2 * Copyright (c) 1991-1994 by Xerox Corporation. All rights reserved.
3 * Copyright (c) 1996-1999 by Silicon Graphics. All rights reserved.
4 * Copyright (c) 1999-2003 by Hewlett-Packard Company. All rights reserved.
7 * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY EXPRESSED
8 * OR IMPLIED. ANY USE IS AT YOUR OWN RISK.
10 * Permission is hereby granted to use or copy this program
11 * for any purpose, provided the above notices are retained on all copies.
12 * Permission to modify the code and to distribute modified code is granted,
13 * provided the above notices are retained, and a notice that the code was
14 * modified is included with the above copyright notice.
18 #include "../read_ordered.h"
20 #include "../test_and_set_t_is_ao_t.h" /* Probably suboptimal */
22 /* NEC LE-IT: ARMv6 is the first architecture providing support for */
23 /* simple LL/SC. A data memory barrier must be raised via CP15 command */
24 /* (see documentation). */
25 /* ARMv7 is compatible to ARMv6 but has a simpler command for issuing */
26 /* a memory barrier (DMB). Raising it via CP15 should still work as */
27 /* told me by the support engineers. If it turns out to be much quicker */
28 /* than we should implement custom code for ARMv7 using the asm { dmb } */
30 /* If only a single processor is used, we can define AO_UNIPROCESSOR */
31 /* and do not need to access CP15 for ensuring a DMB. */
33 /* NEC LE-IT: gcc has no way to easily check the arm architecture */
34 /* but it defines only one of __ARM_ARCH_x__ to be true. */
35 #if !defined(__ARM_ARCH_2__) && !defined(__ARM_ARCH_3__) \
36 && !defined(__ARM_ARCH_3M__) && !defined(__ARM_ARCH_4__) \
37 && !defined(__ARM_ARCH_4T__) && !defined(__ARM_ARCH_5__) \
38 && !defined(__ARM_ARCH_5E__) && !defined(__ARM_ARCH_5T__) \
39 && !defined(__ARM_ARCH_5TE__) && !defined(__ARM_ARCH_5TEJ__)
41 #include "../standard_ao_double_t.h"
46 #ifndef AO_UNIPROCESSOR
47 /* Issue a data memory barrier (keeps ordering of memory */
48 /* transactions before and after this operation). */
50 __asm__ __volatile__("mcr p15,0,%0,c7,c10,5"
51 : "=&r"(dest) : : "memory");
54 #define AO_HAVE_nop_full
56 /* NEC LE-IT: AO_t load is simple reading */
58 AO_load(const volatile AO_t *addr)
60 /* Cast away the volatile for architectures like IA64 where */
61 /* volatile adds barrier semantics. */
62 return (*(const AO_t *)addr);
66 /* NEC LE-IT: atomic "store" - according to ARM documentation this is
67 * the only safe way to set variables also used in LL/SC environment.
68 * A direct write won't be recognized by the LL/SC construct on the _same_ CPU.
69 * Support engineers response for behaviour of ARMv6:
72 ===================================
75 -----------------------------------
79 -----------------------------------
83 -----------------------------------
85 * ARMv7 behaves similar, see documentation CortexA8 TRM, point 8.5
87 * HB: I think this is only a problem if interrupt handlers do not clear
88 * the reservation, as they almost certainly should. Probably change this back
91 AO_INLINE void AO_store(volatile AO_t *addr, AO_t value)
95 __asm__ __volatile__("@AO_store\n"
97 " strex %0, %3, [%2]\n"
100 : "=&r"(flag), "+m"(*addr)
101 : "r" (addr), "r"(value)
104 #define AO_HAVE_store
106 /* NEC LE-IT: replace the SWAP as recommended by ARM:
107 "Applies to: ARM11 Cores
108 Though the SWP instruction will still work with ARM V6 cores, it is
109 recommended to use the new V6 synchronization instructions. The SWP
110 instruction produces 'locked' read and write accesses which are atomic,
111 i.e. another operation cannot be done between these locked accesses which
112 ties up external bus (AHB,AXI) bandwidth and can increase worst case
113 interrupt latencies. LDREX,STREX are more flexible, other instructions
114 can be done between the LDREX and STREX accesses."
117 AO_test_and_set(volatile AO_TS_t *addr)
123 __asm__ __volatile__("@AO_test_and_set\n"
124 "1: ldrex %0, [%3]\n"
125 " strex %1, %4, [%3]\n"
128 : "=&r"(oldval),"=&r"(flag), "+m"(*addr)
134 #define AO_HAVE_test_and_set
136 /* NEC LE-IT: fetch and add for ARMv6 */
138 AO_fetch_and_add(volatile AO_t *p, AO_t incr)
140 unsigned long flag,tmp;
143 __asm__ __volatile__("@AO_fetch_and_add\n"
144 "1: ldrex %0, [%5]\n" /* get original */
145 " add %2, %0, %4\n" /* sum up in incr */
146 " strex %1, %2, [%5]\n" /* store them */
149 : "=&r"(result),"=&r"(flag),"=&r"(tmp),"+m"(*p) /* 0..3 */
150 : "r"(incr), "r"(p) /* 4..5 */
155 #define AO_HAVE_fetch_and_add
157 /* NEC LE-IT: fetch and add1 for ARMv6 */
159 AO_fetch_and_add1(volatile AO_t *p)
161 unsigned long flag,tmp;
164 __asm__ __volatile__("@AO_fetch_and_add1\n"
165 "1: ldrex %0, [%4]\n" /* get original */
166 " add %1, %0, #1\n" /* increment */
167 " strex %2, %1, [%4]\n" /* store them */
170 : "=&r"(result), "=&r"(tmp), "=&r"(flag), "+m"(*p)
176 #define AO_HAVE_fetch_and_add1
178 /* NEC LE-IT: fetch and sub for ARMv6 */
180 AO_fetch_and_sub1(volatile AO_t *p)
182 unsigned long flag,tmp;
185 __asm__ __volatile__("@AO_fetch_and_sub1\n"
186 "1: ldrex %0, [%4]\n" /* get original */
187 " sub %1, %0, #1\n" /* decrement */
188 " strex %2, %1, [%4]\n" /* store them */
191 : "=&r"(result), "=&r"(tmp), "=&r"(flag), "+m"(*p)
197 #define AO_HAVE_fetch_and_sub1
199 /* NEC LE-IT: compare and swap */
200 /* Returns nonzero if the comparison succeeded. */
202 AO_compare_and_swap(volatile AO_t *addr, AO_t old_val, AO_t new_val)
206 __asm__ __volatile__("@AO_compare_and_swap\n"
207 "1: mov %0, #2\n" /* store a flag */
208 " ldrex %1, [%3]\n" /* get original */
209 " teq %1, %4\n" /* see if match */
213 " strexeq %0, %5, [%3]\n" /* store new one if matched */
215 " beq 1b\n" /* if update failed, repeat */
216 : "=&r"(result), "=&r"(tmp), "+m"(*addr)
217 : "r"(addr), "r"(old_val), "r"(new_val)
220 return !(result&2); /* if succeded, return 1, else 0 */
222 #define AO_HAVE_compare_and_swap
224 #if !defined(__ARM_ARCH_6__) && !defined(__ARM_ARCH_6J__) \
225 && !defined(__ARM_ARCH_6T2__) && !defined(__ARM_ARCH_6Z__) \
226 && !defined(__ARM_ARCH_6ZT2__)
227 /* ldrexd/strexd present in ARMv6K/M+ (see gas/config/tc-arm.c) */
229 AO_compare_double_and_swap_double(volatile AO_double_t *addr,
230 AO_t old_val1, AO_t old_val2,
231 AO_t new_val1, AO_t new_val2)
233 double_ptr_storage old_val =
234 ((double_ptr_storage)old_val2 << 32) | old_val1;
235 double_ptr_storage new_val =
236 ((double_ptr_storage)new_val2 << 32) | new_val1;
237 double_ptr_storage tmp;
241 __asm__ __volatile__("@AO_compare_double_and_swap_double\n"
242 " ldrexd %0, [%1]\n" /* get original to r1 & r2 */
248 __asm__ __volatile__(
249 " strexd %0, %2, [%3]\n" /* store new one if matched */
250 : "=&r"(result),"+m"(*addr)
251 : "r"(new_val), "r"(addr)
256 # define AO_HAVE_compare_double_and_swap_double
260 /* pre ARMv6 architectures ... */
262 /* I found a slide set that, if I read it correctly, claims that */
263 /* Loads followed by either a Load or Store are ordered, but nothing */
265 /* It appears that SWP is the only simple memory barrier. */
266 #include "../all_atomic_load_store.h"
268 #if !defined(__ARM_ARCH_2__)
269 AO_INLINE AO_TS_VAL_t
270 AO_test_and_set_full(volatile AO_TS_t *addr)
273 /* SWP on ARM is very similar to XCHG on x86. */
274 /* The first operand is the result, the second the value */
275 /* to be stored. Both registers must be different from addr. */
276 /* Make the address operand an early clobber output so it */
277 /* doesn't overlap with the other operands. The early clobber */
278 /* on oldval is necessary to prevent the compiler allocating */
279 /* them to the same register if they are both unused. */
280 __asm__ __volatile__("swp %0, %2, [%3]"
281 : "=&r"(oldval), "=&r"(addr)
286 # define AO_HAVE_test_and_set_full
287 #endif /* !__ARM_ARCH_2__ */
289 #endif /* __ARM_ARCH_x */