* intl/dcigettext.c (_nl_find_msg): Free encoding if __gconv_open

	failed.
	* intl/finddomain.c (_nl_find_domain): Free normalized_codeset
	on failure.

	* elf/dl-load.c (decompose_rpath): Free copy if result couldn't be
	allocated.

2007-08-03  Jakub Jelinek  <jakub@redhat.com>

19 files changed, 581 insertions, 413 deletions

diff --git a/nptl/sysdeps/unix/sysv/linux/sh/libc-lowlevellock.S b/nptl/sysdeps/unix/sysv/linux/sh/libc-lowlevellock.S
index be85ab7414..feb82110cf 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/libc-lowlevellock.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/libc-lowlevellock.S
@@ -16,32 +16,4 @@
    Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
    02111-1307 USA.  */
 
-#include <kernel-features.h>
-
-/* All locks in libc are private.  Use the kernel feature if possible.  */
-#define FUTEX_PRIVATE_FLAG	128
-#ifdef __ASSUME_PRIVATE_FUTEX
-# define FUTEX_WAIT		(0 | FUTEX_PRIVATE_FLAG)
-# define FUTEX_WAKE		(1 | FUTEX_PRIVATE_FLAG)
-#else
-# define LOAD_FUTEX_WAIT(reg,tmp) \
-	stc	gbr, tmp	; \
-	mov.w	99f, reg	; \
-	add	reg, tmp 	; \
-	bra	98f		; \
-	 mov.l	@tmp, reg	; \
-99:	.word	PRIVATE_FUTEX - TLS_PRE_TCB_SIZE ; \
-98:
-	
-# define LOAD_FUTEX_WAKE(reg,tmp) \
-	stc	gbr, tmp	; \
-	mov.w	99f, reg	; \
-	add	reg, tmp 	; \
-	mov.l	@tmp, reg	; \
-	bra	98f		; \
-	 mov	#FUTEX_WAKE, tmp ; \
-99:	.word	PRIVATE_FUTEX - TLS_PRE_TCB_SIZE ; \
-98:	or	tmp, reg
-#endif
-
 #include "lowlevellock.S"
diff --git a/nptl/sysdeps/unix/sysv/linux/sh/lowlevellock.S b/nptl/sysdeps/unix/sysv/linux/sh/lowlevellock.S
index b10903bcd7..e929e28499 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/lowlevellock.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/lowlevellock.S
@@ -18,45 +18,112 @@
 
 #include <sysdep.h>
 #include <pthread-errnos.h>
+#include <kernel-features.h>
+#include <lowlevellock.h>
 #include "lowlevel-atomic.h"
 
 	.text
 
-#define SYS_gettimeofday	__NR_gettimeofday
-#define SYS_futex		240
-#ifndef FUTEX_WAIT
-# define FUTEX_WAIT		0
-# define FUTEX_WAKE		1
-#endif
-
-#ifndef LOAD_FUTEX_WAIT
+#ifdef __ASSUME_PRIVATE_FUTEX
+# define LOAD_PRIVATE_FUTEX_WAIT(reg,tmp,tmp2) \
+	mov	#(FUTEX_WAIT | FUTEX_PRIVATE_FLAG), reg; \
+	extu.b	reg, reg
+# define LOAD_PRIVATE_FUTEX_WAKE(reg,tmp,tmp2) \
+	mov	#(FUTEX_WAKE | FUTEX_PRIVATE_FLAG), reg; \
+	extu.b	reg, reg
+# define LOAD_FUTEX_WAIT(reg,tmp,tmp2) \
+	mov	#(FUTEX_WAIT | FUTEX_PRIVATE_FLAG), tmp; \
+	extu.b	tmp, tmp; \
+	xor	tmp, reg
+# define LOAD_FUTEX_WAKE(reg,tmp,tmp2) \
+	mov	#(FUTEX_WAKE | FUTEX_PRIVATE_FLAG), tmp; \
+	extu.b	tmp, tmp; \
+	xor	tmp, reg
+#else
 # if FUTEX_WAIT == 0
-#  define LOAD_FUTEX_WAIT(reg,tmp) \
-	xor	reg, reg
+#  define LOAD_PRIVATE_FUTEX_WAIT(reg,tmp,tmp2) \
+	stc	gbr, tmp	; \
+	mov.w	99f, reg	; \
+	add	reg, tmp 	; \
+	bra	98f		; \
+	 mov.l	@tmp, reg	; \
+99:	.word	PRIVATE_FUTEX - TLS_PRE_TCB_SIZE ; \
+98:
 # else
-#  define LOAD_FUTEX_WAIT(reg,tmp) \
-	mov	#FUTEX_WAIT, reg; \
-	extu.b	reg, reg
+#  define LOAD_PRIVATE_FUTEX_WAIT(reg,tmp,tmp2) \
+	stc	gbr, tmp	; \
+	mov.w	99f, reg	; \
+	add	reg, tmp 	; \
+	mov.l	@tmp, reg	; \
+	bra	98f		; \
+	 mov	#FUTEX_WAIT, tmp ; \
+99:	.word	PRIVATE_FUTEX - TLS_PRE_TCB_SIZE ; \
+98:	or	tmp, reg
+# endif
+# define LOAD_PRIVATE_FUTEX_WAKE(reg,tmp,tmp2) \
+	stc	gbr, tmp	; \
+	mov.w	99f, reg	; \
+	add	reg, tmp 	; \
+	mov.l	@tmp, reg	; \
+	bra	98f		; \
+	 mov	#FUTEX_WAKE, tmp ; \
+99:	.word	PRIVATE_FUTEX - TLS_PRE_TCB_SIZE ; \
+98:	or	tmp, reg
+# if FUTEX_WAIT == 0
+#  define LOAD_FUTEX_WAIT(reg,tmp,tmp2) \
+	stc	gbr, tmp	; \
+	mov.w	99f, tmp2	; \
+	add	tmp2, tmp 	; \
+	mov.l	@tmp, tmp2	; \
+	bra	98f		; \
+	 mov	#FUTEX_PRIVATE_FLAG, tmp
+99:	.word	PRIVATE_FUTEX - TLS_PRE_TCB_SIZE ; \
+98:	extu.b	tmp, tmp	; \
+	xor	tmp, reg	; \
+	and	tmp2, reg
+# else
+#  define LOAD_FUTEX_WAIT(reg,tmp,tmp2) \
+	stc	gbr, tmp	; \
+	mov.w	99f, tmp2	; \
+	add	tmp2, tmp 	; \
+	mov.l	@tmp, tmp2	; \
+	bra	98f		; \
+	 mov	#FUTEX_PRIVATE_FLAG, tmp
+99:	.word	PRIVATE_FUTEX - TLS_PRE_TCB_SIZE ; \
+98:	extu.b	tmp, tmp	; \
+	xor	tmp, reg	; \
+	and	tmp2, reg	; \
+	mov	#FUTEX_WAIT, tmp ; \
+	or	tmp, reg
 # endif
 # define LOAD_FUTEX_WAKE(reg,tmp) \
-	mov	#FUTEX_WAKE, reg; \
-	extu.b	reg, reg
+	stc	gbr, tmp	; \
+	mov.w	99f, tmp2	; \
+	add	tmp2, tmp 	; \
+	mov.l	@tmp, tmp2	; \
+	bra	98f		; \
+	 mov	#FUTEX_PRIVATE_FLAG, tmp
+99:	.word	PRIVATE_FUTEX - TLS_PRE_TCB_SIZE ; \
+98:	extu.b	tmp, tmp	; \
+	xor	tmp, reg	; \
+	and	tmp2, reg	; \
+	mov	#FUTEX_WAKE, tmp ; \
+	or	tmp, reg
 #endif
 
-
-	.globl	__lll_mutex_lock_wait
-	.type	__lll_mutex_lock_wait,@function
-	.hidden	__lll_mutex_lock_wait
+	.globl	__lll_lock_wait_private
+	.type	__lll_lock_wait_private,@function
+	.hidden	__lll_lock_wait_private
 	.align	5
 	cfi_startproc
-__lll_mutex_lock_wait:
+__lll_lock_wait_private:
 	mov.l	r8, @-r15
 	cfi_adjust_cfa_offset(4)
 	cfi_rel_offset (r8, 0)
 	mov	r4, r6
 	mov	r5, r8
 	mov	#0, r7		/* No timeout.  */
-	LOAD_FUTEX_WAIT (r5, r0)
+	LOAD_PRIVATE_FUTEX_WAIT (r5, r0, r1)
 
 	mov	#2, r4
 	cmp/eq	r4, r6
@@ -79,22 +146,67 @@ __lll_mutex_lock_wait:
 	ret
 	 mov	r2, r0
 	cfi_endproc
-	.size	__lll_mutex_lock_wait,.-__lll_mutex_lock_wait
-
+	.size	__lll_lock_wait_private,.-__lll_lock_wait_private
 
 #ifdef NOT_IN_libc
-	.globl	__lll_mutex_timedlock_wait
-	.type	__lll_mutex_timedlock_wait,@function
-	.hidden	__lll_mutex_timedlock_wait
+	.globl	__lll_lock_wait
+	.type	__lll_lock_wait,@function
+	.hidden	__lll_lock_wait
 	.align	5
 	cfi_startproc
-__lll_mutex_timedlock_wait:
+__lll_lock_wait:
+	mov.l	r9, @-r15
+	cfi_adjust_cfa_offset(4)
+	cfi_rel_offset (r9, 0)
+	mov.l	r8, @-r15
+	cfi_adjust_cfa_offset(4)
+	cfi_rel_offset (r8, 0)
+	mov	r6, r9
+	mov	r4, r6
+	mov	r5, r8
+	mov	#0, r7		/* No timeout.  */
+	mov	r9, r5
+	LOAD_FUTEX_WAIT (r5, r0, r1)
+
+	mov	#2, r4
+	cmp/eq	r4, r6
+	bf	2f
+
+1:
+	mov	r8, r4
+	mov	#SYS_futex, r3
+	extu.b	r3, r3
+	trapa	#0x14
+	SYSCALL_INST_PAD
+
+2:
+	mov	#2, r6
+	XCHG (r6, @r8, r2)
+	tst	r2, r2
+	bf	1b
+
+	mov.l	@r15+, r8
+	mov.l	@r15+, r9
+	ret
+	 mov	r2, r0
+	cfi_endproc
+	.size	__lll_lock_wait,.-__lll_lock_wait
+
+	.globl	__lll_timedlock_wait
+	.type	__lll_timedlock_wait,@function
+	.hidden	__lll_timedlock_wait
+	.align	5
+	cfi_startproc
+__lll_timedlock_wait:
 	/* Check for a valid timeout value.  */
 	mov.l	@(4,r6), r1
 	mov.l	.L1g, r0
 	cmp/hs	r0, r1
 	bt	3f
 
+	mov.l	r11, @-r15
+	cfi_adjust_cfa_offset(4)
+	cfi_rel_offset (r11, 0)
 	mov.l	r10, @-r15
 	cfi_adjust_cfa_offset(4)
 	cfi_rel_offset (r10, 0)
@@ -104,6 +216,7 @@ __lll_mutex_timedlock_wait:
 	mov.l	r8, @-r15
 	cfi_adjust_cfa_offset(4)
 	cfi_rel_offset (r8, 0)
+	mov	r7, r11
 	mov	r4, r10
 	mov	r6, r9
 	mov	r5, r8
@@ -116,7 +229,7 @@ __lll_mutex_timedlock_wait:
 	/* Get current time.  */
 	mov	r15, r4
 	mov	#0, r5
-	mov	#SYS_gettimeofday, r3
+	mov	#__NR_gettimeofday, r3
 	trapa	#0x12
 	SYSCALL_INST_PAD
 
@@ -149,7 +262,8 @@ __lll_mutex_timedlock_wait:
 	bt	8f
 
 	mov	r8, r4
-	LOAD_FUTEX_WAIT (r5, r0)
+	mov	r11, r5
+	LOAD_FUTEX_WAIT (r5, r0, r1)
 	mov	r10, r6
 	mov	r15, r7
 	mov	#SYS_futex, r3
@@ -169,8 +283,9 @@ __lll_mutex_timedlock_wait:
 	add	#8, r15
 	mov.l	@r15+, r8
 	mov.l	@r15+, r9
+	mov.l	@r15+, r10
 	rts
-	 mov.l	@r15+, r10
+	 mov.l	@r15+, r11
 7:
 	/* Check whether the time expired.  */
 	mov	#-ETIMEDOUT, r1
@@ -198,17 +313,16 @@ __lll_mutex_timedlock_wait:
 .L1g:
 	.long	1000000000
 
-	.size	__lll_mutex_timedlock_wait,.-__lll_mutex_timedlock_wait
+	.size	__lll_timedlock_wait,.-__lll_timedlock_wait
 #endif
 
-
-	.globl	__lll_mutex_unlock_wake
-	.type	__lll_mutex_unlock_wake,@function
-	.hidden	__lll_mutex_unlock_wake
+	.globl	__lll_unlock_wake_private
+	.type	__lll_unlock_wake_private,@function
+	.hidden	__lll_unlock_wake_private
 	.align	5
 	cfi_startproc
-__lll_mutex_unlock_wake:
-	LOAD_FUTEX_WAKE (r5, r0)
+__lll_unlock_wake_private:
+	LOAD_PRIVATE_FUTEX_WAKE (r5, r0, r1)
 	mov	#1, r6		/* Wake one thread.  */
 	mov	#0, r7
 	mov.l	r7, @r4		/* Stores 0.  */
@@ -219,10 +333,28 @@ __lll_mutex_unlock_wake:
 	rts
 	 nop
 	cfi_endproc
-	.size	__lll_mutex_unlock_wake,.-__lll_mutex_unlock_wake
-
+	.size	__lll_unlock_wake_private,.-__lll_unlock_wake_private
 
 #ifdef NOT_IN_libc
+	.globl	__lll_unlock_wake
+	.type	__lll_unlock_wake,@function
+	.hidden	__lll_unlock_wake
+	.align	5
+	cfi_startproc
+__lll_unlock_wake:
+	LOAD_FUTEX_WAKE (r5, r0, r1)
+	mov	#1, r6		/* Wake one thread.  */
+	mov	#0, r7
+	mov.l	r7, @r4		/* Stores 0.  */
+	mov	#SYS_futex, r3
+	extu.b	r3, r3
+	trapa	#0x14
+	SYSCALL_INST_PAD
+	rts
+	 nop
+	cfi_endproc
+	.size	__lll_unlock_wake,.-__lll_unlock_wake
+
 	.globl	__lll_timedwait_tid
 	.type	__lll_timedwait_tid,@function
 	.hidden	__lll_timedwait_tid
@@ -246,7 +378,7 @@ __lll_timedwait_tid:
 	/* Get current time.  */
 	mov	r15, r4
 	mov	#0, r5
-	mov	#SYS_gettimeofday, r3
+	mov	#__NR_gettimeofday, r3
 	trapa	#0x12
 	SYSCALL_INST_PAD
 
diff --git a/nptl/sysdeps/unix/sysv/linux/sh/lowlevellock.h b/nptl/sysdeps/unix/sysv/linux/sh/lowlevellock.h
index be47bd752e..88e94b9905 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/lowlevellock.h
+++ b/nptl/sysdeps/unix/sysv/linux/sh/lowlevellock.h
@@ -19,19 +19,24 @@
 #ifndef _LOWLEVELLOCK_H
 #define _LOWLEVELLOCK_H	1
 
+#ifndef __ASSEMBLER__
 #include <time.h>
 #include <sys/param.h>
 #include <bits/pthreadtypes.h>
 #include <kernel-features.h>
+#endif
 
 #define SYS_futex		240
 #define FUTEX_WAIT		0
 #define FUTEX_WAKE		1
+#define FUTEX_CMP_REQUEUE	4
+#define FUTEX_WAKE_OP		5
 #define FUTEX_LOCK_PI		6
 #define FUTEX_UNLOCK_PI		7
 #define FUTEX_TRYLOCK_PI	8
 #define FUTEX_PRIVATE_FLAG	128
 
+#define FUTEX_OP_CLEAR_WAKE_IF_GT_ONE	((4 << 24) | 1)
 
 /* Values for 'private' parameter of locking macros.  Yes, the
    definition seems to be backwards.  But it is not.  The bit will be
@@ -64,20 +69,30 @@
 # endif	      
 #endif
 
+#ifndef __ASSEMBLER__
 
 /* Initializer for compatibility lock.  */
-#define LLL_MUTEX_LOCK_INITIALIZER		(0)
-#define LLL_MUTEX_LOCK_INITIALIZER_LOCKED	(1)
-#define LLL_MUTEX_LOCK_INITIALIZER_WAITERS	(2)
-
-extern int __lll_mutex_lock_wait (int val, int *__futex) attribute_hidden;
-extern int __lll_mutex_timedlock_wait (int val, int *__futex,
-				       const struct timespec *abstime)
-     attribute_hidden;
-extern int __lll_mutex_unlock_wake (int *__futex) attribute_hidden;
-
-
-#define lll_mutex_trylock(futex) \
+#define LLL_LOCK_INITIALIZER		(0)
+#define LLL_LOCK_INITIALIZER_LOCKED	(1)
+#define LLL_LOCK_INITIALIZER_WAITERS	(2)
+
+extern int __lll_lock_wait_private (int val, int *__futex)
+  attribute_hidden;
+extern int __lll_lock_wait (int val, int *__futex, int private)
+  attribute_hidden;
+extern int __lll_timedlock_wait (int val, int *__futex,
+				 const struct timespec *abstime, int private)
+  attribute_hidden;
+extern int __lll_robust_lock_wait (int val, int *__futex, int private)
+  attribute_hidden;
+extern int __lll_robust_timedlock_wait (int val, int *__futex,
+					const struct timespec *abstime,
+					int private)
+  attribute_hidden;
+extern int __lll_unlock_wake_private (int *__futex) attribute_hidden;
+extern int __lll_unlock_wake (int *__futex, int private) attribute_hidden;
+
+#define lll_trylock(futex) \
   ({ unsigned char __result; \
      __asm __volatile ("\
 	.align 2\n\
@@ -94,12 +109,12 @@ extern int __lll_mutex_unlock_wake (int *__futex) attribute_hidden;
 	negc %0,%0"\
 	: "=r" (__result) \
 	: "r" (&(futex)), \
-	  "r" (LLL_MUTEX_LOCK_INITIALIZER_LOCKED), \
-	  "r" (LLL_MUTEX_LOCK_INITIALIZER) \
+	  "r" (LLL_LOCK_INITIALIZER_LOCKED), \
+	  "r" (LLL_LOCK_INITIALIZER) \
 	: "r0", "r1", "r2", "t", "memory"); \
      __result; })
 
-#define lll_robust_mutex_trylock(futex, id)	\
+#define lll_robust_trylock(futex, id)	\
   ({ unsigned char __result; \
      __asm __volatile ("\
 	.align 2\n\
@@ -117,11 +132,11 @@ extern int __lll_mutex_unlock_wake (int *__futex) attribute_hidden;
 	: "=r" (__result) \
 	: "r" (&(futex)), \
 	  "r" (id), \
-	  "r" (LLL_MUTEX_LOCK_INITIALIZER) \
+	  "r" (LLL_LOCK_INITIALIZER) \
 	: "r0", "r1", "r2", "t", "memory"); \
      __result; })
 
-#define lll_mutex_cond_trylock(futex) \
+#define lll_cond_trylock(futex) \
   ({ unsigned char __result; \
      __asm __volatile ("\
 	.align 2\n\
@@ -138,13 +153,13 @@ extern int __lll_mutex_unlock_wake (int *__futex) attribute_hidden;
 	negc %0,%0"\
 	: "=r" (__result) \
 	: "r" (&(futex)), \
-	  "r" (LLL_MUTEX_LOCK_INITIALIZER_WAITERS), \
-	  "r" (LLL_MUTEX_LOCK_INITIALIZER) \
+	  "r" (LLL_LOCK_INITIALIZER_WAITERS), \
+	  "r" (LLL_LOCK_INITIALIZER) \
 	: "r0", "r1", "r2", "t", "memory"); \
      __result; })
 
-#define lll_mutex_lock(futex) \
-  (void) ({ int __result, val, *__futex = &(futex); \
+#define lll_lock(futex, private) \
+  (void) ({ int __result, *__futex = &(futex); \
 	    __asm __volatile ("\
 		.align 2\n\
 		mova 1f,r0\n\
@@ -159,10 +174,17 @@ extern int __lll_mutex_unlock_wake (int *__futex) attribute_hidden;
 		: "=&r" (__result) : "r" (1), "r" (__futex) \
 		: "r0", "r1", "t", "memory"); \
 	    if (__result) \
-	      __lll_mutex_lock_wait (__result, __futex); })
-
-#define lll_robust_mutex_lock(futex, id)		    \
-  ({ int __result, val, *__futex = &(futex); \
+	      { \
+		if (__builtin_constant_p (private) \
+		    && (private) == LLL_PRIVATE) \
+		  __lll_lock_wait_private (__result, __futex); \
+	        else \
+		  __lll_lock_wait (__result, __futex, (private));	\
+	      } \
+    })
+
+#define lll_robust_lock(futex, id, private) \
+  ({ int __result, *__futex = &(futex); \
      __asm __volatile ("\
 	.align 2\n\
 	mova 1f,r0\n\
@@ -177,13 +199,13 @@ extern int __lll_mutex_unlock_wake (int *__futex) attribute_hidden;
 	: "=&r" (__result) : "r" (id), "r" (__futex) \
 	: "r0", "r1", "t", "memory"); \
      if (__result) \
-       __result = __lll_robust_mutex_lock_wait (__result, __futex); \
+       __result = __lll_robust_lock_wait (__result, __futex, private); \
      __result; })
 
 /* Special version of lll_mutex_lock which causes the unlock function to
    always wakeup waiters.  */
-#define lll_mutex_cond_lock(futex) \
-  (void) ({ int __result, val, *__futex = &(futex); \
+#define lll_cond_lock(futex, private) \
+  (void) ({ int __result, *__futex = &(futex); \
 	    __asm __volatile ("\
 		.align 2\n\
 		mova 1f,r0\n\
@@ -198,10 +220,10 @@ extern int __lll_mutex_unlock_wake (int *__futex) attribute_hidden;
 		: "=&r" (__result) : "r" (2), "r" (__futex) \
 		: "r0", "r1", "t", "memory"); \
 	    if (__result) \
-	      __lll_mutex_lock_wait (__result, __futex); })
+	      __lll_lock_wait (__result, __futex, private); })
 
-#define lll_robust_mutex_cond_lock(futex, id)	    \
-  ({ int __result, val, *__futex = &(futex); \
+#define lll_robust_cond_lock(futex, id, private) \
+  ({ int __result, *__futex = &(futex); \
      __asm __volatile ("\
 	.align 2\n\
 	mova 1f,r0\n\
@@ -216,11 +238,11 @@ extern int __lll_mutex_unlock_wake (int *__futex) attribute_hidden;
 	: "=&r" (__result) : "r" (id | FUTEX_WAITERS), "r" (__futex) \
 	: "r0", "r1", "t", "memory"); \
       if (__result) \
-	__result = __lll_robust_mutex_lock_wait (__result, __futex); \
+	__result = __lll_robust_lock_wait (__result, __futex, private); \
       __result; })
 
-#define lll_mutex_timedlock(futex, timeout) \
-  ({ int __result, val, *__futex = &(futex); \
+#define lll_timedlock(futex, timeout, private) \
+  ({ int __result, *__futex = &(futex); \
      __asm __volatile ("\
 	.align 2\n\
 	mova 1f,r0\n\
@@ -235,11 +257,11 @@ extern int __lll_mutex_unlock_wake (int *__futex) attribute_hidden;
 	: "=&r" (__result) : "r" (1), "r" (__futex) \
 	: "r0", "r1", "t", "memory"); \
     if (__result) \
-      __result = __lll_mutex_timedlock_wait (__result, __futex, timeout); \
+      __result = __lll_timedlock_wait (__result, __futex, timeout, private); \
     __result; })
 
-#define lll_robust_mutex_timedlock(futex, timeout, id)	\
-  ({ int __result, val, *__futex = &(futex); \
+#define lll_robust_timedlock(futex, timeout, id, private) \
+  ({ int __result, *__futex = &(futex); \
      __asm __volatile ("\
 	.align 2\n\
 	mova 1f,r0\n\
@@ -254,11 +276,11 @@ extern int __lll_mutex_unlock_wake (int *__futex) attribute_hidden;
 	: "=&r" (__result) : "r" (id), "r" (__futex) \
 	: "r0", "r1", "t", "memory"); \
     if (__result) \
-      __result = __lll_robust_mutex_timedlock_wait (__result, __futex, \
-						    timeout);	       \
+      __result = __lll_robust_timedlock_wait (__result, __futex, \
+					      timeout, private); \
     __result; })
 
-#define lll_mutex_unlock(futex) \
+#define lll_unlock(futex, private) \
   (void) ({ int __result, *__futex = &(futex); \
 	    __asm __volatile ("\
 		.align 2\n\
@@ -272,9 +294,16 @@ extern int __lll_mutex_unlock_wake (int *__futex) attribute_hidden;
 		: "=&r" (__result) : "r" (__futex) \
 		: "r0", "r1", "memory"); \
 	    if (__result) \
-	      __lll_mutex_unlock_wake (__futex); })
-
-#define lll_robust_mutex_unlock(futex) \
+	      { \
+		if (__builtin_constant_p (private) \
+		    && (private) == LLL_PRIVATE) \
+		  __lll_unlock_wake_private (__futex); \
+	        else \
+		  __lll_unlock_wake (__futex, (private)); \
+	      } \
+    })
+
+#define lll_robust_unlock(futex, private) \
   (void) ({ int __result, *__futex = &(futex); \
 	    __asm __volatile ("\
 		.align 2\n\
@@ -288,9 +317,9 @@ extern int __lll_mutex_unlock_wake (int *__futex) attribute_hidden;
 		: "=&r" (__result) : "r" (__futex), "r" (FUTEX_WAITERS) \
 		: "r0", "r1", "memory");	\
 	    if (__result) \
-	      __lll_mutex_unlock_wake (__futex); })
+	      __lll_unlock_wake (__futex, private); })
 
-#define lll_robust_mutex_dead(futex) \
+#define lll_robust_dead(futex, private)		       \
   (void) ({ int __ignore, *__futex = &(futex); \
 	    __asm __volatile ("\
 		.align 2\n\
@@ -303,22 +332,7 @@ extern int __lll_mutex_unlock_wake (int *__futex) attribute_hidden;
 	     1: mov r1,r15"\
 		: "=&r" (__ignore) : "r" (__futex), "r" (FUTEX_OWNER_DIED) \
 		: "r0", "r1", "memory");	\
-	    lll_futex_wake (__futex, 1, LLL_SHARED); })
-
-#define lll_mutex_islocked(futex) \
-  (futex != 0)
-
-
-/* We have a separate internal lock implementation which is not tied
-   to binary compatibility.  */
-
-/* Type for lock object.  */
-typedef int lll_lock_t;
-
-/* Initializers for lock.  */
-#define LLL_LOCK_INITIALIZER		(0)
-#define LLL_LOCK_INITIALIZER_LOCKED	(1)
-
+	    lll_futex_wake (__futex, 1, private); })
 
 # ifdef NEED_SYSCALL_INST_PAD
 #  define SYSCALL_WITH_INST_PAD "\
@@ -367,25 +381,14 @@ typedef int lll_lock_t;
   } while (0)
 
 
-/* The states of a lock are:
-    0  -  untaken
-    1  -  taken by one user
-    2  -  taken by more users */
-
-#define lll_trylock(futex) lll_mutex_trylock (futex)
-#define lll_lock(futex) lll_mutex_lock (futex)
-#define lll_unlock(futex) lll_mutex_unlock (futex)
-
 #define lll_islocked(futex) \
   (futex != LLL_LOCK_INITIALIZER)
 
-
 /* The kernel notifies a process with uses CLONE_CLEARTID via futex
    wakeup when the clone terminates.  The memory location contains the
    thread ID while the clone is running and is reset to zero
    afterwards.  */
 
-extern int __lll_wait_tid (int *tid) attribute_hidden;
 #define lll_wait_tid(tid) \
   do {									      \
     __typeof (tid) *__tid = &(tid);					      \
@@ -407,24 +410,6 @@ extern int __lll_timedwait_tid (int *tid, const struct timespec *abstime)
       }									      \
     __result; })
 
-
-/* Conditional variable handling.  */
-
-extern void __lll_cond_wait (pthread_cond_t *cond) attribute_hidden;
-extern int __lll_cond_timedwait (pthread_cond_t *cond,
-				 const struct timespec *abstime)
-     attribute_hidden;
-extern void __lll_cond_wake (pthread_cond_t *cond) attribute_hidden;
-extern void __lll_cond_broadcast (pthread_cond_t *cond) attribute_hidden;
-
-
-#define lll_cond_wait(cond) \
-  __lll_cond_wait (cond)
-#define lll_cond_timedwait(cond, abstime) \
-  __lll_cond_timedwait (cond, abstime)
-#define lll_cond_wake(cond) \
-  __lll_cond_wake (cond)
-#define lll_cond_broadcast(cond) \
-  __lll_cond_broadcast (cond)
+#endif  /* !__ASSEMBLER__ */
 
 #endif  /* lowlevellock.h */
diff --git a/nptl/sysdeps/unix/sysv/linux/sh/lowlevelrobustlock.S b/nptl/sysdeps/unix/sysv/linux/sh/lowlevelrobustlock.S
index c57d3cff18..0ebfbfe8c9 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/lowlevelrobustlock.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/lowlevelrobustlock.S
@@ -1,4 +1,5 @@
-/* Copyright (C) 2003, 2004, 2005, 2006 Free Software Foundation, Inc.
+/* Copyright (C) 2003, 2004, 2005, 2006, 2007
+   Free Software Foundation, Inc.
    This file is part of the GNU C Library.
 
    The GNU C Library is free software; you can redistribute it and/or
@@ -18,31 +19,64 @@
 
 #include <sysdep.h>
 #include <pthread-errnos.h>
+#include <lowlevellock.h>
 #include <lowlevelrobustlock.h>
+#include <kernel-features.h>
 #include "lowlevel-atomic.h"
 
 	.text
 
-#define SYS_gettimeofday	__NR_gettimeofday
-#define SYS_futex		240
-#define FUTEX_WAIT		0
-#define FUTEX_WAKE		1
 #define FUTEX_WAITERS		0x80000000
 #define FUTEX_OWNER_DIED	0x40000000
 
+#ifdef __ASSUME_PRIVATE_FUTEX
+# define LOAD_FUTEX_WAIT(reg,tmp,tmp2) \
+	mov	#(FUTEX_WAIT | FUTEX_PRIVATE_FLAG), tmp; \
+	extu.b	tmp, tmp; \
+	xor	tmp, reg
+#else
+# if FUTEX_WAIT == 0
+#  define LOAD_FUTEX_WAIT(reg,tmp,tmp2) \
+	stc	gbr, tmp	; \
+	mov.w	99f, tmp2	; \
+	add	tmp2, tmp 	; \
+	mov.l	@tmp, tmp2	; \
+	bra	98f		; \
+	 mov	#FUTEX_PRIVATE_FLAG, tmp
+99:	.word	PRIVATE_FUTEX - TLS_PRE_TCB_SIZE ; \
+98:	extu.b	tmp, tmp	; \
+	xor	tmp, reg	; \
+	and	tmp2, reg
+# else
+#  define LOAD_FUTEX_WAIT(reg,tmp,tmp2) \
+	stc	gbr, tmp	; \
+	mov.w	99f, tmp2	; \
+	add	tmp2, tmp 	; \
+	mov.l	@tmp, tmp2	; \
+	bra	98f		; \
+	 mov	#FUTEX_PRIVATE_FLAG, tmp
+99:	.word	PRIVATE_FUTEX - TLS_PRE_TCB_SIZE ; \
+98:	extu.b	tmp, tmp	; \
+	xor	tmp, reg	; \
+	and	tmp2, reg	; \
+	mov	#FUTEX_WAIT, tmp ; \
+	or	tmp, reg
+# endif
+#endif
 
-	.globl	__lll_robust_mutex_lock_wait
-	.type	__lll_robust_mutex_lock_wait,@function
-	.hidden	__lll_robust_mutex_lock_wait
+	.globl	__lll_robust_lock_wait
+	.type	__lll_robust_lock_wait,@function
+	.hidden	__lll_robust_lock_wait
 	.align	5
 	cfi_startproc
-__lll_robust_mutex_lock_wait:
+__lll_robust_lock_wait:
 	mov.l	r8, @-r15
 	cfi_adjust_cfa_offset(4)
 	cfi_rel_offset (r8, 0)
 	mov	r5, r8
 	mov	#0, r7		/* No timeout.  */
-	mov	#FUTEX_WAIT, r5
+	mov	r6, r5
+	LOAD_FUTEX_WAIT (r5, r0, r1)
 
 4:
 	mov	r4, r6
@@ -90,21 +124,24 @@ __lll_robust_mutex_lock_wait:
 	.long	FUTEX_WAITERS
 .Ltidoff:
 	.word	TID - TLS_PRE_TCB_SIZE
-	.size	__lll_robust_mutex_lock_wait,.-__lll_robust_mutex_lock_wait
+	.size	__lll_robust_lock_wait,.-__lll_robust_lock_wait
 
 
-	.globl	__lll_robust_mutex_timedlock_wait
-	.type	__lll_robust_mutex_timedlock_wait,@function
-	.hidden	__lll_robust_mutex_timedlock_wait
+	.globl	__lll_robust_timedlock_wait
+	.type	__lll_robust_timedlock_wait,@function
+	.hidden	__lll_robust_timedlock_wait
 	.align	5
 	cfi_startproc
-__lll_robust_mutex_timedlock_wait:
+__lll_robust_timedlock_wait:
 	/* Check for a valid timeout value.  */
 	mov.l	@(4,r6), r1
 	mov.l	.L1g, r0
 	cmp/hs	r0, r1
 	bt	3f
 
+	mov.l	r11, @-r15
+	cfi_adjust_cfa_offset(4)
+	cfi_rel_offset (r11, 0)
 	mov.l	r10, @-r15
 	cfi_adjust_cfa_offset(4)
 	cfi_rel_offset (r10, 0)
@@ -114,6 +151,7 @@ __lll_robust_mutex_timedlock_wait:
 	mov.l	r8, @-r15
 	cfi_adjust_cfa_offset(4)
 	cfi_rel_offset (r8, 0)
+	mov	r7, r11
 	mov	r4, r10
 	mov	r6, r9
 	mov	r5, r8
@@ -126,7 +164,7 @@ __lll_robust_mutex_timedlock_wait:
 	/* Get current time.  */
 	mov	r15, r4
 	mov	#0, r5
-	mov	#SYS_gettimeofday, r3
+	mov	#__NR_gettimeofday, r3
 	trapa	#0x12
 	SYSCALL_INST_PAD
 
@@ -167,7 +205,8 @@ __lll_robust_mutex_timedlock_wait:
 
 2:
 	mov	r8, r4
-	mov	#FUTEX_WAIT, r5
+	mov	r11, r5
+	LOAD_FUTEX_WAIT (r5, r0, r1)
 	mov	r10, r6
 	mov	r15, r7
 	mov	#SYS_futex, r3
@@ -196,8 +235,9 @@ __lll_robust_mutex_timedlock_wait:
 	add	#8, r15
 	mov.l	@r15+, r8
 	mov.l	@r15+, r9
+	mov.l	@r15+, r10
 	rts
-	 mov.l	@r15+, r10
+	 mov.l	@r15+, r11
 
 7:
 	/* Check whether the time expired.  */
@@ -221,4 +261,4 @@ __lll_robust_mutex_timedlock_wait:
 	.word	TID - TLS_PRE_TCB_SIZE
 .L1k:
 	.word	1000
-	.size	__lll_robust_mutex_timedlock_wait,.-__lll_robust_mutex_timedlock_wait
+	.size	__lll_robust_timedlock_wait,.-__lll_robust_timedlock_wait
diff --git a/nptl/sysdeps/unix/sysv/linux/sh/pthread_barrier_wait.S b/nptl/sysdeps/unix/sysv/linux/sh/pthread_barrier_wait.S
index 35dd607fc1..c7d7d2278d 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/pthread_barrier_wait.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/pthread_barrier_wait.S
@@ -17,14 +17,10 @@
    02111-1307 USA.  */
 
 #include <sysdep.h>
+#include <lowlevellock.h>
 #include <lowlevelbarrier.h>
 #include "lowlevel-atomic.h"
 
-#define SYS_futex	240
-#define FUTEX_WAIT	0
-#define FUTEX_WAKE	1
-
-
 	.text
 
 	.globl	pthread_barrier_wait
@@ -152,6 +148,10 @@ pthread_barrier_wait:
 	 mov.l	@r15+, r9
 
 1:
+	mov.l	@(PRIVATE,r8), r6
+	mov	#LLL_SHARED, r0
+	extu.b	r0, r0
+	xor	r0, r6
 	mov	r2, r4
 	mov	r8, r5
 	mov.l	.Lwait0, r1
@@ -162,6 +162,10 @@ pthread_barrier_wait:
 	 nop
 
 4:
+	mov.l	@(PRIVATE,r8), r5
+	mov	#LLL_SHARED, r0
+	extu.b	r0, r0
+	xor	r0, r5
 	mov	r8, r4
 	mov.l	.Lwake0, r1
 	bsrf	r1
@@ -172,6 +176,10 @@ pthread_barrier_wait:
 
 6:
 	mov	r6, r9
+	mov.l	@(PRIVATE,r8), r5
+	mov	#LLL_SHARED, r0
+	extu.b	r0, r0
+	xor	r0, r5
 	mov	r8, r4
 	mov.l	.Lwake1, r1
 	bsrf	r1
@@ -182,6 +190,10 @@ pthread_barrier_wait:
 
 9:	
 	mov	r6, r9
+	mov.l	@(PRIVATE,r8), r5
+	mov	#LLL_SHARED, r0
+	extu.b	r0, r0
+	xor	r0, r5
 	mov	r8, r4
 	mov.l	.Lwake2, r1
 	bsrf	r1
@@ -194,11 +206,11 @@ pthread_barrier_wait:
 .Lall:
 	.long	0x7fffffff
 .Lwait0:
-	.long	__lll_mutex_lock_wait-.Lwait0b
+	.long	__lll_lock_wait-.Lwait0b
 .Lwake0:
-	.long	__lll_mutex_unlock_wake-.Lwake0b
+	.long	__lll_unlock_wake-.Lwake0b
 .Lwake1:
-	.long	__lll_mutex_unlock_wake-.Lwake1b
+	.long	__lll_unlock_wake-.Lwake1b
 .Lwake2:
-	.long	__lll_mutex_unlock_wake-.Lwake2b
+	.long	__lll_unlock_wake-.Lwake2b
 	.size	pthread_barrier_wait,.-pthread_barrier_wait
diff --git a/nptl/sysdeps/unix/sysv/linux/sh/pthread_cond_broadcast.S b/nptl/sysdeps/unix/sysv/linux/sh/pthread_cond_broadcast.S
index 56f0aa95de..b173f2d8bf 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/pthread_cond_broadcast.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/pthread_cond_broadcast.S
@@ -1,4 +1,4 @@
-/* Copyright (C) 2003, 2004, 2006 Free Software Foundation, Inc.
+/* Copyright (C) 2003, 2004, 2006, 2007 Free Software Foundation, Inc.
    This file is part of the GNU C Library.
 
    The GNU C Library is free software; you can redistribute it and/or
@@ -18,19 +18,13 @@
 
 #include <sysdep.h>
 #include <shlib-compat.h>
+#include <lowlevellock.h>
 #include <lowlevelcond.h>
 #include <kernel-features.h>
 #include <pthread-pi-defines.h>
+#include <pthread-errnos.h>
 #include "lowlevel-atomic.h"
 
-#define SYS_futex		240
-#define FUTEX_WAIT		0
-#define FUTEX_WAKE		1
-#define FUTEX_REQUEUE		3
-#define FUTEX_CMP_REQUEUE	4
-
-#define EINVAL			22
-
 	.text
 
 	/* int pthread_cond_broadcast (pthread_cond_t *cond) */
@@ -162,10 +156,12 @@ __pthread_cond_broadcast:
 #if cond_lock != 0
 	add	#cond_lock, r5
 #endif
-	mov.l	.Lmwait5, r1
+	mov	#LLL_SHARED, r6
+	extu.b	r6, r6
+	mov.l	.Lwait5, r1
 	bsrf	r1
 	 mov	r2, r4
-.Lmwait5b:
+.Lwait5b:
 	bra	2b
 	 nop
 
@@ -175,10 +171,11 @@ __pthread_cond_broadcast:
 #if cond_lock != 0
 	add	#cond_lock, r4
 #endif
-	mov.l	.Lmwake5, r1
+	mov	#LLL_SHARED, r5
+	mov.l	.Lwake5, r1
 	bsrf	r1
-	 nop
-.Lmwake5b:
+	 extu.b	r5, r5
+.Lwake5b:
 	bra	6b
 	 nop
 
@@ -188,10 +185,11 @@ __pthread_cond_broadcast:
 #if cond_lock != 0
 	add	#cond_lock, r4
 #endif
-	mov.l	.Lmwake6, r1
+	mov	#LLL_SHARED, r5
+	mov.l	.Lwake6, r1
 	bsrf	r1
-	 nop
-.Lmwake6b:
+	 extu.b	r5, r5
+.Lwake6b:
 	bra	8b
 	 nop
 
@@ -208,12 +206,12 @@ __pthread_cond_broadcast:
 	 nop
 
 	.align	2
-.Lmwait5:
-	.long	__lll_mutex_lock_wait-.Lmwait5b
-.Lmwake5:
-	.long	__lll_mutex_unlock_wake-.Lmwake5b
-.Lmwake6:
-	.long	__lll_mutex_unlock_wake-.Lmwake6b
+.Lwait5:
+	.long	__lll_lock_wait-.Lwait5b
+.Lwake5:
+	.long	__lll_unlock_wake-.Lwake5b
+.Lwake6:
+	.long	__lll_unlock_wake-.Lwake6b
 	.size	__pthread_cond_broadcast, .-__pthread_cond_broadcast
 versioned_symbol (libpthread, __pthread_cond_broadcast, pthread_cond_broadcast,
 		  GLIBC_2_3_2)
diff --git a/nptl/sysdeps/unix/sysv/linux/sh/pthread_cond_signal.S b/nptl/sysdeps/unix/sysv/linux/sh/pthread_cond_signal.S
index 8b0196fdca..3ef2d6ee14 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/pthread_cond_signal.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/pthread_cond_signal.S
@@ -1,4 +1,4 @@
-/* Copyright (C) 2003, 2004 Free Software Foundation, Inc.
+/* Copyright (C) 2003, 2004, 2007 Free Software Foundation, Inc.
    This file is part of the GNU C Library.
 
    The GNU C Library is free software; you can redistribute it and/or
@@ -18,17 +18,12 @@
 
 #include <sysdep.h>
 #include <shlib-compat.h>
+#include <lowlevellock.h>
 #include <lowlevelcond.h>
 #include <kernel-features.h>
+#include <pthread-errnos.h>
 #include "lowlevel-atomic.h"
 
-#define SYS_futex		240
-#define FUTEX_WAIT		0
-#define FUTEX_WAKE		1
-#define FUTEX_REQUEUE		3
-
-#define EINVAL			22
-
 	.text
 
 	/* int pthread_cond_signal (pthread_cond_t *cond) */
@@ -108,10 +103,12 @@ __pthread_cond_signal:
 #if cond_lock != 0
 	add	#cond_lock, r5
 #endif
-	mov.l	.Lmwait4, r1
+	mov	#LLL_SHARED, r6
+	extu.b	r6, r6
+	mov.l	.Lwait4, r1
 	bsrf	r1
 	 mov	r2, r4
-.Lmwait4b:
+.Lwait4b:
 	bra	2b
 	 nop
 
@@ -121,18 +118,19 @@ __pthread_cond_signal:
 #if cond_lock != 0
 	add	#cond_lock, r4
 #endif
-	mov.l	.Lmwake4, r1
+	mov	#LLL_SHARED, r5
+	mov.l	.Lwake4, r1
 	bsrf	r1
-	 nop
-.Lmwake4b:
+	 extu.b	r5, r5
+.Lwake4b:
 	bra	6b
 	 nop
 
 	.align	2
-.Lmwait4:
-	.long	__lll_mutex_lock_wait-.Lmwait4b
-.Lmwake4:
-	.long	__lll_mutex_unlock_wake-.Lmwake4b
+.Lwait4:
+	.long	__lll_lock_wait-.Lwait4b
+.Lwake4:
+	.long	__lll_unlock_wake-.Lwake4b
 	.size	__pthread_cond_signal, .-__pthread_cond_signal
 versioned_symbol (libpthread, __pthread_cond_signal, pthread_cond_signal,
 		  GLIBC_2_3_2)
diff --git a/nptl/sysdeps/unix/sysv/linux/sh/pthread_cond_timedwait.S b/nptl/sysdeps/unix/sysv/linux/sh/pthread_cond_timedwait.S
index 0173cfb7d8..17c1e6f567 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/pthread_cond_timedwait.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/pthread_cond_timedwait.S
@@ -18,16 +18,11 @@
 
 #include <sysdep.h>
 #include <shlib-compat.h>
+#include <lowlevellock.h>
 #include <lowlevelcond.h>
 #include <pthread-errnos.h>
 #include "lowlevel-atomic.h"
 
-#define SYS_gettimeofday	__NR_gettimeofday
-#define SYS_futex		240
-#define FUTEX_WAIT		0
-#define FUTEX_WAKE		1
-
-
 	.text
 
 /* int pthread_cond_timedwait (pthread_cond_t *cond, pthread_mutex_t *mutex,
@@ -164,7 +159,7 @@ __pthread_cond_timedwait:
 	mov	r15, r4
 	add	#16, r4
 	mov	#0, r5
-	mov	#SYS_gettimeofday, r3
+	mov	#__NR_gettimeofday, r3
 	trapa	#0x12
 	SYSCALL_INST_PAD
 
@@ -182,7 +177,7 @@ __pthread_cond_timedwait:
 	mov	r15, r4
 	add	#16, r4
 	mov	#0, r5
-	mov	#SYS_gettimeofday, r3
+	mov	#__NR_gettimeofday, r3
 	trapa	#0x12
 	SYSCALL_INST_PAD
 
@@ -403,10 +398,12 @@ __pthread_cond_timedwait:
 #if cond_lock != 0
 	add	#cond_lock, r5
 #endif
-	mov.l	.Lmwait2, r1
+	mov	#LLL_SHARED, r6
+	extu.b	r6, r6
+	mov.l	.Lwait2, r1
 	bsrf	r1
 	 mov	r2, r4
-.Lmwait2b:
+.Lwait2b:
 	bra	2b
 	 nop
 
@@ -416,10 +413,11 @@ __pthread_cond_timedwait:
 #if cond_lock != 0
 	add	#cond_lock, r4
 #endif
-	mov.l	.Lmwake2, r1
+	mov	#LLL_SHARED, r5
+	mov.l	.Lmwait2, r1
 	bsrf	r1
-	 nop
-.Lmwake2b:
+	 extu.b	r5, r5
+.Lmwait2b:
 	bra	4b
 	 nop
 
@@ -429,10 +427,12 @@ __pthread_cond_timedwait:
 #if cond_lock != 0
 	add	#cond_lock, r5
 #endif
-	mov.l	.Lmwait3, r1
+	mov	#LLL_SHARED, r6
+	extu.b	r6, r6
+	mov.l	.Lwait3, r1
 	bsrf	r1
 	 mov	r2, r4
-.Lmwait3b:
+.Lwait3b:
 	bra	6b
 	 nop
 
@@ -442,10 +442,11 @@ __pthread_cond_timedwait:
 #if cond_lock != 0
 	add	#cond_lock, r4
 #endif
-	mov.l	.Lmwake3, r1
+	mov	#LLL_SHARED, r5
+	mov.l	.Lmwait3, r1
 	bsrf	r1
-	 nop
-.Lmwake3b:
+	 extu.b	r5, r5
+.Lmwait3b:
 	bra	11b
 	 nop
 
@@ -464,25 +465,26 @@ __pthread_cond_timedwait:
 #if cond_lock != 0
 	add	#cond_lock, r4
 #endif
-	mov.l	.Lmwake4, r1
+	mov	#LLL_SHARED, r5
+	mov.l	.Lmwait4, r1
 	bsrf	r1
-	 nop
-.Lmwake4b:
+	 extu.b	r5, r5
+.Lmwait4b:
 17:
 	bra	18b
 	 mov.l	@(24,r15), r0
 
 	.align	2
+.Lwait2:
+	.long	__lll_lock_wait-.Lwait2b
 .Lmwait2:
-	.long	__lll_mutex_lock_wait-.Lmwait2b
-.Lmwake2:
-	.long	__lll_mutex_unlock_wake-.Lmwake2b
+	.long	__lll_unlock_wake-.Lmwait2b
+.Lwait3:
+	.long	__lll_lock_wait-.Lwait3b
 .Lmwait3:
-	.long	__lll_mutex_lock_wait-.Lmwait3b
-.Lmwake3:
-	.long	__lll_mutex_unlock_wake-.Lmwake3b
-.Lmwake4:
-	.long	__lll_mutex_unlock_wake-.Lmwake4b
+	.long	__lll_unlock_wake-.Lmwait3b
+.Lmwait4:
+	.long	__lll_unlock_wake-.Lmwait4b
 	.size	__pthread_cond_timedwait, .-__pthread_cond_timedwait
 versioned_symbol (libpthread, __pthread_cond_timedwait, pthread_cond_timedwait,
 		  GLIBC_2_3_2)
@@ -507,10 +509,12 @@ __condvar_tw_cleanup:
 #if cond_lock != 0
 	add	#cond_lock, r5
 #endif
-	mov.l	.Lmwait5, r1
+	mov	#LLL_SHARED, r6
+	extu.b	r6, r6
+	mov.l	.Lwait5, r1
 	bsrf	r1
 	 mov	r2, r4
-.Lmwait5b:
+.Lwait5b:
 
 1:
 	mov.l	@(broadcast_seq,r8), r0
@@ -600,10 +604,11 @@ __condvar_tw_cleanup:
 #if cond_lock != 0
 	add	#cond_lock, r4
 #endif
-	mov.l	.Lmwake5, r1
+	mov	#LLL_SHARED, r5
+	mov.l	.Lmwait5, r1
 	bsrf	r1
-	 nop
-.Lmwake5b:
+	 extu.b	r5, r5
+.Lmwait5b:
 
 2:
 	/* Wake up all waiters to make sure no signal gets lost.  */
@@ -636,10 +641,10 @@ __condvar_tw_cleanup:
 	sleep
 
 	.align	2
+.Lwait5:
+	.long   __lll_lock_wait-.Lwait5b
 .Lmwait5:
-	.long   __lll_mutex_lock_wait-.Lmwait5b
-.Lmwake5:
-        .long   __lll_mutex_unlock_wake-.Lmwake5b
+        .long   __lll_unlock_wake-.Lmwait5b
 .Lmlocki5:
 	.long   __pthread_mutex_cond_lock-.Lmlocki5b
 .Lresume:
diff --git a/nptl/sysdeps/unix/sysv/linux/sh/pthread_cond_wait.S b/nptl/sysdeps/unix/sysv/linux/sh/pthread_cond_wait.S
index 5eb332e484..9ca4c35edc 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/pthread_cond_wait.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/pthread_cond_wait.S
@@ -18,14 +18,10 @@
 
 #include <sysdep.h>
 #include <shlib-compat.h>
+#include <lowlevellock.h>
 #include <lowlevelcond.h>
 #include "lowlevel-atomic.h"
 
-#define SYS_futex		240
-#define FUTEX_WAIT		0
-#define FUTEX_WAKE		1
-
-
 	.text
 
 /* int pthread_cond_wait (pthread_cond_t *cond, pthread_mutex_t *mutex)  */
@@ -267,10 +263,12 @@ __pthread_cond_wait:
 #if cond_lock != 0
 	add	#cond_lock, r5
 #endif
-	mov.l	.Lmwait0, r1
+	mov	#LLL_SHARED, r6
+	extu.b	r6, r6
+	mov.l	.Lwait0, r1
 	bsrf	r1
 	 mov	r2, r4
-.Lmwait0b:
+.Lwait0b:
 	bra	2b
 	 nop
 3:
@@ -279,10 +277,11 @@ __pthread_cond_wait:
 #if cond_lock != 0
 	add	#cond_lock, r4
 #endif
-	mov.l	.Lmwake0, r1
+	mov	#LLL_SHARED, r5
+	mov.l	.Lwake0, r1
 	bsrf	r1
-	 nop
-.Lmwake0b:
+	 extu.b	r5, r5
+.Lwake0b:
 	bra	4b
 	 nop
 
@@ -292,10 +291,12 @@ __pthread_cond_wait:
 #if cond_lock != 0
 	add	#cond_lock, r5
 #endif
-	mov.l	.Lmwait1, r1
+	mov	#LLL_SHARED, r6
+	extu.b	r6, r6
+	mov.l	.Lwait1, r1
 	bsrf	r1
 	 mov	r2, r4
-.Lmwait1b:
+.Lwait1b:
 	bra	6b
 	 nop
 
@@ -305,10 +306,11 @@ __pthread_cond_wait:
 #if cond_lock != 0
 	add	#cond_lock, r4
 #endif
-	mov.l	.Lmwake1, r1
+	mov	#LLL_SHARED, r5
+	mov.l	.Lwake1, r1
 	bsrf	r1
-	 nop
-.Lmwake1b:
+	 extu.b	r5, r5
+.Lwake1b:
 	bra	11b
 	 nop
 
@@ -327,26 +329,27 @@ __pthread_cond_wait:
 #if cond_lock != 0
 	add	#cond_lock, r4
 #endif
-	mov.l	.Lmwake2, r1
+	mov	#LLL_SHARED, r5
+	mov.l	.Lwake2, r1
 	bsrf	r1
-	 nop
-.Lmwake2b:
+	 extu.b	r5, r5
+.Lwake2b:
 
 13:
 	bra	14b
 	 mov.l	@(12,r15), r0
 
 	.align	2
-.Lmwait0:
-	.long	__lll_mutex_lock_wait-.Lmwait0b
-.Lmwake0:
-	.long	__lll_mutex_unlock_wake-.Lmwake0b
-.Lmwait1:
-	.long	__lll_mutex_lock_wait-.Lmwait1b
-.Lmwake1:
-	.long	__lll_mutex_unlock_wake-.Lmwake1b
-.Lmwake2:
-	.long	__lll_mutex_unlock_wake-.Lmwake2b
+.Lwait0:
+	.long	__lll_lock_wait-.Lwait0b
+.Lwake0:
+	.long	__lll_unlock_wake-.Lwake0b
+.Lwait1:
+	.long	__lll_lock_wait-.Lwait1b
+.Lwake1:
+	.long	__lll_unlock_wake-.Lwake1b
+.Lwake2:
+	.long	__lll_unlock_wake-.Lwake2b
 	.size	__pthread_cond_wait, .-__pthread_cond_wait
 versioned_symbol (libpthread, __pthread_cond_wait, pthread_cond_wait,
 		  GLIBC_2_3_2)
@@ -371,10 +374,12 @@ __condvar_w_cleanup:
 #if cond_lock != 0
 	add	#cond_lock, r5
 #endif
-	mov.l	.Lmwait3, r1
+	mov	#LLL_SHARED, r6
+	extu.b	r6, r6
+	mov.l	.Lwait3, r1
 	bsrf	r1
 	 mov	r2, r4
-.Lmwait3b:
+.Lwait3b:
 
 1:
 	mov.l	@(broadcast_seq,r8), r0
@@ -464,10 +469,11 @@ __condvar_w_cleanup:
 #if cond_lock != 0
 	add	#cond_lock, r4
 #endif
-	mov.l	.Lmwake3, r1
+	mov	#LLL_SHARED, r5
+	mov.l	.Lwake3, r1
 	bsrf	r1
-	 nop
-.Lmwake3b:
+	 extu.b	r5, r5
+.Lwake3b:
 
 2:
 	/* Wake up all waiters to make sure no signal gets lost.  */
@@ -500,10 +506,10 @@ __condvar_w_cleanup:
 	sleep
 
 	.align	2
-.Lmwait3:
-	.long   __lll_mutex_lock_wait-.Lmwait3b
-.Lmwake3:
-        .long   __lll_mutex_unlock_wake-.Lmwake3b
+.Lwait3:
+	.long   __lll_lock_wait-.Lwait3b
+.Lwake3:
+        .long   __lll_unlock_wake-.Lwake3b
 .Lmlocki3:
 	.long   __pthread_mutex_cond_lock-.Lmlocki3b
 .Lresume:
diff --git a/nptl/sysdeps/unix/sysv/linux/sh/pthread_once.S b/nptl/sysdeps/unix/sysv/linux/sh/pthread_once.S
index 3d694d8376..439907502a 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/pthread_once.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/pthread_once.S
@@ -19,12 +19,9 @@
 #include <unwindbuf.h>
 #include <sysdep.h>
 #include <kernel-features.h>
+#include <lowlevellock.h>
 #include "lowlevel-atomic.h"
 
-#define SYS_futex		240
-#define FUTEX_WAIT		0
-#define FUTEX_WAKE		1
-#define FUTEX_PRIVATE_FLAG	128
 
 	.comm	__fork_generation, 4, 4
 
diff --git a/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_rdlock.S b/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_rdlock.S
index f1795131f8..9e4ba959b5 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_rdlock.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_rdlock.S
@@ -17,17 +17,13 @@
    02111-1307 USA.  */
 
 #include <sysdep.h>
+#include <lowlevellock.h>
 #include <lowlevelrwlock.h>
 #include <pthread-errnos.h>
 #include <tcb-offsets.h>
 #include <kernel-features.h>
 #include "lowlevel-atomic.h"
 
-#define SYS_futex		240
-#define FUTEX_WAIT		0
-#define FUTEX_WAKE		1
-#define FUTEX_PRIVATE_FLAG	128
-
 
 	.text
 
@@ -156,10 +152,12 @@ __pthread_rwlock_rdlock:
 #if MUTEX != 0
 	add	#MUTEX, r5
 #endif
-	mov	r2, r4
+	mov	#PSHARED, r0
+	mov.b	@(r0,r8), r6
+	extu.b	r6, r6
 	mov.l	.Lwait0, r1
 	bsrf	r1
-	 nop
+	 mov	r2, r4
 .Lwait0b:
 	bra	2b
 	 nop
@@ -182,6 +180,9 @@ __pthread_rwlock_rdlock:
 #if MUTEX != 0
 	add	#MUTEX, r4
 #endif
+	mov	#PSHARED, r0
+	mov.b	@(r0,r8), r5
+	extu.b	r5, r5
 	mov.l	.Lwake0, r1
 	bsrf	r1
 	 nop
@@ -210,6 +211,9 @@ __pthread_rwlock_rdlock:
 #if MUTEX != 0
 	add	#MUTEX, r4
 #endif
+	mov	#PSHARED, r0
+	mov.b	@(r0,r8), r5
+	extu.b	r5, r5
 	mov.l	.Lwake1, r1
 	bsrf	r1
 	 nop
@@ -222,23 +226,25 @@ __pthread_rwlock_rdlock:
 #if MUTEX != 0
 	add	#MUTEX, r5
 #endif
-	mov	r2, r4
+	mov	#PSHARED, r0
+	mov.b	@(r0,r8), r6
+	extu.b	r6, r6
 	mov.l	.Lwait1, r1
 	bsrf	r1
-	 nop
+	 mov	r2, r4
 .Lwait1b:
 	bra	13b
 	 nop
 
 	.align	2
 .Lwait0:
-	.long	__lll_mutex_lock_wait-.Lwait0b
+	.long	__lll_lock_wait-.Lwait0b
 .Lwake0:
-	.long	__lll_mutex_unlock_wake-.Lwake0b
+	.long	__lll_unlock_wake-.Lwake0b
 .Lwait1:
-	.long	__lll_mutex_lock_wait-.Lwait1b
+	.long	__lll_lock_wait-.Lwait1b
 .Lwake1:
-	.long	__lll_mutex_unlock_wake-.Lwake1b
+	.long	__lll_unlock_wake-.Lwake1b
 	.size	__pthread_rwlock_rdlock,.-__pthread_rwlock_rdlock
 
 	.globl	pthread_rwlock_rdlock
diff --git a/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_timedrdlock.S b/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_timedrdlock.S
index e87326e9bd..b06fd960cf 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_timedrdlock.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_timedrdlock.S
@@ -17,18 +17,13 @@
    02111-1307 USA.  */
 
 #include <sysdep.h>
+#include <lowlevellock.h>
 #include <lowlevelrwlock.h>
 #include <pthread-errnos.h>
 #include <tcb-offsets.h>
 #include <kernel-features.h>
 #include "lowlevel-atomic.h"
 
-#define SYS_gettimeofday	__NR_gettimeofday
-#define SYS_futex		240
-#define FUTEX_WAIT		0
-#define FUTEX_WAKE		1
-#define FUTEX_PRIVATE_FLAG	128
-
 
 	.text
 
@@ -92,7 +87,7 @@ pthread_rwlock_timedrdlock:
 	/* Get current time.  */
 	mov	r15, r4
 	mov	#0, r5
-	mov	#SYS_gettimeofday, r3
+	mov	#__NR_gettimeofday, r3
 	trapa	#0x12
 	SYSCALL_INST_PAD
 
@@ -213,10 +208,12 @@ pthread_rwlock_timedrdlock:
 #if MUTEX != 0
 	add	#MUTEX, r5
 #endif
-	mov	r2, r4
+	mov	#PSHARED, r0
+	mov.b	@(r0,r8), r6
+	extu.b	r6, r6
 	mov.l	.Lwait2, r1
 	bsrf	r1
-	 nop
+	 mov	r2, r4
 .Lwait2b:
 	bra	2b
 	 nop
@@ -239,6 +236,9 @@ pthread_rwlock_timedrdlock:
 #if MUTEX != 0
 	add	#MUTEX, r4
 #endif
+	mov	#PSHARED, r0
+	mov.b	@(r0,r8), r5
+	extu.b	r5, r5
 	mov.l	.Lwake2, r1
 	bsrf	r1
 	 nop
@@ -267,6 +267,9 @@ pthread_rwlock_timedrdlock:
 #if MUTEX != 0
 	add	#MUTEX, r4
 #endif
+	mov	#PSHARED, r0
+	mov.b	@(r0,r8), r5
+	extu.b	r5, r5
 	mov.l	.Lwake3, r1
 	bsrf	r1
 	 nop
@@ -279,10 +282,12 @@ pthread_rwlock_timedrdlock:
 #if MUTEX != 0
 	add	#MUTEX, r5
 #endif
-	mov	r2, r4
+	mov	#PSHARED, r0
+	mov.b	@(r0,r8), r6
+	extu.b	r6, r6
 	mov.l	.Lwait3, r1
 	bsrf	r1
-	 nop
+	 mov	r2, r4
 .Lwait3b:
 	bra	13b
 	 nop
@@ -297,11 +302,11 @@ pthread_rwlock_timedrdlock:
 
 	.align	2
 .Lwait2:
-	.long	__lll_mutex_lock_wait-.Lwait2b
+	.long	__lll_lock_wait-.Lwait2b
 .Lwake2:
-	.long	__lll_mutex_unlock_wake-.Lwake2b
+	.long	__lll_unlock_wake-.Lwake2b
 .Lwait3:
-	.long	__lll_mutex_lock_wait-.Lwait3b
+	.long	__lll_lock_wait-.Lwait3b
 .Lwake3:
-	.long	__lll_mutex_unlock_wake-.Lwake3b
+	.long	__lll_unlock_wake-.Lwake3b
 	.size	pthread_rwlock_timedrdlock,.-pthread_rwlock_timedrdlock
diff --git a/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_timedwrlock.S b/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_timedwrlock.S
index 18641fe9df..4591309ecc 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_timedwrlock.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_timedwrlock.S
@@ -17,18 +17,13 @@
    02111-1307 USA.  */
 
 #include <sysdep.h>
+#include <lowlevellock.h>
 #include <lowlevelrwlock.h>
 #include <pthread-errnos.h>
 #include <tcb-offsets.h>
 #include <kernel-features.h>
 #include "lowlevel-atomic.h"
 
-#define SYS_gettimeofday	__NR_gettimeofday
-#define SYS_futex		240
-#define FUTEX_WAIT		0
-#define FUTEX_WAKE		1
-#define FUTEX_PRIVATE_FLAG    128
-
 
 	.text
 
@@ -88,7 +83,7 @@ pthread_rwlock_timedwrlock:
 	/* Get current time.  */
 	mov	r15, r4
 	mov	#0, r5
-	mov	#SYS_gettimeofday, r3
+	mov	#__NR_gettimeofday, r3
 	trapa	#0x12
 	SYSCALL_INST_PAD
 
@@ -211,10 +206,12 @@ pthread_rwlock_timedwrlock:
 #if MUTEX != 0
 	add	#MUTEX, r5
 #endif
-	mov	r2, r4
+	mov	#PSHARED, r0
+	mov.b	@(r0,r8), r6
+	extu.b	r6, r6
 	mov.l	.Lwait6, r1
 	bsrf	r1
-	 nop
+	 mov	r2, r4
 .Lwait6b:
 	bra	2b
 	 nop
@@ -232,6 +229,9 @@ pthread_rwlock_timedwrlock:
 #if MUTEX != 0
 	add	#MUTEX, r4
 #endif
+	mov	#PSHARED, r0
+	mov.b	@(r0,r8), r5
+	extu.b	r5, r5
 	mov.l	.Lwake6, r1
 	bsrf	r1
 	 nop
@@ -255,6 +255,9 @@ pthread_rwlock_timedwrlock:
 #if MUTEX != 0
 	add	#MUTEX, r4
 #endif
+	mov	#PSHARED, r0
+	mov.b	@(r0,r8), r5
+	extu.b	r5, r5
 	mov.l	.Lwake7, r1
 	bsrf	r1
 	 nop
@@ -267,10 +270,12 @@ pthread_rwlock_timedwrlock:
 #if MUTEX != 0
 	add	#MUTEX, r5
 #endif
-	mov	r2, r4
+	mov	#PSHARED, r0
+	mov.b	@(r0,r8), r6
+	extu.b	r6, r6
 	mov.l	.Lwait7, r1
 	bsrf	r1
-	 nop
+	 mov	r2, r4
 .Lwait7b:
 	bra	13b
 	 nop
@@ -281,11 +286,11 @@ pthread_rwlock_timedwrlock:
 
 	.align	2
 .Lwait6:
-	.long	__lll_mutex_lock_wait-.Lwait6b
+	.long	__lll_lock_wait-.Lwait6b
 .Lwake6:
-	.long	__lll_mutex_unlock_wake-.Lwake6b
+	.long	__lll_unlock_wake-.Lwake6b
 .Lwait7:
-	.long	__lll_mutex_lock_wait-.Lwait7b
+	.long	__lll_lock_wait-.Lwait7b
 .Lwake7:
-	.long	__lll_mutex_unlock_wake-.Lwake7b
+	.long	__lll_unlock_wake-.Lwake7b
 	.size	pthread_rwlock_timedwrlock,.-pthread_rwlock_timedwrlock
diff --git a/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_unlock.S b/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_unlock.S
index df4df60d7f..1504c1aed1 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_unlock.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_unlock.S
@@ -17,15 +17,11 @@
    02111-1307 USA.  */
 
 #include <sysdep.h>
+#include <lowlevellock.h>
 #include <lowlevelrwlock.h>
 #include <kernel-features.h>
 #include "lowlevel-atomic.h"
 
-#define SYS_futex		240
-#define FUTEX_WAIT		0
-#define FUTEX_WAKE		1
-#define FUTEX_PRIVATE_FLAG	128
-
 
 	.text
 
@@ -138,10 +134,12 @@ __pthread_rwlock_unlock:
 #if MUTEX != 0
 	add	#MUTEX, r5
 #endif
-	mov	r2, r4
+	mov	#PSHARED, r0
+	mov.b	@(r0,r8), r6
+	extu.b	r6, r6
 	mov.l	.Lwait8, r1
 	bsrf	r1
-	 nop
+	 mov	r2, r4
 .Lwait8b:
 	bra	2b
 	 nop
@@ -150,6 +148,9 @@ __pthread_rwlock_unlock:
 #if MUTEX != 0
 	add	#MUTEX, r4
 #endif
+	mov	#PSHARED, r0
+	mov.b	@(r0,r8), r5
+	extu.b	r5, r5
 	mov.l	.Lwake8, r1
 	bsrf	r1
 	 nop
@@ -164,6 +165,9 @@ __pthread_rwlock_unlock:
 #if MUTEX != 0
 	add	#MUTEX, r4
 #endif
+	mov	#PSHARED, r0
+	mov.b	@(r0,r8), r5
+	extu.b	r5, r5
 	mov.l	.Lwake9, r1
 	bsrf	r1
 	 nop
@@ -179,11 +183,11 @@ __pthread_rwlock_unlock:
 #endif
 	.align	2
 .Lwait8:	
-	.long	__lll_mutex_lock_wait-.Lwait8b
+	.long	__lll_lock_wait-.Lwait8b
 .Lwake8:
-	.long	__lll_mutex_unlock_wake-.Lwake8b
+	.long	__lll_unlock_wake-.Lwake8b
 .Lwake9:
-	.long	__lll_mutex_unlock_wake-.Lwake9b
+	.long	__lll_unlock_wake-.Lwake9b
 	.size	__pthread_rwlock_unlock,.-__pthread_rwlock_unlock
 
 	.globl	pthread_rwlock_unlock
diff --git a/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_wrlock.S b/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_wrlock.S
index 13a2fda9fd..25dffe4bf9 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_wrlock.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_wrlock.S
@@ -17,17 +17,13 @@
    02111-1307 USA.  */
 
 #include <sysdep.h>
+#include <lowlevellock.h>
 #include <lowlevelrwlock.h>
 #include <pthread-errnos.h>
 #include <tcb-offsets.h>
 #include <kernel-features.h>
 #include "lowlevel-atomic.h"
 
-#define SYS_futex		240
-#define FUTEX_WAIT		0
-#define FUTEX_WAKE		1
-#define FUTEX_PRIVATE_FLAG	128
-
 
 	.text
 
@@ -145,10 +141,12 @@ __pthread_rwlock_wrlock:
 #if MUTEX != 0
 	add	#MUTEX, r5
 #endif
-	mov	r2, r4
+	mov	#PSHARED, r0
+	mov.b	@(r0,r8), r6
+	extu.b	r6, r6
 	mov.l	.Lwait4, r1
 	bsrf	r1
-	 nop
+	 mov	r2, r4
 .Lwait4b:
 	bra	2b
 	 nop
@@ -166,6 +164,9 @@ __pthread_rwlock_wrlock:
 #if MUTEX != 0
 	add	#MUTEX, r4
 #endif
+	mov	#PSHARED, r0
+	mov.b	@(r0,r8), r5
+	extu.b	r5, r5
 	mov.l	.Lwake4, r1
 	bsrf	r1
 	 nop
@@ -192,6 +193,9 @@ __pthread_rwlock_wrlock:
 #if MUTEX != 0
 	add	#MUTEX, r4
 #endif
+	mov	#PSHARED, r0
+	mov.b	@(r0,r8), r5
+	extu.b	r5, r5
 	mov.l	.Lwake5, r1
 	bsrf	r1
 	 nop
@@ -204,23 +208,25 @@ __pthread_rwlock_wrlock:
 #if MUTEX != 0
 	add	#MUTEX, r5
 #endif
-	mov	r2, r4
+	mov	#PSHARED, r0
+	mov.b	@(r0,r8), r6
+	extu.b	r6, r6
 	mov.l	.Lwait5, r1
 	bsrf	r1
-	 nop
+	 mov	r2, r4
 .Lwait5b:
 	bra	13b
 	 nop
 
 	.align	2
 .Lwait4:
-	.long	__lll_mutex_lock_wait-.Lwait4b
+	.long	__lll_lock_wait-.Lwait4b
 .Lwake4:
-	.long	__lll_mutex_unlock_wake-.Lwake4b
+	.long	__lll_unlock_wake-.Lwake4b
 .Lwait5:
-	.long	__lll_mutex_lock_wait-.Lwait5b
+	.long	__lll_lock_wait-.Lwait5b
 .Lwake5:
-	.long	__lll_mutex_unlock_wake-.Lwake5b
+	.long	__lll_unlock_wake-.Lwake5b
 	.globl	pthread_rwlock_wrlock
 pthread_rwlock_wrlock = __pthread_rwlock_wrlock
 
diff --git a/nptl/sysdeps/unix/sysv/linux/sh/sem_post.S b/nptl/sysdeps/unix/sysv/linux/sh/sem_post.S
index a63233feab..a3a7710ef9 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/sem_post.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/sem_post.S
@@ -20,14 +20,10 @@
 #include <shlib-compat.h>
 #include <pthread-errnos.h>
 #include <structsem.h>
+#include <lowlevellock.h>
 #include "lowlevel-atomic.h"
 
 
-#define SYS_gettimeofday	__NR_gettimeofday
-#define SYS_futex		240
-#define FUTEX_WAIT		0
-#define FUTEX_WAKE		1
-
 	.text
 
 	.globl	__new_sem_post
@@ -65,7 +61,12 @@ __new_sem_post:
 	mov.l	.Lerrno3, r0
 	stc	gbr, r1
 	mov.l	@(r0, r12), r0
-	add	r1, r0
+	bra	.Lexit
+	 add	r1, r0
+	.align	2
+.Lerrno3:
+	.long	errno@GOTTPOFF
+.Lexit:
 #else
 	mov.l	.Lerrloc3, r1
 	bsrf	r1
@@ -81,10 +82,7 @@ __new_sem_post:
 	.align	2
 .Lgot3:
 	.long	_GLOBAL_OFFSET_TABLE_
-#if USE___THREAD
-.Lerrno3:
-	.long	errno@GOTTPOFF
-#else
+#if !USE___THREAD
 .Lerrloc3:
 	.long	__errno_location@PLT-(.Lerrloc3b-.)
 #endif
diff --git a/nptl/sysdeps/unix/sysv/linux/sh/sem_timedwait.S b/nptl/sysdeps/unix/sysv/linux/sh/sem_timedwait.S
index e5e064b3a5..7924e87594 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/sem_timedwait.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/sem_timedwait.S
@@ -21,13 +21,10 @@
 #include <pthread-errnos.h>
 #include <tcb-offsets.h>
 #include <structsem.h>
+#include <lowlevellock.h>
 #include "lowlevel-atomic.h"
 
 
-#define SYS_gettimeofday	__NR_gettimeofday
-#define SYS_futex		240
-#define FUTEX_WAIT		0
-
 #if VALUE != 0
 # error "code needs to be rewritten for VALUE != 0"
 #endif
@@ -81,7 +78,7 @@ sem_timedwait:
 	/* Compute relative timeout.  */
 	mov	r15, r4
 	mov	#0, r5
-	mov	#SYS_gettimeofday, r3
+	mov	#__NR_gettimeofday, r3
 	trapa	#0x12
 	SYSCALL_INST_PAD
 
@@ -180,15 +177,19 @@ sem_timedwait:
 	mov.l	.Lerrno2, r0
 	stc	gbr, r1
 	mov.l	@(r0, r12), r0
-	add	r1, r0
-	mov.l	r10, @r0
+	bra	.Lexit
+	 add	r1, r0
+	.align	2
+.Lerrno2:
+ 	.long	errno@GOTTPOFF
+.Lexit:
 #else
 	mov.l	.Lerrloc2, r1
 	bsrf	r1
 	 nop
 .Lerrloc2b:
-	mov.l	r10, @r0
 #endif
+	mov.l	r10, @r0
 	DEC (@(NWAITERS,r8), r2)
 	bra	10b
 	 mov	#-1, r0
@@ -200,10 +201,7 @@ sem_timedwait:
 	.long	1000000000
 .Lgot2:
 	.long	_GLOBAL_OFFSET_TABLE_
-#if USE___THREAD
-.Lerrno2:
-	.long	errno@GOTTPOFF
-#else
+#if !USE___THREAD
 .Lerrloc2:
 	.long	__errno_location@PLT-(.Lerrloc2b-.)
 #endif
diff --git a/nptl/sysdeps/unix/sysv/linux/sh/sem_trywait.S b/nptl/sysdeps/unix/sysv/linux/sh/sem_trywait.S
index ccdf3a0dbd..835e37db9f 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/sem_trywait.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/sem_trywait.S
@@ -1,4 +1,4 @@
-/* Copyright (C) 2003, 2004 Free Software Foundation, Inc.
+/* Copyright (C) 2003, 2004, 2007 Free Software Foundation, Inc.
    This file is part of the GNU C Library.
 
    The GNU C Library is free software; you can redistribute it and/or
@@ -19,6 +19,7 @@
 #include <sysdep.h>
 #include <shlib-compat.h>
 #include <pthread-errnos.h>
+#include <lowlevellock.h>
 #include "lowlevel-atomic.h"
 
 
@@ -60,15 +61,19 @@ __new_sem_trywait:
 	mov.l	.Lerrno1, r0
 	stc	gbr, r1
 	mov.l	@(r0, r12), r0
-	add	r1, r0
-	mov.l	r8, @r0
+	bra	.Lexit
+	 add	r1, r0
+	.align	2
+.Lerrno1:
+	.long	errno@GOTTPOFF
+.Lexit:
 #else
 	mov.l	.Lerrloc1, r1
 	bsrf	r1
 	 nop
 .Lerrloc1b:
-	mov.l	r8, @r0
 #endif
+	mov.l	r8, @r0
 	lds.l	@r15+, pr
 	mov.l	@r15+, r8
 	mov.l	@r15+, r12
@@ -78,10 +83,7 @@ __new_sem_trywait:
 	.align	2
 .Lgot1:
 	.long	_GLOBAL_OFFSET_TABLE_
-#if USE___THREAD
-.Lerrno1:
-	.long	errno@GOTTPOFF
-#else
+#if !USE___THREAD
 .Lerrloc1:
 	.long	__errno_location@PLT-(.Lerrloc1b-.)
 #endif
diff --git a/nptl/sysdeps/unix/sysv/linux/sh/sem_wait.S b/nptl/sysdeps/unix/sysv/linux/sh/sem_wait.S
index 85ce909259..6ef7fcf0cb 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/sem_wait.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/sem_wait.S
@@ -1,4 +1,4 @@
-/* Copyright (C) 2003, 2004 Free Software Foundation, Inc.
+/* Copyright (C) 2003, 2004, 2007 Free Software Foundation, Inc.
    This file is part of the GNU C Library.
 
    The GNU C Library is free software; you can redistribute it and/or
@@ -21,13 +21,10 @@
 #include <pthread-errnos.h>
 #include <tcb-offsets.h>
 #include <structsem.h>
+#include <lowlevellock.h>
 #include "lowlevel-atomic.h"
 
 
-#define SYS_gettimeofday	__NR_gettimeofday
-#define SYS_futex		240
-#define FUTEX_WAIT		0
-
 #if VALUE != 0
 # error "code needs to be rewritten for VALUE != 0"
 #endif
@@ -138,24 +135,26 @@ __new_sem_wait:
 	mov.l	.Lerrno0, r0
 	stc	gbr, r1
 	mov.l	@(r0, r12), r0
-	add	r1, r0
-	mov.l	r8, @r0
+	bra	.Lexit
+	 add	r1, r0
+	.align	2
+.Lerrno0:
+	.long	errno@GOTTPOFF
+.Lexit:
 #else
 	mov.l	.Lerrloc0, r1
 	bsrf	r1
 	 nop
 .Lerrloc0b:
-	mov.l	r8, @r0
 #endif
+	mov.l	r8, @r0
 	bra	9b
 	 mov	#-1, r0
 
+	.align	2
 .Lgot0:
 	.long	_GLOBAL_OFFSET_TABLE_
-#if USE___THREAD
-.Lerrno0:
-	.long	errno@GOTTPOFF
-#else
+#if !USE___THREAD
 .Lerrloc0:
 	.long	__errno_location@PLT-(.Lerrloc0b-.)
 #endif