Update.

2003-03-28  Kaz Kojima  <kkojima@rr.iij4u.or.jp>

	* sysdeps/sh/bits/atomic.h (__arch_compare_and_exchange_val_8_acq):
	Return old value. Make asm output reg constraint earlyclobber.
	Renamed from...
	(__arch_compare_and_exchange_8_acq): ... this.
	(__arch_compare_and_exchange_val_16_acq):
	Return old value. Make asm output reg constraint earlyclobber.
	Renamed from...
	(__arch_compare_and_exchange_16_acq): ... this.
	(__arch_compare_and_exchange_val_32_acq):
	Return old value. Make asm output reg constraint earlyclobber.
	Renamed from...
	(__arch_compare_and_exchange_32_acq): ... this.
	(__arch_compare_and_exchange_val_64_acq):
	Renamed from...
	(__arch_compare_and_exchange_64_acq): ... this.
	(atomic_exchange_and_add): Use local variables and
	__arch_compare_and_exchange_val_64_acq.
	(atomic_add): Likewise.
	(atomic_add_negative, atomic_add_zero): Use local variables.

	* Makefile: Remove libmd5crypt goal.

17 files changed, 382 insertions, 160 deletions

diff --git a/ChangeLog b/ChangeLog
index 7dbe76da04..052274ed3e 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -16,15 +16,35 @@
 	* sysdeps/powerpc/bits/atomic.h [! __powerpc64__]
 	(__arch_atomic_decrement_if_positive_64): Fix bogus definition.
 
+2003-03-28  Kaz Kojima  <kkojima@rr.iij4u.or.jp>
+
+	* sysdeps/sh/bits/atomic.h (__arch_compare_and_exchange_val_8_acq):
+	Return old value. Make asm output reg constraint earlyclobber.
+	Renamed from...
+	(__arch_compare_and_exchange_8_acq): ... this.
+	(__arch_compare_and_exchange_val_16_acq):
+	Return old value. Make asm output reg constraint earlyclobber.
+	Renamed from...
+	(__arch_compare_and_exchange_16_acq): ... this.
+	(__arch_compare_and_exchange_val_32_acq):
+	Return old value. Make asm output reg constraint earlyclobber.
+	Renamed from...
+	(__arch_compare_and_exchange_32_acq): ... this.
+	(__arch_compare_and_exchange_val_64_acq):
+	Renamed from...
+	(__arch_compare_and_exchange_64_acq): ... this.
+	(atomic_exchange_and_add): Use local variables and
+	__arch_compare_and_exchange_val_64_acq.
+	(atomic_add): Likewise.
+	(atomic_add_negative, atomic_add_zero): Use local variables.
+
 2003-03-28  Alexandre Oliva  <aoliva@redhat.com>
 
 	* sysdeps/unix/mips/sysdep.S: Include sys/asm.h.
 
-	* sysdeps/unix/sysv/linux/mips/configure: Rebuilt.
-
 2003-03-27  Ulrich Drepper  <drepper@redhat.com>
 
-	* Makefile: Remove libmd5script goal.
+	* Makefile: Remove libmd5crypt goal.
 
 2003-03-25  Jakub Jelinek  <jakub@redhat.com>
 
diff --git a/nptl/ChangeLog b/nptl/ChangeLog
index 3bf0064ce6..0724d7214f 100644
--- a/nptl/ChangeLog
+++ b/nptl/ChangeLog
@@ -1,3 +1,36 @@
+2003-03-28  Kaz Kojima  <kkojima@rr.iij4u.or.jp>
+
+	* sysdeps/sh/tls.h: Include nptl/descr.h after the definition
+	of TLS_DTV_AT_TP.
+	(INSTALL_DTV): Add parens.
+	(THREAD_GETMEM, THREAD_GETMEM_NC, THREAD_SETMEM, THREAD_SETMEM_NC):
+	Use passed descr instead of THREAD_SELF.
+	* sysdeps/unix/sysv/linux/sh/lowlevelmutex.S
+	(__lll_mutex_timedlock_wait): Correct expected value after
+	spurious wakeup.
+	* sysdeps/unix/sysv/linux/sh/pthread_cond_broadcast.S:
+	Release lock before waking up the waiters.
+	* sysdeps/unix/sysv/linux/sh/pthread_cond_wait.S: Correct exit
+	criteria.  Reorderstruct passed to cleanup handler.  Fix
+	handling of cancellation and failung pthread_mutex_unlock call.
+	Use __pthread_enable_asynccancel_2 instead of
+	__pthread_enable_asynccancel.
+	* sysdeps/unix/sysv/linux/sh/pthread_cond_timedwait.S: Likewise.
+	Return result of lock re-get if it fails.
+	* sysdeps/unix/sysv/linux/sh/pthread_once.S: Fix wrong argument
+	for __pthread_cleanup_push.
+	* sysdeps/unix/sysv/linux/sh/pthread_rwlock_rdlock.S: Fix
+	completely broken rwlock implementation.
+	* sysdeps/unix/sysv/linux/sh/pthread_rwlock_wrlock.S: Likewise.
+	* sysdeps/unix/sysv/linux/sh/pthread_rwlock_timedrdlock.S: Likewise.
+	* sysdeps/unix/sysv/linux/sh/pthread_rwlock_timedwrlock.S: Likewise.
+	* sysdeps/unix/sysv/linux/sh/pthread_rwlock_unlock.S: Likewise.
+	* sysdeps/unix/sysv/linux/sh/pthread_rwlock_wrlock.S: Likewise.
+	* sysdeps/unix/sysv/linux/sh/sem_post.S: Fix error value.  Use
+	versioned_symbol macro.
+	* sysdeps/unix/sysv/linux/sh/sem_trywait.S: Use	versioned_symbol macro.
+	* sysdeps/unix/sysv/linux/sh/sem_wait.S: Likewise.
+
 2003-03-27  Ulrich Drepper  <drepper@redhat.com>
 
 	* sysdeps/unix/sysv/linux/kernel-posix-timers.h: Don't declare
diff --git a/nptl/sysdeps/unix/sysv/linux/sh/lowlevellock.S b/nptl/sysdeps/unix/sysv/linux/sh/lowlevellock.S
index ecb0059036..115c5292dd 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/lowlevellock.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/lowlevellock.S
@@ -177,8 +177,6 @@ __lll_timedwait_tid:
 	trapa	#0x14
 	SYSCALL_INST_PAD
 
-	mov	r0, r1
-
 	mov.l	@r8, r0
 	tst	r0, r0
 	bf	1f
diff --git a/nptl/sysdeps/unix/sysv/linux/sh/lowlevelmutex.S b/nptl/sysdeps/unix/sysv/linux/sh/lowlevelmutex.S
index 6df513cbec..0aa2048d11 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/lowlevelmutex.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/lowlevelmutex.S
@@ -73,12 +73,13 @@ __lll_mutex_timedlock_wait:
 	mov	r5, r8
 	mov	r6, r9
 	mov	r4, r10
-	add	#1, r10
 
 	/* Stack frame for the timespec and timeval structs.  */
 	add	#-8, r15
 
 1:
+	add	#1, r10
+
 	/* Get current time.  */
 	mov	r15, r4
 	mov	#0, r5
@@ -116,10 +117,11 @@ __lll_mutex_timedlock_wait:
 	extu.b	r3, r3
 	trapa	#0x14
 	SYSCALL_INST_PAD
+	mov	r0, r4
 
-	mov	#1, r3
-	XADD (r3, @r8, r10)
-	tst	r10, r10
+	mov	#1, r10
+	XADD (r10, @r8, r3)
+	tst	r3, r3
 	bf	7f
 
 	mov	#2, r1
@@ -134,7 +136,7 @@ __lll_mutex_timedlock_wait:
 7:
 	/* Check whether the time expired.  */
 	mov	#-ETIMEDOUT, r1
-	cmp/eq	r0, r1
+	cmp/eq	r4, r1
 	bt	5f
 	bra	1b
 	 nop
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 93818f96df..8dd31aac29 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/pthread_cond_broadcast.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/pthread_cond_broadcast.S
@@ -59,10 +59,20 @@ __pthread_cond_broadcast:
 	bf	4f
 
 3:
-	/* Case all currently waiting threads to wake up.  */
+	/* Cause all currently waiting threads to recognize they are
+	   woken up.  */
 	mov.l	r1, @(wakeup_seq,r8)
 	mov.l	r0, @(wakeup_seq+4,r8)
 
+	/* Unlock.  */
+#if cond_lock != 0
+	DEC (@(cond_lock,r8), r2)
+#else
+	DEC (@r8, r2)
+#endif
+	tst	r2, r2
+	bf	7f
+8:
 	/* Wake up all threads.  */
 	mov	r8, r4
 	add	#wakeup_seq, r4
@@ -75,6 +85,11 @@ __pthread_cond_broadcast:
 	trapa	#0x14
 	SYSCALL_INST_PAD
 
+	mov	#0, r0
+	lds.l	@r15+, pr
+	rts
+	 mov.l	@r15+, r8
+
 4:
 	/* Unlock.  */
 #if cond_lock != 0
@@ -104,7 +119,7 @@ __pthread_cond_broadcast:
 	 nop
 
 5:
-	/* Unlock in loop requires waekup.  */
+	/* Unlock in loop requires wakeup.  */
 	mov	r8, r4
 #if cond_lock != 0
 	add	#cond_lock, r4
@@ -116,11 +131,26 @@ __pthread_cond_broadcast:
 	bra	6b
 	 nop
 
+7:
+	/* Unlock in loop requires wakeup.  */
+	mov	r8, r4
+#if cond_lock != 0
+	add	#cond_lock, r4
+#endif
+	mov.l	.Lmwake6, r1
+	bsrf	r1
+	 nop
+.Lmwake6b:
+	bra	8b
+	 nop
+
 	.align	2
 .Lmwait5:
 	.long	__lll_mutex_lock_wait-.Lmwait5b
 .Lmwake5:
 	.long	__lll_mutex_unlock_wake-.Lmwake5b
+.Lmwake6:
+	.long	__lll_mutex_unlock_wake-.Lmwake6b
 	.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_timedwait.S b/nptl/sysdeps/unix/sysv/linux/sh/pthread_cond_timedwait.S
index 59aeff70fb..6ad987d005 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/pthread_cond_timedwait.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/pthread_cond_timedwait.S
@@ -41,7 +41,7 @@ __pthread_cond_timedwait:
 	mov.l	r9, @-r15
 	mov.l	r8, @-r15
 	sts.l	pr, @-r15
-	add	#-48, r15
+	add	#-64, r15
 	mov	r4, r8
 	mov	r5, r9
 	mov	r6, r10
@@ -64,6 +64,9 @@ __pthread_cond_timedwait:
 	 mov	r9, r4
 .Lmunlock1b:
 
+	tst	r0, r0
+	bf	16f
+
 	mov	#1, r2
 	mov	#0, r3
 
@@ -86,18 +89,21 @@ __pthread_cond_timedwait:
 	mov.l	.Lccleanup1, r5
 #endif
 	mov	r15, r4
-	add	#28, r4
+	add	#36, r4
 
 	mov.l	.Lccpush1, r1
 	bsrf	r1
-	 mov	r8, r6
+	 mov	r15, r6
 .Lccpush1b:
 
 	/* Get and store current wakeup_seq value.  */
 	mov.l	@(wakeup_seq,r8), r0
 	mov.l	@(wakeup_seq+4,r8), r1
-	mov.l	r0, @(12,r15)
-	mov.l	r1, @(16,r15)
+	mov.l	r0, @(20,r15)
+	mov.l	r1, @(24,r15)
+	/* Prepare structure passed to cancellation handler.  */
+	mov.l	r9, @r15
+	mov.l	r8, @(4,r15)
 
 	/* Unlock.  */
 8:
@@ -111,27 +117,28 @@ __pthread_cond_timedwait:
 	bra	3f
 	 nop
 4:
+	mov	r15, r4
 	mov.l	.Lenable1, r1
 	bsrf	r1
-	 nop
+	 add	#8, r4
+
 .Lenable1b:
-	mov.l	r0, @r15
 
 	/* Get current time.  */
 	mov	r15, r4
-	add	#4, r4
+	add	#12, r4
 	mov	#0, r5
 	mov	#SYS_gettimeofday, r3
 	trapa	#0x12
 	SYSCALL_INST_PAD
 
 	/* Compute relative timeout.  */
-	mov.l	@(8,r15), r0
+	mov.l	@(16,r15), r0
 	mov.w	.L1k, r1
 	dmulu.l	r0, r1		/* Milli seconds to nano seconds.  */
 	mov.l	@r10, r2
 	mov.l	@(4,r10), r3
-	mov.l	@(4,r15), r0
+	mov.l	@(12,r15), r0
 	sts	macl, r1
 	sub	r0, r2
 	clrt
@@ -145,24 +152,24 @@ __pthread_cond_timedwait:
 	bf	13f		/* Time is already up.  */
 
 	/* Store relative timeout.  */
-	mov.l	r2, @(4,r15)
-	mov.l	r3, @(8,r15)
+	mov.l	r2, @(12,r15)
+	mov.l	r3, @(16,r15)
 
 	mov	r15, r7
-	add	#4, r7
+	add	#12, r7
 	mov	#FUTEX_WAIT, r5
-	mov.l	@(12,r15), r6
+	mov.l	@(20,r15), r6
 	mov	r8, r4
 	add	#wakeup_seq, r4
 	mov	#SYS_futex, r3
 	extu.b	r3, r3
 	trapa	#0x14
 	SYSCALL_INST_PAD
-	mov.l	r0, @(20,r15)
+	mov.l	r0, @(28,r15)
 
 	mov.l	.Ldisable1, r1
 	bsrf	r1
-	 mov.l	@r15, r4
+	 mov.l	@(8,r15), r4
 .Ldisable1b:
 
 	/* Lock.  */
@@ -181,14 +188,14 @@ __pthread_cond_timedwait:
 	mov.l	@(wakeup_seq,r8), r2
 	mov.l	@(wakeup_seq+4,r8), r3
 
-	mov.l	@(16,r15), r5
-	cmp/hi	r5, r1
+	mov.l	@(24,r15), r5
+	cmp/hi	r5, r3
 	bt	7f
-	cmp/hi	r1, r5
+	cmp/hi	r3, r5
 	bt	15f
 
-	mov.l	@(12,r15), r5
-	cmp/hi	r0, r5
+	mov.l	@(20,r15), r5
+	cmp/hs	r2, r5
 	bt	15f
 7:
 	cmp/hi	r1, r3
@@ -198,7 +205,7 @@ __pthread_cond_timedwait:
 	cmp/hi	r0, r2
 	bt	9f
 15:
-	mov.l	@(20,r15),r0
+	mov.l	@(28,r15),r0
 	cmp/eq	#-ETIMEDOUT, r0
 	bf	8b
 13:
@@ -214,11 +221,11 @@ __pthread_cond_timedwait:
 	mov.l	r1,@(wakeup_seq+4,r8)
 	mov	#ETIMEDOUT, r0
 	bra	14f
-	 mov.l	r0, @(24,r15)
+	 mov.l	r0, @(32,r15)
 
 9:
 	mov	#0, r0
-	mov.l	r0, @(24,r15)
+	mov.l	r0, @(32,r15)
 14:
 	mov	#1, r2
 	mov	#0, r3
@@ -242,7 +249,7 @@ __pthread_cond_timedwait:
 11:
 	/* Remove cancellation handler.  */
 	mov	r15, r4
-	add	#28, r4
+	add	#36, r4
 	mov.l	.Lcpop1, r1
 	bsrf	r1
 	 mov	#0, r5
@@ -254,11 +261,13 @@ __pthread_cond_timedwait:
 	 mov	#0, r5
 .Lmlocki1b:
 
-	mov.l	@(24,r15), r0
+	/* We return the result of the mutex_lock operation if it failed.  */
+	tst	r0, r0
+	bf	18f
+	mov.l	@(32,r15), r0
 
-	add	#48, r15
-
-	/* We return the result of the mutex_lock operation.  */
+18:	
+	add	#64, r15
 	lds.l	@r15+, pr
 	mov.l	@r15+, r8
 	mov.l	@r15+, r9
@@ -284,7 +293,7 @@ __pthread_cond_timedwait:
 .Lccpush1:
 	.long	__pthread_cleanup_push-.Lccpush1b
 .Lenable1:
-	.long	__pthread_enable_asynccancel-.Lenable1b
+	.long	__pthread_enable_asynccancel_2-.Lenable1b
 .Ldisable1:
 	.long	__pthread_disable_asynccancel-.Ldisable1b
 .Lcpop1:
@@ -346,6 +355,29 @@ __pthread_cond_timedwait:
 	bra	11b
 	 nop
 
+16:
+	/* The initial unlocking of the mutex failed.  */
+	mov.l	r0, @(32,r15)
+#if cond_lock != 0
+	DEC (@(cond_lock,r8), r2)
+#else
+	DEC (@r8, r2)
+#endif
+	tst	r2, r2
+	bf	17f
+
+	mov	r8, r4
+#if cond_lock != 0
+	add	#cond_lock, r4
+#endif
+	mov.l	.Lmwake4, r1
+	bsrf	r1
+	 nop
+.Lmwake4b:
+17:
+	bra	18b
+	 mov.l	@(32,r15), r0
+
 	.align	2
 .Lmwait2:
 	.long	__lll_mutex_lock_wait-.Lmwait2b
@@ -355,6 +387,8 @@ __pthread_cond_timedwait:
 	.long	__lll_mutex_lock_wait-.Lmwait3b
 .Lmwake3:
 	.long	__lll_mutex_unlock_wake-.Lmwake3b
+.Lmwake4:
+	.long	__lll_mutex_unlock_wake-.Lmwake4b
 	.size	__pthread_cond_timedwait, .-__pthread_cond_timedwait
 versioned_symbol (libpthread, __pthread_cond_timedwait, pthread_cond_timedwait,
 		  GLIBC_2_3_2)
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 342e78ed6e..a32df21f29 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/pthread_cond_wait.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/pthread_cond_wait.S
@@ -34,8 +34,10 @@
 	.hidden	__condvar_cleanup
 __condvar_cleanup:
 	mov.l	r8, @-r15
+	mov.l	r9, @-r15
 	sts.l	pr, @-r15
-	mov	r4, r8
+	mov	r4, r9
+	mov.l	@(4,r9), r8
 
 	/* Get internal lock.  */
 	mov	#1, r3
@@ -93,7 +95,33 @@ __condvar_cleanup:
 	 nop
 .Lwake0b:
 2:
+
+	/* Wake up all waiters to make sure no signal gets lost.  */
+	mov	r8, r4
+	add	#wakeup_seq, r4
+	mov	#FUTEX_WAKE, r5
+	mov	#-1, r6
+	shlr	r6		/* r6 = 0x7fffffff */
+	mov	#0, r7
+	mov	#SYS_futex, r3
+	extu.b	r3, r3
+	trapa	#0x14
+	SYSCALL_INST_PAD
+
+	/* Lock the mutex unless asynchronous cancellation is in effect.  */
+	mov.l	@(8,r9), r0
+	and	#2, r0
+	tst	r0, r0
+	bf	3f
+
+	mov.l	.Lmlocki1, r1
+	bsrf	r1
+	 mov.l	@r9, r4
+.Lmlocki1b:
+
+3:
 	lds.l	@r15+, pr
+	mov.l	@r15+, r9
 	rts
 	 mov.l	@r15+, r8
 
@@ -102,6 +130,8 @@ __condvar_cleanup:
 	.long	__lll_mutex_lock_wait-.Lwait0b
 .Lwake0:
 	.long	__lll_mutex_unlock_wake-.Lwake0b
+.Lmlocki1:
+	.long	__pthread_mutex_lock_internal-.Lmlocki1b
 	.size	__condvar_cleanup, .-__condvar_cleanup
 
 
@@ -114,7 +144,7 @@ __pthread_cond_wait:
 	mov.l	r9, @-r15
 	mov.l	r8, @-r15
 	sts.l	pr, @-r15
-	add	#-32, r15
+	add	#-48, r15
 	mov	r4, r8
 	mov	r5, r9
 
@@ -134,6 +164,9 @@ __pthread_cond_wait:
 	 mov	r9, r4
 .Lmunlock0b:
 
+	tst	r0, r0
+	bf	12f
+
 	mov	#1, r2
 	mov	#0, r3
 
@@ -156,18 +189,21 @@ __pthread_cond_wait:
 	mov.l	.Lccleanup0, r5
 #endif
 	mov	r15, r4
-	add	#12, r4
+	add	#20, r4
 
 	mov.l	.Lccpush0, r1
 	bsrf	r1
-	 mov	r8, r6
+	 mov	r15, r6
 .Lccpush0b:
 
 	/* Get and store current wakeup_seq value.  */
 	mov.l	@(wakeup_seq,r8), r0
 	mov.l	@(wakeup_seq+4,r8), r1
-	mov.l	r0, @(4,r15)
-	mov.l	r1, @(8,r15)
+	mov.l	r0, @(12,r15)
+	mov.l	r1, @(16,r15)
+	/* Prepare structure passed to cancellation handler.  */
+	mov.l	r9, @r15
+	mov.l	r8, @(4,r15)
 
 8:
 	/* Unlock.  */
@@ -179,15 +215,15 @@ __pthread_cond_wait:
 	tst	r2, r2
 	bf	3f
 4:
+	mov	r15, r4
 	mov.l	.Lenable0, r1
 	bsrf	r1
-	 nop
+	 add	#8, r4
 .Lenable0b:
-	mov.l	r0, @r15
 
 	mov	#0, r7
 	mov	#FUTEX_WAIT, r5
-	mov.l	@(4,r15), r6
+	mov.l	@(12,r15), r6
 	mov	r8, r4
 	add	#wakeup_seq, r4
 	mov	#SYS_futex, r3
@@ -197,7 +233,7 @@ __pthread_cond_wait:
 
 	mov.l	.Ldisable0, r1
 	bsrf	r1
-	 mov.l	@r15, r4
+	 mov.l	@(8,r15), r4
 .Ldisable0b:	
 
 	/* Lock.  */
@@ -216,14 +252,14 @@ __pthread_cond_wait:
 	mov.l	@(wakeup_seq,r8), r2
 	mov.l	@(wakeup_seq+4,r8), r3
 
-	mov.l	@(8,r15), r5
-	cmp/hi	r5, r1
+	mov.l	@(16,r15), r5
+	cmp/hi	r5, r3
 	bt	7f
-	cmp/hi	r1, r5
+	cmp/hi	r3, r5
 	bt	8b
 
-	mov.l	@(4,r15), r5
-	cmp/hi	r0, r5
+	mov.l	@(12,r15), r5
+	cmp/hs	r2, r5
 	bt	8b
 7:
 	cmp/hi	r1, r3
@@ -255,7 +291,7 @@ __pthread_cond_wait:
 11:
 	/* Remove cancellation handler.  */
 	mov	r15, r4
-	add	#12, r4
+	add	#20, r4
 	mov.l	.Lcpop0, r1
 	bsrf	r1
 	 mov	#0, r5
@@ -266,10 +302,10 @@ __pthread_cond_wait:
 	bsrf	r1
 	 mov	#0, r5
 .Lmlocki0b:
-
-	add	#32, r15
-
 	/* We return the result of the mutex_lock operation.  */
+
+14:
+	add	#48, r15
 	lds.l	@r15+, pr
 	mov.l	@r15+, r8
 	mov.l	@r15+, r9
@@ -291,7 +327,7 @@ __pthread_cond_wait:
 .Lccpush0:
 	.long	__pthread_cleanup_push-.Lccpush0b
 .Lenable0:
-	.long	__pthread_enable_asynccancel-.Lenable0b
+	.long	__pthread_enable_asynccancel_2-.Lenable0b
 .Ldisable0:
 	.long	__pthread_disable_asynccancel-.Ldisable0b
 .Lcpop0:
@@ -350,6 +386,30 @@ __pthread_cond_wait:
 	bra	11b
 	 nop
 
+12:
+	/* The initial unlocking of the mutex failed.  */
+	mov.l	r0, @-r15
+#if cond_lock != 0
+	DEC (@(cond_lock,r8), r2)
+#else
+	DEC (@r8, r2)
+#endif
+	tst	r2, r2
+	bf	13f
+
+	mov	r8, r4
+#if cond_lock != 0
+	add	#cond_lock, r4
+#endif
+	mov.l	.Lmwake2, r1
+	bsrf	r1
+	 nop
+.Lmwake2b:
+
+13:
+	bra	14b
+	 mov.l	@r15+, r0
+
 	.align	2
 .Lmwait0:
 	.long	__lll_mutex_lock_wait-.Lmwait0b
@@ -359,6 +419,8 @@ __pthread_cond_wait:
 	.long	__lll_mutex_lock_wait-.Lmwait1b
 .Lmwake1:
 	.long	__lll_mutex_unlock_wake-.Lmwake1b
+.Lmwake2:
+	.long	__lll_mutex_unlock_wake-.Lmwake2b
 	.size	__pthread_cond_wait, .-__pthread_cond_wait
 versioned_symbol (libpthread, __pthread_cond_wait, pthread_cond_wait,
 		  GLIBC_2_3_2)
diff --git a/nptl/sysdeps/unix/sysv/linux/sh/pthread_once.S b/nptl/sysdeps/unix/sysv/linux/sh/pthread_once.S
index c17701cffc..c292af02c0 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/pthread_once.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/pthread_once.S
@@ -122,7 +122,7 @@ __pthread_once:
 #endif
 	mov.l	.Lcpush, r1
 	bsrf	r1
-	 mov	r3, r6
+	 mov	r9, r6
 .Lcpush0:
 	jsr	@r8
 	 nop
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 1aa1c72bbf..27aecad314 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_rdlock.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_rdlock.S
@@ -75,7 +75,8 @@ __pthread_rwlock_rdlock:
 	mov	r8, r4
 	add	#READERS_WAKEUP, r4
 	mov	#FUTEX_WAIT, r5
-	mov	#0, r6
+	mov.l	@(READERS_WAKEUP,r8), r6
+	mov	#0, r7
 	mov	#SYS_futex, r3
 	extu.b	r3, r3
 	trapa	#0x14
@@ -93,11 +94,8 @@ __pthread_rwlock_rdlock:
 13:
 	mov.l	@(READERS_QUEUED,r8), r0
 	add	#-1, r0
-	mov.l	r0, @(READERS_QUEUED,r8)
-	tst	r0, r0
-	bf	2b
 	bra	2b
-	 mov.l	r0, @(READERS_WAKEUP,r8)
+	 mov.l	r0, @(READERS_QUEUED,r8)
 
 5:
 	mov	#0, r3
@@ -138,7 +136,6 @@ __pthread_rwlock_rdlock:
 	stc	gbr, r1
 	mov.w	.Ltcboff,r2
 	sub	r2,r1
-	mov.l	@(8,r1),r1
 	cmp/eq	r1, r0
 	bf	3b
 	/* Deadlock detected.  */
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 84b5cdda3b..6f3f67e07b 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_timedrdlock.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_timedrdlock.S
@@ -62,6 +62,12 @@ pthread_rwlock_timedrdlock:
 	tst	r0, r0
 	bt	5f
 3:
+	/* Check the value of the timeout parameter.  */
+	mov.l	.L1g0, r1
+	mov.l	@(4,r9), r0
+	cmp/hs	r1, r0
+	bt	19f
+
 	mov.l	@(READERS_QUEUED,r8), r0
 	add	#1, r0
 	mov.l	r0, @(READERS_QUEUED,r8)
@@ -109,7 +115,7 @@ pthread_rwlock_timedrdlock:
 	/* Futex call.  */
 	mov	r15, r7
 	mov	#FUTEX_WAIT, r5
-	mov	#0, r6
+	mov.l	@(READERS_WAKEUP,r8), r6
 	mov	r8, r4
 	add	#READERS_WAKEUP, r4
 	mov	#SYS_futex, r3
@@ -131,16 +137,16 @@ pthread_rwlock_timedrdlock:
 	bf	12f
 
 13:
-	mov	#-ETIMEDOUT, r0
-	cmp/eq	r0, r3
-	bt	18f
 	mov.l	@(READERS_QUEUED,r8), r0
 	add	#-1, r0
 	mov.l	r0, @(READERS_QUEUED,r8)
-	tst	r0, r0
+	mov	#-ETIMEDOUT, r0
+	cmp/eq	r0, r3
 	bf	2b
-	bra	2b
-	 mov.l	r0, @(READERS_WAKEUP,r8)
+
+18:
+	bra	9f
+	 mov	#ETIMEDOUT, r3
 
 5:
 	mov	#0, r3
@@ -189,7 +195,6 @@ pthread_rwlock_timedrdlock:
 	stc	gbr, r1
 	mov.w	.Ltcboff,r2
 	sub	r2,r1
-	mov.l	@(8,r1),r1
 	cmp/eq	r1, r0
 	bf	3b
 	/* Deadlock detected.  */
@@ -253,9 +258,9 @@ pthread_rwlock_timedrdlock:
 	bra	17b
 	 mov	#-ETIMEDOUT, r3
 
-18:
+19:
 	bra	9b
-	 mov	#ETIMEDOUT, r3
+	 mov	#EINVAL, r3
 
 .Ltcboff:
 	.word	TLS_PRE_TCB_SIZE
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 60ff48f90a..89bdc76d37 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_timedwrlock.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_timedwrlock.S
@@ -59,6 +59,12 @@ pthread_rwlock_timedwrlock:
 	tst	r0, r0
 	bt	5f
 3:
+	/* Check the value of the timeout parameter.  */
+	mov.l	.L1g1, r1
+	mov.l	@(4,r9), r0
+	cmp/hs	r1, r0
+	bt	19f
+
 	mov.l	@(WRITERS_QUEUED,r8), r0
 	add	#1, r0
 	mov.l	r0, @(WRITERS_QUEUED,r8)
@@ -106,7 +112,7 @@ pthread_rwlock_timedwrlock:
 	/* Futex call.  */
 	mov	r15, r7
 	mov	#FUTEX_WAIT, r5
-	mov	#0, r6
+	mov.l	@(WRITERS_WAKEUP,r8), r6
 	mov	r8, r4
 	add	#WRITERS_WAKEUP, r4
 	mov	#SYS_futex, r3
@@ -128,22 +134,26 @@ pthread_rwlock_timedwrlock:
 	bf	12f
 
 13:
-	mov	#-ETIMEDOUT, r0
-	cmp/eq	r0, r3
-	bt	18f
 	mov.l	@(WRITERS_QUEUED,r8), r0
 	add	#-1, r0
 	mov.l	r0, @(WRITERS_QUEUED,r8)
-	mov	#0, r0
-	bra	2b
-	 mov.l	r0, @(WRITERS_WAKEUP,r8)
+	mov	#-ETIMEDOUT, r0
+	cmp/eq	r0, r3
+	bf	2b
+
+18:
+	bra	9f
+	 mov	#ETIMEDOUT, r3
+
+19:
+	bra	9f
+	 mov	#EINVAL, r3
 
 5:
 	mov	#0, r3
-	stc	gbr, r1
+	stc	gbr, r0
 	mov.w	.Ltcboff,r2
-	sub	r2,r1
-	mov.l	@(8,r1),r0
+	sub	r2,r0
 	mov.l	r0, @(WRITER,r8)
 9:
 #if MUTEX == 0
@@ -184,7 +194,6 @@ pthread_rwlock_timedwrlock:
 	stc	gbr, r1
 	mov.w	.Ltcboff,r2
 	sub	r2,r1
-	mov.l	@(8,r1),r1
 	cmp/eq	r1, r0
 	bf	3b
 	bra	9b
@@ -238,10 +247,6 @@ pthread_rwlock_timedwrlock:
 	bra	17b
 	 mov	#-ETIMEDOUT, r3
 
-18:
-	bra	9b
-	 mov	#ETIMEDOUT, r3
-
 .Ltcboff:
 	.word	TLS_PRE_TCB_SIZE
 	.align	2
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 af046aad47..8303a984ae 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_unlock.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_unlock.S
@@ -57,24 +57,48 @@ __pthread_rwlock_unlock:
 5:
 	mov	#0, r0
 	mov.l	r0, @(WRITER,r8)
+	mov	#1, r6
 	mov	r8, r4
+	add	#WRITERS_WAKEUP, r4
 	mov.l	@(WRITERS_QUEUED,r8), r0
 	tst	r0, r0
-	bf	9f
+	bf	0f
+
+	/* If also no readers waiting nothing to do.  */
+	mov.l	@(READERS_QUEUED,r8), r0
+	tst	r0, r0
+	bt	6f
+
 	mov	#-1, r6
 	shlr	r6		/* r6 = 0x7fffffff */
-	bra	0f
-	 add	#READERS_WAKEUP, r4
-9:
-	mov	#1, r6
-	add	#WRITERS_WAKEUP, r4
+	mov	r8, r4
+	add	#READERS_WAKEUP, r4
+
 0:
+	mov.l	@r4, r0
+	add	#1, r0
+	mov.l	r0, @r4
+#if MUTEX == 0
+	DEC (@r8, r2)
+#else
+	DEC (@(MUTEX,r8), r2)
+#endif
+	tst	r2, r2
+	bf	7f
+
+8:
 	mov	#FUTEX_WAKE, r5
 	mov	#SYS_futex, r3
+	mov	#0, r7
 	extu.b	r3, r3
 	trapa	#0x14
 	SYSCALL_INST_PAD
 
+	lds.l	@r15+, pr
+	mov.l	@r15+, r8
+	mov.l	@r15+, r12
+	rts
+	 mov	#0, r0
 6:
 #if MUTEX == 0
 	DEC (@r8, r2)
@@ -114,11 +138,29 @@ __pthread_rwlock_unlock:
 	bra	4b
 	 nop
 
+7:
+	mov.l	r4, @-r15
+	mov.l	r6, @-r15
+	mov	r8, r4
+#if MUTEX != 0
+	add	#MUTEX, r4
+#endif
+	mov.l	.Lwake9, r1
+	bsrf	r1
+	 nop
+.Lwake9b:
+
+	mov.l	@r15+, r6
+	bra	8b
+	 mov.l	@r15+, r4
+
 	.align	2
 .Lwait8:	
 	.long	__lll_mutex_lock_wait-.Lwait8b
 .Lwake8:
 	.long	__lll_mutex_unlock_wake-.Lwake8b
+.Lwake9:
+	.long	__lll_mutex_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 06d56e04cb..1456292a73 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_wrlock.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/pthread_rwlock_wrlock.S
@@ -72,7 +72,8 @@ __pthread_rwlock_wrlock:
 	mov	r8, r4
 	add	#WRITERS_WAKEUP, r4
 	mov	#FUTEX_WAIT, r5
-	mov	#0, r6
+	mov.l	@(WRITERS_WAKEUP,r8), r6
+	mov	#0, r7
 	mov	#SYS_futex, r3
 	extu.b	r3, r3
 	trapa	#0x14
@@ -88,19 +89,16 @@ __pthread_rwlock_wrlock:
 	tst	r2, r2
 	bf	12f
 13:
-	mov.l	@(READERS_QUEUED,r8), r0
+	mov.l	@(WRITERS_QUEUED,r8), r0
 	add	#-1, r0
-	mov.l	r0, @(READERS_QUEUED,r8)
-	mov	#0, r0
 	bra	2b
-	 mov.l	r0, @(WRITERS_WAKEUP,r8)
+	 mov.l	r0, @(WRITERS_QUEUED,r8)
 
 5:
 	mov	#0, r3
-	stc	gbr, r1
+	stc	gbr, r0
 	mov.w	.Ltcboff,r2
-	sub	r2,r1
-	mov.l	@(8,r1),r0
+	sub	r2,r0
 	mov.l	r0, @(WRITER,r8)
 9:
 #if MUTEX == 0
@@ -133,7 +131,6 @@ __pthread_rwlock_wrlock:
 	stc	gbr, r1
 	mov.w	.Ltcboff,r2
 	sub	r2,r1
-	mov.l	@(8,r1),r1
 	cmp/eq	r1, r0
 	bf	3b
 	bra	9b
diff --git a/nptl/sysdeps/unix/sysv/linux/sh/sem_post.S b/nptl/sysdeps/unix/sysv/linux/sh/sem_post.S
index 90e63ed184..a080e17abe 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/sem_post.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/sem_post.S
@@ -51,7 +51,7 @@ __new_sem_post:
 	 mov	#0, r0
 
 1:
-	mov	#EAGAIN, r2
+	mov	#EINVAL, r2
 	mova	.Lgot3, r0
 	mov.l	.Lgot3, r12
 	add	r0, r12
@@ -84,4 +84,4 @@ __new_sem_post:
 	.long	__errno_location@PLT-(.Lerrloc3b+2-.)
 #endif
 	.size	__new_sem_post,.-__new_sem_post
-	.symver	__new_sem_post, sem_post@@GLIBC_2.2
+	versioned_symbol(libpthread, __new_sem_post, sem_post, GLIBC_2_1)
diff --git a/nptl/sysdeps/unix/sysv/linux/sh/sem_trywait.S b/nptl/sysdeps/unix/sysv/linux/sh/sem_trywait.S
index c80839ef30..fc505b4af5 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/sem_trywait.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/sem_trywait.S
@@ -91,4 +91,4 @@ __new_sem_trywait:
 	.long	__errno_location@PLT-(.Lerrloc1b+2-.)
 #endif
 	.size	__new_sem_trywait,.-__new_sem_trywait
-	.symver	__new_sem_trywait, sem_trywait@@GLIBC_2.2
+	versioned_symbol(libpthread, __new_sem_trywait, sem_trywait, GLIBC_2_1)
diff --git a/nptl/sysdeps/unix/sysv/linux/sh/sem_wait.S b/nptl/sysdeps/unix/sysv/linux/sh/sem_wait.S
index f7bca99fc3..49b76c6d8c 100644
--- a/nptl/sysdeps/unix/sysv/linux/sh/sem_wait.S
+++ b/nptl/sysdeps/unix/sysv/linux/sh/sem_wait.S
@@ -104,4 +104,4 @@ __new_sem_wait:
 	.long	__errno_location@PLT-(.Lerrloc0b+2-.)
 #endif
 	.size	__new_sem_wait,.-__new_sem_wait
-	.symver	__new_sem_wait, sem_wait@@GLIBC_2.2
+	versioned_symbol(libpthread, __new_sem_wait, sem_wait, GLIBC_2_1)
diff --git a/sysdeps/sh/bits/atomic.h b/sysdeps/sh/bits/atomic.h
index de176ca2c5..d3725fc66b 100644
--- a/sysdeps/sh/bits/atomic.h
+++ b/sysdeps/sh/bits/atomic.h
@@ -45,61 +45,55 @@ typedef intmax_t atomic_max_t;
 typedef uintmax_t uatomic_max_t;
 
 
-#define __arch_compare_and_exchange_8_acq(mem, newval, oldval) \
-  ({ unsigned char __result; \
+#define __arch_compare_and_exchange_val_8_acq(mem, newval, oldval) \
+  ({ __typeof (*mem) __result; \
      __asm __volatile ("\
 	.align 2\n\
 	mova 1f,r0\n\
 	nop\n\
 	mov r15,r1\n\
 	mov #-8,r15\n\
-     0: mov.b @%1,r2\n\
-	cmp/eq r2,%3\n\
+     0: mov.b @%1,%0\n\
+	cmp/eq %0,%3\n\
 	bf 1f\n\
 	mov.b %2,@%1\n\
-     1: mov r1,r15\n\
-	mov #-1,%0\n\
-	negc %0,%0"\
-	: "=r" (__result) : "r" (mem), "r" (newval), "r" (oldval) \
-	: "r0", "r1", "r2", "t", "memory"); \
+     1: mov r1,r15"\
+	: "=&r" (__result) : "r" (mem), "r" (newval), "r" (oldval) \
+	: "r0", "r1", "t", "memory"); \
      __result; })
 
-#define __arch_compare_and_exchange_16_acq(mem, newval, oldval) \
-  ({ unsigned char __result; \
+#define __arch_compare_and_exchange_val_16_acq(mem, newval, oldval) \
+  ({ __typeof (*mem) __result; \
      __asm __volatile ("\
 	.align 2\n\
 	mova 1f,r0\n\
 	nop\n\
 	mov r15,r1\n\
 	mov #-8,r15\n\
-     0: mov.w @%1,r2\n\
-	cmp/eq r2,%3\n\
+     0: mov.w @%1,%0\n\
+	cmp/eq %0,%3\n\
 	bf 1f\n\
 	mov.w %2,@%1\n\
-     1: mov r1,r15\n\
-	mov #-1,%0\n\
-	negc %0,%0"\
-	: "=r" (__result) : "r" (mem), "r" (newval), "r" (oldval) \
-	: "r0", "r1", "r2", "t", "memory"); \
+     1: mov r1,r15"\
+	: "=&r" (__result) : "r" (mem), "r" (newval), "r" (oldval) \
+	: "r0", "r1", "t", "memory"); \
      __result; })
 
-#define __arch_compare_and_exchange_32_acq(mem, newval, oldval) \
-  ({ unsigned char __result; \
+#define __arch_compare_and_exchange_val_32_acq(mem, newval, oldval) \
+  ({ __typeof (*mem) __result; \
      __asm __volatile ("\
 	.align 2\n\
 	mova 1f,r0\n\
 	nop\n\
 	mov r15,r1\n\
 	mov #-8,r15\n\
-     0: mov.l @%1,r2\n\
-	cmp/eq r2,%3\n\
+     0: mov.l @%1,%0\n\
+	cmp/eq %0,%3\n\
 	bf 1f\n\
 	mov.l %2,@%1\n\
-     1: mov r1,r15\n\
-	mov #-1,%0\n\
-	negc %0,%0"\
-	: "=r" (__result) : "r" (mem), "r" (newval), "r" (oldval) \
-	: "r0", "r1", "r2", "t", "memory"); \
+     1: mov r1,r15"\
+	: "=&r" (__result) : "r" (mem), "r" (newval), "r" (oldval) \
+	: "r0", "r1", "t", "memory"); \
      __result; })
 
 /* XXX We do not really need 64-bit compare-and-exchange.  At least
@@ -107,11 +101,12 @@ typedef uintmax_t uatomic_max_t;
    problems since not many other 32-bit architectures have support for
    such an operation.  So don't define any code for now.  */
 
-# define __arch_compare_and_exchange_64_acq(mem, newval, oldval) \
+# define __arch_compare_and_exchange_val_64_acq(mem, newval, oldval) \
   (abort (), 0)
 
 #define atomic_exchange_and_add(mem, value) \
   ({ __typeof (*mem) __result; \
+     __typeof (value) __value; \
      if (sizeof (*mem) == 1) \
        __asm __volatile ("\
 	  .align 2\n\
@@ -122,7 +117,7 @@ typedef uintmax_t uatomic_max_t;
 	  add %0,%1\n\
 	  mov.b %1,@%2\n\
        1: mov r1,r15"\
-	: "=&r" (__result), "=&r" (value) : "r" (mem), "1" (value) \
+	: "=&r" (__result), "=&r" (__value) : "r" (mem), "1" (value) \
 	: "r0", "r1", "memory"); \
      else if (sizeof (*mem) == 2) \
        __asm __volatile ("\
@@ -134,7 +129,7 @@ typedef uintmax_t uatomic_max_t;
 	  add %0,%1\n\
 	  mov.w %1,@%2\n\
        1: mov r1,r15"\
-	: "=&r" (__result), "=&r" (value) : "r" (mem), "1" (value) \
+	: "=&r" (__result), "=&r" (__value) : "r" (mem), "1" (value) \
 	: "r0", "r1", "memory"); \
      else if (sizeof (*mem) == 4) \
        __asm __volatile ("\
@@ -146,22 +141,23 @@ typedef uintmax_t uatomic_max_t;
 	  add %0,%1\n\
 	  mov.l %1,@%2\n\
        1: mov r1,r15"\
-	: "=&r" (__result), "=&r" (value) : "r" (mem), "1" (value) \
+	: "=&r" (__result), "=&r" (__value) : "r" (mem), "1" (value) \
 	: "r0", "r1", "memory"); \
      else \
        { \
 	 __typeof (value) addval = (value); \
-	 __typeof (*mem) oldval; \
 	 __typeof (mem) memp = (mem); \
 	 do \
-	   __result = (oldval = *memp) + addval; \
-	 while (!__arch_compare_and_exchange_64_acq (memp, __result, oldval));\
+	   __result = *memp; \
+	 while (__arch_compare_and_exchange_val_64_acq \
+		 (memp,	__result + addval, __result) == __result); \
 	 (void) addval; \
        } \
      __result; })
 
 #define atomic_add(mem, value) \
-  (void) ({ if (sizeof (*mem) == 1) \
+  (void) ({ __typeof (value) __value; \
+	    if (sizeof (*mem) == 1) \
 	      __asm __volatile ("\
 		.align 2\n\
 		mova 1f,r0\n\
@@ -171,7 +167,7 @@ typedef uintmax_t uatomic_max_t;
 		add r2,%0\n\
 		mov.b %0,@%1\n\
 	     1: mov r1,r15"\
-		: "=&r" (value) : "r" (mem), "0" (value) \
+		: "=&r" (__value) : "r" (mem), "0" (value) \
 		: "r0", "r1", "r2", "memory"); \
 	    else if (sizeof (*mem) == 2) \
 	      __asm __volatile ("\
@@ -183,7 +179,7 @@ typedef uintmax_t uatomic_max_t;
 		add r2,%0\n\
 		mov.w %0,@%1\n\
 	     1: mov r1,r15"\
-		: "=&r" (value) : "r" (mem), "0" (value) \
+		: "=&r" (__value) : "r" (mem), "0" (value) \
 		: "r0", "r1", "r2", "memory"); \
 	    else if (sizeof (*mem) == 4) \
 	      __asm __volatile ("\
@@ -195,7 +191,7 @@ typedef uintmax_t uatomic_max_t;
 		add r2,%0\n\
 		mov.l %0,@%1\n\
 	     1: mov r1,r15"\
-		: "=&r" (value) : "r" (mem), "0" (value) \
+		: "=&r" (__value) : "r" (mem), "0" (value) \
 		: "r0", "r1", "r2", "memory"); \
 	    else \
 	      { \
@@ -204,15 +200,15 @@ typedef uintmax_t uatomic_max_t;
 		__typeof (mem) memp = (mem); \
 		do \
 		  oldval = *memp; \
-		while (! __arch_compare_and_exchange_64_acq (memp, \
-							     oldval + addval, \
-							     oldval)); \
+		while (__arch_compare_and_exchange_val_64_acq \
+			(memp, oldval + addval, oldval) == oldval); \
 		(void) addval; \
 	      } \
 	    })
 
 #define atomic_add_negative(mem, value) \
   ({ unsigned char __result; \
+     __typeof (value) __value; \
      if (sizeof (*mem) == 1) \
        __asm __volatile ("\
 	  .align 2\n\
@@ -225,7 +221,7 @@ typedef uintmax_t uatomic_max_t;
        1: mov r1,r15\n\
 	  shal %1\n\
 	  movt %0"\
-	: "=r" (__result), "=&r" (value) : "r" (mem), "1" (value) \
+	: "=r" (__result), "=&r" (__value) : "r" (mem), "1" (value) \
 	: "r0", "r1", "r2", "t", "memory"); \
      else if (sizeof (*mem) == 2) \
        __asm __volatile ("\
@@ -239,7 +235,7 @@ typedef uintmax_t uatomic_max_t;
        1: mov r1,r15\n\
 	  shal %1\n\
 	  movt %0"\
-	: "=r" (__result), "=&r" (value) : "r" (mem), "1" (value) \
+	: "=r" (__result), "=&r" (__value) : "r" (mem), "1" (value) \
 	: "r0", "r1", "r2", "t", "memory"); \
      else if (sizeof (*mem) == 4) \
        __asm __volatile ("\
@@ -253,7 +249,7 @@ typedef uintmax_t uatomic_max_t;
        1: mov r1,r15\n\
 	  shal %1\n\
 	  movt %0"\
-	: "=r" (__result), "=&r" (value) : "r" (mem), "1" (value) \
+	: "=r" (__result), "=&r" (__value) : "r" (mem), "1" (value) \
 	: "r0", "r1", "r2", "t", "memory"); \
      else \
        abort (); \
@@ -261,6 +257,7 @@ typedef uintmax_t uatomic_max_t;
 
 #define atomic_add_zero(mem, value) \
   ({ unsigned char __result; \
+     __typeof (value) __value; \
      if (sizeof (*mem) == 1) \
        __asm __volatile ("\
 	  .align 2\n\
@@ -273,7 +270,7 @@ typedef uintmax_t uatomic_max_t;
        1: mov r1,r15\n\
 	  tst %1,%1\n\
 	  movt %0"\
-	: "=r" (__result), "=&r" (value) : "r" (mem), "1" (value) \
+	: "=r" (__result), "=&r" (__value) : "r" (mem), "1" (value) \
 	: "r0", "r1", "r2", "t", "memory"); \
      else if (sizeof (*mem) == 2) \
        __asm __volatile ("\
@@ -287,7 +284,7 @@ typedef uintmax_t uatomic_max_t;
        1: mov r1,r15\n\
 	  tst %1,%1\n\
 	  movt %0"\
-	: "=r" (__result), "=&r" (value) : "r" (mem), "1" (value) \
+	: "=r" (__result), "=&r" (__value) : "r" (mem), "1" (value) \
 	: "r0", "r1", "r2", "t", "memory"); \
      else if (sizeof (*mem) == 4) \
        __asm __volatile ("\
@@ -301,7 +298,7 @@ typedef uintmax_t uatomic_max_t;
        1: mov r1,r15\n\
 	  tst %1,%1\n\
 	  movt %0"\
-	: "=r" (__result), "=&r" (value) : "r" (mem), "1" (value) \
+	: "=r" (__result), "=&r" (__value) : "r" (mem), "1" (value) \
 	: "r0", "r1", "r2", "t", "memory"); \
      else \
        abort (); \