[AVR32] Don't use __builtin_xchg()

The implementation of __builtin_xchg() in at least some versions of
avr32 gcc is buggy. Rather than find out exactly which versions that
have this bug, let's just avoid the problem altogether by implementing
xchg() in inline assembly.

Also, in most architectures, xchg() seems to imply a memory barrier,
while the existing avr32 implementation did not. This patch also fixes
that discrepancy.
Signed-off-by: Haavard Skinnemoen <hskinnemoen@atmel.com>

include/asm-avr32/system.h

@@ -73,11 +73,16 @@ extern struct task_struct *__switch_to(struct task_struct *,
 
 extern void __xchg_called_with_bad_pointer(void);
 
-#ifdef __CHECKER__
-extern unsigned long __builtin_xchg(void *ptr, unsigned long x);
-#endif
+static inline unsigned long xchg_u32(u32 val, volatile u32 *m)
+{
+	u32 ret;
 
-#define xchg_u32(val, m) __builtin_xchg((void *)m, val)
+	asm volatile("xchg %[ret], %[m], %[val]"
+			: [ret] "=&r"(ret), "=m"(*m)
+			: "m"(*m), [m] "r"(m), [val] "r"(val)
+			: "memory");
+	return ret;
+}
 
 static inline unsigned long __xchg(unsigned long x,
 				       volatile void *ptr,