deanm@google.com | 1579ec7 | 2008-08-05 18:57:36 +0900 | [diff] [blame^] | 1 | // Copyright 2008, Google Inc. |
| 2 | // All rights reserved. |
| 3 | // |
| 4 | // Redistribution and use in source and binary forms, with or without |
| 5 | // modification, are permitted provided that the following conditions are |
| 6 | // met: |
| 7 | // |
| 8 | // * Redistributions of source code must retain the above copyright |
| 9 | // notice, this list of conditions and the following disclaimer. |
| 10 | // * Redistributions in binary form must reproduce the above |
| 11 | // copyright notice, this list of conditions and the following disclaimer |
| 12 | // in the documentation and/or other materials provided with the |
| 13 | // distribution. |
| 14 | // * Neither the name of Google Inc. nor the names of its |
| 15 | // contributors may be used to endorse or promote products derived from |
| 16 | // this software without specific prior written permission. |
| 17 | // |
| 18 | // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS |
| 19 | // "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT |
| 20 | // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR |
| 21 | // A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT |
| 22 | // OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, |
| 23 | // SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT |
| 24 | // LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, |
| 25 | // DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY |
| 26 | // THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT |
| 27 | // (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE |
| 28 | // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
| 29 | |
| 30 | // This file is an internal atomic implementation, use base/atomicops.h instead. |
| 31 | |
| 32 | #ifndef BASE_ATOMICOPS_INTERNALS_X86_GCC_H_ |
| 33 | #define BASE_ATOMICOPS_INTERNALS_X86_GCC_H_ |
| 34 | |
| 35 | // This struct is not part of the public API of this module; clients may not |
| 36 | // use it. |
| 37 | // Features of this x86. Values may not be correct before main() is run, |
| 38 | // but are set conservatively. |
| 39 | struct AtomicOps_x86CPUFeatureStruct { |
| 40 | bool has_amd_lock_mb_bug; // Processor has AMD memory-barrier bug; do lfence |
| 41 | // after acquire compare-and-swap. |
| 42 | bool has_sse2; // Processor has SSE2. |
| 43 | }; |
| 44 | extern struct AtomicOps_x86CPUFeatureStruct AtomicOps_Internalx86CPUFeatures; |
| 45 | |
| 46 | #define ATOMICOPS_COMPILER_BARRIER() __asm__ __volatile__("" : : : "memory") |
| 47 | |
| 48 | namespace base { |
| 49 | namespace subtle { |
| 50 | |
| 51 | // 32-bit low-level operations on any platform. |
| 52 | |
| 53 | inline Atomic32 NoBarrier_CompareAndSwap(volatile Atomic32* ptr, |
| 54 | Atomic32 old_value, |
| 55 | Atomic32 new_value) { |
| 56 | Atomic32 prev; |
| 57 | __asm__ __volatile__("lock; cmpxchgl %1,%2" |
| 58 | : "=a" (prev) |
| 59 | : "q" (new_value), "m" (*ptr), "0" (old_value) |
| 60 | : "memory"); |
| 61 | return prev; |
| 62 | } |
| 63 | |
| 64 | inline Atomic32 NoBarrier_AtomicExchange(volatile Atomic32* ptr, |
| 65 | Atomic32 new_value) { |
| 66 | __asm__ __volatile__("xchgl %1,%0" // The lock prefix is implicit for xchg. |
| 67 | : "=r" (new_value) |
| 68 | : "m" (*ptr), "0" (new_value) |
| 69 | : "memory"); |
| 70 | return new_value; // Now it's the previous value. |
| 71 | } |
| 72 | |
| 73 | inline Atomic32 NoBarrier_AtomicIncrement(volatile Atomic32* ptr, |
| 74 | Atomic32 increment) { |
| 75 | Atomic32 temp = increment; |
| 76 | __asm__ __volatile__("lock; xaddl %0,%1" |
| 77 | : "+r" (temp), "+m" (*ptr) |
| 78 | : : "memory"); |
| 79 | // temp now holds the old value of *ptr |
| 80 | return temp + increment; |
| 81 | } |
| 82 | |
| 83 | inline Atomic32 Barrier_AtomicIncrement(volatile Atomic32* ptr, |
| 84 | Atomic32 increment) { |
| 85 | Atomic32 temp = increment; |
| 86 | __asm__ __volatile__("lock; xaddl %0,%1" |
| 87 | : "+r" (temp), "+m" (*ptr) |
| 88 | : : "memory"); |
| 89 | // temp now holds the old value of *ptr |
| 90 | if (AtomicOps_Internalx86CPUFeatures.has_amd_lock_mb_bug) { |
| 91 | __asm__ __volatile__("lfence" : : : "memory"); |
| 92 | } |
| 93 | return temp + increment; |
| 94 | } |
| 95 | |
| 96 | inline Atomic32 Acquire_CompareAndSwap(volatile Atomic32* ptr, |
| 97 | Atomic32 old_value, |
| 98 | Atomic32 new_value) { |
| 99 | Atomic32 x = NoBarrier_CompareAndSwap(ptr, old_value, new_value); |
| 100 | if (AtomicOps_Internalx86CPUFeatures.has_amd_lock_mb_bug) { |
| 101 | __asm__ __volatile__("lfence" : : : "memory"); |
| 102 | } |
| 103 | return x; |
| 104 | } |
| 105 | |
| 106 | inline Atomic32 Release_CompareAndSwap(volatile Atomic32* ptr, |
| 107 | Atomic32 old_value, |
| 108 | Atomic32 new_value) { |
| 109 | return NoBarrier_CompareAndSwap(ptr, old_value, new_value); |
| 110 | } |
| 111 | |
| 112 | inline void NoBarrier_Store(volatile Atomic32* ptr, Atomic32 value) { |
| 113 | *ptr = value; |
| 114 | } |
| 115 | |
| 116 | #if defined(__x86_64__) |
| 117 | |
| 118 | // 64-bit implementations of memory barrier can be simpler, because it |
| 119 | // "mfence" is guaranteed to exist. |
| 120 | inline void MemoryBarrier() { |
| 121 | __asm__ __volatile__("mfence" : : : "memory"); |
| 122 | } |
| 123 | |
| 124 | inline void Acquire_Store(volatile Atomic32* ptr, Atomic32 value) { |
| 125 | *ptr = value; |
| 126 | MemoryBarrier(); |
| 127 | } |
| 128 | |
| 129 | #else |
| 130 | |
| 131 | inline void MemoryBarrier() { |
| 132 | if (AtomicOps_Internalx86CPUFeatures.has_sse2) { |
| 133 | __asm__ __volatile__("mfence" : : : "memory"); |
| 134 | } else { // mfence is faster but not present on PIII |
| 135 | Atomic32 x = 0; |
| 136 | NoBarrier_AtomicExchange(&x, 0); // acts as a barrier on PIII |
| 137 | } |
| 138 | } |
| 139 | |
| 140 | inline void Acquire_Store(volatile Atomic32* ptr, Atomic32 value) { |
| 141 | if (AtomicOps_Internalx86CPUFeatures.has_sse2) { |
| 142 | *ptr = value; |
| 143 | __asm__ __volatile__("mfence" : : : "memory"); |
| 144 | } else { |
| 145 | NoBarrier_AtomicExchange(ptr, value); |
| 146 | // acts as a barrier on PIII |
| 147 | } |
| 148 | } |
| 149 | #endif |
| 150 | |
| 151 | inline void Release_Store(volatile Atomic32* ptr, Atomic32 value) { |
| 152 | ATOMICOPS_COMPILER_BARRIER(); |
| 153 | *ptr = value; // An x86 store acts as a release barrier. |
| 154 | // See comments in Atomic64 version of Release_Store(), below. |
| 155 | } |
| 156 | |
| 157 | inline Atomic32 NoBarrier_Load(volatile const Atomic32* ptr) { |
| 158 | return *ptr; |
| 159 | } |
| 160 | |
| 161 | inline Atomic32 Acquire_Load(volatile const Atomic32* ptr) { |
| 162 | Atomic32 value = *ptr; // An x86 load acts as a acquire barrier. |
| 163 | // See comments in Atomic64 version of Release_Store(), below. |
| 164 | ATOMICOPS_COMPILER_BARRIER(); |
| 165 | return value; |
| 166 | } |
| 167 | |
| 168 | inline Atomic32 Release_Load(volatile const Atomic32* ptr) { |
| 169 | MemoryBarrier(); |
| 170 | return *ptr; |
| 171 | } |
| 172 | |
| 173 | #if defined(__x86_64__) |
| 174 | |
| 175 | // 64-bit low-level operations on 64-bit platform. |
| 176 | |
| 177 | inline Atomic64 NoBarrier_CompareAndSwap(volatile Atomic64* ptr, |
| 178 | Atomic64 old_value, |
| 179 | Atomic64 new_value) { |
| 180 | Atomic64 prev; |
| 181 | __asm__ __volatile__("lock; cmpxchgq %1,%2" |
| 182 | : "=a" (prev) |
| 183 | : "q" (new_value), "m" (*ptr), "0" (old_value) |
| 184 | : "memory"); |
| 185 | return prev; |
| 186 | } |
| 187 | |
| 188 | inline Atomic64 NoBarrier_AtomicExchange(volatile Atomic64* ptr, |
| 189 | Atomic64 new_value) { |
| 190 | __asm__ __volatile__("xchgq %1,%0" // The lock prefix is implicit for xchg. |
| 191 | : "=r" (new_value) |
| 192 | : "m" (*ptr), "0" (new_value) |
| 193 | : "memory"); |
| 194 | return new_value; // Now it's the previous value. |
| 195 | } |
| 196 | |
| 197 | inline Atomic64 NoBarrier_AtomicIncrement(volatile Atomic64* ptr, |
| 198 | Atomic64 increment) { |
| 199 | Atomic64 temp = increment; |
| 200 | __asm__ __volatile__("lock; xaddq %0,%1" |
| 201 | : "+r" (temp), "+m" (*ptr) |
| 202 | : : "memory"); |
| 203 | // temp now contains the previous value of *ptr |
| 204 | return temp + increment; |
| 205 | } |
| 206 | |
| 207 | inline Atomic64 Barrier_AtomicIncrement(volatile Atomic64* ptr, |
| 208 | Atomic64 increment) { |
| 209 | Atomic64 temp = increment; |
| 210 | __asm__ __volatile__("lock; xaddq %0,%1" |
| 211 | : "+r" (temp), "+m" (*ptr) |
| 212 | : : "memory"); |
| 213 | // temp now contains the previous value of *ptr |
| 214 | if (AtomicOps_Internalx86CPUFeatures.has_amd_lock_mb_bug) { |
| 215 | __asm__ __volatile__("lfence" : : : "memory"); |
| 216 | } |
| 217 | return temp + increment; |
| 218 | } |
| 219 | |
| 220 | inline void NoBarrier_Store(volatile Atomic64* ptr, Atomic64 value) { |
| 221 | *ptr = value; |
| 222 | } |
| 223 | |
| 224 | inline void Acquire_Store(volatile Atomic64* ptr, Atomic64 value) { |
| 225 | *ptr = value; |
| 226 | MemoryBarrier(); |
| 227 | } |
| 228 | |
| 229 | inline void Release_Store(volatile Atomic64* ptr, Atomic64 value) { |
| 230 | ATOMICOPS_COMPILER_BARRIER(); |
| 231 | |
| 232 | *ptr = value; // An x86 store acts as a release barrier |
| 233 | // for current AMD/Intel chips as of Jan 2008. |
| 234 | // See also Acquire_Load(), below. |
| 235 | |
| 236 | // When new chips come out, check: |
| 237 | // IA-32 Intel Architecture Software Developer's Manual, Volume 3: |
| 238 | // System Programming Guide, Chatper 7: Multiple-processor management, |
| 239 | // Section 7.2, Memory Ordering. |
| 240 | // Last seen at: |
| 241 | // http://developer.intel.com/design/pentium4/manuals/index_new.htm |
| 242 | // |
| 243 | // x86 stores/loads fail to act as barriers for a few instructions (clflush |
| 244 | // maskmovdqu maskmovq movntdq movnti movntpd movntps movntq) but these are |
| 245 | // not generated by the compiler, and are rare. Users of these instructions |
| 246 | // need to know about cache behaviour in any case since all of these involve |
| 247 | // either flushing cache lines or non-temporal cache hints. |
| 248 | } |
| 249 | |
| 250 | inline Atomic64 NoBarrier_Load(volatile const Atomic64* ptr) { |
| 251 | return *ptr; |
| 252 | } |
| 253 | |
| 254 | inline Atomic64 Acquire_Load(volatile const Atomic64* ptr) { |
| 255 | Atomic64 value = *ptr; // An x86 load acts as a acquire barrier, |
| 256 | // for current AMD/Intel chips as of Jan 2008. |
| 257 | // See also Release_Store(), above. |
| 258 | ATOMICOPS_COMPILER_BARRIER(); |
| 259 | return value; |
| 260 | } |
| 261 | |
| 262 | inline Atomic64 Release_Load(volatile const Atomic64* ptr) { |
| 263 | MemoryBarrier(); |
| 264 | return *ptr; |
| 265 | } |
| 266 | #endif // defined(__x86_64__) |
| 267 | |
| 268 | } // namespace base::subtle |
| 269 | } // namespace base |
| 270 | |
| 271 | #undef ATOMICOPS_COMPILER_BARRIER |
| 272 | |
| 273 | #endif // BASE_ATOMICOPS_INTERNALS_X86_GCC_H_ |