This patch optimizes the generic spinlock code.
The type pthread_spinlock_t is a typedef to volatile int on all archs.
Passing a volatile pointer to the atomic macros which are not mapped to the
C11 atomic builtins can lead to extra stores and loads to stack if such
a macro creates a temporary variable by using "__typeof (*(mem)) tmp;".
Thus, those macros which are used by spinlock code - atomic_exchange_acquire,
atomic_load_relaxed, atomic_compare_exchange_weak - have to be adjusted.
According to the comment from Szabolcs Nagy, the type of a cast expression is
unqualified (see http://www.open-std.org/jtc1/sc22/wg14/www/docs/dr_423.htm):
__typeof ((__typeof (*(mem)) *(mem)) tmp;
Thus from spinlock perspective the variable tmp is of type int instead of
type volatile int. This patch adjusts those macros in include/atomic.h.
With this construct GCC >= 5 omits the extra stores and loads.
The atomic macros are replaced by the C11 like atomic macros and thus
the code is aligned to it. The pthread_spin_unlock implementation is now
using release memory order instead of sequentially consistent memory order.
The issue with passed volatile int pointers applies to the C11 like atomic
macros as well as the ones used before.
I've added a glibc_likely hint to the first atomic exchange in
pthread_spin_lock in order to return immediately to the caller if the lock is
free. Without the hint, there is an additional jump if the lock is free.
I've added the atomic_spin_nop macro within the loop of plain reads.
The plain reads are also realized by C11 like atomic_load_relaxed macro.
The new define ATOMIC_EXCHANGE_USES_CAS determines if the first try to acquire
the spinlock in pthread_spin_lock or pthread_spin_trylock is an exchange
or a CAS. This is defined in atomic-machine.h for all architectures.
The define SPIN_LOCK_READS_BETWEEN_CMPXCHG is now removed.
There is no technical reason for throwing in a CAS every now and then,
and so far we have no evidence that it can improve performance.
If that would be the case, we have to adjust other spin-waiting loops
elsewhere, too! Using a CAS loop without plain reads is not a good idea
on many targets and wasn't used by one. Thus there is now no option to
do so.
Architectures are now using the generic spinlock automatically if they
do not provide an own implementation. Thus the pthread_spin_lock.c files
in sysdeps folder are deleted.
ChangeLog:
* NEWS: Mention new spinlock implementation.
* include/atomic.h:
(__atomic_val_bysize): Cast type to omit volatile qualifier.
(atomic_exchange_acq): Likewise.
(atomic_load_relaxed): Likewise.
(ATOMIC_EXCHANGE_USES_CAS): Check definition.
* nptl/pthread_spin_init.c (pthread_spin_init):
Use atomic_store_relaxed.
* nptl/pthread_spin_lock.c (pthread_spin_lock):
Use C11-like atomic macros.
* nptl/pthread_spin_trylock.c (pthread_spin_trylock):
Likewise.
* nptl/pthread_spin_unlock.c (pthread_spin_unlock):
Use atomic_store_release.
* sysdeps/aarch64/nptl/pthread_spin_lock.c: Delete File.
* sysdeps/arm/nptl/pthread_spin_lock.c: Likewise.
* sysdeps/hppa/nptl/pthread_spin_lock.c: Likewise.
* sysdeps/m68k/nptl/pthread_spin_lock.c: Likewise.
* sysdeps/microblaze/nptl/pthread_spin_lock.c: Likewise.
* sysdeps/mips/nptl/pthread_spin_lock.c: Likewise.
* sysdeps/nios2/nptl/pthread_spin_lock.c: Likewise.
* sysdeps/aarch64/atomic-machine.h (ATOMIC_EXCHANGE_USES_CAS): Define.
* sysdeps/alpha/atomic-machine.h: Likewise.
* sysdeps/arm/atomic-machine.h: Likewise.
* sysdeps/i386/atomic-machine.h: Likewise.
* sysdeps/ia64/atomic-machine.h: Likewise.
* sysdeps/m68k/coldfire/atomic-machine.h: Likewise.
* sysdeps/m68k/m680x0/m68020/atomic-machine.h: Likewise.
* sysdeps/microblaze/atomic-machine.h: Likewise.
* sysdeps/mips/atomic-machine.h: Likewise.
* sysdeps/powerpc/powerpc32/atomic-machine.h: Likewise.
* sysdeps/powerpc/powerpc64/atomic-machine.h: Likewise.
* sysdeps/s390/atomic-machine.h: Likewise.
* sysdeps/sparc/sparc32/atomic-machine.h: Likewise.
* sysdeps/sparc/sparc32/sparcv9/atomic-machine.h: Likewise.
* sysdeps/sparc/sparc64/atomic-machine.h: Likewise.
* sysdeps/tile/tilegx/atomic-machine.h: Likewise.
* sysdeps/tile/tilepro/atomic-machine.h: Likewise.
* sysdeps/unix/sysv/linux/hppa/atomic-machine.h: Likewise.
* sysdeps/unix/sysv/linux/m68k/coldfire/atomic-machine.h: Likewise.
* sysdeps/unix/sysv/linux/nios2/atomic-machine.h: Likewise.
* sysdeps/unix/sysv/linux/sh/atomic-machine.h: Likewise.
* sysdeps/x86_64/atomic-machine.h: Likewise.
As was done in b224637928, check for large size causing an overflow
in the loop that walks over the array.
Branching out of line here is the fastest approach for handling this
problem, since tile can bundle the instructions to compute the branch
test in parallel with doing the required memchr loop setup computation.
Unfortunately, the existing saturated ops (e.g. tilegx addxsc) are
all signed saturing ops, so don't help with unsigned saturation.
It was noted in
<https://sourceware.org/ml/libc-alpha/2012-09/msg00305.html> that the
bits/*.h naming scheme should only be used for installed headers.
This patch renames bits/atomic.h to atomic-machine.h to follow that
convention.
This is the only change in this series that needs to change the
filename rather than simply removing a directory level (because both
atomic.h and bits/atomic.h exist at present).
Tested for x86_64 (testsuite, and that installed stripped shared
libraries are unchanged by the patch).
[BZ #14912]
* sysdeps/aarch64/bits/atomic.h: Move to ...
* sysdeps/aarch64/atomic-machine.h: ...here.
(_AARCH64_BITS_ATOMIC_H): Rename macro to
_AARCH64_ATOMIC_MACHINE_H.
* sysdeps/alpha/bits/atomic.h: Move to ...
* sysdeps/alpha/atomic-machine.h: ...here.
* sysdeps/arm/bits/atomic.h: Move to ...
* sysdeps/arm/atomic-machine.h: ...here. Update comments.
* bits/atomic.h: Move to ...
* sysdeps/generic/atomic-machine.h: ...here.
(_BITS_ATOMIC_H): Rename macro to _ATOMIC_MACHINE_H.
* sysdeps/i386/bits/atomic.h: Move to ...
* sysdeps/i386/atomic-machine.h: ...here.
* sysdeps/ia64/bits/atomic.h: Move to ...
* sysdeps/ia64/atomic-machine.h: ...here.
* sysdeps/m68k/coldfire/bits/atomic.h: Move to ...
* sysdeps/m68k/coldfire/atomic-machine.h: ...here.
(_BITS_ATOMIC_H): Rename macro to _ATOMIC_MACHINE_H.
* sysdeps/m68k/m680x0/m68020/bits/atomic.h: Move to ...
* sysdeps/m68k/m680x0/m68020/atomic-machine.h: ...here.
* sysdeps/microblaze/bits/atomic.h: Move to ...
* sysdeps/microblaze/atomic-machine.h: ...here.
* sysdeps/mips/bits/atomic.h: Move to ...
* sysdeps/mips/atomic-machine.h: ...here.
(_MIPS_BITS_ATOMIC_H): Rename macro to _MIPS_ATOMIC_MACHINE_H.
* sysdeps/powerpc/bits/atomic.h: Move to ...
* sysdeps/powerpc/atomic-machine.h: ...here. Update comments.
* sysdeps/powerpc/powerpc32/bits/atomic.h: Move to ...
* sysdeps/powerpc/powerpc32/atomic-machine.h: ...here. Update
comments. Include <atomic-machine.h> instead of <bits/atomic.h>.
* sysdeps/powerpc/powerpc64/bits/atomic.h: Move to ...
* sysdeps/powerpc/powerpc64/atomic-machine.h: ...here. Include
<atomic-machine.h> instead of <bits/atomic.h>.
* sysdeps/s390/bits/atomic.h: Move to ...
* sysdeps/s390/atomic-machine.h: ...here.
* sysdeps/sparc/sparc32/bits/atomic.h: Move to ...
* sysdeps/sparc/sparc32/atomic-machine.h: ...here.
(_BITS_ATOMIC_H): Rename macro to _ATOMIC_MACHINE_H.
* sysdeps/sparc/sparc32/sparcv9/bits/atomic.h: Move to ...
* sysdeps/sparc/sparc32/sparcv9/atomic-machine.h: ...here.
* sysdeps/sparc/sparc64/bits/atomic.h: Move to ...
* sysdeps/sparc/sparc64/atomic-machine.h: ...here.
* sysdeps/tile/bits/atomic.h: Move to ...
* sysdeps/tile/atomic-machine.h: ...here.
* sysdeps/tile/tilegx/bits/atomic.h: Move to ...
* sysdeps/tile/tilegx/atomic-machine.h: ...here. Include
<sysdeps/tile/atomic-machine.h> instead of
<sysdeps/tile/bits/atomic.h>.
(_BITS_ATOMIC_H): Rename macro to _ATOMIC_MACHINE_H.
* sysdeps/tile/tilepro/bits/atomic.h: Move to ...
* sysdeps/tile/tilepro/atomic-machine.h: ...here. Include
<sysdeps/tile/atomic-machine.h> instead of
<sysdeps/tile/bits/atomic.h>.
(_BITS_ATOMIC_H): Rename macro to _ATOMIC_MACHINE_H.
* sysdeps/unix/sysv/linux/arm/bits/atomic.h: Move to ...
* sysdeps/unix/sysv/linux/arm/atomic-machine.h: ...here. Include
<sysdeps/arm/atomic-machine.h> instead of
<sysdeps/arm/bits/atomic.h>.
* sysdeps/unix/sysv/linux/hppa/bits/atomic.h: Move to ...
* sysdeps/unix/sysv/linux/hppa/atomic-machine.h: ...here.
(_BITS_ATOMIC_H): Rename macro to _ATOMIC_MACHINE_H.
* sysdeps/unix/sysv/linux/m68k/coldfire/bits/atomic.h: Move to ...
* sysdeps/unix/sysv/linux/m68k/coldfire/atomic-machine.h: ...here.
(_BITS_ATOMIC_H): Rename macro to _ATOMIC_MACHINE_H.
* sysdeps/unix/sysv/linux/nios2/bits/atomic.h: Move to ...
* sysdeps/unix/sysv/linux/nios2/atomic-machine.h: ...here.
(_NIOS2_BITS_ATOMIC_H): Rename macro to _NIOS2_ATOMIC_MACHINE_H.
* sysdeps/unix/sysv/linux/sh/bits/atomic.h: Move to ...
* sysdeps/unix/sysv/linux/sh/atomic-machine.h: ...here.
* sysdeps/x86_64/bits/atomic.h: Move to ...
* sysdeps/x86_64/atomic-machine.h: ...here.
* include/atomic.h: Include <atomic-machine.h> instead of
<bits/atomic.h>.
Various code in glibc uses __strnlen instead of strnlen for namespace
reasons. However, __strnlen does not use libc_hidden_proto /
libc_hidden_def (as is normally done for any function defined and
called within the same library, whether or not exported from the
library and whatever namespace it is in), so the compiler does not
know that those calls are to a function within libc.
This patch uses libc_hidden_proto / libc_hidden_def with __strnlen.
On x86_64, it makes no difference to the installed stripped shared
libraries. On 32-bit x86, it causes __strnlen calls to go to the same
place as strnlen calls (the fallback strnlen implementation), rather
than through a PLT entry for the strnlen IFUNC; I'm not sure of the
logic behind when calls from within libc should use IFUNCs versus when
they should go direct to a particular function implementation, but
clearly it doesn't make sense for strnlen and __strnlen to be handled
differently in this regard.
Tested for x86_64 and x86 (testsuite, and comparison of installed
shared libraries as described above).
* string/strnlen.c [!STRNLEN] (__strnlen): Use libc_hidden_def.
* include/string.h (__strnlen): Use libc_hidden_proto.
* sysdeps/aarch64/strnlen.S (__strnlen): Use libc_hidden_def.
* sysdeps/i386/i686/multiarch/strnlen-c.c [SHARED]
(libc_hidden_def): Define __GI___strnlen as well as __GI_strnlen.
* sysdeps/powerpc/powerpc32/power4/multiarch/strnlen-power7.S
(libc_hidden_def): Undefine and redefine.
* sysdeps/powerpc/powerpc32/power4/multiarch/strnlen-ppc32.c
[SHARED] (libc_hidden_def): Define __GI___strnlen as well as
__GI_strnlen.
* sysdeps/powerpc/powerpc32/power7/strnlen.S (__strnlen): Use
libc_hidden_def.
* sysdeps/tile/tilegx/strnlen.c (__strnlen): Likewise.
This is because of alignment issues in the sem_t support.
tilegx32 does in fact support 64-bit atomics and we will need
to revisit this after the 2.21 freeze.
I missed this during the initial port. Some testing shows that
enabling this mode does, unsurprisingly, yield some nice speedups
on the math functions in question.
[BZ #17746]
The __builtin_expect() truncated a uint64_t to a 32-bit long
in ILP32 mode, discarding the high 32 bits, and potentially
missing the NUL terminator that we were searching for with SIMD
operations. Explicitly compare to zero to fix the problem.
The two_way_short_needle() routine included from str-two-way.h
is not used, so mark it so to avoid compiler warnings.
Calling strnlen() breaks linknamespace tests, so change it
to __strnlen().
Without this, on gcc 4.8.2 the built glibc crashes when memcpy
or memset are invoked, since they call themselves recursively.
See commit 85c2e6110c for the generic inhibit_loop_to_libcall.
This sets __HAVE_64B_ATOMICS if provided. It also sets
USE_ATOMIC_COMPILER_BUILTINS to true if the existing atomic ops use the
__atomic* builtins (aarch64, mips partially) or if this has been
tested (x86_64); otherwise, this is set to false so that C11 atomics will
be based on the existing atomic operations.
strnlen() is based on the existing tile strlen() with length
checking added. It speeds up by up to 5x, but on average across
the benchtest corpus by around 35%. No regressions are seen.
strstr() does 8-byte aligned loads and compares using a 2-byte
filter on the first two bytes of the needle and then testing
the remaining bytes in needle using memcmp(). It speeds up
about 5x in the best case (for "found" needles), about 2x looking
at benchtest as a whole, with some slowdowns as much as 45%.
on a few cases (including the "fail" case for 128KB search).
strcasestr() is based on strstr() but uses a SIMD tolower
routine to convert 8-bytes to lower case in 5 instructions.
It also uses a 2-byte filter and then strncasecmp() for the
remaining bytes. strncasecmp() is not optimized for SIMD, so
there is futher room for improvement. However, it is still up
to 16x faster for "found" needles, averaging 2x faster on the
whole corpus of benchtests. It does slow down by up to 35%
on a few cases, similarly to strstr().
Define MEMCPY_OK_FOR_FWD_MEMMOVE in memcopy.h and let arch-specific
implementations of that file override the value if necessary. This
override is only useful for tile and moving this macro to memcopy.h
allows us to remove the tile-specific memmove.c.
I've moved the TILE-Gx and TILEPro ports to the main sysdeps hierarchy,
along with the linux-generic ports infrastructure. Beyond the README
update, the move was just
git mv ports/sysdeps/tile sysdeps/tile
git mv ports/sysdeps/unix/sysv/linux/tile \
sysdeps/unix/sysv/linux/tile
git mv ports/sysdeps/unix/sysv/linux/generic \
sysdeps/unix/sysv/linux/generic
I updated the relevant ChangeLogs along the lines of the ARM move
in commit c6bfe5c4d7 and tested the 64-bit tilegx build to confirm that
there were no changes in "objdump -dr" output in the shared objects.
