utils: Use GCC's __atomic built-ins if available

These are available since GCC 4.7 and will eventually replace the __sync
operations.  They support the memory model defined by C++11. For instance,
by using __ATOMIC_RELAXED for some operations on the reference counters we
can avoid memory barriers, which are required by __sync operations (whose
memory model essentially is __ATOMIC_SEQ_CST).
This commit is contained in:
Tobias Brunner 2014-04-11 16:07:32 +02:00
parent efedd0d21e
commit 0f603d425d
3 changed files with 47 additions and 10 deletions

View File

@ -667,21 +667,39 @@ AC_COMPILE_IFELSE(
[AC_MSG_RESULT([no])]
)
AC_MSG_CHECKING([for gcc atomic operations])
AC_MSG_CHECKING([for GCC __atomic operations])
AC_RUN_IFELSE([AC_LANG_SOURCE(
[[
int main() {
volatile int ref = 1;
__sync_fetch_and_add (&ref, 1);
__sync_sub_and_fetch (&ref, 1);
/* Make sure test fails if operations are not supported */
__sync_val_compare_and_swap(&ref, 1, 0);
int ref = 1, val;
__atomic_fetch_add(&ref, 1, __ATOMIC_RELAXED);
val = __atomic_sub_fetch(&ref, 1, __ATOMIC_RELAXED);
__atomic_compare_exchange_n(&ref, &val, 0, 0, __ATOMIC_RELAXED,
__ATOMIC_RELAXED);
return ref;
}
]])],
[AC_MSG_RESULT([yes]);
AC_DEFINE([HAVE_GCC_ATOMIC_OPERATIONS], [],
[have GCC __sync_* atomic operations])],
[have GCC __atomic_* operations])],
[AC_MSG_RESULT([no])],
[AC_MSG_RESULT([no])]
)
AC_MSG_CHECKING([for GCC __sync operations])
AC_RUN_IFELSE([AC_LANG_SOURCE(
[[
int main() {
int ref = 1;
__sync_fetch_and_add (&ref, 1);
__sync_sub_and_fetch (&ref, 1);
__sync_val_compare_and_swap(&ref, 1, 0);
return ref;
}
]])],
[AC_MSG_RESULT([yes]);
AC_DEFINE([HAVE_GCC_SYNC_OPERATIONS], [],
[have GCC __sync_* operations])],
[AC_MSG_RESULT([no])],
[AC_MSG_RESULT([no])]
)

View File

@ -511,7 +511,7 @@ void nop()
{
}
#ifndef HAVE_GCC_ATOMIC_OPERATIONS
#if !defined(HAVE_GCC_ATOMIC_OPERATIONS) && !defined(HAVE_GCC_SYNC_OPERATIONS)
/**
* We use a single mutex for all refcount variables.
@ -578,7 +578,7 @@ bool cas_##name(type *ptr, type oldval, type newval) \
_cas_impl(bool, bool)
_cas_impl(ptr, void*)
#endif /* HAVE_GCC_ATOMIC_OPERATIONS */
#endif /* !HAVE_GCC_ATOMIC_OPERATIONS && !HAVE_GCC_SYNC_OPERATIONS */
#ifdef HAVE_FMEMOPEN_FALLBACK

View File

@ -750,6 +750,25 @@ typedef u_int refcount_t;
#ifdef HAVE_GCC_ATOMIC_OPERATIONS
#define ref_get(ref) __atomic_add_fetch(ref, 1, __ATOMIC_RELAXED)
/* The relaxed memory model works fine for increments as these (usually) don't
* change the state of refcounted objects. But here we have to ensure that we
* free the right stuff if ref counted objects are mutable. So we have to sync
* with other threads that call ref_put(). It would be sufficient to use
* __ATOMIC_RELEASE here and then call __atomic_thread_fence() with
* __ATOMIC_ACQUIRE if we reach 0, but since we don't have control over the use
* of ref_put() we have to make sure. */
#define ref_put(ref) (!__atomic_sub_fetch(ref, 1, __ATOMIC_ACQ_REL))
#define ref_cur(ref) __atomic_load_n(ref, __ATOMIC_RELAXED)
#define _cas_impl(ptr, oldval, newval) ({ typeof(oldval) _old = oldval; \
__atomic_compare_exchange_n(ptr, &_old, newval, FALSE, \
__ATOMIC_SEQ_CST, __ATOMIC_RELAXED); })
#define cas_bool(ptr, oldval, newval) _cas_impl(ptr, oldval, newval)
#define cas_ptr(ptr, oldval, newval) _cas_impl(ptr, oldval, newval)
#elif defined(HAVE_GCC_SYNC_OPERATIONS)
#define ref_get(ref) __sync_add_and_fetch(ref, 1)
#define ref_put(ref) (!__sync_sub_and_fetch(ref, 1))
#define ref_cur(ref) __sync_fetch_and_add(ref, 0)
@ -759,7 +778,7 @@ typedef u_int refcount_t;
#define cas_ptr(ptr, oldval, newval) \
(__sync_bool_compare_and_swap(ptr, oldval, newval))
#else /* !HAVE_GCC_ATOMIC_OPERATIONS */
#else /* !HAVE_GCC_ATOMIC_OPERATIONS && !HAVE_GCC_SYNC_OPERATIONS */
/**
* Get a new reference.