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