1 /*
   2  * Copyright (c) 1999, 2019, 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_RUNTIME_ATOMIC_HPP
  26 #define SHARE_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 atomic_memory_order {
  41   // The modes that align with C++11 are intended to
  42   // follow the same semantics.
  43   memory_order_relaxed = 0,
  44   memory_order_acquire = 2,
  45   memory_order_release = 3,
  46   memory_order_acq_rel = 4,
  47   // Strong two-way memory barrier.
  48   memory_order_conservative = 8
  49 };
  50 
  51 class Atomic : AllStatic {
  52 public:
  53   // Atomic operations on int64 types are not available on all 32-bit
  54   // platforms. If atomic ops on int64 are defined here they must only
  55   // be used from code that verifies they are available at runtime and
  56   // can provide an alternative action if not - see supports_cx8() for
  57   // a means to test availability.
  58 
  59   // The memory operations that are mentioned with each of the atomic
  60   // function families come from src/share/vm/runtime/orderAccess.hpp,
  61   // e.g., <fence> is described in that file and is implemented by the
  62   // OrderAccess::fence() function. See that file for the gory details
  63   // on the Memory Access Ordering Model.
  64 
  65   // All of the atomic operations that imply a read-modify-write action
  66   // guarantee a two-way memory barrier across that operation. Historically
  67   // these semantics reflect the strength of atomic operations that are
  68   // provided on SPARC/X86. We assume that strength is necessary unless
  69   // we can prove that a weaker form is sufficiently safe.
  70 
  71   // Atomically store to a location
  72   // The type T must be either a pointer type convertible to or equal
  73   // to D, an integral/enum type equal to D, or a type equal to D that
  74   // is primitive convertible using PrimitiveConversions.
  75   template<typename T, typename D>
  76   inline static void store(T store_value, volatile D* dest);
  77 
  78   // Atomically load from a location
  79   // The type T must be either a pointer type, an integral/enum type,
  80   // or a type that is primitive convertible using PrimitiveConversions.
  81   template<typename T>
  82   inline static T load(const volatile T* dest);
  83 
  84   // Atomically add to a location. Returns updated value. add*() provide:
  85   // <fence> add-value-to-dest <membar StoreLoad|StoreStore>
  86 
  87   template<typename I, typename D>
  88   inline static D add(I add_value, D volatile* dest,
  89                       atomic_memory_order order = memory_order_conservative);
  90 
  91   template<typename I, typename D>
  92   inline static D sub(I sub_value, D volatile* dest,
  93                       atomic_memory_order order = memory_order_conservative);
  94 
  95   // Atomically increment location. inc() provide:
  96   // <fence> increment-dest <membar StoreLoad|StoreStore>
  97   // The type D may be either a pointer type, or an integral
  98   // type. If it is a pointer type, then the increment is
  99   // scaled to the size of the type pointed to by the pointer.
 100   template<typename D>
 101   inline static void inc(D volatile* dest,
 102                          atomic_memory_order order = memory_order_conservative);
 103 
 104   // Atomically decrement a location. dec() provide:
 105   // <fence> decrement-dest <membar StoreLoad|StoreStore>
 106   // The type D may be either a pointer type, or an integral
 107   // type. If it is a pointer type, then the decrement is
 108   // scaled to the size of the type pointed to by the pointer.
 109   template<typename D>
 110   inline static void dec(D volatile* dest,
 111                          atomic_memory_order order = memory_order_conservative);
 112 
 113   // Performs atomic exchange of *dest with exchange_value. Returns old
 114   // prior value of *dest. xchg*() provide:
 115   // <fence> exchange-value-with-dest <membar StoreLoad|StoreStore>
 116   // The type T must be either a pointer type convertible to or equal
 117   // to D, an integral/enum type equal to D, or a type equal to D that
 118   // is primitive convertible using PrimitiveConversions.
 119   template<typename T, typename D>
 120   inline static D xchg(T exchange_value, volatile D* dest,
 121                        atomic_memory_order order = memory_order_conservative);
 122 
 123   // Performs atomic compare of *dest and compare_value, and exchanges
 124   // *dest with exchange_value if the comparison succeeded. Returns prior
 125   // value of *dest. cmpxchg*() provide:
 126   // <fence> compare-and-exchange <membar StoreLoad|StoreStore>
 127 
 128   template<typename T, typename D, typename U>
 129   inline static D cmpxchg(T exchange_value,
 130                           D volatile* dest,
 131                           U compare_value,
 132                           atomic_memory_order order = memory_order_conservative);
 133 
 134   // Performs atomic compare of *dest and NULL, and replaces *dest
 135   // with exchange_value if the comparison succeeded.  Returns true if
 136   // the comparison succeeded and the exchange occurred.  This is
 137   // often used as part of lazy initialization, as a lock-free
 138   // alternative to the Double-Checked Locking Pattern.
 139   template<typename T, typename D>
 140   inline static bool replace_if_null(T* value, D* volatile* dest,
 141                                      atomic_memory_order order = memory_order_conservative);
 142 
 143 private:
 144 WINDOWS_ONLY(public:) // VS2017 warns (C2027) use of undefined type if IsPointerConvertible is declared private
 145   // Test whether From is implicitly convertible to To.
 146   // From and To must be pointer types.
 147   // Note: Provides the limited subset of C++11 std::is_convertible
 148   // that is needed here.
 149   template<typename From, typename To> struct IsPointerConvertible;
 150 
 151 protected:
 152   // Dispatch handler for store.  Provides type-based validity
 153   // checking and limited conversions around calls to the platform-
 154   // specific implementation layer provided by PlatformOp.
 155   template<typename T, typename D, typename PlatformOp, typename Enable = void>
 156   struct StoreImpl;
 157 
 158   // Platform-specific implementation of store.  Support for sizes
 159   // of 1, 2, 4, and (if different) pointer size bytes are required.
 160   // The class is a function object that must be default constructable,
 161   // with these requirements:
 162   //
 163   // either:
 164   // - dest is of type D*, an integral, enum or pointer type.
 165   // - new_value are of type T, an integral, enum or pointer type D or
 166   //   pointer type convertible to D.
 167   // or:
 168   // - T and D are the same and are primitive convertible using PrimitiveConversions
 169   // and either way:
 170   // - platform_store is an object of type PlatformStore<sizeof(T)>.
 171   //
 172   // Then
 173   //   platform_store(new_value, dest)
 174   // must be a valid expression.
 175   //
 176   // The default implementation is a volatile store. If a platform
 177   // requires more for e.g. 64 bit stores, a specialization is required
 178   template<size_t byte_size> struct PlatformStore;
 179 
 180   // Dispatch handler for load.  Provides type-based validity
 181   // checking and limited conversions around calls to the platform-
 182   // specific implementation layer provided by PlatformOp.
 183   template<typename T, typename PlatformOp, typename Enable = void>
 184   struct LoadImpl;
 185 
 186   // Platform-specific implementation of load. Support for sizes of
 187   // 1, 2, 4 bytes and (if different) pointer size bytes are required.
 188   // The class is a function object that must be default
 189   // constructable, with these requirements:
 190   //
 191   // - dest is of type T*, an integral, enum or pointer type, or
 192   //   T is convertible to a primitive type using PrimitiveConversions
 193   // - platform_load is an object of type PlatformLoad<sizeof(T)>.
 194   //
 195   // Then
 196   //   platform_load(src)
 197   // must be a valid expression, returning a result convertible to T.
 198   //
 199   // The default implementation is a volatile load. If a platform
 200   // requires more for e.g. 64 bit loads, a specialization is required
 201   template<size_t byte_size> struct PlatformLoad;
 202 
 203 private:
 204   // Dispatch handler for add.  Provides type-based validity checking
 205   // and limited conversions around calls to the platform-specific
 206   // implementation layer provided by PlatformAdd.
 207   template<typename I, typename D, typename Enable = void>
 208   struct AddImpl;
 209 
 210   // Platform-specific implementation of add.  Support for sizes of 4
 211   // bytes and (if different) pointer size bytes are required.  The
 212   // class is a function object that must be default constructable,
 213   // with these requirements:
 214   //
 215   // - dest is of type D*, an integral or pointer type.
 216   // - add_value is of type I, an integral type.
 217   // - sizeof(I) == sizeof(D).
 218   // - if D is an integral type, I == D.
 219   // - platform_add is an object of type PlatformAdd<sizeof(D)>.
 220   //
 221   // Then
 222   //   platform_add(add_value, dest)
 223   // must be a valid expression, returning a result convertible to D.
 224   //
 225   // No definition is provided; all platforms must explicitly define
 226   // this class and any needed specializations.
 227   template<size_t byte_size> struct PlatformAdd;
 228 
 229   // Helper base classes for defining PlatformAdd.  To use, define
 230   // PlatformAdd or a specialization that derives from one of these,
 231   // and include in the PlatformAdd definition the support function
 232   // (described below) required by the base class.
 233   //
 234   // These classes implement the required function object protocol for
 235   // PlatformAdd, using a support function template provided by the
 236   // derived class.  Let add_value (of type I) and dest (of type D) be
 237   // the arguments the object is called with.  If D is a pointer type
 238   // P*, then let addend (of type I) be add_value * sizeof(P);
 239   // otherwise, addend is add_value.
 240   //
 241   // FetchAndAdd requires the derived class to provide
 242   //   fetch_and_add(addend, dest)
 243   // atomically adding addend to the value of dest, and returning the
 244   // old value.
 245   //
 246   // AddAndFetch requires the derived class to provide
 247   //   add_and_fetch(addend, dest)
 248   // atomically adding addend to the value of dest, and returning the
 249   // new value.
 250   //
 251   // When D is a pointer type P*, both fetch_and_add and add_and_fetch
 252   // treat it as if it were a uintptr_t; they do not perform any
 253   // scaling of the addend, as that has already been done by the
 254   // caller.
 255 public: // Temporary, can't be private: C++03 11.4/2. Fixed by C++11.
 256   template<typename Derived> struct FetchAndAdd;
 257   template<typename Derived> struct AddAndFetch;
 258 private:
 259 
 260   // Support for platforms that implement some variants of add using a
 261   // (typically out of line) non-template helper function.  The
 262   // generic arguments passed to PlatformAdd need to be translated to
 263   // the appropriate type for the helper function, the helper function
 264   // invoked on the translated arguments, and the result translated
 265   // back.  Type is the parameter / return type of the helper
 266   // function.  No scaling of add_value is performed when D is a pointer
 267   // type, so this function can be used to implement the support function
 268   // required by AddAndFetch.
 269   template<typename Type, typename Fn, typename I, typename D>
 270   static D add_using_helper(Fn fn, I add_value, D volatile* dest);
 271 
 272   // Dispatch handler for cmpxchg.  Provides type-based validity
 273   // checking and limited conversions around calls to the
 274   // platform-specific implementation layer provided by
 275   // PlatformCmpxchg.
 276   template<typename T, typename D, typename U, typename Enable = void>
 277   struct CmpxchgImpl;
 278 
 279   // Platform-specific implementation of cmpxchg.  Support for sizes
 280   // of 1, 4, and 8 are required.  The class is a function object that
 281   // must be default constructable, with these requirements:
 282   //
 283   // - dest is of type T*.
 284   // - exchange_value and compare_value are of type T.
 285   // - order is of type atomic_memory_order.
 286   // - platform_cmpxchg is an object of type PlatformCmpxchg<sizeof(T)>.
 287   //
 288   // Then
 289   //   platform_cmpxchg(exchange_value, dest, compare_value, order)
 290   // must be a valid expression, returning a result convertible to T.
 291   //
 292   // A default definition is provided, which declares a function template
 293   //   T operator()(T, T volatile*, T, atomic_memory_order) const
 294   //
 295   // For each required size, a platform must either provide an
 296   // appropriate definition of that function, or must entirely
 297   // specialize the class template for that size.
 298   template<size_t byte_size> struct PlatformCmpxchg;
 299 
 300   // Support for platforms that implement some variants of cmpxchg
 301   // using a (typically out of line) non-template helper function.
 302   // The generic arguments passed to PlatformCmpxchg need to be
 303   // translated to the appropriate type for the helper function, the
 304   // helper invoked on the translated arguments, and the result
 305   // translated back.  Type is the parameter / return type of the
 306   // helper function.
 307   template<typename Type, typename Fn, typename T>
 308   static T cmpxchg_using_helper(Fn fn,
 309                                 T exchange_value,
 310                                 T volatile* dest,
 311                                 T compare_value);
 312 
 313   // Support platforms that do not provide Read-Modify-Write
 314   // byte-level atomic access. To use, derive PlatformCmpxchg<1> from
 315   // this class.
 316 public: // Temporary, can't be private: C++03 11.4/2. Fixed by C++11.
 317   struct CmpxchgByteUsingInt;
 318 private:
 319 
 320   // Dispatch handler for xchg.  Provides type-based validity
 321   // checking and limited conversions around calls to the
 322   // platform-specific implementation layer provided by
 323   // PlatformXchg.
 324   template<typename T, typename D, typename Enable = void>
 325   struct XchgImpl;
 326 
 327   // Platform-specific implementation of xchg.  Support for sizes
 328   // of 4, and sizeof(intptr_t) are required.  The class is a function
 329   // object that must be default constructable, with these requirements:
 330   //
 331   // - dest is of type T*.
 332   // - exchange_value is of type T.
 333   // - platform_xchg is an object of type PlatformXchg<sizeof(T)>.
 334   //
 335   // Then
 336   //   platform_xchg(exchange_value, dest)
 337   // must be a valid expression, returning a result convertible to T.
 338   //
 339   // A default definition is provided, which declares a function template
 340   //   T operator()(T, T volatile*, T, atomic_memory_order) const
 341   //
 342   // For each required size, a platform must either provide an
 343   // appropriate definition of that function, or must entirely
 344   // specialize the class template for that size.
 345   template<size_t byte_size> struct PlatformXchg;
 346 
 347   // Support for platforms that implement some variants of xchg
 348   // using a (typically out of line) non-template helper function.
 349   // The generic arguments passed to PlatformXchg need to be
 350   // translated to the appropriate type for the helper function, the
 351   // helper invoked on the translated arguments, and the result
 352   // translated back.  Type is the parameter / return type of the
 353   // helper function.
 354   template<typename Type, typename Fn, typename T>
 355   static T xchg_using_helper(Fn fn,
 356                              T exchange_value,
 357                              T volatile* dest);
 358 };
 359 
 360 template<typename From, typename To>
 361 struct Atomic::IsPointerConvertible<From*, To*> : AllStatic {
 362   // Determine whether From* is implicitly convertible to To*, using
 363   // the "sizeof trick".
 364   typedef char yes;
 365   typedef char (&no)[2];
 366 
 367   static yes test(To*);
 368   static no test(...);
 369   static From* test_value;
 370 
 371   static const bool value = (sizeof(yes) == sizeof(test(test_value)));
 372 };
 373 
 374 // Handle load for pointer, integral and enum types.
 375 template<typename T, typename PlatformOp>
 376 struct Atomic::LoadImpl<
 377   T,
 378   PlatformOp,
 379   typename EnableIf<IsIntegral<T>::value || IsRegisteredEnum<T>::value || IsPointer<T>::value>::type>
 380 {
 381   T operator()(T const volatile* dest) const {
 382     // Forward to the platform handler for the size of T.
 383     return PlatformOp()(dest);
 384   }
 385 };
 386 
 387 // Handle load for types that have a translator.
 388 //
 389 // All the involved types must be identical.
 390 //
 391 // This translates the original call into a call on the decayed
 392 // arguments, and returns the recovered result of that translated
 393 // call.
 394 template<typename T, typename PlatformOp>
 395 struct Atomic::LoadImpl<
 396   T,
 397   PlatformOp,
 398   typename EnableIf<PrimitiveConversions::Translate<T>::value>::type>
 399 {
 400   T operator()(T const volatile* dest) const {
 401     typedef PrimitiveConversions::Translate<T> Translator;
 402     typedef typename Translator::Decayed Decayed;
 403     STATIC_ASSERT(sizeof(T) == sizeof(Decayed));
 404     Decayed result = PlatformOp()(reinterpret_cast<Decayed const volatile*>(dest));
 405     return Translator::recover(result);
 406   }
 407 };
 408 
 409 // Default implementation of atomic load if a specific platform
 410 // does not provide a specialization for a certain size class.
 411 // For increased safety, the default implementation only allows
 412 // load types that are pointer sized or smaller. If a platform still
 413 // supports wide atomics, then it has to use specialization
 414 // of Atomic::PlatformLoad for that wider size class.
 415 template<size_t byte_size>
 416 struct Atomic::PlatformLoad {
 417   template<typename T>
 418   T operator()(T const volatile* dest) const {
 419     STATIC_ASSERT(sizeof(T) <= sizeof(void*)); // wide atomics need specialization
 420     return *dest;
 421   }
 422 };
 423 
 424 // Handle store for integral and enum types.
 425 //
 426 // All the involved types must be identical.
 427 template<typename T, typename PlatformOp>
 428 struct Atomic::StoreImpl<
 429   T, T,
 430   PlatformOp,
 431   typename EnableIf<IsIntegral<T>::value || IsRegisteredEnum<T>::value>::type>
 432 {
 433   void operator()(T new_value, T volatile* dest) const {
 434     // Forward to the platform handler for the size of T.
 435     PlatformOp()(new_value, dest);
 436   }
 437 };
 438 
 439 // Handle store for pointer types.
 440 //
 441 // The new_value must be implicitly convertible to the
 442 // destination's type; it must be type-correct to store the
 443 // new_value in the destination.
 444 template<typename T, typename D, typename PlatformOp>
 445 struct Atomic::StoreImpl<
 446   T*, D*,
 447   PlatformOp,
 448   typename EnableIf<Atomic::IsPointerConvertible<T*, D*>::value>::type>
 449 {
 450   void operator()(T* new_value, D* volatile* dest) const {
 451     // Allow derived to base conversion, and adding cv-qualifiers.
 452     D* value = new_value;
 453     PlatformOp()(value, dest);
 454   }
 455 };
 456 
 457 // Handle store for types that have a translator.
 458 //
 459 // All the involved types must be identical.
 460 //
 461 // This translates the original call into a call on the decayed
 462 // arguments.
 463 template<typename T, typename PlatformOp>
 464 struct Atomic::StoreImpl<
 465   T, T,
 466   PlatformOp,
 467   typename EnableIf<PrimitiveConversions::Translate<T>::value>::type>
 468 {
 469   void operator()(T new_value, T volatile* dest) const {
 470     typedef PrimitiveConversions::Translate<T> Translator;
 471     typedef typename Translator::Decayed Decayed;
 472     STATIC_ASSERT(sizeof(T) == sizeof(Decayed));
 473     PlatformOp()(Translator::decay(new_value),
 474                  reinterpret_cast<Decayed volatile*>(dest));
 475   }
 476 };
 477 
 478 // Default implementation of atomic store if a specific platform
 479 // does not provide a specialization for a certain size class.
 480 // For increased safety, the default implementation only allows
 481 // storing types that are pointer sized or smaller. If a platform still
 482 // supports wide atomics, then it has to use specialization
 483 // of Atomic::PlatformStore for that wider size class.
 484 template<size_t byte_size>
 485 struct Atomic::PlatformStore {
 486   template<typename T>
 487   void operator()(T new_value,
 488                   T volatile* dest) const {
 489     STATIC_ASSERT(sizeof(T) <= sizeof(void*)); // wide atomics need specialization
 490     (void)const_cast<T&>(*dest = new_value);
 491   }
 492 };
 493 
 494 // Define FetchAndAdd and AddAndFetch helper classes before including
 495 // platform file, which may use these as base classes, requiring they
 496 // be complete.
 497 
 498 template<typename Derived>
 499 struct Atomic::FetchAndAdd {
 500   template<typename I, typename D>
 501   D operator()(I add_value, D volatile* dest, atomic_memory_order order) const;
 502 };
 503 
 504 template<typename Derived>
 505 struct Atomic::AddAndFetch {
 506   template<typename I, typename D>
 507   D operator()(I add_value, D volatile* dest, atomic_memory_order order) const;
 508 };
 509 
 510 template<typename D>
 511 inline void Atomic::inc(D volatile* dest, atomic_memory_order order) {
 512   STATIC_ASSERT(IsPointer<D>::value || IsIntegral<D>::value);
 513   typedef typename Conditional<IsPointer<D>::value, ptrdiff_t, D>::type I;
 514   Atomic::add(I(1), dest, order);
 515 }
 516 
 517 template<typename D>
 518 inline void Atomic::dec(D volatile* dest, atomic_memory_order order) {
 519   STATIC_ASSERT(IsPointer<D>::value || IsIntegral<D>::value);
 520   typedef typename Conditional<IsPointer<D>::value, ptrdiff_t, D>::type I;
 521   // Assumes two's complement integer representation.
 522   #pragma warning(suppress: 4146)
 523   Atomic::add(I(-1), dest, order);
 524 }
 525 
 526 template<typename I, typename D>
 527 inline D Atomic::sub(I sub_value, D volatile* dest, atomic_memory_order order) {
 528   STATIC_ASSERT(IsPointer<D>::value || IsIntegral<D>::value);
 529   STATIC_ASSERT(IsIntegral<I>::value);
 530   // If D is a pointer type, use [u]intptr_t as the addend type,
 531   // matching signedness of I.  Otherwise, use D as the addend type.
 532   typedef typename Conditional<IsSigned<I>::value, intptr_t, uintptr_t>::type PI;
 533   typedef typename Conditional<IsPointer<D>::value, PI, D>::type AddendType;
 534   // Only allow conversions that can't change the value.
 535   STATIC_ASSERT(IsSigned<I>::value == IsSigned<AddendType>::value);
 536   STATIC_ASSERT(sizeof(I) <= sizeof(AddendType));
 537   AddendType addend = sub_value;
 538   // Assumes two's complement integer representation.
 539   #pragma warning(suppress: 4146) // In case AddendType is not signed.
 540   return Atomic::add(-addend, dest, order);
 541 }
 542 
 543 // Define the class before including platform file, which may specialize
 544 // the operator definition.  No generic definition of specializations
 545 // of the operator template are provided, nor are there any generic
 546 // specializations of the class.  The platform file is responsible for
 547 // providing those.
 548 template<size_t byte_size>
 549 struct Atomic::PlatformCmpxchg {
 550   template<typename T>
 551   T operator()(T exchange_value,
 552                T volatile* dest,
 553                T compare_value,
 554                atomic_memory_order order) const;
 555 };
 556 
 557 // Define the class before including platform file, which may use this
 558 // as a base class, requiring it be complete.  The definition is later
 559 // in this file, near the other definitions related to cmpxchg.
 560 struct Atomic::CmpxchgByteUsingInt {
 561   template<typename T>
 562   T operator()(T exchange_value,
 563                T volatile* dest,
 564                T compare_value,
 565                atomic_memory_order order) const;
 566 };
 567 
 568 // Define the class before including platform file, which may specialize
 569 // the operator definition.  No generic definition of specializations
 570 // of the operator template are provided, nor are there any generic
 571 // specializations of the class.  The platform file is responsible for
 572 // providing those.
 573 template<size_t byte_size>
 574 struct Atomic::PlatformXchg {
 575   template<typename T>
 576   T operator()(T exchange_value,
 577                T volatile* dest,
 578                atomic_memory_order order) const;
 579 };
 580 
 581 // platform specific in-line definitions - must come before shared definitions
 582 
 583 #include OS_CPU_HEADER(atomic)
 584 
 585 // shared in-line definitions
 586 
 587 // size_t casts...
 588 #if (SIZE_MAX != UINTPTR_MAX)
 589 #error size_t is not WORD_SIZE, interesting platform, but missing implementation here
 590 #endif
 591 
 592 template<typename T>
 593 inline T Atomic::load(const volatile T* dest) {
 594   return LoadImpl<T, PlatformLoad<sizeof(T)> >()(dest);
 595 }
 596 
 597 template<typename T, typename D>
 598 inline void Atomic::store(T store_value, volatile D* dest) {
 599   StoreImpl<T, D, PlatformStore<sizeof(D)> >()(store_value, dest);
 600 }
 601 
 602 template<typename I, typename D>
 603 inline D Atomic::add(I add_value, D volatile* dest,
 604                      atomic_memory_order order) {
 605   return AddImpl<I, D>()(add_value, dest, order);
 606 }
 607 
 608 template<typename I, typename D>
 609 struct Atomic::AddImpl<
 610   I, D,
 611   typename EnableIf<IsIntegral<I>::value &&
 612                     IsIntegral<D>::value &&
 613                     (sizeof(I) <= sizeof(D)) &&
 614                     (IsSigned<I>::value == IsSigned<D>::value)>::type>
 615 {
 616   D operator()(I add_value, D volatile* dest, atomic_memory_order order) const {
 617     D addend = add_value;
 618     return PlatformAdd<sizeof(D)>()(addend, dest, order);
 619   }
 620 };
 621 
 622 template<typename I, typename P>
 623 struct Atomic::AddImpl<
 624   I, P*,
 625   typename EnableIf<IsIntegral<I>::value && (sizeof(I) <= sizeof(P*))>::type>
 626 {
 627   P* operator()(I add_value, P* volatile* dest, atomic_memory_order order) const {
 628     STATIC_ASSERT(sizeof(intptr_t) == sizeof(P*));
 629     STATIC_ASSERT(sizeof(uintptr_t) == sizeof(P*));
 630     typedef typename Conditional<IsSigned<I>::value,
 631                                  intptr_t,
 632                                  uintptr_t>::type CI;
 633     CI addend = add_value;
 634     return PlatformAdd<sizeof(P*)>()(addend, dest, order);
 635   }
 636 };
 637 
 638 template<typename Derived>
 639 template<typename I, typename D>
 640 inline D Atomic::FetchAndAdd<Derived>::operator()(I add_value, D volatile* dest,
 641                                                   atomic_memory_order order) const {
 642   I addend = add_value;
 643   // If D is a pointer type P*, scale by sizeof(P).
 644   if (IsPointer<D>::value) {
 645     addend *= sizeof(typename RemovePointer<D>::type);
 646   }
 647   D old = static_cast<const Derived*>(this)->fetch_and_add(addend, dest, order);
 648   return old + add_value;
 649 }
 650 
 651 template<typename Derived>
 652 template<typename I, typename D>
 653 inline D Atomic::AddAndFetch<Derived>::operator()(I add_value, D volatile* dest,
 654                                                   atomic_memory_order order) const {
 655   // If D is a pointer type P*, scale by sizeof(P).
 656   if (IsPointer<D>::value) {
 657     add_value *= sizeof(typename RemovePointer<D>::type);
 658   }
 659   return static_cast<const Derived*>(this)->add_and_fetch(add_value, dest, order);
 660 }
 661 
 662 template<typename Type, typename Fn, typename I, typename D>
 663 inline D Atomic::add_using_helper(Fn fn, I add_value, D volatile* dest) {
 664   return PrimitiveConversions::cast<D>(
 665     fn(PrimitiveConversions::cast<Type>(add_value),
 666        reinterpret_cast<Type volatile*>(dest)));
 667 }
 668 
 669 template<typename T, typename D, typename U>
 670 inline D Atomic::cmpxchg(T exchange_value,
 671                          D volatile* dest,
 672                          U compare_value,
 673                          atomic_memory_order order) {
 674   return CmpxchgImpl<T, D, U>()(exchange_value, dest, compare_value, order);
 675 }
 676 
 677 template<typename T, typename D>
 678 inline bool Atomic::replace_if_null(T* value, D* volatile* dest,
 679                                     atomic_memory_order order) {
 680   // Presently using a trivial implementation in terms of cmpxchg.
 681   // Consider adding platform support, to permit the use of compiler
 682   // intrinsics like gcc's __sync_bool_compare_and_swap.
 683   D* expected_null = NULL;
 684   return expected_null == cmpxchg(value, dest, expected_null, order);
 685 }
 686 
 687 // Handle cmpxchg for integral and enum types.
 688 //
 689 // All the involved types must be identical.
 690 template<typename T>
 691 struct Atomic::CmpxchgImpl<
 692   T, T, T,
 693   typename EnableIf<IsIntegral<T>::value || IsRegisteredEnum<T>::value>::type>
 694 {
 695   T operator()(T exchange_value, T volatile* dest, T compare_value,
 696                atomic_memory_order order) const {
 697     // Forward to the platform handler for the size of T.
 698     return PlatformCmpxchg<sizeof(T)>()(exchange_value,
 699                                         dest,
 700                                         compare_value,
 701                                         order);
 702   }
 703 };
 704 
 705 // Handle cmpxchg for pointer types.
 706 //
 707 // The destination's type and the compare_value type must be the same,
 708 // ignoring cv-qualifiers; we don't care about the cv-qualifiers of
 709 // the compare_value.
 710 //
 711 // The exchange_value must be implicitly convertible to the
 712 // destination's type; it must be type-correct to store the
 713 // exchange_value in the destination.
 714 template<typename T, typename D, typename U>
 715 struct Atomic::CmpxchgImpl<
 716   T*, D*, U*,
 717   typename EnableIf<Atomic::IsPointerConvertible<T*, D*>::value &&
 718                     IsSame<typename RemoveCV<D>::type,
 719                            typename RemoveCV<U>::type>::value>::type>
 720 {
 721   D* operator()(T* exchange_value, D* volatile* dest, U* compare_value,
 722                atomic_memory_order order) const {
 723     // Allow derived to base conversion, and adding cv-qualifiers.
 724     D* new_value = exchange_value;
 725     // Don't care what the CV qualifiers for compare_value are,
 726     // but we need to match D* when calling platform support.
 727     D* old_value = const_cast<D*>(compare_value);
 728     return PlatformCmpxchg<sizeof(D*)>()(new_value, dest, old_value, order);
 729   }
 730 };
 731 
 732 // Handle cmpxchg for types that have a translator.
 733 //
 734 // All the involved types must be identical.
 735 //
 736 // This translates the original call into a call on the decayed
 737 // arguments, and returns the recovered result of that translated
 738 // call.
 739 template<typename T>
 740 struct Atomic::CmpxchgImpl<
 741   T, T, T,
 742   typename EnableIf<PrimitiveConversions::Translate<T>::value>::type>
 743 {
 744   T operator()(T exchange_value, T volatile* dest, T compare_value,
 745                atomic_memory_order order) const {
 746     typedef PrimitiveConversions::Translate<T> Translator;
 747     typedef typename Translator::Decayed Decayed;
 748     STATIC_ASSERT(sizeof(T) == sizeof(Decayed));
 749     return Translator::recover(
 750       cmpxchg(Translator::decay(exchange_value),
 751               reinterpret_cast<Decayed volatile*>(dest),
 752               Translator::decay(compare_value),
 753               order));
 754   }
 755 };
 756 
 757 template<typename Type, typename Fn, typename T>
 758 inline T Atomic::cmpxchg_using_helper(Fn fn,
 759                                       T exchange_value,
 760                                       T volatile* dest,
 761                                       T compare_value) {
 762   STATIC_ASSERT(sizeof(Type) == sizeof(T));
 763   return PrimitiveConversions::cast<T>(
 764     fn(PrimitiveConversions::cast<Type>(exchange_value),
 765        reinterpret_cast<Type volatile*>(dest),
 766        PrimitiveConversions::cast<Type>(compare_value)));
 767 }
 768 
 769 template<typename T>
 770 inline T Atomic::CmpxchgByteUsingInt::operator()(T exchange_value,
 771                                                  T volatile* dest,
 772                                                  T compare_value,
 773                                                  atomic_memory_order order) const {
 774   STATIC_ASSERT(sizeof(T) == sizeof(uint8_t));
 775   uint8_t canon_exchange_value = exchange_value;
 776   uint8_t canon_compare_value = compare_value;
 777   volatile uint32_t* aligned_dest
 778     = reinterpret_cast<volatile uint32_t*>(align_down(dest, sizeof(uint32_t)));
 779   size_t offset = pointer_delta(dest, aligned_dest, 1);
 780   uint32_t cur = *aligned_dest;
 781   uint8_t* cur_as_bytes = reinterpret_cast<uint8_t*>(&cur);
 782 
 783   // current value may not be what we are looking for, so force it
 784   // to that value so the initial cmpxchg will fail if it is different
 785   cur_as_bytes[offset] = canon_compare_value;
 786 
 787   // always execute a real cmpxchg so that we get the required memory
 788   // barriers even on initial failure
 789   do {
 790     // value to swap in matches current value ...
 791     uint32_t new_value = cur;
 792     // ... except for the one byte we want to update
 793     reinterpret_cast<uint8_t*>(&new_value)[offset] = canon_exchange_value;
 794 
 795     uint32_t res = cmpxchg(new_value, aligned_dest, cur, order);
 796     if (res == cur) break;      // success
 797 
 798     // at least one byte in the int changed value, so update
 799     // our view of the current int
 800     cur = res;
 801     // if our byte is still as cur we loop and try again
 802   } while (cur_as_bytes[offset] == canon_compare_value);
 803 
 804   return PrimitiveConversions::cast<T>(cur_as_bytes[offset]);
 805 }
 806 
 807 // Handle xchg for integral and enum types.
 808 //
 809 // All the involved types must be identical.
 810 template<typename T>
 811 struct Atomic::XchgImpl<
 812   T, T,
 813   typename EnableIf<IsIntegral<T>::value || IsRegisteredEnum<T>::value>::type>
 814 {
 815   T operator()(T exchange_value, T volatile* dest, atomic_memory_order order) const {
 816     // Forward to the platform handler for the size of T.
 817     return PlatformXchg<sizeof(T)>()(exchange_value, dest, order);
 818   }
 819 };
 820 
 821 // Handle xchg for pointer types.
 822 //
 823 // The exchange_value must be implicitly convertible to the
 824 // destination's type; it must be type-correct to store the
 825 // exchange_value in the destination.
 826 template<typename T, typename D>
 827 struct Atomic::XchgImpl<
 828   T*, D*,
 829   typename EnableIf<Atomic::IsPointerConvertible<T*, D*>::value>::type>
 830 {
 831   D* operator()(T* exchange_value, D* volatile* dest, atomic_memory_order order) const {
 832     // Allow derived to base conversion, and adding cv-qualifiers.
 833     D* new_value = exchange_value;
 834     return PlatformXchg<sizeof(D*)>()(new_value, dest, order);
 835   }
 836 };
 837 
 838 // Handle xchg for types that have a translator.
 839 //
 840 // All the involved types must be identical.
 841 //
 842 // This translates the original call into a call on the decayed
 843 // arguments, and returns the recovered result of that translated
 844 // call.
 845 template<typename T>
 846 struct Atomic::XchgImpl<
 847   T, T,
 848   typename EnableIf<PrimitiveConversions::Translate<T>::value>::type>
 849 {
 850   T operator()(T exchange_value, T volatile* dest, atomic_memory_order order) const {
 851     typedef PrimitiveConversions::Translate<T> Translator;
 852     typedef typename Translator::Decayed Decayed;
 853     STATIC_ASSERT(sizeof(T) == sizeof(Decayed));
 854     return Translator::recover(
 855       xchg(Translator::decay(exchange_value),
 856            reinterpret_cast<Decayed volatile*>(dest),
 857            order));
 858   }
 859 };
 860 
 861 template<typename Type, typename Fn, typename T>
 862 inline T Atomic::xchg_using_helper(Fn fn,
 863                                    T exchange_value,
 864                                    T volatile* dest) {
 865   STATIC_ASSERT(sizeof(Type) == sizeof(T));
 866   return PrimitiveConversions::cast<T>(
 867     fn(PrimitiveConversions::cast<Type>(exchange_value),
 868        reinterpret_cast<Type volatile*>(dest)));
 869 }
 870 
 871 template<typename T, typename D>
 872 inline D Atomic::xchg(T exchange_value, volatile D* dest, atomic_memory_order order) {
 873   return XchgImpl<T, D>()(exchange_value, dest, order);
 874 }
 875 
 876 #endif // SHARE_RUNTIME_ATOMIC_HPP