atomic: revisit the atomic operation checks and their usage

* configure.ac:
Check for working Solaris/NetBSD-style atomic.h independly on OS.
Add check for GCC-style __sync_*() builtins.
New defines: XS_ATOMIC_GCC_SYNC, XS_ATOMIC_SOLARIS.
Removed define: XS_FORCE_MUTEXES.

* src/atomic_counter.hpp:
(atomic_counter_t::add):
(atomic_counter_t::sub):
* src/atomic_ptr.hpp:
(atomic_ptr_t::xchg):
(atomic_ptr_t::cas):
Use result of these checks.
Preference order:
  1. GCC-style __sync_*() builtins
  2. Inline asm (x86, x86-64, armv7a)
  3. Solaris/NetBSD-style atomic.h, Windows-specific API
  4. Fallback to mutex-based implementation

3 files changed, 129 insertions, 146 deletions

diff --git a/configure.ac b/configure.ac
index decbd6c..0d8b404 100644
--- a/configure.ac
+++ b/configure.ac
@@ -106,19 +106,6 @@ AS_CASE(["${host_os}"],
         AC_DEFINE([XS_HAVE_SOLARIS], [1], [Have Solaris OS])
         AC_CHECK_LIB([socket], [socket])
         AC_CHECK_LIB([nsl], [gethostbyname])
-        AC_MSG_CHECKING([whether atomic operations can be used])
-        AC_COMPILE_IFELSE([AC_LANG_PROGRAM(
-            [[#include <atomic.h>]],
-            [[uint32_t value;
-              atomic_cas_32 (&value, 0, 0);
-              return 0;]])],
-            [solaris_has_atomic=yes],
-            [solaris_has_atomic=no])
-        AC_MSG_RESULT([$solaris_has_atomic])
-        # Solaris 8 does not have atomic operations exported to user space.
-        AS_IF([test "x$solaris_has_atomic" = "xno"], [
-            AC_DEFINE([XS_FORCE_MUTEXES], [1], [Force to use mutexes])
-        ])
     ],
     [*freebsd*], [
         AC_DEFINE([XS_HAVE_FREEBSD], [1], [Have FreeBSD OS])
@@ -133,24 +120,6 @@ AS_CASE(["${host_os}"],
      ],
     [*netbsd*], [
         AC_DEFINE([XS_HAVE_NETBSD], [1], [Have NetBSD OS])
-        # NetBSD 5.0 and newer provides atomic operations but we can
-        # only use these on systems where PR #42842 has been fixed so
-        # we must try and link a test program using C++.
-        libxs_netbsd_has_atomic=no
-        AC_MSG_CHECKING([whether atomic operations can be used])
-        AC_LANG_PUSH([C++])
-        AC_LINK_IFELSE([AC_LANG_PROGRAM(
-            [[#include <atomic.h>]],
-            [[uint32_t value;
-              atomic_cas_32 (&value, 0, 0);
-              return 0;]])],
-            [libxs_netbsd_has_atomic=yes],
-            [libxs_netbsd_has_atomic=no])
-        AC_LANG_POP([C++])
-        AC_MSG_RESULT([$libxs_netbsd_has_atomic])
-        AS_IF([test "x$libxs_netbsd_has_atomic" = "xno"], [
-            AC_DEFINE([XS_FORCE_MUTEXES], [1], [Force to use mutexes])
-        ])
     ],
     [openbsd*], [
         AC_DEFINE([XS_HAVE_OPENBSD], [1], [Have OpenBSD OS])
@@ -241,6 +210,53 @@ AC_CHECK_HEADERS([ifaddrs.h],
     [AC_DEFINE([XS_HAVE_IFADDRS], [1], [Have ifaddrs.h header.])])
 
 ###############################################################################
+# Check available atomic implementations
+#
+# We have following variants:
+#   o  gcc __sync_*() [XS_ATOMIC_GCC_SYNC]
+#   o  optional atomic.h with atomic_cas_32() on solaris and netbsd [XS_ATOMIC_SOLARIS]
+#   o  windows specific [XS_HAVE_WINDOWS]
+#   o  handwritten inline asm for x86 and arm [no specific define]
+#   o  fallback to mutexes (or force by hands) [no specific define]
+
+#
+# Check for Solaris and NetBSD style atomic.h
+#
+AC_LANG_PUSH([C++])
+AC_MSG_CHECKING([for Solaris/NetBSD-style atomic.h])
+AC_COMPILE_IFELSE([AC_LANG_PROGRAM(
+    [[#include <atomic.h>]],
+    [[uint32_t value;
+      atomic_cas_32 (&value, 0, 0);
+      return 0;]])],
+    [solaris_style_atomic=yes],
+    [solaris_style_atomic=no])
+AC_MSG_RESULT([$solaris_style_atomic])
+AS_IF([test "x$solaris_style_atomic" = "xyes"], [
+    AC_DEFINE([XS_ATOMIC_SOLARIS], [1], [Solaris/NetBSD-style atomic.h with atomic_cas_32()])
+])
+AC_LANG_POP([C++])
+
+#
+# Check for GCC-style __sync_*() builtins
+#
+AC_LANG_PUSH([C++])
+AC_MSG_CHECKING([for GCC-style __sync_*() atomic builtins])
+AC_LINK_IFELSE([AC_LANG_PROGRAM(
+    [volatile void* p;],
+    [[  int r;
+        r = __sync_bool_compare_and_swap(&p, (void*) 0x12345, (void*) 0);
+        return r;]])],
+    [gcc_style_sync_atomic=yes],
+    [gcc_style_sync_atomic=no])
+AC_MSG_RESULT([$gcc_style_sync_atomic])
+AS_IF([test "x$gcc_style_sync_atomic" = "xyes"], [
+    AC_DEFINE([XS_ATOMIC_GCC_SYNC], [1], [GCC-style __sync_*() atomic builtins])
+])
+AC_LANG_POP([C++])
+
+
+###############################################################################
 # Check polling system                                                        #
 ###############################################################################
 
diff --git a/src/atomic_counter.hpp b/src/atomic_counter.hpp
index bd89f72..32395d9 100644
--- a/src/atomic_counter.hpp
+++ b/src/atomic_counter.hpp
@@ -25,26 +25,16 @@
 #include "stdint.hpp"
 #include "platform.hpp"
 
-#if defined XS_FORCE_MUTEXES
-#define XS_ATOMIC_COUNTER_MUTEX
-#elif (defined __i386__ || defined __x86_64__) && defined __GNUC__
-#define XS_ATOMIC_COUNTER_X86
-#elif defined __ARM_ARCH_7A__ && defined __GNUC__
-#define XS_ATOMIC_COUNTER_ARM
-#elif defined XS_HAVE_WINDOWS
-#define XS_ATOMIC_COUNTER_WINDOWS
-#elif (defined XS_HAVE_SOLARIS || defined XS_HAVE_NETBSD)
-#define XS_ATOMIC_COUNTER_ATOMIC_H
+#if defined(XS_ATOMIC_GCC_SYNC)
+#elif (defined(__GNUC__) && (defined(__i386__) || defined(__x86_64__)))
+#elif (defined(__GNUC__) && defined(__ARM_ARCH_7A__))
+#elif defined(XS_ATOMIC_SOLARIS)
+#   include <atomic.h>
+#elif defined(XS_HAVE_WINDOWS)
+#   include "windows.hpp"
 #else
-#define XS_ATOMIC_COUNTER_MUTEX
-#endif
-
-#if defined XS_ATOMIC_COUNTER_MUTEX
-#include "mutex.hpp"
-#elif defined XS_ATOMIC_COUNTER_WINDOWS
-#include "windows.hpp"
-#elif defined XS_ATOMIC_COUNTER_ATOMIC_H
-#include <atomic.h>
+#   define XS_ATOMIC_OVER_MUTEX 1
+#   include "mutex.hpp"
 #endif
 
 namespace xs
@@ -79,20 +69,17 @@ namespace xs
         {
             integer_t old_value;
 
-#if defined XS_ATOMIC_COUNTER_WINDOWS
-            old_value = InterlockedExchangeAdd ((LONG*) &value, increment_);
-#elif defined __GNUC__ && !defined XS_DISABLE_GCC_SYNC_BUILTINS
+#if defined(XS_ATOMIC_GCC_SYNC)
             old_value = __sync_fetch_and_add (&value, increment_);
-#elif defined XS_ATOMIC_COUNTER_ATOMIC_H
-            integer_t new_value = atomic_add_32_nv (&value, increment_);
-            old_value = new_value - increment_;
-#elif defined XS_ATOMIC_COUNTER_X86
+
+#elif (defined(__GNUC__) && (defined(__i386__) || defined(__x86_64__)))
             __asm__ volatile (
                 "lock; xadd %0, %1 \n\t"
                 : "=r" (old_value), "=m" (value)
                 : "0" (increment_), "m" (value)
                 : "cc", "memory");
-#elif defined XS_ATOMIC_COUNTER_ARM
+
+#elif (defined(__GNUC__) && defined(__ARM_ARCH_7A__))
             integer_t flag, tmp;
             __asm__ volatile (
                 "       dmb     sy\n\t"
@@ -105,13 +92,19 @@ namespace xs
                 : "=&r"(old_value), "=&r"(flag), "=&r"(tmp), "+Qo"(value)
                 : "Ir"(increment_), "r"(&value)
                 : "cc");
-#elif defined XS_ATOMIC_COUNTER_MUTEX
+
+#elif defined(XS_ATOMIC_SOLARIS)
+            integer_t new_value = atomic_add_32_nv (&value, increment_);
+            old_value = new_value - increment_;
+
+#elif defined(XS_HAVE_WINDOWS)
+            old_value = InterlockedExchangeAdd ((LONG*) &value, increment_);
+
+#else
             sync.lock ();
             old_value = value;
             value += increment_;
             sync.unlock ();
-#else
-#error atomic_counter is not implemented for this platform
 #endif
             return old_value;
         }
@@ -119,18 +112,11 @@ namespace xs
         //  Atomic subtraction. Returns false if the counter drops to zero.
         inline bool sub (integer_t decrement)
         {
-#if defined XS_ATOMIC_COUNTER_WINDOWS
-            LONG delta = - ((LONG) decrement);
-            integer_t old = InterlockedExchangeAdd ((LONG*) &value, delta);
-            return old - decrement != 0;
-#elif defined __GNUC__ && !defined XS_DISABLE_GCC_SYNC_BUILTINS
+#if defined(XS_ATOMIC_GCC_SYNC)
             integer_t new_value = __sync_sub_and_fetch (&value, decrement);
             return (new_value != 0);
-#elif defined XS_ATOMIC_COUNTER_ATOMIC_H
-            int32_t delta = - ((int32_t) decrement);
-            integer_t nv = atomic_add_32_nv (&value, delta);
-            return nv != 0;
-#elif defined XS_ATOMIC_COUNTER_X86
+
+#elif (defined(__GNUC__) && (defined(__i386__) || defined(__x86_64__)))
             integer_t oldval = -decrement;
             volatile integer_t *val = &value;
             __asm__ volatile ("lock; xaddl %0,%1"
@@ -138,7 +124,8 @@ namespace xs
                 : "0" (oldval), "m" (*val)
                 : "cc", "memory");
             return oldval != decrement;
-#elif defined XS_ATOMIC_COUNTER_ARM
+
+#elif (defined(__GNUC__) && defined(__ARM_ARCH_7A__))
             integer_t old_value, flag, tmp;
             __asm__ volatile (
                 "       dmb     sy\n\t"
@@ -152,14 +139,23 @@ namespace xs
                 : "Ir"(decrement), "r"(&value)
                 : "cc");
             return old_value - decrement != 0;
-#elif defined XS_ATOMIC_COUNTER_MUTEX
+
+#elif defined(XS_ATOMIC_SOLARIS)
+            int32_t delta = - ((int32_t) decrement);
+            integer_t nv = atomic_add_32_nv (&value, delta);
+            return nv != 0;
+
+#elif defined(XS_HAVE_WINDOWS)
+            LONG delta = - ((LONG) decrement);
+            integer_t old = InterlockedExchangeAdd ((LONG*) &value, delta);
+            return old - decrement != 0;
+
+#else
             sync.lock ();
             value -= decrement;
             bool result = value ? true : false;
             sync.unlock ();
             return result;
-#else
-#error atomic_counter is not implemented for this platform
 #endif
         }
 
@@ -171,7 +167,7 @@ namespace xs
     private:
 
         volatile integer_t value;
-#if defined XS_ATOMIC_COUNTER_MUTEX
+#if defined(XS_ATOMIC_OVER_MUTEX)
         mutex_t sync;
 #endif
 
@@ -182,20 +178,8 @@ namespace xs
 }
 
 //  Remove macros local to this file.
-#if defined XS_ATOMIC_COUNTER_WINDOWS
-#undef XS_ATOMIC_COUNTER_WINDOWS
-#endif
-#if defined XS_ATOMIC_COUNTER_ATOMIC_H
-#undef XS_ATOMIC_COUNTER_ATOMIC_H
-#endif
-#if defined XS_ATOMIC_COUNTER_X86
-#undef XS_ATOMIC_COUNTER_X86
-#endif
-#if defined XS_ATOMIC_COUNTER_ARM
-#undef XS_ATOMIC_COUNTER_ARM
-#endif
-#if defined XS_ATOMIC_COUNTER_MUTEX
-#undef XS_ATOMIC_COUNTER_MUTEX
+#if defined(XS_ATOMIC_OVER_MUTEX)
+#   undef XS_ATOMIC_OVER_MUTEX
 #endif
 
 #endif
diff --git a/src/atomic_ptr.hpp b/src/atomic_ptr.hpp
index 2e585b7..08418d3 100644
--- a/src/atomic_ptr.hpp
+++ b/src/atomic_ptr.hpp
@@ -26,26 +26,16 @@
 
 #include "platform.hpp"
 
-#if defined XS_FORCE_MUTEXES
-#define XS_ATOMIC_PTR_MUTEX
-#elif (defined __i386__ || defined __x86_64__) && defined __GNUC__
-#define XS_ATOMIC_PTR_X86
-#elif defined __ARM_ARCH_7A__ && defined __GNUC__
-#define XS_ATOMIC_PTR_ARM
-#elif defined XS_HAVE_WINDOWS
-#define XS_ATOMIC_PTR_WINDOWS
-#elif (defined XS_HAVE_SOLARIS || defined XS_HAVE_NETBSD)
-#define XS_ATOMIC_PTR_ATOMIC_H
+#if defined(XS_ATOMIC_GCC_SYNC)
+#elif (defined(__GNUC__) && (defined(__i386__) || defined(__x86_64__)))
+#elif (defined(__GNUC__) && defined(__ARM_ARCH_7A__))
+#elif defined(XS_ATOMIC_SOLARIS)
+#   include <atomic.h>
+#elif defined(XS_HAVE_WINDOWS)
+#   include "windows.hpp"
 #else
-#define XS_ATOMIC_PTR_MUTEX
-#endif
-
-#if defined XS_ATOMIC_PTR_MUTEX
-#include "mutex.hpp"
-#elif defined XS_ATOMIC_PTR_WINDOWS
-#include "windows.hpp"
-#elif defined XS_ATOMIC_PTR_ATOMIC_H
-#include <atomic.h>
+#   define XS_ATOMIC_OVER_MUTEX 1
+#   include "mutex.hpp"
 #endif
 
 namespace xs
@@ -80,9 +70,7 @@ namespace xs
         //  to the 'val' value. Old value is returned.
         inline T *xchg (T *val_)
         {
-#if defined XS_ATOMIC_PTR_WINDOWS
-            return (T*) InterlockedExchangePointer ((PVOID*) &ptr, val_);
-#elif defined __GNUC__ && !defined XS_DISABLE_GCC_SYNC_BUILTINS
+#if defined(XS_ATOMIC_GCC_SYNC)
             {
                 T* ov;
                 do
@@ -91,16 +79,16 @@ namespace xs
                 } while (!__sync_bool_compare_and_swap (&ptr, ov, val_));
                 return ov;
             }
-#elif defined XS_ATOMIC_PTR_ATOMIC_H
-            return (T*) atomic_swap_ptr (&ptr, val_);
-#elif defined XS_ATOMIC_PTR_X86
+
+#elif (defined(__GNUC__) && (defined(__i386__) || defined(__x86_64__)))
             T *old;
             __asm__ volatile (
                 "lock; xchg %0, %2"
                 : "=r" (old), "=m" (ptr)
                 : "m" (ptr), "0" (val_));
             return old;
-#elif defined XS_ATOMIC_PTR_ARM
+
+#elif (defined(__GNUC__) && defined(__ARM_ARCH_7A__))
             T* old;
             unsigned int flag;
             __asm__ volatile (
@@ -114,14 +102,19 @@ namespace xs
                 : "r"(&ptr), "r"(val_)
                 : "cc");
             return old;
-#elif defined XS_ATOMIC_PTR_MUTEX
+
+#elif defined(XS_ATOMIC_SOLARIS)
+            return (T*) atomic_swap_ptr (&ptr, val_);
+
+#elif defined(XS_HAVE_WINDOWS)
+            return (T*) InterlockedExchangePointer ((PVOID*) &ptr, val_);
+
+#else
             sync.lock ();
             T *old = (T*) ptr;
             ptr = val_;
             sync.unlock ();
             return old;
-#else
-#error atomic_ptr is not implemented for this platform
 #endif
         }
 
@@ -131,14 +124,10 @@ namespace xs
         //  is returned.
         inline T *cas (T *cmp_, T *val_)
         {
-#if defined XS_ATOMIC_PTR_WINDOWS
-            return (T*) InterlockedCompareExchangePointer (
-                (volatile PVOID*) &ptr, val_, cmp_);
-#elif defined __GNUC__ && !defined XS_DISABLE_GCC_SYNC_BUILTINS
+#if defined(XS_ATOMIC_GCC_SYNC)
             return (T*) __sync_val_compare_and_swap (&ptr, cmp_, val_);
-#elif defined XS_ATOMIC_PTR_ATOMIC_H
-            return (T*) atomic_cas_ptr (&ptr, cmp_, val_);
-#elif defined XS_ATOMIC_PTR_X86
+
+#elif (defined(__GNUC__) && (defined(__i386__) || defined(__x86_64__)))
             T *old;
             __asm__ volatile (
                 "lock; cmpxchg %2, %3"
@@ -146,7 +135,8 @@ namespace xs
                 : "r" (val_), "m" (ptr), "0" (cmp_)
                 : "cc");
             return old;
-#elif defined XS_ATOMIC_PTR_ARM
+
+#elif (defined(__GNUC__) && defined(__ARM_ARCH_7A__))
             T *old;
             unsigned int flag;
             __asm__ volatile (
@@ -162,22 +152,28 @@ namespace xs
                 : "r"(&ptr), "r"(cmp_), "r"(val_)
                 : "cc");
             return old;
-#elif defined XS_ATOMIC_PTR_MUTEX
+
+#elif defined(XS_ATOMIC_SOLARIS)
+            return (T*) atomic_cas_ptr (&ptr, cmp_, val_);
+
+#elif defined(XS_HAVE_WINDOWS)
+            return (T*) InterlockedCompareExchangePointer (
+                (volatile PVOID*) &ptr, val_, cmp_);
+
+#else
             sync.lock ();
             T *old = (T*) ptr;
             if (ptr == cmp_)
                 ptr = val_;
             sync.unlock ();
             return old;
-#else
-#error atomic_ptr is not implemented for this platform
 #endif
         }
 
     private:
 
         volatile T *ptr;
-#if defined XS_ATOMIC_PTR_MUTEX
+#if defined XS_ATOMIC_OVER_MUTEX
         mutex_t sync;
 #endif
 
@@ -188,21 +184,8 @@ namespace xs
 }
 
 //  Remove macros local to this file.
-#if defined XS_ATOMIC_PTR_WINDOWS
-#undef XS_ATOMIC_PTR_WINDOWS
-#endif
-#if defined XS_ATOMIC_PTR_ATOMIC_H
-#undef XS_ATOMIC_PTR_ATOMIC_H
-#endif
-#if defined XS_ATOMIC_PTR_X86
-#undef XS_ATOMIC_PTR_X86
-#endif
-#if defined XS_ATOMIC_PTR_ARM
-#undef XS_ATOMIC_PTR_ARM
-#endif
-#if defined XS_ATOMIC_PTR_MUTEX
-#undef XS_ATOMIC_PTR_MUTEX
+#if defined(XS_ATOMIC_OVER_MUTEX)
+#   undef XS_ATOMIC_OVER_MUTEX
 #endif
 
 #endif
-