[PATCH] ppc64: speedup cmpxchg
authorAnton Blanchard <anton@samba.org>
Tue, 6 Sep 2005 03:05:58 +0000 (13:05 +1000)
committerPaul Mackerras <paulus@samba.org>
Tue, 6 Sep 2005 06:07:53 +0000 (16:07 +1000)
cmpxchg has the following code:

__typeof__(*(ptr)) _o_ = (o);
__typeof__(*(ptr)) _n_ = (n);

Unfortunately it makes gcc 4.0 store and load the variables to the stack.
Eg in atomic_dec_and_test we get:

  stw     r10,112(r1)
  stw     r9,116(r1)
  lwz     r9,112(r1)
  lwz     r0,116(r1)

x86 is just casting the values so do that instead. Also change __xchg*
and __cmpxchg* to take unsigned values, removing a few sign extensions.

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
include/asm-ppc64/system.h

index b9e1835351e98d7874f1c9f06cc7a4fccc3f4904..c0396428cc3c0c908bb723c111adff2a1d44ebac 100644 (file)
@@ -158,7 +158,7 @@ static inline int __is_processor(unsigned long pv)
  * is more like most of the other architectures.
  */
 static __inline__ unsigned long
-__xchg_u32(volatile int *m, unsigned long val)
+__xchg_u32(volatile unsigned int *m, unsigned long val)
 {
        unsigned long dummy;
 
@@ -200,7 +200,7 @@ __xchg_u64(volatile long *m, unsigned long val)
 extern void __xchg_called_with_bad_pointer(void);
 
 static __inline__ unsigned long
-__xchg(volatile void *ptr, unsigned long x, int size)
+__xchg(volatile void *ptr, unsigned long x, unsigned int size)
 {
        switch (size) {
        case 4:
@@ -223,7 +223,7 @@ __xchg(volatile void *ptr, unsigned long x, int size)
 #define __HAVE_ARCH_CMPXCHG    1
 
 static __inline__ unsigned long
-__cmpxchg_u32(volatile int *p, int old, int new)
+__cmpxchg_u32(volatile unsigned int *p, unsigned long old, unsigned long new)
 {
        unsigned int prev;
 
@@ -271,7 +271,8 @@ __cmpxchg_u64(volatile long *p, unsigned long old, unsigned long new)
 extern void __cmpxchg_called_with_bad_pointer(void);
 
 static __inline__ unsigned long
-__cmpxchg(volatile void *ptr, unsigned long old, unsigned long new, int size)
+__cmpxchg(volatile void *ptr, unsigned long old, unsigned long new,
+         unsigned int size)
 {
        switch (size) {
        case 4:
@@ -283,13 +284,9 @@ __cmpxchg(volatile void *ptr, unsigned long old, unsigned long new, int size)
        return old;
 }
 
-#define cmpxchg(ptr,o,n)                                                \
-  ({                                                                    \
-     __typeof__(*(ptr)) _o_ = (o);                                      \
-     __typeof__(*(ptr)) _n_ = (n);                                      \
-     (__typeof__(*(ptr))) __cmpxchg((ptr), (unsigned long)_o_,          \
-                                   (unsigned long)_n_, sizeof(*(ptr))); \
-  })
+#define cmpxchg(ptr,o,n)\
+       ((__typeof__(*(ptr)))__cmpxchg((ptr),(unsigned long)(o),\
+       (unsigned long)(n),sizeof(*(ptr))))
 
 /*
  * We handle most unaligned accesses in hardware. On the other hand