a reserved
namespace for this header); the API should be consistent between
architectures in the absence of a clear reason for it to differ (and
given that a symbol is part of the API on two architectures, I don't
see any reason for the feature test macros required ever to differ
between those architectures), so this patch makes the alpha version
define it unconditionally as well. Two non-POSIX macros alongside it,
IMAXBEL and IUTF8, are also defined unconditionally on other
architectures, so this patch makes them consistent by defining them
unconditionally on alpha as well.
Tested (compilation only) with build-many-glibcs.py.
[BZ #21277]
* sysdeps/unix/sysv/linux/alpha/bits/termios.h (IUCLC): Define
unconditionally.
(IMAXBEL): Likewise.
(IUTF8): Likewise.
Since commit 8b9e9c3c0b, security_level replaces
is_secure. There were some old files need to be updated.
2017-03-23 Sunyeop Lee <sunyeop97@gmail.com>
* README.tunables: Updated descriptions.
* elf/dl-tunables.list: Fixed typo: SXID_NONE -> NONE.
* scripts/gen-tunables.awk: Updated the code related to the commit.
of the size and alignment is based on a trace of SPEC2006. Instead of
repeating the same copy over and over again like the existing tests, it times
several thousand different copies to more accurately estimate the overhead of
branch prediction.
* benchtests/Makefile (string-benchset): Add memcpy-random.
* benchtests/bench-memcpy-random.c: New file.
ISO C++ section 8.3.5 [dcl.fct] requires exception specifications
to appear before attribute specifiers in function declarations.
This patch fixes issues reported by stdio-common/check-installed-headers-cxx.
* stdio-common/printf.h (register_printf_modifier): Change the
order of __wur and __THROW.
(register_printf_type): Likewise.
* stdio-common/bug25.c: Include stdlib.h.
* support/tst-support_format_dns_packet.c: Include stdio.h,
stdlib.h, and string.h.
* support/tst-support_record_failure.c: Include string.h.
* support/tst-support_record_failure-2.sh: Adjust line number
expectations and correct a typo in an error message.
On Skylake server, _dl_runtime_resolve_avx512_opt is used to preserve
the first 8 vector registers. The code layout is
if only %xmm0 - %xmm7 registers are used
preserve %xmm0 - %xmm7 registers
if only %ymm0 - %ymm7 registers are used
preserve %ymm0 - %ymm7 registers
preserve %zmm0 - %zmm7 registers
Branch predication always executes the fallthrough code path to preserve
%zmm0 - %zmm7 registers speculatively, even though only %xmm0 - %xmm7
registers are used. This leads to lower CPU frequency on Skylake
server. This patch changes the fallthrough code path to preserve
%xmm0 - %xmm7 registers instead:
if whole %zmm0 - %zmm7 registers are used
preserve %zmm0 - %zmm7 registers
if only %ymm0 - %ymm7 registers are used
preserve %ymm0 - %ymm7 registers
preserve %xmm0 - %xmm7 registers
Tested on Skylake server.
[BZ #21258]
* sysdeps/x86_64/dl-trampoline.S (_dl_runtime_resolve_opt):
Define only if _dl_runtime_resolve is defined to
_dl_runtime_resolve_sse_vex.
* sysdeps/x86_64/dl-trampoline.h (_dl_runtime_resolve_opt):
Fallthrough to _dl_runtime_resolve_sse_vex.
The test malloc/tst-interpose-nothread fails on s390x if built
with GCC 7 and glibc commit "Remove the str(n)dup inlines
from string/bits/string2.h. Although inlining"
(ae65d4f3c3) with output:
error: free: 0x3fffdffa010: invalid allocation index: 0 (not less than 0)
The destructor check_for_allocations in malloc/tst-interpose-aux.c is
called twice. One time after the test-child-process has finished successfully
and once after the test-parent-process finishes.
During the latter invocation, allocation_index == 0. GCC 7 is now inlining the
free function and calls unconditionally fail in get_header as
header->allocation_index (type == size_t) is always >= allocation_index (= 0).
Before the mentioned commit above, strdup was replaced by strlen, malloc and
memcpy. The malloc call was also inlined and allocation_index was set to one.
This patch moves the already existing compiler barrier before the invocation
of free.
ChangeLog:
* malloc/tst-interpose-aux.c (check_for_allocations):
Move compiler barrier before free.
