Optimize generic spinlock code and use C11 like atomic macros.

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.

4 files changed, 100 insertions, 34 deletions

diff --git a/nptl/pthread_spin_init.c b/nptl/pthread_spin_init.c
index 01dec5e..fe30913 100644
--- a/nptl/pthread_spin_init.c
+++ b/nptl/pthread_spin_init.c
@@ -22,6 +22,7 @@
 int
 pthread_spin_init (pthread_spinlock_t *lock, int pshared)
 {
-  *lock = 0;
+  /* Relaxed MO is fine because this is an initializing store.  */
+  atomic_store_relaxed (lock, 0);
   return 0;
 }
diff --git a/nptl/pthread_spin_lock.c b/nptl/pthread_spin_lock.c
index 4d03b78..682af80 100644
--- a/nptl/pthread_spin_lock.c
+++ b/nptl/pthread_spin_lock.c
@@ -19,27 +19,35 @@
 #include <atomic.h>
 #include "pthreadP.h"
 
-/* A machine-specific version can define SPIN_LOCK_READS_BETWEEN_CMPXCHG
-  to the number of plain reads that it's optimal to spin on between uses
-  of atomic_compare_and_exchange_val_acq.  If spinning forever is optimal
-  then use -1.  If no plain reads here would ever be optimal, use 0.  */
-#ifndef SPIN_LOCK_READS_BETWEEN_CMPXCHG
-# warning machine-dependent file should define SPIN_LOCK_READS_BETWEEN_CMPXCHG
-# define SPIN_LOCK_READS_BETWEEN_CMPXCHG 1000
-#endif
-
 int
 pthread_spin_lock (pthread_spinlock_t *lock)
 {
-  /* atomic_exchange usually takes less instructions than
-     atomic_compare_and_exchange.  On the other hand,
-     atomic_compare_and_exchange potentially generates less bus traffic
-     when the lock is locked.
-     We assume that the first try mostly will be successful, and we use
-     atomic_exchange.  For the subsequent tries we use
-     atomic_compare_and_exchange.  */
-  if (atomic_exchange_acq (lock, 1) == 0)
+  int val = 0;
+
+  /* We assume that the first try mostly will be successful, thus we use
+     atomic_exchange if it is not implemented by a CAS loop (we also assume
+     that atomic_exchange can be faster if it succeeds, see
+     ATOMIC_EXCHANGE_USES_CAS).  Otherwise, we use a weak CAS and not an
+     exchange so we bail out after the first failed attempt to change the
+     state.  For the subsequent attempts we use atomic_compare_and_exchange
+     after we observe that the lock is not acquired.
+     See also comment in pthread_spin_trylock.
+     We use acquire MO to synchronize-with the release MO store in
+     pthread_spin_unlock, and thus ensure that prior critical sections
+     happen-before this critical section.  */
+#if ! ATOMIC_EXCHANGE_USES_CAS
+  /* Try to acquire the lock with an exchange instruction as this architecture
+     has such an instruction and we assume it is faster than a CAS.
+     The acquisition succeeds if the lock is not in an acquired state.  */
+  if (__glibc_likely (atomic_exchange_acquire (lock, 1) == 0))
     return 0;
+#else
+  /* Try to acquire the lock with a CAS instruction as this architecture
+     has no exchange instruction.  The acquisition succeeds if the lock is not
+     acquired.  */
+  if (__glibc_likely (atomic_compare_exchange_weak_acquire (lock, &val, 1)))
+    return 0;
+#endif
 
   do
     {
@@ -47,23 +55,26 @@ pthread_spin_lock (pthread_spinlock_t *lock)
 	 to cmpxchg is not a good idea on many targets as that will force
 	 expensive memory synchronizations among processors and penalize other
 	 running threads.
-	 On the other hand, we do want to update memory state on the local core
-	 once in a while to avoid spinning indefinitely until some event that
-	 will happen to update local memory as a side-effect.  */
-      if (SPIN_LOCK_READS_BETWEEN_CMPXCHG >= 0)
+	 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!
+	 Thus we use relaxed MO reads until we observe the lock to not be
+	 acquired anymore.  */
+      do
 	{
-	  int wait = SPIN_LOCK_READS_BETWEEN_CMPXCHG;
+	  /* TODO Back-off.  */
 
-	  while (*lock != 0 && wait > 0)
-	    --wait;
-	}
-      else
-	{
-	  while (*lock != 0)
-	    ;
+	  atomic_spin_nop ();
+
+	  val = atomic_load_relaxed (lock);
 	}
+      while (val != 0);
+
+      /* We need acquire memory order here for the same reason as mentioned
+	 for the first try to lock the spinlock.  */
     }
-  while (atomic_compare_and_exchange_val_acq (lock, 1, 0) != 0);
+  while (!atomic_compare_exchange_weak_acquire (lock, &val, 1));
 
   return 0;
 }
diff --git a/nptl/pthread_spin_trylock.c b/nptl/pthread_spin_trylock.c
index 593bba3..83921b0 100644
--- a/nptl/pthread_spin_trylock.c
+++ b/nptl/pthread_spin_trylock.c
@@ -23,5 +23,57 @@
 int
 pthread_spin_trylock (pthread_spinlock_t *lock)
 {
-  return atomic_exchange_acq (lock, 1) ? EBUSY : 0;
+  /* For the spin try lock, we have the following possibilities:
+
+     1) If we assume that trylock will most likely succeed in practice:
+     * We just do an exchange.
+
+     2) If we want to bias towards cases where trylock succeeds, but don't
+     rule out contention:
+     * If exchange is not implemented by a CAS loop, and exchange is faster
+     than CAS, do an exchange.
+     * If exchange is implemented by a CAS loop, use a weak CAS and not an
+     exchange so we bail out after the first failed attempt to change the state.
+
+     3) If we expect contention to be likely:
+     * If CAS always brings the cache line into an exclusive state even if the
+     spinlock is already acquired, then load the value first with
+     atomic_load_relaxed and test if lock is not acquired. Then do 2).
+
+     We assume that 2) is the common case, and that this won't be slower than
+     1) in the common case.
+
+     We use acquire MO to synchronize-with the release MO store in
+     pthread_spin_unlock, and thus ensure that prior critical sections
+     happen-before this critical section.  */
+#if ! ATOMIC_EXCHANGE_USES_CAS
+  /* Try to acquire the lock with an exchange instruction as this architecture
+     has such an instruction and we assume it is faster than a CAS.
+     The acquisition succeeds if the lock is not in an acquired state.  */
+  if (atomic_exchange_acquire (lock, 1) == 0)
+    return 0;
+#else
+  /* Try to acquire the lock with a CAS instruction as this architecture
+     has no exchange instruction.  The acquisition succeeds if the lock is not
+     acquired.  */
+  do
+    {
+      int val = 0;
+      if (atomic_compare_exchange_weak_acquire (lock, &val, 1))
+	return 0;
+    }
+  /* atomic_compare_exchange_weak_acquire can fail spuriously.  Whereas
+     C++11 and C11 make it clear that trylock operations can fail spuriously,
+     POSIX does not explicitly specify this; it only specifies that failing
+     synchronization operations do not need to have synchronization effects
+     themselves, but a spurious failure is something that could contradict a
+     happens-before established earlier (e.g., that we need to observe that
+     the lock is acquired).  Therefore, we emulate a strong CAS by simply
+     checking with a relaxed MO load that the lock is really acquired before
+     returning EBUSY; the additional overhead this may cause is on the slow
+     path.  */
+  while (atomic_load_relaxed (lock) == 0);
+#endif
+
+  return EBUSY;
 }
diff --git a/nptl/pthread_spin_unlock.c b/nptl/pthread_spin_unlock.c
index 5fd73e5..f83b696 100644
--- a/nptl/pthread_spin_unlock.c
+++ b/nptl/pthread_spin_unlock.c
@@ -23,7 +23,9 @@
 int
 pthread_spin_unlock (pthread_spinlock_t *lock)
 {
-  atomic_full_barrier ();
-  *lock = 0;
+  /* The atomic_store_release synchronizes-with the atomic_exchange_acquire
+     or atomic_compare_exchange_weak_acquire in pthread_spin_lock /
+     pthread_spin_trylock.  */
+  atomic_store_release (lock, 0);
   return 0;
 }