OLD | NEW |
(Empty) | |
| 1 // Copyright 2010 the V8 project 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 atomicops.h instead. |
| 6 |
| 7 #ifndef V8_BASE_ATOMICOPS_INTERNALS_X86_GCC_H_ |
| 8 #define V8_BASE_ATOMICOPS_INTERNALS_X86_GCC_H_ |
| 9 |
| 10 #include "src/base/base-export.h" |
| 11 |
| 12 namespace v8 { |
| 13 namespace base { |
| 14 |
| 15 // This struct is not part of the public API of this module; clients may not |
| 16 // use it. |
| 17 // Features of this x86. Values may not be correct before main() is run, |
| 18 // but are set conservatively. |
| 19 struct AtomicOps_x86CPUFeatureStruct { |
| 20 bool has_amd_lock_mb_bug; // Processor has AMD memory-barrier bug; do lfence |
| 21 // after acquire compare-and-swap. |
| 22 #if !defined(__SSE2__) |
| 23 bool has_sse2; // Processor has SSE2. |
| 24 #endif |
| 25 }; |
| 26 V8_BASE_EXPORT extern struct AtomicOps_x86CPUFeatureStruct |
| 27 AtomicOps_Internalx86CPUFeatures; |
| 28 |
| 29 #define ATOMICOPS_COMPILER_BARRIER() __asm__ __volatile__("" : : : "memory") |
| 30 |
| 31 // 32-bit low-level operations on any platform. |
| 32 |
| 33 inline Atomic32 NoBarrier_CompareAndSwap(volatile Atomic32* ptr, |
| 34 Atomic32 old_value, |
| 35 Atomic32 new_value) { |
| 36 Atomic32 prev; |
| 37 __asm__ __volatile__("lock; cmpxchgl %1,%2" |
| 38 : "=a" (prev) |
| 39 : "q" (new_value), "m" (*ptr), "0" (old_value) |
| 40 : "memory"); |
| 41 return prev; |
| 42 } |
| 43 |
| 44 inline Atomic32 NoBarrier_AtomicExchange(volatile Atomic32* ptr, |
| 45 Atomic32 new_value) { |
| 46 __asm__ __volatile__("xchgl %1,%0" // The lock prefix is implicit for xchg. |
| 47 : "=r" (new_value) |
| 48 : "m" (*ptr), "0" (new_value) |
| 49 : "memory"); |
| 50 return new_value; // Now it's the previous value. |
| 51 } |
| 52 |
| 53 inline Atomic32 NoBarrier_AtomicIncrement(volatile Atomic32* ptr, |
| 54 Atomic32 increment) { |
| 55 Atomic32 temp = increment; |
| 56 __asm__ __volatile__("lock; xaddl %0,%1" |
| 57 : "+r" (temp), "+m" (*ptr) |
| 58 : : "memory"); |
| 59 // temp now holds the old value of *ptr |
| 60 return temp + increment; |
| 61 } |
| 62 |
| 63 inline Atomic32 Barrier_AtomicIncrement(volatile Atomic32* ptr, |
| 64 Atomic32 increment) { |
| 65 Atomic32 temp = increment; |
| 66 __asm__ __volatile__("lock; xaddl %0,%1" |
| 67 : "+r" (temp), "+m" (*ptr) |
| 68 : : "memory"); |
| 69 // temp now holds the old value of *ptr |
| 70 if (AtomicOps_Internalx86CPUFeatures.has_amd_lock_mb_bug) { |
| 71 __asm__ __volatile__("lfence" : : : "memory"); |
| 72 } |
| 73 return temp + increment; |
| 74 } |
| 75 |
| 76 inline Atomic32 Acquire_CompareAndSwap(volatile Atomic32* ptr, |
| 77 Atomic32 old_value, |
| 78 Atomic32 new_value) { |
| 79 Atomic32 x = NoBarrier_CompareAndSwap(ptr, old_value, new_value); |
| 80 if (AtomicOps_Internalx86CPUFeatures.has_amd_lock_mb_bug) { |
| 81 __asm__ __volatile__("lfence" : : : "memory"); |
| 82 } |
| 83 return x; |
| 84 } |
| 85 |
| 86 inline Atomic32 Release_CompareAndSwap(volatile Atomic32* ptr, |
| 87 Atomic32 old_value, |
| 88 Atomic32 new_value) { |
| 89 return NoBarrier_CompareAndSwap(ptr, old_value, new_value); |
| 90 } |
| 91 |
| 92 inline void NoBarrier_Store(volatile Atomic8* ptr, Atomic8 value) { |
| 93 *ptr = value; |
| 94 } |
| 95 |
| 96 inline void NoBarrier_Store(volatile Atomic32* ptr, Atomic32 value) { |
| 97 *ptr = value; |
| 98 } |
| 99 |
| 100 #if defined(__x86_64__) || defined(__SSE2__) |
| 101 |
| 102 // 64-bit implementations of memory barrier can be simpler, because it |
| 103 // "mfence" is guaranteed to exist. |
| 104 inline void MemoryBarrier() { |
| 105 __asm__ __volatile__("mfence" : : : "memory"); |
| 106 } |
| 107 |
| 108 inline void Acquire_Store(volatile Atomic32* ptr, Atomic32 value) { |
| 109 *ptr = value; |
| 110 MemoryBarrier(); |
| 111 } |
| 112 |
| 113 #else |
| 114 |
| 115 inline void MemoryBarrier() { |
| 116 if (AtomicOps_Internalx86CPUFeatures.has_sse2) { |
| 117 __asm__ __volatile__("mfence" : : : "memory"); |
| 118 } else { // mfence is faster but not present on PIII |
| 119 Atomic32 x = 0; |
| 120 NoBarrier_AtomicExchange(&x, 0); // acts as a barrier on PIII |
| 121 } |
| 122 } |
| 123 |
| 124 inline void Acquire_Store(volatile Atomic32* ptr, Atomic32 value) { |
| 125 if (AtomicOps_Internalx86CPUFeatures.has_sse2) { |
| 126 *ptr = value; |
| 127 __asm__ __volatile__("mfence" : : : "memory"); |
| 128 } else { |
| 129 NoBarrier_AtomicExchange(ptr, value); |
| 130 // acts as a barrier on PIII |
| 131 } |
| 132 } |
| 133 #endif |
| 134 |
| 135 inline void Release_Store(volatile Atomic32* ptr, Atomic32 value) { |
| 136 ATOMICOPS_COMPILER_BARRIER(); |
| 137 *ptr = value; // An x86 store acts as a release barrier. |
| 138 // See comments in Atomic64 version of Release_Store(), below. |
| 139 } |
| 140 |
| 141 inline Atomic8 NoBarrier_Load(volatile const Atomic8* ptr) { |
| 142 return *ptr; |
| 143 } |
| 144 |
| 145 inline Atomic32 NoBarrier_Load(volatile const Atomic32* ptr) { |
| 146 return *ptr; |
| 147 } |
| 148 |
| 149 inline Atomic32 Acquire_Load(volatile const Atomic32* ptr) { |
| 150 Atomic32 value = *ptr; // An x86 load acts as a acquire barrier. |
| 151 // See comments in Atomic64 version of Release_Store(), below. |
| 152 ATOMICOPS_COMPILER_BARRIER(); |
| 153 return value; |
| 154 } |
| 155 |
| 156 inline Atomic32 Release_Load(volatile const Atomic32* ptr) { |
| 157 MemoryBarrier(); |
| 158 return *ptr; |
| 159 } |
| 160 |
| 161 #if defined(__x86_64__) && defined(V8_HOST_ARCH_64_BIT) |
| 162 |
| 163 // 64-bit low-level operations on 64-bit platform. |
| 164 |
| 165 inline Atomic64 NoBarrier_CompareAndSwap(volatile Atomic64* ptr, |
| 166 Atomic64 old_value, |
| 167 Atomic64 new_value) { |
| 168 Atomic64 prev; |
| 169 __asm__ __volatile__("lock; cmpxchgq %1,%2" |
| 170 : "=a" (prev) |
| 171 : "q" (new_value), "m" (*ptr), "0" (old_value) |
| 172 : "memory"); |
| 173 return prev; |
| 174 } |
| 175 |
| 176 inline Atomic64 NoBarrier_AtomicExchange(volatile Atomic64* ptr, |
| 177 Atomic64 new_value) { |
| 178 __asm__ __volatile__("xchgq %1,%0" // The lock prefix is implicit for xchg. |
| 179 : "=r" (new_value) |
| 180 : "m" (*ptr), "0" (new_value) |
| 181 : "memory"); |
| 182 return new_value; // Now it's the previous value. |
| 183 } |
| 184 |
| 185 inline Atomic64 NoBarrier_AtomicIncrement(volatile Atomic64* ptr, |
| 186 Atomic64 increment) { |
| 187 Atomic64 temp = increment; |
| 188 __asm__ __volatile__("lock; xaddq %0,%1" |
| 189 : "+r" (temp), "+m" (*ptr) |
| 190 : : "memory"); |
| 191 // temp now contains the previous value of *ptr |
| 192 return temp + increment; |
| 193 } |
| 194 |
| 195 inline Atomic64 Barrier_AtomicIncrement(volatile Atomic64* ptr, |
| 196 Atomic64 increment) { |
| 197 Atomic64 temp = increment; |
| 198 __asm__ __volatile__("lock; xaddq %0,%1" |
| 199 : "+r" (temp), "+m" (*ptr) |
| 200 : : "memory"); |
| 201 // temp now contains the previous value of *ptr |
| 202 if (AtomicOps_Internalx86CPUFeatures.has_amd_lock_mb_bug) { |
| 203 __asm__ __volatile__("lfence" : : : "memory"); |
| 204 } |
| 205 return temp + increment; |
| 206 } |
| 207 |
| 208 inline void NoBarrier_Store(volatile Atomic64* ptr, Atomic64 value) { |
| 209 *ptr = value; |
| 210 } |
| 211 |
| 212 inline void Acquire_Store(volatile Atomic64* ptr, Atomic64 value) { |
| 213 *ptr = value; |
| 214 MemoryBarrier(); |
| 215 } |
| 216 |
| 217 inline void Release_Store(volatile Atomic64* ptr, Atomic64 value) { |
| 218 ATOMICOPS_COMPILER_BARRIER(); |
| 219 |
| 220 *ptr = value; // An x86 store acts as a release barrier |
| 221 // for current AMD/Intel chips as of Jan 2008. |
| 222 // See also Acquire_Load(), below. |
| 223 |
| 224 // When new chips come out, check: |
| 225 // IA-32 Intel Architecture Software Developer's Manual, Volume 3: |
| 226 // System Programming Guide, Chatper 7: Multiple-processor management, |
| 227 // Section 7.2, Memory Ordering. |
| 228 // Last seen at: |
| 229 // http://developer.intel.com/design/pentium4/manuals/index_new.htm |
| 230 // |
| 231 // x86 stores/loads fail to act as barriers for a few instructions (clflush |
| 232 // maskmovdqu maskmovq movntdq movnti movntpd movntps movntq) but these are |
| 233 // not generated by the compiler, and are rare. Users of these instructions |
| 234 // need to know about cache behaviour in any case since all of these involve |
| 235 // either flushing cache lines or non-temporal cache hints. |
| 236 } |
| 237 |
| 238 inline Atomic64 NoBarrier_Load(volatile const Atomic64* ptr) { |
| 239 return *ptr; |
| 240 } |
| 241 |
| 242 inline Atomic64 Acquire_Load(volatile const Atomic64* ptr) { |
| 243 Atomic64 value = *ptr; // An x86 load acts as a acquire barrier, |
| 244 // for current AMD/Intel chips as of Jan 2008. |
| 245 // See also Release_Store(), above. |
| 246 ATOMICOPS_COMPILER_BARRIER(); |
| 247 return value; |
| 248 } |
| 249 |
| 250 inline Atomic64 Release_Load(volatile const Atomic64* ptr) { |
| 251 MemoryBarrier(); |
| 252 return *ptr; |
| 253 } |
| 254 |
| 255 inline Atomic64 Acquire_CompareAndSwap(volatile Atomic64* ptr, |
| 256 Atomic64 old_value, |
| 257 Atomic64 new_value) { |
| 258 Atomic64 x = NoBarrier_CompareAndSwap(ptr, old_value, new_value); |
| 259 if (AtomicOps_Internalx86CPUFeatures.has_amd_lock_mb_bug) { |
| 260 __asm__ __volatile__("lfence" : : : "memory"); |
| 261 } |
| 262 return x; |
| 263 } |
| 264 |
| 265 inline Atomic64 Release_CompareAndSwap(volatile Atomic64* ptr, |
| 266 Atomic64 old_value, |
| 267 Atomic64 new_value) { |
| 268 return NoBarrier_CompareAndSwap(ptr, old_value, new_value); |
| 269 } |
| 270 |
| 271 #endif // defined(__x86_64__) |
| 272 |
| 273 } // namespace base |
| 274 } // namespace v8 |
| 275 |
| 276 #undef ATOMICOPS_COMPILER_BARRIER |
| 277 |
| 278 #endif // V8_BASE_ATOMICOPS_INTERNALS_X86_GCC_H_ |
OLD | NEW |