1 /*
2 * Copyright (c) 1999, 2017, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #ifndef SHARE_VM_RUNTIME_ATOMIC_HPP
26 #define SHARE_VM_RUNTIME_ATOMIC_HPP
27
28 #include "memory/allocation.hpp"
29 #include "metaprogramming/integerTypes.hpp"
30 #include "metaprogramming/isIntegral.hpp"
31 #include "metaprogramming/isPointer.hpp"
32 #include "utilities/align.hpp"
33 #include "utilities/debug.hpp"
34 #include "utilities/macros.hpp"
35
36 enum cmpxchg_memory_order {
37 memory_order_relaxed,
38 // Use value which doesn't interfere with C++2011. We need to be more conservative.
39 memory_order_conservative = 8
40 };
41
42 class Atomic : AllStatic {
43 template<typename T> class Never: public FalseType {};
44
45 template <typename T>
46 inline static void specialized_store(T store_value, volatile T* dest) {
47 STATIC_ASSERT(sizeof(T) <= size_t(BytesPerWord)); // Does the machine support atomic wide accesses?
48 (void)const_cast<T&>(*dest = store_value);
49 }
50
51 template <typename T>
52 inline static T specialized_load(const volatile T* dest) {
53 STATIC_ASSERT(sizeof(T) <= size_t(BytesPerWord)); // Does the machine support atomic wide accesses?
54 return *dest;
55 }
56
57 template <typename T>
58 inline static T specialized_add(T add_value, volatile T* dest) {
59 STATIC_ASSERT(Never<T>::value);
60 return add_value;
61 }
62
63 template <typename T>
64 inline static void specialized_inc(volatile T* dest) {
65 add(1, dest);
66 }
67
68 template <typename T>
69 inline static void specialized_dec(volatile T* dest) {
70 add(-1, dest);
71 }
72
73 template <typename T>
74 inline static T specialized_xchg(T exchange_value, volatile T* dest) {
75 STATIC_ASSERT(Never<T>::value);
76 return exchange_value;
77 }
78
79 template <typename T>
80 inline static T specialized_cmpxchg(T exchange_value, volatile T* dest, T compare_value, cmpxchg_memory_order order) {
81 STATIC_ASSERT(Never<T>::value);
82 return exchange_value;
83 }
84
85 public:
86 // Atomic operations on 64-bit types are not available on all 32-bit
87 // platforms. If atomic ops on 64-bit types are defined here they must only
88 // be used from code that verifies they are available at runtime and
89 // can provide an alternative action if not - see supports_cx8() for
90 // a means to test availability.
91
92 // The memory operations that are mentioned with each of the atomic
93 // function families come from src/share/vm/runtime/orderAccess.hpp,
94 // e.g., <fence> is described in that file and is implemented by the
95 // OrderAccess::fence() function. See that file for the gory details
96 // on the Memory Access Ordering Model.
97
98 // All of the atomic operations that imply a read-modify-write action
99 // guarantee a two-way memory barrier across that operation. Historically
100 // these semantics reflect the strength of atomic operations that are
101 // provided on SPARC/X86. We assume that strength is necessary unless
102 // we can prove that a weaker form is sufficiently safe.
103
104 // Atomically store to a location
105 // See comment above about using 64-bit atomics on 32-bit platforms
106 template <typename T, typename U>
107 inline static void store(T store_value, volatile U* dest);
108
109 // The store_ptr() member functions are deprecated. Use store() instead.
110 static void store_ptr(intptr_t store_value, volatile intptr_t* dest) {
111 store(store_value, dest);
112 }
113
114 static void store_ptr(void* store_value, volatile void* dest) {
115 store((intptr_t)store_value, (volatile intptr_t*)dest);
116 }
117
118 // Atomically load from a location
119 // See comment above about using 64-bit atomics on 32-bit platforms
120 template <typename T>
121 inline static T load(volatile T* src);
122
123 // Atomically add to a location. Returns updated value. add*() provide:
124 // <fence> add-value-to-dest <membar StoreLoad|StoreStore>
125 // add(I1 v, I* d)
126 // add(I1 v, P* d)
127 // where I, I1 are integral types, P is a pointer type.
128 // Functional behavior is modelled on *dest += add_value.
129 template <typename T, typename U>
130 inline static U add(T add_value, volatile U* dst);
131
132 template <typename T, typename U>
133 inline static U* add(T add_value, U* volatile* dst);
134
135 // The add_ptr() member functions are deprecated. Use add() instead.
136 static intptr_t add_ptr(intptr_t add_value, volatile intptr_t* dest) {
137 return add(add_value, dest);
138 }
139
140 static void* add_ptr(intptr_t add_value, volatile void* dest) {
141 return (void*)add(add_value, (volatile intptr_t*)dest);
142 }
143
144 // Atomically increment location. inc*() provide:
145 // <fence> increment-dest <membar StoreLoad|StoreStore>
146 // Functional behavior is modelled on *dest++
147 template <typename T>
148 inline static void inc(volatile T* dest);
149
150 template <typename T>
151 inline static void inc(T* volatile* dest);
152
153 // The inc_ptr member functions are deprecated. Use inc() instead.
154 static void inc_ptr(volatile intptr_t* dest) {
155 inc(dest);
156 }
157
158 static void inc_ptr(volatile void* dest) {
159 inc((volatile intptr_t*)dest);
160 }
161
162 // Atomically decrement a location. dec*() provide:
163 // <fence> decrement-dest <membar StoreLoad|StoreStore>
164 // Functional behavior is modelled on *dest--
165 template <typename T>
166 inline static void dec(volatile T* dest);
167
168 template <typename T>
169 inline static void dec(T* volatile* dest);
170
171 // The dec_ptr member functions are deprecated. Use dec() instead.
172 static void dec_ptr(volatile intptr_t* dest) {
173 dec(dest);
174 }
175
176 static void dec_ptr(volatile void* dest) {
177 dec((volatile intptr_t*)dest);
178 }
179
180 // Performs atomic exchange of *dest with exchange_value. Returns old
181 // prior value of *dest. xchg*() provide:
182 // <fence> exchange-value-with-dest <membar StoreLoad|StoreStore>
183 template <typename T, typename U>
184 inline static U xchg(T exchange_value, volatile U* dest);
185
186 // The xchg_ptr() member functions are deprecated. Use xchg() instead.
187 static intptr_t xchg_ptr(intptr_t exchange_value, volatile intptr_t* dest) {
188 return xchg(exchange_value, dest);
189 }
190
191 static void* xchg_ptr(void* exchange_value, volatile void* dest) {
192 return (void*)xchg((intptr_t)exchange_value, (volatile intptr_t*)dest);
193 }
194
195 // Performs atomic compare of *dest and compare_value, and exchanges
196 // *dest with exchange_value if the comparison succeeded. Returns prior
197 // value of *dest. cmpxchg*() provide:
198 // <fence> compare-and-exchange <membar StoreLoad|StoreStore>
199 // See comment above about using 64-bit atomics on 32-bit platforms
200 template <typename T, typename U, typename V>
201 inline static U cmpxchg(T exchange_value, volatile U* dest, V compare_value, cmpxchg_memory_order order = memory_order_conservative);
202
203 // The cmpxchg_ptr member functions are deprecated. Use cmpxchg() instead.
204 inline static intptr_t cmpxchg_ptr(intptr_t exchange_value, volatile intptr_t* dest,
205 intptr_t compare_value, cmpxchg_memory_order order = memory_order_conservative) {
206 return cmpxchg(exchange_value, dest, compare_value, order);
207
208 }
209
210 inline static void* cmpxchg_ptr(void* exchange_value, volatile void* dest,
211 void* compare_value, cmpxchg_memory_order order = memory_order_conservative) {
212 return (void*)cmpxchg((intptr_t)exchange_value, (volatile intptr_t*)dest, (intptr_t)compare_value, order);
213 }
214 };
215
216 // internal implementation
217
218 template <typename T, typename U>
219 inline void Atomic::store(T store_value, volatile U* dest) {
220 typedef typename IntegerTypes::Signed<U>::type Raw;
221 U store_value_cast = store_value;
222 specialized_store(IntegerTypes::cast_to_signed(store_value_cast), reinterpret_cast<volatile Raw*>(dest));
223 }
224
225 template <typename T>
226 inline T Atomic::load(volatile T* src) {
227 typedef typename IntegerTypes::Signed<T>::type Raw;
228 return IntegerTypes::cast<T>(specialized_load(reinterpret_cast<const volatile Raw*>(src)));
229 }
230
231 template <typename T, typename U>
232 inline U Atomic::add(T add_value, volatile U* dst) {
233 STATIC_ASSERT(IsIntegral<T>::value);
234 STATIC_ASSERT(IsIntegral<U>::value);
235 typedef typename IntegerTypes::Signed<U>::type Raw;
236 // Allow -Wconversion or the like to complain about unsafe conversions.
237 U value = add_value;
238 Raw raw_value = IntegerTypes::cast_to_signed(value);
239 Raw result = specialized_add(raw_value, reinterpret_cast<volatile Raw*>(dst));
240 return IntegerTypes::cast<U>(result);
241 }
242
243 template <typename T, typename U>
244 inline U* Atomic::add(T add_value, U* volatile* dst) {
245 STATIC_ASSERT(IsIntegral<T>::value);
246 typedef typename IntegerTypes::Signed<U*>::type Raw;
247 ptrdiff_t value = add_value;
248 Raw raw_value = IntegerTypes::cast_to_signed(value * sizeof(U));
249 Raw result = specialized_add(raw_value, reinterpret_cast<volatile Raw*>(dst));
250 return IntegerTypes::cast<U*>(result);
251 }
252
253 template <typename T>
254 inline void Atomic::inc(volatile T* src) {
255 STATIC_ASSERT(IsIntegral<T>::value);
256 typedef typename IntegerTypes::Signed<T>::type Raw;
257 specialized_inc(reinterpret_cast<volatile Raw*>(src));
258 }
259
260 template <typename T>
261 inline void Atomic::inc(T* volatile* src) {
262 if (sizeof(T) != 1) {
263 add(1, src);
264 } else {
265 typedef typename IntegerTypes::Signed<T*>::type Raw;
266 specialized_inc(reinterpret_cast<volatile Raw*>(src));
267 }
268 }
269
270 template <typename T>
271 inline void Atomic::dec(volatile T* src) {
272 STATIC_ASSERT(IsIntegral<T>::value);
273 typedef typename IntegerTypes::Signed<T>::type Raw;
274 specialized_dec(reinterpret_cast<volatile Raw*>(src));
275 }
276
277 template <typename T>
278 inline void Atomic::dec(T* volatile* src) {
279 if (sizeof(T) != 1) {
280 add(-1, src);
281 } else {
282 typedef typename IntegerTypes::Signed<T*>::type Raw;
283 specialized_dec(reinterpret_cast<volatile Raw*>(src));
284 }
285 }
286
287 template <typename T, typename U>
288 inline U Atomic::xchg(T exchange_value, volatile U* dest) {
289 typedef typename IntegerTypes::Signed<U>::type Raw;
290 U exchange_value_cast = exchange_value;
291 Raw result = specialized_xchg(IntegerTypes::cast_to_signed(exchange_value_cast),
292 reinterpret_cast<volatile Raw*>(dest));
293 return IntegerTypes::cast<U>(result);
294 }
295
296 template <typename T, typename U, typename V>
297 inline U Atomic::cmpxchg(T exchange_value, volatile U* dest, V compare_value, cmpxchg_memory_order order) {
298 typedef typename IntegerTypes::Signed<U>::type Raw;
299 U exchange_value_cast = exchange_value;
300 U compare_value_cast = compare_value;
301 Raw result = specialized_cmpxchg(IntegerTypes::cast_to_signed(exchange_value_cast),
302 reinterpret_cast<volatile Raw*>(dest),
303 IntegerTypes::cast_to_signed(compare_value_cast), order);
304 return IntegerTypes::cast<U>(result);
305 }
306
307 // platform specific in-line definitions - must come before shared definitions
308
309 #include OS_CPU_HEADER(atomic)
310
311 // shared in-line definitions
312
313 #ifndef VM_HAS_SPECIALIZED_CMPXCHG_BYTE
314 /*
315 * This is the default implementation of byte-sized cmpxchg. It emulates 8-bit-sized cmpxchg
316 * in terms of 32-bit-sized cmpxchg. Platforms may override this by defining their own inline definition
317 * as well as defining VM_HAS_SPECIALIZED_CMPXCHG_BYTE. This will cause the platform specific
318 * implementation to be used instead.
319 */
320 template <>
321 inline int8_t Atomic::specialized_cmpxchg<int8_t>(int8_t exchange_value, volatile int8_t* dest,
322 int8_t compare_value, cmpxchg_memory_order order) {
323 volatile int32_t* dest_int =
324 reinterpret_cast<volatile int32_t*>(align_down(dest, sizeof(int32_t)));
325 size_t offset = pointer_delta(dest, dest_int, 1);
326 int32_t cur = *dest_int;
327 int8_t* cur_as_bytes = reinterpret_cast<int8_t*>(&cur);
328
329 // current value may not be what we are looking for, so force it
330 // to that value so the initial cmpxchg will fail if it is different
331 cur_as_bytes[offset] = compare_value;
332
333 // always execute a real cmpxchg so that we get the required memory
334 // barriers even on initial failure
335 do {
336 // value to swap in matches current value ...
337 int32_t new_value = cur;
338 // ... except for the one byte we want to update
339 reinterpret_cast<int8_t*>(&new_value)[offset] = exchange_value;
340
341 int32_t res = cmpxchg(new_value, dest_int, cur, order);
342 if (res == cur) break; // success
343
344 // at least one byte in the int changed value, so update
345 // our view of the current int
346 cur = res;
347 // if our byte is still as cur we loop and try again
348 } while (cur_as_bytes[offset] == compare_value);
349
350 return cur_as_bytes[offset];
351 }
352
353 #endif // VM_HAS_SPECIALIZED_CMPXCHG_BYTE
354
355 template <>
356 inline int16_t Atomic::specialized_add<int16_t>(int16_t add_value, volatile int16_t* dest) {
357 // Most platforms do not support atomic add on a 2-byte value. However,
358 // if the value occupies the most significant 16 bits of an aligned 32-bit
359 // word, then we can do this with an atomic add of (add_value << 16)
360 // to the 32-bit word.
361 //
362 // The least significant parts of this 32-bit word will never be affected, even
363 // in case of overflow/underflow.
364 //
365 // Use the ATOMIC_SHORT_PAIR macro (see macros.hpp) to get the desired alignment.
366 #ifdef VM_LITTLE_ENDIAN
367 assert((intx(dest) & 0x03) == 0x02, "wrong alignment");
368 int32_t new_value = Atomic::add(int32_t(add_value) << 16, (volatile int32_t*)(dest-1));
369 #else
370 assert((intx(dest) & 0x03) == 0x00, "wrong alignment");
371 int32_t new_value = Atomic::add(int32_t(add_value) << 16, (volatile int32_t*)(dest));
372 #endif
373 return (int16_t)(new_value >> 16); // preserves sign
374 }
375
376 template <>
377 inline void Atomic::specialized_inc<int16_t>(volatile int16_t* dest) {
378 (void)add(int16_t(1), dest);
379 }
380
381 template <>
382 inline void Atomic::specialized_dec<int16_t>(volatile int16_t* dest) {
383 (void)add(int16_t(-1), dest);
384 }
385
386 #endif // SHARE_VM_RUNTIME_ATOMIC_HPP