ipc/chromium/src/base/atomicops_internals_x86_gcc.h

Wed, 31 Dec 2014 06:09:35 +0100

author
Michael Schloh von Bennewitz <michael@schloh.com>
date
Wed, 31 Dec 2014 06:09:35 +0100
changeset 0
6474c204b198
permissions
-rw-r--r--

Cloned upstream origin tor-browser at tor-browser-31.3.0esr-4.5-1-build1
revision ID fc1c9ff7c1b2defdbc039f12214767608f46423f for hacking purpose.

michael@0 1 // Copyright (c) 2006-2008 The Chromium Authors. All rights reserved.
michael@0 2 // Use of this source code is governed by a BSD-style license that can be
michael@0 3 // found in the LICENSE file.
michael@0 4
michael@0 5 // This file is an internal atomic implementation, use base/atomicops.h instead.
michael@0 6
michael@0 7 #ifndef BASE_ATOMICOPS_INTERNALS_X86_GCC_H_
michael@0 8 #define BASE_ATOMICOPS_INTERNALS_X86_GCC_H_
michael@0 9
michael@0 10 // This struct is not part of the public API of this module; clients may not
michael@0 11 // use it.
michael@0 12 // Features of this x86. Values may not be correct before main() is run,
michael@0 13 // but are set conservatively.
michael@0 14 struct AtomicOps_x86CPUFeatureStruct {
michael@0 15 bool has_amd_lock_mb_bug; // Processor has AMD memory-barrier bug; do lfence
michael@0 16 // after acquire compare-and-swap.
michael@0 17 bool has_sse2; // Processor has SSE2.
michael@0 18 };
michael@0 19 extern struct AtomicOps_x86CPUFeatureStruct AtomicOps_Internalx86CPUFeatures;
michael@0 20
michael@0 21 #define ATOMICOPS_COMPILER_BARRIER() __asm__ __volatile__("" : : : "memory")
michael@0 22
michael@0 23 namespace base {
michael@0 24 namespace subtle {
michael@0 25
michael@0 26 // 32-bit low-level operations on any platform.
michael@0 27
michael@0 28 inline Atomic32 NoBarrier_CompareAndSwap(volatile Atomic32* ptr,
michael@0 29 Atomic32 old_value,
michael@0 30 Atomic32 new_value) {
michael@0 31 Atomic32 prev;
michael@0 32 __asm__ __volatile__("lock; cmpxchgl %1,%2"
michael@0 33 : "=a" (prev)
michael@0 34 : "q" (new_value), "m" (*ptr), "0" (old_value)
michael@0 35 : "memory");
michael@0 36 return prev;
michael@0 37 }
michael@0 38
michael@0 39 inline Atomic32 NoBarrier_AtomicExchange(volatile Atomic32* ptr,
michael@0 40 Atomic32 new_value) {
michael@0 41 __asm__ __volatile__("xchgl %1,%0" // The lock prefix is implicit for xchg.
michael@0 42 : "=r" (new_value)
michael@0 43 : "m" (*ptr), "0" (new_value)
michael@0 44 : "memory");
michael@0 45 return new_value; // Now it's the previous value.
michael@0 46 }
michael@0 47
michael@0 48 inline Atomic32 NoBarrier_AtomicIncrement(volatile Atomic32* ptr,
michael@0 49 Atomic32 increment) {
michael@0 50 Atomic32 temp = increment;
michael@0 51 __asm__ __volatile__("lock; xaddl %0,%1"
michael@0 52 : "+r" (temp), "+m" (*ptr)
michael@0 53 : : "memory");
michael@0 54 // temp now holds the old value of *ptr
michael@0 55 return temp + increment;
michael@0 56 }
michael@0 57
michael@0 58 inline Atomic32 Barrier_AtomicIncrement(volatile Atomic32* ptr,
michael@0 59 Atomic32 increment) {
michael@0 60 Atomic32 temp = increment;
michael@0 61 __asm__ __volatile__("lock; xaddl %0,%1"
michael@0 62 : "+r" (temp), "+m" (*ptr)
michael@0 63 : : "memory");
michael@0 64 // temp now holds the old value of *ptr
michael@0 65 if (AtomicOps_Internalx86CPUFeatures.has_amd_lock_mb_bug) {
michael@0 66 __asm__ __volatile__("lfence" : : : "memory");
michael@0 67 }
michael@0 68 return temp + increment;
michael@0 69 }
michael@0 70
michael@0 71 inline Atomic32 Acquire_CompareAndSwap(volatile Atomic32* ptr,
michael@0 72 Atomic32 old_value,
michael@0 73 Atomic32 new_value) {
michael@0 74 Atomic32 x = NoBarrier_CompareAndSwap(ptr, old_value, new_value);
michael@0 75 if (AtomicOps_Internalx86CPUFeatures.has_amd_lock_mb_bug) {
michael@0 76 __asm__ __volatile__("lfence" : : : "memory");
michael@0 77 }
michael@0 78 return x;
michael@0 79 }
michael@0 80
michael@0 81 inline Atomic32 Release_CompareAndSwap(volatile Atomic32* ptr,
michael@0 82 Atomic32 old_value,
michael@0 83 Atomic32 new_value) {
michael@0 84 return NoBarrier_CompareAndSwap(ptr, old_value, new_value);
michael@0 85 }
michael@0 86
michael@0 87 inline void NoBarrier_Store(volatile Atomic32* ptr, Atomic32 value) {
michael@0 88 *ptr = value;
michael@0 89 }
michael@0 90
michael@0 91 #if defined(__x86_64__)
michael@0 92
michael@0 93 // 64-bit implementations of memory barrier can be simpler, because it
michael@0 94 // "mfence" is guaranteed to exist.
michael@0 95 inline void MemoryBarrier() {
michael@0 96 __asm__ __volatile__("mfence" : : : "memory");
michael@0 97 }
michael@0 98
michael@0 99 inline void Acquire_Store(volatile Atomic32* ptr, Atomic32 value) {
michael@0 100 *ptr = value;
michael@0 101 MemoryBarrier();
michael@0 102 }
michael@0 103
michael@0 104 #else
michael@0 105
michael@0 106 inline void MemoryBarrier() {
michael@0 107 if (AtomicOps_Internalx86CPUFeatures.has_sse2) {
michael@0 108 __asm__ __volatile__("mfence" : : : "memory");
michael@0 109 } else { // mfence is faster but not present on PIII
michael@0 110 Atomic32 x = 0;
michael@0 111 NoBarrier_AtomicExchange(&x, 0); // acts as a barrier on PIII
michael@0 112 }
michael@0 113 }
michael@0 114
michael@0 115 inline void Acquire_Store(volatile Atomic32* ptr, Atomic32 value) {
michael@0 116 if (AtomicOps_Internalx86CPUFeatures.has_sse2) {
michael@0 117 *ptr = value;
michael@0 118 __asm__ __volatile__("mfence" : : : "memory");
michael@0 119 } else {
michael@0 120 NoBarrier_AtomicExchange(ptr, value);
michael@0 121 // acts as a barrier on PIII
michael@0 122 }
michael@0 123 }
michael@0 124 #endif
michael@0 125
michael@0 126 inline void Release_Store(volatile Atomic32* ptr, Atomic32 value) {
michael@0 127 ATOMICOPS_COMPILER_BARRIER();
michael@0 128 *ptr = value; // An x86 store acts as a release barrier.
michael@0 129 // See comments in Atomic64 version of Release_Store(), below.
michael@0 130 }
michael@0 131
michael@0 132 inline Atomic32 NoBarrier_Load(volatile const Atomic32* ptr) {
michael@0 133 return *ptr;
michael@0 134 }
michael@0 135
michael@0 136 inline Atomic32 Acquire_Load(volatile const Atomic32* ptr) {
michael@0 137 Atomic32 value = *ptr; // An x86 load acts as a acquire barrier.
michael@0 138 // See comments in Atomic64 version of Release_Store(), below.
michael@0 139 ATOMICOPS_COMPILER_BARRIER();
michael@0 140 return value;
michael@0 141 }
michael@0 142
michael@0 143 inline Atomic32 Release_Load(volatile const Atomic32* ptr) {
michael@0 144 MemoryBarrier();
michael@0 145 return *ptr;
michael@0 146 }
michael@0 147
michael@0 148 #if defined(__x86_64__)
michael@0 149
michael@0 150 // 64-bit low-level operations on 64-bit platform.
michael@0 151
michael@0 152 inline Atomic64 NoBarrier_CompareAndSwap(volatile Atomic64* ptr,
michael@0 153 Atomic64 old_value,
michael@0 154 Atomic64 new_value) {
michael@0 155 Atomic64 prev;
michael@0 156 __asm__ __volatile__("lock; cmpxchgq %1,%2"
michael@0 157 : "=a" (prev)
michael@0 158 : "q" (new_value), "m" (*ptr), "0" (old_value)
michael@0 159 : "memory");
michael@0 160 return prev;
michael@0 161 }
michael@0 162
michael@0 163 inline Atomic64 NoBarrier_AtomicExchange(volatile Atomic64* ptr,
michael@0 164 Atomic64 new_value) {
michael@0 165 __asm__ __volatile__("xchgq %1,%0" // The lock prefix is implicit for xchg.
michael@0 166 : "=r" (new_value)
michael@0 167 : "m" (*ptr), "0" (new_value)
michael@0 168 : "memory");
michael@0 169 return new_value; // Now it's the previous value.
michael@0 170 }
michael@0 171
michael@0 172 inline Atomic64 NoBarrier_AtomicIncrement(volatile Atomic64* ptr,
michael@0 173 Atomic64 increment) {
michael@0 174 Atomic64 temp = increment;
michael@0 175 __asm__ __volatile__("lock; xaddq %0,%1"
michael@0 176 : "+r" (temp), "+m" (*ptr)
michael@0 177 : : "memory");
michael@0 178 // temp now contains the previous value of *ptr
michael@0 179 return temp + increment;
michael@0 180 }
michael@0 181
michael@0 182 inline Atomic64 Barrier_AtomicIncrement(volatile Atomic64* ptr,
michael@0 183 Atomic64 increment) {
michael@0 184 Atomic64 temp = increment;
michael@0 185 __asm__ __volatile__("lock; xaddq %0,%1"
michael@0 186 : "+r" (temp), "+m" (*ptr)
michael@0 187 : : "memory");
michael@0 188 // temp now contains the previous value of *ptr
michael@0 189 if (AtomicOps_Internalx86CPUFeatures.has_amd_lock_mb_bug) {
michael@0 190 __asm__ __volatile__("lfence" : : : "memory");
michael@0 191 }
michael@0 192 return temp + increment;
michael@0 193 }
michael@0 194
michael@0 195 inline void NoBarrier_Store(volatile Atomic64* ptr, Atomic64 value) {
michael@0 196 *ptr = value;
michael@0 197 }
michael@0 198
michael@0 199 inline Atomic64 Acquire_CompareAndSwap(volatile Atomic64* ptr,
michael@0 200 Atomic64 old_value,
michael@0 201 Atomic64 new_value) {
michael@0 202 Atomic64 x = NoBarrier_CompareAndSwap(ptr, old_value, new_value);
michael@0 203 /* XXX/cjones: no idea if this is necessary... */
michael@0 204 if (AtomicOps_Internalx86CPUFeatures.has_amd_lock_mb_bug) {
michael@0 205 __asm__ __volatile__("lfence" : : : "memory");
michael@0 206 }
michael@0 207 return x;
michael@0 208 }
michael@0 209
michael@0 210 inline void Acquire_Store(volatile Atomic64* ptr, Atomic64 value) {
michael@0 211 *ptr = value;
michael@0 212 MemoryBarrier();
michael@0 213 }
michael@0 214
michael@0 215 inline void Release_Store(volatile Atomic64* ptr, Atomic64 value) {
michael@0 216 ATOMICOPS_COMPILER_BARRIER();
michael@0 217
michael@0 218 *ptr = value; // An x86 store acts as a release barrier
michael@0 219 // for current AMD/Intel chips as of Jan 2008.
michael@0 220 // See also Acquire_Load(), below.
michael@0 221
michael@0 222 // When new chips come out, check:
michael@0 223 // IA-32 Intel Architecture Software Developer's Manual, Volume 3:
michael@0 224 // System Programming Guide, Chatper 7: Multiple-processor management,
michael@0 225 // Section 7.2, Memory Ordering.
michael@0 226 // Last seen at:
michael@0 227 // http://developer.intel.com/design/pentium4/manuals/index_new.htm
michael@0 228 //
michael@0 229 // x86 stores/loads fail to act as barriers for a few instructions (clflush
michael@0 230 // maskmovdqu maskmovq movntdq movnti movntpd movntps movntq) but these are
michael@0 231 // not generated by the compiler, and are rare. Users of these instructions
michael@0 232 // need to know about cache behaviour in any case since all of these involve
michael@0 233 // either flushing cache lines or non-temporal cache hints.
michael@0 234 }
michael@0 235
michael@0 236 inline Atomic64 NoBarrier_Load(volatile const Atomic64* ptr) {
michael@0 237 return *ptr;
michael@0 238 }
michael@0 239
michael@0 240 inline Atomic64 Acquire_Load(volatile const Atomic64* ptr) {
michael@0 241 Atomic64 value = *ptr; // An x86 load acts as a acquire barrier,
michael@0 242 // for current AMD/Intel chips as of Jan 2008.
michael@0 243 // See also Release_Store(), above.
michael@0 244 ATOMICOPS_COMPILER_BARRIER();
michael@0 245 return value;
michael@0 246 }
michael@0 247
michael@0 248 inline Atomic64 Release_Load(volatile const Atomic64* ptr) {
michael@0 249 MemoryBarrier();
michael@0 250 return *ptr;
michael@0 251 }
michael@0 252 #endif // defined(__x86_64__)
michael@0 253
michael@0 254 } // namespace base::subtle
michael@0 255 } // namespace base
michael@0 256
michael@0 257 #undef ATOMICOPS_COMPILER_BARRIER
michael@0 258
michael@0 259 #endif // BASE_ATOMICOPS_INTERNALS_X86_GCC_H_

mercurial