OLD | NEW |
(Empty) | |
| 1 // Copyright (c) 2011 The Chromium Authors. All rights reserved. |
| 2 // Use of this source code is governed by a BSD-style license that can be |
| 3 // found in the LICENSE file. |
| 4 |
| 5 // This file is an internal atomic implementation, use base/atomicops.h instead. |
| 6 |
| 7 #ifndef BASE_ATOMICOPS_INTERNALS_X86_GCC_H_ |
| 8 #define BASE_ATOMICOPS_INTERNALS_X86_GCC_H_ |
| 9 |
| 10 #define ATOMICOPS_COMPILER_BARRIER() __asm__ __volatile__("" : : : "memory") |
| 11 |
| 12 namespace base { |
| 13 namespace subtle { |
| 14 |
| 15 // 32-bit low-level operations on any platform. |
| 16 |
| 17 inline Atomic32 NoBarrier_CompareAndSwap(volatile Atomic32* ptr, |
| 18 Atomic32 old_value, |
| 19 Atomic32 new_value) { |
| 20 Atomic32 prev; |
| 21 __asm__ __volatile__("lock; cmpxchgl %1,%2" |
| 22 : "=a" (prev) |
| 23 : "q" (new_value), "m" (*ptr), "0" (old_value) |
| 24 : "memory"); |
| 25 return prev; |
| 26 } |
| 27 |
| 28 inline Atomic32 NoBarrier_AtomicExchange(volatile Atomic32* ptr, |
| 29 Atomic32 new_value) { |
| 30 __asm__ __volatile__("xchgl %1,%0" // The lock prefix is implicit for xchg. |
| 31 : "=r" (new_value) |
| 32 : "m" (*ptr), "0" (new_value) |
| 33 : "memory"); |
| 34 return new_value; // Now it's the previous value. |
| 35 } |
| 36 |
| 37 inline Atomic32 NoBarrier_AtomicIncrement(volatile Atomic32* ptr, |
| 38 Atomic32 increment) { |
| 39 Atomic32 temp = increment; |
| 40 __asm__ __volatile__("lock; xaddl %0,%1" |
| 41 : "+r" (temp), "+m" (*ptr) |
| 42 : : "memory"); |
| 43 // temp now holds the old value of *ptr |
| 44 return temp + increment; |
| 45 } |
| 46 |
| 47 inline Atomic32 Barrier_AtomicIncrement(volatile Atomic32* ptr, |
| 48 Atomic32 increment) { |
| 49 Atomic32 temp = increment; |
| 50 __asm__ __volatile__("lock; xaddl %0,%1" |
| 51 : "+r" (temp), "+m" (*ptr) |
| 52 : : "memory"); |
| 53 // temp now holds the old value of *ptr |
| 54 if (AtomicOps_Internalx86CPUFeatures.has_amd_lock_mb_bug) { |
| 55 __asm__ __volatile__("lfence" : : : "memory"); |
| 56 } |
| 57 return temp + increment; |
| 58 } |
| 59 |
| 60 inline Atomic32 Acquire_CompareAndSwap(volatile Atomic32* ptr, |
| 61 Atomic32 old_value, |
| 62 Atomic32 new_value) { |
| 63 Atomic32 x = NoBarrier_CompareAndSwap(ptr, old_value, new_value); |
| 64 if (AtomicOps_Internalx86CPUFeatures.has_amd_lock_mb_bug) { |
| 65 __asm__ __volatile__("lfence" : : : "memory"); |
| 66 } |
| 67 return x; |
| 68 } |
| 69 |
| 70 inline Atomic32 Release_CompareAndSwap(volatile Atomic32* ptr, |
| 71 Atomic32 old_value, |
| 72 Atomic32 new_value) { |
| 73 return NoBarrier_CompareAndSwap(ptr, old_value, new_value); |
| 74 } |
| 75 |
| 76 inline void NoBarrier_Store(volatile Atomic32* ptr, Atomic32 value) { |
| 77 *ptr = value; |
| 78 } |
| 79 |
| 80 inline void MemoryBarrier() { |
| 81 __asm__ __volatile__("mfence" : : : "memory"); |
| 82 } |
| 83 |
| 84 inline void Acquire_Store(volatile Atomic32* ptr, Atomic32 value) { |
| 85 *ptr = value; |
| 86 MemoryBarrier(); |
| 87 } |
| 88 |
| 89 inline void Release_Store(volatile Atomic32* ptr, Atomic32 value) { |
| 90 ATOMICOPS_COMPILER_BARRIER(); |
| 91 *ptr = value; // An x86 store acts as a release barrier. |
| 92 // See comments in Atomic64 version of Release_Store(), below. |
| 93 } |
| 94 |
| 95 inline Atomic32 NoBarrier_Load(volatile const Atomic32* ptr) { |
| 96 return *ptr; |
| 97 } |
| 98 |
| 99 inline Atomic32 Acquire_Load(volatile const Atomic32* ptr) { |
| 100 Atomic32 value = *ptr; // An x86 load acts as a acquire barrier. |
| 101 // See comments in Atomic64 version of Release_Store(), below. |
| 102 ATOMICOPS_COMPILER_BARRIER(); |
| 103 return value; |
| 104 } |
| 105 |
| 106 inline Atomic32 Release_Load(volatile const Atomic32* ptr) { |
| 107 MemoryBarrier(); |
| 108 return *ptr; |
| 109 } |
| 110 |
| 111 #if defined(__x86_64__) |
| 112 |
| 113 // 64-bit low-level operations on 64-bit platform. |
| 114 |
| 115 inline Atomic64 NoBarrier_CompareAndSwap(volatile Atomic64* ptr, |
| 116 Atomic64 old_value, |
| 117 Atomic64 new_value) { |
| 118 Atomic64 prev; |
| 119 __asm__ __volatile__("lock; cmpxchgq %1,%2" |
| 120 : "=a" (prev) |
| 121 : "q" (new_value), "m" (*ptr), "0" (old_value) |
| 122 : "memory"); |
| 123 return prev; |
| 124 } |
| 125 |
| 126 inline Atomic64 NoBarrier_AtomicExchange(volatile Atomic64* ptr, |
| 127 Atomic64 new_value) { |
| 128 __asm__ __volatile__("xchgq %1,%0" // The lock prefix is implicit for xchg. |
| 129 : "=r" (new_value) |
| 130 : "m" (*ptr), "0" (new_value) |
| 131 : "memory"); |
| 132 return new_value; // Now it's the previous value. |
| 133 } |
| 134 |
| 135 inline Atomic64 NoBarrier_AtomicIncrement(volatile Atomic64* ptr, |
| 136 Atomic64 increment) { |
| 137 Atomic64 temp = increment; |
| 138 __asm__ __volatile__("lock; xaddq %0,%1" |
| 139 : "+r" (temp), "+m" (*ptr) |
| 140 : : "memory"); |
| 141 // temp now contains the previous value of *ptr |
| 142 return temp + increment; |
| 143 } |
| 144 |
| 145 inline Atomic64 Barrier_AtomicIncrement(volatile Atomic64* ptr, |
| 146 Atomic64 increment) { |
| 147 Atomic64 temp = increment; |
| 148 __asm__ __volatile__("lock; xaddq %0,%1" |
| 149 : "+r" (temp), "+m" (*ptr) |
| 150 : : "memory"); |
| 151 // temp now contains the previous value of *ptr |
| 152 if (AtomicOps_Internalx86CPUFeatures.has_amd_lock_mb_bug) { |
| 153 __asm__ __volatile__("lfence" : : : "memory"); |
| 154 } |
| 155 return temp + increment; |
| 156 } |
| 157 |
| 158 inline void NoBarrier_Store(volatile Atomic64* ptr, Atomic64 value) { |
| 159 *ptr = value; |
| 160 } |
| 161 |
| 162 inline void Acquire_Store(volatile Atomic64* ptr, Atomic64 value) { |
| 163 *ptr = value; |
| 164 MemoryBarrier(); |
| 165 } |
| 166 |
| 167 inline void Release_Store(volatile Atomic64* ptr, Atomic64 value) { |
| 168 ATOMICOPS_COMPILER_BARRIER(); |
| 169 |
| 170 *ptr = value; // An x86 store acts as a release barrier |
| 171 // for current AMD/Intel chips as of Jan 2008. |
| 172 // See also Acquire_Load(), below. |
| 173 |
| 174 // When new chips come out, check: |
| 175 // IA-32 Intel Architecture Software Developer's Manual, Volume 3: |
| 176 // System Programming Guide, Chatper 7: Multiple-processor management, |
| 177 // Section 7.2, Memory Ordering. |
| 178 // Last seen at: |
| 179 // http://developer.intel.com/design/pentium4/manuals/index_new.htm |
| 180 // |
| 181 // x86 stores/loads fail to act as barriers for a few instructions (clflush |
| 182 // maskmovdqu maskmovq movntdq movnti movntpd movntps movntq) but these are |
| 183 // not generated by the compiler, and are rare. Users of these instructions |
| 184 // need to know about cache behaviour in any case since all of these involve |
| 185 // either flushing cache lines or non-temporal cache hints. |
| 186 } |
| 187 |
| 188 inline Atomic64 NoBarrier_Load(volatile const Atomic64* ptr) { |
| 189 return *ptr; |
| 190 } |
| 191 |
| 192 inline Atomic64 Acquire_Load(volatile const Atomic64* ptr) { |
| 193 Atomic64 value = *ptr; // An x86 load acts as a acquire barrier, |
| 194 // for current AMD/Intel chips as of Jan 2008. |
| 195 // See also Release_Store(), above. |
| 196 ATOMICOPS_COMPILER_BARRIER(); |
| 197 return value; |
| 198 } |
| 199 |
| 200 inline Atomic64 Release_Load(volatile const Atomic64* ptr) { |
| 201 MemoryBarrier(); |
| 202 return *ptr; |
| 203 } |
| 204 |
| 205 inline Atomic64 Acquire_CompareAndSwap(volatile Atomic64* ptr, |
| 206 Atomic64 old_value, |
| 207 Atomic64 new_value) { |
| 208 Atomic64 x = NoBarrier_CompareAndSwap(ptr, old_value, new_value); |
| 209 if (AtomicOps_Internalx86CPUFeatures.has_amd_lock_mb_bug) { |
| 210 __asm__ __volatile__("lfence" : : : "memory"); |
| 211 } |
| 212 return x; |
| 213 } |
| 214 |
| 215 inline Atomic64 Release_CompareAndSwap(volatile Atomic64* ptr, |
| 216 Atomic64 old_value, |
| 217 Atomic64 new_value) { |
| 218 return NoBarrier_CompareAndSwap(ptr, old_value, new_value); |
| 219 } |
| 220 |
| 221 #endif // defined(__x86_64__) |
| 222 |
| 223 } // namespace subtle |
| 224 } // namespace base |
| 225 |
| 226 #undef ATOMICOPS_COMPILER_BARRIER |
| 227 |
| 228 #endif // BASE_ATOMICOPS_INTERNALS_X86_GCC_H_ |
OLD | NEW |