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   inline static void store    (jbyte    store_value, jbyte*    dest);
  68   inline static void store    (jshort   store_value, jshort*   dest);
  69   inline static void store    (jint     store_value, jint*     dest);
  70   // See comment above about using jlong atomics on 32-bit platforms
  71   inline static void store    (jlong    store_value, jlong*    dest);
  72   inline static void store_ptr(intptr_t store_value, intptr_t* dest);
  73   inline static void store_ptr(void*    store_value, void*     dest);
  74 
  75   inline static void store    (jbyte    store_value, volatile jbyte*    dest);
  76   inline static void store    (jshort   store_value, volatile jshort*   dest);
  77   inline static void store    (jint     store_value, volatile jint*     dest);
  78   // See comment above about using jlong atomics on 32-bit platforms
  79   inline static void store    (jlong    store_value, volatile jlong*    dest);
  80   inline static void store_ptr(intptr_t store_value, volatile intptr_t* dest);
  81   inline static void store_ptr(void*    store_value, volatile void*     dest);
  82 
  83   // See comment above about using jlong atomics on 32-bit platforms
  84   inline static jlong load(const volatile jlong* src);
  85 
  86   // Atomically add to a location. Returns updated value. add*() provide:
  87   // <fence> add-value-to-dest <membar StoreLoad|StoreStore>
  88 
  89   template<typename I, typename D>
  90   inline static D add(I add_value, D volatile* dest);
  91 
  92   inline static intptr_t add_ptr(intptr_t add_value, volatile intptr_t* dest) {
  93     return add(add_value, dest);
  94   }
  95 
  96   inline static void* add_ptr(intptr_t add_value, volatile void* dest) {
  97     return add(add_value, reinterpret_cast<char* volatile*>(dest));
  98   }
  99 
 100   // Atomically increment location. inc() provide:
 101   // <fence> increment-dest <membar StoreLoad|StoreStore>
 102   // The type D may be either a pointer type, or an integral
 103   // type. If it is a pointer type, then the increment is
 104   // scaled to the size of the type pointed to by the pointer.
 105   template<typename D>
 106   inline static void inc(D volatile* dest);
 107 
 108   // Atomically decrement a location. dec() provide:
 109   // <fence> decrement-dest <membar StoreLoad|StoreStore>
 110   // The type D may be either a pointer type, or an integral
 111   // type. If it is a pointer type, then the decrement is
 112   // scaled to the size of the type pointed to by the pointer.
 113   template<typename D>
 114   inline static void dec(D volatile* dest);
 115 
 116   // Performs atomic exchange of *dest with exchange_value. Returns old
 117   // prior value of *dest. xchg*() provide:
 118   // <fence> exchange-value-with-dest <membar StoreLoad|StoreStore>
 119   // The type T must be either a pointer type convertible to or equal
 120   // to D, an integral/enum type equal to D, or a type equal to D that
 121   // is primitive convertible using PrimitiveConversions.
 122   template<typename T, typename D>
 123   inline static D xchg(T exchange_value, volatile D* dest);
 124 
 125   inline static intptr_t xchg_ptr(intptr_t exchange_value, volatile intptr_t* dest) {
 126     return xchg(exchange_value, dest);
 127   }
 128 
 129   inline static void*    xchg_ptr(void*    exchange_value, volatile void*     dest) {
 130     return xchg(exchange_value, reinterpret_cast<void* volatile*>(dest));
 131   }
 132 
 133   // Performs atomic compare of *dest and compare_value, and exchanges
 134   // *dest with exchange_value if the comparison succeeded. Returns prior
 135   // value of *dest. cmpxchg*() provide:
 136   // <fence> compare-and-exchange <membar StoreLoad|StoreStore>
 137 
 138   template<typename T, typename D, typename U>
 139   inline static D cmpxchg(T exchange_value,
 140                           D volatile* dest,
 141                           U compare_value,
 142                           cmpxchg_memory_order order = memory_order_conservative);
 143 
 144   // Performs atomic compare of *dest and NULL, and replaces *dest
 145   // with exchange_value if the comparison succeeded.  Returns true if
 146   // the comparison succeeded and the exchange occurred.  This is
 147   // often used as part of lazy initialization, as a lock-free
 148   // alternative to the Double-Checked Locking Pattern.
 149   template<typename T, typename D>
 150   inline static bool replace_if_null(T* value, D* volatile* dest,
 151                                      cmpxchg_memory_order order = memory_order_conservative);
 152 
 153   inline static intptr_t cmpxchg_ptr(intptr_t exchange_value,
 154                                      volatile intptr_t* dest,
 155                                      intptr_t compare_value,
 156                                      cmpxchg_memory_order order = memory_order_conservative) {
 157     return cmpxchg(exchange_value, dest, compare_value, order);
 158   }
 159 
 160   inline static void* cmpxchg_ptr(void* exchange_value,
 161                                   volatile void* dest,
 162                                   void* compare_value,
 163                                   cmpxchg_memory_order order = memory_order_conservative) {
 164     return cmpxchg(exchange_value,
 165                    reinterpret_cast<void* volatile*>(dest),
 166                    compare_value,
 167                    order);
 168   }
 169 
 170 private:
 171   // Test whether From is implicitly convertible to To.
 172   // From and To must be pointer types.
 173   // Note: Provides the limited subset of C++11 std::is_convertible
 174   // that is needed here.
 175   template<typename From, typename To> struct IsPointerConvertible;
 176 
 177   // Dispatch handler for add.  Provides type-based validity checking
 178   // and limited conversions around calls to the platform-specific
 179   // implementation layer provided by PlatformAdd.
 180   template<typename I, typename D, typename Enable = void>
 181   struct AddImpl;
 182 
 183   // Platform-specific implementation of add.  Support for sizes of 4
 184   // bytes and (if different) pointer size bytes are required.  The
 185   // class is a function object that must be default constructable,
 186   // with these requirements:
 187   //
 188   // - dest is of type D*, an integral or pointer type.
 189   // - add_value is of type I, an integral type.
 190   // - sizeof(I) == sizeof(D).
 191   // - if D is an integral type, I == D.
 192   // - platform_add is an object of type PlatformAdd<sizeof(D)>.
 193   //
 194   // Then
 195   //   platform_add(add_value, dest)
 196   // must be a valid expression, returning a result convertible to D.
 197   //
 198   // No definition is provided; all platforms must explicitly define
 199   // this class and any needed specializations.
 200   template<size_t byte_size> struct PlatformAdd;
 201 
 202   // Helper base classes for defining PlatformAdd.  To use, define
 203   // PlatformAdd or a specialization that derives from one of these,
 204   // and include in the PlatformAdd definition the support function
 205   // (described below) required by the base class.
 206   //
 207   // These classes implement the required function object protocol for
 208   // PlatformAdd, using a support function template provided by the
 209   // derived class.  Let add_value (of type I) and dest (of type D) be
 210   // the arguments the object is called with.  If D is a pointer type
 211   // P*, then let addend (of type I) be add_value * sizeof(P);
 212   // otherwise, addend is add_value.
 213   //
 214   // FetchAndAdd requires the derived class to provide
 215   //   fetch_and_add(addend, dest)
 216   // atomically adding addend to the value of dest, and returning the
 217   // old value.
 218   //
 219   // AddAndFetch requires the derived class to provide
 220   //   add_and_fetch(addend, dest)
 221   // atomically adding addend to the value of dest, and returning the
 222   // new value.
 223   //
 224   // When D is a pointer type P*, both fetch_and_add and add_and_fetch
 225   // treat it as if it were a uintptr_t; they do not perform any
 226   // scaling of the addend, as that has already been done by the
 227   // caller.
 228 public: // Temporary, can't be private: C++03 11.4/2. Fixed by C++11.
 229   template<typename Derived> struct FetchAndAdd;
 230   template<typename Derived> struct AddAndFetch;
 231 private:
 232 
 233   // Support for platforms that implement some variants of add using a
 234   // (typically out of line) non-template helper function.  The
 235   // generic arguments passed to PlatformAdd need to be translated to
 236   // the appropriate type for the helper function, the helper function
 237   // invoked on the translated arguments, and the result translated
 238   // back.  Type is the parameter / return type of the helper
 239   // function.  No scaling of add_value is performed when D is a pointer
 240   // type, so this function can be used to implement the support function
 241   // required by AddAndFetch.
 242   template<typename Type, typename Fn, typename I, typename D>
 243   static D add_using_helper(Fn fn, I add_value, D volatile* dest);
 244 
 245   // Dispatch handler for cmpxchg.  Provides type-based validity
 246   // checking and limited conversions around calls to the
 247   // platform-specific implementation layer provided by
 248   // PlatformCmpxchg.
 249   template<typename T, typename D, typename U, typename Enable = void>
 250   struct CmpxchgImpl;
 251 
 252   // Platform-specific implementation of cmpxchg.  Support for sizes
 253   // of 1, 4, and 8 are required.  The class is a function object that
 254   // must be default constructable, with these requirements:
 255   //
 256   // - dest is of type T*.
 257   // - exchange_value and compare_value are of type T.
 258   // - order is of type cmpxchg_memory_order.
 259   // - platform_cmpxchg is an object of type PlatformCmpxchg<sizeof(T)>.
 260   //
 261   // Then
 262   //   platform_cmpxchg(exchange_value, dest, compare_value, order)
 263   // must be a valid expression, returning a result convertible to T.
 264   //
 265   // A default definition is provided, which declares a function template
 266   //   T operator()(T, T volatile*, T, cmpxchg_memory_order) const
 267   //
 268   // For each required size, a platform must either provide an
 269   // appropriate definition of that function, or must entirely
 270   // specialize the class template for that size.
 271   template<size_t byte_size> struct PlatformCmpxchg;
 272 
 273   // Support for platforms that implement some variants of cmpxchg
 274   // using a (typically out of line) non-template helper function.
 275   // The generic arguments passed to PlatformCmpxchg need to be
 276   // translated to the appropriate type for the helper function, the
 277   // helper invoked on the translated arguments, and the result
 278   // translated back.  Type is the parameter / return type of the
 279   // helper function.
 280   template<typename Type, typename Fn, typename T>
 281   static T cmpxchg_using_helper(Fn fn,
 282                                 T exchange_value,
 283                                 T volatile* dest,
 284                                 T compare_value);
 285 
 286   // Support platforms that do not provide Read-Modify-Write
 287   // byte-level atomic access. To use, derive PlatformCmpxchg<1> from
 288   // this class.
 289 public: // Temporary, can't be private: C++03 11.4/2. Fixed by C++11.
 290   struct CmpxchgByteUsingInt;
 291 private:
 292 
 293   // Dispatch handler for xchg.  Provides type-based validity
 294   // checking and limited conversions around calls to the
 295   // platform-specific implementation layer provided by
 296   // PlatformXchg.
 297   template<typename T, typename D, typename Enable = void>
 298   struct XchgImpl;
 299 
 300   // Platform-specific implementation of xchg.  Support for sizes
 301   // of 4, and sizeof(intptr_t) are required.  The class is a function
 302   // object that must be default constructable, with these requirements:
 303   //
 304   // - dest is of type T*.
 305   // - exchange_value is of type T.
 306   // - platform_xchg is an object of type PlatformXchg<sizeof(T)>.
 307   //
 308   // Then
 309   //   platform_xchg(exchange_value, dest)
 310   // must be a valid expression, returning a result convertible to T.
 311   //
 312   // A default definition is provided, which declares a function template
 313   //   T operator()(T, T volatile*, T, cmpxchg_memory_order) const
 314   //
 315   // For each required size, a platform must either provide an
 316   // appropriate definition of that function, or must entirely
 317   // specialize the class template for that size.
 318   template<size_t byte_size> struct PlatformXchg;
 319 
 320   // Support for platforms that implement some variants of xchg
 321   // using a (typically out of line) non-template helper function.
 322   // The generic arguments passed to PlatformXchg need to be
 323   // translated to the appropriate type for the helper function, the
 324   // helper invoked on the translated arguments, and the result
 325   // translated back.  Type is the parameter / return type of the
 326   // helper function.
 327   template<typename Type, typename Fn, typename T>
 328   static T xchg_using_helper(Fn fn,
 329                              T exchange_value,
 330                              T volatile* dest);
 331 };
 332 
 333 template<typename From, typename To>
 334 struct Atomic::IsPointerConvertible<From*, To*> : AllStatic {
 335   // Determine whether From* is implicitly convertible to To*, using
 336   // the "sizeof trick".
 337   typedef char yes;
 338   typedef char (&no)[2];
 339 
 340   static yes test(To*);
 341   static no test(...);
 342   static From* test_value;
 343 
 344   static const bool value = (sizeof(yes) == sizeof(test(test_value)));
 345 };
 346 
 347 // Define FetchAndAdd and AddAndFetch helper classes before including
 348 // platform file, which may use these as base classes, requiring they
 349 // be complete.
 350 
 351 template<typename Derived>
 352 struct Atomic::FetchAndAdd VALUE_OBJ_CLASS_SPEC {
 353   template<typename I, typename D>
 354   D operator()(I add_value, D volatile* dest) const;
 355 };
 356 
 357 template<typename Derived>
 358 struct Atomic::AddAndFetch VALUE_OBJ_CLASS_SPEC {
 359   template<typename I, typename D>
 360   D operator()(I add_value, D volatile* dest) const;
 361 };
 362 
 363 template<typename D>
 364 inline void Atomic::inc(D volatile* dest) {
 365   STATIC_ASSERT(IsPointer<D>::value || IsIntegral<D>::value);
 366   typedef typename Conditional<IsPointer<D>::value, ptrdiff_t, D>::type I;
 367   Atomic::add(I(1), dest);
 368 }
 369 
 370 template<typename D>
 371 inline void Atomic::dec(D volatile* dest) {
 372   STATIC_ASSERT(IsPointer<D>::value || IsIntegral<D>::value);
 373   typedef typename Conditional<IsPointer<D>::value, ptrdiff_t, D>::type I;
 374   // Assumes two's complement integer representation.
 375   #pragma warning(suppress: 4146)
 376   Atomic::add(I(-1), dest);
 377 }
 378 
 379 // Define the class before including platform file, which may specialize
 380 // the operator definition.  No generic definition of specializations
 381 // of the operator template are provided, nor are there any generic
 382 // specializations of the class.  The platform file is responsible for
 383 // providing those.
 384 template<size_t byte_size>
 385 struct Atomic::PlatformCmpxchg VALUE_OBJ_CLASS_SPEC {
 386   template<typename T>
 387   T operator()(T exchange_value,
 388                T volatile* dest,
 389                T compare_value,
 390                cmpxchg_memory_order order) const;
 391 };
 392 
 393 // Define the class before including platform file, which may use this
 394 // as a base class, requiring it be complete.  The definition is later
 395 // in this file, near the other definitions related to cmpxchg.
 396 struct Atomic::CmpxchgByteUsingInt VALUE_OBJ_CLASS_SPEC {
 397   template<typename T>
 398   T operator()(T exchange_value,
 399                T volatile* dest,
 400                T compare_value,
 401                cmpxchg_memory_order order) const;
 402 };
 403 
 404 // Define the class before including platform file, which may specialize
 405 // the operator definition.  No generic definition of specializations
 406 // of the operator template are provided, nor are there any generic
 407 // specializations of the class.  The platform file is responsible for
 408 // providing those.
 409 template<size_t byte_size>
 410 struct Atomic::PlatformXchg VALUE_OBJ_CLASS_SPEC {
 411   template<typename T>
 412   T operator()(T exchange_value,
 413                T volatile* dest) const;
 414 };
 415 
 416 // platform specific in-line definitions - must come before shared definitions
 417 
 418 #include OS_CPU_HEADER(atomic)
 419 
 420 // shared in-line definitions
 421 
 422 // size_t casts...
 423 #if (SIZE_MAX != UINTPTR_MAX)
 424 #error size_t is not WORD_SIZE, interesting platform, but missing implementation here
 425 #endif
 426 
 427 template<typename I, typename D>
 428 inline D Atomic::add(I add_value, D volatile* dest) {
 429   return AddImpl<I, D>()(add_value, dest);
 430 }
 431 
 432 template<typename I, typename D>
 433 struct Atomic::AddImpl<
 434   I, D,
 435   typename EnableIf<IsIntegral<I>::value &&
 436                     IsIntegral<D>::value &&
 437                     (sizeof(I) <= sizeof(D)) &&
 438                     (IsSigned<I>::value == IsSigned<D>::value)>::type>
 439   VALUE_OBJ_CLASS_SPEC
 440 {
 441   D operator()(I add_value, D volatile* dest) const {
 442     D addend = add_value;
 443     return PlatformAdd<sizeof(D)>()(addend, dest);
 444   }
 445 };
 446 
 447 template<typename I, typename P>
 448 struct Atomic::AddImpl<
 449   I, P*,
 450   typename EnableIf<IsIntegral<I>::value && (sizeof(I) <= sizeof(P*))>::type>
 451   VALUE_OBJ_CLASS_SPEC
 452 {
 453   P* operator()(I add_value, P* volatile* dest) const {
 454     STATIC_ASSERT(sizeof(intptr_t) == sizeof(P*));
 455     STATIC_ASSERT(sizeof(uintptr_t) == sizeof(P*));
 456     typedef typename Conditional<IsSigned<I>::value,
 457                                  intptr_t,
 458                                  uintptr_t>::type CI;
 459     CI addend = add_value;
 460     return PlatformAdd<sizeof(P*)>()(addend, dest);
 461   }
 462 };
 463 
 464 // Most platforms do not support atomic add on a 2-byte value. However,
 465 // if the value occupies the most significant 16 bits of an aligned 32-bit
 466 // word, then we can do this with an atomic add of (add_value << 16)
 467 // to the 32-bit word.
 468 //
 469 // The least significant parts of this 32-bit word will never be affected, even
 470 // in case of overflow/underflow.
 471 //
 472 // Use the ATOMIC_SHORT_PAIR macro (see macros.hpp) to get the desired alignment.
 473 template<>
 474 struct Atomic::AddImpl<jshort, jshort> VALUE_OBJ_CLASS_SPEC {
 475   jshort operator()(jshort add_value, jshort volatile* dest) const {
 476 #ifdef VM_LITTLE_ENDIAN
 477     assert((intx(dest) & 0x03) == 0x02, "wrong alignment");
 478     jint new_value = Atomic::add(add_value << 16, (volatile jint*)(dest-1));
 479 #else
 480     assert((intx(dest) & 0x03) == 0x00, "wrong alignment");
 481     jint new_value = Atomic::add(add_value << 16, (volatile jint*)(dest));
 482 #endif
 483     return (jshort)(new_value >> 16); // preserves sign
 484   }
 485 };
 486 
 487 template<typename Derived>
 488 template<typename I, typename D>
 489 inline D Atomic::FetchAndAdd<Derived>::operator()(I add_value, D volatile* dest) const {
 490   I addend = add_value;
 491   // If D is a pointer type P*, scale by sizeof(P).
 492   if (IsPointer<D>::value) {
 493     addend *= sizeof(typename RemovePointer<D>::type);
 494   }
 495   D old = static_cast<const Derived*>(this)->fetch_and_add(addend, dest);
 496   return old + add_value;
 497 }
 498 
 499 template<typename Derived>
 500 template<typename I, typename D>
 501 inline D Atomic::AddAndFetch<Derived>::operator()(I add_value, D volatile* dest) const {
 502   // If D is a pointer type P*, scale by sizeof(P).
 503   if (IsPointer<D>::value) {
 504     add_value *= sizeof(typename RemovePointer<D>::type);
 505   }
 506   return static_cast<const Derived*>(this)->add_and_fetch(add_value, dest);
 507 }
 508 
 509 template<typename Type, typename Fn, typename I, typename D>
 510 inline D Atomic::add_using_helper(Fn fn, I add_value, D volatile* dest) {
 511   return PrimitiveConversions::cast<D>(
 512     fn(PrimitiveConversions::cast<Type>(add_value),
 513        reinterpret_cast<Type volatile*>(dest)));
 514 }
 515 
 516 template<typename T, typename D, typename U>
 517 inline D Atomic::cmpxchg(T exchange_value,
 518                          D volatile* dest,
 519                          U compare_value,
 520                          cmpxchg_memory_order order) {
 521   return CmpxchgImpl<T, D, U>()(exchange_value, dest, compare_value, order);
 522 }
 523 
 524 template<typename T, typename D>
 525 inline bool Atomic::replace_if_null(T* value, D* volatile* dest,
 526                                     cmpxchg_memory_order order) {
 527   // Presently using a trivial implementation in terms of cmpxchg.
 528   // Consider adding platform support, to permit the use of compiler
 529   // intrinsics like gcc's __sync_bool_compare_and_swap.
 530   D* expected_null = NULL;
 531   return expected_null == cmpxchg(value, dest, expected_null, order);
 532 }
 533 
 534 // Handle cmpxchg for integral and enum types.
 535 //
 536 // All the involved types must be identical.
 537 template<typename T>
 538 struct Atomic::CmpxchgImpl<
 539   T, T, T,
 540   typename EnableIf<IsIntegral<T>::value || IsRegisteredEnum<T>::value>::type>
 541   VALUE_OBJ_CLASS_SPEC
 542 {
 543   T operator()(T exchange_value, T volatile* dest, T compare_value,
 544                cmpxchg_memory_order order) const {
 545     // Forward to the platform handler for the size of T.
 546     return PlatformCmpxchg<sizeof(T)>()(exchange_value,
 547                                         dest,
 548                                         compare_value,
 549                                         order);
 550   }
 551 };
 552 
 553 // Handle cmpxchg for pointer types.
 554 //
 555 // The destination's type and the compare_value type must be the same,
 556 // ignoring cv-qualifiers; we don't care about the cv-qualifiers of
 557 // the compare_value.
 558 //
 559 // The exchange_value must be implicitly convertible to the
 560 // destination's type; it must be type-correct to store the
 561 // exchange_value in the destination.
 562 template<typename T, typename D, typename U>
 563 struct Atomic::CmpxchgImpl<
 564   T*, D*, U*,
 565   typename EnableIf<Atomic::IsPointerConvertible<T*, D*>::value &&
 566                     IsSame<typename RemoveCV<D>::type,
 567                            typename RemoveCV<U>::type>::value>::type>
 568   VALUE_OBJ_CLASS_SPEC
 569 {
 570   D* operator()(T* exchange_value, D* volatile* dest, U* compare_value,
 571                cmpxchg_memory_order order) const {
 572     // Allow derived to base conversion, and adding cv-qualifiers.
 573     D* new_value = exchange_value;
 574     // Don't care what the CV qualifiers for compare_value are,
 575     // but we need to match D* when calling platform support.
 576     D* old_value = const_cast<D*>(compare_value);
 577     return PlatformCmpxchg<sizeof(D*)>()(new_value, dest, old_value, order);
 578   }
 579 };
 580 
 581 // Handle cmpxchg for types that have a translator.
 582 //
 583 // All the involved types must be identical.
 584 //
 585 // This translates the original call into a call on the decayed
 586 // arguments, and returns the recovered result of that translated
 587 // call.
 588 template<typename T>
 589 struct Atomic::CmpxchgImpl<
 590   T, T, T,
 591   typename EnableIf<PrimitiveConversions::Translate<T>::value>::type>
 592   VALUE_OBJ_CLASS_SPEC
 593 {
 594   T operator()(T exchange_value, T volatile* dest, T compare_value,
 595                cmpxchg_memory_order order) const {
 596     typedef PrimitiveConversions::Translate<T> Translator;
 597     typedef typename Translator::Decayed Decayed;
 598     STATIC_ASSERT(sizeof(T) == sizeof(Decayed));
 599     return Translator::recover(
 600       cmpxchg(Translator::decay(exchange_value),
 601               reinterpret_cast<Decayed volatile*>(dest),
 602               Translator::decay(compare_value),
 603               order));
 604   }
 605 };
 606 
 607 template<typename Type, typename Fn, typename T>
 608 inline T Atomic::cmpxchg_using_helper(Fn fn,
 609                                       T exchange_value,
 610                                       T volatile* dest,
 611                                       T compare_value) {
 612   STATIC_ASSERT(sizeof(Type) == sizeof(T));
 613   return PrimitiveConversions::cast<T>(
 614     fn(PrimitiveConversions::cast<Type>(exchange_value),
 615        reinterpret_cast<Type volatile*>(dest),
 616        PrimitiveConversions::cast<Type>(compare_value)));
 617 }
 618 
 619 template<typename T>
 620 inline T Atomic::CmpxchgByteUsingInt::operator()(T exchange_value,
 621                                                  T volatile* dest,
 622                                                  T compare_value,
 623                                                  cmpxchg_memory_order order) const {
 624   STATIC_ASSERT(sizeof(T) == sizeof(uint8_t));
 625   uint8_t canon_exchange_value = exchange_value;
 626   uint8_t canon_compare_value = compare_value;
 627   volatile uint32_t* aligned_dest
 628     = reinterpret_cast<volatile uint32_t*>(align_down(dest, sizeof(uint32_t)));
 629   size_t offset = pointer_delta(dest, aligned_dest, 1);
 630   uint32_t cur = *aligned_dest;
 631   uint8_t* cur_as_bytes = reinterpret_cast<uint8_t*>(&cur);
 632 
 633   // current value may not be what we are looking for, so force it
 634   // to that value so the initial cmpxchg will fail if it is different
 635   cur_as_bytes[offset] = canon_compare_value;
 636 
 637   // always execute a real cmpxchg so that we get the required memory
 638   // barriers even on initial failure
 639   do {
 640     // value to swap in matches current value ...
 641     uint32_t new_value = cur;
 642     // ... except for the one jbyte we want to update
 643     reinterpret_cast<uint8_t*>(&new_value)[offset] = canon_exchange_value;
 644 
 645     uint32_t res = cmpxchg(new_value, aligned_dest, cur, order);
 646     if (res == cur) break;      // success
 647 
 648     // at least one byte in the int changed value, so update
 649     // our view of the current int
 650     cur = res;
 651     // if our byte is still as cur we loop and try again
 652   } while (cur_as_bytes[offset] == canon_compare_value);
 653 
 654   return PrimitiveConversions::cast<T>(cur_as_bytes[offset]);
 655 }
 656 
 657 // Handle xchg for integral and enum types.
 658 //
 659 // All the involved types must be identical.
 660 template<typename T>
 661 struct Atomic::XchgImpl<
 662   T, T,
 663   typename EnableIf<IsIntegral<T>::value || IsRegisteredEnum<T>::value>::type>
 664   VALUE_OBJ_CLASS_SPEC
 665 {
 666   T operator()(T exchange_value, T volatile* dest) const {
 667     // Forward to the platform handler for the size of T.
 668     return PlatformXchg<sizeof(T)>()(exchange_value, dest);
 669   }
 670 };
 671 
 672 // Handle xchg for pointer types.
 673 //
 674 // The exchange_value must be implicitly convertible to the
 675 // destination's type; it must be type-correct to store the
 676 // exchange_value in the destination.
 677 template<typename T, typename D>
 678 struct Atomic::XchgImpl<
 679   T*, D*,
 680   typename EnableIf<Atomic::IsPointerConvertible<T*, D*>::value>::type>
 681   VALUE_OBJ_CLASS_SPEC
 682 {
 683   D* operator()(T* exchange_value, D* volatile* dest) const {
 684     // Allow derived to base conversion, and adding cv-qualifiers.
 685     D* new_value = exchange_value;
 686     return PlatformXchg<sizeof(D*)>()(new_value, dest);
 687   }
 688 };
 689 
 690 // Handle xchg for types that have a translator.
 691 //
 692 // All the involved types must be identical.
 693 //
 694 // This translates the original call into a call on the decayed
 695 // arguments, and returns the recovered result of that translated
 696 // call.
 697 template<typename T>
 698 struct Atomic::XchgImpl<
 699   T, T,
 700   typename EnableIf<PrimitiveConversions::Translate<T>::value>::type>
 701   VALUE_OBJ_CLASS_SPEC
 702 {
 703   T operator()(T exchange_value, T volatile* dest) const {
 704     typedef PrimitiveConversions::Translate<T> Translator;
 705     typedef typename Translator::Decayed Decayed;
 706     STATIC_ASSERT(sizeof(T) == sizeof(Decayed));
 707     return Translator::recover(
 708       xchg(Translator::decay(exchange_value),
 709            reinterpret_cast<Decayed volatile*>(dest)));
 710   }
 711 };
 712 
 713 template<typename Type, typename Fn, typename T>
 714 inline T Atomic::xchg_using_helper(Fn fn,
 715                                    T exchange_value,
 716                                    T volatile* dest) {
 717   STATIC_ASSERT(sizeof(Type) == sizeof(T));
 718   return PrimitiveConversions::cast<T>(
 719     fn(PrimitiveConversions::cast<Type>(exchange_value),
 720        reinterpret_cast<Type volatile*>(dest)));
 721 }
 722 
 723 template<typename T, typename D>
 724 inline D Atomic::xchg(T exchange_value, volatile D* dest) {
 725   return XchgImpl<T, D>()(exchange_value, dest);
 726 }
 727 
 728 #endif // SHARE_VM_RUNTIME_ATOMIC_HPP