powerpc: use __NO_LWSYNC__ check to use appropriate lwsync/sync opcode
[urcu.git] / urcu / uatomic / ppc.h
1 #ifndef _URCU_ARCH_UATOMIC_PPC_H
2 #define _URCU_ARCH_UATOMIC_PPC_H
3
4 /*
5 * Copyright (c) 1991-1994 by Xerox Corporation. All rights reserved.
6 * Copyright (c) 1996-1999 by Silicon Graphics. All rights reserved.
7 * Copyright (c) 1999-2004 Hewlett-Packard Development Company, L.P.
8 * Copyright (c) 2009 Mathieu Desnoyers
9 *
10 * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY EXPRESSED
11 * OR IMPLIED. ANY USE IS AT YOUR OWN RISK.
12 *
13 * Permission is hereby granted to use or copy this program
14 * for any purpose, provided the above notices are retained on all copies.
15 * Permission to modify the code and to distribute modified code is granted,
16 * provided the above notices are retained, and a notice that the code was
17 * modified is included with the above copyright notice.
18 *
19 * Code inspired from libuatomic_ops-1.2, inherited in part from the
20 * Boehm-Demers-Weiser conservative garbage collector.
21 */
22
23 #include <urcu/compiler.h>
24 #include <urcu/system.h>
25
26 #ifdef __cplusplus
27 extern "C" {
28 #endif
29
30 #define ILLEGAL_INSTR ".long 0xd00d00"
31
32 /*
33 * Providing sequential consistency semantic with respect to other
34 * instructions for cmpxchg and add_return family of atomic primitives.
35 *
36 * This is achieved with:
37 * lwsync (prior loads can be reordered after following load)
38 * lwarx
39 * stwcx.
40 * test if success (retry)
41 * sync
42 *
43 * Explanation of the sequential consistency provided by this scheme
44 * from Paul E. McKenney:
45 *
46 * The reason we can get away with the lwsync before is that if a prior
47 * store reorders with the lwarx, then you have to store to the atomic
48 * variable from some other CPU to detect it.
49 *
50 * And if you do that, the lwarx will lose its reservation, so the stwcx
51 * will fail. The atomic operation will retry, so that the caller won't be
52 * able to see the misordering.
53 */
54
55 /* xchg */
56
57 static inline __attribute__((always_inline))
58 unsigned long _uatomic_exchange(void *addr, unsigned long val, int len)
59 {
60 switch (len) {
61 case 4:
62 {
63 unsigned int result;
64
65 __asm__ __volatile__(
66 LWSYNC_OPCODE
67 "1:\t" "lwarx %0,0,%1\n" /* load and reserve */
68 "stwcx. %2,0,%1\n" /* else store conditional */
69 "bne- 1b\n" /* retry if lost reservation */
70 "sync\n"
71 : "=&r"(result)
72 : "r"(addr), "r"(val)
73 : "memory", "cc");
74
75 return result;
76 }
77 #if (CAA_BITS_PER_LONG == 64)
78 case 8:
79 {
80 unsigned long result;
81
82 __asm__ __volatile__(
83 LWSYNC_OPCODE
84 "1:\t" "ldarx %0,0,%1\n" /* load and reserve */
85 "stdcx. %2,0,%1\n" /* else store conditional */
86 "bne- 1b\n" /* retry if lost reservation */
87 "sync\n"
88 : "=&r"(result)
89 : "r"(addr), "r"(val)
90 : "memory", "cc");
91
92 return result;
93 }
94 #endif
95 }
96 /* generate an illegal instruction. Cannot catch this with linker tricks
97 * when optimizations are disabled. */
98 __asm__ __volatile__(ILLEGAL_INSTR);
99 return 0;
100 }
101
102 #define uatomic_xchg(addr, v) \
103 ((__typeof__(*(addr))) _uatomic_exchange((addr), (unsigned long)(v), \
104 sizeof(*(addr))))
105 /* cmpxchg */
106
107 static inline __attribute__((always_inline))
108 unsigned long _uatomic_cmpxchg(void *addr, unsigned long old,
109 unsigned long _new, int len)
110 {
111 switch (len) {
112 case 4:
113 {
114 unsigned int old_val;
115
116 __asm__ __volatile__(
117 LWSYNC_OPCODE
118 "1:\t" "lwarx %0,0,%1\n" /* load and reserve */
119 "cmpw %0,%3\n" /* if load is not equal to */
120 "bne 2f\n" /* old, fail */
121 "stwcx. %2,0,%1\n" /* else store conditional */
122 "bne- 1b\n" /* retry if lost reservation */
123 "sync\n"
124 "2:\n"
125 : "=&r"(old_val)
126 : "r"(addr), "r"((unsigned int)_new),
127 "r"((unsigned int)old)
128 : "memory", "cc");
129
130 return old_val;
131 }
132 #if (CAA_BITS_PER_LONG == 64)
133 case 8:
134 {
135 unsigned long old_val;
136
137 __asm__ __volatile__(
138 LWSYNC_OPCODE
139 "1:\t" "ldarx %0,0,%1\n" /* load and reserve */
140 "cmpd %0,%3\n" /* if load is not equal to */
141 "bne 2f\n" /* old, fail */
142 "stdcx. %2,0,%1\n" /* else store conditional */
143 "bne- 1b\n" /* retry if lost reservation */
144 "sync\n"
145 "2:\n"
146 : "=&r"(old_val)
147 : "r"(addr), "r"((unsigned long)_new),
148 "r"((unsigned long)old)
149 : "memory", "cc");
150
151 return old_val;
152 }
153 #endif
154 }
155 /* generate an illegal instruction. Cannot catch this with linker tricks
156 * when optimizations are disabled. */
157 __asm__ __volatile__(ILLEGAL_INSTR);
158 return 0;
159 }
160
161
162 #define uatomic_cmpxchg(addr, old, _new) \
163 ((__typeof__(*(addr))) _uatomic_cmpxchg((addr), (unsigned long)(old),\
164 (unsigned long)(_new), \
165 sizeof(*(addr))))
166
167 /* uatomic_add_return */
168
169 static inline __attribute__((always_inline))
170 unsigned long _uatomic_add_return(void *addr, unsigned long val,
171 int len)
172 {
173 switch (len) {
174 case 4:
175 {
176 unsigned int result;
177
178 __asm__ __volatile__(
179 LWSYNC_OPCODE
180 "1:\t" "lwarx %0,0,%1\n" /* load and reserve */
181 "add %0,%2,%0\n" /* add val to value loaded */
182 "stwcx. %0,0,%1\n" /* store conditional */
183 "bne- 1b\n" /* retry if lost reservation */
184 "sync\n"
185 : "=&r"(result)
186 : "r"(addr), "r"(val)
187 : "memory", "cc");
188
189 return result;
190 }
191 #if (CAA_BITS_PER_LONG == 64)
192 case 8:
193 {
194 unsigned long result;
195
196 __asm__ __volatile__(
197 LWSYNC_OPCODE
198 "1:\t" "ldarx %0,0,%1\n" /* load and reserve */
199 "add %0,%2,%0\n" /* add val to value loaded */
200 "stdcx. %0,0,%1\n" /* store conditional */
201 "bne- 1b\n" /* retry if lost reservation */
202 "sync\n"
203 : "=&r"(result)
204 : "r"(addr), "r"(val)
205 : "memory", "cc");
206
207 return result;
208 }
209 #endif
210 }
211 /* generate an illegal instruction. Cannot catch this with linker tricks
212 * when optimizations are disabled. */
213 __asm__ __volatile__(ILLEGAL_INSTR);
214 return 0;
215 }
216
217
218 #define uatomic_add_return(addr, v) \
219 ((__typeof__(*(addr))) _uatomic_add_return((addr), \
220 (unsigned long)(v), \
221 sizeof(*(addr))))
222
223 #ifdef __cplusplus
224 }
225 #endif
226
227 #include <urcu/uatomic/generic.h>
228
229 #endif /* _URCU_ARCH_UATOMIC_PPC_H */
This page took 0.047312 seconds and 4 git commands to generate.