eal/arm: fix clang build of native target
[dpdk.git] / lib / librte_eal / arm / include / rte_atomic_64.h
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2015 Cavium, Inc
3  * Copyright(c) 2020 Arm Limited
4  */
5
6 #ifndef _RTE_ATOMIC_ARM64_H_
7 #define _RTE_ATOMIC_ARM64_H_
8
9 #ifndef RTE_FORCE_INTRINSICS
10 #  error Platform must be built with RTE_FORCE_INTRINSICS
11 #endif
12
13 #ifdef __cplusplus
14 extern "C" {
15 #endif
16
17 #include "generic/rte_atomic.h"
18 #include <rte_branch_prediction.h>
19 #include <rte_compat.h>
20 #include <rte_debug.h>
21
22 #define rte_mb() asm volatile("dmb osh" : : : "memory")
23
24 #define rte_wmb() asm volatile("dmb oshst" : : : "memory")
25
26 #define rte_rmb() asm volatile("dmb oshld" : : : "memory")
27
28 #define rte_smp_mb() asm volatile("dmb ish" : : : "memory")
29
30 #define rte_smp_wmb() asm volatile("dmb ishst" : : : "memory")
31
32 #define rte_smp_rmb() asm volatile("dmb ishld" : : : "memory")
33
34 #define rte_io_mb() rte_mb()
35
36 #define rte_io_wmb() rte_wmb()
37
38 #define rte_io_rmb() rte_rmb()
39
40 static __rte_always_inline void
41 rte_atomic_thread_fence(int memorder)
42 {
43         __atomic_thread_fence(memorder);
44 }
45
46 /*------------------------ 128 bit atomic operations -------------------------*/
47
48 #if defined(__ARM_FEATURE_ATOMICS) || defined(RTE_ARM_FEATURE_ATOMICS)
49 #define __LSE_PREAMBLE  ".arch armv8-a+lse\n"
50
51 #define __ATOMIC128_CAS_OP(cas_op_name, op_string)                          \
52 static __rte_noinline rte_int128_t                                          \
53 cas_op_name(rte_int128_t *dst, rte_int128_t old, rte_int128_t updated)      \
54 {                                                                           \
55         /* caspX instructions register pair must start from even-numbered
56          * register at operand 1.
57          * So, specify registers for local variables here.
58          */                                                                 \
59         register uint64_t x0 __asm("x0") = (uint64_t)old.val[0];            \
60         register uint64_t x1 __asm("x1") = (uint64_t)old.val[1];            \
61         register uint64_t x2 __asm("x2") = (uint64_t)updated.val[0];        \
62         register uint64_t x3 __asm("x3") = (uint64_t)updated.val[1];        \
63         asm volatile(                                                       \
64                 __LSE_PREAMBLE                                              \
65                 op_string " %[old0], %[old1], %[upd0], %[upd1], [%[dst]]"   \
66                 : [old0] "+r" (x0),                                         \
67                 [old1] "+r" (x1)                                            \
68                 : [upd0] "r" (x2),                                          \
69                 [upd1] "r" (x3),                                            \
70                 [dst] "r" (dst)                                             \
71                 : "memory");                                                \
72         old.val[0] = x0;                                                    \
73         old.val[1] = x1;                                                    \
74         return old;                                                         \
75 }
76
77 __ATOMIC128_CAS_OP(__cas_128_relaxed, "casp")
78 __ATOMIC128_CAS_OP(__cas_128_acquire, "caspa")
79 __ATOMIC128_CAS_OP(__cas_128_release, "caspl")
80 __ATOMIC128_CAS_OP(__cas_128_acq_rel, "caspal")
81
82 #undef __LSE_PREAMBLE
83 #undef __ATOMIC128_CAS_OP
84
85 #endif
86
87 __rte_experimental
88 static inline int
89 rte_atomic128_cmp_exchange(rte_int128_t *dst, rte_int128_t *exp,
90                 const rte_int128_t *src, unsigned int weak, int success,
91                 int failure)
92 {
93         /* Always do strong CAS */
94         RTE_SET_USED(weak);
95         /* Ignore memory ordering for failure, memory order for
96          * success must be stronger or equal
97          */
98         RTE_SET_USED(failure);
99         /* Find invalid memory order */
100         RTE_ASSERT(success == __ATOMIC_RELAXED ||
101                 success == __ATOMIC_ACQUIRE ||
102                 success == __ATOMIC_RELEASE ||
103                 success == __ATOMIC_ACQ_REL ||
104                 success == __ATOMIC_SEQ_CST);
105
106         rte_int128_t expected = *exp;
107         rte_int128_t desired = *src;
108         rte_int128_t old;
109
110 #if defined(__ARM_FEATURE_ATOMICS) || defined(RTE_ARM_FEATURE_ATOMICS)
111         if (success == __ATOMIC_RELAXED)
112                 old = __cas_128_relaxed(dst, expected, desired);
113         else if (success == __ATOMIC_ACQUIRE)
114                 old = __cas_128_acquire(dst, expected, desired);
115         else if (success == __ATOMIC_RELEASE)
116                 old = __cas_128_release(dst, expected, desired);
117         else
118                 old = __cas_128_acq_rel(dst, expected, desired);
119 #else
120 #define __HAS_ACQ(mo) ((mo) != __ATOMIC_RELAXED && (mo) != __ATOMIC_RELEASE)
121 #define __HAS_RLS(mo) ((mo) == __ATOMIC_RELEASE || (mo) == __ATOMIC_ACQ_REL || \
122                 (mo) == __ATOMIC_SEQ_CST)
123
124         int ldx_mo = __HAS_ACQ(success) ? __ATOMIC_ACQUIRE : __ATOMIC_RELAXED;
125         int stx_mo = __HAS_RLS(success) ? __ATOMIC_RELEASE : __ATOMIC_RELAXED;
126
127 #undef __HAS_ACQ
128 #undef __HAS_RLS
129
130         uint32_t ret = 1;
131
132         /* ldx128 can not guarantee atomic,
133          * Must write back src or old to verify atomicity of ldx128;
134          */
135         do {
136
137 #define __LOAD_128(op_string, src, dst) { \
138         asm volatile(                     \
139                 op_string " %0, %1, %2"   \
140                 : "=&r" (dst.val[0]),     \
141                   "=&r" (dst.val[1])      \
142                 : "Q" (src->val[0])       \
143                 : "memory"); }
144
145                 if (ldx_mo == __ATOMIC_RELAXED)
146                         __LOAD_128("ldxp", dst, old)
147                 else
148                         __LOAD_128("ldaxp", dst, old)
149
150 #undef __LOAD_128
151
152 #define __STORE_128(op_string, dst, src, ret) { \
153         asm volatile(                           \
154                 op_string " %w0, %1, %2, %3"    \
155                 : "=&r" (ret)                   \
156                 : "r" (src.val[0]),             \
157                   "r" (src.val[1]),             \
158                   "Q" (dst->val[0])             \
159                 : "memory"); }
160
161                 if (likely(old.int128 == expected.int128)) {
162                         if (stx_mo == __ATOMIC_RELAXED)
163                                 __STORE_128("stxp", dst, desired, ret)
164                         else
165                                 __STORE_128("stlxp", dst, desired, ret)
166                 } else {
167                         /* In the failure case (since 'weak' is ignored and only
168                          * weak == 0 is implemented), expected should contain
169                          * the atomically read value of dst. This means, 'old'
170                          * needs to be stored back to ensure it was read
171                          * atomically.
172                          */
173                         if (stx_mo == __ATOMIC_RELAXED)
174                                 __STORE_128("stxp", dst, old, ret)
175                         else
176                                 __STORE_128("stlxp", dst, old, ret)
177                 }
178
179 #undef __STORE_128
180
181         } while (unlikely(ret));
182 #endif
183
184         /* Unconditionally updating expected removes an 'if' statement.
185          * expected should already be in register if not in the cache.
186          */
187         *exp = old;
188
189         return (old.int128 == expected.int128);
190 }
191
192 #ifdef __cplusplus
193 }
194 #endif
195
196 #endif /* _RTE_ATOMIC_ARM64_H_ */