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/conditional.hpp"
  30 #include "metaprogramming/enableIf.hpp"
  31 #include "metaprogramming/isIntegral.hpp"
  32 #include "metaprogramming/isPointer.hpp"
  33 #include "metaprogramming/isSame.hpp"
  34 #include "metaprogramming/primitiveConversions.hpp"
  35 #include "metaprogramming/removeCV.hpp"
  36 #include "metaprogramming/removePointer.hpp"
  37 #include "utilities/align.hpp"
  38 #include "utilities/macros.hpp"
  39 
  40 enum cmpxchg_memory_order {
  41   memory_order_relaxed,
  42   // Use value which doesn't interfere with C++2011. We need to be more conservative.
  43   memory_order_conservative = 8
  44 };
  45 
  46 class Atomic : AllStatic {
  47 public:
  48   // Atomic operations on jlong types are not available on all 32-bit
  49   // platforms. If atomic ops on jlongs are defined here they must only
  50   // be used from code that verifies they are available at runtime and
  51   // can provide an alternative action if not - see supports_cx8() for
  52   // a means to test availability.
  53 
  54   // The memory operations that are mentioned with each of the atomic
  55   // function families come from src/share/vm/runtime/orderAccess.hpp,
  56   // e.g., <fence> is described in that file and is implemented by the
  57   // OrderAccess::fence() function. See that file for the gory details
  58   // on the Memory Access Ordering Model.
  59 
  60   // All of the atomic operations that imply a read-modify-write action
  61   // guarantee a two-way memory barrier across that operation. Historically
  62   // these semantics reflect the strength of atomic operations that are
  63   // provided on SPARC/X86. We assume that strength is necessary unless
  64   // we can prove that a weaker form is sufficiently safe.
  65 
  66   // Atomically store to a location
  67   // The type T must be either a pointer type convertible to or equal
  68   // to D, an integral/enum type equal to D, or a type equal to D that
  69   // is primitive convertible using PrimitiveConversions.
  70   template<typename T, typename D>
  71   inline static void store(T store_value, volatile D* dest);
  72 
  73   // Atomically load from a location
  74   // The type T must be either a pointer type, an integral/enum type,
  75   // or a type that is primitive convertible using PrimitiveConversions.
  76   template<typename T>
  77   inline static T load(const volatile T* dest);
  78 
  79   // Atomically add to a location. Returns updated value. add*() provide:
  80   // <fence> add-value-to-dest <membar StoreLoad|StoreStore>
  81 
  82   template<typename I, typename D>
  83   inline static D add(I add_value, D volatile* dest);
  84 
  85   template<typename I, typename D>
  86   inline static D sub(I sub_value, D volatile* dest);
  87 
  88   // Atomically increment location. inc() provide:
  89   // <fence> increment-dest <membar StoreLoad|StoreStore>
  90   // The type D may be either a pointer type, or an integral
  91   // type. If it is a pointer type, then the increment is
  92   // scaled to the size of the type pointed to by the pointer.
  93   template<typename D>
  94   inline static void inc(D volatile* dest);
  95 
  96   // Atomically decrement a location. dec() provide:
  97   // <fence> decrement-dest <membar StoreLoad|StoreStore>
  98   // The type D may be either a pointer type, or an integral
  99   // type. If it is a pointer type, then the decrement is
 100   // scaled to the size of the type pointed to by the pointer.
 101   template<typename D>
 102   inline static void dec(D volatile* dest);
 103 
 104   // Performs atomic exchange of *dest with exchange_value. Returns old
 105   // prior value of *dest. xchg*() provide:
 106   // <fence> exchange-value-with-dest <membar StoreLoad|StoreStore>
 107   // The type T must be either a pointer type convertible to or equal
 108   // to D, an integral/enum type equal to D, or a type equal to D that
 109   // is primitive convertible using PrimitiveConversions.
 110   template<typename T, typename D>
 111   inline static D xchg(T exchange_value, volatile D* dest);
 112 
 113   // Performs atomic compare of *dest and compare_value, and exchanges
 114   // *dest with exchange_value if the comparison succeeded. Returns prior
 115   // value of *dest. cmpxchg*() provide:
 116   // <fence> compare-and-exchange <membar StoreLoad|StoreStore>
 117 
 118   template<typename T, typename D, typename U>
 119   inline static D cmpxchg(T exchange_value,
 120                           D volatile* dest,
 121                           U compare_value,
 122                           cmpxchg_memory_order order = memory_order_conservative);
 123 
 124   // Performs atomic compare of *dest and NULL, and replaces *dest
 125   // with exchange_value if the comparison succeeded.  Returns true if
 126   // the comparison succeeded and the exchange occurred.  This is
 127   // often used as part of lazy initialization, as a lock-free
 128   // alternative to the Double-Checked Locking Pattern.
 129   template<typename T, typename D>
 130   inline static bool replace_if_null(T* value, D* volatile* dest,
 131                                      cmpxchg_memory_order order = memory_order_conservative);
 132 
 133 private:
 134   // Test whether From is implicitly convertible to To.
 135   // From and To must be pointer types.
 136   // Note: Provides the limited subset of C++11 std::is_convertible
 137   // that is needed here.
 138   template<typename From, typename To> struct IsPointerConvertible;
 139 
 140 protected:
 141   // Dispatch handler for store.  Provides type-based validity
 142   // checking and limited conversions around calls to the platform-
 143   // specific implementation layer provided by PlatformOp.
 144   template<typename T, typename D, typename PlatformOp, typename Enable = void>
 145   struct StoreImpl;
 146 
 147   // Platform-specific implementation of store.  Support for sizes
 148   // of 1, 2, 4, and (if different) pointer size bytes are required.
 149   // The class is a function object that must be default constructable,
 150   // with these requirements:
 151   //
 152   // either:
 153   // - dest is of type D*, an integral, enum or pointer type.
 154   // - new_value are of type T, an integral, enum or pointer type D or
 155   //   pointer type convertible to D.
 156   // or:
 157   // - T and D are the same and are primitive convertible using PrimitiveConversions
 158   // and either way:
 159   // - platform_store is an object of type PlatformStore<sizeof(T)>.
 160   //
 161   // Then
 162   //   platform_store(new_value, dest)
 163   // must be a valid expression.
 164   //
 165   // The default implementation is a volatile store. If a platform
 166   // requires more for e.g. 64 bit stores, a specialization is required
 167   template<size_t byte_size> struct PlatformStore;
 168 
 169   // Dispatch handler for load.  Provides type-based validity
 170   // checking and limited conversions around calls to the platform-
 171   // specific implementation layer provided by PlatformOp.
 172   template<typename T, typename PlatformOp, typename Enable = void>
 173   struct LoadImpl;
 174 
 175   // Platform-specific implementation of load. Support for sizes of
 176   // 1, 2, 4 bytes and (if different) pointer size bytes are required.
 177   // The class is a function object that must be default
 178   // constructable, with these requirements:
 179   //
 180   // - dest is of type T*, an integral, enum or pointer type, or
 181   //   T is convertible to a primitive type using PrimitiveConversions
 182   // - platform_load is an object of type PlatformLoad<sizeof(T)>.
 183   //
 184   // Then
 185   //   platform_load(src)
 186   // must be a valid expression, returning a result convertible to T.
 187   //
 188   // The default implementation is a volatile load. If a platform
 189   // requires more for e.g. 64 bit loads, a specialization is required
 190   template<size_t byte_size> struct PlatformLoad;
 191 
 192 private:
 193   // Dispatch handler for add.  Provides type-based validity checking
 194   // and limited conversions around calls to the platform-specific
 195   // implementation layer provided by PlatformAdd.
 196   template<typename I, typename D, typename Enable = void>
 197   struct AddImpl;
 198 
 199   // Platform-specific implementation of add.  Support for sizes of 4
 200   // bytes and (if different) pointer size bytes are required.  The
 201   // class is a function object that must be default constructable,
 202   // with these requirements:
 203   //
 204   // - dest is of type D*, an integral or pointer type.
 205   // - add_value is of type I, an integral type.
 206   // - sizeof(I) == sizeof(D).
 207   // - if D is an integral type, I == D.
 208   // - platform_add is an object of type PlatformAdd<sizeof(D)>.
 209   //
 210   // Then
 211   //   platform_add(add_value, dest)
 212   // must be a valid expression, returning a result convertible to D.
 213   //
 214   // No definition is provided; all platforms must explicitly define
 215   // this class and any needed specializations.
 216   template<size_t byte_size> struct PlatformAdd;
 217 
 218   // Helper base classes for defining PlatformAdd.  To use, define
 219   // PlatformAdd or a specialization that derives from one of these,
 220   // and include in the PlatformAdd definition the support function
 221   // (described below) required by the base class.
 222   //
 223   // These classes implement the required function object protocol for
 224   // PlatformAdd, using a support function template provided by the
 225   // derived class.  Let add_value (of type I) and dest (of type D) be
 226   // the arguments the object is called with.  If D is a pointer type
 227   // P*, then let addend (of type I) be add_value * sizeof(P);
 228   // otherwise, addend is add_value.
 229   //
 230   // FetchAndAdd requires the derived class to provide
 231   //   fetch_and_add(addend, dest)
 232   // atomically adding addend to the value of dest, and returning the
 233   // old value.
 234   //
 235   // AddAndFetch requires the derived class to provide
 236   //   add_and_fetch(addend, dest)
 237   // atomically adding addend to the value of dest, and returning the
 238   // new value.
 239   //
 240   // When D is a pointer type P*, both fetch_and_add and add_and_fetch
 241   // treat it as if it were a uintptr_t; they do not perform any
 242   // scaling of the addend, as that has already been done by the
 243   // caller.
 244 public: // Temporary, can't be private: C++03 11.4/2. Fixed by C++11.
 245   template<typename Derived> struct FetchAndAdd;
 246   template<typename Derived> struct AddAndFetch;
 247 private:
 248 
 249   // Support for platforms that implement some variants of add using a
 250   // (typically out of line) non-template helper function.  The
 251   // generic arguments passed to PlatformAdd need to be translated to
 252   // the appropriate type for the helper function, the helper function
 253   // invoked on the translated arguments, and the result translated
 254   // back.  Type is the parameter / return type of the helper
 255   // function.  No scaling of add_value is performed when D is a pointer
 256   // type, so this function can be used to implement the support function
 257   // required by AddAndFetch.
 258   template<typename Type, typename Fn, typename I, typename D>
 259   static D add_using_helper(Fn fn, I add_value, D volatile* dest);
 260 
 261   // Dispatch handler for cmpxchg.  Provides type-based validity
 262   // checking and limited conversions around calls to the
 263   // platform-specific implementation layer provided by
 264   // PlatformCmpxchg.
 265   template<typename T, typename D, typename U, typename Enable = void>
 266   struct CmpxchgImpl;
 267 
 268   // Platform-specific implementation of cmpxchg.  Support for sizes
 269   // of 1, 4, and 8 are required.  The class is a function object that
 270   // must be default constructable, with these requirements:
 271   //
 272   // - dest is of type T*.
 273   // - exchange_value and compare_value are of type T.
 274   // - order is of type cmpxchg_memory_order.
 275   // - platform_cmpxchg is an object of type PlatformCmpxchg<sizeof(T)>.
 276   //
 277   // Then
 278   //   platform_cmpxchg(exchange_value, dest, compare_value, order)
 279   // must be a valid expression, returning a result convertible to T.
 280   //
 281   // A default definition is provided, which declares a function template
 282   //   T operator()(T, T volatile*, T, cmpxchg_memory_order) const
 283   //
 284   // For each required size, a platform must either provide an
 285   // appropriate definition of that function, or must entirely
 286   // specialize the class template for that size.
 287   template<size_t byte_size> struct PlatformCmpxchg;
 288 
 289   // Support for platforms that implement some variants of cmpxchg
 290   // using a (typically out of line) non-template helper function.
 291   // The generic arguments passed to PlatformCmpxchg need to be
 292   // translated to the appropriate type for the helper function, the
 293   // helper invoked on the translated arguments, and the result
 294   // translated back.  Type is the parameter / return type of the
 295   // helper function.
 296   template<typename Type, typename Fn, typename T>
 297   static T cmpxchg_using_helper(Fn fn,
 298                                 T exchange_value,
 299                                 T volatile* dest,
 300                                 T compare_value);
 301 
 302   // Support platforms that do not provide Read-Modify-Write
 303   // byte-level atomic access. To use, derive PlatformCmpxchg<1> from
 304   // this class.
 305 public: // Temporary, can't be private: C++03 11.4/2. Fixed by C++11.
 306   struct CmpxchgByteUsingInt;
 307 private:
 308 
 309   // Dispatch handler for xchg.  Provides type-based validity
 310   // checking and limited conversions around calls to the
 311   // platform-specific implementation layer provided by
 312   // PlatformXchg.
 313   template<typename T, typename D, typename Enable = void>
 314   struct XchgImpl;
 315 
 316   // Platform-specific implementation of xchg.  Support for sizes
 317   // of 4, and sizeof(intptr_t) are required.  The class is a function
 318   // object that must be default constructable, with these requirements:
 319   //
 320   // - dest is of type T*.
 321   // - exchange_value is of type T.
 322   // - platform_xchg is an object of type PlatformXchg<sizeof(T)>.
 323   //
 324   // Then
 325   //   platform_xchg(exchange_value, dest)
 326   // must be a valid expression, returning a result convertible to T.
 327   //
 328   // A default definition is provided, which declares a function template
 329   //   T operator()(T, T volatile*, T, cmpxchg_memory_order) const
 330   //
 331   // For each required size, a platform must either provide an
 332   // appropriate definition of that function, or must entirely
 333   // specialize the class template for that size.
 334   template<size_t byte_size> struct PlatformXchg;
 335 
 336   // Support for platforms that implement some variants of xchg
 337   // using a (typically out of line) non-template helper function.
 338   // The generic arguments passed to PlatformXchg need to be
 339   // translated to the appropriate type for the helper function, the
 340   // helper invoked on the translated arguments, and the result
 341   // translated back.  Type is the parameter / return type of the
 342   // helper function.
 343   template<typename Type, typename Fn, typename T>
 344   static T xchg_using_helper(Fn fn,
 345                              T exchange_value,
 346                              T volatile* dest);
 347 };
 348 
 349 template<typename From, typename To>
 350 struct Atomic::IsPointerConvertible<From*, To*> : AllStatic {
 351   // Determine whether From* is implicitly convertible to To*, using
 352   // the "sizeof trick".
 353   typedef char yes;
 354   typedef char (&no)[2];
 355 
 356   static yes test(To*);
 357   static no test(...);
 358   static From* test_value;
 359 
 360   static const bool value = (sizeof(yes) == sizeof(test(test_value)));
 361 };
 362 
 363 // Handle load for pointer, integral and enum types.
 364 template<typename T, typename PlatformOp>
 365 struct Atomic::LoadImpl<
 366   T,
 367   PlatformOp,
 368   typename EnableIf<IsIntegral<T>::value || IsRegisteredEnum<T>::value || IsPointer<T>::value>::type>
 369   VALUE_OBJ_CLASS_SPEC
 370 {
 371   T operator()(T const volatile* dest) const {
 372     // Forward to the platform handler for the size of T.
 373     return PlatformOp()(dest);
 374   }
 375 };
 376 
 377 // Handle load for types that have a translator.
 378 //
 379 // All the involved types must be identical.
 380 //
 381 // This translates the original call into a call on the decayed
 382 // arguments, and returns the recovered result of that translated
 383 // call.
 384 template<typename T, typename PlatformOp>
 385 struct Atomic::LoadImpl<
 386   T,
 387   PlatformOp,
 388   typename EnableIf<PrimitiveConversions::Translate<T>::value>::type>
 389   VALUE_OBJ_CLASS_SPEC
 390 {
 391   T operator()(T const volatile* dest) const {
 392     typedef PrimitiveConversions::Translate<T> Translator;
 393     typedef typename Translator::Decayed Decayed;
 394     STATIC_ASSERT(sizeof(T) == sizeof(Decayed));
 395     Decayed result = PlatformOp()(reinterpret_cast<Decayed const volatile*>(dest));
 396     return Translator::recover(result);
 397   }
 398 };
 399 
 400 // Default implementation of atomic load if a specific platform
 401 // does not provide a specialization for a certain size class.
 402 // For increased safety, the default implementation only allows
 403 // load types that are pointer sized or smaller. If a platform still
 404 // supports wide atomics, then it has to use specialization
 405 // of Atomic::PlatformLoad for that wider size class.
 406 template<size_t byte_size>
 407 struct Atomic::PlatformLoad VALUE_OBJ_CLASS_SPEC {
 408   template<typename T>
 409   T operator()(T const volatile* dest) const {
 410     STATIC_ASSERT(sizeof(T) <= sizeof(void*)); // wide atomics need specialization
 411     return *dest;
 412   }
 413 };
 414 
 415 // Handle store for integral and enum types.
 416 //
 417 // All the involved types must be identical.
 418 template<typename T, typename PlatformOp>
 419 struct Atomic::StoreImpl<
 420   T, T,
 421   PlatformOp,
 422   typename EnableIf<IsIntegral<T>::value || IsRegisteredEnum<T>::value>::type>
 423   VALUE_OBJ_CLASS_SPEC
 424 {
 425   void operator()(T new_value, T volatile* dest) const {
 426     // Forward to the platform handler for the size of T.
 427     PlatformOp()(new_value, dest);
 428   }
 429 };
 430 
 431 // Handle store for pointer types.
 432 //
 433 // The new_value must be implicitly convertible to the
 434 // destination's type; it must be type-correct to store the
 435 // new_value in the destination.
 436 template<typename T, typename D, typename PlatformOp>
 437 struct Atomic::StoreImpl<
 438   T*, D*,
 439   PlatformOp,
 440   typename EnableIf<Atomic::IsPointerConvertible<T*, D*>::value>::type>
 441   VALUE_OBJ_CLASS_SPEC
 442 {
 443   void operator()(T* new_value, D* volatile* dest) const {
 444     // Allow derived to base conversion, and adding cv-qualifiers.
 445     D* value = new_value;
 446     PlatformOp()(value, dest);
 447   }
 448 };
 449 
 450 // Handle store for types that have a translator.
 451 //
 452 // All the involved types must be identical.
 453 //
 454 // This translates the original call into a call on the decayed
 455 // arguments.
 456 template<typename T, typename PlatformOp>
 457 struct Atomic::StoreImpl<
 458   T, T,
 459   PlatformOp,
 460   typename EnableIf<PrimitiveConversions::Translate<T>::value>::type>
 461   VALUE_OBJ_CLASS_SPEC
 462 {
 463   void operator()(T new_value, T volatile* dest) const {
 464     typedef PrimitiveConversions::Translate<T> Translator;
 465     typedef typename Translator::Decayed Decayed;
 466     STATIC_ASSERT(sizeof(T) == sizeof(Decayed));
 467     PlatformOp()(Translator::decay(new_value),
 468                  reinterpret_cast<Decayed volatile*>(dest));
 469   }
 470 };
 471 
 472 // Default implementation of atomic store if a specific platform
 473 // does not provide a specialization for a certain size class.
 474 // For increased safety, the default implementation only allows
 475 // storing types that are pointer sized or smaller. If a platform still
 476 // supports wide atomics, then it has to use specialization
 477 // of Atomic::PlatformStore for that wider size class.
 478 template<size_t byte_size>
 479 struct Atomic::PlatformStore VALUE_OBJ_CLASS_SPEC {
 480   template<typename T>
 481   void operator()(T new_value,
 482                   T volatile* dest) const {
 483     STATIC_ASSERT(sizeof(T) <= sizeof(void*)); // wide atomics need specialization
 484     (void)const_cast<T&>(*dest = new_value);
 485   }
 486 };
 487 
 488 // Define FetchAndAdd and AddAndFetch helper classes before including
 489 // platform file, which may use these as base classes, requiring they
 490 // be complete.
 491 
 492 template<typename Derived>
 493 struct Atomic::FetchAndAdd VALUE_OBJ_CLASS_SPEC {
 494   template<typename I, typename D>
 495   D operator()(I add_value, D volatile* dest) const;
 496 };
 497 
 498 template<typename Derived>
 499 struct Atomic::AddAndFetch VALUE_OBJ_CLASS_SPEC {
 500   template<typename I, typename D>
 501   D operator()(I add_value, D volatile* dest) const;
 502 };
 503 
 504 template<typename D>
 505 inline void Atomic::inc(D volatile* dest) {
 506   STATIC_ASSERT(IsPointer<D>::value || IsIntegral<D>::value);
 507   typedef typename Conditional<IsPointer<D>::value, ptrdiff_t, D>::type I;
 508   Atomic::add(I(1), dest);
 509 }
 510 
 511 template<typename D>
 512 inline void Atomic::dec(D volatile* dest) {
 513   STATIC_ASSERT(IsPointer<D>::value || IsIntegral<D>::value);
 514   typedef typename Conditional<IsPointer<D>::value, ptrdiff_t, D>::type I;
 515   // Assumes two's complement integer representation.
 516   #pragma warning(suppress: 4146)
 517   Atomic::add(I(-1), dest);
 518 }
 519 
 520 template<typename I, typename D>
 521 inline D Atomic::sub(I sub_value, D volatile* dest) {
 522   STATIC_ASSERT(IsPointer<D>::value || IsIntegral<D>::value);
 523   STATIC_ASSERT(IsIntegral<I>::value);
 524   // If D is a pointer type, use [u]intptr_t as the addend type,
 525   // matching signedness of I.  Otherwise, use D as the addend type.
 526   typedef typename Conditional<IsSigned<I>::value, intptr_t, uintptr_t>::type PI;
 527   typedef typename Conditional<IsPointer<D>::value, PI, D>::type AddendType;
 528   // Only allow conversions that can't change the value.
 529   STATIC_ASSERT(IsSigned<I>::value == IsSigned<AddendType>::value); 
 530   STATIC_ASSERT(sizeof(I) <= sizeof(AddendType));
 531   AddendType addend = sub_value;
 532   // Assumes two's complement integer representation.
 533   #pragma warning(suppress: 4146) // In case AddendType is not signed.
 534   return Atomic::add(-addend, dest);
 535 }
 536 
 537 // Define the class before including platform file, which may specialize
 538 // the operator definition.  No generic definition of specializations
 539 // of the operator template are provided, nor are there any generic
 540 // specializations of the class.  The platform file is responsible for
 541 // providing those.
 542 template<size_t byte_size>
 543 struct Atomic::PlatformCmpxchg VALUE_OBJ_CLASS_SPEC {
 544   template<typename T>
 545   T operator()(T exchange_value,
 546                T volatile* dest,
 547                T compare_value,
 548                cmpxchg_memory_order order) const;
 549 };
 550 
 551 // Define the class before including platform file, which may use this
 552 // as a base class, requiring it be complete.  The definition is later
 553 // in this file, near the other definitions related to cmpxchg.
 554 struct Atomic::CmpxchgByteUsingInt VALUE_OBJ_CLASS_SPEC {
 555   template<typename T>
 556   T operator()(T exchange_value,
 557                T volatile* dest,
 558                T compare_value,
 559                cmpxchg_memory_order order) const;
 560 };
 561 
 562 // Define the class before including platform file, which may specialize
 563 // the operator definition.  No generic definition of specializations
 564 // of the operator template are provided, nor are there any generic
 565 // specializations of the class.  The platform file is responsible for
 566 // providing those.
 567 template<size_t byte_size>
 568 struct Atomic::PlatformXchg VALUE_OBJ_CLASS_SPEC {
 569   template<typename T>
 570   T operator()(T exchange_value,
 571                T volatile* dest) const;
 572 };
 573 
 574 // platform specific in-line definitions - must come before shared definitions
 575 
 576 #include OS_CPU_HEADER(atomic)
 577 
 578 // shared in-line definitions
 579 
 580 // size_t casts...
 581 #if (SIZE_MAX != UINTPTR_MAX)
 582 #error size_t is not WORD_SIZE, interesting platform, but missing implementation here
 583 #endif
 584 
 585 template<typename T>
 586 inline T Atomic::load(const volatile T* dest) {
 587   return LoadImpl<T, PlatformLoad<sizeof(T)> >()(dest);
 588 }
 589 
 590 template<typename T, typename D>
 591 inline void Atomic::store(T store_value, volatile D* dest) {
 592   StoreImpl<T, D, PlatformStore<sizeof(D)> >()(store_value, dest);
 593 }
 594 
 595 template<typename I, typename D>
 596 inline D Atomic::add(I add_value, D volatile* dest) {
 597   return AddImpl<I, D>()(add_value, dest);
 598 }
 599 
 600 template<typename I, typename D>
 601 struct Atomic::AddImpl<
 602   I, D,
 603   typename EnableIf<IsIntegral<I>::value &&
 604                     IsIntegral<D>::value &&
 605                     (sizeof(I) <= sizeof(D)) &&
 606                     (IsSigned<I>::value == IsSigned<D>::value)>::type>
 607   VALUE_OBJ_CLASS_SPEC
 608 {
 609   D operator()(I add_value, D volatile* dest) const {
 610     D addend = add_value;
 611     return PlatformAdd<sizeof(D)>()(addend, dest);
 612   }
 613 };
 614 
 615 template<typename I, typename P>
 616 struct Atomic::AddImpl<
 617   I, P*,
 618   typename EnableIf<IsIntegral<I>::value && (sizeof(I) <= sizeof(P*))>::type>
 619   VALUE_OBJ_CLASS_SPEC
 620 {
 621   P* operator()(I add_value, P* volatile* dest) const {
 622     STATIC_ASSERT(sizeof(intptr_t) == sizeof(P*));
 623     STATIC_ASSERT(sizeof(uintptr_t) == sizeof(P*));
 624     typedef typename Conditional<IsSigned<I>::value,
 625                                  intptr_t,
 626                                  uintptr_t>::type CI;
 627     CI addend = add_value;
 628     return PlatformAdd<sizeof(P*)>()(addend, dest);
 629   }
 630 };
 631 
 632 // Most platforms do not support atomic add on a 2-byte value. However,
 633 // if the value occupies the most significant 16 bits of an aligned 32-bit
 634 // word, then we can do this with an atomic add of (add_value << 16)
 635 // to the 32-bit word.
 636 //
 637 // The least significant parts of this 32-bit word will never be affected, even
 638 // in case of overflow/underflow.
 639 //
 640 // Use the ATOMIC_SHORT_PAIR macro (see macros.hpp) to get the desired alignment.
 641 template<>
 642 struct Atomic::AddImpl<jshort, jshort> VALUE_OBJ_CLASS_SPEC {
 643   jshort operator()(jshort add_value, jshort volatile* dest) const {
 644 #ifdef VM_LITTLE_ENDIAN
 645     assert((intx(dest) & 0x03) == 0x02, "wrong alignment");
 646     jint new_value = Atomic::add(add_value << 16, (volatile jint*)(dest-1));
 647 #else
 648     assert((intx(dest) & 0x03) == 0x00, "wrong alignment");
 649     jint new_value = Atomic::add(add_value << 16, (volatile jint*)(dest));
 650 #endif
 651     return (jshort)(new_value >> 16); // preserves sign
 652   }
 653 };
 654 
 655 template<typename Derived>
 656 template<typename I, typename D>
 657 inline D Atomic::FetchAndAdd<Derived>::operator()(I add_value, D volatile* dest) const {
 658   I addend = add_value;
 659   // If D is a pointer type P*, scale by sizeof(P).
 660   if (IsPointer<D>::value) {
 661     addend *= sizeof(typename RemovePointer<D>::type);
 662   }
 663   D old = static_cast<const Derived*>(this)->fetch_and_add(addend, dest);
 664   return old + add_value;
 665 }
 666 
 667 template<typename Derived>
 668 template<typename I, typename D>
 669 inline D Atomic::AddAndFetch<Derived>::operator()(I add_value, D volatile* dest) const {
 670   // If D is a pointer type P*, scale by sizeof(P).
 671   if (IsPointer<D>::value) {
 672     add_value *= sizeof(typename RemovePointer<D>::type);
 673   }
 674   return static_cast<const Derived*>(this)->add_and_fetch(add_value, dest);
 675 }
 676 
 677 template<typename Type, typename Fn, typename I, typename D>
 678 inline D Atomic::add_using_helper(Fn fn, I add_value, D volatile* dest) {
 679   return PrimitiveConversions::cast<D>(
 680     fn(PrimitiveConversions::cast<Type>(add_value),
 681        reinterpret_cast<Type volatile*>(dest)));
 682 }
 683 
 684 template<typename T, typename D, typename U>
 685 inline D Atomic::cmpxchg(T exchange_value,
 686                          D volatile* dest,
 687                          U compare_value,
 688                          cmpxchg_memory_order order) {
 689   return CmpxchgImpl<T, D, U>()(exchange_value, dest, compare_value, order);
 690 }
 691 
 692 template<typename T, typename D>
 693 inline bool Atomic::replace_if_null(T* value, D* volatile* dest,
 694                                     cmpxchg_memory_order order) {
 695   // Presently using a trivial implementation in terms of cmpxchg.
 696   // Consider adding platform support, to permit the use of compiler
 697   // intrinsics like gcc's __sync_bool_compare_and_swap.
 698   D* expected_null = NULL;
 699   return expected_null == cmpxchg(value, dest, expected_null, order);
 700 }
 701 
 702 // Handle cmpxchg for integral and enum types.
 703 //
 704 // All the involved types must be identical.
 705 template<typename T>
 706 struct Atomic::CmpxchgImpl<
 707   T, T, T,
 708   typename EnableIf<IsIntegral<T>::value || IsRegisteredEnum<T>::value>::type>
 709   VALUE_OBJ_CLASS_SPEC
 710 {
 711   T operator()(T exchange_value, T volatile* dest, T compare_value,
 712                cmpxchg_memory_order order) const {
 713     // Forward to the platform handler for the size of T.
 714     return PlatformCmpxchg<sizeof(T)>()(exchange_value,
 715                                         dest,
 716                                         compare_value,
 717                                         order);
 718   }
 719 };
 720 
 721 // Handle cmpxchg for pointer types.
 722 //
 723 // The destination's type and the compare_value type must be the same,
 724 // ignoring cv-qualifiers; we don't care about the cv-qualifiers of
 725 // the compare_value.
 726 //
 727 // The exchange_value must be implicitly convertible to the
 728 // destination's type; it must be type-correct to store the
 729 // exchange_value in the destination.
 730 template<typename T, typename D, typename U>
 731 struct Atomic::CmpxchgImpl<
 732   T*, D*, U*,
 733   typename EnableIf<Atomic::IsPointerConvertible<T*, D*>::value &&
 734                     IsSame<typename RemoveCV<D>::type,
 735                            typename RemoveCV<U>::type>::value>::type>
 736   VALUE_OBJ_CLASS_SPEC
 737 {
 738   D* operator()(T* exchange_value, D* volatile* dest, U* compare_value,
 739                cmpxchg_memory_order order) const {
 740     // Allow derived to base conversion, and adding cv-qualifiers.
 741     D* new_value = exchange_value;
 742     // Don't care what the CV qualifiers for compare_value are,
 743     // but we need to match D* when calling platform support.
 744     D* old_value = const_cast<D*>(compare_value);
 745     return PlatformCmpxchg<sizeof(D*)>()(new_value, dest, old_value, order);
 746   }
 747 };
 748 
 749 // Handle cmpxchg for types that have a translator.
 750 //
 751 // All the involved types must be identical.
 752 //
 753 // This translates the original call into a call on the decayed
 754 // arguments, and returns the recovered result of that translated
 755 // call.
 756 template<typename T>
 757 struct Atomic::CmpxchgImpl<
 758   T, T, T,
 759   typename EnableIf<PrimitiveConversions::Translate<T>::value>::type>
 760   VALUE_OBJ_CLASS_SPEC
 761 {
 762   T operator()(T exchange_value, T volatile* dest, T compare_value,
 763                cmpxchg_memory_order order) const {
 764     typedef PrimitiveConversions::Translate<T> Translator;
 765     typedef typename Translator::Decayed Decayed;
 766     STATIC_ASSERT(sizeof(T) == sizeof(Decayed));
 767     return Translator::recover(
 768       cmpxchg(Translator::decay(exchange_value),
 769               reinterpret_cast<Decayed volatile*>(dest),
 770               Translator::decay(compare_value),
 771               order));
 772   }
 773 };
 774 
 775 template<typename Type, typename Fn, typename T>
 776 inline T Atomic::cmpxchg_using_helper(Fn fn,
 777                                       T exchange_value,
 778                                       T volatile* dest,
 779                                       T compare_value) {
 780   STATIC_ASSERT(sizeof(Type) == sizeof(T));
 781   return PrimitiveConversions::cast<T>(
 782     fn(PrimitiveConversions::cast<Type>(exchange_value),
 783        reinterpret_cast<Type volatile*>(dest),
 784        PrimitiveConversions::cast<Type>(compare_value)));
 785 }
 786 
 787 template<typename T>
 788 inline T Atomic::CmpxchgByteUsingInt::operator()(T exchange_value,
 789                                                  T volatile* dest,
 790                                                  T compare_value,
 791                                                  cmpxchg_memory_order order) const {
 792   STATIC_ASSERT(sizeof(T) == sizeof(uint8_t));
 793   uint8_t canon_exchange_value = exchange_value;
 794   uint8_t canon_compare_value = compare_value;
 795   volatile uint32_t* aligned_dest
 796     = reinterpret_cast<volatile uint32_t*>(align_down(dest, sizeof(uint32_t)));
 797   size_t offset = pointer_delta(dest, aligned_dest, 1);
 798   uint32_t cur = *aligned_dest;
 799   uint8_t* cur_as_bytes = reinterpret_cast<uint8_t*>(&cur);
 800 
 801   // current value may not be what we are looking for, so force it
 802   // to that value so the initial cmpxchg will fail if it is different
 803   cur_as_bytes[offset] = canon_compare_value;
 804 
 805   // always execute a real cmpxchg so that we get the required memory
 806   // barriers even on initial failure
 807   do {
 808     // value to swap in matches current value ...
 809     uint32_t new_value = cur;
 810     // ... except for the one jbyte we want to update
 811     reinterpret_cast<uint8_t*>(&new_value)[offset] = canon_exchange_value;
 812 
 813     uint32_t res = cmpxchg(new_value, aligned_dest, cur, order);
 814     if (res == cur) break;      // success
 815 
 816     // at least one byte in the int changed value, so update
 817     // our view of the current int
 818     cur = res;
 819     // if our byte is still as cur we loop and try again
 820   } while (cur_as_bytes[offset] == canon_compare_value);
 821 
 822   return PrimitiveConversions::cast<T>(cur_as_bytes[offset]);
 823 }
 824 
 825 // Handle xchg for integral and enum types.
 826 //
 827 // All the involved types must be identical.
 828 template<typename T>
 829 struct Atomic::XchgImpl<
 830   T, T,
 831   typename EnableIf<IsIntegral<T>::value || IsRegisteredEnum<T>::value>::type>
 832   VALUE_OBJ_CLASS_SPEC
 833 {
 834   T operator()(T exchange_value, T volatile* dest) const {
 835     // Forward to the platform handler for the size of T.
 836     return PlatformXchg<sizeof(T)>()(exchange_value, dest);
 837   }
 838 };
 839 
 840 // Handle xchg for pointer types.
 841 //
 842 // The exchange_value must be implicitly convertible to the
 843 // destination's type; it must be type-correct to store the
 844 // exchange_value in the destination.
 845 template<typename T, typename D>
 846 struct Atomic::XchgImpl<
 847   T*, D*,
 848   typename EnableIf<Atomic::IsPointerConvertible<T*, D*>::value>::type>
 849   VALUE_OBJ_CLASS_SPEC
 850 {
 851   D* operator()(T* exchange_value, D* volatile* dest) const {
 852     // Allow derived to base conversion, and adding cv-qualifiers.
 853     D* new_value = exchange_value;
 854     return PlatformXchg<sizeof(D*)>()(new_value, dest);
 855   }
 856 };
 857 
 858 // Handle xchg for types that have a translator.
 859 //
 860 // All the involved types must be identical.
 861 //
 862 // This translates the original call into a call on the decayed
 863 // arguments, and returns the recovered result of that translated
 864 // call.
 865 template<typename T>
 866 struct Atomic::XchgImpl<
 867   T, T,
 868   typename EnableIf<PrimitiveConversions::Translate<T>::value>::type>
 869   VALUE_OBJ_CLASS_SPEC
 870 {
 871   T operator()(T exchange_value, T volatile* dest) const {
 872     typedef PrimitiveConversions::Translate<T> Translator;
 873     typedef typename Translator::Decayed Decayed;
 874     STATIC_ASSERT(sizeof(T) == sizeof(Decayed));
 875     return Translator::recover(
 876       xchg(Translator::decay(exchange_value),
 877            reinterpret_cast<Decayed volatile*>(dest)));
 878   }
 879 };
 880 
 881 template<typename Type, typename Fn, typename T>
 882 inline T Atomic::xchg_using_helper(Fn fn,
 883                                    T exchange_value,
 884                                    T volatile* dest) {
 885   STATIC_ASSERT(sizeof(Type) == sizeof(T));
 886   return PrimitiveConversions::cast<T>(
 887     fn(PrimitiveConversions::cast<Type>(exchange_value),
 888        reinterpret_cast<Type volatile*>(dest)));
 889 }
 890 
 891 template<typename T, typename D>
 892 inline D Atomic::xchg(T exchange_value, volatile D* dest) {
 893   return XchgImpl<T, D>()(exchange_value, dest);
 894 }
 895 
 896 #endif // SHARE_VM_RUNTIME_ATOMIC_HPP