WL#5138 merged to mysql-next-mr

7 files changed, 207 insertions, 150 deletions

diff --git a/include/atomic/gcc_builtins.h b/include/atomic/gcc_builtins.h
index 01ebc38707e..100ff80cacd 100644
--- a/include/atomic/gcc_builtins.h
+++ b/include/atomic/gcc_builtins.h
@@ -18,7 +18,7 @@
 
 #define make_atomic_add_body(S)                     \
   v= __sync_fetch_and_add(a, v);
-#define make_atomic_swap_body(S)                    \
+#define make_atomic_fas_body(S)                     \
   v= __sync_lock_test_and_set(a, v);
 #define make_atomic_cas_body(S)                     \
   int ## S sav;                                     \
@@ -28,7 +28,10 @@
 #ifdef MY_ATOMIC_MODE_DUMMY
 #define make_atomic_load_body(S)   ret= *a
 #define make_atomic_store_body(S)  *a= v
+#define MY_ATOMIC_MODE "gcc-builtins-up"
+
 #else
+#define MY_ATOMIC_MODE "gcc-builtins-smp"
 #define make_atomic_load_body(S)                    \
   ret= __sync_fetch_and_or(a, 0);
 #define make_atomic_store_body(S)                   \
diff --git a/include/atomic/generic-msvc.h b/include/atomic/generic-msvc.h
new file mode 100644
index 00000000000..f1e1b0e88c9
--- /dev/null
+++ b/include/atomic/generic-msvc.h
@@ -0,0 +1,116 @@
+/* Copyright (C) 2006-2008 MySQL AB, 2008-2009 Sun Microsystems, Inc.
+
+   This program is free software; you can redistribute it and/or modify
+   it under the terms of the GNU General Public License as published by
+   the Free Software Foundation; version 2 of the License.
+
+   This program is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+   GNU General Public License for more details.
+
+   You should have received a copy of the GNU General Public License
+   along with this program; if not, write to the Free Software
+   Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA */
+
+#ifndef _atomic_h_cleanup_
+#define _atomic_h_cleanup_ "atomic/generic-msvc.h"
+
+/*
+  We don't implement anything specific for MY_ATOMIC_MODE_DUMMY, always use
+  intrinsics.
+  8 and 16-bit atomics are not implemented, but it can be done if necessary.
+*/
+#undef MY_ATOMIC_HAS_8_16
+
+/*
+  x86 compilers (both VS2003 or VS2005) never use instrinsics, but generate 
+  function calls to kernel32 instead, even in the optimized build. 
+  We force intrinsics as described in MSDN documentation for 
+  _InterlockedCompareExchange.
+*/
+#ifdef _M_IX86
+
+#if (_MSC_VER >= 1500)
+#include <intrin.h>
+#else
+C_MODE_START
+/*Visual Studio 2003 and earlier do not have prototypes for atomic intrinsics*/
+LONG _InterlockedExchange (LONG volatile *Target,LONG Value);
+LONG _InterlockedCompareExchange (LONG volatile *Target, LONG Value, LONG Comp);
+LONG _InterlockedExchangeAdd (LONG volatile *Addend, LONG Value);
+C_MODE_END
+
+#pragma intrinsic(_InterlockedExchangeAdd)
+#pragma intrinsic(_InterlockedCompareExchange)
+#pragma intrinsic(_InterlockedExchange)
+#endif
+
+#define InterlockedExchange        _InterlockedExchange
+#define InterlockedExchangeAdd     _InterlockedExchangeAdd
+#define InterlockedCompareExchange _InterlockedCompareExchange
+/*
+ No need to do something special for InterlockedCompareExchangePointer
+ as it is a #define to InterlockedCompareExchange. The same applies to
+ InterlockedExchangePointer. 
+*/
+#endif /*_M_IX86*/
+
+#define MY_ATOMIC_MODE "msvc-intrinsics"
+#define IL_EXCHG_ADD32(X,Y)     InterlockedExchangeAdd((volatile LONG *)(X),(Y))
+#define IL_COMP_EXCHG32(X,Y,Z)  InterlockedCompareExchange((volatile LONG *)(X),(Y),(Z))
+#define IL_COMP_EXCHGptr        InterlockedCompareExchangePointer
+#define IL_EXCHG32(X,Y)         InterlockedExchange((volatile LONG *)(X),(Y))
+#define IL_EXCHGptr             InterlockedExchangePointer
+#define make_atomic_add_body(S) \
+  v= IL_EXCHG_ADD ## S (a, v)
+#define make_atomic_cas_body(S)                                 \
+  int ## S initial_cmp= *cmp;                                   \
+  int ## S initial_a= IL_COMP_EXCHG ## S (a, set, initial_cmp); \
+  if (!(ret= (initial_a == initial_cmp))) *cmp= initial_a;
+#define make_atomic_swap_body(S) \
+  v= IL_EXCHG ## S (a, v)
+#define make_atomic_load_body(S)       \
+  ret= 0; /* avoid compiler warning */ \
+  ret= IL_COMP_EXCHG ## S (a, ret, ret);
+
+/*
+  my_yield_processor (equivalent of x86 PAUSE instruction) should be used
+  to improve performance on hyperthreaded CPUs. Intel recommends to use it in
+  spin loops also on non-HT machines to reduce power consumption (see e.g 
+  http://softwarecommunity.intel.com/articles/eng/2004.htm)
+
+  Running benchmarks for spinlocks implemented with InterlockedCompareExchange
+  and YieldProcessor shows that much better performance is achieved by calling
+  YieldProcessor in a loop - that is, yielding longer. On Intel boxes setting
+  loop count in the range 200-300 brought best results.
+ */
+#ifndef YIELD_LOOPS
+#define YIELD_LOOPS 200
+#endif
+
+static __inline int my_yield_processor()
+{
+  int i;
+  for(i=0; i<YIELD_LOOPS; i++)
+  {
+#if (_MSC_VER <= 1310)
+    /* On older compilers YieldProcessor is not available, use inline assembly*/
+    __asm { rep nop }
+#else
+    YieldProcessor();
+#endif
+  }
+  return 1;
+}
+
+#define LF_BACKOFF my_yield_processor()
+#else /* cleanup */
+
+#undef IL_EXCHG_ADD32
+#undef IL_COMP_EXCHG32
+#undef IL_COMP_EXCHGptr
+#undef IL_EXCHG32
+#undef IL_EXCHGptr
+
+#endif
diff --git a/include/atomic/nolock.h b/include/atomic/nolock.h
index d7d87167ade..e4cd9ab9896 100644
--- a/include/atomic/nolock.h
+++ b/include/atomic/nolock.h
@@ -16,43 +16,36 @@
    along with this program; if not, write to the Free Software
    Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA */
 
-#if defined(__i386__) || defined(_M_IX86) || defined(HAVE_GCC_ATOMIC_BUILTINS)
-
-#ifdef MY_ATOMIC_MODE_DUMMY
-#  define LOCK ""
-#else
-#  define LOCK "lock"
-#endif
-
-#ifdef HAVE_GCC_ATOMIC_BUILTINS
-#include "gcc_builtins.h"
-#elif __GNUC__
-#include "x86-gcc.h"
-#elif defined(_MSC_VER)
-#include "x86-msvc.h"
-#endif
-
+#if defined(__i386__) || defined(_MSC_VER) || defined(__x86_64__)   \
+    || defined(HAVE_GCC_ATOMIC_BUILTINS)
+
+#  ifdef MY_ATOMIC_MODE_DUMMY
+#    define LOCK_prefix ""
+#  else
+#    define LOCK_prefix "lock"
+#  endif
+
+#  ifdef HAVE_GCC_ATOMIC_BUILTINS
+#    include "gcc_builtins.h"
+#  elif __GNUC__
+#    include "x86-gcc.h"
+#  elif defined(_MSC_VER)
+#    include "generic-msvc.h"
+#  endif
 #elif defined(HAVE_SOLARIS_ATOMIC)
-
 #include "solaris.h"
-
-#endif /* __i386__ || _M_IX86 || HAVE_GCC_ATOMIC_BUILTINS */
+#endif
 
 #if defined(make_atomic_cas_body) || defined(MY_ATOMICS_MADE)
 /*
  * We have atomics that require no locking
  */
 #define	MY_ATOMIC_NOLOCK
-
-#ifdef __SUNPRO_C
 /*
- * Sun Studio 12 (and likely earlier) does not accept a typedef struct {}
- */
-typedef char my_atomic_rwlock_t;
-#else
-typedef struct { } my_atomic_rwlock_t;
-#endif
-
+  Type not used so minimal size (emptry struct has different size between C
+  and C++, zero-length array is gcc-specific).
+*/
+typedef char my_atomic_rwlock_t __attribute__ ((unused));
 #define my_atomic_rwlock_destroy(name)
 #define my_atomic_rwlock_init(name)
 #define my_atomic_rwlock_rdlock(name)
diff --git a/include/atomic/rwlock.h b/include/atomic/rwlock.h
index 0ff4d16c545..29e22fcb3d5 100644
--- a/include/atomic/rwlock.h
+++ b/include/atomic/rwlock.h
@@ -16,7 +16,8 @@
    along with this program; if not, write to the Free Software
    Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA */
 
-typedef struct {pthread_rwlock_t rw;} my_atomic_rwlock_t;
+typedef struct {pthread_mutex_t rw;} my_atomic_rwlock_t;
+#define MY_ATOMIC_MODE_RWLOCKS 1
 
 #ifdef MY_ATOMIC_MODE_DUMMY
 /*
@@ -34,17 +35,25 @@ typedef struct {pthread_rwlock_t rw;} my_atomic_rwlock_t;
 #define my_atomic_rwlock_wrunlock(name)
 #define MY_ATOMIC_MODE "dummy (non-atomic)"
 #else
-#define my_atomic_rwlock_destroy(name)     pthread_rwlock_destroy(& (name)->rw)
-#define my_atomic_rwlock_init(name)        pthread_rwlock_init(& (name)->rw, 0)
-#define my_atomic_rwlock_rdlock(name)      pthread_rwlock_rdlock(& (name)->rw)
-#define my_atomic_rwlock_wrlock(name)      pthread_rwlock_wrlock(& (name)->rw)
-#define my_atomic_rwlock_rdunlock(name)    pthread_rwlock_unlock(& (name)->rw)
-#define my_atomic_rwlock_wrunlock(name)    pthread_rwlock_unlock(& (name)->rw)
-#define MY_ATOMIC_MODE "rwlocks"
+/*
+  we're using read-write lock macros but map them to mutex locks, and they're
+  faster. Still, having semantically rich API we can change the
+  underlying implementation, if necessary.
+*/
+#define my_atomic_rwlock_destroy(name)     pthread_mutex_destroy(& (name)->rw)
+#define my_atomic_rwlock_init(name)        pthread_mutex_init(& (name)->rw, 0)
+#define my_atomic_rwlock_rdlock(name)      pthread_mutex_lock(& (name)->rw)
+#define my_atomic_rwlock_wrlock(name)      pthread_mutex_lock(& (name)->rw)
+#define my_atomic_rwlock_rdunlock(name)    pthread_mutex_unlock(& (name)->rw)
+#define my_atomic_rwlock_wrunlock(name)    pthread_mutex_unlock(& (name)->rw)
+#define MY_ATOMIC_MODE "mutex"
+#ifndef MY_ATOMIC_MODE_RWLOCKS
+#define MY_ATOMIC_MODE_RWLOCKS 1
+#endif
 #endif
 
 #define make_atomic_add_body(S)     int ## S sav; sav= *a; *a+= v; v=sav;
-#define make_atomic_swap_body(S)    int ## S sav; sav= *a; *a= v; v=sav;
+#define make_atomic_fas_body(S)     int ## S sav; sav= *a; *a= v; v=sav;
 #define make_atomic_cas_body(S)     if ((ret= (*a == *cmp))) *a= set; else *cmp=*a;
 #define make_atomic_load_body(S)    ret= *a;
 #define make_atomic_store_body(S)   *a= v;
diff --git a/include/atomic/solaris.h b/include/atomic/solaris.h
index 4c51253d2d5..45efd9faaba 100644
--- a/include/atomic/solaris.h
+++ b/include/atomic/solaris.h
@@ -186,25 +186,25 @@ my_atomic_storeptr(void * volatile *a, void *v)
 /* ------------------------------------------------------------------------ */
 
 STATIC_INLINE int8
-my_atomic_swap8(int8 volatile *a, int8 v)
+my_atomic_fas8(int8 volatile *a, int8 v)
 {
 	return ((int8) atomic_swap_8((volatile uint8_t *)a, (uint8_t)v));
 }
 
 STATIC_INLINE int16
-my_atomic_swap16(int16 volatile *a, int16 v)
+my_atomic_fas16(int16 volatile *a, int16 v)
 {
 	return ((int16) atomic_swap_16((volatile uint16_t *)a, (uint16_t)v));
 }
 
 STATIC_INLINE int32
-my_atomic_swap32(int32 volatile *a, int32 v)
+my_atomic_fas32(int32 volatile *a, int32 v)
 {
 	return ((int32) atomic_swap_32((volatile uint32_t *)a, (uint32_t)v));
 }
 
 STATIC_INLINE void *
-my_atomic_swapptr(void * volatile *a, void *v)
+my_atomic_fasptr(void * volatile *a, void *v)
 {
 	return (atomic_swap_ptr(a, v));
 }
diff --git a/include/atomic/x86-gcc.h b/include/atomic/x86-gcc.h
index c3029f9c1b4..cbdb2d11e68 100644
--- a/include/atomic/x86-gcc.h
+++ b/include/atomic/x86-gcc.h
@@ -22,10 +22,18 @@
   architectures support double-word (128-bit) cas.
 */
 
-#ifdef MY_ATOMIC_NO_XADD
-#define MY_ATOMIC_MODE "gcc-x86" LOCK "-no-xadd"
+#ifdef __x86_64__
+#  ifdef MY_ATOMIC_NO_XADD
+#    define MY_ATOMIC_MODE "gcc-amd64" LOCK_prefix "-no-xadd"
+#  else
+#    define MY_ATOMIC_MODE "gcc-amd64" LOCK_prefix
+#  endif
 #else
-#define MY_ATOMIC_MODE "gcc-x86" LOCK
+#  ifdef MY_ATOMIC_NO_XADD
+#    define MY_ATOMIC_MODE "gcc-x86" LOCK_prefix "-no-xadd"
+#  else
+#    define MY_ATOMIC_MODE "gcc-x86" LOCK_prefix
+#  endif
 #endif
 
 /* fix -ansi errors while maintaining readability */
@@ -34,29 +42,53 @@
 #endif
 
 #ifndef MY_ATOMIC_NO_XADD
-#define make_atomic_add_body(S)					\
-  asm volatile (LOCK "; xadd %0, %1;" : "+r" (v) , "+m" (*a))
+#define make_atomic_add_body(S)         make_atomic_add_body ## S
+#define make_atomic_cas_body(S)         make_atomic_cas_body ## S
 #endif
-#define make_atomic_swap_body(S)				\
-  asm volatile ("; xchg %0, %1;" : "+q" (v) , "+m" (*a))
-#define make_atomic_cas_body(S)					\
-  asm volatile (LOCK "; cmpxchg %3, %0; setz %2;"		\
+
+#define make_atomic_add_body32                                  \
+  asm volatile (LOCK_prefix "; xadd %0, %1;" : "+r" (v) , "+m" (*a))
+
+#define make_atomic_cas_body32                                  \
+  asm volatile (LOCK_prefix "; cmpxchg %3, %0; setz %2;"	\
                : "+m" (*a), "+a" (*cmp), "=q" (ret): "r" (set))
 
+#define make_atomic_cas_bodyptr make_atomic_cas_body32
+
+#ifndef __x86_64__
+#define make_atomic_add_body64 make_atomic_add_body32
+#define make_atomic_cas_body64 make_atomic_cas_body32
+#else
+#define make_atomic_add_body64                                  \
+  int64 tmp=*a;                                                 \
+  while (!my_atomic_cas64(a, &tmp, tmp+v));                     \
+  v=tmp;
+#define make_atomic_cas_body64                                  \
+  int32 ebx=(set & 0xFFFFFFFF), ecx=(set >> 32);                \
+  asm volatile (LOCK_prefix "; cmpxchg8b %0; setz %2;"         \
+               : "+m" (*a), "+A" (*cmp), "=q" (ret)             \
+               :"b" (ebx), "c" (ecx))
+#endif
+
+#define make_atomic_fas_body(S)                                        \
+  asm volatile ("xchg %0, %1;" : "+r" (v) , "+m" (*a))
+
 #ifdef MY_ATOMIC_MODE_DUMMY
 #define make_atomic_load_body(S)   ret=*a
 #define make_atomic_store_body(S)  *a=v
 #else
 /*
   Actually 32-bit reads/writes are always atomic on x86
-  But we add LOCK here anyway to force memory barriers
+  But we add LOCK_prefix here anyway to force memory barriers
 */
 #define make_atomic_load_body(S)				\
   ret=0;							\
-  asm volatile (LOCK "; cmpxchg %2, %0"				\
+  asm volatile (LOCK_prefix "; cmpxchg %2, %0"			\
                : "+m" (*a), "+a" (ret): "r" (ret))
 #define make_atomic_store_body(S)				\
-  asm volatile ("; xchg %0, %1;" : "+m" (*a) : "r" (v))
+  asm volatile ("; xchg %0, %1;" : "+m" (*a), "+r" (v))
 #endif
 
+/* TODO test on intel whether the below helps. on AMD it makes no difference */
+//#define LF_BACKOFF ({asm volatile ("rep; nop"); 1; })
 #endif /* ATOMIC_X86_GCC_INCLUDED */
diff --git a/include/atomic/x86-msvc.h b/include/atomic/x86-msvc.h
deleted file mode 100644
index c4885bb8451..00000000000
--- a/include/atomic/x86-msvc.h
+++ /dev/null
@@ -1,96 +0,0 @@
-/* Copyright (C) 2006 MySQL AB
-
-   This program is free software; you can redistribute it and/or modify
-   it under the terms of the GNU General Public License as published by
-   the Free Software Foundation; version 2 of the License.
-
-   This program is distributed in the hope that it will be useful,
-   but WITHOUT ANY WARRANTY; without even the implied warranty of
-   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-   GNU General Public License for more details.
-
-   You should have received a copy of the GNU General Public License
-   along with this program; if not, write to the Free Software
-   Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA */
-
-/*
-  XXX 64-bit atomic operations can be implemented using
-  cmpxchg8b, if necessary
-*/
-
-// Would it be better to use intrinsics ?
-// (InterlockedCompareExchange, InterlockedCompareExchange16
-// InterlockedExchangeAdd, InterlockedExchange)
-
-#ifndef _atomic_h_cleanup_
-#define _atomic_h_cleanup_ "atomic/x86-msvc.h"
-
-#define MY_ATOMIC_MODE "msvc-x86" LOCK
-
-#define make_atomic_add_body(S)				\
-  _asm {						\
-    _asm mov   reg_ ## S, v				\
-    _asm LOCK  xadd *a, reg_ ## S			\
-    _asm movzx v, reg_ ## S				\
-  }
-#define make_atomic_cas_body(S)				\
-  _asm {						\
-    _asm mov    areg_ ## S, *cmp			\
-    _asm mov    reg2_ ## S, set				\
-    _asm LOCK cmpxchg *a, reg2_ ## S			\
-    _asm mov    *cmp, areg_ ## S			\
-    _asm setz   al					\
-    _asm movzx  ret, al					\
-  }
-#define make_atomic_swap_body(S)			\
-  _asm {						\
-    _asm mov    reg_ ## S, v				\
-    _asm xchg   *a, reg_ ## S				\
-    _asm mov    v, reg_ ## S				\
-  }
-
-#ifdef MY_ATOMIC_MODE_DUMMY
-#define make_atomic_load_body(S)        ret=*a
-#define make_atomic_store_body(S)       *a=v
-#else
-/*
-  Actually 32-bit reads/writes are always atomic on x86
-  But we add LOCK here anyway to force memory barriers
-*/
-#define make_atomic_load_body(S)			\
-  _asm {						\
-    _asm mov    areg_ ## S, 0				\
-    _asm mov    reg2_ ## S, areg_ ## S			\
-    _asm LOCK cmpxchg *a, reg2_ ## S			\
-    _asm mov    ret, areg_ ## S				\
-  }
-#define make_atomic_store_body(S)			\
-  _asm {						\
-    _asm mov    reg_ ## S, v				\
-    _asm xchg   *a, reg_ ## S				\
-  }
-#endif
-
-#define reg_8           al
-#define reg_16          ax
-#define reg_32          eax
-#define areg_8          al
-#define areg_16         ax
-#define areg_32         eax
-#define reg2_8          bl
-#define reg2_16         bx
-#define reg2_32         ebx
-
-#else /* cleanup */
-
-#undef reg_8
-#undef reg_16
-#undef reg_32
-#undef areg_8
-#undef areg_16
-#undef areg_32
-#undef reg2_8
-#undef reg2_16
-#undef reg2_32
-#endif
-