diff --git a/include/gcc/x86/ck_pr.h b/include/gcc/x86/ck_pr.h
index 43eda0a..5194dee 100644
--- a/include/gcc/x86/ck_pr.h
+++ b/include/gcc/x86/ck_pr.h
@@ -307,7 +307,7 @@ CK_PR_GENERATE(xor)
 #undef CK_PR_BINARY
 
 /*
- * Atomic compare and swap.
+ * Atomic compare and swap, with a variant that sets *v to the old value of target.
  */
 #ifdef __GCC_ASM_FLAG_OUTPUTS__
 #define CK_PR_CAS(S, M, T, C, I)						\
@@ -323,6 +323,20 @@ CK_PR_GENERATE(xor)
 					: "q"     (set)				\
 					: "memory", "cc");			\
 		return z;							\
+	}									\
+										\
+	CK_CC_INLINE static bool						\
+	ck_pr_cas_##S##_value(M *target, T compare, T set, M *v)		\
+	{									\
+		bool z;								\
+		__asm__ __volatile__(CK_PR_LOCK_PREFIX I " %3, %0;"		\
+					: "+m"    (*(C *)target),		\
+					  "=@ccz" (z),				\
+					  "+a"    (compare)			\
+					: "q"     (set)				\
+					: "memory", "cc");			\
+		*(T *)v = compare;						\
+		return z;							\
 	}
 #else
 #define CK_PR_CAS(S, M, T, C, I)						\
@@ -337,48 +351,13 @@ CK_PR_GENERATE(xor)
 					  "a"   (compare)			\
 					: "memory", "cc");			\
 		return z;							\
-	}
-#endif
-
-CK_PR_CAS(ptr, void, void *, char, "cmpxchgl")
-
-#define CK_PR_CAS_S(S, T, I) CK_PR_CAS(S, T, T, T, I)
-
-CK_PR_CAS_S(char, char, "cmpxchgb")
-CK_PR_CAS_S(int, int, "cmpxchgl")
-CK_PR_CAS_S(uint, unsigned int, "cmpxchgl")
-CK_PR_CAS_S(32, uint32_t, "cmpxchgl")
-CK_PR_CAS_S(16, uint16_t, "cmpxchgw")
-CK_PR_CAS_S(8,  uint8_t,  "cmpxchgb")
-
-#undef CK_PR_CAS_S
-#undef CK_PR_CAS
-
-/*
- * Compare and swap, set *v to old value of target.
- */
-#ifdef __GCC_ASM_FLAG_OUTPUTS__
-#define CK_PR_CAS_O(S, M, T, C, I)						\
-	CK_CC_INLINE static bool						\
-	ck_pr_cas_##S##_value(M *target, T compare, T set, M *v)		\
-	{									\
-		bool z;								\
-		__asm__ __volatile__(CK_PR_LOCK_PREFIX "cmpxchg" I " %3, %0;"	\
-					: "+m"    (*(C *)target),		\
-					  "=@ccz" (z),				\
-					  "+a"    (compare)			\
-					: "q"     (set)				\
-					: "memory", "cc");			\
-		*(T *)v = compare;						\
-		return z;							\
-	}
-#else
-#define CK_PR_CAS_O(S, M, T, C, I)						\
+	}									\
+										\
 	CK_CC_INLINE static bool						\
 	ck_pr_cas_##S##_value(M *target, T compare, T set, M *v)		\
 	{									\
 		bool z;								\
-		__asm__ __volatile__(CK_PR_LOCK_PREFIX "cmpxchg" I " %3, %0;"	\
+		__asm__ __volatile__(CK_PR_LOCK_PREFIX I " %3, %0;"		\
 				     "setz %1;"					\
 					: "+m"  (*(C *)target),			\
 					  "=q"  (z),				\
@@ -390,20 +369,19 @@ CK_PR_CAS_S(8,  uint8_t,  "cmpxchgb")
 	}
 #endif
 
-CK_PR_CAS_O(ptr, void, void *, char, "l")
+CK_PR_CAS(ptr, void, void *, char, "cmpxchgl")
 
-#define CK_PR_CAS_O_S(S, T, I)	\
-	CK_PR_CAS_O(S, T, T, T, I)
+#define CK_PR_CAS_S(S, T, I) CK_PR_CAS(S, T, T, T, I)
 
-CK_PR_CAS_O_S(char, char, "b")
-CK_PR_CAS_O_S(int, int, "l")
-CK_PR_CAS_O_S(uint, unsigned int, "l")
-CK_PR_CAS_O_S(32, uint32_t, "l")
-CK_PR_CAS_O_S(16, uint16_t, "w")
-CK_PR_CAS_O_S(8,  uint8_t,  "b")
+CK_PR_CAS_S(char, char, "cmpxchgb")
+CK_PR_CAS_S(int, int, "cmpxchgl")
+CK_PR_CAS_S(uint, unsigned int, "cmpxchgl")
+CK_PR_CAS_S(32, uint32_t, "cmpxchgl")
+CK_PR_CAS_S(16, uint16_t, "cmpxchgw")
+CK_PR_CAS_S(8,  uint8_t,  "cmpxchgb")
 
-#undef CK_PR_CAS_O_S
-#undef CK_PR_CAS_O
+#undef CK_PR_CAS_S
+#undef CK_PR_CAS
 
 /*
  * Atomic bit test operations.
diff --git a/include/gcc/x86_64/ck_pr.h b/include/gcc/x86_64/ck_pr.h
index 85916d9..4222729 100644
--- a/include/gcc/x86_64/ck_pr.h
+++ b/include/gcc/x86_64/ck_pr.h
@@ -408,7 +408,7 @@ CK_PR_GENERATE(xor)
 #undef CK_PR_BINARY
 
 /*
- * Atomic compare and swap.
+ * Atomic compare and swap, with a variant that sets *v to the old value of target.
  */
 #ifdef __GCC_ASM_FLAG_OUTPUTS__
 #define CK_PR_CAS(S, M, T, C, I)						\
@@ -424,6 +424,20 @@ CK_PR_GENERATE(xor)
 					: "q"     (set)				\
 					: "memory", "cc");			\
 		return z;							\
+	}									\
+										\
+	CK_CC_INLINE static bool						\
+	ck_pr_cas_##S##_value(M *target, T compare, T set, M *v)		\
+	{									\
+		bool z;								\
+		__asm__ __volatile__(CK_PR_LOCK_PREFIX I " %3, %0;"		\
+					: "+m"    (*(C *)target),		\
+					  "=@ccz" (z),				\
+					  "+a"    (compare)			\
+					: "q"     (set)				\
+					: "memory", "cc");			\
+		*(T *)v = compare;						\
+		return z;							\
 	}
 #else
 #define CK_PR_CAS(S, M, T, C, I)						\
@@ -438,52 +452,13 @@ CK_PR_GENERATE(xor)
 					  "a"   (compare)			\
 					: "memory", "cc");			\
 		return z;							\
-	}
-#endif
-
-CK_PR_CAS(ptr, void, void *, char, "cmpxchgq")
-
-#define CK_PR_CAS_S(S, T, I) CK_PR_CAS(S, T, T, T, I)
-
-CK_PR_CAS_S(char, char, "cmpxchgb")
-CK_PR_CAS_S(int, int, "cmpxchgl")
-CK_PR_CAS_S(uint, unsigned int, "cmpxchgl")
-#ifndef CK_PR_DISABLE_DOUBLE
-CK_PR_CAS_S(double, double, "cmpxchgq")
-#endif
-CK_PR_CAS_S(64, uint64_t, "cmpxchgq")
-CK_PR_CAS_S(32, uint32_t, "cmpxchgl")
-CK_PR_CAS_S(16, uint16_t, "cmpxchgw")
-CK_PR_CAS_S(8,  uint8_t,  "cmpxchgb")
-
-#undef CK_PR_CAS_S
-#undef CK_PR_CAS
-
-/*
- * Compare and swap, set *v to old value of target.
- */
-#ifdef __GCC_ASM_FLAG_OUTPUTS__
-#define CK_PR_CAS_O(S, M, T, C, I)						\
-	CK_CC_INLINE static bool						\
-	ck_pr_cas_##S##_value(M *target, T compare, T set, M *v)		\
-	{									\
-		bool z;								\
-		__asm__ __volatile__(CK_PR_LOCK_PREFIX "cmpxchg" I " %3, %0;"	\
-					: "+m"    (*(C *)target),		\
-					  "=@ccz" (z),				\
-					  "+a"    (compare)			\
-					: "q"     (set)				\
-					: "memory", "cc");			\
-		*(T *)v = compare;						\
-		return z;							\
-	}
-#else
-#define CK_PR_CAS_O(S, M, T, C, I)						\
+	}									\
+										\
 	CK_CC_INLINE static bool						\
 	ck_pr_cas_##S##_value(M *target, T compare, T set, M *v)		\
 	{									\
 		bool z;								\
-		__asm__ __volatile__(CK_PR_LOCK_PREFIX "cmpxchg" I " %3, %0;"	\
+		__asm__ __volatile__(CK_PR_LOCK_PREFIX I " %3, %0;"		\
 				     "setz %1;"					\
 					: "+m"  (*(C *)target),			\
 					  "=q"  (z),				\
@@ -495,24 +470,23 @@ CK_PR_CAS_S(8,  uint8_t,  "cmpxchgb")
 	}
 #endif
 
-CK_PR_CAS_O(ptr, void, void *, char, "q")
+CK_PR_CAS(ptr, void, void *, char, "cmpxchgq")
 
-#define CK_PR_CAS_O_S(S, T, I)	\
-	CK_PR_CAS_O(S, T, T, T, I)
+#define CK_PR_CAS_S(S, T, I) CK_PR_CAS(S, T, T, T, I)
 
-CK_PR_CAS_O_S(char, char, "b")
-CK_PR_CAS_O_S(int, int, "l")
-CK_PR_CAS_O_S(uint, unsigned int, "l")
+CK_PR_CAS_S(char, char, "cmpxchgb")
+CK_PR_CAS_S(int, int, "cmpxchgl")
+CK_PR_CAS_S(uint, unsigned int, "cmpxchgl")
 #ifndef CK_PR_DISABLE_DOUBLE
-CK_PR_CAS_O_S(double, double, "q")
+CK_PR_CAS_S(double, double, "cmpxchgq")
 #endif
-CK_PR_CAS_O_S(64, uint64_t, "q")
-CK_PR_CAS_O_S(32, uint32_t, "l")
-CK_PR_CAS_O_S(16, uint16_t, "w")
-CK_PR_CAS_O_S(8,  uint8_t,  "b")
+CK_PR_CAS_S(64, uint64_t, "cmpxchgq")
+CK_PR_CAS_S(32, uint32_t, "cmpxchgl")
+CK_PR_CAS_S(16, uint16_t, "cmpxchgw")
+CK_PR_CAS_S(8,  uint8_t,  "cmpxchgb")
 
-#undef CK_PR_CAS_O_S
-#undef CK_PR_CAS_O
+#undef CK_PR_CAS_S
+#undef CK_PR_CAS
 
 /*
  * Contrary to C-interface, alignment requirements are that of uint64_t[2].