Auto-merge from mysql-next-mr.

11 files changed, 712 insertions, 218 deletions

diff --git a/include/Makefile.am b/include/Makefile.am
index 0a7a35bd6fa..83a22f1beec 100644
--- a/include/Makefile.am
+++ b/include/Makefile.am
@@ -30,7 +30,7 @@ pkginclude_HEADERS =	$(HEADERS_ABI) my_dbug.h m_string.h my_sys.h \
 			m_ctype.h my_attribute.h $(HEADERS_GEN_CONFIGURE) \
 			$(HEADERS_GEN_MAKE) probes_mysql.h probes_mysql_nodtrace.h
 
-noinst_HEADERS =	config-win.h config-netware.h my_bit.h \
+noinst_HEADERS =	config-win.h config-netware.h lf.h my_bit.h \
 			heap.h my_bitmap.h my_uctype.h \
 			myisam.h myisampack.h myisammrg.h ft_global.h\
 			mysys_err.h my_base.h help_start.h help_end.h \
@@ -39,9 +39,9 @@ noinst_HEADERS =	config-win.h config-netware.h my_bit.h \
 			thr_lock.h t_ctype.h violite.h my_md5.h base64.h \
 			my_handler.h my_time.h service_versions.h \
 			my_vle.h my_user.h my_atomic.h atomic/nolock.h \
-			atomic/rwlock.h atomic/x86-gcc.h atomic/x86-msvc.h \
-                        atomic/solaris.h \
-			atomic/gcc_builtins.h my_libwrap.h my_stacktrace.h
+			atomic/rwlock.h atomic/x86-gcc.h atomic/generic-msvc.h \
+			atomic/gcc_builtins.h my_libwrap.h my_stacktrace.h \
+			atomic/solaris.h
 
 EXTRA_DIST =        mysql.h.pp mysql/plugin.h.pp probes_mysql.d.base 
 
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..a31f8ed6ca1 100644
--- a/include/atomic/rwlock.h
+++ b/include/atomic/rwlock.h
@@ -1,7 +1,7 @@
 #ifndef ATOMIC_RWLOCK_INCLUDED
 #define ATOMIC_RWLOCK_INCLUDED
 
-/* Copyright (C) 2006 MySQL AB
+/* Copyright (C) 2006 MySQL AB, 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
@@ -16,7 +16,7 @@
    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;
+#define MY_ATOMIC_MODE_RWLOCKS 1
 
 #ifdef MY_ATOMIC_MODE_DUMMY
 /*
@@ -26,6 +26,9 @@ typedef struct {pthread_rwlock_t rw;} my_atomic_rwlock_t;
   implementations (another way is to run a UP build on an SMP box).
 */
 #warning MY_ATOMIC_MODE_DUMMY and MY_ATOMIC_MODE_RWLOCKS are incompatible
+
+typedef char my_atomic_rwlock_t;
+
 #define my_atomic_rwlock_destroy(name)
 #define my_atomic_rwlock_init(name)
 #define my_atomic_rwlock_rdlock(name)
@@ -33,18 +36,63 @@ typedef struct {pthread_rwlock_t rw;} my_atomic_rwlock_t;
 #define my_atomic_rwlock_rdunlock(name)
 #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"
+#else /* not MY_ATOMIC_MODE_DUMMY */
+
+typedef struct {pthread_mutex_t rw;} my_atomic_rwlock_t;
+
+#ifndef SAFE_MUTEX
+
+/*
+  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)
+
+#else /* SAFE_MUTEX */
+
+/*
+  SAFE_MUTEX pollutes the compiling name space with macros
+  that alter pthread_mutex_t, pthread_mutex_init, etc.
+  Atomic operations should never use the safe mutex wrappers.
+  Unfortunately, there is no way to have both:
+  - safe mutex macros expanding pthread_mutex_lock to safe_mutex_lock
+  - my_atomic macros expanding to unmodified pthread_mutex_lock
+  inlined in the same compilation unit.
+  So, in case of SAFE_MUTEX, a function call is required.
+  Given that SAFE_MUTEX is a debugging facility,
+  this extra function call is not a performance concern for
+  production builds.
+*/
+C_MODE_START
+extern void plain_pthread_mutex_init(safe_mutex_t *);
+extern void plain_pthread_mutex_destroy(safe_mutex_t *);
+extern void plain_pthread_mutex_lock(safe_mutex_t *);
+extern void plain_pthread_mutex_unlock(safe_mutex_t *);
+C_MODE_END
+
+#define my_atomic_rwlock_destroy(name)     plain_pthread_mutex_destroy(&(name)->rw)
+#define my_atomic_rwlock_init(name)        plain_pthread_mutex_init(&(name)->rw)
+#define my_atomic_rwlock_rdlock(name)      plain_pthread_mutex_lock(&(name)->rw)
+#define my_atomic_rwlock_wrlock(name)      plain_pthread_mutex_lock(&(name)->rw)
+#define my_atomic_rwlock_rdunlock(name)    plain_pthread_mutex_unlock(&(name)->rw)
+#define my_atomic_rwlock_wrunlock(name)    plain_pthread_mutex_unlock(&(name)->rw)
+
+#endif /* SAFE_MUTEX */
+
+#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..59090bc26b7 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 */
@@ -35,12 +43,12 @@
 
 #ifndef MY_ATOMIC_NO_XADD
 #define make_atomic_add_body(S)					\
-  asm volatile (LOCK "; xadd %0, %1;" : "+r" (v) , "+m" (*a))
+  asm volatile (LOCK_prefix "; xadd %0, %1;" : "+r" (v) , "+m" (*a))
 #endif
-#define make_atomic_swap_body(S)				\
-  asm volatile ("; xchg %0, %1;" : "+q" (v) , "+m" (*a))
+#define make_atomic_fas_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;"		\
+  asm volatile (LOCK_prefix "; cmpxchg %3, %0; setz %2;"	\
                : "+m" (*a), "+a" (*cmp), "=q" (ret): "r" (set))
 
 #ifdef MY_ATOMIC_MODE_DUMMY
@@ -49,14 +57,14 @@
 #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
 
 #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
-
diff --git a/include/lf.h b/include/lf.h
new file mode 100644
index 00000000000..7e8f05f4ada
--- /dev/null
+++ b/include/lf.h
@@ -0,0 +1,268 @@
+/* Copyright (C) 2007-2008 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 _lf_h
+#define _lf_h
+
+#include <my_atomic.h>
+
+C_MODE_START
+
+/*
+  Helpers to define both func() and _func(), where
+  func() is a _func() protected by my_atomic_rwlock_wrlock()
+*/
+
+#define lock_wrap(f, t, proto_args, args, lock) \
+t _ ## f proto_args;                            \
+static inline t f  proto_args                   \
+{                                               \
+  t ret;                                        \
+  my_atomic_rwlock_wrlock(lock);                \
+  ret= _ ## f args;                             \
+  my_atomic_rwlock_wrunlock(lock);              \
+  return ret;                                   \
+}
+
+#define lock_wrap_void(f, proto_args, args, lock) \
+void _ ## f proto_args;                         \
+static inline void f proto_args                 \
+{                                               \
+  my_atomic_rwlock_wrlock(lock);                \
+  _ ## f args;                                  \
+  my_atomic_rwlock_wrunlock(lock);              \
+}
+
+#define nolock_wrap(f, t, proto_args, args)     \
+t _ ## f proto_args;                            \
+static inline t f  proto_args                   \
+{                                               \
+  return _ ## f args;                           \
+}
+
+#define nolock_wrap_void(f, proto_args, args)   \
+void _ ## f proto_args;                         \
+static inline void f proto_args                 \
+{                                               \
+  _ ## f args;                                  \
+}
+
+/*
+  wait-free dynamic array, see lf_dynarray.c
+
+  4 levels of 256 elements each mean 4311810304 elements in an array - it
+  should be enough for a while
+*/
+#define LF_DYNARRAY_LEVEL_LENGTH 256
+#define LF_DYNARRAY_LEVELS       4
+
+typedef struct {
+  void * volatile level[LF_DYNARRAY_LEVELS];
+  uint size_of_element;
+  my_atomic_rwlock_t lock;
+} LF_DYNARRAY;
+
+typedef int (*lf_dynarray_func)(void *, void *);
+
+void lf_dynarray_init(LF_DYNARRAY *array, uint element_size);
+void lf_dynarray_destroy(LF_DYNARRAY *array);
+
+nolock_wrap(lf_dynarray_value, void *,
+            (LF_DYNARRAY *array, uint idx),
+            (array, idx))
+lock_wrap(lf_dynarray_lvalue, void *,
+          (LF_DYNARRAY *array, uint idx),
+          (array, idx),
+          &array->lock)
+nolock_wrap(lf_dynarray_iterate, int,
+            (LF_DYNARRAY *array, lf_dynarray_func func, void *arg),
+            (array, func, arg))
+
+/*
+  pin manager for memory allocator, lf_alloc-pin.c
+*/
+
+#define LF_PINBOX_PINS 4
+#define LF_PURGATORY_SIZE 10
+
+typedef void lf_pinbox_free_func(void *, void *, void*);
+
+typedef struct {
+  LF_DYNARRAY pinarray;
+  lf_pinbox_free_func *free_func;
+  void *free_func_arg;
+  uint free_ptr_offset;
+  uint32 volatile pinstack_top_ver;         /* this is a versioned pointer */
+  uint32 volatile pins_in_array;            /* number of elements in array */
+} LF_PINBOX;
+
+typedef struct {
+  void * volatile pin[LF_PINBOX_PINS];
+  LF_PINBOX *pinbox;
+  void  **stack_ends_here;
+  void  *purgatory;
+  uint32 purgatory_count;
+  uint32 volatile link;
+/* we want sizeof(LF_PINS) to be 64 to avoid false sharing */
+#if SIZEOF_INT*2+SIZEOF_CHARP*(LF_PINBOX_PINS+3) != 64
+  char pad[64-sizeof(uint32)*2-sizeof(void*)*(LF_PINBOX_PINS+3)];
+#endif
+} LF_PINS;
+
+/*
+  shortcut macros to do an atomic_wrlock on a structure that uses pins
+  (e.g. lf_hash).
+*/
+#define lf_rwlock_by_pins(PINS)   \
+  my_atomic_rwlock_wrlock(&(PINS)->pinbox->pinarray.lock)
+#define lf_rwunlock_by_pins(PINS) \
+  my_atomic_rwlock_wrunlock(&(PINS)->pinbox->pinarray.lock)
+
+/*
+  compile-time assert, to require "no less than N" pins
+  it's enough if it'll fail on at least one compiler, so
+  we'll enable it on GCC only, which supports zero-length arrays.
+*/
+#if defined(__GNUC__) && defined(MY_LF_EXTRA_DEBUG)
+#define LF_REQUIRE_PINS(N)                                      \
+  static const char require_pins[LF_PINBOX_PINS-N]              \
+                             __attribute__ ((unused));          \
+  static const int LF_NUM_PINS_IN_THIS_FILE= N;
+#define _lf_pin(PINS, PIN, ADDR)                                \
+  (                                                             \
+    assert(PIN < LF_NUM_PINS_IN_THIS_FILE),                     \
+    my_atomic_storeptr(&(PINS)->pin[PIN], (ADDR))               \
+  )
+#else
+#define LF_REQUIRE_PINS(N)
+#define _lf_pin(PINS, PIN, ADDR)  my_atomic_storeptr(&(PINS)->pin[PIN], (ADDR))
+#endif
+
+#define _lf_unpin(PINS, PIN)      _lf_pin(PINS, PIN, NULL)
+#define lf_pin(PINS, PIN, ADDR)   \
+  do {                            \
+    lf_rwlock_by_pins(PINS);      \
+    _lf_pin(PINS, PIN, ADDR);     \
+    lf_rwunlock_by_pins(PINS);    \
+  } while (0)
+#define lf_unpin(PINS, PIN)  lf_pin(PINS, PIN, NULL)
+#define _lf_assert_pin(PINS, PIN) assert((PINS)->pin[PIN] != 0)
+#define _lf_assert_unpin(PINS, PIN) assert((PINS)->pin[PIN] == 0)
+
+void lf_pinbox_init(LF_PINBOX *pinbox, uint free_ptr_offset,
+                    lf_pinbox_free_func *free_func, void * free_func_arg);
+void lf_pinbox_destroy(LF_PINBOX *pinbox);
+
+lock_wrap(lf_pinbox_get_pins, LF_PINS *,
+          (LF_PINBOX *pinbox),
+          (pinbox),
+          &pinbox->pinarray.lock)
+lock_wrap_void(lf_pinbox_put_pins,
+               (LF_PINS *pins),
+               (pins),
+               &pins->pinbox->pinarray.lock)
+lock_wrap_void(lf_pinbox_free,
+               (LF_PINS *pins, void *addr),
+               (pins, addr),
+               &pins->pinbox->pinarray.lock)
+
+/*
+  memory allocator, lf_alloc-pin.c
+*/
+
+typedef struct st_lf_allocator {
+  LF_PINBOX pinbox;
+  uchar * volatile top;
+  uint element_size;
+  uint32 volatile mallocs;
+  void (*constructor)(uchar *); /* called, when an object is malloc()'ed */
+  void (*destructor)(uchar *);  /* called, when an object is free()'d    */
+} LF_ALLOCATOR;
+
+void lf_alloc_init(LF_ALLOCATOR *allocator, uint size, uint free_ptr_offset);
+void lf_alloc_destroy(LF_ALLOCATOR *allocator);
+uint lf_alloc_pool_count(LF_ALLOCATOR *allocator);
+/*
+  shortcut macros to access underlying pinbox functions from an LF_ALLOCATOR
+  see _lf_pinbox_get_pins() and _lf_pinbox_put_pins()
+*/
+#define _lf_alloc_free(PINS, PTR)     _lf_pinbox_free((PINS), (PTR))
+#define lf_alloc_free(PINS, PTR)       lf_pinbox_free((PINS), (PTR))
+#define _lf_alloc_get_pins(A)         _lf_pinbox_get_pins(&(A)->pinbox)
+#define lf_alloc_get_pins(A)           lf_pinbox_get_pins(&(A)->pinbox)
+#define _lf_alloc_put_pins(PINS)      _lf_pinbox_put_pins(PINS)
+#define lf_alloc_put_pins(PINS)        lf_pinbox_put_pins(PINS)
+#define lf_alloc_direct_free(ALLOC, ADDR) my_free((uchar*)(ADDR), MYF(0))
+
+lock_wrap(lf_alloc_new, void *,
+          (LF_PINS *pins),
+          (pins),
+          &pins->pinbox->pinarray.lock)
+
+C_MODE_END
+
+/*
+  extendible hash, lf_hash.c
+*/
+#include <hash.h>
+
+C_MODE_START
+
+#define LF_HASH_UNIQUE 1
+
+/* lf_hash overhead per element (that is, sizeof(LF_SLIST) */
+extern const int LF_HASH_OVERHEAD;
+
+typedef struct {
+  LF_DYNARRAY array;                    /* hash itself */
+  LF_ALLOCATOR alloc;                   /* allocator for elements */
+  my_hash_get_key get_key;              /* see HASH */
+  CHARSET_INFO *charset;                /* see HASH */
+  uint key_offset, key_length;          /* see HASH */
+  uint element_size;                    /* size of memcpy'ed area on insert */
+  uint flags;                           /* LF_HASH_UNIQUE, etc */
+  int32 volatile size;                  /* size of array */
+  int32 volatile count;                 /* number of elements in the hash */
+} LF_HASH;
+
+void lf_hash_init(LF_HASH *hash, uint element_size, uint flags,
+                  uint key_offset, uint key_length, my_hash_get_key get_key,
+                  CHARSET_INFO *charset);
+void lf_hash_destroy(LF_HASH *hash);
+int lf_hash_insert(LF_HASH *hash, LF_PINS *pins, const void *data);
+void *lf_hash_search(LF_HASH *hash, LF_PINS *pins, const void *key, uint keylen);
+int lf_hash_delete(LF_HASH *hash, LF_PINS *pins, const void *key, uint keylen);
+/*
+  shortcut macros to access underlying pinbox functions from an LF_HASH
+  see _lf_pinbox_get_pins() and _lf_pinbox_put_pins()
+*/
+#define _lf_hash_get_pins(HASH)     _lf_alloc_get_pins(&(HASH)->alloc)
+#define lf_hash_get_pins(HASH)       lf_alloc_get_pins(&(HASH)->alloc)
+#define _lf_hash_put_pins(PINS)     _lf_pinbox_put_pins(PINS)
+#define lf_hash_put_pins(PINS)       lf_pinbox_put_pins(PINS)
+#define lf_hash_search_unpin(PINS)   lf_unpin((PINS), 2)
+/*
+  cleanup
+*/
+
+#undef lock_wrap_void
+#undef lock_wrap
+#undef nolock_wrap_void
+#undef nolock_wrap
+
+C_MODE_END
+
+#endif
+
diff --git a/include/my_atomic.h b/include/my_atomic.h
index f5da6e6a0d9..85cf87165fb 100644
--- a/include/my_atomic.h
+++ b/include/my_atomic.h
@@ -16,9 +16,51 @@
    along with this program; if not, write to the Free Software
    Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA */
 
+/*
+  This header defines five atomic operations:
+
+  my_atomic_add#(&var, what)
+    add 'what' to *var, and return the old value of *var
+
+  my_atomic_fas#(&var, what)
+    'Fetch And Store'
+    store 'what' in *var, and return the old value of *var
+
+  my_atomic_cas#(&var, &old, new)
+    'Compare And Swap'
+    if *var is equal to *old, then store 'new' in *var, and return TRUE
+    otherwise store *var in *old, and return FALSE
+
+  my_atomic_load#(&var)
+    return *var
+
+  my_atomic_store#(&var, what)
+    store 'what' in *var
+
+  '#' is substituted by a size suffix - 8, 16, 32, or ptr
+  (e.g. my_atomic_add8, my_atomic_fas32, my_atomic_casptr).
+
+  NOTE This operations are not always atomic, so they always must be
+  enclosed in my_atomic_rwlock_rdlock(lock)/my_atomic_rwlock_rdunlock(lock)
+  or my_atomic_rwlock_wrlock(lock)/my_atomic_rwlock_wrunlock(lock).
+  Hint: if a code block makes intensive use of atomic ops, it make sense
+  to take/release rwlock once for the whole block, not for every statement.
+
+  On architectures where these operations are really atomic, rwlocks will
+  be optimized away.
+  8- and 16-bit atomics aren't implemented for windows (see generic-msvc.h),
+  but can be added, if necessary.
+*/
+
 #ifndef my_atomic_rwlock_init
 
 #define intptr         void *
+/**
+  On most platforms we implement 8-bit, 16-bit, 32-bit and "pointer"
+  operations. Thus the symbol below is defined by default; platforms
+  where we leave out 8-bit or 16-bit operations should undefine it.
+*/
+#define MY_ATOMIC_HAS_8_16 1
 
 #ifndef MY_ATOMIC_MODE_RWLOCKS
 /*
@@ -27,124 +69,223 @@
 #include "atomic/nolock.h"
 #endif
 
-#ifndef MY_ATOMIC_NOLOCK
-/*
- * Have to use rw-locks for atomic ops
- */
+#ifndef make_atomic_cas_body
+/* nolock.h was not able to generate even a CAS function, fall back */
 #include "atomic/rwlock.h"
-#endif
-
-#ifndef MY_ATOMICS_MADE
-
+#else
+/* define missing functions by using the already generated ones */
 #ifndef make_atomic_add_body
-#define make_atomic_add_body(S)					\
+#define make_atomic_add_body(S)                                 \
   int ## S tmp=*a;                                              \
   while (!my_atomic_cas ## S(a, &tmp, tmp+v));                  \
   v=tmp;
 #endif
+#ifndef make_atomic_fas_body
+#define make_atomic_fas_body(S)                                 \
+  int ## S tmp=*a;                                              \
+  while (!my_atomic_cas ## S(a, &tmp, v));                      \
+  v=tmp;
+#endif
+#ifndef make_atomic_load_body
+#define make_atomic_load_body(S)                                \
+  ret= 0; /* avoid compiler warning */                          \
+  (void)(my_atomic_cas ## S(a, &ret, ret));
+#endif
+#ifndef make_atomic_store_body
+#define make_atomic_store_body(S)                               \
+  (void)(my_atomic_fas ## S (a, v));
+#endif
+#endif
+
+/*
+  transparent_union doesn't work in g++
+  Bug ?
+
+  Darwin's gcc doesn't want to put pointers in a transparent_union
+  when built with -arch ppc64. Complains:
+  warning: 'transparent_union' attribute ignored
+*/
+#if defined(__GNUC__) && !defined(__cplusplus) && \
+      ! (defined(__APPLE__) && defined(_ARCH_PPC64))
+/*
+  we want to be able to use my_atomic_xxx functions with
+  both signed and unsigned integers. But gcc will issue a warning
+  "passing arg N of `my_atomic_XXX' as [un]signed due to prototype"
+  if the signedness of the argument doesn't match the prototype, or
+  "pointer targets in passing argument N of my_atomic_XXX differ in signedness"
+  if int* is used where uint* is expected (or vice versa).
+  Let's shut these warnings up
+*/
+#define make_transparent_unions(S)                              \
+        typedef union {                                         \
+          int  ## S  i;                                         \
+          uint ## S  u;                                         \
+        } U_ ## S   __attribute__ ((transparent_union));        \
+        typedef union {                                         \
+          int  ## S volatile *i;                                \
+          uint ## S volatile *u;                                \
+        } Uv_ ## S   __attribute__ ((transparent_union));
+#define uintptr intptr
+make_transparent_unions(8)
+make_transparent_unions(16)
+make_transparent_unions(32)
+make_transparent_unions(ptr)
+#undef uintptr
+#undef make_transparent_unions
+#define a       U_a.i
+#define cmp     U_cmp.i
+#define v       U_v.i
+#define set     U_set.i
+#else
+#define U_8    int8
+#define U_16   int16
+#define U_32   int32
+#define U_ptr  intptr
+#define Uv_8   int8
+#define Uv_16  int16
+#define Uv_32  int32
+#define Uv_ptr intptr
+#define U_a    volatile *a
+#define U_cmp  *cmp
+#define U_v    v
+#define U_set  set
+#endif /* __GCC__ transparent_union magic */
 
 #ifdef HAVE_INLINE
 
-#define make_atomic_add(S)					\
-STATIC_INLINE int ## S my_atomic_add ## S(			\
-                        int ## S volatile *a, int ## S v)	\
-{								\
-  make_atomic_add_body(S);					\
-  return v;							\
+#define make_atomic_cas(S)                                      \
+STATIC_INLINE int my_atomic_cas ## S(Uv_ ## S U_a,              \
+                            Uv_ ## S U_cmp, U_ ## S U_set)      \
+{                                                               \
+  int8 ret;                                                     \
+  make_atomic_cas_body(S);                                      \
+  return ret;                                                   \
 }
 
-#define make_atomic_swap(S)					\
-STATIC_INLINE int ## S my_atomic_swap ## S(			\
-                         int ## S volatile *a, int ## S v)	\
-{								\
-  make_atomic_swap_body(S);					\
-  return v;							\
+#define make_atomic_add(S)                                      \
+STATIC_INLINE int ## S my_atomic_add ## S(                      \
+                        Uv_ ## S U_a, U_ ## S U_v)              \
+{                                                               \
+  make_atomic_add_body(S);                                      \
+  return v;                                                     \
 }
 
-#define make_atomic_cas(S)					\
-STATIC_INLINE int my_atomic_cas ## S(int ## S volatile *a,	\
-                            int ## S *cmp, int ## S set)	\
-{								\
-  int8 ret;							\
-  make_atomic_cas_body(S);					\
-  return ret;							\
+#define make_atomic_fas(S)                                      \
+STATIC_INLINE int ## S my_atomic_fas ## S(                      \
+                         Uv_ ## S U_a, U_ ## S U_v)             \
+{                                                               \
+  make_atomic_fas_body(S);                                      \
+  return v;                                                     \
 }
 
-#define make_atomic_load(S)					\
-STATIC_INLINE int ## S my_atomic_load ## S(int ## S volatile *a) \
-{								\
-  int ## S ret;						\
-  make_atomic_load_body(S);					\
-  return ret;							\
+#define make_atomic_load(S)                                     \
+STATIC_INLINE int ## S my_atomic_load ## S(Uv_ ## S U_a)        \
+{                                                               \
+  int ## S ret;                                                 \
+  make_atomic_load_body(S);                                     \
+  return ret;                                                   \
 }
 
-#define make_atomic_store(S)					\
-STATIC_INLINE void my_atomic_store ## S(			\
-                     int ## S volatile *a, int ## S v)	\
-{								\
-  make_atomic_store_body(S);					\
+#define make_atomic_store(S)                                    \
+STATIC_INLINE void my_atomic_store ## S(                        \
+                     Uv_ ## S U_a, U_ ## S U_v)                 \
+{                                                               \
+  make_atomic_store_body(S);                                    \
 }
 
 #else /* no inline functions */
 
-#define make_atomic_add(S)					\
-extern int ## S my_atomic_add ## S(int ## S volatile *a, int ## S v);
+#define make_atomic_add(S)                                      \
+extern int ## S my_atomic_add ## S(Uv_ ## S U_a, U_ ## S U_v);
 
-#define make_atomic_swap(S)					\
-extern int ## S my_atomic_swap ## S(int ## S volatile *a, int ## S v);
+#define make_atomic_fas(S)                                      \
+extern int ## S my_atomic_fas ## S(Uv_ ## S U_a, U_ ## S U_v);
 
-#define make_atomic_cas(S)					\
-extern int my_atomic_cas ## S(int ## S volatile *a, int ## S *cmp, int ## S set);
+#define make_atomic_cas(S)                                      \
+extern int my_atomic_cas ## S(Uv_ ## S U_a, Uv_ ## S U_cmp, U_ ## S U_set);
 
-#define make_atomic_load(S)					\
-extern int ## S my_atomic_load ## S(int ## S volatile *a);
+#define make_atomic_load(S)                                     \
+extern int ## S my_atomic_load ## S(Uv_ ## S U_a);
 
-#define make_atomic_store(S)					\
-extern void my_atomic_store ## S(int ## S volatile *a, int ## S v);
+#define make_atomic_store(S)                                    \
+extern void my_atomic_store ## S(Uv_ ## S U_a, U_ ## S U_v);
 
 #endif /* HAVE_INLINE */
 
-make_atomic_cas( 8)
+#ifdef MY_ATOMIC_HAS_8_16
+make_atomic_cas(8)
 make_atomic_cas(16)
+#endif
 make_atomic_cas(32)
 make_atomic_cas(ptr)
 
-make_atomic_add( 8)
+#ifdef MY_ATOMIC_HAS_8_16
+make_atomic_add(8)
 make_atomic_add(16)
+#endif
 make_atomic_add(32)
 
-make_atomic_load( 8)
+#ifdef MY_ATOMIC_HAS_8_16
+make_atomic_load(8)
 make_atomic_load(16)
+#endif
 make_atomic_load(32)
 make_atomic_load(ptr)
 
-make_atomic_store( 8)
+#ifdef MY_ATOMIC_HAS_8_16
+make_atomic_fas(8)
+make_atomic_fas(16)
+#endif
+make_atomic_fas(32)
+make_atomic_fas(ptr)
+
+#ifdef MY_ATOMIC_HAS_8_16
+make_atomic_store(8)
 make_atomic_store(16)
+#endif
 make_atomic_store(32)
 make_atomic_store(ptr)
 
-make_atomic_swap( 8)
-make_atomic_swap(16)
-make_atomic_swap(32)
-make_atomic_swap(ptr)
+#ifdef _atomic_h_cleanup_
+#include _atomic_h_cleanup_
+#undef _atomic_h_cleanup_
+#endif
 
+#undef U_8
+#undef U_16
+#undef U_32
+#undef U_ptr
+#undef Uv_8
+#undef Uv_16
+#undef Uv_32
+#undef Uv_ptr
+#undef a
+#undef cmp
+#undef v
+#undef set
+#undef U_a
+#undef U_cmp
+#undef U_v
+#undef U_set
 #undef make_atomic_add
 #undef make_atomic_cas
 #undef make_atomic_load
 #undef make_atomic_store
-#undef make_atomic_swap
+#undef make_atomic_fas
 #undef make_atomic_add_body
 #undef make_atomic_cas_body
 #undef make_atomic_load_body
 #undef make_atomic_store_body
-#undef make_atomic_swap_body
+#undef make_atomic_fas_body
 #undef intptr
 
-#endif /* MY_ATOMICS_MADE */
-
-#ifdef _atomic_h_cleanup_
-#include _atomic_h_cleanup_
-#undef _atomic_h_cleanup_
+/*
+  the macro below defines (as an expression) the code that
+  will be run in spin-loops. Intel manuals recummend to have PAUSE there.
+  It is expected to be defined in include/atomic/ *.h files
+*/
+#ifndef LF_BACKOFF
+#define LF_BACKOFF (1)
 #endif
 
 #define MY_ATOMIC_OK       0
diff --git a/include/my_pthread.h b/include/my_pthread.h
index eff6d654f2e..3b95e2adb73 100644
--- a/include/my_pthread.h
+++ b/include/my_pthread.h
@@ -148,6 +148,7 @@ int pthread_join(pthread_t thread, void **value_ptr);
 #define pthread_detach_this_thread()
 #define pthread_condattr_init(A)
 #define pthread_condattr_destroy(A)
+#define pthread_yield() SwitchToThread()
 
 /* per the platform's documentation */
 #define pthread_yield() Sleep(0)
@@ -354,6 +355,17 @@ void my_pthread_attr_getstacksize(pthread_attr_t *attrib, size_t *size);
 int my_pthread_mutex_trylock(pthread_mutex_t *mutex);
 #endif
 
+#if !defined(HAVE_PTHREAD_YIELD_ONE_ARG) && !defined(HAVE_PTHREAD_YIELD_ZERO_ARG)
+/* no pthread_yield() available */
+#ifdef HAVE_SCHED_YIELD
+#define pthread_yield() sched_yield()
+#elif defined(HAVE_PTHREAD_YIELD_NP) /* can be Mac OS X */
+#define pthread_yield() pthread_yield_np()
+#elif defined(HAVE_THR_YIELD)
+#define pthread_yield() thr_yield()
+#endif
+#endif
+
 /*
   The defines set_timespec and set_timespec_nsec should be used
   for calculating an absolute time at which
@@ -632,6 +644,7 @@ struct st_my_thread_var
   my_bool init;
   struct st_my_thread_var *next,**prev;
   void *opt_info;
+  void  *stack_ends_here;
 #ifndef DBUG_OFF
   void *dbug;
   char name[THREAD_NAME_SIZE+1];