Implement _atomic_lock in assembler to make it more readable.

plus some various assembler improvements.
Tested by various people a few months ago and in my tree for almost
6 months now.

2 files changed, 19 insertions, 37 deletions

diff --git a/lib/libpthread/arch/alpha/_atomic_lock.c b/lib/libpthread/arch/alpha/_atomic_lock.c
index 5f11655ab1f..912d7c97709 100644
--- a/lib/libpthread/arch/alpha/_atomic_lock.c
+++ b/lib/libpthread/arch/alpha/_atomic_lock.c
@@ -1,42 +1,11 @@
-/*	$OpenBSD: _atomic_lock.c,v 1.5 1999/05/26 00:11:27 d Exp $	*/
+/*	$OpenBSD: _atomic_lock.c,v 1.6 2002/06/06 15:43:04 art Exp $	*/
 /*
- * Atomic lock for alpha
+ * Atomi lock for alpha.
  */
 
 #include "spinlock.h"
 
-int
-_atomic_lock(volatile _spinlock_lock_t * lock)
-{
-	_spinlock_lock_t old;
-	_spinlock_lock_t new;
-	int success;
-
-	do {
-		/* load the value of the thread-lock (lock mem on load) */
-		__asm__( "ldq_l %0, %1" : "=r"(old) : "m"(*lock) );
-		if (old) 
-			new = old;	/* locked: no change */
-		else
-			new = _SPINLOCK_LOCKED;	/* unlocked: grab it */
-
-		success = 0;
-		/* store the new value of the thrd-lock (unlock mem on store) */
-		/*
-		 * XXX may need to add *large* branch forward for main line
-		 * branch prediction to be right :( [this note from linux]
-		 */
-		__asm__( "stq_c	%2, %0\n"
-			 "beq	%2, 1f\n"
-			 "mb\n"		
-			 "mov	1, %1\n"
-			 "1:"
-			: "=m"(*lock), "=r"(success)
-			: "r"(new) );
-	} while (!success);
-
-	return (old != _SPINLOCK_UNLOCKED);
-}
+/* _atomic lock is implemented in assembler. */
 
 int
 _atomic_is_locked(volatile _spinlock_lock_t * lock)
diff --git a/lib/libpthread/arch/alpha/uthread_machdep_asm.S b/lib/libpthread/arch/alpha/uthread_machdep_asm.S
index c702ea9cd47..ce5378d280f 100644
--- a/lib/libpthread/arch/alpha/uthread_machdep_asm.S
+++ b/lib/libpthread/arch/alpha/uthread_machdep_asm.S
@@ -1,4 +1,4 @@
-/*	$OpenBSD: uthread_machdep_asm.S,v 1.1 2000/09/25 05:44:13 d Exp $	*/
+/*	$OpenBSD: uthread_machdep_asm.S,v 1.2 2002/06/06 15:43:04 art Exp $	*/
 /* David Leonard, <d@csee.uq.edu.au>. Public domain. */
 
 #include <machine/asm.h>
@@ -8,11 +8,12 @@
 #define ALIGN(x)	(((x)+15)&~15)
 #define FRAMESIZE	ALIGN(FPOFF(8))
 
+	.set	noreorder
 	.globl	_thread_machdep_switch
 	.ent	_thread_machdep_switch, 2
 _thread_machdep_switch:
 	.frame	sp, FRAMESIZE, ra
-
+	ldgp	gp, 0(pv)
 	lda	sp, -FRAMESIZE(sp)
 
 	stq	ra, INTOFF(0)(sp)
@@ -35,6 +36,7 @@ _thread_machdep_switch:
 	stt	fs7, FPOFF(7)(sp)
 
 	stq	sp, 0(a1)
+	or	a0, zero, pv
 	ldq	sp, 0(a0)
 
 	ldt	fs7, FPOFF(7)(sp)
@@ -57,7 +59,18 @@ _thread_machdep_switch:
 	ldq	ra, INTOFF(0)(sp)
 
 	lda	sp,FRAMESIZE(sp)
-
 	RET
 
 	.end	_thread_machdep_switch
+
+LEAF(_atomic_lock,1)
+        LDGP(pv)
+
+0:      ldq_l   v0, 0(a0)       /* read existing lock value */
+        mov     1, t0           /* locked value to store */
+        stq_c   t0, 0(a0)       /* attempt to store, status in t0 */
+        beq     t0, 1f          /* branch foward to optimise prediction */
+        mb                      /* sync with other processors */
+        RET                     /* return with v0==0 if lock obtained */
+1:      br      0b              /* loop to try again */
+END(_atomic_lock)