base & arm: simplify and optimize cmpxchg

The old inline assembly provided two output operands that were afterwards
combined for a return value. However, the second output operand isn't
necessary when using the "Acquiring a Lock" example-code of the ARM manual
"Barrier Litmus Tests and Cookbook". This saves two logical operations
per acquisition try. Additionally better documentation is now provided.

Fixes #1292

repos/base/include/spec/arm/cpu/atomic.h

@@ -2,6 +2,7 @@
  * \brief  Atomic operations for ARM
  * \author Norman Feske
  * \author Stefan Kalkowski
+ * \author Martin Stein
  * \date   2007-04-28
  */
 
@@ -34,25 +35,33 @@ namespace Genode {
 	 */
 	inline int cmpxchg(volatile int *dest, int cmp_val, int new_val)
 	{
-		unsigned long equal, not_exclusive;
+		int result;
+		asm volatile (
 
-		__asm__ __volatile__(
-			"@ cmpxchg\n"
-			"  1:                  \n"
-			"  ldrex   %0, [%2]    \n"
-			"  cmp     %0, %3      \n"
-			"  bne     2f          \n"
-			"  strexeq %0, %4, [%2]\n"
-			"  teqeq   %0, #0      \n"
-			"  bne     1b          \n"
-			"  moveq   %1, #1      \n"
-			"  2:                  \n"
-			"  movne   %1, #0      \n"
-			: "=&r" (not_exclusive), "=&r" (equal)
-			: "r" (dest), "r" (cmp_val), "r" (new_val)
+			/* compare values */
+			"1:\n"
+			"ldrex %0, [%1]\n"
+			"cmp %0, %2\n"
+
+			/* if not equal, return with result 0 */
+			"movne %0, #0\n"
+			"bne 2f\n"
+
+			/* if equal, try to override memory value exclusively */
+			"strex %0, %3, [%1]\n"
+			"cmp %0, #0\n"
+
+			/* if access wasn't exclusive, go back to comparison */
+			"bne 1b\n"
+
+			/* if access was exclusive, return with result 1 */
+			"mov %0, #1\n"
+			"2:\n"
+			: "=&r" (result) : "r" (dest), "r" (cmp_val), "r" (new_val)
 			: "cc");
+
 		Genode::memory_barrier();
-		return equal && !not_exclusive;
+		return result;
 	}
 }