The update of *adapt_count after the release of the lock causes a race
condition when thread A unlocks, thread B continues and destroys the
mutex, and thread A writes to *adapt_count.
Work around a GCC behavior with hardware transactional memory built-ins.
GCC doesn't treat the PowerPC transactional built-ins as compiler
barriers, moving instructions past the transaction boundaries and
altering their atomicity.
With TLE enabled, the adapt count variable update incurs
an 8% overhead before entering the critical section of an
elided mutex.
Instead, if it is done right after leaving the critical
section, this serialization can be avoided.
This alters the existing behavior of __lll_trylock_elision
as it will only decrement the adapt_count if it successfully
acquires the lock.
* sysdeps/unix/sysv/linux/powerpc/elision-lock.c
(__lll_lock_elision): Remove adapt_count decrement...
* sysdeps/unix/sysv/linux/powerpc/elision-trylock.c
(__lll_trylock_elision): Likewise.
* sysdeps/unix/sysv/linux/powerpc/elision-unlock.c
(__lll_unlock_elision): ... to here. And utilize
new adapt_count parameter.
* sysdeps/unix/sysv/linux/powerpc/lowlevellock.h
(__lll_unlock_elision): Update to include adapt_count
parameter.
(lll_unlock_elision): Pass pointer to adapt_count
variable.
This patch adds support for lock elision using ISA 2.07 hardware
transactional memory instructions for pthread_mutex primitives.
Similar to s390 version, the for elision logic defined in
'force-elision.h' is only enabled if ENABLE_LOCK_ELISION is defined.
Also, the lock elision code should be able to be built even with
a compiler that does not provide HTM support with builtins.
However I have noted the performance is sub-optimal due scheduling
pressures.
